3 #include "isl_map_private.h"
4 #include "isl_equalities.h"
6 /* Given a set of modulo constraints
10 * this function computes a particular solution y_0
12 * The input is given as a matrix B = [ c A ] and a vector d.
14 * The output is matrix containing the solution y_0 or
15 * a zero-column matrix if the constraints admit no integer solution.
17 * The given set of constrains is equivalent to
21 * with D = diag d and x a fresh set of variables.
22 * Reducing both c and A modulo d does not change the
23 * value of y in the solution and may lead to smaller coefficients.
24 * Let M = [ D A ] and [ H 0 ] = M U, the Hermite normal form of M.
30 * [ H 0 ] U^{-1} [ y ] = - c
33 * [ B ] = U^{-1} [ y ]
37 * so B may be chosen arbitrarily, e.g., B = 0, and then
40 * U^{-1} [ y ] = [ 0 ]
48 * If any of the coordinates of this y are non-integer
49 * then the constraints admit no integer solution and
50 * a zero-column matrix is returned.
52 static struct isl_mat *particular_solution(struct isl_mat *B, struct isl_vec *d)
55 struct isl_mat *M = NULL;
56 struct isl_mat *C = NULL;
57 struct isl_mat *U = NULL;
58 struct isl_mat *H = NULL;
59 struct isl_mat *cst = NULL;
60 struct isl_mat *T = NULL;
62 M = isl_mat_alloc(B->ctx, B->n_row, B->n_row + B->n_col - 1);
63 C = isl_mat_alloc(B->ctx, 1 + B->n_row, 1);
66 isl_int_set_si(C->row[0][0], 1);
67 for (i = 0; i < B->n_row; ++i) {
68 isl_seq_clr(M->row[i], B->n_row);
69 isl_int_set(M->row[i][i], d->block.data[i]);
70 isl_int_neg(C->row[1 + i][0], B->row[i][0]);
71 isl_int_fdiv_r(C->row[1+i][0], C->row[1+i][0], M->row[i][i]);
72 for (j = 0; j < B->n_col - 1; ++j)
73 isl_int_fdiv_r(M->row[i][B->n_row + j],
74 B->row[i][1 + j], M->row[i][i]);
76 M = isl_mat_left_hermite(M, 0, &U, NULL);
79 H = isl_mat_sub_alloc(B->ctx, M->row, 0, B->n_row, 0, B->n_row);
80 H = isl_mat_lin_to_aff(H);
81 C = isl_mat_inverse_product(H, C);
84 for (i = 0; i < B->n_row; ++i) {
85 if (!isl_int_is_divisible_by(C->row[1+i][0], C->row[0][0]))
87 isl_int_divexact(C->row[1+i][0], C->row[1+i][0], C->row[0][0]);
90 cst = isl_mat_alloc(B->ctx, B->n_row, 0);
92 cst = isl_mat_sub_alloc(C->ctx, C->row, 1, B->n_row, 0, 1);
93 T = isl_mat_sub_alloc(U->ctx, U->row, B->n_row, B->n_col - 1, 0, B->n_row);
94 cst = isl_mat_product(T, cst);
106 /* Compute and return the matrix
108 * U_1^{-1} diag(d_1, 1, ..., 1)
110 * with U_1 the unimodular completion of the first (and only) row of B.
111 * The columns of this matrix generate the lattice that satisfies
112 * the single (linear) modulo constraint.
114 static struct isl_mat *parameter_compression_1(
115 struct isl_mat *B, struct isl_vec *d)
119 U = isl_mat_alloc(B->ctx, B->n_col - 1, B->n_col - 1);
122 isl_seq_cpy(U->row[0], B->row[0] + 1, B->n_col - 1);
123 U = isl_mat_unimodular_complete(U, 1);
124 U = isl_mat_right_inverse(U);
127 isl_mat_col_mul(U, 0, d->block.data[0], 0);
128 U = isl_mat_lin_to_aff(U);
135 /* Compute a common lattice of solutions to the linear modulo
136 * constraints specified by B and d.
137 * See also the documentation of isl_mat_parameter_compression.
140 * A = [ L_1^{-T} L_2^{-T} ... L_k^{-T} ]
142 * on a common denominator. This denominator D is the lcm of modulos d.
143 * Since L_i = U_i^{-1} diag(d_i, 1, ... 1), we have
144 * L_i^{-T} = U_i^T diag(d_i, 1, ... 1)^{-T} = U_i^T diag(1/d_i, 1, ..., 1).
145 * Putting this on the common denominator, we have
146 * D * L_i^{-T} = U_i^T diag(D/d_i, D, ..., D).
148 static struct isl_mat *parameter_compression_multi(
149 struct isl_mat *B, struct isl_vec *d)
153 struct isl_mat *A = NULL, *U = NULL;
162 A = isl_mat_alloc(B->ctx, size, B->n_row * size);
163 U = isl_mat_alloc(B->ctx, size, size);
166 for (i = 0; i < B->n_row; ++i) {
167 isl_seq_cpy(U->row[0], B->row[i] + 1, size);
168 U = isl_mat_unimodular_complete(U, 1);
171 isl_int_divexact(D, D, d->block.data[i]);
172 for (k = 0; k < U->n_col; ++k)
173 isl_int_mul(A->row[k][i*size+0], D, U->row[0][k]);
174 isl_int_mul(D, D, d->block.data[i]);
175 for (j = 1; j < U->n_row; ++j)
176 for (k = 0; k < U->n_col; ++k)
177 isl_int_mul(A->row[k][i*size+j],
180 A = isl_mat_left_hermite(A, 0, NULL, NULL);
181 T = isl_mat_sub_alloc(A->ctx, A->row, 0, A->n_row, 0, A->n_row);
182 T = isl_mat_lin_to_aff(T);
183 isl_int_set(T->row[0][0], D);
184 T = isl_mat_right_inverse(T);
185 isl_assert(T->ctx, isl_int_is_one(T->row[0][0]), goto error);
186 T = isl_mat_transpose(T);
199 /* Given a set of modulo constraints
203 * this function returns an affine transformation T,
207 * that bijectively maps the integer vectors y' to integer
208 * vectors y that satisfy the modulo constraints.
210 * This function is inspired by Section 2.5.3
211 * of B. Meister, "Stating and Manipulating Periodicity in the Polytope
212 * Model. Applications to Program Analysis and Optimization".
213 * However, the implementation only follows the algorithm of that
214 * section for computing a particular solution and not for computing
215 * a general homogeneous solution. The latter is incomplete and
216 * may remove some valid solutions.
217 * Instead, we use an adaptation of the algorithm in Section 7 of
218 * B. Meister, S. Verdoolaege, "Polynomial Approximations in the Polytope
219 * Model: Bringing the Power of Quasi-Polynomials to the Masses".
221 * The input is given as a matrix B = [ c A ] and a vector d.
222 * Each element of the vector d corresponds to a row in B.
223 * The output is a lower triangular matrix.
224 * If no integer vector y satisfies the given constraints then
225 * a matrix with zero columns is returned.
227 * We first compute a particular solution y_0 to the given set of
228 * modulo constraints in particular_solution. If no such solution
229 * exists, then we return a zero-columned transformation matrix.
230 * Otherwise, we compute the generic solution to
234 * That is we want to compute G such that
238 * with y'' integer, describes the set of solutions.
240 * We first remove the common factors of each row.
241 * In particular if gcd(A_i,d_i) != 1, then we divide the whole
242 * row i (including d_i) by this common factor. If afterwards gcd(A_i) != 1,
243 * then we divide this row of A by the common factor, unless gcd(A_i) = 0.
244 * In the later case, we simply drop the row (in both A and d).
246 * If there are no rows left in A, the G is the identity matrix. Otherwise,
247 * for each row i, we now determine the lattice of integer vectors
248 * that satisfies this row. Let U_i be the unimodular extension of the
249 * row A_i. This unimodular extension exists because gcd(A_i) = 1.
250 * The first component of
254 * needs to be a multiple of d_i. Let y' = diag(d_i, 1, ..., 1) y''.
257 * y = U_i^{-1} diag(d_i, 1, ..., 1) y''
259 * for arbitrary integer vectors y''. That is, y belongs to the lattice
260 * generated by the columns of L_i = U_i^{-1} diag(d_i, 1, ..., 1).
261 * If there is only one row, then G = L_1.
263 * If there is more than one row left, we need to compute the intersection
264 * of the lattices. That is, we need to compute an L such that
266 * L = L_i L_i' for all i
268 * with L_i' some integer matrices. Let A be constructed as follows
270 * A = [ L_1^{-T} L_2^{-T} ... L_k^{-T} ]
272 * and computed the Hermite Normal Form of A = [ H 0 ] U
275 * L_i^{-T} = H U_{1,i}
279 * H^{-T} = L_i U_{1,i}^T
281 * In other words G = L = H^{-T}.
282 * To ensure that G is lower triangular, we compute and use its Hermite
285 * The affine transformation matrix returned is then
290 * as any y = y_0 + G y' with y' integer is a solution to the original
291 * modulo constraints.
293 struct isl_mat *isl_mat_parameter_compression(
294 struct isl_mat *B, struct isl_vec *d)
297 struct isl_mat *cst = NULL;
298 struct isl_mat *T = NULL;
303 isl_assert(B->ctx, B->n_row == d->size, goto error);
304 cst = particular_solution(B, d);
307 if (cst->n_col == 0) {
308 T = isl_mat_alloc(B->ctx, B->n_col, 0);
315 /* Replace a*g*row = 0 mod g*m by row = 0 mod m */
316 for (i = 0; i < B->n_row; ++i) {
317 isl_seq_gcd(B->row[i] + 1, B->n_col - 1, &D);
318 if (isl_int_is_one(D))
320 if (isl_int_is_zero(D)) {
321 B = isl_mat_drop_rows(B, i, 1);
325 isl_seq_cpy(d->block.data+i, d->block.data+i+1,
334 isl_seq_scale_down(B->row[i] + 1, B->row[i] + 1, D, B->n_col-1);
335 isl_int_gcd(D, D, d->block.data[i]);
339 isl_int_divexact(d->block.data[i], d->block.data[i], D);
343 T = isl_mat_identity(B->ctx, B->n_col);
344 else if (B->n_row == 1)
345 T = parameter_compression_1(B, d);
347 T = parameter_compression_multi(B, d);
348 T = isl_mat_left_hermite(T, 0, NULL, NULL);
351 isl_mat_sub_copy(T->ctx, T->row + 1, cst->row, cst->n_row, 0, 0, 1);
365 /* Given a set of equalities
369 * this function computes unimodular transformation from a lower-dimensional
370 * space to the original space that bijectively maps the integer points x'
371 * in the lower-dimensional space to the integer points x in the original
372 * space that satisfy the equalities.
374 * The input is given as a matrix B = [ -c M ] and the out is a
375 * matrix that maps [1 x'] to [1 x].
376 * If T2 is not NULL, then *T2 is set to a matrix mapping [1 x] to [1 x'].
378 * First compute the (left) Hermite normal form of M,
380 * M [U1 U2] = M U = H = [H1 0]
382 * M = H Q = [H1 0] [Q1]
385 * with U, Q unimodular, Q = U^{-1} (and H lower triangular).
386 * Define the transformed variables as
388 * x = [U1 U2] [ x1' ] = [U1 U2] [Q1] x
391 * The equalities then become
393 * H1 x1' - c = 0 or x1' = H1^{-1} c = c'
395 * If any of the c' is non-integer, then the original set has no
396 * integer solutions (since the x' are a unimodular transformation
398 * Otherwise, the transformation is given by
400 * x = U1 H1^{-1} c + U2 x2'
402 * The inverse transformation is simply
406 struct isl_mat *isl_mat_variable_compression(struct isl_mat *B,
410 struct isl_mat *H = NULL, *C = NULL, *H1, *U = NULL, *U1, *U2, *TC;
419 H = isl_mat_sub_alloc(B->ctx, B->row, 0, B->n_row, 1, dim);
420 H = isl_mat_left_hermite(H, 0, &U, T2);
421 if (!H || !U || (T2 && !*T2))
424 *T2 = isl_mat_drop_rows(*T2, 0, B->n_row);
425 *T2 = isl_mat_lin_to_aff(*T2);
429 C = isl_mat_alloc(B->ctx, 1+B->n_row, 1);
432 isl_int_set_si(C->row[0][0], 1);
433 isl_mat_sub_neg(C->ctx, C->row+1, B->row, B->n_row, 0, 0, 1);
434 H1 = isl_mat_sub_alloc(H->ctx, H->row, 0, H->n_row, 0, H->n_row);
435 H1 = isl_mat_lin_to_aff(H1);
436 TC = isl_mat_inverse_product(H1, C);
440 if (!isl_int_is_one(TC->row[0][0])) {
441 for (i = 0; i < B->n_row; ++i) {
442 if (!isl_int_is_divisible_by(TC->row[1+i][0], TC->row[0][0])) {
443 struct isl_ctx *ctx = B->ctx;
451 return isl_mat_alloc(ctx, 1 + dim, 0);
453 isl_seq_scale_down(TC->row[1+i], TC->row[1+i], TC->row[0][0], 1);
455 isl_int_set_si(TC->row[0][0], 1);
457 U1 = isl_mat_sub_alloc(U->ctx, U->row, 0, U->n_row, 0, B->n_row);
458 U1 = isl_mat_lin_to_aff(U1);
459 U2 = isl_mat_sub_alloc(U->ctx, U->row, 0, U->n_row,
460 B->n_row, U->n_row - B->n_row);
461 U2 = isl_mat_lin_to_aff(U2);
463 TC = isl_mat_product(U1, TC);
464 TC = isl_mat_aff_direct_sum(TC, U2);
480 /* Use the n equalities of bset to unimodularly transform the
481 * variables x such that n transformed variables x1' have a constant value
482 * and rewrite the constraints of bset in terms of the remaining
483 * transformed variables x2'. The matrix pointed to by T maps
484 * the new variables x2' back to the original variables x, while T2
485 * maps the original variables to the new variables.
487 static struct isl_basic_set *compress_variables(
488 struct isl_basic_set *bset, struct isl_mat **T, struct isl_mat **T2)
490 struct isl_mat *B, *TC;
499 isl_assert(bset->ctx, isl_basic_set_n_param(bset) == 0, goto error);
500 isl_assert(bset->ctx, bset->n_div == 0, goto error);
501 dim = isl_basic_set_n_dim(bset);
502 isl_assert(bset->ctx, bset->n_eq <= dim, goto error);
506 B = isl_mat_sub_alloc(bset->ctx, bset->eq, 0, bset->n_eq, 0, 1 + dim);
507 TC = isl_mat_variable_compression(B, T2);
510 if (TC->n_col == 0) {
516 return isl_basic_set_set_to_empty(bset);
519 bset = isl_basic_set_preimage(bset, T ? isl_mat_copy(TC) : TC);
524 isl_basic_set_free(bset);
528 struct isl_basic_set *isl_basic_set_remove_equalities(
529 struct isl_basic_set *bset, struct isl_mat **T, struct isl_mat **T2)
537 isl_assert(bset->ctx, isl_basic_set_n_param(bset) == 0, goto error);
538 bset = isl_basic_set_gauss(bset, NULL);
539 if (ISL_F_ISSET(bset, ISL_BASIC_SET_EMPTY))
541 bset = compress_variables(bset, T, T2);
544 isl_basic_set_free(bset);
549 /* Check if dimension dim belongs to a residue class
550 * i_dim \equiv r mod m
551 * with m != 1 and if so return m in *modulo and r in *residue.
552 * As a special case, when i_dim has a fixed value v, then
553 * *modulo is set to 0 and *residue to v.
555 * If i_dim does not belong to such a residue class, then *modulo
556 * is set to 1 and *residue is set to 0.
558 int isl_basic_set_dim_residue_class(struct isl_basic_set *bset,
559 int pos, isl_int *modulo, isl_int *residue)
562 struct isl_mat *H = NULL, *U = NULL, *C, *H1, *U1;
566 if (!bset || !modulo || !residue)
569 if (isl_basic_set_fast_dim_is_fixed(bset, pos, residue)) {
570 isl_int_set_si(*modulo, 0);
575 total = isl_basic_set_total_dim(bset);
576 nparam = isl_basic_set_n_param(bset);
577 H = isl_mat_sub_alloc(bset->ctx, bset->eq, 0, bset->n_eq, 1, total);
578 H = isl_mat_left_hermite(H, 0, &U, NULL);
582 isl_seq_gcd(U->row[nparam + pos]+bset->n_eq,
583 total-bset->n_eq, modulo);
584 if (isl_int_is_zero(*modulo))
585 isl_int_set_si(*modulo, 1);
586 if (isl_int_is_one(*modulo)) {
587 isl_int_set_si(*residue, 0);
593 C = isl_mat_alloc(bset->ctx, 1+bset->n_eq, 1);
596 isl_int_set_si(C->row[0][0], 1);
597 isl_mat_sub_neg(C->ctx, C->row+1, bset->eq, bset->n_eq, 0, 0, 1);
598 H1 = isl_mat_sub_alloc(H->ctx, H->row, 0, H->n_row, 0, H->n_row);
599 H1 = isl_mat_lin_to_aff(H1);
600 C = isl_mat_inverse_product(H1, C);
602 U1 = isl_mat_sub_alloc(U->ctx, U->row, nparam+pos, 1, 0, bset->n_eq);
603 U1 = isl_mat_lin_to_aff(U1);
605 C = isl_mat_product(U1, C);
608 if (!isl_int_is_divisible_by(C->row[1][0], C->row[0][0])) {
609 bset = isl_basic_set_copy(bset);
610 bset = isl_basic_set_set_to_empty(bset);
611 isl_basic_set_free(bset);
612 isl_int_set_si(*modulo, 1);
613 isl_int_set_si(*residue, 0);
616 isl_int_divexact(*residue, C->row[1][0], C->row[0][0]);
617 isl_int_fdiv_r(*residue, *residue, *modulo);