/* Construct a basic set described by the "n" equalities of "bset" starting
* at "first".
*/
static __isl_give isl_basic_set *copy_equalities(__isl_keep isl_basic_set *bset,
unsigned first, unsigned n)
{
int i, k;
isl_basic_set *eq;
unsigned total;
isl_assert(bset->ctx, bset->n_div == 0, return NULL);
total = isl_basic_set_total_dim(bset);
eq = isl_basic_set_alloc_space(isl_space_copy(bset->dim), 0, n, 0);
if (!eq)
return NULL;
for (i = 0; i < n; ++i) {
k = isl_basic_set_alloc_equality(eq);
if (k < 0)
goto error;
isl_seq_cpy(eq->eq[k], bset->eq[first + k], 1 + total);
}
return eq;
error:
isl_basic_set_free(eq);
return NULL;
}
/* Return 1 if "bmap" contains the point "point".
* "bmap" is assumed to have known divs.
* The point is first extended with the divs and then passed
* to basic_map_contains.
*/
int isl_basic_map_contains_point(__isl_keep isl_basic_map *bmap,
__isl_keep isl_point *point)
{
int i;
struct isl_vec *vec;
unsigned dim;
int contains;
if (!bmap || !point)
return -1;
isl_assert(bmap->ctx, isl_dim_equal(bmap->dim, point->dim), return -1);
if (bmap->n_div == 0)
return isl_basic_map_contains(bmap, point->vec);
dim = isl_basic_map_total_dim(bmap) - bmap->n_div;
vec = isl_vec_alloc(bmap->ctx, 1 + dim + bmap->n_div);
if (!vec)
return -1;
isl_seq_cpy(vec->el, point->vec->el, point->vec->size);
for (i = 0; i < bmap->n_div; ++i) {
isl_seq_inner_product(bmap->div[i] + 1, vec->el,
1 + dim + i, &vec->el[1+dim+i]);
isl_int_fdiv_q(vec->el[1+dim+i], vec->el[1+dim+i],
bmap->div[i][0]);
}
contains = isl_basic_map_contains(bmap, vec);
isl_vec_free(vec);
return contains;
}
/* Compute a common lattice of solutions to the linear modulo
* constraints specified by B and d.
* See also the documentation of isl_mat_parameter_compression.
* We put the matrix
*
* A = [ L_1^{-T} L_2^{-T} ... L_k^{-T} ]
*
* on a common denominator. This denominator D is the lcm of modulos d.
* Since L_i = U_i^{-1} diag(d_i, 1, ... 1), we have
* L_i^{-T} = U_i^T diag(d_i, 1, ... 1)^{-T} = U_i^T diag(1/d_i, 1, ..., 1).
* Putting this on the common denominator, we have
* D * L_i^{-T} = U_i^T diag(D/d_i, D, ..., D).
*/
static struct isl_mat *parameter_compression_multi(
struct isl_mat *B, struct isl_vec *d)
{
int i, j, k;
isl_int D;
struct isl_mat *A = NULL, *U = NULL;
struct isl_mat *T;
unsigned size;
isl_int_init(D);
isl_vec_lcm(d, &D);
size = B->n_col - 1;
A = isl_mat_alloc(B->ctx, size, B->n_row * size);
U = isl_mat_alloc(B->ctx, size, size);
if (!U || !A)
goto error;
for (i = 0; i < B->n_row; ++i) {
isl_seq_cpy(U->row[0], B->row[i] + 1, size);
U = isl_mat_unimodular_complete(U, 1);
if (!U)
goto error;
isl_int_divexact(D, D, d->block.data[i]);
for (k = 0; k < U->n_col; ++k)
isl_int_mul(A->row[k][i*size+0], D, U->row[0][k]);
isl_int_mul(D, D, d->block.data[i]);
for (j = 1; j < U->n_row; ++j)
for (k = 0; k < U->n_col; ++k)
isl_int_mul(A->row[k][i*size+j],
D, U->row[j][k]);
}
A = isl_mat_left_hermite(A, 0, NULL, NULL);
T = isl_mat_sub_alloc(A, 0, A->n_row, 0, A->n_row);
T = isl_mat_lin_to_aff(T);
if (!T)
goto error;
isl_int_set(T->row[0][0], D);
T = isl_mat_right_inverse(T);
if (!T)
goto error;
isl_assert(T->ctx, isl_int_is_one(T->row[0][0]), goto error);
T = isl_mat_transpose(T);
isl_mat_free(A);
isl_mat_free(U);
isl_int_clear(D);
return T;
error:
isl_mat_free(A);
isl_mat_free(U);
isl_int_clear(D);
return NULL;
}
static void expand_constraint(isl_vec *v, unsigned dim,
isl_int *c, int *div_map, unsigned n_div)
{
int i;
isl_seq_cpy(v->el, c, 1 + dim);
isl_seq_clr(v->el + 1 + dim, v->size - (1 + dim));
for (i = 0; i < n_div; ++i)
isl_int_set(v->el[1 + dim + div_map[i]], c[1 + dim + i]);
}
static int tab_add_divs(struct isl_tab *tab, __isl_keep isl_basic_map *bmap,
int **div_map)
{
int i, j;
struct isl_vec *vec;
unsigned total;
unsigned dim;
if (!bmap)
return -1;
if (!bmap->n_div)
return 0;
if (!*div_map)
*div_map = isl_alloc_array(bmap->ctx, int, bmap->n_div);
if (!*div_map)
return -1;
total = isl_basic_map_total_dim(tab->bmap);
dim = total - tab->bmap->n_div;
vec = isl_vec_alloc(bmap->ctx, 2 + total + bmap->n_div);
if (!vec)
return -1;
for (i = 0; i < bmap->n_div; ++i) {
isl_seq_cpy(vec->el, bmap->div[i], 2 + dim);
isl_seq_clr(vec->el + 2 + dim, tab->bmap->n_div);
for (j = 0; j < i; ++j)
isl_int_set(vec->el[2 + dim + (*div_map)[j]],
bmap->div[i][2 + dim + j]);
for (j = 0; j < tab->bmap->n_div; ++j)
if (isl_seq_eq(tab->bmap->div[j],
vec->el, 2 + dim + tab->bmap->n_div))
break;
(*div_map)[i] = j;
if (j == tab->bmap->n_div) {
vec->size = 2 + dim + tab->bmap->n_div;
if (isl_tab_add_div(tab, vec) < 0)
goto error;
}
}
isl_vec_free(vec);
return 0;
error:
isl_vec_free(vec);
return -1;
}
/* Compute and return the matrix
*
* U_1^{-1} diag(d_1, 1, ..., 1)
*
* with U_1 the unimodular completion of the first (and only) row of B.
* The columns of this matrix generate the lattice that satisfies
* the single (linear) modulo constraint.
*/
static struct isl_mat *parameter_compression_1(
struct isl_mat *B, struct isl_vec *d)
{
struct isl_mat *U;
U = isl_mat_alloc(B->ctx, B->n_col - 1, B->n_col - 1);
if (!U)
return NULL;
isl_seq_cpy(U->row[0], B->row[0] + 1, B->n_col - 1);
U = isl_mat_unimodular_complete(U, 1);
U = isl_mat_right_inverse(U);
if (!U)
return NULL;
isl_mat_col_mul(U, 0, d->block.data[0], 0);
U = isl_mat_lin_to_aff(U);
return U;
}
static struct isl_vec *isl_vec_lin_to_aff(struct isl_vec *vec)
{
struct isl_vec *aff;
if (!vec)
return NULL;
aff = isl_vec_alloc(vec->ctx, 1 + vec->size);
if (!aff)
goto error;
isl_int_set_si(aff->el[0], 0);
isl_seq_cpy(aff->el + 1, vec->el, vec->size);
isl_vec_free(vec);
return aff;
error:
isl_vec_free(vec);
return NULL;
}
/* Detect and make explicit all equalities satisfied by the (integer)
* points in bmap.
*/
struct isl_basic_map *isl_basic_map_detect_equalities(
struct isl_basic_map *bmap)
{
int i, j;
struct isl_basic_set *hull = NULL;
if (!bmap)
return NULL;
if (bmap->n_ineq == 0)
return bmap;
if (ISL_F_ISSET(bmap, ISL_BASIC_MAP_EMPTY))
return bmap;
if (ISL_F_ISSET(bmap, ISL_BASIC_MAP_ALL_EQUALITIES))
return bmap;
if (ISL_F_ISSET(bmap, ISL_BASIC_MAP_RATIONAL))
return isl_basic_map_implicit_equalities(bmap);
hull = equalities_in_underlying_set(isl_basic_map_copy(bmap));
if (!hull)
goto error;
if (ISL_F_ISSET(hull, ISL_BASIC_SET_EMPTY)) {
isl_basic_set_free(hull);
return isl_basic_map_set_to_empty(bmap);
}
bmap = isl_basic_map_extend_dim(bmap, isl_dim_copy(bmap->dim), 0,
hull->n_eq, 0);
for (i = 0; i < hull->n_eq; ++i) {
j = isl_basic_map_alloc_equality(bmap);
if (j < 0)
goto error;
isl_seq_cpy(bmap->eq[j], hull->eq[i],
1 + isl_basic_set_total_dim(hull));
}
isl_vec_free(bmap->sample);
bmap->sample = isl_vec_copy(hull->sample);
isl_basic_set_free(hull);
ISL_F_SET(bmap, ISL_BASIC_MAP_NO_IMPLICIT | ISL_BASIC_MAP_ALL_EQUALITIES);
bmap = isl_basic_map_simplify(bmap);
return isl_basic_map_finalize(bmap);
error:
isl_basic_set_free(hull);
isl_basic_map_free(bmap);
return NULL;
}
static int scan_samples_add_sample(struct isl_scan_callback *cb,
__isl_take isl_vec *sample)
{
struct scan_samples *ss = (struct scan_samples *)cb;
ss->samples = isl_mat_extend(ss->samples, ss->samples->n_row + 1,
ss->samples->n_col);
if (!ss->samples)
goto error;
isl_seq_cpy(ss->samples->row[ss->samples->n_row - 1],
sample->el, sample->size);
isl_vec_free(sample);
return 0;
error:
isl_vec_free(sample);
return -1;
}
/* Given an unbounded tableau and an integer point satisfying the tableau,
* construct an initial affine hull containing the recession cone
* shifted to the given point.
*
* The unbounded directions are taken from the last rows of the basis,
* which is assumed to have been initialized appropriately.
*/
static __isl_give isl_basic_set *initial_hull(struct isl_tab *tab,
__isl_take isl_vec *vec)
{
int i;
int k;
struct isl_basic_set *bset = NULL;
struct isl_ctx *ctx;
unsigned dim;
if (!vec || !tab)
return NULL;
ctx = vec->ctx;
isl_assert(ctx, vec->size != 0, goto error);
bset = isl_basic_set_alloc(ctx, 0, vec->size - 1, 0, vec->size - 1, 0);
if (!bset)
goto error;
dim = isl_basic_set_n_dim(bset) - tab->n_unbounded;
for (i = 0; i < dim; ++i) {
k = isl_basic_set_alloc_equality(bset);
if (k < 0)
goto error;
isl_seq_cpy(bset->eq[k] + 1, tab->basis->row[1 + i] + 1,
vec->size - 1);
isl_seq_inner_product(bset->eq[k] + 1, vec->el +1,
vec->size - 1, &bset->eq[k][0]);
isl_int_neg(bset->eq[k][0], bset->eq[k][0]);
}
bset->sample = vec;
bset = isl_basic_set_gauss(bset, NULL);
return bset;
error:
isl_basic_set_free(bset);
isl_vec_free(vec);
return NULL;
}
/* Add stride constraints to "bset" based on the inverse mapping
* that was plugged in. In particular, if morph maps x' to x,
* the the constraints of the original input
*
* A x' + b >= 0
*
* have been rewritten to
*
* A inv x + b >= 0
*
* However, this substitution may loose information on the integrality of x',
* so we need to impose that
*
* inv x
*
* is integral. If inv = B/d, this means that we need to impose that
*
* B x = 0 mod d
*
* or
*
* exists alpha in Z^m: B x = d alpha
*
*/
static __isl_give isl_basic_set *add_strides(__isl_take isl_basic_set *bset,
__isl_keep isl_morph *morph)
{
int i, div, k;
isl_int gcd;
if (isl_int_is_one(morph->inv->row[0][0]))
return bset;
isl_int_init(gcd);
for (i = 0; 1 + i < morph->inv->n_row; ++i) {
isl_seq_gcd(morph->inv->row[1 + i], morph->inv->n_col, &gcd);
if (isl_int_is_divisible_by(gcd, morph->inv->row[0][0]))
continue;
div = isl_basic_set_alloc_div(bset);
if (div < 0)
goto error;
k = isl_basic_set_alloc_equality(bset);
if (k < 0)
goto error;
isl_seq_cpy(bset->eq[k], morph->inv->row[1 + i],
morph->inv->n_col);
isl_seq_clr(bset->eq[k] + morph->inv->n_col, bset->n_div);
isl_int_set(bset->eq[k][morph->inv->n_col + div],
morph->inv->row[0][0]);
}
isl_int_clear(gcd);
return bset;
error:
isl_int_clear(gcd);
isl_basic_set_free(bset);
return NULL;
}
/* Apply the morphism to the basic set.
* We basically just compute the preimage of "bset" under the inverse mapping
* in morph, add in stride constraints and intersect with the range
* of the morphism.
*/
__isl_give isl_basic_set *isl_morph_basic_set(__isl_take isl_morph *morph,
__isl_take isl_basic_set *bset)
{
isl_basic_set *res = NULL;
isl_mat *mat = NULL;
int i, k;
int max_stride;
if (!morph || !bset)
goto error;
isl_assert(bset->ctx, isl_space_is_equal(bset->dim, morph->dom->dim),
goto error);
max_stride = morph->inv->n_row - 1;
if (isl_int_is_one(morph->inv->row[0][0]))
max_stride = 0;
res = isl_basic_set_alloc_space(isl_space_copy(morph->ran->dim),
bset->n_div + max_stride, bset->n_eq + max_stride, bset->n_ineq);
for (i = 0; i < bset->n_div; ++i)
if (isl_basic_set_alloc_div(res) < 0)
goto error;
mat = isl_mat_sub_alloc6(bset->ctx, bset->eq, 0, bset->n_eq,
0, morph->inv->n_row);
mat = isl_mat_product(mat, isl_mat_copy(morph->inv));
if (!mat)
goto error;
for (i = 0; i < bset->n_eq; ++i) {
k = isl_basic_set_alloc_equality(res);
if (k < 0)
goto error;
isl_seq_cpy(res->eq[k], mat->row[i], mat->n_col);
isl_seq_scale(res->eq[k] + mat->n_col, bset->eq[i] + mat->n_col,
morph->inv->row[0][0], bset->n_div);
}
isl_mat_free(mat);
mat = isl_mat_sub_alloc6(bset->ctx, bset->ineq, 0, bset->n_ineq,
0, morph->inv->n_row);
mat = isl_mat_product(mat, isl_mat_copy(morph->inv));
if (!mat)
goto error;
for (i = 0; i < bset->n_ineq; ++i) {
k = isl_basic_set_alloc_inequality(res);
if (k < 0)
goto error;
isl_seq_cpy(res->ineq[k], mat->row[i], mat->n_col);
isl_seq_scale(res->ineq[k] + mat->n_col,
bset->ineq[i] + mat->n_col,
morph->inv->row[0][0], bset->n_div);
}
isl_mat_free(mat);
mat = isl_mat_sub_alloc6(bset->ctx, bset->div, 0, bset->n_div,
1, morph->inv->n_row);
mat = isl_mat_product(mat, isl_mat_copy(morph->inv));
if (!mat)
goto error;
for (i = 0; i < bset->n_div; ++i) {
isl_int_mul(res->div[i][0],
morph->inv->row[0][0], bset->div[i][0]);
isl_seq_cpy(res->div[i] + 1, mat->row[i], mat->n_col);
isl_seq_scale(res->div[i] + 1 + mat->n_col,
bset->div[i] + 1 + mat->n_col,
morph->inv->row[0][0], bset->n_div);
}
isl_mat_free(mat);
res = add_strides(res, morph);
if (isl_basic_set_is_rational(bset))
res = isl_basic_set_set_rational(res);
res = isl_basic_set_simplify(res);
res = isl_basic_set_finalize(res);
res = isl_basic_set_intersect(res, isl_basic_set_copy(morph->ran));
isl_morph_free(morph);
isl_basic_set_free(bset);
return res;
error:
isl_mat_free(mat);
isl_morph_free(morph);
isl_basic_set_free(bset);
isl_basic_set_free(res);
return NULL;
}
/* Given a set of modulo constraints
*
* c + A y = 0 mod d
*
* this function returns an affine transformation T,
*
* y = T y'
*
* that bijectively maps the integer vectors y' to integer
* vectors y that satisfy the modulo constraints.
*
* This function is inspired by Section 2.5.3
* of B. Meister, "Stating and Manipulating Periodicity in the Polytope
* Model. Applications to Program Analysis and Optimization".
* However, the implementation only follows the algorithm of that
* section for computing a particular solution and not for computing
* a general homogeneous solution. The latter is incomplete and
* may remove some valid solutions.
* Instead, we use an adaptation of the algorithm in Section 7 of
* B. Meister, S. Verdoolaege, "Polynomial Approximations in the Polytope
* Model: Bringing the Power of Quasi-Polynomials to the Masses".
*
* The input is given as a matrix B = [ c A ] and a vector d.
* Each element of the vector d corresponds to a row in B.
* The output is a lower triangular matrix.
* If no integer vector y satisfies the given constraints then
* a matrix with zero columns is returned.
*
* We first compute a particular solution y_0 to the given set of
* modulo constraints in particular_solution. If no such solution
* exists, then we return a zero-columned transformation matrix.
* Otherwise, we compute the generic solution to
*
* A y = 0 mod d
*
* That is we want to compute G such that
*
* y = G y''
*
* with y'' integer, describes the set of solutions.
*
* We first remove the common factors of each row.
* In particular if gcd(A_i,d_i) != 1, then we divide the whole
* row i (including d_i) by this common factor. If afterwards gcd(A_i) != 1,
* then we divide this row of A by the common factor, unless gcd(A_i) = 0.
* In the later case, we simply drop the row (in both A and d).
*
* If there are no rows left in A, then G is the identity matrix. Otherwise,
* for each row i, we now determine the lattice of integer vectors
* that satisfies this row. Let U_i be the unimodular extension of the
* row A_i. This unimodular extension exists because gcd(A_i) = 1.
* The first component of
*
* y' = U_i y
*
* needs to be a multiple of d_i. Let y' = diag(d_i, 1, ..., 1) y''.
* Then,
*
* y = U_i^{-1} diag(d_i, 1, ..., 1) y''
*
* for arbitrary integer vectors y''. That is, y belongs to the lattice
* generated by the columns of L_i = U_i^{-1} diag(d_i, 1, ..., 1).
* If there is only one row, then G = L_1.
*
* If there is more than one row left, we need to compute the intersection
* of the lattices. That is, we need to compute an L such that
*
* L = L_i L_i' for all i
*
* with L_i' some integer matrices. Let A be constructed as follows
*
* A = [ L_1^{-T} L_2^{-T} ... L_k^{-T} ]
*
* and computed the Hermite Normal Form of A = [ H 0 ] U
* Then,
*
* L_i^{-T} = H U_{1,i}
*
* or
*
* H^{-T} = L_i U_{1,i}^T
*
* In other words G = L = H^{-T}.
* To ensure that G is lower triangular, we compute and use its Hermite
* normal form.
*
* The affine transformation matrix returned is then
*
* [ 1 0 ]
* [ y_0 G ]
*
* as any y = y_0 + G y' with y' integer is a solution to the original
* modulo constraints.
*/
struct isl_mat *isl_mat_parameter_compression(
struct isl_mat *B, struct isl_vec *d)
{
int i;
struct isl_mat *cst = NULL;
struct isl_mat *T = NULL;
isl_int D;
if (!B || !d)
goto error;
isl_assert(B->ctx, B->n_row == d->size, goto error);
cst = particular_solution(B, d);
if (!cst)
goto error;
if (cst->n_col == 0) {
T = isl_mat_alloc(B->ctx, B->n_col, 0);
isl_mat_free(cst);
isl_mat_free(B);
isl_vec_free(d);
return T;
}
isl_int_init(D);
/* Replace a*g*row = 0 mod g*m by row = 0 mod m */
for (i = 0; i < B->n_row; ++i) {
isl_seq_gcd(B->row[i] + 1, B->n_col - 1, &D);
if (isl_int_is_one(D))
continue;
if (isl_int_is_zero(D)) {
B = isl_mat_drop_rows(B, i, 1);
d = isl_vec_cow(d);
if (!B || !d)
goto error2;
isl_seq_cpy(d->block.data+i, d->block.data+i+1,
d->size - (i+1));
d->size--;
i--;
continue;
}
B = isl_mat_cow(B);
if (!B)
goto error2;
isl_seq_scale_down(B->row[i] + 1, B->row[i] + 1, D, B->n_col-1);
isl_int_gcd(D, D, d->block.data[i]);
d = isl_vec_cow(d);
if (!d)
goto error2;
isl_int_divexact(d->block.data[i], d->block.data[i], D);
}
isl_int_clear(D);
if (B->n_row == 0)
T = isl_mat_identity(B->ctx, B->n_col);
else if (B->n_row == 1)
T = parameter_compression_1(B, d);
else
T = parameter_compression_multi(B, d);
T = isl_mat_left_hermite(T, 0, NULL, NULL);
if (!T)
goto error;
isl_mat_sub_copy(T->ctx, T->row + 1, cst->row, cst->n_row, 0, 0, 1);
isl_mat_free(cst);
isl_mat_free(B);
isl_vec_free(d);
return T;
error2:
isl_int_clear(D);
error:
isl_mat_free(cst);
isl_mat_free(B);
isl_vec_free(d);
return NULL;
}
/* Given a tableau of a set and a tableau of the corresponding
* recession cone, detect and add all equalities to the tableau.
* If the tableau is bounded, then we can simply keep the
* tableau in its state after the return from extend_affine_hull.
* However, if the tableau is unbounded, then
* isl_tab_set_initial_basis_with_cone will add some additional
* constraints to the tableau that have to be removed again.
* In this case, we therefore rollback to the state before
* any constraints were added and then add the equalities back in.
*/
struct isl_tab *isl_tab_detect_equalities(struct isl_tab *tab,
struct isl_tab *tab_cone)
{
int j;
struct isl_vec *sample;
struct isl_basic_set *hull;
struct isl_tab_undo *snap;
if (!tab || !tab_cone)
goto error;
snap = isl_tab_snap(tab);
isl_mat_free(tab->basis);
tab->basis = NULL;
isl_assert(tab->mat->ctx, tab->bmap, goto error);
isl_assert(tab->mat->ctx, tab->samples, goto error);
isl_assert(tab->mat->ctx, tab->samples->n_col == 1 + tab->n_var, goto error);
isl_assert(tab->mat->ctx, tab->n_sample > tab->n_outside, goto error);
if (isl_tab_set_initial_basis_with_cone(tab, tab_cone) < 0)
goto error;
sample = isl_vec_alloc(tab->mat->ctx, 1 + tab->n_var);
if (!sample)
goto error;
isl_seq_cpy(sample->el, tab->samples->row[tab->n_outside], sample->size);
isl_vec_free(tab->bmap->sample);
tab->bmap->sample = isl_vec_copy(sample);
if (tab->n_unbounded == 0)
hull = isl_basic_set_from_vec(isl_vec_copy(sample));
else
hull = initial_hull(tab, isl_vec_copy(sample));
for (j = tab->n_outside + 1; j < tab->n_sample; ++j) {
isl_seq_cpy(sample->el, tab->samples->row[j], sample->size);
hull = affine_hull(hull,
isl_basic_set_from_vec(isl_vec_copy(sample)));
}
isl_vec_free(sample);
hull = extend_affine_hull(tab, hull);
if (!hull)
goto error;
if (tab->n_unbounded == 0) {
isl_basic_set_free(hull);
return tab;
}
if (isl_tab_rollback(tab, snap) < 0)
goto error;
if (hull->n_eq > tab->n_zero) {
for (j = 0; j < hull->n_eq; ++j) {
isl_seq_normalize(tab->mat->ctx, hull->eq[j], 1 + tab->n_var);
if (isl_tab_add_eq(tab, hull->eq[j]) < 0)
goto error;
}
}
isl_basic_set_free(hull);
return tab;
error:
isl_tab_free(tab);
return NULL;
}