add isl_aff_mod_val
[platform/upstream/isl.git] / isl_transitive_closure.c
index c55b979..70d15b9 100644 (file)
 /*
  * Copyright 2010      INRIA Saclay
  *
- * 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, INRIA Saclay - Ile-de-France,
  * Parc Club Orsay Universite, ZAC des vignes, 4 rue Jacques Monod,
  * 91893 Orsay, France 
  */
 
-#include "isl_map.h"
-#include "isl_map_private.h"
-#include "isl_seq.h"
+#include <isl_ctx_private.h>
+#include <isl_map_private.h>
+#include <isl/map.h>
+#include <isl/seq.h>
+#include <isl_space_private.h>
+#include <isl/lp.h>
+#include <isl/union_map.h>
+#include <isl_mat_private.h>
+#include <isl_options_private.h>
+#include <isl_tarjan.h>
+
+int isl_map_is_transitively_closed(__isl_keep isl_map *map)
+{
+       isl_map *map2;
+       int closed;
+
+       map2 = isl_map_apply_range(isl_map_copy(map), isl_map_copy(map));
+       closed = isl_map_is_subset(map2, map);
+       isl_map_free(map2);
+
+       return closed;
+}
+
+int isl_union_map_is_transitively_closed(__isl_keep isl_union_map *umap)
+{
+       isl_union_map *umap2;
+       int closed;
+
+       umap2 = isl_union_map_apply_range(isl_union_map_copy(umap),
+                                         isl_union_map_copy(umap));
+       closed = isl_union_map_is_subset(umap2, umap);
+       isl_union_map_free(umap2);
+
+       return closed;
+}
  
+/* Given a map that represents a path with the length of the path
+ * encoded as the difference between the last output coordindate
+ * and the last input coordinate, set this length to either
+ * exactly "length" (if "exactly" is set) or at least "length"
+ * (if "exactly" is not set).
+ */
+static __isl_give isl_map *set_path_length(__isl_take isl_map *map,
+       int exactly, int length)
+{
+       isl_space *dim;
+       struct isl_basic_map *bmap;
+       unsigned d;
+       unsigned nparam;
+       int k;
+       isl_int *c;
+
+       if (!map)
+               return NULL;
+
+       dim = isl_map_get_space(map);
+       d = isl_space_dim(dim, isl_dim_in);
+       nparam = isl_space_dim(dim, isl_dim_param);
+       bmap = isl_basic_map_alloc_space(dim, 0, 1, 1);
+       if (exactly) {
+               k = isl_basic_map_alloc_equality(bmap);
+               c = bmap->eq[k];
+       } else {
+               k = isl_basic_map_alloc_inequality(bmap);
+               c = bmap->ineq[k];
+       }
+       if (k < 0)
+               goto error;
+       isl_seq_clr(c, 1 + isl_basic_map_total_dim(bmap));
+       isl_int_set_si(c[0], -length);
+       isl_int_set_si(c[1 + nparam + d - 1], -1);
+       isl_int_set_si(c[1 + nparam + d + d - 1], 1);
+
+       bmap = isl_basic_map_finalize(bmap);
+       map = isl_map_intersect(map, isl_map_from_basic_map(bmap));
+
+       return map;
+error:
+       isl_basic_map_free(bmap);
+       isl_map_free(map);
+       return NULL;
+}
+
+/* Check whether the overapproximation of the power of "map" is exactly
+ * the power of "map".  Let R be "map" and A_k the overapproximation.
+ * The approximation is exact if
+ *
+ *     A_1 = R
+ *     A_k = A_{k-1} \circ R                   k >= 2
+ *
+ * Since A_k is known to be an overapproximation, we only need to check
+ *
+ *     A_1 \subset R
+ *     A_k \subset A_{k-1} \circ R             k >= 2
+ *
+ * In practice, "app" has an extra input and output coordinate
+ * to encode the length of the path.  So, we first need to add
+ * this coordinate to "map" and set the length of the path to
+ * one.
+ */
+static int check_power_exactness(__isl_take isl_map *map,
+       __isl_take isl_map *app)
+{
+       int exact;
+       isl_map *app_1;
+       isl_map *app_2;
+
+       map = isl_map_add_dims(map, isl_dim_in, 1);
+       map = isl_map_add_dims(map, isl_dim_out, 1);
+       map = set_path_length(map, 1, 1);
+
+       app_1 = set_path_length(isl_map_copy(app), 1, 1);
+
+       exact = isl_map_is_subset(app_1, map);
+       isl_map_free(app_1);
+
+       if (!exact || exact < 0) {
+               isl_map_free(app);
+               isl_map_free(map);
+               return exact;
+       }
+
+       app_1 = set_path_length(isl_map_copy(app), 0, 1);
+       app_2 = set_path_length(app, 0, 2);
+       app_1 = isl_map_apply_range(map, app_1);
+
+       exact = isl_map_is_subset(app_2, app_1);
+
+       isl_map_free(app_1);
+       isl_map_free(app_2);
+
+       return exact;
+}
+
+/* Check whether the overapproximation of the power of "map" is exactly
+ * the power of "map", possibly after projecting out the power (if "project"
+ * is set).
+ *
+ * If "project" is set and if "steps" can only result in acyclic paths,
+ * then we check
+ *
+ *     A = R \cup (A \circ R)
+ *
+ * where A is the overapproximation with the power projected out, i.e.,
+ * an overapproximation of the transitive closure.
+ * More specifically, since A is known to be an overapproximation, we check
+ *
+ *     A \subset R \cup (A \circ R)
+ *
+ * Otherwise, we check if the power is exact.
+ *
+ * Note that "app" has an extra input and output coordinate to encode
+ * the length of the part.  If we are only interested in the transitive
+ * closure, then we can simply project out these coordinates first.
+ */
+static int check_exactness(__isl_take isl_map *map, __isl_take isl_map *app,
+       int project)
+{
+       isl_map *test;
+       int exact;
+       unsigned d;
+
+       if (!project)
+               return check_power_exactness(map, app);
+
+       d = isl_map_dim(map, isl_dim_in);
+       app = set_path_length(app, 0, 1);
+       app = isl_map_project_out(app, isl_dim_in, d, 1);
+       app = isl_map_project_out(app, isl_dim_out, d, 1);
+
+       app = isl_map_reset_space(app, isl_map_get_space(map));
+
+       test = isl_map_apply_range(isl_map_copy(map), isl_map_copy(app));
+       test = isl_map_union(test, isl_map_copy(map));
+
+       exact = isl_map_is_subset(app, test);
+
+       isl_map_free(app);
+       isl_map_free(test);
+
+       isl_map_free(map);
+
+       return exact;
+}
+
 /*
  * The transitive closure implementation is based on the paper
  * "Computing the Transitive Closure of a Union of Affine Integer
  * For any element in this relation, the number of steps taken
  * is equal to the difference in the final coordinates.
  */
-static __isl_give isl_map *path_along_steps(__isl_take isl_dim *dim,
+static __isl_give isl_map *path_along_steps(__isl_take isl_space *dim,
        __isl_keep isl_mat *steps)
 {
        int i, j, k;
@@ -43,11 +224,11 @@ static __isl_give isl_map *path_along_steps(__isl_take isl_dim *dim,
        if (!dim || !steps)
                goto error;
 
-       d = isl_dim_size(dim, isl_dim_in);
+       d = isl_space_dim(dim, isl_dim_in);
        n = steps->n_row;
-       nparam = isl_dim_size(dim, isl_dim_param);
+       nparam = isl_space_dim(dim, isl_dim_param);
 
-       path = isl_basic_map_alloc_dim(isl_dim_copy(dim), n, d, n);
+       path = isl_basic_map_alloc_space(isl_space_copy(dim), n, d, n);
 
        for (i = 0; i < n; ++i) {
                k = isl_basic_map_alloc_div(path);
@@ -81,13 +262,13 @@ static __isl_give isl_map *path_along_steps(__isl_take isl_dim *dim,
                isl_int_set_si(path->ineq[k][1 + nparam + 2 * d + i], 1);
        }
 
-       isl_dim_free(dim);
+       isl_space_free(dim);
 
        path = isl_basic_map_simplify(path);
        path = isl_basic_map_finalize(path);
        return isl_map_from_basic_map(path);
 error:
-       isl_dim_free(dim);
+       isl_space_free(dim);
        isl_basic_map_free(path);
        return NULL;
 }
@@ -95,28 +276,231 @@ error:
 #define IMPURE         0
 #define PURE_PARAM     1
 #define PURE_VAR       2
+#define MIXED          3
+
+/* Check whether the parametric constant term of constraint c is never
+ * positive in "bset".
+ */
+static int parametric_constant_never_positive(__isl_keep isl_basic_set *bset,
+       isl_int *c, int *div_purity)
+{
+       unsigned d;
+       unsigned n_div;
+       unsigned nparam;
+       int i;
+       int k;
+       int empty;
+
+       n_div = isl_basic_set_dim(bset, isl_dim_div);
+       d = isl_basic_set_dim(bset, isl_dim_set);
+       nparam = isl_basic_set_dim(bset, isl_dim_param);
+
+       bset = isl_basic_set_copy(bset);
+       bset = isl_basic_set_cow(bset);
+       bset = isl_basic_set_extend_constraints(bset, 0, 1);
+       k = isl_basic_set_alloc_inequality(bset);
+       if (k < 0)
+               goto error;
+       isl_seq_clr(bset->ineq[k], 1 + isl_basic_set_total_dim(bset));
+       isl_seq_cpy(bset->ineq[k], c, 1 + nparam);
+       for (i = 0; i < n_div; ++i) {
+               if (div_purity[i] != PURE_PARAM)
+                       continue;
+               isl_int_set(bset->ineq[k][1 + nparam + d + i],
+                           c[1 + nparam + d + i]);
+       }
+       isl_int_sub_ui(bset->ineq[k][0], bset->ineq[k][0], 1);
+       empty = isl_basic_set_is_empty(bset);
+       isl_basic_set_free(bset);
+
+       return empty;
+error:
+       isl_basic_set_free(bset);
+       return -1;
+}
 
 /* Return PURE_PARAM if only the coefficients of the parameters are non-zero.
  * Return PURE_VAR if only the coefficients of the set variables are non-zero.
+ * Return MIXED if only the coefficients of the parameters and the set
+ *     variables are non-zero and if moreover the parametric constant
+ *     can never attain positive values.
  * Return IMPURE otherwise.
+ *
+ * If div_purity is NULL then we are dealing with a non-parametric set
+ * and so the constraint is obviously PURE_VAR.
  */
-static int purity(__isl_keep isl_basic_set *bset, isl_int *c)
+static int purity(__isl_keep isl_basic_set *bset, isl_int *c, int *div_purity,
+       int eq)
 {
        unsigned d;
        unsigned n_div;
        unsigned nparam;
+       int empty;
+       int i;
+       int p = 0, v = 0;
+
+       if (!div_purity)
+               return PURE_VAR;
 
        n_div = isl_basic_set_dim(bset, isl_dim_div);
        d = isl_basic_set_dim(bset, isl_dim_set);
        nparam = isl_basic_set_dim(bset, isl_dim_param);
 
-       if (isl_seq_first_non_zero(c + 1 + nparam + d, n_div) != -1)
-               return IMPURE;
-       if (isl_seq_first_non_zero(c + 1, nparam) == -1)
+       for (i = 0; i < n_div; ++i) {
+               if (isl_int_is_zero(c[1 + nparam + d + i]))
+                       continue;
+               switch (div_purity[i]) {
+               case PURE_PARAM: p = 1; break;
+               case PURE_VAR: v = 1; break;
+               default: return IMPURE;
+               }
+       }
+       if (!p && isl_seq_first_non_zero(c + 1, nparam) == -1)
                return PURE_VAR;
-       if (isl_seq_first_non_zero(c + 1 + nparam, d) == -1)
+       if (!v && isl_seq_first_non_zero(c + 1 + nparam, d) == -1)
                return PURE_PARAM;
-       return IMPURE;
+
+       empty = parametric_constant_never_positive(bset, c, div_purity);
+       if (eq && empty >= 0 && !empty) {
+               isl_seq_neg(c, c, 1 + nparam + d + n_div);
+               empty = parametric_constant_never_positive(bset, c, div_purity);
+       }
+
+       return empty < 0 ? -1 : empty ? MIXED : IMPURE;
+}
+
+/* Return an array of integers indicating the type of each div in bset.
+ * If the div is (recursively) defined in terms of only the parameters,
+ * then the type is PURE_PARAM.
+ * If the div is (recursively) defined in terms of only the set variables,
+ * then the type is PURE_VAR.
+ * Otherwise, the type is IMPURE.
+ */
+static __isl_give int *get_div_purity(__isl_keep isl_basic_set *bset)
+{
+       int i, j;
+       int *div_purity;
+       unsigned d;
+       unsigned n_div;
+       unsigned nparam;
+
+       if (!bset)
+               return NULL;
+
+       n_div = isl_basic_set_dim(bset, isl_dim_div);
+       d = isl_basic_set_dim(bset, isl_dim_set);
+       nparam = isl_basic_set_dim(bset, isl_dim_param);
+
+       div_purity = isl_alloc_array(bset->ctx, int, n_div);
+       if (!div_purity)
+               return NULL;
+
+       for (i = 0; i < bset->n_div; ++i) {
+               int p = 0, v = 0;
+               if (isl_int_is_zero(bset->div[i][0])) {
+                       div_purity[i] = IMPURE;
+                       continue;
+               }
+               if (isl_seq_first_non_zero(bset->div[i] + 2, nparam) != -1)
+                       p = 1;
+               if (isl_seq_first_non_zero(bset->div[i] + 2 + nparam, d) != -1)
+                       v = 1;
+               for (j = 0; j < i; ++j) {
+                       if (isl_int_is_zero(bset->div[i][2 + nparam + d + j]))
+                               continue;
+                       switch (div_purity[j]) {
+                       case PURE_PARAM: p = 1; break;
+                       case PURE_VAR: v = 1; break;
+                       default: p = v = 1; break;
+                       }
+               }
+               div_purity[i] = v ? p ? IMPURE : PURE_VAR : PURE_PARAM;
+       }
+
+       return div_purity;
+}
+
+/* Given a path with the as yet unconstrained length at position "pos",
+ * check if setting the length to zero results in only the identity
+ * mapping.
+ */
+static int empty_path_is_identity(__isl_keep isl_basic_map *path, unsigned pos)
+{
+       isl_basic_map *test = NULL;
+       isl_basic_map *id = NULL;
+       int k;
+       int is_id;
+
+       test = isl_basic_map_copy(path);
+       test = isl_basic_map_extend_constraints(test, 1, 0);
+       k = isl_basic_map_alloc_equality(test);
+       if (k < 0)
+               goto error;
+       isl_seq_clr(test->eq[k], 1 + isl_basic_map_total_dim(test));
+       isl_int_set_si(test->eq[k][pos], 1);
+       id = isl_basic_map_identity(isl_basic_map_get_space(path));
+       is_id = isl_basic_map_is_equal(test, id);
+       isl_basic_map_free(test);
+       isl_basic_map_free(id);
+       return is_id;
+error:
+       isl_basic_map_free(test);
+       return -1;
+}
+
+/* If any of the constraints is found to be impure then this function
+ * sets *impurity to 1.
+ */
+static __isl_give isl_basic_map *add_delta_constraints(
+       __isl_take isl_basic_map *path,
+       __isl_keep isl_basic_set *delta, unsigned off, unsigned nparam,
+       unsigned d, int *div_purity, int eq, int *impurity)
+{
+       int i, k;
+       int n = eq ? delta->n_eq : delta->n_ineq;
+       isl_int **delta_c = eq ? delta->eq : delta->ineq;
+       unsigned n_div;
+
+       n_div = isl_basic_set_dim(delta, isl_dim_div);
+
+       for (i = 0; i < n; ++i) {
+               isl_int *path_c;
+               int p = purity(delta, delta_c[i], div_purity, eq);
+               if (p < 0)
+                       goto error;
+               if (p != PURE_VAR && p != PURE_PARAM && !*impurity)
+                       *impurity = 1;
+               if (p == IMPURE)
+                       continue;
+               if (eq && p != MIXED) {
+                       k = isl_basic_map_alloc_equality(path);
+                       path_c = path->eq[k];
+               } else {
+                       k = isl_basic_map_alloc_inequality(path);
+                       path_c = path->ineq[k];
+               }
+               if (k < 0)
+                       goto error;
+               isl_seq_clr(path_c, 1 + isl_basic_map_total_dim(path));
+               if (p == PURE_VAR) {
+                       isl_seq_cpy(path_c + off,
+                                   delta_c[i] + 1 + nparam, d);
+                       isl_int_set(path_c[off + d], delta_c[i][0]);
+               } else if (p == PURE_PARAM) {
+                       isl_seq_cpy(path_c, delta_c[i], 1 + nparam);
+               } else {
+                       isl_seq_cpy(path_c + off,
+                                   delta_c[i] + 1 + nparam, d);
+                       isl_seq_cpy(path_c, delta_c[i], 1 + nparam);
+               }
+               isl_seq_cpy(path_c + off - n_div,
+                           delta_c[i] + 1 + nparam + d, n_div);
+       }
+
+       return path;
+error:
+       isl_basic_map_free(path);
+       return NULL;
 }
 
 /* Given a set of offsets "delta", construct a relation of the
@@ -135,19 +519,46 @@ static int purity(__isl_keep isl_basic_set *bset, isl_int *c)
  *
  * In particular, let delta be defined as
  *
- *     \delta = [p] -> { [x] : A x + a >= and B p + b >= 0 and
- *                             C x + C'p + c >= 0 }
+ *     \delta = [p] -> { [x] : A x + a >= 0 and B p + b >= 0 and
+ *                             C x + C'p + c >= 0 and
+ *                             D x + D'p + d >= 0 }
  *
- * then the relation is constructed as
+ * where the constraints C x + C'p + c >= 0 are such that the parametric
+ * constant term of each constraint j, "C_j x + C'_j p + c_j",
+ * can never attain positive values, then the relation is constructed as
  *
  *     { [x] -> [y] : exists [f, k] \in Z^{n+1} : y = x + f and
- *                     A f + k a >= 0 and B p + b >= 0 and k >= 1 }
+ *                     A f + k a >= 0 and B p + b >= 0 and
+ *                     C f + C'p + c >= 0 and k >= 1 }
  *     union { [x] -> [x] }
  *
- * Existentially quantified variables in \delta are currently ignored.
+ * If the zero-length paths happen to correspond exactly to the identity
+ * mapping, then we return
+ *
+ *     { [x] -> [y] : exists [f, k] \in Z^{n+1} : y = x + f and
+ *                     A f + k a >= 0 and B p + b >= 0 and
+ *                     C f + C'p + c >= 0 and k >= 0 }
+ *
+ * instead.
+ *
+ * Existentially quantified variables in \delta are handled by
+ * classifying them as independent of the parameters, purely
+ * parameter dependent and others.  Constraints containing
+ * any of the other existentially quantified variables are removed.
  * This is safe, but leads to an additional overapproximation.
+ *
+ * If there are any impure constraints, then we also eliminate
+ * the parameters from \delta, resulting in a set
+ *
+ *     \delta' = { [x] : E x + e >= 0 }
+ *
+ * and add the constraints
+ *
+ *                     E f + k e >= 0
+ *
+ * to the constructed relation.
  */
-static __isl_give isl_map *path_along_delta(__isl_take isl_dim *dim,
+static __isl_give isl_map *path_along_delta(__isl_take isl_space *dim,
        __isl_take isl_basic_set *delta)
 {
        isl_basic_map *path = NULL;
@@ -156,14 +567,17 @@ static __isl_give isl_map *path_along_delta(__isl_take isl_dim *dim,
        unsigned nparam;
        unsigned off;
        int i, k;
+       int is_id;
+       int *div_purity = NULL;
+       int impurity = 0;
 
        if (!delta)
                goto error;
        n_div = isl_basic_set_dim(delta, isl_dim_div);
        d = isl_basic_set_dim(delta, isl_dim_set);
        nparam = isl_basic_set_dim(delta, isl_dim_param);
-       path = isl_basic_map_alloc_dim(isl_dim_copy(dim), n_div + d + 1,
-                       d + 1 + delta->n_eq, delta->n_ineq + 1);
+       path = isl_basic_map_alloc_space(isl_space_copy(dim), n_div + d + 1,
+                       d + 1 + delta->n_eq, delta->n_eq + delta->n_ineq + 1);
        off = 1 + nparam + 2 * (d + 1) + n_div;
 
        for (i = 0; i < n_div + d + 1; ++i) {
@@ -183,64 +597,67 @@ static __isl_give isl_map *path_along_delta(__isl_take isl_dim *dim,
                isl_int_set_si(path->eq[k][off + i], 1);
        }
 
-       for (i = 0; i < delta->n_eq; ++i) {
-               int p = purity(delta, delta->eq[i]);
-               if (p == IMPURE)
-                       continue;
-               k = isl_basic_map_alloc_equality(path);
-               if (k < 0)
-                       goto error;
-               isl_seq_clr(path->eq[k], 1 + isl_basic_map_total_dim(path));
-               if (p == PURE_VAR) {
-                       isl_seq_cpy(path->eq[k] + off,
-                                   delta->eq[i] + 1 + nparam, d);
-                       isl_int_set(path->eq[k][off + d], delta->eq[i][0]);
-               } else
-                       isl_seq_cpy(path->eq[k], delta->eq[i], 1 + nparam);
-       }
+       div_purity = get_div_purity(delta);
+       if (!div_purity)
+               goto error;
 
-       for (i = 0; i < delta->n_ineq; ++i) {
-               int p = purity(delta, delta->ineq[i]);
-               if (p == IMPURE)
-                       continue;
-               k = isl_basic_map_alloc_inequality(path);
-               if (k < 0)
+       path = add_delta_constraints(path, delta, off, nparam, d,
+                                    div_purity, 1, &impurity);
+       path = add_delta_constraints(path, delta, off, nparam, d,
+                                    div_purity, 0, &impurity);
+       if (impurity) {
+               isl_space *dim = isl_basic_set_get_space(delta);
+               delta = isl_basic_set_project_out(delta,
+                                                 isl_dim_param, 0, nparam);
+               delta = isl_basic_set_add_dims(delta, isl_dim_param, nparam);
+               delta = isl_basic_set_reset_space(delta, dim);
+               if (!delta)
                        goto error;
-               isl_seq_clr(path->ineq[k], 1 + isl_basic_map_total_dim(path));
-               if (p == PURE_VAR) {
-                       isl_seq_cpy(path->ineq[k] + off,
-                                   delta->ineq[i] + 1 + nparam, d);
-                       isl_int_set(path->ineq[k][off + d], delta->ineq[i][0]);
-               } else
-                       isl_seq_cpy(path->ineq[k], delta->ineq[i], 1 + nparam);
+               path = isl_basic_map_extend_constraints(path, delta->n_eq,
+                                                       delta->n_ineq + 1);
+               path = add_delta_constraints(path, delta, off, nparam, d,
+                                            NULL, 1, &impurity);
+               path = add_delta_constraints(path, delta, off, nparam, d,
+                                            NULL, 0, &impurity);
+               path = isl_basic_map_gauss(path, NULL);
        }
 
+       is_id = empty_path_is_identity(path, off + d);
+       if (is_id < 0)
+               goto error;
+
        k = isl_basic_map_alloc_inequality(path);
        if (k < 0)
                goto error;
        isl_seq_clr(path->ineq[k], 1 + isl_basic_map_total_dim(path));
-       isl_int_set_si(path->ineq[k][0], -1);
+       if (!is_id)
+               isl_int_set_si(path->ineq[k][0], -1);
        isl_int_set_si(path->ineq[k][off + d], 1);
                        
+       free(div_purity);
        isl_basic_set_free(delta);
        path = isl_basic_map_finalize(path);
-       return isl_basic_map_union(path,
-                               isl_basic_map_identity(isl_dim_domain(dim)));
+       if (is_id) {
+               isl_space_free(dim);
+               return isl_map_from_basic_map(path);
+       }
+       return isl_basic_map_union(path, isl_basic_map_identity(dim));
 error:
-       isl_dim_free(dim);
+       free(div_purity);
+       isl_space_free(dim);
        isl_basic_set_free(delta);
        isl_basic_map_free(path);
        return NULL;
 }
 
-/* Given a dimenion specification Z^{n+1} -> Z^{n+1} and a parameter "param",
+/* Given a dimension specification Z^{n+1} -> Z^{n+1} and a parameter "param",
  * construct a map that equates the parameter to the difference
  * in the final coordinates and imposes that this difference is positive.
  * That is, construct
  *
  *     { [x,x_s] -> [y,y_s] : k = y_s - x_s > 0 }
  */
-static __isl_give isl_map *equate_parameter_to_length(__isl_take isl_dim *dim,
+static __isl_give isl_map *equate_parameter_to_length(__isl_take isl_space *dim,
        unsigned param)
 {
        struct isl_basic_map *bmap;
@@ -248,9 +665,9 @@ static __isl_give isl_map *equate_parameter_to_length(__isl_take isl_dim *dim,
        unsigned nparam;
        int k;
 
-       d = isl_dim_size(dim, isl_dim_in);
-       nparam = isl_dim_size(dim, isl_dim_param);
-       bmap = isl_basic_map_alloc_dim(dim, 0, 1, 1);
+       d = isl_space_dim(dim, isl_dim_in);
+       nparam = isl_space_dim(dim, isl_dim_param);
+       bmap = isl_basic_map_alloc_space(dim, 0, 1, 1);
        k = isl_basic_map_alloc_equality(bmap);
        if (k < 0)
                goto error;
@@ -329,7 +746,7 @@ static int is_acyclic(__isl_take isl_map *path)
  * Since each of these paths performs an addition, composition is
  * symmetric and we can simply compose all resulting paths in any order.
  */
-static __isl_give isl_map *construct_extended_path(__isl_take isl_dim *dim,
+static __isl_give isl_map *construct_extended_path(__isl_take isl_space *dim,
        __isl_keep isl_map *map, int *project)
 {
        struct isl_mat *steps = NULL;
@@ -339,7 +756,7 @@ static __isl_give isl_map *construct_extended_path(__isl_take isl_dim *dim,
 
        d = isl_map_dim(map, isl_dim_in);
 
-       path = isl_map_identity(isl_dim_domain(isl_dim_copy(dim)));
+       path = isl_map_identity(isl_space_copy(dim));
 
        steps = isl_mat_alloc(map->ctx, map->n, d);
        if (!steps)
@@ -354,7 +771,7 @@ static __isl_give isl_map *construct_extended_path(__isl_take isl_dim *dim,
                for (j = 0; j < d; ++j) {
                        int fixed;
 
-                       fixed = isl_basic_set_fast_dim_is_fixed(delta, j,
+                       fixed = isl_basic_set_plain_dim_is_fixed(delta, j,
                                                            &steps->row[n][j]);
                        if (fixed < 0) {
                                isl_basic_set_free(delta);
@@ -367,7 +784,8 @@ static __isl_give isl_map *construct_extended_path(__isl_take isl_dim *dim,
 
                if (j < d) {
                        path = isl_map_apply_range(path,
-                               path_along_delta(isl_dim_copy(dim), delta));
+                               path_along_delta(isl_space_copy(dim), delta));
+                       path = isl_map_coalesce(path);
                } else {
                        isl_basic_set_free(delta);
                        ++n;
@@ -377,7 +795,7 @@ static __isl_give isl_map *construct_extended_path(__isl_take isl_dim *dim,
        if (n > 0) {
                steps->n_row = n;
                path = isl_map_apply_range(path,
-                               path_along_steps(isl_dim_copy(dim), steps));
+                               path_along_steps(isl_space_copy(dim), steps));
        }
 
        if (project && *project) {
@@ -386,23 +804,39 @@ static __isl_give isl_map *construct_extended_path(__isl_take isl_dim *dim,
                        goto error;
        }
 
-       isl_dim_free(dim);
+       isl_space_free(dim);
        isl_mat_free(steps);
        return path;
 error:
-       isl_dim_free(dim);
+       isl_space_free(dim);
        isl_mat_free(steps);
        isl_map_free(path);
        return NULL;
 }
 
-/* Given a union of basic maps R = \cup_i R_i \subseteq D \times D,
+static int isl_set_overlaps(__isl_keep isl_set *set1, __isl_keep isl_set *set2)
+{
+       isl_set *i;
+       int no_overlap;
+
+       if (!isl_space_tuple_match(set1->dim, isl_dim_set, set2->dim, isl_dim_set))
+               return 0;
+
+       i = isl_set_intersect(isl_set_copy(set1), isl_set_copy(set2));
+       no_overlap = isl_set_is_empty(i);
+       isl_set_free(i);
+
+       return no_overlap < 0 ? -1 : !no_overlap;
+}
+
+/* Given a union of basic maps R = \cup_i R_i \subseteq D \times D
+ * and a dimension specification (Z^{n+1} -> Z^{n+1}),
  * construct a map that is an overapproximation of the map
- * that takes an element from the space D to another
- * element from the same space, such that the difference between
- * them is a strictly positive sum of differences between images
- * and pre-images in one of the R_i.
- * The number of differences in the sum is equated to parameter "param".
+ * that takes an element from the dom R \times Z to an
+ * element from ran R \times Z, such that the first n coordinates of the
+ * difference between them is a sum of differences between images
+ * and pre-images in one of the R_i and such that the last coordinate
+ * is equal to the number of steps taken.
  * That is, let
  *
  *     \Delta_i = { y - x | (x, y) in R_i }
@@ -410,273 +844,2135 @@ error:
  * then the constructed map is an overapproximation of
  *
  *     { (x) -> (x + d) | \exists k_i >= 0, \delta_i \in \Delta_i :
- *                             d = \sum_i k_i \delta_i and k = \sum_i k_i > 0 }
- *
- * We first construct an extended mapping with an extra coordinate
- * that indicates the number of steps taken.  In particular,
- * the difference in the last coordinate is equal to the number
- * of steps taken to move from a domain element to the corresponding
- * image element(s).
- * In the final step, this difference is equated to the parameter "param"
- * and made positive.  The extra coordinates are subsequently projected out.
+ *                             d = (\sum_i k_i \delta_i, \sum_i k_i) and
+ *                             x in dom R and x + d in ran R and
+ *                             \sum_i k_i >= 1 }
  */
-static __isl_give isl_map *construct_path(__isl_keep isl_map *map,
-       unsigned param, int *project)
+static __isl_give isl_map *construct_component(__isl_take isl_space *dim,
+       __isl_keep isl_map *map, int *exact, int project)
 {
-       struct isl_map *path = NULL;
-       struct isl_map *diff;
-       struct isl_dim *dim = NULL;
-       unsigned d;
-
-       if (!map)
-               return NULL;
-
-       dim = isl_map_get_dim(map);
+       struct isl_set *domain = NULL;
+       struct isl_set *range = NULL;
+       struct isl_map *app = NULL;
+       struct isl_map *path = NULL;
 
-       d = isl_dim_size(dim, isl_dim_in);
-       dim = isl_dim_add(dim, isl_dim_in, 1);
-       dim = isl_dim_add(dim, isl_dim_out, 1);
+       domain = isl_map_domain(isl_map_copy(map));
+       domain = isl_set_coalesce(domain);
+       range = isl_map_range(isl_map_copy(map));
+       range = isl_set_coalesce(range);
+       if (!isl_set_overlaps(domain, range)) {
+               isl_set_free(domain);
+               isl_set_free(range);
+               isl_space_free(dim);
+
+               map = isl_map_copy(map);
+               map = isl_map_add_dims(map, isl_dim_in, 1);
+               map = isl_map_add_dims(map, isl_dim_out, 1);
+               map = set_path_length(map, 1, 1);
+               return map;
+       }
+       app = isl_map_from_domain_and_range(domain, range);
+       app = isl_map_add_dims(app, isl_dim_in, 1);
+       app = isl_map_add_dims(app, isl_dim_out, 1);
 
-       path = construct_extended_path(isl_dim_copy(dim), map, project);
+       path = construct_extended_path(isl_space_copy(dim), map,
+                                       exact && *exact ? &project : NULL);
+       app = isl_map_intersect(app, path);
 
-       diff = equate_parameter_to_length(dim, param);
-       path = isl_map_intersect(path, diff);
-       path = isl_map_project_out(path, isl_dim_in, d, 1);
-       path = isl_map_project_out(path, isl_dim_out, d, 1);
+       if (exact && *exact &&
+           (*exact = check_exactness(isl_map_copy(map), isl_map_copy(app),
+                                     project)) < 0)
+               goto error;
 
-       return path;
+       isl_space_free(dim);
+       app = set_path_length(app, 0, 1);
+       return app;
+error:
+       isl_space_free(dim);
+       isl_map_free(app);
+       return NULL;
 }
 
-/* Shift variable at position "pos" up by one.
- * That is, replace the corresponding variable v by v - 1.
+/* Call construct_component and, if "project" is set, project out
+ * the final coordinates.
  */
-static __isl_give isl_basic_map *basic_map_shift_pos(
-       __isl_take isl_basic_map *bmap, unsigned pos)
+static __isl_give isl_map *construct_projected_component(
+       __isl_take isl_space *dim,
+       __isl_keep isl_map *map, int *exact, int project)
 {
-       int i;
+       isl_map *app;
+       unsigned d;
 
-       bmap = isl_basic_map_cow(bmap);
-       if (!bmap)
+       if (!dim)
                return NULL;
+       d = isl_space_dim(dim, isl_dim_in);
 
-       for (i = 0; i < bmap->n_eq; ++i)
-               isl_int_sub(bmap->eq[i][0], bmap->eq[i][0], bmap->eq[i][pos]);
-
-       for (i = 0; i < bmap->n_ineq; ++i)
-               isl_int_sub(bmap->ineq[i][0],
-                               bmap->ineq[i][0], bmap->ineq[i][pos]);
-
-       for (i = 0; i < bmap->n_div; ++i) {
-               if (isl_int_is_zero(bmap->div[i][0]))
-                       continue;
-               isl_int_sub(bmap->div[i][1],
-                               bmap->div[i][1], bmap->div[i][1 + pos]);
+       app = construct_component(dim, map, exact, project);
+       if (project) {
+               app = isl_map_project_out(app, isl_dim_in, d - 1, 1);
+               app = isl_map_project_out(app, isl_dim_out, d - 1, 1);
        }
+       return app;
+}
+
+/* Compute an extended version, i.e., with path lengths, of
+ * an overapproximation of the transitive closure of "bmap"
+ * with path lengths greater than or equal to zero and with
+ * domain and range equal to "dom".
+ */
+static __isl_give isl_map *q_closure(__isl_take isl_space *dim,
+       __isl_take isl_set *dom, __isl_keep isl_basic_map *bmap, int *exact)
+{
+       int project = 1;
+       isl_map *path;
+       isl_map *map;
+       isl_map *app;
+
+       dom = isl_set_add_dims(dom, isl_dim_set, 1);
+       app = isl_map_from_domain_and_range(dom, isl_set_copy(dom));
+       map = isl_map_from_basic_map(isl_basic_map_copy(bmap));
+       path = construct_extended_path(dim, map, &project);
+       app = isl_map_intersect(app, path);
+
+       if ((*exact = check_exactness(map, isl_map_copy(app), project)) < 0)
+               goto error;
 
-       return bmap;
+       return app;
+error:
+       isl_map_free(app);
+       return NULL;
 }
 
-/* Shift variable at position "pos" up by one.
- * That is, replace the corresponding variable v by v - 1.
+/* Check whether qc has any elements of length at least one
+ * with domain and/or range outside of dom and ran.
  */
-static __isl_give isl_map *map_shift_pos(__isl_take isl_map *map, unsigned pos)
+static int has_spurious_elements(__isl_keep isl_map *qc,
+       __isl_keep isl_set *dom, __isl_keep isl_set *ran)
 {
-       int i;
+       isl_set *s;
+       int subset;
+       unsigned d;
 
-       map = isl_map_cow(map);
-       if (!map)
-               return NULL;
+       if (!qc || !dom || !ran)
+               return -1;
 
-       for (i = 0; i < map->n; ++i) {
-               map->p[i] = basic_map_shift_pos(map->p[i], pos);
-               if (!map->p[i])
-                       goto error;
+       d = isl_map_dim(qc, isl_dim_in);
+
+       qc = isl_map_copy(qc);
+       qc = set_path_length(qc, 0, 1);
+       qc = isl_map_project_out(qc, isl_dim_in, d - 1, 1);
+       qc = isl_map_project_out(qc, isl_dim_out, d - 1, 1);
+
+       s = isl_map_domain(isl_map_copy(qc));
+       subset = isl_set_is_subset(s, dom);
+       isl_set_free(s);
+       if (subset < 0)
+               goto error;
+       if (!subset) {
+               isl_map_free(qc);
+               return 1;
        }
-       ISL_F_CLR(map, ISL_MAP_NORMALIZED);
-       return map;
+
+       s = isl_map_range(qc);
+       subset = isl_set_is_subset(s, ran);
+       isl_set_free(s);
+
+       return subset < 0 ? -1 : !subset;
 error:
-       isl_map_free(map);
-       return NULL;
+       isl_map_free(qc);
+       return -1;
 }
 
-/* Check whether the overapproximation of the power of "map" is exactly
- * the power of "map".  Let R be "map" and A_k the overapproximation.
- * The approximation is exact if
+#define LEFT   2
+#define RIGHT  1
+
+/* For each basic map in "map", except i, check whether it combines
+ * with the transitive closure that is reflexive on C combines
+ * to the left and to the right.
  *
- *     A_1 = R
- *     A_k = A_{k-1} \circ R                   k >= 2
+ * In particular, if
  *
- * Since A_k is known to be an overapproximation, we only need to check
+ *     dom map_j \subseteq C
  *
- *     A_1 \subset R
- *     A_k \subset A_{k-1} \circ R             k >= 2
+ * then right[j] is set to 1.  Otherwise, if
+ *
+ *     ran map_i \cap dom map_j = \emptyset
+ *
+ * then right[j] is set to 0.  Otherwise, composing to the right
+ * is impossible.
  *
+ * Similar, for composing to the left, we have if
+ *
+ *     ran map_j \subseteq C
+ *
+ * then left[j] is set to 1.  Otherwise, if
+ *
+ *     dom map_i \cap ran map_j = \emptyset
+ *
+ * then left[j] is set to 0.  Otherwise, composing to the left
+ * is impossible.
+ *
+ * The return value is or'd with LEFT if composing to the left
+ * is possible and with RIGHT if composing to the right is possible.
  */
-static int check_power_exactness(__isl_take isl_map *map,
-       __isl_take isl_map *app, unsigned param)
+static int composability(__isl_keep isl_set *C, int i,
+       isl_set **dom, isl_set **ran, int *left, int *right,
+       __isl_keep isl_map *map)
 {
-       int exact;
-       isl_map *app_1;
-       isl_map *app_2;
+       int j;
+       int ok;
 
-       app_1 = isl_map_fix_si(isl_map_copy(app), isl_dim_param, param, 1);
+       ok = LEFT | RIGHT;
+       for (j = 0; j < map->n && ok; ++j) {
+               int overlaps, subset;
+               if (j == i)
+                       continue;
 
-       exact = isl_map_is_subset(app_1, map);
-       isl_map_free(app_1);
+               if (ok & RIGHT) {
+                       if (!dom[j])
+                               dom[j] = isl_set_from_basic_set(
+                                       isl_basic_map_domain(
+                                               isl_basic_map_copy(map->p[j])));
+                       if (!dom[j])
+                               return -1;
+                       overlaps = isl_set_overlaps(ran[i], dom[j]);
+                       if (overlaps < 0)
+                               return -1;
+                       if (!overlaps)
+                               right[j] = 0;
+                       else {
+                               subset = isl_set_is_subset(dom[j], C);
+                               if (subset < 0)
+                                       return -1;
+                               if (subset)
+                                       right[j] = 1;
+                               else
+                                       ok &= ~RIGHT;
+                       }
+               }
 
-       if (!exact || exact < 0) {
-               isl_map_free(app);
-               isl_map_free(map);
-               return exact;
+               if (ok & LEFT) {
+                       if (!ran[j])
+                               ran[j] = isl_set_from_basic_set(
+                                       isl_basic_map_range(
+                                               isl_basic_map_copy(map->p[j])));
+                       if (!ran[j])
+                               return -1;
+                       overlaps = isl_set_overlaps(dom[i], ran[j]);
+                       if (overlaps < 0)
+                               return -1;
+                       if (!overlaps)
+                               left[j] = 0;
+                       else {
+                               subset = isl_set_is_subset(ran[j], C);
+                               if (subset < 0)
+                                       return -1;
+                               if (subset)
+                                       left[j] = 1;
+                               else
+                                       ok &= ~LEFT;
+                       }
+               }
        }
 
-       app_2 = isl_map_lower_bound_si(isl_map_copy(app),
-                                       isl_dim_param, param, 2);
-       app_1 = map_shift_pos(app, 1 + param);
-       app_1 = isl_map_apply_range(map, app_1);
+       return ok;
+}
 
-       exact = isl_map_is_subset(app_2, app_1);
+static __isl_give isl_map *anonymize(__isl_take isl_map *map)
+{
+       map = isl_map_reset(map, isl_dim_in);
+       map = isl_map_reset(map, isl_dim_out);
+       return map;
+}
 
-       isl_map_free(app_1);
-       isl_map_free(app_2);
+/* Return a map that is a union of the basic maps in "map", except i,
+ * composed to left and right with qc based on the entries of "left"
+ * and "right".
+ */
+static __isl_give isl_map *compose(__isl_keep isl_map *map, int i,
+       __isl_take isl_map *qc, int *left, int *right)
+{
+       int j;
+       isl_map *comp;
 
-       return exact;
+       comp = isl_map_empty(isl_map_get_space(map));
+       for (j = 0; j < map->n; ++j) {
+               isl_map *map_j;
+
+               if (j == i)
+                       continue;
+
+               map_j = isl_map_from_basic_map(isl_basic_map_copy(map->p[j]));
+               map_j = anonymize(map_j);
+               if (left && left[j])
+                       map_j = isl_map_apply_range(map_j, isl_map_copy(qc));
+               if (right && right[j])
+                       map_j = isl_map_apply_range(isl_map_copy(qc), map_j);
+               comp = isl_map_union(comp, map_j);
+       }
+
+       comp = isl_map_compute_divs(comp);
+       comp = isl_map_coalesce(comp);
+
+       isl_map_free(qc);
+
+       return comp;
 }
 
-/* Check whether the overapproximation of the power of "map" is exactly
- * the power of "map", possibly after projecting out the power (if "project"
- * is set).
+/* Compute the transitive closure of "map" incrementally by
+ * computing
  *
- * If "project" is set and if "steps" can only result in acyclic paths,
- * then we check
+ *     map_i^+ \cup qc^+
  *
- *     A = R \cup (A \circ R)
+ * or
  *
- * where A is the overapproximation with the power projected out, i.e.,
- * an overapproximation of the transitive closure.
- * More specifically, since A is known to be an overapproximation, we check
+ *     map_i^+ \cup ((id \cup map_i^) \circ qc^+)
  *
- *     A \subset R \cup (A \circ R)
+ * or
  *
- * Otherwise, we check if the power is exact.
+ *     map_i^+ \cup (qc^+ \circ (id \cup map_i^))
+ *
+ * depending on whether left or right are NULL.
  */
-static int check_exactness(__isl_take isl_map *map, __isl_take isl_map *app,
-       unsigned param, int project)
+static __isl_give isl_map *compute_incremental(
+       __isl_take isl_space *dim, __isl_keep isl_map *map,
+       int i, __isl_take isl_map *qc, int *left, int *right, int *exact)
 {
-       isl_map *test;
-       int exact;
+       isl_map *map_i;
+       isl_map *tc;
+       isl_map *rtc = NULL;
 
-       if (!project)
-               return check_power_exactness(map, app, param);
+       if (!map)
+               goto error;
+       isl_assert(map->ctx, left || right, goto error);
 
-       map = isl_map_project_out(map, isl_dim_param, param, 1);
-       app = isl_map_project_out(app, isl_dim_param, param, 1);
+       map_i = isl_map_from_basic_map(isl_basic_map_copy(map->p[i]));
+       tc = construct_projected_component(isl_space_copy(dim), map_i,
+                                               exact, 1);
+       isl_map_free(map_i);
 
-       test = isl_map_apply_range(isl_map_copy(map), isl_map_copy(app));
-       test = isl_map_union(test, isl_map_copy(map));
+       if (*exact)
+               qc = isl_map_transitive_closure(qc, exact);
 
-       exact = isl_map_is_subset(app, test);
+       if (!*exact) {
+               isl_space_free(dim);
+               isl_map_free(tc);
+               isl_map_free(qc);
+               return isl_map_universe(isl_map_get_space(map));
+       }
 
-       isl_map_free(app);
-       isl_map_free(test);
+       if (!left || !right)
+               rtc = isl_map_union(isl_map_copy(tc),
+                                   isl_map_identity(isl_map_get_space(tc)));
+       if (!right)
+               qc = isl_map_apply_range(rtc, qc);
+       if (!left)
+               qc = isl_map_apply_range(qc, rtc);
+       qc = isl_map_union(tc, qc);
 
-       isl_map_free(map);
+       isl_space_free(dim);
 
-       return exact;
+       return qc;
 error:
-       isl_map_free(app);
-       isl_map_free(map);
+       isl_space_free(dim);
+       isl_map_free(qc);
+       return NULL;
+}
+
+/* Given a map "map", try to find a basic map such that
+ * map^+ can be computed as
+ *
+ * map^+ = map_i^+ \cup
+ *    \bigcup_j ((map_i^+ \cup Id_C)^+ \circ map_j \circ (map_i^+ \cup Id_C))^+
+ *
+ * with C the simple hull of the domain and range of the input map.
+ * map_i^ \cup Id_C is computed by allowing the path lengths to be zero
+ * and by intersecting domain and range with C.
+ * Of course, we need to check that this is actually equal to map_i^ \cup Id_C.
+ * Also, we only use the incremental computation if all the transitive
+ * closures are exact and if the number of basic maps in the union,
+ * after computing the integer divisions, is smaller than the number
+ * of basic maps in the input map.
+ */
+static int incemental_on_entire_domain(__isl_keep isl_space *dim,
+       __isl_keep isl_map *map,
+       isl_set **dom, isl_set **ran, int *left, int *right,
+       __isl_give isl_map **res)
+{
+       int i;
+       isl_set *C;
+       unsigned d;
+
+       *res = NULL;
+
+       C = isl_set_union(isl_map_domain(isl_map_copy(map)),
+                         isl_map_range(isl_map_copy(map)));
+       C = isl_set_from_basic_set(isl_set_simple_hull(C));
+       if (!C)
+               return -1;
+       if (C->n != 1) {
+               isl_set_free(C);
+               return 0;
+       }
+
+       d = isl_map_dim(map, isl_dim_in);
+
+       for (i = 0; i < map->n; ++i) {
+               isl_map *qc;
+               int exact_i, spurious;
+               int j;
+               dom[i] = isl_set_from_basic_set(isl_basic_map_domain(
+                                       isl_basic_map_copy(map->p[i])));
+               ran[i] = isl_set_from_basic_set(isl_basic_map_range(
+                                       isl_basic_map_copy(map->p[i])));
+               qc = q_closure(isl_space_copy(dim), isl_set_copy(C),
+                               map->p[i], &exact_i);
+               if (!qc)
+                       goto error;
+               if (!exact_i) {
+                       isl_map_free(qc);
+                       continue;
+               }
+               spurious = has_spurious_elements(qc, dom[i], ran[i]);
+               if (spurious) {
+                       isl_map_free(qc);
+                       if (spurious < 0)
+                               goto error;
+                       continue;
+               }
+               qc = isl_map_project_out(qc, isl_dim_in, d, 1);
+               qc = isl_map_project_out(qc, isl_dim_out, d, 1);
+               qc = isl_map_compute_divs(qc);
+               for (j = 0; j < map->n; ++j)
+                       left[j] = right[j] = 1;
+               qc = compose(map, i, qc, left, right);
+               if (!qc)
+                       goto error;
+               if (qc->n >= map->n) {
+                       isl_map_free(qc);
+                       continue;
+               }
+               *res = compute_incremental(isl_space_copy(dim), map, i, qc,
+                               left, right, &exact_i);
+               if (!*res)
+                       goto error;
+               if (exact_i)
+                       break;
+               isl_map_free(*res);
+               *res = NULL;
+       }
+
+       isl_set_free(C);
+
+       return *res != NULL;
+error:
+       isl_set_free(C);
        return -1;
 }
 
-/* Compute the positive powers of "map", or an overapproximation.
- * The power is given by parameter "param".  If the result is exact,
- * then *exact is set to 1.
- * If project is set, then we are actually interested in the transitive
- * closure, so we can use a more relaxed exactness check.
+/* Try and compute the transitive closure of "map" as
+ *
+ * map^+ = map_i^+ \cup
+ *    \bigcup_j ((map_i^+ \cup Id_C)^+ \circ map_j \circ (map_i^+ \cup Id_C))^+
+ *
+ * with C either the simple hull of the domain and range of the entire
+ * map or the simple hull of domain and range of map_i.
  */
-static __isl_give isl_map *map_power(__isl_take isl_map *map, unsigned param,
-       int *exact, int project)
+static __isl_give isl_map *incremental_closure(__isl_take isl_space *dim,
+       __isl_keep isl_map *map, int *exact, int project)
 {
-       struct isl_set *domain = NULL;
-       struct isl_set *range = NULL;
-       struct isl_map *app = NULL;
-       struct isl_map *path = NULL;
+       int i;
+       isl_set **dom = NULL;
+       isl_set **ran = NULL;
+       int *left = NULL;
+       int *right = NULL;
+       isl_set *C;
+       unsigned d;
+       isl_map *res = NULL;
 
-       if (exact)
-               *exact = 1;
+       if (!project)
+               return construct_projected_component(dim, map, exact, project);
 
-       map = isl_map_remove_empty_parts(map);
        if (!map)
-               return NULL;
+               goto error;
+       if (map->n <= 1)
+               return construct_projected_component(dim, map, exact, project);
 
-       if (isl_map_fast_is_empty(map))
-               return map;
+       d = isl_map_dim(map, isl_dim_in);
 
-       isl_assert(map->ctx, param < isl_map_dim(map, isl_dim_param), goto error);
-       isl_assert(map->ctx,
-               isl_map_dim(map, isl_dim_in) == isl_map_dim(map, isl_dim_out),
-               goto error);
+       dom = isl_calloc_array(map->ctx, isl_set *, map->n);
+       ran = isl_calloc_array(map->ctx, isl_set *, map->n);
+       left = isl_calloc_array(map->ctx, int, map->n);
+       right = isl_calloc_array(map->ctx, int, map->n);
+       if (!ran || !dom || !left || !right)
+               goto error;
 
-       domain = isl_map_domain(isl_map_copy(map));
-       domain = isl_set_coalesce(domain);
-       range = isl_map_range(isl_map_copy(map));
-       range = isl_set_coalesce(range);
-       app = isl_map_from_domain_and_range(isl_set_copy(domain),
-                                           isl_set_copy(range));
+       if (incemental_on_entire_domain(dim, map, dom, ran, left, right, &res) < 0)
+               goto error;
 
-       path = construct_path(map, param, exact ? &project : NULL);
-       app = isl_map_intersect(app, isl_map_copy(path));
+       for (i = 0; !res && i < map->n; ++i) {
+               isl_map *qc;
+               int exact_i, spurious, comp;
+               if (!dom[i])
+                       dom[i] = isl_set_from_basic_set(
+                                       isl_basic_map_domain(
+                                               isl_basic_map_copy(map->p[i])));
+               if (!dom[i])
+                       goto error;
+               if (!ran[i])
+                       ran[i] = isl_set_from_basic_set(
+                                       isl_basic_map_range(
+                                               isl_basic_map_copy(map->p[i])));
+               if (!ran[i])
+                       goto error;
+               C = isl_set_union(isl_set_copy(dom[i]),
+                                     isl_set_copy(ran[i]));
+               C = isl_set_from_basic_set(isl_set_simple_hull(C));
+               if (!C)
+                       goto error;
+               if (C->n != 1) {
+                       isl_set_free(C);
+                       continue;
+               }
+               comp = composability(C, i, dom, ran, left, right, map);
+               if (!comp || comp < 0) {
+                       isl_set_free(C);
+                       if (comp < 0)
+                               goto error;
+                       continue;
+               }
+               qc = q_closure(isl_space_copy(dim), C, map->p[i], &exact_i);
+               if (!qc)
+                       goto error;
+               if (!exact_i) {
+                       isl_map_free(qc);
+                       continue;
+               }
+               spurious = has_spurious_elements(qc, dom[i], ran[i]);
+               if (spurious) {
+                       isl_map_free(qc);
+                       if (spurious < 0)
+                               goto error;
+                       continue;
+               }
+               qc = isl_map_project_out(qc, isl_dim_in, d, 1);
+               qc = isl_map_project_out(qc, isl_dim_out, d, 1);
+               qc = isl_map_compute_divs(qc);
+               qc = compose(map, i, qc, (comp & LEFT) ? left : NULL,
+                               (comp & RIGHT) ? right : NULL);
+               if (!qc)
+                       goto error;
+               if (qc->n >= map->n) {
+                       isl_map_free(qc);
+                       continue;
+               }
+               res = compute_incremental(isl_space_copy(dim), map, i, qc,
+                               (comp & LEFT) ? left : NULL,
+                               (comp & RIGHT) ? right : NULL, &exact_i);
+               if (!res)
+                       goto error;
+               if (exact_i)
+                       break;
+               isl_map_free(res);
+               res = NULL;
+       }
 
-       if (exact &&
-           (*exact = check_exactness(isl_map_copy(map), isl_map_copy(app),
-                                     param, project)) < 0)
-               goto error;
+       for (i = 0; i < map->n; ++i) {
+               isl_set_free(dom[i]);
+               isl_set_free(ran[i]);
+       }
+       free(dom);
+       free(ran);
+       free(left);
+       free(right);
+
+       if (res) {
+               isl_space_free(dim);
+               return res;
+       }
 
-       isl_set_free(domain);
-       isl_set_free(range);
-       isl_map_free(path);
-       isl_map_free(map);
-       return app;
+       return construct_projected_component(dim, map, exact, project);
 error:
-       isl_set_free(domain);
-       isl_set_free(range);
-       isl_map_free(path);
-       isl_map_free(map);
-       isl_map_free(app);
+       if (dom)
+               for (i = 0; i < map->n; ++i)
+                       isl_set_free(dom[i]);
+       free(dom);
+       if (ran)
+               for (i = 0; i < map->n; ++i)
+                       isl_set_free(ran[i]);
+       free(ran);
+       free(left);
+       free(right);
+       isl_space_free(dim);
        return NULL;
 }
 
-/* Compute the positive powers of "map", or an overapproximation.
- * The power is given by parameter "param".  If the result is exact,
- * then *exact is set to 1.
+/* Given an array of sets "set", add "dom" at position "pos"
+ * and search for elements at earlier positions that overlap with "dom".
+ * If any can be found, then merge all of them, together with "dom", into
+ * a single set and assign the union to the first in the array,
+ * which becomes the new group leader for all groups involved in the merge.
+ * During the search, we only consider group leaders, i.e., those with
+ * group[i] = i, as the other sets have already been combined
+ * with one of the group leaders.
  */
-__isl_give isl_map *isl_map_power(__isl_take isl_map *map, unsigned param,
-       int *exact)
+static int merge(isl_set **set, int *group, __isl_take isl_set *dom, int pos)
 {
-       return map_power(map, param, exact, 0);
+       int i;
+
+       group[pos] = pos;
+       set[pos] = isl_set_copy(dom);
+
+       for (i = pos - 1; i >= 0; --i) {
+               int o;
+
+               if (group[i] != i)
+                       continue;
+
+               o = isl_set_overlaps(set[i], dom);
+               if (o < 0)
+                       goto error;
+               if (!o)
+                       continue;
+
+               set[i] = isl_set_union(set[i], set[group[pos]]);
+               set[group[pos]] = NULL;
+               if (!set[i])
+                       goto error;
+               group[group[pos]] = i;
+               group[pos] = i;
+       }
+
+       isl_set_free(dom);
+       return 0;
+error:
+       isl_set_free(dom);
+       return -1;
 }
 
-/* Compute the transitive closure  of "map", or an overapproximation.
- * If the result is exact, then *exact is set to 1.
- * Simply compute the powers of map and then project out the parameter
- * describing the power.
+/* Replace each entry in the n by n grid of maps by the cross product
+ * with the relation { [i] -> [i + 1] }.
  */
-__isl_give isl_map *isl_map_transitive_closure(__isl_take isl_map *map,
-       int *exact)
+static int add_length(__isl_keep isl_map *map, isl_map ***grid, int n)
 {
-       unsigned param;
+       int i, j, k;
+       isl_space *dim;
+       isl_basic_map *bstep;
+       isl_map *step;
+       unsigned nparam;
 
        if (!map)
-               goto error;
+               return -1;
+
+       dim = isl_map_get_space(map);
+       nparam = isl_space_dim(dim, isl_dim_param);
+       dim = isl_space_drop_dims(dim, isl_dim_in, 0, isl_space_dim(dim, isl_dim_in));
+       dim = isl_space_drop_dims(dim, isl_dim_out, 0, isl_space_dim(dim, isl_dim_out));
+       dim = isl_space_add_dims(dim, isl_dim_in, 1);
+       dim = isl_space_add_dims(dim, isl_dim_out, 1);
+       bstep = isl_basic_map_alloc_space(dim, 0, 1, 0);
+       k = isl_basic_map_alloc_equality(bstep);
+       if (k < 0) {
+               isl_basic_map_free(bstep);
+               return -1;
+       }
+       isl_seq_clr(bstep->eq[k], 1 + isl_basic_map_total_dim(bstep));
+       isl_int_set_si(bstep->eq[k][0], 1);
+       isl_int_set_si(bstep->eq[k][1 + nparam], 1);
+       isl_int_set_si(bstep->eq[k][1 + nparam + 1], -1);
+       bstep = isl_basic_map_finalize(bstep);
+       step = isl_map_from_basic_map(bstep);
 
-       param = isl_map_dim(map, isl_dim_param);
-       map = isl_map_add(map, isl_dim_param, 1);
-       map = map_power(map, param, exact, 1);
-       map = isl_map_project_out(map, isl_dim_param, param, 1);
+       for (i = 0; i < n; ++i)
+               for (j = 0; j < n; ++j)
+                       grid[i][j] = isl_map_product(grid[i][j],
+                                                    isl_map_copy(step));
 
-       return map;
-error:
-       isl_map_free(map);
-       return NULL;
+       isl_map_free(step);
+
+       return 0;
+}
+
+/* The core of the Floyd-Warshall algorithm.
+ * Updates the given n x x matrix of relations in place.
+ *
+ * The algorithm iterates over all vertices.  In each step, the whole
+ * matrix is updated to include all paths that go to the current vertex,
+ * possibly stay there a while (including passing through earlier vertices)
+ * and then come back.  At the start of each iteration, the diagonal
+ * element corresponding to the current vertex is replaced by its
+ * transitive closure to account for all indirect paths that stay
+ * in the current vertex.
+ */
+static void floyd_warshall_iterate(isl_map ***grid, int n, int *exact)
+{
+       int r, p, q;
+
+       for (r = 0; r < n; ++r) {
+               int r_exact;
+               grid[r][r] = isl_map_transitive_closure(grid[r][r],
+                               (exact && *exact) ? &r_exact : NULL);
+               if (exact && *exact && !r_exact)
+                       *exact = 0;
+
+               for (p = 0; p < n; ++p)
+                       for (q = 0; q < n; ++q) {
+                               isl_map *loop;
+                               if (p == r && q == r)
+                                       continue;
+                               loop = isl_map_apply_range(
+                                               isl_map_copy(grid[p][r]),
+                                               isl_map_copy(grid[r][q]));
+                               grid[p][q] = isl_map_union(grid[p][q], loop);
+                               loop = isl_map_apply_range(
+                                               isl_map_copy(grid[p][r]),
+                                       isl_map_apply_range(
+                                               isl_map_copy(grid[r][r]),
+                                               isl_map_copy(grid[r][q])));
+                               grid[p][q] = isl_map_union(grid[p][q], loop);
+                               grid[p][q] = isl_map_coalesce(grid[p][q]);
+                       }
+       }
 }
+
+/* Given a partition of the domains and ranges of the basic maps in "map",
+ * apply the Floyd-Warshall algorithm with the elements in the partition
+ * as vertices.
+ *
+ * In particular, there are "n" elements in the partition and "group" is
+ * an array of length 2 * map->n with entries in [0,n-1].
+ *
+ * We first construct a matrix of relations based on the partition information,
+ * apply Floyd-Warshall on this matrix of relations and then take the
+ * union of all entries in the matrix as the final result.
+ *
+ * If we are actually computing the power instead of the transitive closure,
+ * i.e., when "project" is not set, then the result should have the
+ * path lengths encoded as the difference between an extra pair of
+ * coordinates.  We therefore apply the nested transitive closures
+ * to relations that include these lengths.  In particular, we replace
+ * the input relation by the cross product with the unit length relation
+ * { [i] -> [i + 1] }.
+ */
+static __isl_give isl_map *floyd_warshall_with_groups(__isl_take isl_space *dim,
+       __isl_keep isl_map *map, int *exact, int project, int *group, int n)
+{
+       int i, j, k;
+       isl_map ***grid = NULL;
+       isl_map *app;
+
+       if (!map)
+               goto error;
+
+       if (n == 1) {
+               free(group);
+               return incremental_closure(dim, map, exact, project);
+       }
+
+       grid = isl_calloc_array(map->ctx, isl_map **, n);
+       if (!grid)
+               goto error;
+       for (i = 0; i < n; ++i) {
+               grid[i] = isl_calloc_array(map->ctx, isl_map *, n);
+               if (!grid[i])
+                       goto error;
+               for (j = 0; j < n; ++j)
+                       grid[i][j] = isl_map_empty(isl_map_get_space(map));
+       }
+
+       for (k = 0; k < map->n; ++k) {
+               i = group[2 * k];
+               j = group[2 * k + 1];
+               grid[i][j] = isl_map_union(grid[i][j],
+                               isl_map_from_basic_map(
+                                       isl_basic_map_copy(map->p[k])));
+       }
+
+       if (!project && add_length(map, grid, n) < 0)
+               goto error;
+
+       floyd_warshall_iterate(grid, n, exact);
+
+       app = isl_map_empty(isl_map_get_space(map));
+
+       for (i = 0; i < n; ++i) {
+               for (j = 0; j < n; ++j)
+                       app = isl_map_union(app, grid[i][j]);
+               free(grid[i]);
+       }
+       free(grid);
+
+       free(group);
+       isl_space_free(dim);
+
+       return app;
+error:
+       if (grid)
+               for (i = 0; i < n; ++i) {
+                       if (!grid[i])
+                               continue;
+                       for (j = 0; j < n; ++j)
+                               isl_map_free(grid[i][j]);
+                       free(grid[i]);
+               }
+       free(grid);
+       free(group);
+       isl_space_free(dim);
+       return NULL;
+}
+
+/* Partition the domains and ranges of the n basic relations in list
+ * into disjoint cells.
+ *
+ * To find the partition, we simply consider all of the domains
+ * and ranges in turn and combine those that overlap.
+ * "set" contains the partition elements and "group" indicates
+ * to which partition element a given domain or range belongs.
+ * The domain of basic map i corresponds to element 2 * i in these arrays,
+ * while the domain corresponds to element 2 * i + 1.
+ * During the construction group[k] is either equal to k,
+ * in which case set[k] contains the union of all the domains and
+ * ranges in the corresponding group, or is equal to some l < k,
+ * with l another domain or range in the same group.
+ */
+static int *setup_groups(isl_ctx *ctx, __isl_keep isl_basic_map **list, int n,
+       isl_set ***set, int *n_group)
+{
+       int i;
+       int *group = NULL;
+       int g;
+
+       *set = isl_calloc_array(ctx, isl_set *, 2 * n);
+       group = isl_alloc_array(ctx, int, 2 * n);
+
+       if (!*set || !group)
+               goto error;
+
+       for (i = 0; i < n; ++i) {
+               isl_set *dom;
+               dom = isl_set_from_basic_set(isl_basic_map_domain(
+                               isl_basic_map_copy(list[i])));
+               if (merge(*set, group, dom, 2 * i) < 0)
+                       goto error;
+               dom = isl_set_from_basic_set(isl_basic_map_range(
+                               isl_basic_map_copy(list[i])));
+               if (merge(*set, group, dom, 2 * i + 1) < 0)
+                       goto error;
+       }
+
+       g = 0;
+       for (i = 0; i < 2 * n; ++i)
+               if (group[i] == i) {
+                       if (g != i) {
+                               (*set)[g] = (*set)[i];
+                               (*set)[i] = NULL;
+                       }
+                       group[i] = g++;
+               } else
+                       group[i] = group[group[i]];
+
+       *n_group = g;
+
+       return group;
+error:
+       if (*set) {
+               for (i = 0; i < 2 * n; ++i)
+                       isl_set_free((*set)[i]);
+               free(*set);
+               *set = NULL;
+       }
+       free(group);
+       return NULL;
+}
+
+/* Check if the domains and ranges of the basic maps in "map" can
+ * be partitioned, and if so, apply Floyd-Warshall on the elements
+ * of the partition.  Note that we also apply this algorithm
+ * if we want to compute the power, i.e., when "project" is not set.
+ * However, the results are unlikely to be exact since the recursive
+ * calls inside the Floyd-Warshall algorithm typically result in
+ * non-linear path lengths quite quickly.
+ */
+static __isl_give isl_map *floyd_warshall(__isl_take isl_space *dim,
+       __isl_keep isl_map *map, int *exact, int project)
+{
+       int i;
+       isl_set **set = NULL;
+       int *group = NULL;
+       int n;
+
+       if (!map)
+               goto error;
+       if (map->n <= 1)
+               return incremental_closure(dim, map, exact, project);
+
+       group = setup_groups(map->ctx, map->p, map->n, &set, &n);
+       if (!group)
+               goto error;
+
+       for (i = 0; i < 2 * map->n; ++i)
+               isl_set_free(set[i]);
+
+       free(set);
+
+       return floyd_warshall_with_groups(dim, map, exact, project, group, n);
+error:
+       isl_space_free(dim);
+       return NULL;
+}
+
+/* Structure for representing the nodes of the graph of which
+ * strongly connected components are being computed.
+ *
+ * list contains the actual nodes
+ * check_closed is set if we may have used the fact that
+ * a pair of basic maps can be interchanged
+ */
+struct isl_tc_follows_data {
+       isl_basic_map **list;
+       int check_closed;
+};
+
+/* Check whether in the computation of the transitive closure
+ * "list[i]" (R_1) should follow (or be part of the same component as)
+ * "list[j]" (R_2).
+ *
+ * That is check whether
+ *
+ *     R_1 \circ R_2
+ *
+ * is a subset of
+ *
+ *     R_2 \circ R_1
+ *
+ * If so, then there is no reason for R_1 to immediately follow R_2
+ * in any path.
+ *
+ * *check_closed is set if the subset relation holds while
+ * R_1 \circ R_2 is not empty.
+ */
+static int basic_map_follows(int i, int j, void *user)
+{
+       struct isl_tc_follows_data *data = user;
+       struct isl_map *map12 = NULL;
+       struct isl_map *map21 = NULL;
+       int subset;
+
+       if (!isl_space_tuple_match(data->list[i]->dim, isl_dim_in,
+                                   data->list[j]->dim, isl_dim_out))
+               return 0;
+
+       map21 = isl_map_from_basic_map(
+                       isl_basic_map_apply_range(
+                               isl_basic_map_copy(data->list[j]),
+                               isl_basic_map_copy(data->list[i])));
+       subset = isl_map_is_empty(map21);
+       if (subset < 0)
+               goto error;
+       if (subset) {
+               isl_map_free(map21);
+               return 0;
+       }
+
+       if (!isl_space_tuple_match(data->list[i]->dim, isl_dim_in,
+                                   data->list[i]->dim, isl_dim_out) ||
+           !isl_space_tuple_match(data->list[j]->dim, isl_dim_in,
+                                   data->list[j]->dim, isl_dim_out)) {
+               isl_map_free(map21);
+               return 1;
+       }
+
+       map12 = isl_map_from_basic_map(
+                       isl_basic_map_apply_range(
+                               isl_basic_map_copy(data->list[i]),
+                               isl_basic_map_copy(data->list[j])));
+
+       subset = isl_map_is_subset(map21, map12);
+
+       isl_map_free(map12);
+       isl_map_free(map21);
+
+       if (subset)
+               data->check_closed = 1;
+
+       return subset < 0 ? -1 : !subset;
+error:
+       isl_map_free(map21);
+       return -1;
+}
+
+/* Given a union of basic maps R = \cup_i R_i \subseteq D \times D
+ * and a dimension specification (Z^{n+1} -> Z^{n+1}),
+ * construct a map that is an overapproximation of the map
+ * that takes an element from the dom R \times Z to an
+ * element from ran R \times Z, such that the first n coordinates of the
+ * difference between them is a sum of differences between images
+ * and pre-images in one of the R_i and such that the last coordinate
+ * is equal to the number of steps taken.
+ * If "project" is set, then these final coordinates are not included,
+ * i.e., a relation of type Z^n -> Z^n is returned.
+ * That is, let
+ *
+ *     \Delta_i = { y - x | (x, y) in R_i }
+ *
+ * then the constructed map is an overapproximation of
+ *
+ *     { (x) -> (x + d) | \exists k_i >= 0, \delta_i \in \Delta_i :
+ *                             d = (\sum_i k_i \delta_i, \sum_i k_i) and
+ *                             x in dom R and x + d in ran R }
+ *
+ * or
+ *
+ *     { (x) -> (x + d) | \exists k_i >= 0, \delta_i \in \Delta_i :
+ *                             d = (\sum_i k_i \delta_i) and
+ *                             x in dom R and x + d in ran R }
+ *
+ * if "project" is set.
+ *
+ * We first split the map into strongly connected components, perform
+ * the above on each component and then join the results in the correct
+ * order, at each join also taking in the union of both arguments
+ * to allow for paths that do not go through one of the two arguments.
+ */
+static __isl_give isl_map *construct_power_components(__isl_take isl_space *dim,
+       __isl_keep isl_map *map, int *exact, int project)
+{
+       int i, n, c;
+       struct isl_map *path = NULL;
+       struct isl_tc_follows_data data;
+       struct isl_tarjan_graph *g = NULL;
+       int *orig_exact;
+       int local_exact;
+
+       if (!map)
+               goto error;
+       if (map->n <= 1)
+               return floyd_warshall(dim, map, exact, project);
+
+       data.list = map->p;
+       data.check_closed = 0;
+       g = isl_tarjan_graph_init(map->ctx, map->n, &basic_map_follows, &data);
+       if (!g)
+               goto error;
+
+       orig_exact = exact;
+       if (data.check_closed && !exact)
+               exact = &local_exact;
+
+       c = 0;
+       i = 0;
+       n = map->n;
+       if (project)
+               path = isl_map_empty(isl_map_get_space(map));
+       else
+               path = isl_map_empty(isl_space_copy(dim));
+       path = anonymize(path);
+       while (n) {
+               struct isl_map *comp;
+               isl_map *path_comp, *path_comb;
+               comp = isl_map_alloc_space(isl_map_get_space(map), n, 0);
+               while (g->order[i] != -1) {
+                       comp = isl_map_add_basic_map(comp,
+                                   isl_basic_map_copy(map->p[g->order[i]]));
+                       --n;
+                       ++i;
+               }
+               path_comp = floyd_warshall(isl_space_copy(dim),
+                                               comp, exact, project);
+               path_comp = anonymize(path_comp);
+               path_comb = isl_map_apply_range(isl_map_copy(path),
+                                               isl_map_copy(path_comp));
+               path = isl_map_union(path, path_comp);
+               path = isl_map_union(path, path_comb);
+               isl_map_free(comp);
+               ++i;
+               ++c;
+       }
+
+       if (c > 1 && data.check_closed && !*exact) {
+               int closed;
+
+               closed = isl_map_is_transitively_closed(path);
+               if (closed < 0)
+                       goto error;
+               if (!closed) {
+                       isl_tarjan_graph_free(g);
+                       isl_map_free(path);
+                       return floyd_warshall(dim, map, orig_exact, project);
+               }
+       }
+
+       isl_tarjan_graph_free(g);
+       isl_space_free(dim);
+
+       return path;
+error:
+       isl_tarjan_graph_free(g);
+       isl_space_free(dim);
+       isl_map_free(path);
+       return NULL;
+}
+
+/* Given a union of basic maps R = \cup_i R_i \subseteq D \times D,
+ * construct a map that is an overapproximation of the map
+ * that takes an element from the space D to another
+ * element from the same space, such that the difference between
+ * them is a strictly positive sum of differences between images
+ * and pre-images in one of the R_i.
+ * The number of differences in the sum is equated to parameter "param".
+ * That is, let
+ *
+ *     \Delta_i = { y - x | (x, y) in R_i }
+ *
+ * then the constructed map is an overapproximation of
+ *
+ *     { (x) -> (x + d) | \exists k_i >= 0, \delta_i \in \Delta_i :
+ *                             d = \sum_i k_i \delta_i and k = \sum_i k_i > 0 }
+ * or
+ *
+ *     { (x) -> (x + d) | \exists k_i >= 0, \delta_i \in \Delta_i :
+ *                             d = \sum_i k_i \delta_i and \sum_i k_i > 0 }
+ *
+ * if "project" is set.
+ *
+ * If "project" is not set, then
+ * we construct an extended mapping with an extra coordinate
+ * that indicates the number of steps taken.  In particular,
+ * the difference in the last coordinate is equal to the number
+ * of steps taken to move from a domain element to the corresponding
+ * image element(s).
+ */
+static __isl_give isl_map *construct_power(__isl_keep isl_map *map,
+       int *exact, int project)
+{
+       struct isl_map *app = NULL;
+       isl_space *dim = NULL;
+       unsigned d;
+
+       if (!map)
+               return NULL;
+
+       dim = isl_map_get_space(map);
+
+       d = isl_space_dim(dim, isl_dim_in);
+       dim = isl_space_add_dims(dim, isl_dim_in, 1);
+       dim = isl_space_add_dims(dim, isl_dim_out, 1);
+
+       app = construct_power_components(isl_space_copy(dim), map,
+                                       exact, project);
+
+       isl_space_free(dim);
+
+       return app;
+}
+
+/* Compute the positive powers of "map", or an overapproximation.
+ * If the result is exact, then *exact is set to 1.
+ *
+ * If project is set, then we are actually interested in the transitive
+ * closure, so we can use a more relaxed exactness check.
+ * The lengths of the paths are also projected out instead of being
+ * encoded as the difference between an extra pair of final coordinates.
+ */
+static __isl_give isl_map *map_power(__isl_take isl_map *map,
+       int *exact, int project)
+{
+       struct isl_map *app = NULL;
+
+       if (exact)
+               *exact = 1;
+
+       if (!map)
+               return NULL;
+
+       isl_assert(map->ctx,
+               isl_map_dim(map, isl_dim_in) == isl_map_dim(map, isl_dim_out),
+               goto error);
+
+       app = construct_power(map, exact, project);
+
+       isl_map_free(map);
+       return app;
+error:
+       isl_map_free(map);
+       isl_map_free(app);
+       return NULL;
+}
+
+/* Compute the positive powers of "map", or an overapproximation.
+ * The result maps the exponent to a nested copy of the corresponding power.
+ * If the result is exact, then *exact is set to 1.
+ * map_power constructs an extended relation with the path lengths
+ * encoded as the difference between the final coordinates.
+ * In the final step, this difference is equated to an extra parameter
+ * and made positive.  The extra coordinates are subsequently projected out
+ * and the parameter is turned into the domain of the result.
+ */
+__isl_give isl_map *isl_map_power(__isl_take isl_map *map, int *exact)
+{
+       isl_space *target_dim;
+       isl_space *dim;
+       isl_map *diff;
+       unsigned d;
+       unsigned param;
+
+       if (!map)
+               return NULL;
+
+       d = isl_map_dim(map, isl_dim_in);
+       param = isl_map_dim(map, isl_dim_param);
+
+       map = isl_map_compute_divs(map);
+       map = isl_map_coalesce(map);
+
+       if (isl_map_plain_is_empty(map)) {
+               map = isl_map_from_range(isl_map_wrap(map));
+               map = isl_map_add_dims(map, isl_dim_in, 1);
+               map = isl_map_set_dim_name(map, isl_dim_in, 0, "k");
+               return map;
+       }
+
+       target_dim = isl_map_get_space(map);
+       target_dim = isl_space_from_range(isl_space_wrap(target_dim));
+       target_dim = isl_space_add_dims(target_dim, isl_dim_in, 1);
+       target_dim = isl_space_set_dim_name(target_dim, isl_dim_in, 0, "k");
+
+       map = map_power(map, exact, 0);
+
+       map = isl_map_add_dims(map, isl_dim_param, 1);
+       dim = isl_map_get_space(map);
+       diff = equate_parameter_to_length(dim, param);
+       map = isl_map_intersect(map, diff);
+       map = isl_map_project_out(map, isl_dim_in, d, 1);
+       map = isl_map_project_out(map, isl_dim_out, d, 1);
+       map = isl_map_from_range(isl_map_wrap(map));
+       map = isl_map_move_dims(map, isl_dim_in, 0, isl_dim_param, param, 1);
+
+       map = isl_map_reset_space(map, target_dim);
+
+       return map;
+}
+
+/* Compute a relation that maps each element in the range of the input
+ * relation to the lengths of all paths composed of edges in the input
+ * relation that end up in the given range element.
+ * The result may be an overapproximation, in which case *exact is set to 0.
+ * The resulting relation is very similar to the power relation.
+ * The difference are that the domain has been projected out, the
+ * range has become the domain and the exponent is the range instead
+ * of a parameter.
+ */
+__isl_give isl_map *isl_map_reaching_path_lengths(__isl_take isl_map *map,
+       int *exact)
+{
+       isl_space *dim;
+       isl_map *diff;
+       unsigned d;
+       unsigned param;
+
+       if (!map)
+               return NULL;
+
+       d = isl_map_dim(map, isl_dim_in);
+       param = isl_map_dim(map, isl_dim_param);
+
+       map = isl_map_compute_divs(map);
+       map = isl_map_coalesce(map);
+
+       if (isl_map_plain_is_empty(map)) {
+               if (exact)
+                       *exact = 1;
+               map = isl_map_project_out(map, isl_dim_out, 0, d);
+               map = isl_map_add_dims(map, isl_dim_out, 1);
+               return map;
+       }
+
+       map = map_power(map, exact, 0);
+
+       map = isl_map_add_dims(map, isl_dim_param, 1);
+       dim = isl_map_get_space(map);
+       diff = equate_parameter_to_length(dim, param);
+       map = isl_map_intersect(map, diff);
+       map = isl_map_project_out(map, isl_dim_in, 0, d + 1);
+       map = isl_map_project_out(map, isl_dim_out, d, 1);
+       map = isl_map_reverse(map);
+       map = isl_map_move_dims(map, isl_dim_out, 0, isl_dim_param, param, 1);
+
+       return map;
+}
+
+/* Check whether equality i of bset is a pure stride constraint
+ * on a single dimensions, i.e., of the form
+ *
+ *     v = k e
+ *
+ * with k a constant and e an existentially quantified variable.
+ */
+static int is_eq_stride(__isl_keep isl_basic_set *bset, int i)
+{
+       unsigned nparam;
+       unsigned d;
+       unsigned n_div;
+       int pos1;
+       int pos2;
+
+       if (!bset)
+               return -1;
+
+       if (!isl_int_is_zero(bset->eq[i][0]))
+               return 0;
+
+       nparam = isl_basic_set_dim(bset, isl_dim_param);
+       d = isl_basic_set_dim(bset, isl_dim_set);
+       n_div = isl_basic_set_dim(bset, isl_dim_div);
+
+       if (isl_seq_first_non_zero(bset->eq[i] + 1, nparam) != -1)
+               return 0;
+       pos1 = isl_seq_first_non_zero(bset->eq[i] + 1 + nparam, d);
+       if (pos1 == -1)
+               return 0;
+       if (isl_seq_first_non_zero(bset->eq[i] + 1 + nparam + pos1 + 1, 
+                                       d - pos1 - 1) != -1)
+               return 0;
+
+       pos2 = isl_seq_first_non_zero(bset->eq[i] + 1 + nparam + d, n_div);
+       if (pos2 == -1)
+               return 0;
+       if (isl_seq_first_non_zero(bset->eq[i] + 1 + nparam + d  + pos2 + 1,
+                                  n_div - pos2 - 1) != -1)
+               return 0;
+       if (!isl_int_is_one(bset->eq[i][1 + nparam + pos1]) &&
+           !isl_int_is_negone(bset->eq[i][1 + nparam + pos1]))
+               return 0;
+
+       return 1;
+}
+
+/* Given a map, compute the smallest superset of this map that is of the form
+ *
+ *     { i -> j : L <= j - i <= U and exists a_p: j_p - i_p = M_p a_p }
+ *
+ * (where p ranges over the (non-parametric) dimensions),
+ * compute the transitive closure of this map, i.e.,
+ *
+ *     { i -> j : exists k > 0:
+ *             k L <= j - i <= k U and exists a: j_p - i_p = M_p a_p }
+ *
+ * and intersect domain and range of this transitive closure with
+ * the given domain and range.
+ *
+ * If with_id is set, then try to include as much of the identity mapping
+ * as possible, by computing
+ *
+ *     { i -> j : exists k >= 0:
+ *             k L <= j - i <= k U and exists a: j_p - i_p = M_p a_p }
+ *
+ * instead (i.e., allow k = 0).
+ *
+ * In practice, we compute the difference set
+ *
+ *     delta  = { j - i | i -> j in map },
+ *
+ * look for stride constraint on the individual dimensions and compute
+ * (constant) lower and upper bounds for each individual dimension,
+ * adding a constraint for each bound not equal to infinity.
+ */
+static __isl_give isl_map *box_closure_on_domain(__isl_take isl_map *map,
+       __isl_take isl_set *dom, __isl_take isl_set *ran, int with_id)
+{
+       int i;
+       int k;
+       unsigned d;
+       unsigned nparam;
+       unsigned total;
+       isl_space *dim;
+       isl_set *delta;
+       isl_map *app = NULL;
+       isl_basic_set *aff = NULL;
+       isl_basic_map *bmap = NULL;
+       isl_vec *obj = NULL;
+       isl_int opt;
+
+       isl_int_init(opt);
+
+       delta = isl_map_deltas(isl_map_copy(map));
+
+       aff = isl_set_affine_hull(isl_set_copy(delta));
+       if (!aff)
+               goto error;
+       dim = isl_map_get_space(map);
+       d = isl_space_dim(dim, isl_dim_in);
+       nparam = isl_space_dim(dim, isl_dim_param);
+       total = isl_space_dim(dim, isl_dim_all);
+       bmap = isl_basic_map_alloc_space(dim,
+                                       aff->n_div + 1, aff->n_div, 2 * d + 1);
+       for (i = 0; i < aff->n_div + 1; ++i) {
+               k = isl_basic_map_alloc_div(bmap);
+               if (k < 0)
+                       goto error;
+               isl_int_set_si(bmap->div[k][0], 0);
+       }
+       for (i = 0; i < aff->n_eq; ++i) {
+               if (!is_eq_stride(aff, i))
+                       continue;
+               k = isl_basic_map_alloc_equality(bmap);
+               if (k < 0)
+                       goto error;
+               isl_seq_clr(bmap->eq[k], 1 + nparam);
+               isl_seq_cpy(bmap->eq[k] + 1 + nparam + d,
+                               aff->eq[i] + 1 + nparam, d);
+               isl_seq_neg(bmap->eq[k] + 1 + nparam,
+                               aff->eq[i] + 1 + nparam, d);
+               isl_seq_cpy(bmap->eq[k] + 1 + nparam + 2 * d,
+                               aff->eq[i] + 1 + nparam + d, aff->n_div);
+               isl_int_set_si(bmap->eq[k][1 + total + aff->n_div], 0);
+       }
+       obj = isl_vec_alloc(map->ctx, 1 + nparam + d);
+       if (!obj)
+               goto error;
+       isl_seq_clr(obj->el, 1 + nparam + d);
+       for (i = 0; i < d; ++ i) {
+               enum isl_lp_result res;
+
+               isl_int_set_si(obj->el[1 + nparam + i], 1);
+
+               res = isl_set_solve_lp(delta, 0, obj->el, map->ctx->one, &opt,
+                                       NULL, NULL);
+               if (res == isl_lp_error)
+                       goto error;
+               if (res == isl_lp_ok) {
+                       k = isl_basic_map_alloc_inequality(bmap);
+                       if (k < 0)
+                               goto error;
+                       isl_seq_clr(bmap->ineq[k],
+                                       1 + nparam + 2 * d + bmap->n_div);
+                       isl_int_set_si(bmap->ineq[k][1 + nparam + i], -1);
+                       isl_int_set_si(bmap->ineq[k][1 + nparam + d + i], 1);
+                       isl_int_neg(bmap->ineq[k][1 + nparam + 2 * d + aff->n_div], opt);
+               }
+
+               res = isl_set_solve_lp(delta, 1, obj->el, map->ctx->one, &opt,
+                                       NULL, NULL);
+               if (res == isl_lp_error)
+                       goto error;
+               if (res == isl_lp_ok) {
+                       k = isl_basic_map_alloc_inequality(bmap);
+                       if (k < 0)
+                               goto error;
+                       isl_seq_clr(bmap->ineq[k],
+                                       1 + nparam + 2 * d + bmap->n_div);
+                       isl_int_set_si(bmap->ineq[k][1 + nparam + i], 1);
+                       isl_int_set_si(bmap->ineq[k][1 + nparam + d + i], -1);
+                       isl_int_set(bmap->ineq[k][1 + nparam + 2 * d + aff->n_div], opt);
+               }
+
+               isl_int_set_si(obj->el[1 + nparam + i], 0);
+       }
+       k = isl_basic_map_alloc_inequality(bmap);
+       if (k < 0)
+               goto error;
+       isl_seq_clr(bmap->ineq[k],
+                       1 + nparam + 2 * d + bmap->n_div);
+       if (!with_id)
+               isl_int_set_si(bmap->ineq[k][0], -1);
+       isl_int_set_si(bmap->ineq[k][1 + nparam + 2 * d + aff->n_div], 1);
+
+       app = isl_map_from_domain_and_range(dom, ran);
+
+       isl_vec_free(obj);
+       isl_basic_set_free(aff);
+       isl_map_free(map);
+       bmap = isl_basic_map_finalize(bmap);
+       isl_set_free(delta);
+       isl_int_clear(opt);
+
+       map = isl_map_from_basic_map(bmap);
+       map = isl_map_intersect(map, app);
+
+       return map;
+error:
+       isl_vec_free(obj);
+       isl_basic_map_free(bmap);
+       isl_basic_set_free(aff);
+       isl_set_free(dom);
+       isl_set_free(ran);
+       isl_map_free(map);
+       isl_set_free(delta);
+       isl_int_clear(opt);
+       return NULL;
+}
+
+/* Given a map, compute the smallest superset of this map that is of the form
+ *
+ *     { i -> j : L <= j - i <= U and exists a_p: j_p - i_p = M_p a_p }
+ *
+ * (where p ranges over the (non-parametric) dimensions),
+ * compute the transitive closure of this map, i.e.,
+ *
+ *     { i -> j : exists k > 0:
+ *             k L <= j - i <= k U and exists a: j_p - i_p = M_p a_p }
+ *
+ * and intersect domain and range of this transitive closure with
+ * domain and range of the original map.
+ */
+static __isl_give isl_map *box_closure(__isl_take isl_map *map)
+{
+       isl_set *domain;
+       isl_set *range;
+
+       domain = isl_map_domain(isl_map_copy(map));
+       domain = isl_set_coalesce(domain);
+       range = isl_map_range(isl_map_copy(map));
+       range = isl_set_coalesce(range);
+
+       return box_closure_on_domain(map, domain, range, 0);
+}
+
+/* Given a map, compute the smallest superset of this map that is of the form
+ *
+ *     { i -> j : L <= j - i <= U and exists a_p: j_p - i_p = M_p a_p }
+ *
+ * (where p ranges over the (non-parametric) dimensions),
+ * compute the transitive and partially reflexive closure of this map, i.e.,
+ *
+ *     { i -> j : exists k >= 0:
+ *             k L <= j - i <= k U and exists a: j_p - i_p = M_p a_p }
+ *
+ * and intersect domain and range of this transitive closure with
+ * the given domain.
+ */
+static __isl_give isl_map *box_closure_with_identity(__isl_take isl_map *map,
+       __isl_take isl_set *dom)
+{
+       return box_closure_on_domain(map, dom, isl_set_copy(dom), 1);
+}
+
+/* Check whether app is the transitive closure of map.
+ * In particular, check that app is acyclic and, if so,
+ * check that
+ *
+ *     app \subset (map \cup (map \circ app))
+ */
+static int check_exactness_omega(__isl_keep isl_map *map,
+       __isl_keep isl_map *app)
+{
+       isl_set *delta;
+       int i;
+       int is_empty, is_exact;
+       unsigned d;
+       isl_map *test;
+
+       delta = isl_map_deltas(isl_map_copy(app));
+       d = isl_set_dim(delta, isl_dim_set);
+       for (i = 0; i < d; ++i)
+               delta = isl_set_fix_si(delta, isl_dim_set, i, 0);
+       is_empty = isl_set_is_empty(delta);
+       isl_set_free(delta);
+       if (is_empty < 0)
+               return -1;
+       if (!is_empty)
+               return 0;
+
+       test = isl_map_apply_range(isl_map_copy(app), isl_map_copy(map));
+       test = isl_map_union(test, isl_map_copy(map));
+       is_exact = isl_map_is_subset(app, test);
+       isl_map_free(test);
+
+       return is_exact;
+}
+
+/* Check if basic map M_i can be combined with all the other
+ * basic maps such that
+ *
+ *     (\cup_j M_j)^+
+ *
+ * can be computed as
+ *
+ *     M_i \cup (\cup_{j \ne i} M_i^* \circ M_j \circ M_i^*)^+
+ *
+ * In particular, check if we can compute a compact representation
+ * of
+ *
+ *             M_i^* \circ M_j \circ M_i^*
+ *
+ * for each j != i.
+ * Let M_i^? be an extension of M_i^+ that allows paths
+ * of length zero, i.e., the result of box_closure(., 1).
+ * The criterion, as proposed by Kelly et al., is that
+ * id = M_i^? - M_i^+ can be represented as a basic map
+ * and that
+ *
+ *     id \circ M_j \circ id = M_j
+ *
+ * for each j != i.
+ *
+ * If this function returns 1, then tc and qc are set to
+ * M_i^+ and M_i^?, respectively.
+ */
+static int can_be_split_off(__isl_keep isl_map *map, int i,
+       __isl_give isl_map **tc, __isl_give isl_map **qc)
+{
+       isl_map *map_i, *id = NULL;
+       int j = -1;
+       isl_set *C;
+
+       *tc = NULL;
+       *qc = NULL;
+
+       C = isl_set_union(isl_map_domain(isl_map_copy(map)),
+                         isl_map_range(isl_map_copy(map)));
+       C = isl_set_from_basic_set(isl_set_simple_hull(C));
+       if (!C)
+               goto error;
+
+       map_i = isl_map_from_basic_map(isl_basic_map_copy(map->p[i]));
+       *tc = box_closure(isl_map_copy(map_i));
+       *qc = box_closure_with_identity(map_i, C);
+       id = isl_map_subtract(isl_map_copy(*qc), isl_map_copy(*tc));
+
+       if (!id || !*qc)
+               goto error;
+       if (id->n != 1 || (*qc)->n != 1)
+               goto done;
+
+       for (j = 0; j < map->n; ++j) {
+               isl_map *map_j, *test;
+               int is_ok;
+
+               if (i == j)
+                       continue;
+               map_j = isl_map_from_basic_map(
+                                       isl_basic_map_copy(map->p[j]));
+               test = isl_map_apply_range(isl_map_copy(id),
+                                               isl_map_copy(map_j));
+               test = isl_map_apply_range(test, isl_map_copy(id));
+               is_ok = isl_map_is_equal(test, map_j);
+               isl_map_free(map_j);
+               isl_map_free(test);
+               if (is_ok < 0)
+                       goto error;
+               if (!is_ok)
+                       break;
+       }
+
+done:
+       isl_map_free(id);
+       if (j == map->n)
+               return 1;
+
+       isl_map_free(*qc);
+       isl_map_free(*tc);
+       *qc = NULL;
+       *tc = NULL;
+
+       return 0;
+error:
+       isl_map_free(id);
+       isl_map_free(*qc);
+       isl_map_free(*tc);
+       *qc = NULL;
+       *tc = NULL;
+       return -1;
+}
+
+static __isl_give isl_map *box_closure_with_check(__isl_take isl_map *map,
+       int *exact)
+{
+       isl_map *app;
+
+       app = box_closure(isl_map_copy(map));
+       if (exact)
+               *exact = check_exactness_omega(map, app);
+
+       isl_map_free(map);
+       return app;
+}
+
+/* Compute an overapproximation of the transitive closure of "map"
+ * using a variation of the algorithm from
+ * "Transitive Closure of Infinite Graphs and its Applications"
+ * by Kelly et al.
+ *
+ * We first check whether we can can split of any basic map M_i and
+ * compute
+ *
+ *     (\cup_j M_j)^+
+ *
+ * as
+ *
+ *     M_i \cup (\cup_{j \ne i} M_i^* \circ M_j \circ M_i^*)^+
+ *
+ * using a recursive call on the remaining map.
+ *
+ * If not, we simply call box_closure on the whole map.
+ */
+static __isl_give isl_map *transitive_closure_omega(__isl_take isl_map *map,
+       int *exact)
+{
+       int i, j;
+       int exact_i;
+       isl_map *app;
+
+       if (!map)
+               return NULL;
+       if (map->n == 1)
+               return box_closure_with_check(map, exact);
+
+       for (i = 0; i < map->n; ++i) {
+               int ok;
+               isl_map *qc, *tc;
+               ok = can_be_split_off(map, i, &tc, &qc);
+               if (ok < 0)
+                       goto error;
+               if (!ok)
+                       continue;
+
+               app = isl_map_alloc_space(isl_map_get_space(map), map->n - 1, 0);
+
+               for (j = 0; j < map->n; ++j) {
+                       if (j == i)
+                               continue;
+                       app = isl_map_add_basic_map(app,
+                                               isl_basic_map_copy(map->p[j]));
+               }
+
+               app = isl_map_apply_range(isl_map_copy(qc), app);
+               app = isl_map_apply_range(app, qc);
+
+               app = isl_map_union(tc, transitive_closure_omega(app, NULL));
+               exact_i = check_exactness_omega(map, app);
+               if (exact_i == 1) {
+                       if (exact)
+                               *exact = exact_i;
+                       isl_map_free(map);
+                       return app;
+               }
+               isl_map_free(app);
+               if (exact_i < 0)
+                       goto error;
+       }
+
+       return box_closure_with_check(map, exact);
+error:
+       isl_map_free(map);
+       return NULL;
+}
+
+/* Compute the transitive closure  of "map", or an overapproximation.
+ * If the result is exact, then *exact is set to 1.
+ * Simply use map_power to compute the powers of map, but tell
+ * it to project out the lengths of the paths instead of equating
+ * the length to a parameter.
+ */
+__isl_give isl_map *isl_map_transitive_closure(__isl_take isl_map *map,
+       int *exact)
+{
+       isl_space *target_dim;
+       int closed;
+
+       if (!map)
+               goto error;
+
+       if (map->ctx->opt->closure == ISL_CLOSURE_BOX)
+               return transitive_closure_omega(map, exact);
+
+       map = isl_map_compute_divs(map);
+       map = isl_map_coalesce(map);
+       closed = isl_map_is_transitively_closed(map);
+       if (closed < 0)
+               goto error;
+       if (closed) {
+               if (exact)
+                       *exact = 1;
+               return map;
+       }
+
+       target_dim = isl_map_get_space(map);
+       map = map_power(map, exact, 1);
+       map = isl_map_reset_space(map, target_dim);
+
+       return map;
+error:
+       isl_map_free(map);
+       return NULL;
+}
+
+static int inc_count(__isl_take isl_map *map, void *user)
+{
+       int *n = user;
+
+       *n += map->n;
+
+       isl_map_free(map);
+
+       return 0;
+}
+
+static int collect_basic_map(__isl_take isl_map *map, void *user)
+{
+       int i;
+       isl_basic_map ***next = user;
+
+       for (i = 0; i < map->n; ++i) {
+               **next = isl_basic_map_copy(map->p[i]);
+               if (!**next)
+                       goto error;
+               (*next)++;
+       }
+
+       isl_map_free(map);
+       return 0;
+error:
+       isl_map_free(map);
+       return -1;
+}
+
+/* Perform Floyd-Warshall on the given list of basic relations.
+ * The basic relations may live in different dimensions,
+ * but basic relations that get assigned to the diagonal of the
+ * grid have domains and ranges of the same dimension and so
+ * the standard algorithm can be used because the nested transitive
+ * closures are only applied to diagonal elements and because all
+ * compositions are peformed on relations with compatible domains and ranges.
+ */
+static __isl_give isl_union_map *union_floyd_warshall_on_list(isl_ctx *ctx,
+       __isl_keep isl_basic_map **list, int n, int *exact)
+{
+       int i, j, k;
+       int n_group;
+       int *group = NULL;
+       isl_set **set = NULL;
+       isl_map ***grid = NULL;
+       isl_union_map *app;
+
+       group = setup_groups(ctx, list, n, &set, &n_group);
+       if (!group)
+               goto error;
+
+       grid = isl_calloc_array(ctx, isl_map **, n_group);
+       if (!grid)
+               goto error;
+       for (i = 0; i < n_group; ++i) {
+               grid[i] = isl_calloc_array(ctx, isl_map *, n_group);
+               if (!grid[i])
+                       goto error;
+               for (j = 0; j < n_group; ++j) {
+                       isl_space *dim1, *dim2, *dim;
+                       dim1 = isl_space_reverse(isl_set_get_space(set[i]));
+                       dim2 = isl_set_get_space(set[j]);
+                       dim = isl_space_join(dim1, dim2);
+                       grid[i][j] = isl_map_empty(dim);
+               }
+       }
+
+       for (k = 0; k < n; ++k) {
+               i = group[2 * k];
+               j = group[2 * k + 1];
+               grid[i][j] = isl_map_union(grid[i][j],
+                               isl_map_from_basic_map(
+                                       isl_basic_map_copy(list[k])));
+       }
+       
+       floyd_warshall_iterate(grid, n_group, exact);
+
+       app = isl_union_map_empty(isl_map_get_space(grid[0][0]));
+
+       for (i = 0; i < n_group; ++i) {
+               for (j = 0; j < n_group; ++j)
+                       app = isl_union_map_add_map(app, grid[i][j]);
+               free(grid[i]);
+       }
+       free(grid);
+
+       for (i = 0; i < 2 * n; ++i)
+               isl_set_free(set[i]);
+       free(set);
+
+       free(group);
+       return app;
+error:
+       if (grid)
+               for (i = 0; i < n_group; ++i) {
+                       if (!grid[i])
+                               continue;
+                       for (j = 0; j < n_group; ++j)
+                               isl_map_free(grid[i][j]);
+                       free(grid[i]);
+               }
+       free(grid);
+       if (set) {
+               for (i = 0; i < 2 * n; ++i)
+                       isl_set_free(set[i]);
+               free(set);
+       }
+       free(group);
+       return NULL;
+}
+
+/* Perform Floyd-Warshall on the given union relation.
+ * The implementation is very similar to that for non-unions.
+ * The main difference is that it is applied unconditionally.
+ * We first extract a list of basic maps from the union map
+ * and then perform the algorithm on this list.
+ */
+static __isl_give isl_union_map *union_floyd_warshall(
+       __isl_take isl_union_map *umap, int *exact)
+{
+       int i, n;
+       isl_ctx *ctx;
+       isl_basic_map **list = NULL;
+       isl_basic_map **next;
+       isl_union_map *res;
+
+       n = 0;
+       if (isl_union_map_foreach_map(umap, inc_count, &n) < 0)
+               goto error;
+
+       ctx = isl_union_map_get_ctx(umap);
+       list = isl_calloc_array(ctx, isl_basic_map *, n);
+       if (!list)
+               goto error;
+
+       next = list;
+       if (isl_union_map_foreach_map(umap, collect_basic_map, &next) < 0)
+               goto error;
+
+       res = union_floyd_warshall_on_list(ctx, list, n, exact);
+
+       if (list) {
+               for (i = 0; i < n; ++i)
+                       isl_basic_map_free(list[i]);
+               free(list);
+       }
+
+       isl_union_map_free(umap);
+       return res;
+error:
+       if (list) {
+               for (i = 0; i < n; ++i)
+                       isl_basic_map_free(list[i]);
+               free(list);
+       }
+       isl_union_map_free(umap);
+       return NULL;
+}
+
+/* Decompose the give union relation into strongly connected components.
+ * The implementation is essentially the same as that of
+ * construct_power_components with the major difference that all
+ * operations are performed on union maps.
+ */
+static __isl_give isl_union_map *union_components(
+       __isl_take isl_union_map *umap, int *exact)
+{
+       int i;
+       int n;
+       isl_ctx *ctx;
+       isl_basic_map **list = NULL;
+       isl_basic_map **next;
+       isl_union_map *path = NULL;
+       struct isl_tc_follows_data data;
+       struct isl_tarjan_graph *g = NULL;
+       int c, l;
+       int recheck = 0;
+
+       n = 0;
+       if (isl_union_map_foreach_map(umap, inc_count, &n) < 0)
+               goto error;
+
+       if (n <= 1)
+               return union_floyd_warshall(umap, exact);
+
+       ctx = isl_union_map_get_ctx(umap);
+       list = isl_calloc_array(ctx, isl_basic_map *, n);
+       if (!list)
+               goto error;
+
+       next = list;
+       if (isl_union_map_foreach_map(umap, collect_basic_map, &next) < 0)
+               goto error;
+
+       data.list = list;
+       data.check_closed = 0;
+       g = isl_tarjan_graph_init(ctx, n, &basic_map_follows, &data);
+       if (!g)
+               goto error;
+
+       c = 0;
+       i = 0;
+       l = n;
+       path = isl_union_map_empty(isl_union_map_get_space(umap));
+       while (l) {
+               isl_union_map *comp;
+               isl_union_map *path_comp, *path_comb;
+               comp = isl_union_map_empty(isl_union_map_get_space(umap));
+               while (g->order[i] != -1) {
+                       comp = isl_union_map_add_map(comp,
+                                   isl_map_from_basic_map(
+                                       isl_basic_map_copy(list[g->order[i]])));
+                       --l;
+                       ++i;
+               }
+               path_comp = union_floyd_warshall(comp, exact);
+               path_comb = isl_union_map_apply_range(isl_union_map_copy(path),
+                                               isl_union_map_copy(path_comp));
+               path = isl_union_map_union(path, path_comp);
+               path = isl_union_map_union(path, path_comb);
+               ++i;
+               ++c;
+       }
+
+       if (c > 1 && data.check_closed && !*exact) {
+               int closed;
+
+               closed = isl_union_map_is_transitively_closed(path);
+               if (closed < 0)
+                       goto error;
+               recheck = !closed;
+       }
+
+       isl_tarjan_graph_free(g);
+
+       for (i = 0; i < n; ++i)
+               isl_basic_map_free(list[i]);
+       free(list);
+
+       if (recheck) {
+               isl_union_map_free(path);
+               return union_floyd_warshall(umap, exact);
+       }
+
+       isl_union_map_free(umap);
+
+       return path;
+error:
+       isl_tarjan_graph_free(g);
+       if (list) {
+               for (i = 0; i < n; ++i)
+                       isl_basic_map_free(list[i]);
+               free(list);
+       }
+       isl_union_map_free(umap);
+       isl_union_map_free(path);
+       return NULL;
+}
+
+/* Compute the transitive closure  of "umap", or an overapproximation.
+ * If the result is exact, then *exact is set to 1.
+ */
+__isl_give isl_union_map *isl_union_map_transitive_closure(
+       __isl_take isl_union_map *umap, int *exact)
+{
+       int closed;
+
+       if (!umap)
+               return NULL;
+
+       if (exact)
+               *exact = 1;
+
+       umap = isl_union_map_compute_divs(umap);
+       umap = isl_union_map_coalesce(umap);
+       closed = isl_union_map_is_transitively_closed(umap);
+       if (closed < 0)
+               goto error;
+       if (closed)
+               return umap;
+       umap = union_components(umap, exact);
+       return umap;
+error:
+       isl_union_map_free(umap);
+       return NULL;
+}
+
+struct isl_union_power {
+       isl_union_map *pow;
+       int *exact;
+};
+
+static int power(__isl_take isl_map *map, void *user)
+{
+       struct isl_union_power *up = user;
+
+       map = isl_map_power(map, up->exact);
+       up->pow = isl_union_map_from_map(map);
+
+       return -1;
+}
+
+/* Construct a map [x] -> [x+1], with parameters prescribed by "dim".
+ */
+static __isl_give isl_union_map *increment(__isl_take isl_space *dim)
+{
+       int k;
+       isl_basic_map *bmap;
+
+       dim = isl_space_add_dims(dim, isl_dim_in, 1);
+       dim = isl_space_add_dims(dim, isl_dim_out, 1);
+       bmap = isl_basic_map_alloc_space(dim, 0, 1, 0);
+       k = isl_basic_map_alloc_equality(bmap);
+       if (k < 0)
+               goto error;
+       isl_seq_clr(bmap->eq[k], isl_basic_map_total_dim(bmap));
+       isl_int_set_si(bmap->eq[k][0], 1);
+       isl_int_set_si(bmap->eq[k][isl_basic_map_offset(bmap, isl_dim_in)], 1);
+       isl_int_set_si(bmap->eq[k][isl_basic_map_offset(bmap, isl_dim_out)], -1);
+       return isl_union_map_from_map(isl_map_from_basic_map(bmap));
+error:
+       isl_basic_map_free(bmap);
+       return NULL;
+}
+
+/* Construct a map [[x]->[y]] -> [y-x], with parameters prescribed by "dim".
+ */
+static __isl_give isl_union_map *deltas_map(__isl_take isl_space *dim)
+{
+       isl_basic_map *bmap;
+
+       dim = isl_space_add_dims(dim, isl_dim_in, 1);
+       dim = isl_space_add_dims(dim, isl_dim_out, 1);
+       bmap = isl_basic_map_universe(dim);
+       bmap = isl_basic_map_deltas_map(bmap);
+
+       return isl_union_map_from_map(isl_map_from_basic_map(bmap));
+}
+
+/* Compute the positive powers of "map", or an overapproximation.
+ * The result maps the exponent to a nested copy of the corresponding power.
+ * If the result is exact, then *exact is set to 1.
+ */
+__isl_give isl_union_map *isl_union_map_power(__isl_take isl_union_map *umap,
+       int *exact)
+{
+       int n;
+       isl_union_map *inc;
+       isl_union_map *dm;
+
+       if (!umap)
+               return NULL;
+       n = isl_union_map_n_map(umap);
+       if (n == 0)
+               return umap;
+       if (n == 1) {
+               struct isl_union_power up = { NULL, exact };
+               isl_union_map_foreach_map(umap, &power, &up);
+               isl_union_map_free(umap);
+               return up.pow;
+       }
+       inc = increment(isl_union_map_get_space(umap));
+       umap = isl_union_map_product(inc, umap);
+       umap = isl_union_map_transitive_closure(umap, exact);
+       umap = isl_union_map_zip(umap);
+       dm = deltas_map(isl_union_map_get_space(umap));
+       umap = isl_union_map_apply_domain(umap, dm);
+       
+       return umap;
+}
+
+#undef TYPE
+#define TYPE isl_map
+#include "isl_power_templ.c"
+
+#undef TYPE
+#define TYPE isl_union_map
+#include "isl_power_templ.c"