Home | History | Annotate | Line # | Download | only in dist
      1 /*
      2  * Copyright 2008-2009 Katholieke Universiteit Leuven
      3  * Copyright 2010      INRIA Saclay
      4  *
      5  * Use of this software is governed by the MIT license
      6  *
      7  * Written by Sven Verdoolaege, K.U.Leuven, Departement
      8  * Computerwetenschappen, Celestijnenlaan 200A, B-3001 Leuven, Belgium
      9  * and INRIA Saclay - Ile-de-France, Parc Club Orsay Universite,
     10  * ZAC des vignes, 4 rue Jacques Monod, 91893 Orsay, France
     11  */
     12 
     13 #include <isl_mat_private.h>
     14 #include <isl_vec_private.h>
     15 #include <isl_seq.h>
     16 #include "isl_map_private.h"
     17 #include "isl_equalities.h"
     18 #include <isl_val_private.h>
     19 
     20 /* Given a set of modulo constraints
     21  *
     22  *		c + A y = 0 mod d
     23  *
     24  * this function computes a particular solution y_0
     25  *
     26  * The input is given as a matrix B = [ c A ] and a vector d.
     27  *
     28  * The output is matrix containing the solution y_0 or
     29  * a zero-column matrix if the constraints admit no integer solution.
     30  *
     31  * The given set of constrains is equivalent to
     32  *
     33  *		c + A y = -D x
     34  *
     35  * with D = diag d and x a fresh set of variables.
     36  * Reducing both c and A modulo d does not change the
     37  * value of y in the solution and may lead to smaller coefficients.
     38  * Let M = [ D A ] and [ H 0 ] = M U, the Hermite normal form of M.
     39  * Then
     40  *		  [ x ]
     41  *		M [ y ] = - c
     42  * and so
     43  *		               [ x ]
     44  *		[ H 0 ] U^{-1} [ y ] = - c
     45  * Let
     46  *		[ A ]          [ x ]
     47  *		[ B ] = U^{-1} [ y ]
     48  * then
     49  *		H A + 0 B = -c
     50  *
     51  * so B may be chosen arbitrarily, e.g., B = 0, and then
     52  *
     53  *		       [ x ] = [ -c ]
     54  *		U^{-1} [ y ] = [  0 ]
     55  * or
     56  *		[ x ]     [ -c ]
     57  *		[ y ] = U [  0 ]
     58  * specifically,
     59  *
     60  *		y = U_{2,1} (-c)
     61  *
     62  * If any of the coordinates of this y are non-integer
     63  * then the constraints admit no integer solution and
     64  * a zero-column matrix is returned.
     65  */
     66 static __isl_give isl_mat *particular_solution(__isl_keep isl_mat *B,
     67 	__isl_keep isl_vec *d)
     68 {
     69 	int i, j;
     70 	struct isl_mat *M = NULL;
     71 	struct isl_mat *C = NULL;
     72 	struct isl_mat *U = NULL;
     73 	struct isl_mat *H = NULL;
     74 	struct isl_mat *cst = NULL;
     75 	struct isl_mat *T = NULL;
     76 
     77 	M = isl_mat_alloc(B->ctx, B->n_row, B->n_row + B->n_col - 1);
     78 	C = isl_mat_alloc(B->ctx, 1 + B->n_row, 1);
     79 	if (!M || !C)
     80 		goto error;
     81 	isl_int_set_si(C->row[0][0], 1);
     82 	for (i = 0; i < B->n_row; ++i) {
     83 		isl_seq_clr(M->row[i], B->n_row);
     84 		isl_int_set(M->row[i][i], d->block.data[i]);
     85 		isl_int_neg(C->row[1 + i][0], B->row[i][0]);
     86 		isl_int_fdiv_r(C->row[1+i][0], C->row[1+i][0], M->row[i][i]);
     87 		for (j = 0; j < B->n_col - 1; ++j)
     88 			isl_int_fdiv_r(M->row[i][B->n_row + j],
     89 					B->row[i][1 + j], M->row[i][i]);
     90 	}
     91 	M = isl_mat_left_hermite(M, 0, &U, NULL);
     92 	if (!M || !U)
     93 		goto error;
     94 	H = isl_mat_sub_alloc(M, 0, B->n_row, 0, B->n_row);
     95 	H = isl_mat_lin_to_aff(H);
     96 	C = isl_mat_inverse_product(H, C);
     97 	if (!C)
     98 		goto error;
     99 	for (i = 0; i < B->n_row; ++i) {
    100 		if (!isl_int_is_divisible_by(C->row[1+i][0], C->row[0][0]))
    101 			break;
    102 		isl_int_divexact(C->row[1+i][0], C->row[1+i][0], C->row[0][0]);
    103 	}
    104 	if (i < B->n_row)
    105 		cst = isl_mat_alloc(B->ctx, B->n_row, 0);
    106 	else
    107 		cst = isl_mat_sub_alloc(C, 1, B->n_row, 0, 1);
    108 	T = isl_mat_sub_alloc(U, B->n_row, B->n_col - 1, 0, B->n_row);
    109 	cst = isl_mat_product(T, cst);
    110 	isl_mat_free(M);
    111 	isl_mat_free(C);
    112 	isl_mat_free(U);
    113 	return cst;
    114 error:
    115 	isl_mat_free(M);
    116 	isl_mat_free(C);
    117 	isl_mat_free(U);
    118 	return NULL;
    119 }
    120 
    121 /* Compute and return the matrix
    122  *
    123  *		U_1^{-1} diag(d_1, 1, ..., 1)
    124  *
    125  * with U_1 the unimodular completion of the first (and only) row of B.
    126  * The columns of this matrix generate the lattice that satisfies
    127  * the single (linear) modulo constraint.
    128  */
    129 static __isl_take isl_mat *parameter_compression_1(__isl_keep isl_mat *B,
    130 	__isl_keep isl_vec *d)
    131 {
    132 	struct isl_mat *U;
    133 
    134 	U = isl_mat_alloc(B->ctx, B->n_col - 1, B->n_col - 1);
    135 	if (!U)
    136 		return NULL;
    137 	isl_seq_cpy(U->row[0], B->row[0] + 1, B->n_col - 1);
    138 	U = isl_mat_unimodular_complete(U, 1);
    139 	U = isl_mat_right_inverse(U);
    140 	if (!U)
    141 		return NULL;
    142 	isl_mat_col_mul(U, 0, d->block.data[0], 0);
    143 	U = isl_mat_lin_to_aff(U);
    144 	return U;
    145 }
    146 
    147 /* Compute a common lattice of solutions to the linear modulo
    148  * constraints specified by B and d.
    149  * See also the documentation of isl_mat_parameter_compression.
    150  * We put the matrix
    151  *
    152  *		A = [ L_1^{-T} L_2^{-T} ... L_k^{-T} ]
    153  *
    154  * on a common denominator.  This denominator D is the lcm of modulos d.
    155  * Since L_i = U_i^{-1} diag(d_i, 1, ... 1), we have
    156  * L_i^{-T} = U_i^T diag(d_i, 1, ... 1)^{-T} = U_i^T diag(1/d_i, 1, ..., 1).
    157  * Putting this on the common denominator, we have
    158  * D * L_i^{-T} = U_i^T diag(D/d_i, D, ..., D).
    159  */
    160 static __isl_give isl_mat *parameter_compression_multi(__isl_keep isl_mat *B,
    161 	__isl_keep isl_vec *d)
    162 {
    163 	int i, j, k;
    164 	isl_int D;
    165 	struct isl_mat *A = NULL, *U = NULL;
    166 	struct isl_mat *T;
    167 	unsigned size;
    168 
    169 	isl_int_init(D);
    170 
    171 	isl_vec_lcm(d, &D);
    172 
    173 	size = B->n_col - 1;
    174 	A = isl_mat_alloc(B->ctx, size, B->n_row * size);
    175 	U = isl_mat_alloc(B->ctx, size, size);
    176 	if (!U || !A)
    177 		goto error;
    178 	for (i = 0; i < B->n_row; ++i) {
    179 		isl_seq_cpy(U->row[0], B->row[i] + 1, size);
    180 		U = isl_mat_unimodular_complete(U, 1);
    181 		if (!U)
    182 			goto error;
    183 		isl_int_divexact(D, D, d->block.data[i]);
    184 		for (k = 0; k < U->n_col; ++k)
    185 			isl_int_mul(A->row[k][i*size+0], D, U->row[0][k]);
    186 		isl_int_mul(D, D, d->block.data[i]);
    187 		for (j = 1; j < U->n_row; ++j)
    188 			for (k = 0; k < U->n_col; ++k)
    189 				isl_int_mul(A->row[k][i*size+j],
    190 						D, U->row[j][k]);
    191 	}
    192 	A = isl_mat_left_hermite(A, 0, NULL, NULL);
    193 	T = isl_mat_sub_alloc(A, 0, A->n_row, 0, A->n_row);
    194 	T = isl_mat_lin_to_aff(T);
    195 	if (!T)
    196 		goto error;
    197 	isl_int_set(T->row[0][0], D);
    198 	T = isl_mat_right_inverse(T);
    199 	if (!T)
    200 		goto error;
    201 	isl_assert(T->ctx, isl_int_is_one(T->row[0][0]), goto error);
    202 	T = isl_mat_transpose(T);
    203 	isl_mat_free(A);
    204 	isl_mat_free(U);
    205 
    206 	isl_int_clear(D);
    207 	return T;
    208 error:
    209 	isl_mat_free(A);
    210 	isl_mat_free(U);
    211 	isl_int_clear(D);
    212 	return NULL;
    213 }
    214 
    215 /* Given a set of modulo constraints
    216  *
    217  *		c + A y = 0 mod d
    218  *
    219  * this function returns an affine transformation T,
    220  *
    221  *		y = T y'
    222  *
    223  * that bijectively maps the integer vectors y' to integer
    224  * vectors y that satisfy the modulo constraints.
    225  *
    226  * This function is inspired by Section 2.5.3
    227  * of B. Meister, "Stating and Manipulating Periodicity in the Polytope
    228  * Model.  Applications to Program Analysis and Optimization".
    229  * However, the implementation only follows the algorithm of that
    230  * section for computing a particular solution and not for computing
    231  * a general homogeneous solution.  The latter is incomplete and
    232  * may remove some valid solutions.
    233  * Instead, we use an adaptation of the algorithm in Section 7 of
    234  * B. Meister, S. Verdoolaege, "Polynomial Approximations in the Polytope
    235  * Model: Bringing the Power of Quasi-Polynomials to the Masses".
    236  *
    237  * The input is given as a matrix B = [ c A ] and a vector d.
    238  * Each element of the vector d corresponds to a row in B.
    239  * The output is a lower triangular matrix.
    240  * If no integer vector y satisfies the given constraints then
    241  * a matrix with zero columns is returned.
    242  *
    243  * We first compute a particular solution y_0 to the given set of
    244  * modulo constraints in particular_solution.  If no such solution
    245  * exists, then we return a zero-columned transformation matrix.
    246  * Otherwise, we compute the generic solution to
    247  *
    248  *		A y = 0 mod d
    249  *
    250  * That is we want to compute G such that
    251  *
    252  *		y = G y''
    253  *
    254  * with y'' integer, describes the set of solutions.
    255  *
    256  * We first remove the common factors of each row.
    257  * In particular if gcd(A_i,d_i) != 1, then we divide the whole
    258  * row i (including d_i) by this common factor.  If afterwards gcd(A_i) != 1,
    259  * then we divide this row of A by the common factor, unless gcd(A_i) = 0.
    260  * In the later case, we simply drop the row (in both A and d).
    261  *
    262  * If there are no rows left in A, then G is the identity matrix. Otherwise,
    263  * for each row i, we now determine the lattice of integer vectors
    264  * that satisfies this row.  Let U_i be the unimodular extension of the
    265  * row A_i.  This unimodular extension exists because gcd(A_i) = 1.
    266  * The first component of
    267  *
    268  *		y' = U_i y
    269  *
    270  * needs to be a multiple of d_i.  Let y' = diag(d_i, 1, ..., 1) y''.
    271  * Then,
    272  *
    273  *		y = U_i^{-1} diag(d_i, 1, ..., 1) y''
    274  *
    275  * for arbitrary integer vectors y''.  That is, y belongs to the lattice
    276  * generated by the columns of L_i = U_i^{-1} diag(d_i, 1, ..., 1).
    277  * If there is only one row, then G = L_1.
    278  *
    279  * If there is more than one row left, we need to compute the intersection
    280  * of the lattices.  That is, we need to compute an L such that
    281  *
    282  *		L = L_i L_i'	for all i
    283  *
    284  * with L_i' some integer matrices.  Let A be constructed as follows
    285  *
    286  *		A = [ L_1^{-T} L_2^{-T} ... L_k^{-T} ]
    287  *
    288  * and computed the Hermite Normal Form of A = [ H 0 ] U
    289  * Then,
    290  *
    291  *		L_i^{-T} = H U_{1,i}
    292  *
    293  * or
    294  *
    295  *		H^{-T} = L_i U_{1,i}^T
    296  *
    297  * In other words G = L = H^{-T}.
    298  * To ensure that G is lower triangular, we compute and use its Hermite
    299  * normal form.
    300  *
    301  * The affine transformation matrix returned is then
    302  *
    303  *		[  1   0  ]
    304  *		[ y_0  G  ]
    305  *
    306  * as any y = y_0 + G y' with y' integer is a solution to the original
    307  * modulo constraints.
    308  */
    309 __isl_give isl_mat *isl_mat_parameter_compression(__isl_take isl_mat *B,
    310 	__isl_take isl_vec *d)
    311 {
    312 	int i;
    313 	struct isl_mat *cst = NULL;
    314 	struct isl_mat *T = NULL;
    315 	isl_int D;
    316 
    317 	if (!B || !d)
    318 		goto error;
    319 	isl_assert(B->ctx, B->n_row == d->size, goto error);
    320 	cst = particular_solution(B, d);
    321 	if (!cst)
    322 		goto error;
    323 	if (cst->n_col == 0) {
    324 		T = isl_mat_alloc(B->ctx, B->n_col, 0);
    325 		isl_mat_free(cst);
    326 		isl_mat_free(B);
    327 		isl_vec_free(d);
    328 		return T;
    329 	}
    330 	isl_int_init(D);
    331 	/* Replace a*g*row = 0 mod g*m by row = 0 mod m */
    332 	for (i = 0; i < B->n_row; ++i) {
    333 		isl_seq_gcd(B->row[i] + 1, B->n_col - 1, &D);
    334 		if (isl_int_is_one(D))
    335 			continue;
    336 		if (isl_int_is_zero(D)) {
    337 			B = isl_mat_drop_rows(B, i, 1);
    338 			d = isl_vec_cow(d);
    339 			if (!B || !d)
    340 				goto error2;
    341 			isl_seq_cpy(d->block.data+i, d->block.data+i+1,
    342 							d->size - (i+1));
    343 			d->size--;
    344 			i--;
    345 			continue;
    346 		}
    347 		B = isl_mat_cow(B);
    348 		if (!B)
    349 			goto error2;
    350 		isl_seq_scale_down(B->row[i] + 1, B->row[i] + 1, D, B->n_col-1);
    351 		isl_int_gcd(D, D, d->block.data[i]);
    352 		d = isl_vec_cow(d);
    353 		if (!d)
    354 			goto error2;
    355 		isl_int_divexact(d->block.data[i], d->block.data[i], D);
    356 	}
    357 	isl_int_clear(D);
    358 	if (B->n_row == 0)
    359 		T = isl_mat_identity(B->ctx, B->n_col);
    360 	else if (B->n_row == 1)
    361 		T = parameter_compression_1(B, d);
    362 	else
    363 		T = parameter_compression_multi(B, d);
    364 	T = isl_mat_left_hermite(T, 0, NULL, NULL);
    365 	if (!T)
    366 		goto error;
    367 	isl_mat_sub_copy(T->ctx, T->row + 1, cst->row, cst->n_row, 0, 0, 1);
    368 	isl_mat_free(cst);
    369 	isl_mat_free(B);
    370 	isl_vec_free(d);
    371 	return T;
    372 error2:
    373 	isl_int_clear(D);
    374 error:
    375 	isl_mat_free(cst);
    376 	isl_mat_free(B);
    377 	isl_vec_free(d);
    378 	return NULL;
    379 }
    380 
    381 /* Given a set of equalities
    382  *
    383  *		B(y) + A x = 0						(*)
    384  *
    385  * compute and return an affine transformation T,
    386  *
    387  *		y = T y'
    388  *
    389  * that bijectively maps the integer vectors y' to integer
    390  * vectors y that satisfy the modulo constraints for some value of x.
    391  *
    392  * Let [H 0] be the Hermite Normal Form of A, i.e.,
    393  *
    394  *		A = [H 0] Q
    395  *
    396  * Then y is a solution of (*) iff
    397  *
    398  *		H^-1 B(y) (= - [I 0] Q x)
    399  *
    400  * is an integer vector.  Let d be the common denominator of H^-1.
    401  * We impose
    402  *
    403  *		d H^-1 B(y) = 0 mod d
    404  *
    405  * and compute the solution using isl_mat_parameter_compression.
    406  */
    407 __isl_give isl_mat *isl_mat_parameter_compression_ext(__isl_take isl_mat *B,
    408 	__isl_take isl_mat *A)
    409 {
    410 	isl_ctx *ctx;
    411 	isl_vec *d;
    412 	int n_row, n_col;
    413 
    414 	if (!A)
    415 		return isl_mat_free(B);
    416 
    417 	ctx = isl_mat_get_ctx(A);
    418 	n_row = A->n_row;
    419 	n_col = A->n_col;
    420 	A = isl_mat_left_hermite(A, 0, NULL, NULL);
    421 	A = isl_mat_drop_cols(A, n_row, n_col - n_row);
    422 	A = isl_mat_lin_to_aff(A);
    423 	A = isl_mat_right_inverse(A);
    424 	d = isl_vec_alloc(ctx, n_row);
    425 	if (A)
    426 		d = isl_vec_set(d, A->row[0][0]);
    427 	A = isl_mat_drop_rows(A, 0, 1);
    428 	A = isl_mat_drop_cols(A, 0, 1);
    429 	B = isl_mat_product(A, B);
    430 
    431 	return isl_mat_parameter_compression(B, d);
    432 }
    433 
    434 /* Return a compression matrix that indicates that there are no solutions
    435  * to the original constraints.  In particular, return a zero-column
    436  * matrix with 1 + dim rows.  If "T2" is not NULL, then assign *T2
    437  * the inverse of this matrix.  *T2 may already have been assigned
    438  * matrix, so free it first.
    439  * "free1", "free2" and "free3" are temporary matrices that are
    440  * not useful when an empty compression is returned.  They are
    441  * simply freed.
    442  */
    443 static __isl_give isl_mat *empty_compression(isl_ctx *ctx, unsigned dim,
    444 	__isl_give isl_mat **T2, __isl_take isl_mat *free1,
    445 	__isl_take isl_mat *free2, __isl_take isl_mat *free3)
    446 {
    447 	isl_mat_free(free1);
    448 	isl_mat_free(free2);
    449 	isl_mat_free(free3);
    450 	if (T2) {
    451 		isl_mat_free(*T2);
    452 		*T2 = isl_mat_alloc(ctx, 0, 1 + dim);
    453 	}
    454 	return isl_mat_alloc(ctx, 1 + dim, 0);
    455 }
    456 
    457 /* Given a matrix that maps a (possibly) parametric domain to
    458  * a parametric domain, add in rows that map the "nparam" parameters onto
    459  * themselves.
    460  */
    461 static __isl_give isl_mat *insert_parameter_rows(__isl_take isl_mat *mat,
    462 	unsigned nparam)
    463 {
    464 	int i;
    465 
    466 	if (nparam == 0)
    467 		return mat;
    468 	if (!mat)
    469 		return NULL;
    470 
    471 	mat = isl_mat_insert_rows(mat, 1, nparam);
    472 	if (!mat)
    473 		return NULL;
    474 
    475 	for (i = 0; i < nparam; ++i) {
    476 		isl_seq_clr(mat->row[1 + i], mat->n_col);
    477 		isl_int_set(mat->row[1 + i][1 + i], mat->row[0][0]);
    478 	}
    479 
    480 	return mat;
    481 }
    482 
    483 /* Given a set of equalities
    484  *
    485  *		-C(y) + M x = 0
    486  *
    487  * this function computes a unimodular transformation from a lower-dimensional
    488  * space to the original space that bijectively maps the integer points x'
    489  * in the lower-dimensional space to the integer points x in the original
    490  * space that satisfy the equalities.
    491  *
    492  * The input is given as a matrix B = [ -C M ] and the output is a
    493  * matrix that maps [1 x'] to [1 x].
    494  * The number of equality constraints in B is assumed to be smaller than
    495  * or equal to the number of variables x.
    496  * "first" is the position of the first x variable.
    497  * The preceding variables are considered to be y-variables.
    498  * If T2 is not NULL, then *T2 is set to a matrix mapping [1 x] to [1 x'].
    499  *
    500  * First compute the (left) Hermite normal form of M,
    501  *
    502  *		M [U1 U2] = M U = H = [H1 0]
    503  * or
    504  *		              M = H Q = [H1 0] [Q1]
    505  *                                             [Q2]
    506  *
    507  * with U, Q unimodular, Q = U^{-1} (and H lower triangular).
    508  * Define the transformed variables as
    509  *
    510  *		x = [U1 U2] [ x1' ] = [U1 U2] [Q1] x
    511  *		            [ x2' ]           [Q2]
    512  *
    513  * The equalities then become
    514  *
    515  *		-C(y) + H1 x1' = 0   or   x1' = H1^{-1} C(y) = C'(y)
    516  *
    517  * If the denominator of the constant term does not divide the
    518  * the common denominator of the coefficients of y, then every
    519  * integer point is mapped to a non-integer point and then the original set
    520  * has no integer solutions (since the x' are a unimodular transformation
    521  * of the x).  In this case, a zero-column matrix is returned.
    522  * Otherwise, the transformation is given by
    523  *
    524  *		x = U1 H1^{-1} C(y) + U2 x2'
    525  *
    526  * The inverse transformation is simply
    527  *
    528  *		x2' = Q2 x
    529  */
    530 __isl_give isl_mat *isl_mat_final_variable_compression(__isl_take isl_mat *B,
    531 	int first, __isl_give isl_mat **T2)
    532 {
    533 	int i, n;
    534 	isl_ctx *ctx;
    535 	isl_mat *H = NULL, *C, *H1, *U = NULL, *U1, *U2;
    536 	unsigned dim;
    537 
    538 	if (T2)
    539 		*T2 = NULL;
    540 	if (!B)
    541 		goto error;
    542 
    543 	ctx = isl_mat_get_ctx(B);
    544 	dim = B->n_col - 1;
    545 	n = dim - first;
    546 	if (n < B->n_row)
    547 		isl_die(ctx, isl_error_invalid, "too many equality constraints",
    548 			goto error);
    549 	H = isl_mat_sub_alloc(B, 0, B->n_row, 1 + first, n);
    550 	H = isl_mat_left_hermite(H, 0, &U, T2);
    551 	if (!H || !U || (T2 && !*T2))
    552 		goto error;
    553 	if (T2) {
    554 		*T2 = isl_mat_drop_rows(*T2, 0, B->n_row);
    555 		*T2 = isl_mat_diagonal(isl_mat_identity(ctx, 1 + first), *T2);
    556 		if (!*T2)
    557 			goto error;
    558 	}
    559 	C = isl_mat_alloc(ctx, 1 + B->n_row, 1 + first);
    560 	if (!C)
    561 		goto error;
    562 	isl_int_set_si(C->row[0][0], 1);
    563 	isl_seq_clr(C->row[0] + 1, first);
    564 	isl_mat_sub_neg(ctx, C->row + 1, B->row, B->n_row, 0, 0, 1 + first);
    565 	H1 = isl_mat_sub_alloc(H, 0, H->n_row, 0, H->n_row);
    566 	H1 = isl_mat_lin_to_aff(H1);
    567 	C = isl_mat_inverse_product(H1, C);
    568 	if (!C)
    569 		goto error;
    570 	isl_mat_free(H);
    571 	if (!isl_int_is_one(C->row[0][0])) {
    572 		isl_int g;
    573 
    574 		isl_int_init(g);
    575 		for (i = 0; i < B->n_row; ++i) {
    576 			isl_seq_gcd(C->row[1 + i] + 1, first, &g);
    577 			isl_int_gcd(g, g, C->row[0][0]);
    578 			if (!isl_int_is_divisible_by(C->row[1 + i][0], g))
    579 				break;
    580 		}
    581 		isl_int_clear(g);
    582 
    583 		if (i < B->n_row)
    584 			return empty_compression(ctx, dim, T2, B, C, U);
    585 		C = isl_mat_normalize(C);
    586 	}
    587 	U1 = isl_mat_sub_alloc(U, 0, U->n_row, 0, B->n_row);
    588 	U1 = isl_mat_lin_to_aff(U1);
    589 	U2 = isl_mat_sub_alloc(U, 0, U->n_row, B->n_row, U->n_row - B->n_row);
    590 	U2 = isl_mat_lin_to_aff(U2);
    591 	isl_mat_free(U);
    592 	C = isl_mat_product(U1, C);
    593 	C = isl_mat_aff_direct_sum(C, U2);
    594 	C = insert_parameter_rows(C, first);
    595 
    596 	isl_mat_free(B);
    597 
    598 	return C;
    599 error:
    600 	isl_mat_free(B);
    601 	isl_mat_free(H);
    602 	isl_mat_free(U);
    603 	if (T2) {
    604 		isl_mat_free(*T2);
    605 		*T2 = NULL;
    606 	}
    607 	return NULL;
    608 }
    609 
    610 /* Given a set of equalities
    611  *
    612  *		M x - c = 0
    613  *
    614  * this function computes a unimodular transformation from a lower-dimensional
    615  * space to the original space that bijectively maps the integer points x'
    616  * in the lower-dimensional space to the integer points x in the original
    617  * space that satisfy the equalities.
    618  *
    619  * The input is given as a matrix B = [ -c M ] and the output is a
    620  * matrix that maps [1 x'] to [1 x].
    621  * The number of equality constraints in B is assumed to be smaller than
    622  * or equal to the number of variables x.
    623  * If T2 is not NULL, then *T2 is set to a matrix mapping [1 x] to [1 x'].
    624  */
    625 __isl_give isl_mat *isl_mat_variable_compression(__isl_take isl_mat *B,
    626 	__isl_give isl_mat **T2)
    627 {
    628 	return isl_mat_final_variable_compression(B, 0, T2);
    629 }
    630 
    631 /* Return "bset" and set *T and *T2 to the identity transformation
    632  * on "bset" (provided T and T2 are not NULL).
    633  */
    634 static __isl_give isl_basic_set *return_with_identity(
    635 	__isl_take isl_basic_set *bset, __isl_give isl_mat **T,
    636 	__isl_give isl_mat **T2)
    637 {
    638 	isl_size dim;
    639 	isl_mat *id;
    640 
    641 	dim = isl_basic_set_dim(bset, isl_dim_set);
    642 	if (dim < 0)
    643 		return isl_basic_set_free(bset);
    644 	if (!T && !T2)
    645 		return bset;
    646 
    647 	id = isl_mat_identity(isl_basic_map_get_ctx(bset), 1 + dim);
    648 	if (T)
    649 		*T = isl_mat_copy(id);
    650 	if (T2)
    651 		*T2 = isl_mat_copy(id);
    652 	isl_mat_free(id);
    653 
    654 	return bset;
    655 }
    656 
    657 /* Use the n equalities of bset to unimodularly transform the
    658  * variables x such that n transformed variables x1' have a constant value
    659  * and rewrite the constraints of bset in terms of the remaining
    660  * transformed variables x2'.  The matrix pointed to by T maps
    661  * the new variables x2' back to the original variables x, while T2
    662  * maps the original variables to the new variables.
    663  */
    664 static __isl_give isl_basic_set *compress_variables(
    665 	__isl_take isl_basic_set *bset,
    666 	__isl_give isl_mat **T, __isl_give isl_mat **T2)
    667 {
    668 	struct isl_mat *B, *TC;
    669 	isl_size dim;
    670 
    671 	if (T)
    672 		*T = NULL;
    673 	if (T2)
    674 		*T2 = NULL;
    675 	if (isl_basic_set_check_no_params(bset) < 0 ||
    676 	    isl_basic_set_check_no_locals(bset) < 0)
    677 		return isl_basic_set_free(bset);
    678 	dim = isl_basic_set_dim(bset, isl_dim_set);
    679 	if (dim < 0)
    680 		return isl_basic_set_free(bset);
    681 	isl_assert(bset->ctx, bset->n_eq <= dim, goto error);
    682 	if (bset->n_eq == 0)
    683 		return return_with_identity(bset, T, T2);
    684 
    685 	B = isl_mat_sub_alloc6(bset->ctx, bset->eq, 0, bset->n_eq, 0, 1 + dim);
    686 	TC = isl_mat_variable_compression(B, T2);
    687 	if (!TC)
    688 		goto error;
    689 	if (TC->n_col == 0) {
    690 		isl_mat_free(TC);
    691 		if (T2) {
    692 			isl_mat_free(*T2);
    693 			*T2 = NULL;
    694 		}
    695 		bset = isl_basic_set_set_to_empty(bset);
    696 		return return_with_identity(bset, T, T2);
    697 	}
    698 
    699 	bset = isl_basic_set_preimage(bset, T ? isl_mat_copy(TC) : TC);
    700 	if (T)
    701 		*T = TC;
    702 	return bset;
    703 error:
    704 	isl_basic_set_free(bset);
    705 	return NULL;
    706 }
    707 
    708 __isl_give isl_basic_set *isl_basic_set_remove_equalities(
    709 	__isl_take isl_basic_set *bset, __isl_give isl_mat **T,
    710 	__isl_give isl_mat **T2)
    711 {
    712 	if (T)
    713 		*T = NULL;
    714 	if (T2)
    715 		*T2 = NULL;
    716 	if (isl_basic_set_check_no_params(bset) < 0)
    717 		return isl_basic_set_free(bset);
    718 	bset = isl_basic_set_gauss(bset, NULL);
    719 	if (ISL_F_ISSET(bset, ISL_BASIC_SET_EMPTY))
    720 		return return_with_identity(bset, T, T2);
    721 	bset = compress_variables(bset, T, T2);
    722 	return bset;
    723 }
    724 
    725 /* Check if dimension dim belongs to a residue class
    726  *		i_dim \equiv r mod m
    727  * with m != 1 and if so return m in *modulo and r in *residue.
    728  * As a special case, when i_dim has a fixed value v, then
    729  * *modulo is set to 0 and *residue to v.
    730  *
    731  * If i_dim does not belong to such a residue class, then *modulo
    732  * is set to 1 and *residue is set to 0.
    733  */
    734 isl_stat isl_basic_set_dim_residue_class(__isl_keep isl_basic_set *bset,
    735 	int pos, isl_int *modulo, isl_int *residue)
    736 {
    737 	isl_bool fixed;
    738 	struct isl_ctx *ctx;
    739 	struct isl_mat *H = NULL, *U = NULL, *C, *H1, *U1;
    740 	isl_size total;
    741 	isl_size nparam;
    742 
    743 	if (!bset || !modulo || !residue)
    744 		return isl_stat_error;
    745 
    746 	fixed = isl_basic_set_plain_dim_is_fixed(bset, pos, residue);
    747 	if (fixed < 0)
    748 		return isl_stat_error;
    749 	if (fixed) {
    750 		isl_int_set_si(*modulo, 0);
    751 		return isl_stat_ok;
    752 	}
    753 
    754 	ctx = isl_basic_set_get_ctx(bset);
    755 	total = isl_basic_set_dim(bset, isl_dim_all);
    756 	nparam = isl_basic_set_dim(bset, isl_dim_param);
    757 	if (total < 0 || nparam < 0)
    758 		return isl_stat_error;
    759 	H = isl_mat_sub_alloc6(ctx, bset->eq, 0, bset->n_eq, 1, total);
    760 	H = isl_mat_left_hermite(H, 0, &U, NULL);
    761 	if (!H)
    762 		return isl_stat_error;
    763 
    764 	isl_seq_gcd(U->row[nparam + pos]+bset->n_eq,
    765 			total-bset->n_eq, modulo);
    766 	if (isl_int_is_zero(*modulo))
    767 		isl_int_set_si(*modulo, 1);
    768 	if (isl_int_is_one(*modulo)) {
    769 		isl_int_set_si(*residue, 0);
    770 		isl_mat_free(H);
    771 		isl_mat_free(U);
    772 		return isl_stat_ok;
    773 	}
    774 
    775 	C = isl_mat_alloc(ctx, 1 + bset->n_eq, 1);
    776 	if (!C)
    777 		goto error;
    778 	isl_int_set_si(C->row[0][0], 1);
    779 	isl_mat_sub_neg(ctx, C->row + 1, bset->eq, bset->n_eq, 0, 0, 1);
    780 	H1 = isl_mat_sub_alloc(H, 0, H->n_row, 0, H->n_row);
    781 	H1 = isl_mat_lin_to_aff(H1);
    782 	C = isl_mat_inverse_product(H1, C);
    783 	isl_mat_free(H);
    784 	U1 = isl_mat_sub_alloc(U, nparam+pos, 1, 0, bset->n_eq);
    785 	U1 = isl_mat_lin_to_aff(U1);
    786 	isl_mat_free(U);
    787 	C = isl_mat_product(U1, C);
    788 	if (!C)
    789 		return isl_stat_error;
    790 	if (!isl_int_is_divisible_by(C->row[1][0], C->row[0][0])) {
    791 		bset = isl_basic_set_copy(bset);
    792 		bset = isl_basic_set_set_to_empty(bset);
    793 		isl_basic_set_free(bset);
    794 		isl_int_set_si(*modulo, 1);
    795 		isl_int_set_si(*residue, 0);
    796 		return isl_stat_ok;
    797 	}
    798 	isl_int_divexact(*residue, C->row[1][0], C->row[0][0]);
    799 	isl_int_fdiv_r(*residue, *residue, *modulo);
    800 	isl_mat_free(C);
    801 	return isl_stat_ok;
    802 error:
    803 	isl_mat_free(H);
    804 	isl_mat_free(U);
    805 	return isl_stat_error;
    806 }
    807 
    808 /* Check if dimension dim belongs to a residue class
    809  *		i_dim \equiv r mod m
    810  * with m != 1 and if so return m in *modulo and r in *residue.
    811  * As a special case, when i_dim has a fixed value v, then
    812  * *modulo is set to 0 and *residue to v.
    813  *
    814  * If i_dim does not belong to such a residue class, then *modulo
    815  * is set to 1 and *residue is set to 0.
    816  */
    817 isl_stat isl_set_dim_residue_class(__isl_keep isl_set *set,
    818 	int pos, isl_int *modulo, isl_int *residue)
    819 {
    820 	isl_int m;
    821 	isl_int r;
    822 	int i;
    823 
    824 	if (!set || !modulo || !residue)
    825 		return isl_stat_error;
    826 
    827 	if (set->n == 0) {
    828 		isl_int_set_si(*modulo, 0);
    829 		isl_int_set_si(*residue, 0);
    830 		return isl_stat_ok;
    831 	}
    832 
    833 	if (isl_basic_set_dim_residue_class(set->p[0], pos, modulo, residue)<0)
    834 		return isl_stat_error;
    835 
    836 	if (set->n == 1)
    837 		return isl_stat_ok;
    838 
    839 	if (isl_int_is_one(*modulo))
    840 		return isl_stat_ok;
    841 
    842 	isl_int_init(m);
    843 	isl_int_init(r);
    844 
    845 	for (i = 1; i < set->n; ++i) {
    846 		if (isl_basic_set_dim_residue_class(set->p[i], pos, &m, &r) < 0)
    847 			goto error;
    848 		isl_int_gcd(*modulo, *modulo, m);
    849 		isl_int_sub(m, *residue, r);
    850 		isl_int_gcd(*modulo, *modulo, m);
    851 		if (!isl_int_is_zero(*modulo))
    852 			isl_int_fdiv_r(*residue, *residue, *modulo);
    853 		if (isl_int_is_one(*modulo))
    854 			break;
    855 	}
    856 
    857 	isl_int_clear(m);
    858 	isl_int_clear(r);
    859 
    860 	return isl_stat_ok;
    861 error:
    862 	isl_int_clear(m);
    863 	isl_int_clear(r);
    864 	return isl_stat_error;
    865 }
    866 
    867 /* Check if dimension "dim" belongs to a residue class
    868  *		i_dim \equiv r mod m
    869  * with m != 1 and if so return m in *modulo and r in *residue.
    870  * As a special case, when i_dim has a fixed value v, then
    871  * *modulo is set to 0 and *residue to v.
    872  *
    873  * If i_dim does not belong to such a residue class, then *modulo
    874  * is set to 1 and *residue is set to 0.
    875  */
    876 isl_stat isl_set_dim_residue_class_val(__isl_keep isl_set *set,
    877 	int pos, __isl_give isl_val **modulo, __isl_give isl_val **residue)
    878 {
    879 	*modulo = NULL;
    880 	*residue = NULL;
    881 	if (!set)
    882 		return isl_stat_error;
    883 	*modulo = isl_val_alloc(isl_set_get_ctx(set));
    884 	*residue = isl_val_alloc(isl_set_get_ctx(set));
    885 	if (!*modulo || !*residue)
    886 		goto error;
    887 	if (isl_set_dim_residue_class(set, pos,
    888 					&(*modulo)->n, &(*residue)->n) < 0)
    889 		goto error;
    890 	isl_int_set_si((*modulo)->d, 1);
    891 	isl_int_set_si((*residue)->d, 1);
    892 	return isl_stat_ok;
    893 error:
    894 	isl_val_free(*modulo);
    895 	isl_val_free(*residue);
    896 	return isl_stat_error;
    897 }
    898