X-Git-Url: http://review.tizen.org/git/?a=blobdiff_plain;f=isl_tab_pip.c;h=b63d3b7b40d1cea4eadfb0c473818c573d4e505e;hb=160facc5bac70a0dd6beab2c0147193906796d4b;hp=9f840f04318b0b5c80ce6d8c94a339c72f348014;hpb=ec21e6de21d784ee1ba32689aebcbda9f786fd30;p=platform%2Fupstream%2Fisl.git diff --git a/isl_tab_pip.c b/isl_tab_pip.c index 9f840f0..b63d3b7 100644 --- a/isl_tab_pip.c +++ b/isl_tab_pip.c @@ -1,7 +1,21 @@ +/* + * Copyright 2008-2009 Katholieke Universiteit Leuven + * Copyright 2010 INRIA Saclay + * + * Use of this software is governed by the GNU LGPLv2.1 license + * + * Written by Sven Verdoolaege, K.U.Leuven, Departement + * Computerwetenschappen, Celestijnenlaan 200A, B-3001 Leuven, Belgium + * and INRIA Saclay - Ile-de-France, Parc Club Orsay Universite, + * ZAC des vignes, 4 rue Jacques Monod, 91893 Orsay, France + */ + +#include #include "isl_map_private.h" -#include "isl_seq.h" +#include #include "isl_tab.h" #include "isl_sample.h" +#include /* * The implementation of parametric integer linear programming in this file @@ -26,10 +40,10 @@ * then the initial sample value may be chosen equal to zero. * However, we will not make this assumption. Instead, we apply * the "big parameter" trick. Any variable x is then not directly - * used in the tableau, but instead it its represented by another + * used in the tableau, but instead it is represented by another * variable x' = M + x, where M is an arbitrarily large (positive) * value. x' is therefore always non-negative, whatever the value of x. - * Taking as initial smaple value x' = 0 corresponds to x = -M, + * Taking as initial sample value x' = 0 corresponds to x = -M, * which is always smaller than any possible value of x. * * The big parameter trick is used in the main tableau and @@ -71,9 +85,8 @@ struct isl_context_op { /* return index of a div that corresponds to "div" */ int (*get_div)(struct isl_context *context, struct isl_tab *tab, struct isl_vec *div); - /* add div "div" to context and return index and non-negativity */ - int (*add_div)(struct isl_context *context, struct isl_vec *div, - int *nonneg); + /* add div "div" to context and return non-negativity */ + int (*add_div)(struct isl_context *context, struct isl_vec *div); int (*detect_equalities)(struct isl_context *context, struct isl_tab *tab); /* return row index of "best" split */ @@ -101,6 +114,20 @@ struct isl_context_lex { struct isl_tab *tab; }; +struct isl_partial_sol { + int level; + struct isl_basic_set *dom; + struct isl_mat *M; + + struct isl_partial_sol *next; +}; + +struct isl_sol; +struct isl_sol_callback { + struct isl_tab_callback callback; + struct isl_sol *sol; +}; + /* isl_sol is an interface for constructing a solution to * a parametric integer linear programming problem. * Every time the algorithm reaches a state where a solution @@ -118,57 +145,225 @@ struct isl_context_lex { * the solution. */ struct isl_sol { + int error; + int rational; + int level; + int max; + int n_out; struct isl_context *context; - struct isl_sol *(*add)(struct isl_sol *sol, struct isl_tab *tab); + struct isl_partial_sol *partial; + void (*add)(struct isl_sol *sol, + struct isl_basic_set *dom, struct isl_mat *M); + void (*add_empty)(struct isl_sol *sol, struct isl_basic_set *bset); void (*free)(struct isl_sol *sol); + struct isl_sol_callback dec_level; }; static void sol_free(struct isl_sol *sol) { + struct isl_partial_sol *partial, *next; if (!sol) return; + for (partial = sol->partial; partial; partial = next) { + next = partial->next; + isl_basic_set_free(partial->dom); + isl_mat_free(partial->M); + free(partial); + } sol->free(sol); } -struct isl_sol_map { - struct isl_sol sol; - struct isl_map *map; - struct isl_set *empty; - int max; -}; - -static void sol_map_free(struct isl_sol_map *sol_map) +/* Push a partial solution represented by a domain and mapping M + * onto the stack of partial solutions. + */ +static void sol_push_sol(struct isl_sol *sol, + struct isl_basic_set *dom, struct isl_mat *M) { - if (sol_map->sol.context) - sol_map->sol.context->op->free(sol_map->sol.context); - isl_map_free(sol_map->map); - isl_set_free(sol_map->empty); - free(sol_map); + struct isl_partial_sol *partial; + + if (sol->error || !dom) + goto error; + + partial = isl_alloc_type(dom->ctx, struct isl_partial_sol); + if (!partial) + goto error; + + partial->level = sol->level; + partial->dom = dom; + partial->M = M; + partial->next = sol->partial; + + sol->partial = partial; + + return; +error: + isl_basic_set_free(dom); + sol->error = 1; } -static void sol_map_free_wrap(struct isl_sol *sol) +/* Pop one partial solution from the partial solution stack and + * pass it on to sol->add or sol->add_empty. + */ +static void sol_pop_one(struct isl_sol *sol) { - sol_map_free((struct isl_sol_map *)sol); + struct isl_partial_sol *partial; + + partial = sol->partial; + sol->partial = partial->next; + + if (partial->M) + sol->add(sol, partial->dom, partial->M); + else + sol->add_empty(sol, partial->dom); + free(partial); } -static struct isl_sol_map *add_empty(struct isl_sol_map *sol) +/* Return a fresh copy of the domain represented by the context tableau. + */ +static struct isl_basic_set *sol_domain(struct isl_sol *sol) { struct isl_basic_set *bset; - if (!sol->empty) - return sol; - sol->empty = isl_set_grow(sol->empty, 1); - bset = sol->sol.context->op->peek_basic_set(sol->sol.context); - bset = isl_basic_set_copy(bset); - bset = isl_basic_set_simplify(bset); - bset = isl_basic_set_finalize(bset); - sol->empty = isl_set_add(sol->empty, bset); - if (!sol->empty) - goto error; - return sol; -error: - sol_map_free(sol); - return NULL; + if (sol->error) + return NULL; + + bset = isl_basic_set_dup(sol->context->op->peek_basic_set(sol->context)); + bset = isl_basic_set_update_from_tab(bset, + sol->context->op->peek_tab(sol->context)); + + return bset; +} + +/* Check whether two partial solutions have the same mapping, where n_div + * is the number of divs that the two partial solutions have in common. + */ +static int same_solution(struct isl_partial_sol *s1, struct isl_partial_sol *s2, + unsigned n_div) +{ + int i; + unsigned dim; + + if (!s1->M != !s2->M) + return 0; + if (!s1->M) + return 1; + + dim = isl_basic_set_total_dim(s1->dom) - s1->dom->n_div; + + for (i = 0; i < s1->M->n_row; ++i) { + if (isl_seq_first_non_zero(s1->M->row[i]+1+dim+n_div, + s1->M->n_col-1-dim-n_div) != -1) + return 0; + if (isl_seq_first_non_zero(s2->M->row[i]+1+dim+n_div, + s2->M->n_col-1-dim-n_div) != -1) + return 0; + if (!isl_seq_eq(s1->M->row[i], s2->M->row[i], 1+dim+n_div)) + return 0; + } + return 1; +} + +/* Pop all solutions from the partial solution stack that were pushed onto + * the stack at levels that are deeper than the current level. + * If the two topmost elements on the stack have the same level + * and represent the same solution, then their domains are combined. + * This combined domain is the same as the current context domain + * as sol_pop is called each time we move back to a higher level. + */ +static void sol_pop(struct isl_sol *sol) +{ + struct isl_partial_sol *partial; + unsigned n_div; + + if (sol->error) + return; + + if (sol->level == 0) { + for (partial = sol->partial; partial; partial = sol->partial) + sol_pop_one(sol); + return; + } + + partial = sol->partial; + if (!partial) + return; + + if (partial->level <= sol->level) + return; + + if (partial->next && partial->next->level == partial->level) { + n_div = isl_basic_set_dim( + sol->context->op->peek_basic_set(sol->context), + isl_dim_div); + + if (!same_solution(partial, partial->next, n_div)) { + sol_pop_one(sol); + sol_pop_one(sol); + } else { + struct isl_basic_set *bset; + + bset = sol_domain(sol); + + isl_basic_set_free(partial->next->dom); + partial->next->dom = bset; + partial->next->level = sol->level; + + sol->partial = partial->next; + isl_basic_set_free(partial->dom); + isl_mat_free(partial->M); + free(partial); + } + } else + sol_pop_one(sol); +} + +static void sol_dec_level(struct isl_sol *sol) +{ + if (sol->error) + return; + + sol->level--; + + sol_pop(sol); +} + +static int sol_dec_level_wrap(struct isl_tab_callback *cb) +{ + struct isl_sol_callback *callback = (struct isl_sol_callback *)cb; + + sol_dec_level(callback->sol); + + return callback->sol->error ? -1 : 0; +} + +/* Move down to next level and push callback onto context tableau + * to decrease the level again when it gets rolled back across + * the current state. That is, dec_level will be called with + * the context tableau in the same state as it is when inc_level + * is called. + */ +static void sol_inc_level(struct isl_sol *sol) +{ + struct isl_tab *tab; + + if (sol->error) + return; + + sol->level++; + tab = sol->context->op->peek_tab(sol->context); + if (isl_tab_push_callback(tab, &sol->dec_level.callback) < 0) + sol->error = 1; +} + +static void scale_rows(struct isl_mat *mat, isl_int m, int n_row) +{ + int i; + + if (isl_int_is_one(m)) + return; + + for (i = 0; i < n_row; ++i) + isl_seq_scale(mat->row[i], mat->row[i], m, mat->n_col); } /* Add the solution identified by the tableau and the context tableau. @@ -184,23 +379,23 @@ error: * dimensions in the input map * tab->n_div is equal to the number of divs in the context * - * If there is no solution, then the basic set corresponding to the - * context tableau is added to the set "empty". + * If there is no solution, then call add_empty with a basic set + * that corresponds to the context tableau. (If add_empty is NULL, + * then do nothing). * - * Otherwise, a basic map is constructed with the same parameters - * and divs as the context, the dimensions of the context as input - * dimensions and a number of output dimensions that is equal to - * the number of output dimensions in the input map. + * If there is a solution, then first construct a matrix that maps + * all dimensions of the context to the output variables, i.e., + * the output dimensions in the input map. * The divs in the input map (if any) that do not correspond to any * div in the context do not appear in the solution. * The algorithm will make sure that they have an integer value, * but these values themselves are of no interest. + * We have to be careful not to drop or rearrange any divs in the + * context because that would change the meaning of the matrix. * - * The constraints and divs of the context are simply copied - * fron context_tab->bset. * To extract the value of the output variables, it should be noted - * that we always use a big parameter M and so the variable stored - * in the tableau is not an output variable x itself, but + * that we always use a big parameter M in the main tableau and so + * the variable stored in this tableau is not an output variable x itself, but * x' = M + x (in case of minimization) * or * x' = M - x (in case of maximization) @@ -211,15 +406,182 @@ error: * are bounded, so this cannot occur. * Similarly, when x' appears in a row, then the coefficient of M in that * row is necessarily 1. - * If the row represents + * If the row in the tableau represents * d x' = c + d M + e(y) - * then, in case of minimization, an equality - * c + e(y) - d x' = 0 - * is added, and in case of maximization, - * c + e(y) + d x' = 0 + * then, in case of minimization, the corresponding row in the matrix + * will be + * a c + a e(y) + * with a d = m, the (updated) common denominator of the matrix. + * In case of maximization, the row will be + * -a c - a e(y) */ -static struct isl_sol_map *sol_map_add(struct isl_sol_map *sol, - struct isl_tab *tab) +static void sol_add(struct isl_sol *sol, struct isl_tab *tab) +{ + struct isl_basic_set *bset = NULL; + struct isl_mat *mat = NULL; + unsigned off; + int row, i; + isl_int m; + + if (sol->error || !tab) + goto error; + + if (tab->empty && !sol->add_empty) + return; + + bset = sol_domain(sol); + + if (tab->empty) { + sol_push_sol(sol, bset, NULL); + return; + } + + off = 2 + tab->M; + + mat = isl_mat_alloc(tab->mat->ctx, 1 + sol->n_out, + 1 + tab->n_param + tab->n_div); + if (!mat) + goto error; + + isl_int_init(m); + + isl_seq_clr(mat->row[0] + 1, mat->n_col - 1); + isl_int_set_si(mat->row[0][0], 1); + for (row = 0; row < sol->n_out; ++row) { + int i = tab->n_param + row; + int r, j; + + isl_seq_clr(mat->row[1 + row], mat->n_col); + if (!tab->var[i].is_row) { + if (tab->M) + isl_die(mat->ctx, isl_error_invalid, + "unbounded optimum", goto error2); + continue; + } + + r = tab->var[i].index; + if (tab->M && + isl_int_ne(tab->mat->row[r][2], tab->mat->row[r][0])) + isl_die(mat->ctx, isl_error_invalid, + "unbounded optimum", goto error2); + isl_int_gcd(m, mat->row[0][0], tab->mat->row[r][0]); + isl_int_divexact(m, tab->mat->row[r][0], m); + scale_rows(mat, m, 1 + row); + isl_int_divexact(m, mat->row[0][0], tab->mat->row[r][0]); + isl_int_mul(mat->row[1 + row][0], m, tab->mat->row[r][1]); + for (j = 0; j < tab->n_param; ++j) { + int col; + if (tab->var[j].is_row) + continue; + col = tab->var[j].index; + isl_int_mul(mat->row[1 + row][1 + j], m, + tab->mat->row[r][off + col]); + } + for (j = 0; j < tab->n_div; ++j) { + int col; + if (tab->var[tab->n_var - tab->n_div+j].is_row) + continue; + col = tab->var[tab->n_var - tab->n_div+j].index; + isl_int_mul(mat->row[1 + row][1 + tab->n_param + j], m, + tab->mat->row[r][off + col]); + } + if (sol->max) + isl_seq_neg(mat->row[1 + row], mat->row[1 + row], + mat->n_col); + } + + isl_int_clear(m); + + sol_push_sol(sol, bset, mat); + return; +error2: + isl_int_clear(m); +error: + isl_basic_set_free(bset); + isl_mat_free(mat); + sol->error = 1; +} + +struct isl_sol_map { + struct isl_sol sol; + struct isl_map *map; + struct isl_set *empty; +}; + +static void sol_map_free(struct isl_sol_map *sol_map) +{ + if (!sol_map) + return; + if (sol_map->sol.context) + sol_map->sol.context->op->free(sol_map->sol.context); + isl_map_free(sol_map->map); + isl_set_free(sol_map->empty); + free(sol_map); +} + +static void sol_map_free_wrap(struct isl_sol *sol) +{ + sol_map_free((struct isl_sol_map *)sol); +} + +/* This function is called for parts of the context where there is + * no solution, with "bset" corresponding to the context tableau. + * Simply add the basic set to the set "empty". + */ +static void sol_map_add_empty(struct isl_sol_map *sol, + struct isl_basic_set *bset) +{ + if (!bset) + goto error; + isl_assert(bset->ctx, sol->empty, goto error); + + sol->empty = isl_set_grow(sol->empty, 1); + bset = isl_basic_set_simplify(bset); + bset = isl_basic_set_finalize(bset); + sol->empty = isl_set_add_basic_set(sol->empty, isl_basic_set_copy(bset)); + if (!sol->empty) + goto error; + isl_basic_set_free(bset); + return; +error: + isl_basic_set_free(bset); + sol->sol.error = 1; +} + +static void sol_map_add_empty_wrap(struct isl_sol *sol, + struct isl_basic_set *bset) +{ + sol_map_add_empty((struct isl_sol_map *)sol, bset); +} + +/* Add bset to sol's empty, but only if we are actually collecting + * the empty set. + */ +static void sol_map_add_empty_if_needed(struct isl_sol_map *sol, + struct isl_basic_set *bset) +{ + if (sol->empty) + sol_map_add_empty(sol, bset); + else + isl_basic_set_free(bset); +} + +/* Given a basic map "dom" that represents the context and an affine + * matrix "M" that maps the dimensions of the context to the + * output variables, construct a basic map with the same parameters + * and divs as the context, the dimensions of the context as input + * dimensions and a number of output dimensions that is equal to + * the number of output dimensions in the input map. + * + * The constraints and divs of the context are simply copied + * from "dom". For each row + * x = c + e(y) + * an equality + * c + e(y) - d x = 0 + * is added, with d the common denominator of M. + */ +static void sol_map_add(struct isl_sol_map *sol, + struct isl_basic_set *dom, struct isl_mat *M) { int i; struct isl_basic_map *bmap = NULL; @@ -230,121 +592,79 @@ static struct isl_sol_map *sol_map_add(struct isl_sol_map *sol, unsigned total; unsigned n_div; unsigned n_out; - unsigned off; - if (!sol || !tab) + if (sol->sol.error || !dom || !M) goto error; - if (tab->empty) - return add_empty(sol); - - context_bset = sol->sol.context->op->peek_basic_set(sol->sol.context); - off = 2 + tab->M; - n_out = isl_map_dim(sol->map, isl_dim_out); - n_eq = context_bset->n_eq + n_out; - n_ineq = context_bset->n_ineq; - nparam = tab->n_param; + n_out = sol->sol.n_out; + n_eq = dom->n_eq + n_out; + n_ineq = dom->n_ineq; + n_div = dom->n_div; + nparam = isl_basic_set_total_dim(dom) - n_div; total = isl_map_dim(sol->map, isl_dim_all); bmap = isl_basic_map_alloc_dim(isl_map_get_dim(sol->map), - tab->n_div, n_eq, 2 * tab->n_div + n_ineq); + n_div, n_eq, 2 * n_div + n_ineq); if (!bmap) goto error; - n_div = tab->n_div; - if (tab->rational) + if (sol->sol.rational) ISL_F_SET(bmap, ISL_BASIC_MAP_RATIONAL); - for (i = 0; i < context_bset->n_div; ++i) { + for (i = 0; i < dom->n_div; ++i) { int k = isl_basic_map_alloc_div(bmap); if (k < 0) goto error; - isl_seq_cpy(bmap->div[k], - context_bset->div[i], 1 + 1 + nparam); + isl_seq_cpy(bmap->div[k], dom->div[i], 1 + 1 + nparam); isl_seq_clr(bmap->div[k] + 1 + 1 + nparam, total - nparam); isl_seq_cpy(bmap->div[k] + 1 + 1 + total, - context_bset->div[i] + 1 + 1 + nparam, i); + dom->div[i] + 1 + 1 + nparam, i); } - for (i = 0; i < context_bset->n_eq; ++i) { + for (i = 0; i < dom->n_eq; ++i) { int k = isl_basic_map_alloc_equality(bmap); if (k < 0) goto error; - isl_seq_cpy(bmap->eq[k], context_bset->eq[i], 1 + nparam); + isl_seq_cpy(bmap->eq[k], dom->eq[i], 1 + nparam); isl_seq_clr(bmap->eq[k] + 1 + nparam, total - nparam); isl_seq_cpy(bmap->eq[k] + 1 + total, - context_bset->eq[i] + 1 + nparam, n_div); + dom->eq[i] + 1 + nparam, n_div); } - for (i = 0; i < context_bset->n_ineq; ++i) { + for (i = 0; i < dom->n_ineq; ++i) { int k = isl_basic_map_alloc_inequality(bmap); if (k < 0) goto error; - isl_seq_cpy(bmap->ineq[k], - context_bset->ineq[i], 1 + nparam); + isl_seq_cpy(bmap->ineq[k], dom->ineq[i], 1 + nparam); isl_seq_clr(bmap->ineq[k] + 1 + nparam, total - nparam); isl_seq_cpy(bmap->ineq[k] + 1 + total, - context_bset->ineq[i] + 1 + nparam, n_div); + dom->ineq[i] + 1 + nparam, n_div); } - for (i = tab->n_param; i < total; ++i) { + for (i = 0; i < M->n_row - 1; ++i) { int k = isl_basic_map_alloc_equality(bmap); if (k < 0) goto error; - isl_seq_clr(bmap->eq[k] + 1, isl_basic_map_total_dim(bmap)); - if (!tab->var[i].is_row) { - /* no unbounded */ - isl_assert(bmap->ctx, !tab->M, goto error); - isl_int_set_si(bmap->eq[k][0], 0); - if (sol->max) - isl_int_set_si(bmap->eq[k][1 + i], 1); - else - isl_int_set_si(bmap->eq[k][1 + i], -1); - } else { - int row, j; - row = tab->var[i].index; - /* no unbounded */ - if (tab->M) - isl_assert(bmap->ctx, - isl_int_eq(tab->mat->row[row][2], - tab->mat->row[row][0]), - goto error); - isl_int_set(bmap->eq[k][0], tab->mat->row[row][1]); - for (j = 0; j < tab->n_param; ++j) { - int col; - if (tab->var[j].is_row) - continue; - col = tab->var[j].index; - isl_int_set(bmap->eq[k][1 + j], - tab->mat->row[row][off + col]); - } - for (j = 0; j < tab->n_div; ++j) { - int col; - if (tab->var[tab->n_var - tab->n_div+j].is_row) - continue; - col = tab->var[tab->n_var - tab->n_div+j].index; - isl_int_set(bmap->eq[k][1 + total + j], - tab->mat->row[row][off + col]); - } - if (sol->max) - isl_int_set(bmap->eq[k][1 + i], - tab->mat->row[row][0]); - else - isl_int_neg(bmap->eq[k][1 + i], - tab->mat->row[row][0]); - } + isl_seq_cpy(bmap->eq[k], M->row[1 + i], 1 + nparam); + isl_seq_clr(bmap->eq[k] + 1 + nparam, n_out); + isl_int_neg(bmap->eq[k][1 + nparam + i], M->row[0][0]); + isl_seq_cpy(bmap->eq[k] + 1 + nparam + n_out, + M->row[1 + i] + 1 + nparam, n_div); } bmap = isl_basic_map_simplify(bmap); bmap = isl_basic_map_finalize(bmap); sol->map = isl_map_grow(sol->map, 1); - sol->map = isl_map_add(sol->map, bmap); + sol->map = isl_map_add_basic_map(sol->map, bmap); if (!sol->map) goto error; - return sol; + isl_basic_set_free(dom); + isl_mat_free(M); + return; error: + isl_basic_set_free(dom); + isl_mat_free(M); isl_basic_map_free(bmap); - sol_free(&sol->sol); - return NULL; + sol->sol.error = 1; } -static struct isl_sol *sol_map_add_wrap(struct isl_sol *sol, - struct isl_tab *tab) +static void sol_map_add_wrap(struct isl_sol *sol, + struct isl_basic_set *dom, struct isl_mat *M) { - return (struct isl_sol *)sol_map_add((struct isl_sol_map *)sol, tab); + sol_map_add((struct isl_sol_map *)sol, dom, M); } @@ -757,7 +1077,7 @@ error: } /* Return the first known violated constraint, i.e., a non-negative - * contraint that currently has an either obviously negative value + * constraint that currently has an either obviously negative value * or a previously determined to be negative value. * * If any constraint has a negative coefficient for the big parameter, @@ -771,8 +1091,11 @@ static int first_neg(struct isl_tab *tab) for (row = tab->n_redundant; row < tab->n_row; ++row) { if (!isl_tab_var_from_row(tab, row)->is_nonneg) continue; - if (isl_int_is_neg(tab->mat->row[row][2])) - return row; + if (!isl_int_is_neg(tab->mat->row[row][2])) + continue; + if (tab->row_sign) + tab->row_sign[row] = isl_tab_row_neg; + return row; } for (row = tab->n_redundant; row < tab->n_row; ++row) { if (!isl_tab_var_from_row(tab, row)->is_nonneg) @@ -792,11 +1115,11 @@ static int first_neg(struct isl_tab *tab) /* Resolve all known or obviously violated constraints through pivoting. * In particular, as long as we can find any violated constraint, we - * look for a pivoting column that would result in the lexicographicallly + * look for a pivoting column that would result in the lexicographically * smallest increment in the sample point. If there is no such column * then the tableau is infeasible. */ -static struct isl_tab *restore_lexmin(struct isl_tab *tab); +static struct isl_tab *restore_lexmin(struct isl_tab *tab) WARN_UNUSED; static struct isl_tab *restore_lexmin(struct isl_tab *tab) { int row, col; @@ -807,8 +1130,11 @@ static struct isl_tab *restore_lexmin(struct isl_tab *tab) return tab; while ((row = first_neg(tab)) != -1) { col = lexmin_pivot_col(tab, row); - if (col >= tab->n_col) - return isl_tab_mark_empty(tab); + if (col >= tab->n_col) { + if (isl_tab_mark_empty(tab) < 0) + goto error; + return tab; + } if (col < 0) goto error; if (isl_tab_pivot(tab, row, col) < 0) @@ -901,8 +1227,6 @@ static struct isl_tab *add_lexmin_valid_eq(struct isl_tab *tab, isl_int *eq) if (isl_tab_kill_col(tab, i) < 0) goto error; tab->n_eq++; - - tab = restore_lexmin(tab); } return tab; @@ -930,6 +1254,7 @@ static int is_constant(struct isl_tab *tab, int row) * In the end we try to use one of the two constraints to eliminate * a column. */ +static struct isl_tab *add_lexmin_eq(struct isl_tab *tab, isl_int *eq) WARN_UNUSED; static struct isl_tab *add_lexmin_eq(struct isl_tab *tab, isl_int *eq) { int r1, r2; @@ -949,10 +1274,13 @@ static struct isl_tab *add_lexmin_eq(struct isl_tab *tab, isl_int *eq) row = tab->con[r1].index; if (is_constant(tab, row)) { if (!isl_int_is_zero(tab->mat->row[row][1]) || - (tab->M && !isl_int_is_zero(tab->mat->row[row][2]))) - return isl_tab_mark_empty(tab); - if (isl_tab_rollback(tab, snap) < 0) - goto error; + (tab->M && !isl_int_is_zero(tab->mat->row[row][2]))) { + if (isl_tab_mark_empty(tab) < 0) + goto error; + return tab; + } + if (isl_tab_rollback(tab, snap) < 0) + goto error; return tab; } @@ -993,16 +1321,16 @@ static struct isl_tab *add_lexmin_eq(struct isl_tab *tab, isl_int *eq) } } - if (tab->bset) { - tab->bset = isl_basic_set_add_ineq(tab->bset, eq); - if (isl_tab_push(tab, isl_tab_undo_bset_ineq) < 0) + if (tab->bmap) { + tab->bmap = isl_basic_map_add_ineq(tab->bmap, eq); + if (isl_tab_push(tab, isl_tab_undo_bmap_ineq) < 0) goto error; isl_seq_neg(eq, eq, 1 + tab->n_var); - tab->bset = isl_basic_set_add_ineq(tab->bset, eq); + tab->bmap = isl_basic_map_add_ineq(tab->bmap, eq); isl_seq_neg(eq, eq, 1 + tab->n_var); - if (isl_tab_push(tab, isl_tab_undo_bset_ineq) < 0) + if (isl_tab_push(tab, isl_tab_undo_bmap_ineq) < 0) goto error; - if (!tab->bset) + if (!tab->bmap) goto error; } @@ -1021,11 +1349,11 @@ static struct isl_tab *add_lexmin_ineq(struct isl_tab *tab, isl_int *ineq) if (!tab) return NULL; - if (tab->bset) { - tab->bset = isl_basic_set_add_ineq(tab->bset, ineq); - if (isl_tab_push(tab, isl_tab_undo_bset_ineq) < 0) + if (tab->bmap) { + tab->bmap = isl_basic_map_add_ineq(tab->bmap, ineq); + if (isl_tab_push(tab, isl_tab_undo_bmap_ineq) < 0) goto error; - if (!tab->bset) + if (!tab->bmap) goto error; } r = isl_tab_add_row(tab, ineq); @@ -1086,7 +1414,7 @@ static int integer_variable(struct isl_tab *tab, int row) int i; unsigned off = 2 + tab->M; - for (i = 0; i < tab->n_col; ++i) { + for (i = tab->n_dead; i < tab->n_col; ++i) { if (tab->col_var[i] >= 0 && (tab->col_var[i] < tab->n_param || tab->col_var[i] >= tab->n_var - tab->n_div)) @@ -1110,8 +1438,9 @@ static int integer_constant(struct isl_tab *tab, int row) #define I_PAR 1 << 1 #define I_VAR 1 << 2 -/* Check for first (non-parameter) variable that is non-integer and - * therefore requires a cut. +/* Check for next (non-parameter) variable after "var" (first if var == -1) + * that is non-integer and therefore requires a cut and return + * the index of the variable. * For parametric tableaus, there are three parts in a row, * the constant, the coefficients of the parameters and the rest. * For each part, we check whether the coefficients in that part @@ -1120,16 +1449,16 @@ static int integer_constant(struct isl_tab *tab, int row) * current sample value is integral and no cut is required * (irrespective of whether the variable part is integral). */ -static int first_non_integer(struct isl_tab *tab, int *f) +static int next_non_integer_var(struct isl_tab *tab, int var, int *f) { - int i; + var = var < 0 ? tab->n_param : var + 1; - for (i = tab->n_param; i < tab->n_var - tab->n_div; ++i) { + for (; var < tab->n_var - tab->n_div; ++var) { int flags = 0; int row; - if (!tab->var[i].is_row) + if (!tab->var[var].is_row) continue; - row = tab->var[i].index; + row = tab->var[var].index; if (integer_constant(tab, row)) ISL_FL_SET(flags, I_CST); if (integer_parameter(tab, row)) @@ -1139,11 +1468,28 @@ static int first_non_integer(struct isl_tab *tab, int *f) if (integer_variable(tab, row)) ISL_FL_SET(flags, I_VAR); *f = flags; - return row; + return var; } return -1; } +/* Check for first (non-parameter) variable that is non-integer and + * therefore requires a cut and return the corresponding row. + * For parametric tableaus, there are three parts in a row, + * the constant, the coefficients of the parameters and the rest. + * For each part, we check whether the coefficients in that part + * are all integral and if so, set the corresponding flag in *f. + * If the constant and the parameter part are integral, then the + * current sample value is integral and no cut is required + * (irrespective of whether the variable part is integral). + */ +static int first_non_integer_row(struct isl_tab *tab, int *f) +{ + int var = next_non_integer_var(tab, -1, f); + + return var < 0 ? -1 : tab->var[var].index; +} + /* Add a (non-parametric) cut to cut away the non-integral sample * value of the given row. * @@ -1203,15 +1549,17 @@ static int add_cut(struct isl_tab *tab, int row) * sample point is obtained or until the tableau is determined * to be integer infeasible. * As long as there is any non-integer value in the sample point, - * we add an appropriate cut, if possible and resolve the violated - * cut constraint using restore_lexmin. + * we add appropriate cuts, if possible, for each of these + * non-integer values and then resolve the violated + * cut constraints using restore_lexmin. * If one of the corresponding rows is equal to an integral * combination of variables/constraints plus a non-integral constant, - * then there is no way to obtain an integer point an we return + * then there is no way to obtain an integer point and we return * a tableau that is marked empty. */ static struct isl_tab *cut_to_integer_lexmin(struct isl_tab *tab) { + int var; int row; int flags; @@ -1220,12 +1568,18 @@ static struct isl_tab *cut_to_integer_lexmin(struct isl_tab *tab) if (tab->empty) return tab; - while ((row = first_non_integer(tab, &flags)) != -1) { - if (ISL_FL_ISSET(flags, I_VAR)) - return isl_tab_mark_empty(tab); - row = add_cut(tab, row); - if (row < 0) - goto error; + while ((var = next_non_integer_var(tab, -1, &flags)) != -1) { + do { + if (ISL_FL_ISSET(flags, I_VAR)) { + if (isl_tab_mark_empty(tab) < 0) + goto error; + return tab; + } + row = tab->var[var].index; + row = add_cut(tab, row); + if (row < 0) + goto error; + } while ((var = next_non_integer_var(tab, var, &flags)) != -1); tab = restore_lexmin(tab); if (!tab || tab->empty) break; @@ -1248,7 +1602,7 @@ static struct isl_tab *check_samples(struct isl_tab *tab, isl_int *ineq, int eq) if (!tab) return NULL; - isl_assert(tab->mat->ctx, tab->bset, goto error); + 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); @@ -1347,7 +1701,7 @@ static int tab_has_valid_sample(struct isl_tab *tab, isl_int *ineq, int eq) if (!tab) return -1; - isl_assert(tab->mat->ctx, tab->bset, return -1); + isl_assert(tab->mat->ctx, tab->bmap, return -1); isl_assert(tab->mat->ctx, tab->samples, return -1); isl_assert(tab->mat->ctx, tab->samples->n_col == 1 + tab->n_var, return -1); @@ -1365,78 +1719,21 @@ static int tab_has_valid_sample(struct isl_tab *tab, isl_int *ineq, int eq) return i < tab->n_sample; } -/* For a div d = floor(f/m), add the constraints - * - * f - m d >= 0 - * -(f-(m-1)) + m d >= 0 - * - * Note that the second constraint is the negation of - * - * f - m d >= m - */ -static void add_div_constraints(struct isl_context *context, unsigned div) -{ - unsigned total; - unsigned div_pos; - struct isl_vec *ineq; - struct isl_basic_set *bset; - - bset = context->op->peek_basic_set(context); - if (!bset) - goto error; - - total = isl_basic_set_total_dim(bset); - div_pos = 1 + total - bset->n_div + div; - - ineq = ineq_for_div(bset, div); - if (!ineq) - goto error; - - context->op->add_ineq(context, ineq->el, 0, 0); - - isl_seq_neg(ineq->el, bset->div[div] + 1, 1 + total); - isl_int_set(ineq->el[div_pos], bset->div[div][0]); - isl_int_add(ineq->el[0], ineq->el[0], ineq->el[div_pos]); - isl_int_sub_ui(ineq->el[0], ineq->el[0], 1); - - context->op->add_ineq(context, ineq->el, 0, 0); - - isl_vec_free(ineq); - - return; -error: - context->op->invalidate(context); -} - -/* Add a div specifed by "div" to the tableau "tab" and return - * the index of the new div. *nonneg is set to 1 if the div - * is obviously non-negative. +/* Add a div specified by "div" to the tableau "tab" and return + * 1 if the div is obviously non-negative. */ static int context_tab_add_div(struct isl_tab *tab, struct isl_vec *div, - int *nonneg) + int (*add_ineq)(void *user, isl_int *), void *user) { int i; int r; - int k; struct isl_mat *samples; + int nonneg; - for (i = 0; i < tab->n_var; ++i) { - if (isl_int_is_zero(div->el[2 + i])) - continue; - if (!tab->var[i].is_nonneg) - break; - } - *nonneg = i == tab->n_var; - - if (isl_tab_extend_cons(tab, 3) < 0) - return -1; - if (isl_tab_extend_vars(tab, 1) < 0) - return -1; - r = isl_tab_allocate_var(tab); + r = isl_tab_add_div(tab, div, add_ineq, user); if (r < 0) return -1; - if (*nonneg) - tab->var[r].is_nonneg = 1; + nonneg = tab->var[r].is_nonneg; tab->var[r].frozen = 1; samples = isl_mat_extend(tab->samples, @@ -1451,16 +1748,7 @@ static int context_tab_add_div(struct isl_tab *tab, struct isl_vec *div, samples->row[i][samples->n_col - 1], div->el[0]); } - tab->bset = isl_basic_set_extend_dim(tab->bset, - isl_basic_set_get_dim(tab->bset), 1, 0, 2); - k = isl_basic_set_alloc_div(tab->bset); - if (k < 0) - return -1; - isl_seq_cpy(tab->bset->div[k], div->el, div->size); - if (isl_tab_push(tab, isl_tab_undo_bset_div) < 0) - return -1; - - return k; + return nonneg; } /* Add a div specified by "div" to both the main tableau and @@ -1473,14 +1761,11 @@ static int add_div(struct isl_tab *tab, struct isl_context *context, struct isl_vec *div) { int r; - int k; int nonneg; - k = context->op->add_div(context, div, &nonneg); - if (k < 0) + if ((nonneg = context->op->add_div(context, div)) < 0) goto error; - add_div_constraints(context, k); if (!context->op->is_ok(context)) goto error; @@ -1503,12 +1788,12 @@ error: static int find_div(struct isl_tab *tab, isl_int *div, isl_int denom) { int i; - unsigned total = isl_basic_set_total_dim(tab->bset); + unsigned total = isl_basic_map_total_dim(tab->bmap); - for (i = 0; i < tab->bset->n_div; ++i) { - if (isl_int_ne(tab->bset->div[i][0], denom)) + for (i = 0; i < tab->bmap->n_div; ++i) { + if (isl_int_ne(tab->bmap->div[i][0], denom)) continue; - if (!isl_seq_eq(tab->bset->div[i] + 1, div, total)) + if (!isl_seq_eq(tab->bmap->div[i] + 1, div, 1 + total)) continue; return i; } @@ -1685,8 +1970,11 @@ static struct isl_tab *tab_for_lexmin(struct isl_basic_map *bmap, if (!tab->row_sign) goto error; } - if (ISL_F_ISSET(bmap, ISL_BASIC_MAP_EMPTY)) - return isl_tab_mark_empty(tab); + if (ISL_F_ISSET(bmap, ISL_BASIC_MAP_EMPTY)) { + if (isl_tab_mark_empty(tab) < 0) + goto error; + return tab; + } for (i = tab->n_param; i < tab->n_var - tab->n_div; ++i) { tab->var[i].is_nonneg = 1; @@ -1705,6 +1993,8 @@ static struct isl_tab *tab_for_lexmin(struct isl_basic_map *bmap, if (!tab || tab->empty) return tab; } + if (bmap->n_eq) + tab = restore_lexmin(tab); for (i = 0; i < bmap->n_ineq; ++i) { if (max) isl_seq_neg(bmap->ineq[i] + 1 + tab->n_param, @@ -1757,6 +2047,7 @@ static int best_split(struct isl_tab *tab, struct isl_tab *context_tab) struct isl_tab_undo *snap2; struct isl_vec *ineq = NULL; int r = 0; + int ok; if (!isl_tab_var_from_row(tab, split)->is_nonneg) continue; @@ -1766,8 +2057,10 @@ static int best_split(struct isl_tab *tab, struct isl_tab *context_tab) ineq = get_row_parameter_ineq(tab, split); if (!ineq) return -1; - context_tab = isl_tab_add_ineq(context_tab, ineq->el); + ok = isl_tab_add_ineq(context_tab, ineq->el) >= 0; isl_vec_free(ineq); + if (!ok) + return -1; snap2 = isl_tab_snap(context_tab); @@ -1784,8 +2077,10 @@ static int best_split(struct isl_tab *tab, struct isl_tab *context_tab) ineq = get_row_parameter_ineq(tab, row); if (!ineq) return -1; - context_tab = isl_tab_add_ineq(context_tab, ineq->el); + ok = isl_tab_add_ineq(context_tab, ineq->el) >= 0; isl_vec_free(ineq); + if (!ok) + return -1; var = &context_tab->con[context_tab->n_con - 1]; if (!context_tab->empty && !isl_tab_min_at_most_neg_one(context_tab, var)) @@ -1810,7 +2105,7 @@ static struct isl_basic_set *context_lex_peek_basic_set( struct isl_context_lex *clex = (struct isl_context_lex *)context; if (!clex->tab) return NULL; - return clex->tab->bset; + return isl_tab_peek_bset(clex->tab); } static struct isl_tab *context_lex_peek_tab(struct isl_context *context) @@ -1874,6 +2169,13 @@ error: clex->tab = NULL; } +static int context_lex_add_ineq_wrap(void *user, isl_int *ineq) +{ + struct isl_context *context = (struct isl_context *)user; + context_lex_add_ineq(context, ineq, 0, 0); + return context->op->is_ok(context) ? 0 : -1; +} + /* Check which signs can be obtained by "ineq" on all the currently * active sample values. See row_sign for more information. */ @@ -1883,10 +2185,11 @@ static enum isl_tab_row_sign tab_ineq_sign(struct isl_tab *tab, isl_int *ineq, int i; int sgn; isl_int tmp; - int res = isl_tab_row_unknown; + enum isl_tab_row_sign res = isl_tab_row_unknown; - isl_assert(tab->mat->ctx, tab->samples, return 0); - isl_assert(tab->mat->ctx, tab->samples->n_col == 1 + tab->n_var, return 0); + isl_assert(tab->mat->ctx, tab->samples, return isl_tab_row_unknown); + isl_assert(tab->mat->ctx, tab->samples->n_col == 1 + tab->n_var, + return isl_tab_row_unknown); isl_int_init(tmp); for (i = tab->n_outside; i < tab->n_sample; ++i) { @@ -1955,11 +2258,25 @@ static int context_lex_get_div(struct isl_context *context, struct isl_tab *tab, return get_div(tab, context, div); } -static int context_lex_add_div(struct isl_context *context, struct isl_vec *div, - int *nonneg) +/* Add a div specified by "div" to the context tableau and return + * 1 if the div is obviously non-negative. + * context_tab_add_div will always return 1, because all variables + * in a isl_context_lex tableau are non-negative. + * However, if we are using a big parameter in the context, then this only + * reflects the non-negativity of the variable used to _encode_ the + * div, i.e., div' = M + div, so we can't draw any conclusions. + */ +static int context_lex_add_div(struct isl_context *context, struct isl_vec *div) { struct isl_context_lex *clex = (struct isl_context_lex *)context; - return context_tab_add_div(clex->tab, div, nonneg); + int nonneg; + nonneg = context_tab_add_div(clex->tab, div, + context_lex_add_ineq_wrap, context); + if (nonneg < 0) + return -1; + if (clex->tab->M) + return 0; + return nonneg; } static int context_lex_detect_equalities(struct isl_context *context, @@ -1980,7 +2297,7 @@ static int context_lex_best_split(struct isl_context *context, return -1; r = best_split(tab, clex->tab); - if (isl_tab_rollback(clex->tab, snap) < 0) + if (r >= 0 && isl_tab_rollback(clex->tab, snap) < 0) return -1; return r; @@ -2053,7 +2370,8 @@ static struct isl_tab *tab_detect_nonnegative_parameters(struct isl_tab *tab, isl_seq_clr(ineq->el, ineq->size); for (i = 0; i < context_tab->n_var; ++i) { isl_int_set_si(ineq->el[1 + i], 1); - context_tab = isl_tab_add_ineq(context_tab, ineq->el); + if (isl_tab_add_ineq(context_tab, ineq->el) < 0) + goto error; var = &context_tab->con[context_tab->n_con - 1]; if (!context_tab->empty && !isl_tab_min_at_most_neg_one(context_tab, var)) { @@ -2087,6 +2405,9 @@ static struct isl_tab *context_lex_detect_nonnegative_parameters( struct isl_context_lex *clex = (struct isl_context_lex *)context; struct isl_tab_undo *snap; + if (!tab) + return NULL; + snap = isl_tab_snap(clex->tab); if (isl_tab_push_basis(clex->tab) < 0) goto error; @@ -2146,7 +2467,8 @@ static struct isl_tab *context_tab_for_lexmin(struct isl_basic_set *bset) tab = tab_for_lexmin((struct isl_basic_map *)bset, NULL, 1, 0); if (!tab) goto error; - tab->bset = bset; + if (isl_tab_track_bset(tab, bset) < 0) + goto error; tab = isl_tab_init_samples(tab); return tab; error: @@ -2190,6 +2512,8 @@ static struct isl_tab *context_gbr_detect_nonnegative_parameters( struct isl_context *context, struct isl_tab *tab) { struct isl_context_gbr *cgbr = (struct isl_context_gbr *)context; + if (!tab) + return NULL; return tab_detect_nonnegative_parameters(tab, cgbr->tab); } @@ -2199,7 +2523,7 @@ static struct isl_basic_set *context_gbr_peek_basic_set( struct isl_context_gbr *cgbr = (struct isl_context_gbr *)context; if (!cgbr->tab) return NULL; - return cgbr->tab->bset; + return isl_tab_peek_bset(cgbr->tab); } static struct isl_tab *context_gbr_peek_tab(struct isl_context *context) @@ -2218,7 +2542,7 @@ static void gbr_init_shifted(struct isl_context_gbr *cgbr) { int i, j; struct isl_vec *cst; - struct isl_basic_set *bset = cgbr->tab->bset; + struct isl_basic_set *bset = isl_tab_peek_bset(cgbr->tab); unsigned dim = isl_basic_set_total_dim(bset); cst = isl_vec_alloc(cgbr->tab->mat->ctx, bset->n_ineq); @@ -2282,7 +2606,7 @@ static struct isl_basic_set *drop_constant_terms(struct isl_basic_set *bset) static int use_shifted(struct isl_context_gbr *cgbr) { - return cgbr->tab->bset->n_eq == 0 && cgbr->tab->bset->n_div == 0; + return cgbr->tab->bmap->n_eq == 0 && cgbr->tab->bmap->n_div == 0; } static struct isl_vec *gbr_get_sample(struct isl_context_gbr *cgbr) @@ -2304,13 +2628,14 @@ static struct isl_vec *gbr_get_sample(struct isl_context_gbr *cgbr) } if (!cgbr->cone) { - cgbr->cone = isl_tab_from_recession_cone(cgbr->tab->bset); + bset = isl_tab_peek_bset(cgbr->tab); + cgbr->cone = isl_tab_from_recession_cone(bset, 0); if (!cgbr->cone) return NULL; - cgbr->cone->bset = isl_basic_set_dup(cgbr->tab->bset); + if (isl_tab_track_bset(cgbr->cone, isl_basic_set_dup(bset)) < 0) + return NULL; } - cgbr->cone = isl_tab_detect_implicit_equalities(cgbr->cone); - if (!cgbr->cone) + if (isl_tab_detect_implicit_equalities(cgbr->cone) < 0) return NULL; if (cgbr->cone->n_dead == cgbr->cone->n_col) { @@ -2321,10 +2646,9 @@ static struct isl_vec *gbr_get_sample(struct isl_context_gbr *cgbr) if (cgbr->tab->basis->n_col != 1 + cgbr->tab->n_var) { isl_mat_free(cgbr->tab->basis); cgbr->tab->basis = NULL; - } else { - cgbr->tab->n_zero = 0; - cgbr->tab->n_unbounded = 0; } + cgbr->tab->n_zero = 0; + cgbr->tab->n_unbounded = 0; } snap = isl_tab_snap(cgbr->tab); @@ -2339,13 +2663,13 @@ static struct isl_vec *gbr_get_sample(struct isl_context_gbr *cgbr) return sample; } - cone = isl_basic_set_dup(cgbr->cone->bset); + cone = isl_basic_set_dup(isl_tab_peek_bset(cgbr->cone)); cone = drop_constant_terms(cone); cone = isl_basic_set_update_from_tab(cone, cgbr->cone); cone = isl_basic_set_underlying_set(cone); cone = isl_basic_set_gauss(cone, NULL); - bset = isl_basic_set_dup(cgbr->tab->bset); + bset = isl_basic_set_dup(isl_tab_peek_bset(cgbr->tab)); bset = isl_basic_set_update_from_tab(bset, cgbr->tab); bset = isl_basic_set_underlying_set(bset); bset = isl_basic_set_gauss(bset, NULL); @@ -2369,7 +2693,8 @@ static void check_gbr_integer_feasible(struct isl_context_gbr *cgbr) if (sample->size == 0) { isl_vec_free(sample); - cgbr->tab = isl_tab_mark_empty(cgbr->tab); + if (isl_tab_mark_empty(cgbr->tab) < 0) + goto error; return; } @@ -2391,7 +2716,8 @@ static struct isl_tab *add_gbr_eq(struct isl_tab *tab, isl_int *eq) if (isl_tab_extend_cons(tab, 2) < 0) goto error; - tab = isl_tab_add_eq(tab, eq); + if (isl_tab_add_eq(tab, eq) < 0) + goto error; return tab; error: @@ -2409,7 +2735,8 @@ static void context_gbr_add_eq(struct isl_context *context, isl_int *eq, if (cgbr->cone && cgbr->cone->n_col != cgbr->cone->n_dead) { if (isl_tab_extend_cons(cgbr->cone, 2) < 0) goto error; - cgbr->cone = isl_tab_add_eq(cgbr->cone, eq); + if (isl_tab_add_eq(cgbr->cone, eq) < 0) + goto error; } if (check) { @@ -2435,12 +2762,13 @@ static void add_gbr_ineq(struct isl_context_gbr *cgbr, isl_int *ineq) if (isl_tab_extend_cons(cgbr->tab, 1) < 0) goto error; - cgbr->tab = isl_tab_add_ineq(cgbr->tab, ineq); + if (isl_tab_add_ineq(cgbr->tab, ineq) < 0) + goto error; if (cgbr->shifted && !cgbr->shifted->empty && use_shifted(cgbr)) { int i; unsigned dim; - dim = isl_basic_set_total_dim(cgbr->tab->bset); + dim = isl_basic_map_total_dim(cgbr->tab->bmap); if (isl_tab_extend_cons(cgbr->shifted, 1) < 0) goto error; @@ -2451,7 +2779,8 @@ static void add_gbr_ineq(struct isl_context_gbr *cgbr, isl_int *ineq) isl_int_add(ineq[0], ineq[0], ineq[1 + i]); } - cgbr->shifted = isl_tab_add_ineq(cgbr->shifted, ineq); + if (isl_tab_add_ineq(cgbr->shifted, ineq) < 0) + goto error; for (i = 0; i < dim; ++i) { if (!isl_int_is_neg(ineq[1 + i])) @@ -2463,7 +2792,8 @@ static void add_gbr_ineq(struct isl_context_gbr *cgbr, isl_int *ineq) if (cgbr->cone && cgbr->cone->n_col != cgbr->cone->n_dead) { if (isl_tab_extend_cons(cgbr->cone, 1) < 0) goto error; - cgbr->cone = isl_tab_add_ineq(cgbr->cone, ineq); + if (isl_tab_add_ineq(cgbr->cone, ineq) < 0) + goto error; } return; @@ -2496,6 +2826,13 @@ error: cgbr->tab = NULL; } +static int context_gbr_add_ineq_wrap(void *user, isl_int *ineq) +{ + struct isl_context *context = (struct isl_context *)user; + context_gbr_add_ineq(context, ineq, 0, 0); + return context->op->is_ok(context) ? 0 : -1; +} + static enum isl_tab_row_sign context_gbr_ineq_sign(struct isl_context *context, isl_int *ineq, int strict) { @@ -2599,20 +2936,20 @@ static void propagate_equalities(struct isl_context_gbr *cgbr, if (!eq) goto error; - if (isl_tab_extend_cons(tab, (cgbr->tab->bset->n_ineq - first)/2) < 0) + if (isl_tab_extend_cons(tab, (cgbr->tab->bmap->n_ineq - first)/2) < 0) goto error; isl_seq_clr(eq->el + 1 + tab->n_param, tab->n_var - tab->n_param - tab->n_div); - for (i = first; i < cgbr->tab->bset->n_ineq; i += 2) { + for (i = first; i < cgbr->tab->bmap->n_ineq; i += 2) { int j; int r; struct isl_tab_undo *snap; snap = isl_tab_snap(tab); - isl_seq_cpy(eq->el, cgbr->tab->bset->ineq[i], 1 + tab->n_param); + isl_seq_cpy(eq->el, cgbr->tab->bmap->ineq[i], 1 + tab->n_param); isl_seq_cpy(eq->el + 1 + tab->n_var - tab->n_div, - cgbr->tab->bset->ineq[i] + 1 + tab->n_param, + cgbr->tab->bmap->ineq[i] + 1 + tab->n_param, tab->n_div); r = isl_tab_add_row(tab, eq->el); @@ -2657,16 +2994,19 @@ static int context_gbr_detect_equalities(struct isl_context *context, ctx = cgbr->tab->mat->ctx; if (!cgbr->cone) { - cgbr->cone = isl_tab_from_recession_cone(cgbr->tab->bset); + struct isl_basic_set *bset = isl_tab_peek_bset(cgbr->tab); + cgbr->cone = isl_tab_from_recession_cone(bset, 0); if (!cgbr->cone) goto error; - cgbr->cone->bset = isl_basic_set_dup(cgbr->tab->bset); + if (isl_tab_track_bset(cgbr->cone, isl_basic_set_dup(bset)) < 0) + goto error; } - cgbr->cone = isl_tab_detect_implicit_equalities(cgbr->cone); + if (isl_tab_detect_implicit_equalities(cgbr->cone) < 0) + goto error; - n_ineq = cgbr->tab->bset->n_ineq; + n_ineq = cgbr->tab->bmap->n_ineq; cgbr->tab = isl_tab_detect_equalities(cgbr->tab, cgbr->cone); - if (cgbr->tab && cgbr->tab->bset->n_ineq > n_ineq) + if (cgbr->tab && cgbr->tab->bmap->n_ineq > n_ineq) propagate_equalities(cgbr, tab, n_ineq); return 0; @@ -2682,8 +3022,7 @@ static int context_gbr_get_div(struct isl_context *context, struct isl_tab *tab, return get_div(tab, context, div); } -static int context_gbr_add_div(struct isl_context *context, struct isl_vec *div, - int *nonneg) +static int context_gbr_add_div(struct isl_context *context, struct isl_vec *div) { struct isl_context_gbr *cgbr = (struct isl_context_gbr *)context; if (cgbr->cone) { @@ -2696,16 +3035,17 @@ static int context_gbr_add_div(struct isl_context *context, struct isl_vec *div, if (isl_tab_allocate_var(cgbr->cone) <0) return -1; - cgbr->cone->bset = isl_basic_set_extend_dim(cgbr->cone->bset, - isl_basic_set_get_dim(cgbr->cone->bset), 1, 0, 2); - k = isl_basic_set_alloc_div(cgbr->cone->bset); + cgbr->cone->bmap = isl_basic_map_extend_dim(cgbr->cone->bmap, + isl_basic_map_get_dim(cgbr->cone->bmap), 1, 0, 2); + k = isl_basic_map_alloc_div(cgbr->cone->bmap); if (k < 0) return -1; - isl_seq_cpy(cgbr->cone->bset->div[k], div->el, div->size); - if (isl_tab_push(cgbr->cone, isl_tab_undo_bset_div) < 0) + isl_seq_cpy(cgbr->cone->bmap->div[k], div->el, div->size); + if (isl_tab_push(cgbr->cone, isl_tab_undo_bmap_div) < 0) return -1; } - return context_tab_add_div(cgbr->tab, div, nonneg); + return context_tab_add_div(cgbr->tab, div, + context_gbr_add_ineq_wrap, context); } static int context_gbr_best_split(struct isl_context *context, @@ -2718,7 +3058,7 @@ static int context_gbr_best_split(struct isl_context *context, snap = isl_tab_snap(cgbr->tab); r = best_split(tab, cgbr->tab); - if (isl_tab_rollback(cgbr->tab, snap) < 0) + if (r >= 0 && isl_tab_rollback(cgbr->tab, snap) < 0) return -1; return r; @@ -2864,8 +3204,8 @@ static struct isl_context *isl_context_gbr_alloc(struct isl_basic_set *dom) cgbr->tab = isl_tab_init_samples(cgbr->tab); if (!cgbr->tab) goto error; - cgbr->tab->bset = isl_basic_set_cow(isl_basic_set_copy(dom)); - if (!cgbr->tab->bset) + if (isl_tab_track_bset(cgbr->tab, + isl_basic_set_cow(isl_basic_set_copy(dom))) < 0) goto error; check_gbr_integer_feasible(cgbr); @@ -2880,7 +3220,7 @@ static struct isl_context *isl_context_alloc(struct isl_basic_set *dom) if (!dom) return NULL; - if (dom->ctx->context == ISL_CONTEXT_LEXMIN) + if (dom->ctx->opt->context == ISL_CONTEXT_LEXMIN) return isl_context_lex_alloc(dom); else return isl_context_gbr_alloc(dom); @@ -2896,14 +3236,22 @@ static struct isl_context *isl_context_alloc(struct isl_basic_set *dom) static struct isl_sol_map *sol_map_init(struct isl_basic_map *bmap, struct isl_basic_set *dom, int track_empty, int max) { - struct isl_sol_map *sol_map; + struct isl_sol_map *sol_map = NULL; + + if (!bmap) + goto error; - sol_map = isl_calloc_type(bset->ctx, struct isl_sol_map); + sol_map = isl_calloc_type(bmap->ctx, struct isl_sol_map); if (!sol_map) goto error; - sol_map->max = max; + sol_map->sol.rational = ISL_F_ISSET(bmap, ISL_BASIC_MAP_RATIONAL); + sol_map->sol.dec_level.callback.run = &sol_dec_level_wrap; + sol_map->sol.dec_level.sol = &sol_map->sol; + sol_map->sol.max = max; + sol_map->sol.n_out = isl_basic_map_dim(bmap, isl_dim_out); sol_map->sol.add = &sol_map_add_wrap; + sol_map->sol.add_empty = track_empty ? &sol_map_add_empty_wrap : NULL; sol_map->sol.free = &sol_map_free_wrap; sol_map->map = isl_map_alloc_dim(isl_basic_map_get_dim(bmap), 1, ISL_MAP_DISJOINT); @@ -3033,7 +3381,7 @@ static enum isl_tab_row_sign row_sign(struct isl_tab *tab, struct isl_sol *sol, int row) { struct isl_vec *ineq = NULL; - int res = isl_tab_row_unknown; + enum isl_tab_row_sign res = isl_tab_row_unknown; int critical; int strict; int row2; @@ -3097,10 +3445,10 @@ static enum isl_tab_row_sign row_sign(struct isl_tab *tab, return res; error: isl_vec_free(ineq); - return 0; + return isl_tab_row_unknown; } -static struct isl_sol *find_solutions(struct isl_sol *sol, struct isl_tab *tab); +static void find_solutions(struct isl_sol *sol, struct isl_tab *tab); /* Find solutions for values of the parameters that satisfy the given * inequality. @@ -3116,8 +3464,7 @@ static struct isl_sol *find_solutions(struct isl_sol *sol, struct isl_tab *tab); * and that we need to do this before saving the current basis * such that the basis has been restore before we restore the row signs. */ -static struct isl_sol *find_in_pos(struct isl_sol *sol, - struct isl_tab *tab, isl_int *ineq) +static void find_in_pos(struct isl_sol *sol, struct isl_tab *tab, isl_int *ineq) { void *saved; @@ -3131,25 +3478,25 @@ static struct isl_sol *find_in_pos(struct isl_sol *sol, sol->context->op->add_ineq(sol->context, ineq, 0, 1); - sol = find_solutions(sol, tab); + find_solutions(sol, tab); - sol->context->op->restore(sol->context, saved); - return sol; + if (!sol->error) + sol->context->op->restore(sol->context, saved); + return; error: - sol_free(sol); - return NULL; + sol->error = 1; } /* Record the absence of solutions for those values of the parameters * that do not satisfy the given inequality with equality. */ -static struct isl_sol *no_sol_in_strict(struct isl_sol *sol, +static void no_sol_in_strict(struct isl_sol *sol, struct isl_tab *tab, struct isl_vec *ineq) { int empty; void *saved; - if (!sol->context) + if (!sol->context || sol->error) goto error; saved = sol->context->op->save(sol->context); @@ -3161,16 +3508,15 @@ static struct isl_sol *no_sol_in_strict(struct isl_sol *sol, empty = tab->empty; tab->empty = 1; - sol = sol->add(sol, tab); + sol_add(sol, tab); tab->empty = empty; isl_int_add_ui(ineq->el[0], ineq->el[0], 1); sol->context->op->restore(sol->context, saved); - return sol; + return; error: - sol_free(sol); - return NULL; + sol->error = 1; } /* Compute the lexicographic minimum of the set represented by the main @@ -3247,7 +3593,7 @@ error: * coefficient are integral, then there is nothing that can be done * and the tableau has no integral solution. * If, on the other hand, one or more of the other columns have rational - * coeffcients, but the parameter coefficients are all integral, then + * coefficients, but the parameter coefficients are all integral, then * we can perform a regular (non-parametric) cut. * Finally, if there is any parameter coefficient that is non-integral, * then we need to involve the context tableau. There are two cases here. @@ -3267,11 +3613,11 @@ error: * In the part of the context where this inequality does not hold, the * main tableau is marked as being empty. */ -static struct isl_sol *find_solutions(struct isl_sol *sol, struct isl_tab *tab) +static void find_solutions(struct isl_sol *sol, struct isl_tab *tab) { struct isl_context *context; - if (!tab || !sol) + if (!tab || sol->error) goto error; context = sol->context; @@ -3284,7 +3630,7 @@ static struct isl_sol *find_solutions(struct isl_sol *sol, struct isl_tab *tab) for (; tab && !tab->empty; tab = restore_lexmin(tab)) { int flags; int row; - int sgn; + enum isl_tab_row_sign sgn; int split = -1; int n_split = 0; @@ -3321,25 +3667,28 @@ static struct isl_sol *find_solutions(struct isl_sol *sol, struct isl_tab *tab) tab->row_sign[row] = isl_tab_row_unknown; } tab->row_sign[split] = isl_tab_row_pos; - sol = find_in_pos(sol, tab, ineq->el); + sol_inc_level(sol); + find_in_pos(sol, tab, ineq->el); tab->row_sign[split] = isl_tab_row_neg; row = split; isl_seq_neg(ineq->el, ineq->el, ineq->size); isl_int_sub_ui(ineq->el[0], ineq->el[0], 1); - context->op->add_ineq(context, ineq->el, 0, 1); + if (!sol->error) + context->op->add_ineq(context, ineq->el, 0, 1); isl_vec_free(ineq); - if (!sol) + if (sol->error) goto error; continue; } if (tab->rational) break; - row = first_non_integer(tab, &flags); + row = first_non_integer_row(tab, &flags); if (row < 0) break; if (ISL_FL_ISSET(flags, I_PAR)) { if (ISL_FL_ISSET(flags, I_VAR)) { - tab = isl_tab_mark_empty(tab); + if (isl_tab_mark_empty(tab) < 0) + goto error; break; } row = add_cut(tab, row); @@ -3355,26 +3704,30 @@ static struct isl_sol *find_solutions(struct isl_sol *sol, struct isl_tab *tab) if (d < 0) goto error; ineq = ineq_for_div(context->op->peek_basic_set(context), d); - sol = no_sol_in_strict(sol, tab, ineq); + if (!ineq) + goto error; + sol_inc_level(sol); + no_sol_in_strict(sol, tab, ineq); isl_seq_neg(ineq->el, ineq->el, ineq->size); context->op->add_ineq(context, ineq->el, 1, 1); isl_vec_free(ineq); - if (!sol || !context->op->is_ok(context)) + if (sol->error || !context->op->is_ok(context)) goto error; tab = set_row_cst_to_div(tab, row, d); + if (context->op->is_empty(context)) + break; } else row = add_parametric_cut(tab, row, context); if (row < 0) goto error; } done: - sol = sol->add(sol, tab); + sol_add(sol, tab); isl_tab_free(tab); - return sol; + return; error: isl_tab_free(tab); - sol_free(sol); - return NULL; + sol->error = 1; } /* Compute the lexicographic minimum of the set represented by the main @@ -3388,11 +3741,15 @@ error: * In parts of the context where the added equality does not hold, * the main tableau is marked as being empty. */ -static struct isl_sol *find_solutions_main(struct isl_sol *sol, - struct isl_tab *tab) +static void find_solutions_main(struct isl_sol *sol, struct isl_tab *tab) { int row; + if (!tab) + goto error; + + sol->level = 0; + for (row = tab->n_redundant; row < tab->n_row; ++row) { int p; struct isl_vec *eq; @@ -3409,14 +3766,18 @@ static struct isl_sol *find_solutions_main(struct isl_sol *sol, + tab->n_param - (tab->n_var - tab->n_div); eq = isl_vec_alloc(tab->mat->ctx, 1+tab->n_param+tab->n_div); + if (!eq) + goto error; get_row_parameter_line(tab, row, eq->el); isl_int_neg(eq->el[1 + p], tab->mat->row[row][0]); eq = isl_vec_normalize(eq); - sol = no_sol_in_strict(sol, tab, eq); + sol_inc_level(sol); + no_sol_in_strict(sol, tab, eq); isl_seq_neg(eq->el, eq->el, eq->size); - sol = no_sol_in_strict(sol, tab, eq); + sol_inc_level(sol); + no_sol_in_strict(sol, tab, eq); isl_seq_neg(eq->el, eq->el, eq->size); sol->context->op->add_eq(sol->context, eq->el, 1, 1); @@ -3432,17 +3793,21 @@ static struct isl_sol *find_solutions_main(struct isl_sol *sol, row = tab->n_redundant - 1; } - return find_solutions(sol, tab); + find_solutions(sol, tab); + + sol->level = 0; + sol_pop(sol); + + return; error: isl_tab_free(tab); - sol_free(sol); - return NULL; + sol->error = 1; } -static struct isl_sol_map *sol_map_find_solutions(struct isl_sol_map *sol_map, +static void sol_map_find_solutions(struct isl_sol_map *sol_map, struct isl_tab *tab) { - return (struct isl_sol_map *)find_solutions_main(&sol_map->sol, tab); + find_solutions_main(&sol_map->sol, tab); } /* Check if integer division "div" of "dom" also occurs in "bmap". @@ -3521,40 +3886,22 @@ error: return NULL; } -/* Compute the lexicographic minimum (or maximum if "max" is set) - * of "bmap" over the domain "dom" and return the result as a map. - * If "empty" is not NULL, then *empty is assigned a set that - * contains those parts of the domain where there is no solution. - * If "bmap" is marked as rational (ISL_BASIC_MAP_RATIONAL), - * then we compute the rational optimum. Otherwise, we compute - * the integral optimum. +/* Base case of isl_tab_basic_map_partial_lexopt, after removing + * some obvious symmetries. * - * We perform some preprocessing. As the PILP solver does not - * handle implicit equalities very well, we first make sure all - * the equalities are explicitly available. - * We also make sure the divs in the domain are properly order, + * We make sure the divs in the domain are properly ordered, * because they will be added one by one in the given order * during the construction of the solution map. */ -struct isl_map *isl_tab_basic_map_partial_lexopt( - struct isl_basic_map *bmap, struct isl_basic_set *dom, - struct isl_set **empty, int max) +static __isl_give isl_map *basic_map_partial_lexopt_base( + __isl_take isl_basic_map *bmap, __isl_take isl_basic_set *dom, + __isl_give isl_set **empty, int max) { + isl_map *result = NULL; struct isl_tab *tab; - struct isl_map *result = NULL; struct isl_sol_map *sol_map = NULL; struct isl_context *context; - if (empty) - *empty = NULL; - if (!bmap || !dom) - goto error; - - isl_assert(bmap->ctx, - isl_basic_map_compatible_domain(bmap, dom), goto error); - - bmap = isl_basic_map_detect_equalities(bmap); - if (dom->n_div) { dom = isl_basic_set_order_divs(dom); bmap = align_context_divs(bmap, dom); @@ -3567,24 +3914,531 @@ struct isl_map *isl_tab_basic_map_partial_lexopt( if (isl_basic_set_fast_is_empty(context->op->peek_basic_set(context))) /* nothing */; else if (isl_basic_map_fast_is_empty(bmap)) - sol_map = add_empty(sol_map); + sol_map_add_empty_if_needed(sol_map, + isl_basic_set_copy(context->op->peek_basic_set(context))); else { tab = tab_for_lexmin(bmap, context->op->peek_basic_set(context), 1, max); tab = context->op->detect_nonnegative_parameters(context, tab); - sol_map = sol_map_find_solutions(sol_map, tab); - if (!sol_map) - goto error; + sol_map_find_solutions(sol_map, tab); } + if (sol_map->sol.error) + goto error; result = isl_map_copy(sol_map->map); if (empty) *empty = isl_set_copy(sol_map->empty); - sol_map_free(sol_map); + sol_free(&sol_map->sol); isl_basic_map_free(bmap); return result; error: - sol_map_free(sol_map); + sol_free(&sol_map->sol); + isl_basic_map_free(bmap); + return NULL; +} + +/* Structure used during detection of parallel constraints. + * n_in: number of "input" variables: isl_dim_param + isl_dim_in + * n_out: number of "output" variables: isl_dim_out + isl_dim_div + * val: the coefficients of the output variables + */ +struct isl_constraint_equal_info { + isl_basic_map *bmap; + unsigned n_in; + unsigned n_out; + isl_int *val; +}; + +/* Check whether the coefficients of the output variables + * of the constraint in "entry" are equal to info->val. + */ +static int constraint_equal(const void *entry, const void *val) +{ + isl_int **row = (isl_int **)entry; + const struct isl_constraint_equal_info *info = val; + + return isl_seq_eq((*row) + 1 + info->n_in, info->val, info->n_out); +} + +/* Check whether "bmap" has a pair of constraints that have + * the same coefficients for the output variables. + * Note that the coefficients of the existentially quantified + * variables need to be zero since the existentially quantified + * of the result are usually not the same as those of the input. + * the isl_dim_out and isl_dim_div dimensions. + * If so, return 1 and return the row indices of the two constraints + * in *first and *second. + */ +static int parallel_constraints(__isl_keep isl_basic_map *bmap, + int *first, int *second) +{ + int i; + isl_ctx *ctx = isl_basic_map_get_ctx(bmap); + struct isl_hash_table *table = NULL; + struct isl_hash_table_entry *entry; + struct isl_constraint_equal_info info; + unsigned n_out; + unsigned n_div; + + ctx = isl_basic_map_get_ctx(bmap); + table = isl_hash_table_alloc(ctx, bmap->n_ineq); + if (!table) + goto error; + + info.n_in = isl_basic_map_dim(bmap, isl_dim_param) + + isl_basic_map_dim(bmap, isl_dim_in); + info.bmap = bmap; + n_out = isl_basic_map_dim(bmap, isl_dim_out); + n_div = isl_basic_map_dim(bmap, isl_dim_div); + info.n_out = n_out + n_div; + for (i = 0; i < bmap->n_ineq; ++i) { + uint32_t hash; + + info.val = bmap->ineq[i] + 1 + info.n_in; + if (isl_seq_first_non_zero(info.val, n_out) < 0) + continue; + if (isl_seq_first_non_zero(info.val + n_out, n_div) >= 0) + continue; + hash = isl_seq_get_hash(info.val, info.n_out); + entry = isl_hash_table_find(ctx, table, hash, + constraint_equal, &info, 1); + if (!entry) + goto error; + if (entry->data) + break; + entry->data = &bmap->ineq[i]; + } + + if (i < bmap->n_ineq) { + *first = ((isl_int **)entry->data) - bmap->ineq; + *second = i; + } + + isl_hash_table_free(ctx, table); + + return i < bmap->n_ineq; +error: + isl_hash_table_free(ctx, table); + return -1; +} + +/* Given a set of upper bounds on the last "input" variable m, + * construct a set that assigns the minimal upper bound to m, i.e., + * construct a set that divides the space into cells where one + * of the upper bounds is smaller than all the others and assign + * this upper bound to m. + * + * In particular, if there are n bounds b_i, then the result + * consists of n basic sets, each one of the form + * + * m = b_i + * b_i <= b_j for j > i + * b_i < b_j for j < i + */ +static __isl_give isl_set *set_minimum(__isl_take isl_dim *dim, + __isl_take isl_mat *var) +{ + int i, j, k; + isl_basic_set *bset = NULL; + isl_ctx *ctx; + isl_set *set = NULL; + + if (!dim || !var) + goto error; + + ctx = isl_dim_get_ctx(dim); + set = isl_set_alloc_dim(isl_dim_copy(dim), + var->n_row, ISL_SET_DISJOINT); + + for (i = 0; i < var->n_row; ++i) { + bset = isl_basic_set_alloc_dim(isl_dim_copy(dim), 0, + 1, var->n_row - 1); + k = isl_basic_set_alloc_equality(bset); + if (k < 0) + goto error; + isl_seq_cpy(bset->eq[k], var->row[i], var->n_col); + isl_int_set_si(bset->eq[k][var->n_col], -1); + for (j = 0; j < var->n_row; ++j) { + if (j == i) + continue; + k = isl_basic_set_alloc_inequality(bset); + if (k < 0) + goto error; + isl_seq_combine(bset->ineq[k], ctx->one, var->row[j], + ctx->negone, var->row[i], + var->n_col); + isl_int_set_si(bset->ineq[k][var->n_col], 0); + if (j < i) + isl_int_sub_ui(bset->ineq[k][0], + bset->ineq[k][0], 1); + } + bset = isl_basic_set_finalize(bset); + set = isl_set_add_basic_set(set, bset); + } + + isl_dim_free(dim); + isl_mat_free(var); + return set; +error: + isl_basic_set_free(bset); + isl_set_free(set); + isl_dim_free(dim); + isl_mat_free(var); + return NULL; +} + +/* Given that the last input variable of "bmap" represents the minimum + * of the bounds in "cst", check whether we need to split the domain + * based on which bound attains the minimum. + * + * A split is needed when the minimum appears in an integer division + * or in an equality. Otherwise, it is only needed if it appears in + * an upper bound that is different from the upper bounds on which it + * is defined. + */ +static int need_split_map(__isl_keep isl_basic_map *bmap, + __isl_keep isl_mat *cst) +{ + int i, j; + unsigned total; + unsigned pos; + + pos = cst->n_col - 1; + total = isl_basic_map_dim(bmap, isl_dim_all); + + for (i = 0; i < bmap->n_div; ++i) + if (!isl_int_is_zero(bmap->div[i][2 + pos])) + return 1; + + for (i = 0; i < bmap->n_eq; ++i) + if (!isl_int_is_zero(bmap->eq[i][1 + pos])) + return 1; + + for (i = 0; i < bmap->n_ineq; ++i) { + if (isl_int_is_nonneg(bmap->ineq[i][1 + pos])) + continue; + if (!isl_int_is_negone(bmap->ineq[i][1 + pos])) + return 1; + if (isl_seq_first_non_zero(bmap->ineq[i] + 1 + pos + 1, + total - pos - 1) >= 0) + return 1; + + for (j = 0; j < cst->n_row; ++j) + if (isl_seq_eq(bmap->ineq[i], cst->row[j], cst->n_col)) + break; + if (j >= cst->n_row) + return 1; + } + + return 0; +} + +static int need_split_set(__isl_keep isl_basic_set *bset, + __isl_keep isl_mat *cst) +{ + return need_split_map((isl_basic_map *)bset, cst); +} + +/* Given a set of which the last set variable is the minimum + * of the bounds in "cst", split each basic set in the set + * in pieces where one of the bounds is (strictly) smaller than the others. + * This subdivision is given in "min_expr". + * The variable is subsequently projected out. + * + * We only do the split when it is needed. + * For example if the last input variable m = min(a,b) and the only + * constraints in the given basic set are lower bounds on m, + * i.e., l <= m = min(a,b), then we can simply project out m + * to obtain l <= a and l <= b, without having to split on whether + * m is equal to a or b. + */ +static __isl_give isl_set *split(__isl_take isl_set *empty, + __isl_take isl_set *min_expr, __isl_take isl_mat *cst) +{ + int n_in; + int i; + isl_dim *dim; + isl_set *res; + + if (!empty || !min_expr || !cst) + goto error; + + n_in = isl_set_dim(empty, isl_dim_set); + dim = isl_set_get_dim(empty); + dim = isl_dim_drop(dim, isl_dim_set, n_in - 1, 1); + res = isl_set_empty(dim); + + for (i = 0; i < empty->n; ++i) { + isl_set *set; + + set = isl_set_from_basic_set(isl_basic_set_copy(empty->p[i])); + if (need_split_set(empty->p[i], cst)) + set = isl_set_intersect(set, isl_set_copy(min_expr)); + set = isl_set_remove_dims(set, isl_dim_set, n_in - 1, 1); + + res = isl_set_union_disjoint(res, set); + } + + isl_set_free(empty); + isl_set_free(min_expr); + isl_mat_free(cst); + return res; +error: + isl_set_free(empty); + isl_set_free(min_expr); + isl_mat_free(cst); + return NULL; +} + +/* Given a map of which the last input variable is the minimum + * of the bounds in "cst", split each basic set in the set + * in pieces where one of the bounds is (strictly) smaller than the others. + * This subdivision is given in "min_expr". + * The variable is subsequently projected out. + * + * The implementation is essentially the same as that of "split". + */ +static __isl_give isl_map *split_domain(__isl_take isl_map *opt, + __isl_take isl_set *min_expr, __isl_take isl_mat *cst) +{ + int n_in; + int i; + isl_dim *dim; + isl_map *res; + + if (!opt || !min_expr || !cst) + goto error; + + n_in = isl_map_dim(opt, isl_dim_in); + dim = isl_map_get_dim(opt); + dim = isl_dim_drop(dim, isl_dim_in, n_in - 1, 1); + res = isl_map_empty(dim); + + for (i = 0; i < opt->n; ++i) { + isl_map *map; + + map = isl_map_from_basic_map(isl_basic_map_copy(opt->p[i])); + if (need_split_map(opt->p[i], cst)) + map = isl_map_intersect_domain(map, + isl_set_copy(min_expr)); + map = isl_map_remove_dims(map, isl_dim_in, n_in - 1, 1); + + res = isl_map_union_disjoint(res, map); + } + + isl_map_free(opt); + isl_set_free(min_expr); + isl_mat_free(cst); + return res; +error: + isl_map_free(opt); + isl_set_free(min_expr); + isl_mat_free(cst); + return NULL; +} + +static __isl_give isl_map *basic_map_partial_lexopt( + __isl_take isl_basic_map *bmap, __isl_take isl_basic_set *dom, + __isl_give isl_set **empty, int max); + +/* Given a basic map with at least two parallel constraints (as found + * by the function parallel_constraints), first look for more constraints + * parallel to the two constraint and replace the found list of parallel + * constraints by a single constraint with as "input" part the minimum + * of the input parts of the list of constraints. Then, recursively call + * basic_map_partial_lexopt (possibly finding more parallel constraints) + * and plug in the definition of the minimum in the result. + * + * More specifically, given a set of constraints + * + * a x + b_i(p) >= 0 + * + * Replace this set by a single constraint + * + * a x + u >= 0 + * + * with u a new parameter with constraints + * + * u <= b_i(p) + * + * Any solution to the new system is also a solution for the original system + * since + * + * a x >= -u >= -b_i(p) + * + * Moreover, m = min_i(b_i(p)) satisfies the constraints on u and can + * therefore be plugged into the solution. + */ +static __isl_give isl_map *basic_map_partial_lexopt_symm( + __isl_take isl_basic_map *bmap, __isl_take isl_basic_set *dom, + __isl_give isl_set **empty, int max, int first, int second) +{ + int i, n, k; + int *list = NULL; + unsigned n_in, n_out, n_div; + isl_ctx *ctx; + isl_vec *var = NULL; + isl_mat *cst = NULL; + isl_map *opt; + isl_set *min_expr; + isl_dim *map_dim, *set_dim; + + map_dim = isl_basic_map_get_dim(bmap); + set_dim = empty ? isl_basic_set_get_dim(dom) : NULL; + + n_in = isl_basic_map_dim(bmap, isl_dim_param) + + isl_basic_map_dim(bmap, isl_dim_in); + n_out = isl_basic_map_dim(bmap, isl_dim_all) - n_in; + + ctx = isl_basic_map_get_ctx(bmap); + list = isl_alloc_array(ctx, int, bmap->n_ineq); + var = isl_vec_alloc(ctx, n_out); + if (!list || !var) + goto error; + + list[0] = first; + list[1] = second; + isl_seq_cpy(var->el, bmap->ineq[first] + 1 + n_in, n_out); + for (i = second + 1, n = 2; i < bmap->n_ineq; ++i) { + if (isl_seq_eq(var->el, bmap->ineq[i] + 1 + n_in, n_out)) + list[n++] = i; + } + + cst = isl_mat_alloc(ctx, n, 1 + n_in); + if (!cst) + goto error; + + for (i = 0; i < n; ++i) + isl_seq_cpy(cst->row[i], bmap->ineq[list[i]], 1 + n_in); + + bmap = isl_basic_map_cow(bmap); + if (!bmap) + goto error; + for (i = n - 1; i >= 0; --i) + if (isl_basic_map_drop_inequality(bmap, list[i]) < 0) + goto error; + + bmap = isl_basic_map_add(bmap, isl_dim_in, 1); + bmap = isl_basic_map_extend_constraints(bmap, 0, 1); + k = isl_basic_map_alloc_inequality(bmap); + if (k < 0) + goto error; + isl_seq_clr(bmap->ineq[k], 1 + n_in); + isl_int_set_si(bmap->ineq[k][1 + n_in], 1); + isl_seq_cpy(bmap->ineq[k] + 1 + n_in + 1, var->el, n_out); + bmap = isl_basic_map_finalize(bmap); + + n_div = isl_basic_set_dim(dom, isl_dim_div); + dom = isl_basic_set_add(dom, isl_dim_set, 1); + dom = isl_basic_set_extend_constraints(dom, 0, n); + for (i = 0; i < n; ++i) { + k = isl_basic_set_alloc_inequality(dom); + if (k < 0) + goto error; + isl_seq_cpy(dom->ineq[k], cst->row[i], 1 + n_in); + isl_int_set_si(dom->ineq[k][1 + n_in], -1); + isl_seq_clr(dom->ineq[k] + 1 + n_in + 1, n_div); + } + + min_expr = set_minimum(isl_basic_set_get_dim(dom), isl_mat_copy(cst)); + + isl_vec_free(var); + free(list); + + opt = basic_map_partial_lexopt(bmap, dom, empty, max); + + if (empty) { + *empty = split(*empty, + isl_set_copy(min_expr), isl_mat_copy(cst)); + *empty = isl_set_reset_dim(*empty, set_dim); + } + + opt = split_domain(opt, min_expr, cst); + opt = isl_map_reset_dim(opt, map_dim); + + return opt; +error: + isl_dim_free(map_dim); + isl_dim_free(set_dim); + isl_mat_free(cst); + isl_vec_free(var); + free(list); + isl_basic_set_free(dom); + isl_basic_map_free(bmap); + return NULL; +} + +/* Recursive part of isl_tab_basic_map_partial_lexopt, after detecting + * equalities and removing redundant constraints. + * + * We first check if there are any parallel constraints (left). + * If not, we are in the base case. + * If there are parallel constraints, we replace them by a single + * constraint in basic_map_partial_lexopt_symm and then call + * this function recursively to look for more parallel constraints. + */ +static __isl_give isl_map *basic_map_partial_lexopt( + __isl_take isl_basic_map *bmap, __isl_take isl_basic_set *dom, + __isl_give isl_set **empty, int max) +{ + int par = 0; + int first, second; + + if (!bmap) + goto error; + + if (bmap->ctx->opt->pip_symmetry) + par = parallel_constraints(bmap, &first, &second); + if (par < 0) + goto error; + if (!par) + return basic_map_partial_lexopt_base(bmap, dom, empty, max); + + return basic_map_partial_lexopt_symm(bmap, dom, empty, max, + first, second); +error: + isl_basic_set_free(dom); + isl_basic_map_free(bmap); + return NULL; +} + +/* Compute the lexicographic minimum (or maximum if "max" is set) + * of "bmap" over the domain "dom" and return the result as a map. + * If "empty" is not NULL, then *empty is assigned a set that + * contains those parts of the domain where there is no solution. + * If "bmap" is marked as rational (ISL_BASIC_MAP_RATIONAL), + * then we compute the rational optimum. Otherwise, we compute + * the integral optimum. + * + * We perform some preprocessing. As the PILP solver does not + * handle implicit equalities very well, we first make sure all + * the equalities are explicitly available. + * + * We also add context constraints to the basic map and remove + * redundant constraints. This is only needed because of the + * way we handle simple symmetries. In particular, we currently look + * for symmetries on the constraints, before we set up the main tableau. + * It is then no good to look for symmetries on possibly redundant constraints. + */ +struct isl_map *isl_tab_basic_map_partial_lexopt( + struct isl_basic_map *bmap, struct isl_basic_set *dom, + struct isl_set **empty, int max) +{ + if (empty) + *empty = NULL; + if (!bmap || !dom) + goto error; + + isl_assert(bmap->ctx, + isl_basic_map_compatible_domain(bmap, dom), goto error); + + bmap = isl_basic_map_intersect_domain(bmap, isl_basic_set_copy(dom)); + bmap = isl_basic_map_detect_equalities(bmap); + bmap = isl_basic_map_remove_redundancies(bmap); + + return basic_map_partial_lexopt(bmap, dom, empty, max); +error: + isl_basic_set_free(dom); isl_basic_map_free(bmap); return NULL; } @@ -3594,7 +4448,6 @@ struct isl_sol_for { int (*fn)(__isl_take isl_basic_set *dom, __isl_take isl_mat *map, void *user); void *user; - int max; }; static void sol_for_free(struct isl_sol_for *sol_for) @@ -3611,11 +4464,11 @@ static void sol_for_free_wrap(struct isl_sol *sol) /* Add the solution identified by the tableau and the context tableau. * - * See documentation of sol_map_add for more details. + * See documentation of sol_add for more details. * * Instead of constructing a basic map, this function calls a user * defined function with the current context as a basic set and - * an affine matrix reprenting the relation between the input and output. + * an affine matrix representing the relation between the input and output. * The number of rows in this matrix is equal to one plus the number * of output variables. The number of columns is equal to one plus * the total dimension of the context, i.e., the number of parameters, @@ -3623,88 +4476,31 @@ static void sol_for_free_wrap(struct isl_sol *sol) * may refer to the divs, the basic set is not simplified. * (Simplification may reorder or remove divs.) */ -static struct isl_sol_for *sol_for_add(struct isl_sol_for *sol, - struct isl_tab *tab) +static void sol_for_add(struct isl_sol_for *sol, + struct isl_basic_set *dom, struct isl_mat *M) { - struct isl_basic_set *bset; - struct isl_mat *mat = NULL; - unsigned n_out; - unsigned off; - int row, i; - - if (!sol || !tab) - goto error; - - if (tab->empty) - return sol; - - off = 2 + tab->M; - - n_out = tab->n_var - tab->n_param - tab->n_div; - mat = isl_mat_alloc(tab->mat->ctx, 1 + n_out, 1 + tab->n_param + tab->n_div); - if (!mat) + if (sol->sol.error || !dom || !M) goto error; - isl_seq_clr(mat->row[0] + 1, mat->n_col - 1); - isl_int_set_si(mat->row[0][0], 1); - for (row = 0; row < n_out; ++row) { - int i = tab->n_param + row; - int r, j; - - isl_seq_clr(mat->row[1 + row], mat->n_col); - if (!tab->var[i].is_row) - continue; - - r = tab->var[i].index; - /* no unbounded */ - if (tab->M) - isl_assert(mat->ctx, isl_int_eq(tab->mat->row[r][2], - tab->mat->row[r][0]), - goto error); - isl_int_set(mat->row[1 + row][0], tab->mat->row[r][1]); - for (j = 0; j < tab->n_param; ++j) { - int col; - if (tab->var[j].is_row) - continue; - col = tab->var[j].index; - isl_int_set(mat->row[1 + row][1 + j], - tab->mat->row[r][off + col]); - } - for (j = 0; j < tab->n_div; ++j) { - int col; - if (tab->var[tab->n_var - tab->n_div+j].is_row) - continue; - col = tab->var[tab->n_var - tab->n_div+j].index; - isl_int_set(mat->row[1 + row][1 + tab->n_param + j], - tab->mat->row[r][off + col]); - } - if (!isl_int_is_one(tab->mat->row[r][0])) - isl_seq_scale_down(mat->row[1 + row], mat->row[1 + row], - tab->mat->row[r][0], mat->n_col); - if (sol->max) - isl_seq_neg(mat->row[1 + row], mat->row[1 + row], - mat->n_col); - } - - bset = sol->sol.context->op->peek_basic_set(sol->sol.context); - bset = isl_basic_set_dup(bset); - bset = isl_basic_set_finalize(bset); + dom = isl_basic_set_simplify(dom); + dom = isl_basic_set_finalize(dom); - if (sol->fn(bset, isl_mat_copy(mat), sol->user) < 0) + if (sol->fn(isl_basic_set_copy(dom), isl_mat_copy(M), sol->user) < 0) goto error; - isl_mat_free(mat); - return sol; + isl_basic_set_free(dom); + isl_mat_free(M); + return; error: - isl_mat_free(mat); - sol_free(&sol->sol); - return NULL; + isl_basic_set_free(dom); + isl_mat_free(M); + sol->sol.error = 1; } -static struct isl_sol *sol_for_add_wrap(struct isl_sol *sol, - struct isl_tab *tab) +static void sol_for_add_wrap(struct isl_sol *sol, + struct isl_basic_set *dom, struct isl_mat *M) { - return (struct isl_sol *)sol_for_add((struct isl_sol_for *)sol, tab); + sol_for_add((struct isl_sol_for *)sol, dom, M); } static struct isl_sol_for *sol_for_init(struct isl_basic_map *bmap, int max, @@ -3716,17 +4512,22 @@ static struct isl_sol_for *sol_for_init(struct isl_basic_map *bmap, int max, struct isl_dim *dom_dim; struct isl_basic_set *dom = NULL; - sol_for = isl_calloc_type(bset->ctx, struct isl_sol_for); + sol_for = isl_calloc_type(bmap->ctx, struct isl_sol_for); if (!sol_for) goto error; dom_dim = isl_dim_domain(isl_dim_copy(bmap->dim)); dom = isl_basic_set_universe(dom_dim); + sol_for->sol.rational = ISL_F_ISSET(bmap, ISL_BASIC_MAP_RATIONAL); + sol_for->sol.dec_level.callback.run = &sol_dec_level_wrap; + sol_for->sol.dec_level.sol = &sol_for->sol; sol_for->fn = fn; sol_for->user = user; - sol_for->max = max; + sol_for->sol.max = max; + sol_for->sol.n_out = isl_basic_map_dim(bmap, isl_dim_out); sol_for->sol.add = &sol_for_add_wrap; + sol_for->sol.add_empty = NULL; sol_for->sol.free = &sol_for_free_wrap; sol_for->sol.context = isl_context_alloc(dom); @@ -3741,10 +4542,10 @@ error: return NULL; } -static struct isl_sol_for *sol_for_find_solutions(struct isl_sol_for *sol_for, +static void sol_for_find_solutions(struct isl_sol_for *sol_for, struct isl_tab *tab) { - return (struct isl_sol_for *)find_solutions_main(&sol_for->sol, tab); + find_solutions_main(&sol_for->sol, tab); } int isl_basic_map_foreach_lexopt(__isl_keep isl_basic_map *bmap, int max, @@ -3769,16 +4570,16 @@ int isl_basic_map_foreach_lexopt(__isl_keep isl_basic_map *bmap, int max, tab = tab_for_lexmin(bmap, context->op->peek_basic_set(context), 1, max); tab = context->op->detect_nonnegative_parameters(context, tab); - sol_for = sol_for_find_solutions(sol_for, tab); - if (!sol_for) + sol_for_find_solutions(sol_for, tab); + if (sol_for->sol.error) goto error; } - sol_for_free(sol_for); + sol_free(&sol_for->sol); isl_basic_map_free(bmap); return 0; error: - sol_for_free(sol_for); + sol_free(&sol_for->sol); isl_basic_map_free(bmap); return -1; }