2 * Copyright 2010 INRIA Saclay
4 * Use of this software is governed by the GNU LGPLv2.1 license
6 * Written by Sven Verdoolaege, INRIA Saclay - Ile-de-France,
7 * Parc Club Orsay Universite, ZAC des vignes, 4 rue Jacques Monod,
12 #include "isl_map_private.h"
16 /* Given a map that represents a path with the length of the path
17 * encoded as the difference between the last output coordindate
18 * and the last input coordinate, set this length to either
19 * exactly "length" (if "exactly" is set) or at least "length"
20 * (if "exactly" is not set).
22 static __isl_give isl_map
*set_path_length(__isl_take isl_map
*map
,
23 int exactly
, int length
)
26 struct isl_basic_map
*bmap
;
35 dim
= isl_map_get_dim(map
);
36 d
= isl_dim_size(dim
, isl_dim_in
);
37 nparam
= isl_dim_size(dim
, isl_dim_param
);
38 bmap
= isl_basic_map_alloc_dim(dim
, 0, 1, 1);
40 k
= isl_basic_map_alloc_equality(bmap
);
43 k
= isl_basic_map_alloc_inequality(bmap
);
48 isl_seq_clr(c
, 1 + isl_basic_map_total_dim(bmap
));
49 isl_int_set_si(c
[0], -length
);
50 isl_int_set_si(c
[1 + nparam
+ d
- 1], -1);
51 isl_int_set_si(c
[1 + nparam
+ d
+ d
- 1], 1);
53 bmap
= isl_basic_map_finalize(bmap
);
54 map
= isl_map_intersect(map
, isl_map_from_basic_map(bmap
));
58 isl_basic_map_free(bmap
);
63 /* Check whether the overapproximation of the power of "map" is exactly
64 * the power of "map". Let R be "map" and A_k the overapproximation.
65 * The approximation is exact if
68 * A_k = A_{k-1} \circ R k >= 2
70 * Since A_k is known to be an overapproximation, we only need to check
73 * A_k \subset A_{k-1} \circ R k >= 2
75 * In practice, "app" has an extra input and output coordinate
76 * to encode the length of the path. So, we first need to add
77 * this coordinate to "map" and set the length of the path to
80 static int check_power_exactness(__isl_take isl_map
*map
,
81 __isl_take isl_map
*app
)
87 map
= isl_map_add(map
, isl_dim_in
, 1);
88 map
= isl_map_add(map
, isl_dim_out
, 1);
89 map
= set_path_length(map
, 1, 1);
91 app_1
= set_path_length(isl_map_copy(app
), 1, 1);
93 exact
= isl_map_is_subset(app_1
, map
);
96 if (!exact
|| exact
< 0) {
102 app_1
= set_path_length(isl_map_copy(app
), 0, 1);
103 app_2
= set_path_length(app
, 0, 2);
104 app_1
= isl_map_apply_range(map
, app_1
);
106 exact
= isl_map_is_subset(app_2
, app_1
);
114 /* Check whether the overapproximation of the power of "map" is exactly
115 * the power of "map", possibly after projecting out the power (if "project"
118 * If "project" is set and if "steps" can only result in acyclic paths,
121 * A = R \cup (A \circ R)
123 * where A is the overapproximation with the power projected out, i.e.,
124 * an overapproximation of the transitive closure.
125 * More specifically, since A is known to be an overapproximation, we check
127 * A \subset R \cup (A \circ R)
129 * Otherwise, we check if the power is exact.
131 * Note that "app" has an extra input and output coordinate to encode
132 * the length of the part. If we are only interested in the transitive
133 * closure, then we can simply project out these coordinates first.
135 static int check_exactness(__isl_take isl_map
*map
, __isl_take isl_map
*app
,
143 return check_power_exactness(map
, app
);
145 d
= isl_map_dim(map
, isl_dim_in
);
146 app
= set_path_length(app
, 0, 1);
147 app
= isl_map_project_out(app
, isl_dim_in
, d
, 1);
148 app
= isl_map_project_out(app
, isl_dim_out
, d
, 1);
150 test
= isl_map_apply_range(isl_map_copy(map
), isl_map_copy(app
));
151 test
= isl_map_union(test
, isl_map_copy(map
));
153 exact
= isl_map_is_subset(app
, test
);
168 * The transitive closure implementation is based on the paper
169 * "Computing the Transitive Closure of a Union of Affine Integer
170 * Tuple Relations" by Anna Beletska, Denis Barthou, Wlodzimierz Bielecki and
174 /* Given a set of n offsets v_i (the rows of "steps"), construct a relation
175 * of the given dimension specification (Z^{n+1} -> Z^{n+1})
176 * that maps an element x to any element that can be reached
177 * by taking a non-negative number of steps along any of
178 * the extended offsets v'_i = [v_i 1].
181 * { [x] -> [y] : exists k_i >= 0, y = x + \sum_i k_i v'_i }
183 * For any element in this relation, the number of steps taken
184 * is equal to the difference in the final coordinates.
186 static __isl_give isl_map
*path_along_steps(__isl_take isl_dim
*dim
,
187 __isl_keep isl_mat
*steps
)
190 struct isl_basic_map
*path
= NULL
;
198 d
= isl_dim_size(dim
, isl_dim_in
);
200 nparam
= isl_dim_size(dim
, isl_dim_param
);
202 path
= isl_basic_map_alloc_dim(isl_dim_copy(dim
), n
, d
, n
);
204 for (i
= 0; i
< n
; ++i
) {
205 k
= isl_basic_map_alloc_div(path
);
208 isl_assert(steps
->ctx
, i
== k
, goto error
);
209 isl_int_set_si(path
->div
[k
][0], 0);
212 for (i
= 0; i
< d
; ++i
) {
213 k
= isl_basic_map_alloc_equality(path
);
216 isl_seq_clr(path
->eq
[k
], 1 + isl_basic_map_total_dim(path
));
217 isl_int_set_si(path
->eq
[k
][1 + nparam
+ i
], 1);
218 isl_int_set_si(path
->eq
[k
][1 + nparam
+ d
+ i
], -1);
220 for (j
= 0; j
< n
; ++j
)
221 isl_int_set_si(path
->eq
[k
][1 + nparam
+ 2 * d
+ j
], 1);
223 for (j
= 0; j
< n
; ++j
)
224 isl_int_set(path
->eq
[k
][1 + nparam
+ 2 * d
+ j
],
228 for (i
= 0; i
< n
; ++i
) {
229 k
= isl_basic_map_alloc_inequality(path
);
232 isl_seq_clr(path
->ineq
[k
], 1 + isl_basic_map_total_dim(path
));
233 isl_int_set_si(path
->ineq
[k
][1 + nparam
+ 2 * d
+ i
], 1);
238 path
= isl_basic_map_simplify(path
);
239 path
= isl_basic_map_finalize(path
);
240 return isl_map_from_basic_map(path
);
243 isl_basic_map_free(path
);
252 /* Check whether the parametric constant term of constraint c is never
253 * positive in "bset".
255 static int parametric_constant_never_positive(__isl_keep isl_basic_set
*bset
,
256 isl_int
*c
, int *div_purity
)
265 n_div
= isl_basic_set_dim(bset
, isl_dim_div
);
266 d
= isl_basic_set_dim(bset
, isl_dim_set
);
267 nparam
= isl_basic_set_dim(bset
, isl_dim_param
);
269 bset
= isl_basic_set_copy(bset
);
270 bset
= isl_basic_set_cow(bset
);
271 bset
= isl_basic_set_extend_constraints(bset
, 0, 1);
272 k
= isl_basic_set_alloc_inequality(bset
);
275 isl_seq_clr(bset
->ineq
[k
], 1 + isl_basic_set_total_dim(bset
));
276 isl_seq_cpy(bset
->ineq
[k
], c
, 1 + nparam
);
277 for (i
= 0; i
< n_div
; ++i
) {
278 if (div_purity
[i
] != PURE_PARAM
)
280 isl_int_set(bset
->ineq
[k
][1 + nparam
+ d
+ i
],
281 c
[1 + nparam
+ d
+ i
]);
283 isl_int_sub_ui(bset
->ineq
[k
][0], bset
->ineq
[k
][0], 1);
284 empty
= isl_basic_set_is_empty(bset
);
285 isl_basic_set_free(bset
);
289 isl_basic_set_free(bset
);
293 /* Return PURE_PARAM if only the coefficients of the parameters are non-zero.
294 * Return PURE_VAR if only the coefficients of the set variables are non-zero.
295 * Return MIXED if only the coefficients of the parameters and the set
296 * variables are non-zero and if moreover the parametric constant
297 * can never attain positive values.
298 * Return IMPURE otherwise.
300 static int purity(__isl_keep isl_basic_set
*bset
, isl_int
*c
, int *div_purity
,
310 n_div
= isl_basic_set_dim(bset
, isl_dim_div
);
311 d
= isl_basic_set_dim(bset
, isl_dim_set
);
312 nparam
= isl_basic_set_dim(bset
, isl_dim_param
);
314 for (i
= 0; i
< n_div
; ++i
) {
315 if (isl_int_is_zero(c
[1 + nparam
+ d
+ i
]))
317 switch (div_purity
[i
]) {
318 case PURE_PARAM
: p
= 1; break;
319 case PURE_VAR
: v
= 1; break;
320 default: return IMPURE
;
323 if (!p
&& isl_seq_first_non_zero(c
+ 1, nparam
) == -1)
325 if (!v
&& isl_seq_first_non_zero(c
+ 1 + nparam
, d
) == -1)
328 empty
= parametric_constant_never_positive(bset
, c
, div_purity
);
329 if (eq
&& empty
>= 0 && !empty
) {
330 isl_seq_neg(c
, c
, 1 + nparam
+ d
+ n_div
);
331 empty
= parametric_constant_never_positive(bset
, c
, div_purity
);
334 return empty
< 0 ? -1 : empty
? MIXED
: IMPURE
;
337 /* Return an array of integers indicating the type of each div in bset.
338 * If the div is (recursively) defined in terms of only the parameters,
339 * then the type is PURE_PARAM.
340 * If the div is (recursively) defined in terms of only the set variables,
341 * then the type is PURE_VAR.
342 * Otherwise, the type is IMPURE.
344 static __isl_give
int *get_div_purity(__isl_keep isl_basic_set
*bset
)
355 n_div
= isl_basic_set_dim(bset
, isl_dim_div
);
356 d
= isl_basic_set_dim(bset
, isl_dim_set
);
357 nparam
= isl_basic_set_dim(bset
, isl_dim_param
);
359 div_purity
= isl_alloc_array(bset
->ctx
, int, n_div
);
363 for (i
= 0; i
< bset
->n_div
; ++i
) {
365 if (isl_int_is_zero(bset
->div
[i
][0])) {
366 div_purity
[i
] = IMPURE
;
369 if (isl_seq_first_non_zero(bset
->div
[i
] + 2, nparam
) != -1)
371 if (isl_seq_first_non_zero(bset
->div
[i
] + 2 + nparam
, d
) != -1)
373 for (j
= 0; j
< i
; ++j
) {
374 if (isl_int_is_zero(bset
->div
[i
][2 + nparam
+ d
+ j
]))
376 switch (div_purity
[j
]) {
377 case PURE_PARAM
: p
= 1; break;
378 case PURE_VAR
: v
= 1; break;
379 default: p
= v
= 1; break;
382 div_purity
[i
] = v
? p
? IMPURE
: PURE_VAR
: PURE_PARAM
;
388 /* Given a path with the as yet unconstrained length at position "pos",
389 * check if setting the length to zero results in only the identity
392 int empty_path_is_identity(__isl_keep isl_basic_map
*path
, unsigned pos
)
394 isl_basic_map
*test
= NULL
;
395 isl_basic_map
*id
= NULL
;
399 test
= isl_basic_map_copy(path
);
400 test
= isl_basic_map_extend_constraints(test
, 1, 0);
401 k
= isl_basic_map_alloc_equality(test
);
404 isl_seq_clr(test
->eq
[k
], 1 + isl_basic_map_total_dim(test
));
405 isl_int_set_si(test
->eq
[k
][pos
], 1);
406 id
= isl_basic_map_identity(isl_dim_domain(isl_basic_map_get_dim(path
)));
407 is_id
= isl_basic_map_is_equal(test
, id
);
408 isl_basic_map_free(test
);
409 isl_basic_map_free(id
);
412 isl_basic_map_free(test
);
416 __isl_give isl_basic_map
*add_delta_constraints(__isl_take isl_basic_map
*path
,
417 __isl_keep isl_basic_set
*delta
, unsigned off
, unsigned nparam
,
418 unsigned d
, int *div_purity
, int eq
)
421 int n
= eq
? delta
->n_eq
: delta
->n_ineq
;
422 isl_int
**delta_c
= eq
? delta
->eq
: delta
->ineq
;
425 n_div
= isl_basic_set_dim(delta
, isl_dim_div
);
427 for (i
= 0; i
< n
; ++i
) {
429 int p
= purity(delta
, delta_c
[i
], div_purity
, eq
);
434 if (eq
&& p
!= MIXED
) {
435 k
= isl_basic_map_alloc_equality(path
);
436 path_c
= path
->eq
[k
];
438 k
= isl_basic_map_alloc_inequality(path
);
439 path_c
= path
->ineq
[k
];
443 isl_seq_clr(path_c
, 1 + isl_basic_map_total_dim(path
));
445 isl_seq_cpy(path_c
+ off
,
446 delta_c
[i
] + 1 + nparam
, d
);
447 isl_int_set(path_c
[off
+ d
], delta_c
[i
][0]);
448 } else if (p
== PURE_PARAM
) {
449 isl_seq_cpy(path_c
, delta_c
[i
], 1 + nparam
);
451 isl_seq_cpy(path_c
+ off
,
452 delta_c
[i
] + 1 + nparam
, d
);
453 isl_seq_cpy(path_c
, delta_c
[i
], 1 + nparam
);
455 isl_seq_cpy(path_c
+ off
- n_div
,
456 delta_c
[i
] + 1 + nparam
+ d
, n_div
);
461 isl_basic_map_free(path
);
465 /* Given a set of offsets "delta", construct a relation of the
466 * given dimension specification (Z^{n+1} -> Z^{n+1}) that
467 * is an overapproximation of the relations that
468 * maps an element x to any element that can be reached
469 * by taking a non-negative number of steps along any of
470 * the elements in "delta".
471 * That is, construct an approximation of
473 * { [x] -> [y] : exists f \in \delta, k \in Z :
474 * y = x + k [f, 1] and k >= 0 }
476 * For any element in this relation, the number of steps taken
477 * is equal to the difference in the final coordinates.
479 * In particular, let delta be defined as
481 * \delta = [p] -> { [x] : A x + a >= and B p + b >= 0 and
482 * C x + C'p + c >= 0 and
483 * D x + D'p + d >= 0 }
485 * where the constraints C x + C'p + c >= 0 are such that the parametric
486 * constant term of each constraint j, "C_j x + C'_j p + c_j",
487 * can never attain positive values, then the relation is constructed as
489 * { [x] -> [y] : exists [f, k] \in Z^{n+1} : y = x + f and
490 * A f + k a >= 0 and B p + b >= 0 and
491 * C f + C'p + c >= 0 and k >= 1 }
492 * union { [x] -> [x] }
494 * If the zero-length paths happen to correspond exactly to the identity
495 * mapping, then we return
497 * { [x] -> [y] : exists [f, k] \in Z^{n+1} : y = x + f and
498 * A f + k a >= 0 and B p + b >= 0 and
499 * C f + C'p + c >= 0 and k >= 0 }
503 * Existentially quantified variables in \delta are handled by
504 * classifying them as independent of the parameters, purely
505 * parameter dependent and others. Constraints containing
506 * any of the other existentially quantified variables are removed.
507 * This is safe, but leads to an additional overapproximation.
509 static __isl_give isl_map
*path_along_delta(__isl_take isl_dim
*dim
,
510 __isl_take isl_basic_set
*delta
)
512 isl_basic_map
*path
= NULL
;
519 int *div_purity
= NULL
;
523 n_div
= isl_basic_set_dim(delta
, isl_dim_div
);
524 d
= isl_basic_set_dim(delta
, isl_dim_set
);
525 nparam
= isl_basic_set_dim(delta
, isl_dim_param
);
526 path
= isl_basic_map_alloc_dim(isl_dim_copy(dim
), n_div
+ d
+ 1,
527 d
+ 1 + delta
->n_eq
, delta
->n_eq
+ delta
->n_ineq
+ 1);
528 off
= 1 + nparam
+ 2 * (d
+ 1) + n_div
;
530 for (i
= 0; i
< n_div
+ d
+ 1; ++i
) {
531 k
= isl_basic_map_alloc_div(path
);
534 isl_int_set_si(path
->div
[k
][0], 0);
537 for (i
= 0; i
< d
+ 1; ++i
) {
538 k
= isl_basic_map_alloc_equality(path
);
541 isl_seq_clr(path
->eq
[k
], 1 + isl_basic_map_total_dim(path
));
542 isl_int_set_si(path
->eq
[k
][1 + nparam
+ i
], 1);
543 isl_int_set_si(path
->eq
[k
][1 + nparam
+ d
+ 1 + i
], -1);
544 isl_int_set_si(path
->eq
[k
][off
+ i
], 1);
547 div_purity
= get_div_purity(delta
);
551 path
= add_delta_constraints(path
, delta
, off
, nparam
, d
, div_purity
, 1);
552 path
= add_delta_constraints(path
, delta
, off
, nparam
, d
, div_purity
, 0);
554 is_id
= empty_path_is_identity(path
, off
+ d
);
558 k
= isl_basic_map_alloc_inequality(path
);
561 isl_seq_clr(path
->ineq
[k
], 1 + isl_basic_map_total_dim(path
));
563 isl_int_set_si(path
->ineq
[k
][0], -1);
564 isl_int_set_si(path
->ineq
[k
][off
+ d
], 1);
567 isl_basic_set_free(delta
);
568 path
= isl_basic_map_finalize(path
);
571 return isl_map_from_basic_map(path
);
573 return isl_basic_map_union(path
,
574 isl_basic_map_identity(isl_dim_domain(dim
)));
578 isl_basic_set_free(delta
);
579 isl_basic_map_free(path
);
583 /* Given a dimenion specification Z^{n+1} -> Z^{n+1} and a parameter "param",
584 * construct a map that equates the parameter to the difference
585 * in the final coordinates and imposes that this difference is positive.
588 * { [x,x_s] -> [y,y_s] : k = y_s - x_s > 0 }
590 static __isl_give isl_map
*equate_parameter_to_length(__isl_take isl_dim
*dim
,
593 struct isl_basic_map
*bmap
;
598 d
= isl_dim_size(dim
, isl_dim_in
);
599 nparam
= isl_dim_size(dim
, isl_dim_param
);
600 bmap
= isl_basic_map_alloc_dim(dim
, 0, 1, 1);
601 k
= isl_basic_map_alloc_equality(bmap
);
604 isl_seq_clr(bmap
->eq
[k
], 1 + isl_basic_map_total_dim(bmap
));
605 isl_int_set_si(bmap
->eq
[k
][1 + param
], -1);
606 isl_int_set_si(bmap
->eq
[k
][1 + nparam
+ d
- 1], -1);
607 isl_int_set_si(bmap
->eq
[k
][1 + nparam
+ d
+ d
- 1], 1);
609 k
= isl_basic_map_alloc_inequality(bmap
);
612 isl_seq_clr(bmap
->ineq
[k
], 1 + isl_basic_map_total_dim(bmap
));
613 isl_int_set_si(bmap
->ineq
[k
][1 + param
], 1);
614 isl_int_set_si(bmap
->ineq
[k
][0], -1);
616 bmap
= isl_basic_map_finalize(bmap
);
617 return isl_map_from_basic_map(bmap
);
619 isl_basic_map_free(bmap
);
623 /* Check whether "path" is acyclic, where the last coordinates of domain
624 * and range of path encode the number of steps taken.
625 * That is, check whether
627 * { d | d = y - x and (x,y) in path }
629 * does not contain any element with positive last coordinate (positive length)
630 * and zero remaining coordinates (cycle).
632 static int is_acyclic(__isl_take isl_map
*path
)
637 struct isl_set
*delta
;
639 delta
= isl_map_deltas(path
);
640 dim
= isl_set_dim(delta
, isl_dim_set
);
641 for (i
= 0; i
< dim
; ++i
) {
643 delta
= isl_set_lower_bound_si(delta
, isl_dim_set
, i
, 1);
645 delta
= isl_set_fix_si(delta
, isl_dim_set
, i
, 0);
648 acyclic
= isl_set_is_empty(delta
);
654 /* Given a union of basic maps R = \cup_i R_i \subseteq D \times D
655 * and a dimension specification (Z^{n+1} -> Z^{n+1}),
656 * construct a map that is an overapproximation of the map
657 * that takes an element from the space D \times Z to another
658 * element from the same space, such that the first n coordinates of the
659 * difference between them is a sum of differences between images
660 * and pre-images in one of the R_i and such that the last coordinate
661 * is equal to the number of steps taken.
664 * \Delta_i = { y - x | (x, y) in R_i }
666 * then the constructed map is an overapproximation of
668 * { (x) -> (x + d) | \exists k_i >= 0, \delta_i \in \Delta_i :
669 * d = (\sum_i k_i \delta_i, \sum_i k_i) }
671 * The elements of the singleton \Delta_i's are collected as the
672 * rows of the steps matrix. For all these \Delta_i's together,
673 * a single path is constructed.
674 * For each of the other \Delta_i's, we compute an overapproximation
675 * of the paths along elements of \Delta_i.
676 * Since each of these paths performs an addition, composition is
677 * symmetric and we can simply compose all resulting paths in any order.
679 static __isl_give isl_map
*construct_extended_path(__isl_take isl_dim
*dim
,
680 __isl_keep isl_map
*map
, int *project
)
682 struct isl_mat
*steps
= NULL
;
683 struct isl_map
*path
= NULL
;
687 d
= isl_map_dim(map
, isl_dim_in
);
689 path
= isl_map_identity(isl_dim_domain(isl_dim_copy(dim
)));
691 steps
= isl_mat_alloc(map
->ctx
, map
->n
, d
);
696 for (i
= 0; i
< map
->n
; ++i
) {
697 struct isl_basic_set
*delta
;
699 delta
= isl_basic_map_deltas(isl_basic_map_copy(map
->p
[i
]));
701 for (j
= 0; j
< d
; ++j
) {
704 fixed
= isl_basic_set_fast_dim_is_fixed(delta
, j
,
707 isl_basic_set_free(delta
);
716 path
= isl_map_apply_range(path
,
717 path_along_delta(isl_dim_copy(dim
), delta
));
718 path
= isl_map_coalesce(path
);
720 isl_basic_set_free(delta
);
727 path
= isl_map_apply_range(path
,
728 path_along_steps(isl_dim_copy(dim
), steps
));
731 if (project
&& *project
) {
732 *project
= is_acyclic(isl_map_copy(path
));
747 static int isl_set_overlaps(__isl_keep isl_set
*set1
, __isl_keep isl_set
*set2
)
752 i
= isl_set_intersect(isl_set_copy(set1
), isl_set_copy(set2
));
753 no_overlap
= isl_set_is_empty(i
);
756 return no_overlap
< 0 ? -1 : !no_overlap
;
759 /* Given a union of basic maps R = \cup_i R_i \subseteq D \times D
760 * and a dimension specification (Z^{n+1} -> Z^{n+1}),
761 * construct a map that is an overapproximation of the map
762 * that takes an element from the dom R \times Z to an
763 * element from ran R \times Z, such that the first n coordinates of the
764 * difference between them is a sum of differences between images
765 * and pre-images in one of the R_i and such that the last coordinate
766 * is equal to the number of steps taken.
769 * \Delta_i = { y - x | (x, y) in R_i }
771 * then the constructed map is an overapproximation of
773 * { (x) -> (x + d) | \exists k_i >= 0, \delta_i \in \Delta_i :
774 * d = (\sum_i k_i \delta_i, \sum_i k_i) and
775 * x in dom R and x + d in ran R and
778 static __isl_give isl_map
*construct_component(__isl_take isl_dim
*dim
,
779 __isl_keep isl_map
*map
, int *exact
, int project
)
781 struct isl_set
*domain
= NULL
;
782 struct isl_set
*range
= NULL
;
783 struct isl_map
*app
= NULL
;
784 struct isl_map
*path
= NULL
;
786 domain
= isl_map_domain(isl_map_copy(map
));
787 domain
= isl_set_coalesce(domain
);
788 range
= isl_map_range(isl_map_copy(map
));
789 range
= isl_set_coalesce(range
);
790 if (!isl_set_overlaps(domain
, range
)) {
791 isl_set_free(domain
);
795 map
= isl_map_copy(map
);
796 map
= isl_map_add(map
, isl_dim_in
, 1);
797 map
= isl_map_add(map
, isl_dim_out
, 1);
798 map
= set_path_length(map
, 1, 1);
801 app
= isl_map_from_domain_and_range(domain
, range
);
802 app
= isl_map_add(app
, isl_dim_in
, 1);
803 app
= isl_map_add(app
, isl_dim_out
, 1);
805 path
= construct_extended_path(isl_dim_copy(dim
), map
,
806 exact
&& *exact
? &project
: NULL
);
807 app
= isl_map_intersect(app
, path
);
809 if (exact
&& *exact
&&
810 (*exact
= check_exactness(isl_map_copy(map
), isl_map_copy(app
),
815 app
= set_path_length(app
, 0, 1);
823 /* Call construct_component and, if "project" is set, project out
824 * the final coordinates.
826 static __isl_give isl_map
*construct_projected_component(
827 __isl_take isl_dim
*dim
,
828 __isl_keep isl_map
*map
, int *exact
, int project
)
835 d
= isl_dim_size(dim
, isl_dim_in
);
837 app
= construct_component(dim
, map
, exact
, project
);
839 app
= isl_map_project_out(app
, isl_dim_in
, d
- 1, 1);
840 app
= isl_map_project_out(app
, isl_dim_out
, d
- 1, 1);
845 /* Compute an extended version, i.e., with path lengths, of
846 * an overapproximation of the transitive closure of "bmap"
847 * with path lengths greater than or equal to zero and with
848 * domain and range equal to "dom".
850 static __isl_give isl_map
*q_closure(__isl_take isl_dim
*dim
,
851 __isl_take isl_set
*dom
, __isl_keep isl_basic_map
*bmap
, int *exact
)
858 dom
= isl_set_add(dom
, isl_dim_set
, 1);
859 app
= isl_map_from_domain_and_range(dom
, isl_set_copy(dom
));
860 map
= isl_map_from_basic_map(isl_basic_map_copy(bmap
));
861 path
= construct_extended_path(dim
, map
, &project
);
862 app
= isl_map_intersect(app
, path
);
864 if ((*exact
= check_exactness(map
, isl_map_copy(app
), project
)) < 0)
873 /* Check whether qc has any elements of length at least one
874 * with domain and/or range outside of dom and ran.
876 static int has_spurious_elements(__isl_keep isl_map
*qc
,
877 __isl_keep isl_set
*dom
, __isl_keep isl_set
*ran
)
883 if (!qc
|| !dom
|| !ran
)
886 d
= isl_map_dim(qc
, isl_dim_in
);
888 qc
= isl_map_copy(qc
);
889 qc
= set_path_length(qc
, 0, 1);
890 qc
= isl_map_project_out(qc
, isl_dim_in
, d
- 1, 1);
891 qc
= isl_map_project_out(qc
, isl_dim_out
, d
- 1, 1);
893 s
= isl_map_domain(isl_map_copy(qc
));
894 subset
= isl_set_is_subset(s
, dom
);
903 s
= isl_map_range(qc
);
904 subset
= isl_set_is_subset(s
, ran
);
907 return subset
< 0 ? -1 : !subset
;
916 /* For each basic map in "map", except i, check whether it combines
917 * with the transitive closure that is reflexive on C combines
918 * to the left and to the right.
922 * dom map_j \subseteq C
924 * then right[j] is set to 1. Otherwise, if
926 * ran map_i \cap dom map_j = \emptyset
928 * then right[j] is set to 0. Otherwise, composing to the right
931 * Similar, for composing to the left, we have if
933 * ran map_j \subseteq C
935 * then left[j] is set to 1. Otherwise, if
937 * dom map_i \cap ran map_j = \emptyset
939 * then left[j] is set to 0. Otherwise, composing to the left
942 * The return value is or'd with LEFT if composing to the left
943 * is possible and with RIGHT if composing to the right is possible.
945 static int composability(__isl_keep isl_set
*C
, int i
,
946 isl_set
**dom
, isl_set
**ran
, int *left
, int *right
,
947 __isl_keep isl_map
*map
)
953 for (j
= 0; j
< map
->n
&& ok
; ++j
) {
954 int overlaps
, subset
;
960 dom
[j
] = isl_set_from_basic_set(
961 isl_basic_map_domain(
962 isl_basic_map_copy(map
->p
[j
])));
965 overlaps
= isl_set_overlaps(ran
[i
], dom
[j
]);
971 subset
= isl_set_is_subset(dom
[j
], C
);
983 ran
[j
] = isl_set_from_basic_set(
985 isl_basic_map_copy(map
->p
[j
])));
988 overlaps
= isl_set_overlaps(dom
[i
], ran
[j
]);
994 subset
= isl_set_is_subset(ran
[j
], C
);
1008 /* Return a map that is a union of the basic maps in "map", except i,
1009 * composed to left and right with qc based on the entries of "left"
1012 static __isl_give isl_map
*compose(__isl_keep isl_map
*map
, int i
,
1013 __isl_take isl_map
*qc
, int *left
, int *right
)
1018 comp
= isl_map_empty(isl_map_get_dim(map
));
1019 for (j
= 0; j
< map
->n
; ++j
) {
1025 map_j
= isl_map_from_basic_map(isl_basic_map_copy(map
->p
[j
]));
1026 if (left
&& left
[j
])
1027 map_j
= isl_map_apply_range(map_j
, isl_map_copy(qc
));
1028 if (right
&& right
[j
])
1029 map_j
= isl_map_apply_range(isl_map_copy(qc
), map_j
);
1030 comp
= isl_map_union(comp
, map_j
);
1033 comp
= isl_map_compute_divs(comp
);
1034 comp
= isl_map_coalesce(comp
);
1041 /* Compute the transitive closure of "map" incrementally by
1048 * map_i^+ \cup ((id \cup map_i^) \circ qc^+)
1052 * map_i^+ \cup (qc^+ \circ (id \cup map_i^))
1054 * depending on whether left or right are NULL.
1056 static __isl_give isl_map
*compute_incremental(
1057 __isl_take isl_dim
*dim
, __isl_keep isl_map
*map
,
1058 int i
, __isl_take isl_map
*qc
, int *left
, int *right
, int *exact
)
1062 isl_map
*rtc
= NULL
;
1066 isl_assert(map
->ctx
, left
|| right
, goto error
);
1068 map_i
= isl_map_from_basic_map(isl_basic_map_copy(map
->p
[i
]));
1069 tc
= construct_projected_component(isl_dim_copy(dim
), map_i
,
1071 isl_map_free(map_i
);
1074 qc
= isl_map_transitive_closure(qc
, exact
);
1080 return isl_map_universe(isl_map_get_dim(map
));
1083 if (!left
|| !right
)
1084 rtc
= isl_map_union(isl_map_copy(tc
),
1085 isl_map_identity(isl_dim_domain(isl_map_get_dim(tc
))));
1087 qc
= isl_map_apply_range(rtc
, qc
);
1089 qc
= isl_map_apply_range(qc
, rtc
);
1090 qc
= isl_map_union(tc
, qc
);
1101 /* Given a map "map", try to find a basic map such that
1102 * map^+ can be computed as
1104 * map^+ = map_i^+ \cup
1105 * \bigcup_j ((map_i^+ \cup Id_C)^+ \circ map_j \circ (map_i^+ \cup Id_C))^+
1107 * with C the simple hull of the domain and range of the input map.
1108 * map_i^ \cup Id_C is computed by allowing the path lengths to be zero
1109 * and by intersecting domain and range with C.
1110 * Of course, we need to check that this is actually equal to map_i^ \cup Id_C.
1111 * Also, we only use the incremental computation if all the transitive
1112 * closures are exact and if the number of basic maps in the union,
1113 * after computing the integer divisions, is smaller than the number
1114 * of basic maps in the input map.
1116 static int incemental_on_entire_domain(__isl_keep isl_dim
*dim
,
1117 __isl_keep isl_map
*map
,
1118 isl_set
**dom
, isl_set
**ran
, int *left
, int *right
,
1119 __isl_give isl_map
**res
)
1127 C
= isl_set_union(isl_map_domain(isl_map_copy(map
)),
1128 isl_map_range(isl_map_copy(map
)));
1129 C
= isl_set_from_basic_set(isl_set_simple_hull(C
));
1137 d
= isl_map_dim(map
, isl_dim_in
);
1139 for (i
= 0; i
< map
->n
; ++i
) {
1141 int exact_i
, spurious
;
1143 dom
[i
] = isl_set_from_basic_set(isl_basic_map_domain(
1144 isl_basic_map_copy(map
->p
[i
])));
1145 ran
[i
] = isl_set_from_basic_set(isl_basic_map_range(
1146 isl_basic_map_copy(map
->p
[i
])));
1147 qc
= q_closure(isl_dim_copy(dim
), isl_set_copy(C
),
1148 map
->p
[i
], &exact_i
);
1155 spurious
= has_spurious_elements(qc
, dom
[i
], ran
[i
]);
1162 qc
= isl_map_project_out(qc
, isl_dim_in
, d
, 1);
1163 qc
= isl_map_project_out(qc
, isl_dim_out
, d
, 1);
1164 qc
= isl_map_compute_divs(qc
);
1165 for (j
= 0; j
< map
->n
; ++j
)
1166 left
[j
] = right
[j
] = 1;
1167 qc
= compose(map
, i
, qc
, left
, right
);
1170 if (qc
->n
>= map
->n
) {
1174 *res
= compute_incremental(isl_dim_copy(dim
), map
, i
, qc
,
1175 left
, right
, &exact_i
);
1186 return *res
!= NULL
;
1192 /* Try and compute the transitive closure of "map" as
1194 * map^+ = map_i^+ \cup
1195 * \bigcup_j ((map_i^+ \cup Id_C)^+ \circ map_j \circ (map_i^+ \cup Id_C))^+
1197 * with C either the simple hull of the domain and range of the entire
1198 * map or the simple hull of domain and range of map_i.
1200 static __isl_give isl_map
*incremental_closure(__isl_take isl_dim
*dim
,
1201 __isl_keep isl_map
*map
, int *exact
, int project
)
1204 isl_set
**dom
= NULL
;
1205 isl_set
**ran
= NULL
;
1210 isl_map
*res
= NULL
;
1213 return construct_projected_component(dim
, map
, exact
, project
);
1218 return construct_projected_component(dim
, map
, exact
, project
);
1220 d
= isl_map_dim(map
, isl_dim_in
);
1222 dom
= isl_calloc_array(map
->ctx
, isl_set
*, map
->n
);
1223 ran
= isl_calloc_array(map
->ctx
, isl_set
*, map
->n
);
1224 left
= isl_calloc_array(map
->ctx
, int, map
->n
);
1225 right
= isl_calloc_array(map
->ctx
, int, map
->n
);
1226 if (!ran
|| !dom
|| !left
|| !right
)
1229 if (incemental_on_entire_domain(dim
, map
, dom
, ran
, left
, right
, &res
) < 0)
1232 for (i
= 0; !res
&& i
< map
->n
; ++i
) {
1234 int exact_i
, spurious
, comp
;
1236 dom
[i
] = isl_set_from_basic_set(
1237 isl_basic_map_domain(
1238 isl_basic_map_copy(map
->p
[i
])));
1242 ran
[i
] = isl_set_from_basic_set(
1243 isl_basic_map_range(
1244 isl_basic_map_copy(map
->p
[i
])));
1247 C
= isl_set_union(isl_set_copy(dom
[i
]),
1248 isl_set_copy(ran
[i
]));
1249 C
= isl_set_from_basic_set(isl_set_simple_hull(C
));
1256 comp
= composability(C
, i
, dom
, ran
, left
, right
, map
);
1257 if (!comp
|| comp
< 0) {
1263 qc
= q_closure(isl_dim_copy(dim
), C
, map
->p
[i
], &exact_i
);
1270 spurious
= has_spurious_elements(qc
, dom
[i
], ran
[i
]);
1277 qc
= isl_map_project_out(qc
, isl_dim_in
, d
, 1);
1278 qc
= isl_map_project_out(qc
, isl_dim_out
, d
, 1);
1279 qc
= isl_map_compute_divs(qc
);
1280 qc
= compose(map
, i
, qc
, (comp
& LEFT
) ? left
: NULL
,
1281 (comp
& RIGHT
) ? right
: NULL
);
1284 if (qc
->n
>= map
->n
) {
1288 res
= compute_incremental(isl_dim_copy(dim
), map
, i
, qc
,
1289 (comp
& LEFT
) ? left
: NULL
,
1290 (comp
& RIGHT
) ? right
: NULL
, &exact_i
);
1299 for (i
= 0; i
< map
->n
; ++i
) {
1300 isl_set_free(dom
[i
]);
1301 isl_set_free(ran
[i
]);
1313 return construct_projected_component(dim
, map
, exact
, project
);
1316 for (i
= 0; i
< map
->n
; ++i
)
1317 isl_set_free(dom
[i
]);
1320 for (i
= 0; i
< map
->n
; ++i
)
1321 isl_set_free(ran
[i
]);
1329 /* Given an array of sets "set", add "dom" at position "pos"
1330 * and search for elements at earlier positions that overlap with "dom".
1331 * If any can be found, then merge all of them, together with "dom", into
1332 * a single set and assign the union to the first in the array,
1333 * which becomes the new group leader for all groups involved in the merge.
1334 * During the search, we only consider group leaders, i.e., those with
1335 * group[i] = i, as the other sets have already been combined
1336 * with one of the group leaders.
1338 static int merge(isl_set
**set
, int *group
, __isl_take isl_set
*dom
, int pos
)
1343 set
[pos
] = isl_set_copy(dom
);
1345 for (i
= pos
- 1; i
>= 0; --i
) {
1351 o
= isl_set_overlaps(set
[i
], dom
);
1357 set
[i
] = isl_set_union(set
[i
], set
[group
[pos
]]);
1360 set
[group
[pos
]] = NULL
;
1361 group
[group
[pos
]] = i
;
1372 /* Given a partition of the domains and ranges of the basic maps in "map",
1373 * apply the Floyd-Warshall algorithm with the elements in the partition
1376 * In particular, there are "n" elements in the partition and "group" is
1377 * an array of length 2 * map->n with entries in [0,n-1].
1379 * We first construct a matrix of relations based on the partition information,
1380 * apply Floyd-Warshall on this matrix of relations and then take the
1381 * union of all entries in the matrix as the final result.
1383 * The algorithm iterates over all vertices. In each step, the whole
1384 * matrix is updated to include all paths that go to the current vertex,
1385 * possibly stay there a while (including passing through earlier vertices)
1386 * and then come back. At the start of each iteration, the diagonal
1387 * element corresponding to the current vertex is replaced by its
1388 * transitive closure to account for all indirect paths that stay
1389 * in the current vertex.
1391 static __isl_give isl_map
*floyd_warshall_with_groups(__isl_take isl_dim
*dim
,
1392 __isl_keep isl_map
*map
, int *exact
, int project
, int *group
, int n
)
1396 isl_map
***grid
= NULL
;
1404 return incremental_closure(dim
, map
, exact
, project
);
1407 grid
= isl_calloc_array(map
->ctx
, isl_map
**, n
);
1410 for (i
= 0; i
< n
; ++i
) {
1411 grid
[i
] = isl_calloc_array(map
->ctx
, isl_map
*, n
);
1414 for (j
= 0; j
< n
; ++j
)
1415 grid
[i
][j
] = isl_map_empty(isl_map_get_dim(map
));
1418 for (k
= 0; k
< map
->n
; ++k
) {
1420 j
= group
[2 * k
+ 1];
1421 grid
[i
][j
] = isl_map_union(grid
[i
][j
],
1422 isl_map_from_basic_map(
1423 isl_basic_map_copy(map
->p
[k
])));
1426 for (r
= 0; r
< n
; ++r
) {
1428 grid
[r
][r
] = isl_map_transitive_closure(grid
[r
][r
],
1429 (exact
&& *exact
) ? &r_exact
: NULL
);
1430 if (exact
&& *exact
&& !r_exact
)
1433 for (p
= 0; p
< n
; ++p
)
1434 for (q
= 0; q
< n
; ++q
) {
1436 if (p
== r
&& q
== r
)
1438 loop
= isl_map_apply_range(
1439 isl_map_copy(grid
[p
][r
]),
1440 isl_map_copy(grid
[r
][q
]));
1441 grid
[p
][q
] = isl_map_union(grid
[p
][q
], loop
);
1442 loop
= isl_map_apply_range(
1443 isl_map_copy(grid
[p
][r
]),
1444 isl_map_apply_range(
1445 isl_map_copy(grid
[r
][r
]),
1446 isl_map_copy(grid
[r
][q
])));
1447 grid
[p
][q
] = isl_map_union(grid
[p
][q
], loop
);
1448 grid
[p
][q
] = isl_map_coalesce(grid
[p
][q
]);
1452 app
= isl_map_empty(isl_map_get_dim(map
));
1454 for (i
= 0; i
< n
; ++i
) {
1455 for (j
= 0; j
< n
; ++j
)
1456 app
= isl_map_union(app
, grid
[i
][j
]);
1467 for (i
= 0; i
< n
; ++i
) {
1470 for (j
= 0; j
< n
; ++j
)
1471 isl_map_free(grid
[i
][j
]);
1480 /* Check if the domains and ranges of the basic maps in "map" can
1481 * be partitioned, and if so, apply Floyd-Warshall on the elements
1482 * of the partition. Note that we can only apply this algorithm
1483 * if we want to compute the transitive closure, i.e., when "project"
1484 * is set. If we want to compute the power, we need to keep track
1485 * of the lengths and the recursive calls inside the Floyd-Warshall
1486 * would result in non-linear lengths.
1488 * To find the partition, we simply consider all of the domains
1489 * and ranges in turn and combine those that overlap.
1490 * "set" contains the partition elements and "group" indicates
1491 * to which partition element a given domain or range belongs.
1492 * The domain of basic map i corresponds to element 2 * i in these arrays,
1493 * while the domain corresponds to element 2 * i + 1.
1494 * During the construction group[k] is either equal to k,
1495 * in which case set[k] contains the union of all the domains and
1496 * ranges in the corresponding group, or is equal to some l < k,
1497 * with l another domain or range in the same group.
1499 static __isl_give isl_map
*floyd_warshall(__isl_take isl_dim
*dim
,
1500 __isl_keep isl_map
*map
, int *exact
, int project
)
1503 isl_set
**set
= NULL
;
1509 if (!project
|| map
->n
<= 1)
1510 return incremental_closure(dim
, map
, exact
, project
);
1512 set
= isl_calloc_array(map
->ctx
, isl_set
*, 2 * map
->n
);
1513 group
= isl_alloc_array(map
->ctx
, int, 2 * map
->n
);
1518 for (i
= 0; i
< map
->n
; ++i
) {
1520 dom
= isl_set_from_basic_set(isl_basic_map_domain(
1521 isl_basic_map_copy(map
->p
[i
])));
1522 if (merge(set
, group
, dom
, 2 * i
) < 0)
1524 dom
= isl_set_from_basic_set(isl_basic_map_range(
1525 isl_basic_map_copy(map
->p
[i
])));
1526 if (merge(set
, group
, dom
, 2 * i
+ 1) < 0)
1531 for (i
= 0; i
< 2 * map
->n
; ++i
)
1535 group
[i
] = group
[group
[i
]];
1537 for (i
= 0; i
< 2 * map
->n
; ++i
)
1538 isl_set_free(set
[i
]);
1542 return floyd_warshall_with_groups(dim
, map
, exact
, project
, group
, n
);
1544 for (i
= 0; i
< 2 * map
->n
; ++i
)
1545 isl_set_free(set
[i
]);
1552 /* Structure for representing the nodes in the graph being traversed
1553 * using Tarjan's algorithm.
1554 * index represents the order in which nodes are visited.
1555 * min_index is the index of the root of a (sub)component.
1556 * on_stack indicates whether the node is currently on the stack.
1558 struct basic_map_sort_node
{
1563 /* Structure for representing the graph being traversed
1564 * using Tarjan's algorithm.
1565 * len is the number of nodes
1566 * node is an array of nodes
1567 * stack contains the nodes on the path from the root to the current node
1568 * sp is the stack pointer
1569 * index is the index of the last node visited
1570 * order contains the elements of the components separated by -1
1571 * op represents the current position in order
1573 struct basic_map_sort
{
1575 struct basic_map_sort_node
*node
;
1583 static void basic_map_sort_free(struct basic_map_sort
*s
)
1593 static struct basic_map_sort
*basic_map_sort_alloc(struct isl_ctx
*ctx
, int len
)
1595 struct basic_map_sort
*s
;
1598 s
= isl_calloc_type(ctx
, struct basic_map_sort
);
1602 s
->node
= isl_alloc_array(ctx
, struct basic_map_sort_node
, len
);
1605 for (i
= 0; i
< len
; ++i
)
1606 s
->node
[i
].index
= -1;
1607 s
->stack
= isl_alloc_array(ctx
, int, len
);
1610 s
->order
= isl_alloc_array(ctx
, int, 2 * len
);
1620 basic_map_sort_free(s
);
1624 /* Check whether in the computation of the transitive closure
1625 * "bmap1" (R_1) should follow (or be part of the same component as)
1628 * That is check whether
1636 * If so, then there is no reason for R_1 to immediately follow R_2
1639 static int basic_map_follows(__isl_keep isl_basic_map
*bmap1
,
1640 __isl_keep isl_basic_map
*bmap2
)
1642 struct isl_map
*map12
= NULL
;
1643 struct isl_map
*map21
= NULL
;
1646 map21
= isl_map_from_basic_map(
1647 isl_basic_map_apply_range(
1648 isl_basic_map_copy(bmap2
),
1649 isl_basic_map_copy(bmap1
)));
1650 subset
= isl_map_is_empty(map21
);
1654 isl_map_free(map21
);
1658 map12
= isl_map_from_basic_map(
1659 isl_basic_map_apply_range(
1660 isl_basic_map_copy(bmap1
),
1661 isl_basic_map_copy(bmap2
)));
1663 subset
= isl_map_is_subset(map21
, map12
);
1665 isl_map_free(map12
);
1666 isl_map_free(map21
);
1668 return subset
< 0 ? -1 : !subset
;
1670 isl_map_free(map21
);
1674 /* Perform Tarjan's algorithm for computing the strongly connected components
1675 * in the graph with the disjuncts of "map" as vertices and with an
1676 * edge between any pair of disjuncts such that the first has
1677 * to be applied after the second.
1679 static int power_components_tarjan(struct basic_map_sort
*s
,
1680 __isl_keep isl_map
*map
, int i
)
1684 s
->node
[i
].index
= s
->index
;
1685 s
->node
[i
].min_index
= s
->index
;
1686 s
->node
[i
].on_stack
= 1;
1688 s
->stack
[s
->sp
++] = i
;
1690 for (j
= s
->len
- 1; j
>= 0; --j
) {
1695 if (s
->node
[j
].index
>= 0 &&
1696 (!s
->node
[j
].on_stack
||
1697 s
->node
[j
].index
> s
->node
[i
].min_index
))
1700 f
= basic_map_follows(map
->p
[i
], map
->p
[j
]);
1706 if (s
->node
[j
].index
< 0) {
1707 power_components_tarjan(s
, map
, j
);
1708 if (s
->node
[j
].min_index
< s
->node
[i
].min_index
)
1709 s
->node
[i
].min_index
= s
->node
[j
].min_index
;
1710 } else if (s
->node
[j
].index
< s
->node
[i
].min_index
)
1711 s
->node
[i
].min_index
= s
->node
[j
].index
;
1714 if (s
->node
[i
].index
!= s
->node
[i
].min_index
)
1718 j
= s
->stack
[--s
->sp
];
1719 s
->node
[j
].on_stack
= 0;
1720 s
->order
[s
->op
++] = j
;
1722 s
->order
[s
->op
++] = -1;
1727 /* Given a union of basic maps R = \cup_i R_i \subseteq D \times D
1728 * and a dimension specification (Z^{n+1} -> Z^{n+1}),
1729 * construct a map that is an overapproximation of the map
1730 * that takes an element from the dom R \times Z to an
1731 * element from ran R \times Z, such that the first n coordinates of the
1732 * difference between them is a sum of differences between images
1733 * and pre-images in one of the R_i and such that the last coordinate
1734 * is equal to the number of steps taken.
1735 * If "project" is set, then these final coordinates are not included,
1736 * i.e., a relation of type Z^n -> Z^n is returned.
1739 * \Delta_i = { y - x | (x, y) in R_i }
1741 * then the constructed map is an overapproximation of
1743 * { (x) -> (x + d) | \exists k_i >= 0, \delta_i \in \Delta_i :
1744 * d = (\sum_i k_i \delta_i, \sum_i k_i) and
1745 * x in dom R and x + d in ran R }
1749 * { (x) -> (x + d) | \exists k_i >= 0, \delta_i \in \Delta_i :
1750 * d = (\sum_i k_i \delta_i) and
1751 * x in dom R and x + d in ran R }
1753 * if "project" is set.
1755 * We first split the map into strongly connected components, perform
1756 * the above on each component and then join the results in the correct
1757 * order, at each join also taking in the union of both arguments
1758 * to allow for paths that do not go through one of the two arguments.
1760 static __isl_give isl_map
*construct_power_components(__isl_take isl_dim
*dim
,
1761 __isl_keep isl_map
*map
, int *exact
, int project
)
1764 struct isl_map
*path
= NULL
;
1765 struct basic_map_sort
*s
= NULL
;
1770 return floyd_warshall(dim
, map
, exact
, project
);
1772 s
= basic_map_sort_alloc(map
->ctx
, map
->n
);
1775 for (i
= map
->n
- 1; i
>= 0; --i
) {
1776 if (s
->node
[i
].index
>= 0)
1778 if (power_components_tarjan(s
, map
, i
) < 0)
1785 path
= isl_map_empty(isl_map_get_dim(map
));
1787 path
= isl_map_empty(isl_dim_copy(dim
));
1789 struct isl_map
*comp
;
1790 isl_map
*path_comp
, *path_comb
;
1791 comp
= isl_map_alloc_dim(isl_map_get_dim(map
), n
, 0);
1792 while (s
->order
[i
] != -1) {
1793 comp
= isl_map_add_basic_map(comp
,
1794 isl_basic_map_copy(map
->p
[s
->order
[i
]]));
1798 path_comp
= floyd_warshall(isl_dim_copy(dim
),
1799 comp
, exact
, project
);
1800 path_comb
= isl_map_apply_range(isl_map_copy(path
),
1801 isl_map_copy(path_comp
));
1802 path
= isl_map_union(path
, path_comp
);
1803 path
= isl_map_union(path
, path_comb
);
1808 basic_map_sort_free(s
);
1813 basic_map_sort_free(s
);
1818 /* Given a union of basic maps R = \cup_i R_i \subseteq D \times D,
1819 * construct a map that is an overapproximation of the map
1820 * that takes an element from the space D to another
1821 * element from the same space, such that the difference between
1822 * them is a strictly positive sum of differences between images
1823 * and pre-images in one of the R_i.
1824 * The number of differences in the sum is equated to parameter "param".
1827 * \Delta_i = { y - x | (x, y) in R_i }
1829 * then the constructed map is an overapproximation of
1831 * { (x) -> (x + d) | \exists k_i >= 0, \delta_i \in \Delta_i :
1832 * d = \sum_i k_i \delta_i and k = \sum_i k_i > 0 }
1835 * { (x) -> (x + d) | \exists k_i >= 0, \delta_i \in \Delta_i :
1836 * d = \sum_i k_i \delta_i and \sum_i k_i > 0 }
1838 * if "project" is set.
1840 * If "project" is not set, then
1841 * we first construct an extended mapping with an extra coordinate
1842 * that indicates the number of steps taken. In particular,
1843 * the difference in the last coordinate is equal to the number
1844 * of steps taken to move from a domain element to the corresponding
1846 * In the final step, this difference is equated to the parameter "param"
1847 * and made positive. The extra coordinates are subsequently projected out.
1849 static __isl_give isl_map
*construct_power(__isl_keep isl_map
*map
,
1850 unsigned param
, int *exact
, int project
)
1852 struct isl_map
*app
= NULL
;
1853 struct isl_map
*diff
;
1854 struct isl_dim
*dim
= NULL
;
1860 dim
= isl_map_get_dim(map
);
1862 d
= isl_dim_size(dim
, isl_dim_in
);
1863 dim
= isl_dim_add(dim
, isl_dim_in
, 1);
1864 dim
= isl_dim_add(dim
, isl_dim_out
, 1);
1866 app
= construct_power_components(isl_dim_copy(dim
), map
,
1872 diff
= equate_parameter_to_length(dim
, param
);
1873 app
= isl_map_intersect(app
, diff
);
1874 app
= isl_map_project_out(app
, isl_dim_in
, d
, 1);
1875 app
= isl_map_project_out(app
, isl_dim_out
, d
, 1);
1881 /* Compute the positive powers of "map", or an overapproximation.
1882 * The power is given by parameter "param". If the result is exact,
1883 * then *exact is set to 1.
1885 * If project is set, then we are actually interested in the transitive
1886 * closure, so we can use a more relaxed exactness check.
1887 * The lengths of the paths are also projected out instead of being
1888 * equated to "param" (which is then ignored in this case).
1890 static __isl_give isl_map
*map_power(__isl_take isl_map
*map
, unsigned param
,
1891 int *exact
, int project
)
1893 struct isl_map
*app
= NULL
;
1901 if (isl_map_fast_is_empty(map
))
1904 isl_assert(map
->ctx
, project
|| param
< isl_map_dim(map
, isl_dim_param
),
1906 isl_assert(map
->ctx
,
1907 isl_map_dim(map
, isl_dim_in
) == isl_map_dim(map
, isl_dim_out
),
1910 app
= construct_power(map
, param
, exact
, project
);
1920 /* Compute the positive powers of "map", or an overapproximation.
1921 * The power is given by parameter "param". If the result is exact,
1922 * then *exact is set to 1.
1924 __isl_give isl_map
*isl_map_power(__isl_take isl_map
*map
, unsigned param
,
1927 map
= isl_map_compute_divs(map
);
1928 map
= isl_map_coalesce(map
);
1929 return map_power(map
, param
, exact
, 0);
1932 /* Check whether equality i of bset is a pure stride constraint
1933 * on a single dimensions, i.e., of the form
1937 * with k a constant and e an existentially quantified variable.
1939 static int is_eq_stride(__isl_keep isl_basic_set
*bset
, int i
)
1951 if (!isl_int_is_zero(bset
->eq
[i
][0]))
1954 nparam
= isl_basic_set_dim(bset
, isl_dim_param
);
1955 d
= isl_basic_set_dim(bset
, isl_dim_set
);
1956 n_div
= isl_basic_set_dim(bset
, isl_dim_div
);
1958 if (isl_seq_first_non_zero(bset
->eq
[i
] + 1, nparam
) != -1)
1960 pos1
= isl_seq_first_non_zero(bset
->eq
[i
] + 1 + nparam
, d
);
1963 if (isl_seq_first_non_zero(bset
->eq
[i
] + 1 + nparam
+ pos1
+ 1,
1964 d
- pos1
- 1) != -1)
1967 pos2
= isl_seq_first_non_zero(bset
->eq
[i
] + 1 + nparam
+ d
, n_div
);
1970 if (isl_seq_first_non_zero(bset
->eq
[i
] + 1 + nparam
+ d
+ pos2
+ 1,
1971 n_div
- pos2
- 1) != -1)
1973 if (!isl_int_is_one(bset
->eq
[i
][1 + nparam
+ pos1
]) &&
1974 !isl_int_is_negone(bset
->eq
[i
][1 + nparam
+ pos1
]))
1980 /* Given a map, compute the smallest superset of this map that is of the form
1982 * { i -> j : L <= j - i <= U and exists a_p: j_p - i_p = M_p a_p }
1984 * (where p ranges over the (non-parametric) dimensions),
1985 * compute the transitive closure of this map, i.e.,
1987 * { i -> j : exists k > 0:
1988 * k L <= j - i <= k U and exists a: j_p - i_p = M_p a_p }
1990 * and intersect domain and range of this transitive closure with
1991 * the given domain and range.
1993 * If with_id is set, then try to include as much of the identity mapping
1994 * as possible, by computing
1996 * { i -> j : exists k >= 0:
1997 * k L <= j - i <= k U and exists a: j_p - i_p = M_p a_p }
1999 * instead (i.e., allow k = 0).
2001 * In practice, we compute the difference set
2003 * delta = { j - i | i -> j in map },
2005 * look for stride constraint on the individual dimensions and compute
2006 * (constant) lower and upper bounds for each individual dimension,
2007 * adding a constraint for each bound not equal to infinity.
2009 static __isl_give isl_map
*box_closure_on_domain(__isl_take isl_map
*map
,
2010 __isl_take isl_set
*dom
, __isl_take isl_set
*ran
, int with_id
)
2019 isl_map
*app
= NULL
;
2020 isl_basic_set
*aff
= NULL
;
2021 isl_basic_map
*bmap
= NULL
;
2022 isl_vec
*obj
= NULL
;
2027 delta
= isl_map_deltas(isl_map_copy(map
));
2029 aff
= isl_set_affine_hull(isl_set_copy(delta
));
2032 dim
= isl_map_get_dim(map
);
2033 d
= isl_dim_size(dim
, isl_dim_in
);
2034 nparam
= isl_dim_size(dim
, isl_dim_param
);
2035 total
= isl_dim_total(dim
);
2036 bmap
= isl_basic_map_alloc_dim(dim
,
2037 aff
->n_div
+ 1, aff
->n_div
, 2 * d
+ 1);
2038 for (i
= 0; i
< aff
->n_div
+ 1; ++i
) {
2039 k
= isl_basic_map_alloc_div(bmap
);
2042 isl_int_set_si(bmap
->div
[k
][0], 0);
2044 for (i
= 0; i
< aff
->n_eq
; ++i
) {
2045 if (!is_eq_stride(aff
, i
))
2047 k
= isl_basic_map_alloc_equality(bmap
);
2050 isl_seq_clr(bmap
->eq
[k
], 1 + nparam
);
2051 isl_seq_cpy(bmap
->eq
[k
] + 1 + nparam
+ d
,
2052 aff
->eq
[i
] + 1 + nparam
, d
);
2053 isl_seq_neg(bmap
->eq
[k
] + 1 + nparam
,
2054 aff
->eq
[i
] + 1 + nparam
, d
);
2055 isl_seq_cpy(bmap
->eq
[k
] + 1 + nparam
+ 2 * d
,
2056 aff
->eq
[i
] + 1 + nparam
+ d
, aff
->n_div
);
2057 isl_int_set_si(bmap
->eq
[k
][1 + total
+ aff
->n_div
], 0);
2059 obj
= isl_vec_alloc(map
->ctx
, 1 + nparam
+ d
);
2062 isl_seq_clr(obj
->el
, 1 + nparam
+ d
);
2063 for (i
= 0; i
< d
; ++ i
) {
2064 enum isl_lp_result res
;
2066 isl_int_set_si(obj
->el
[1 + nparam
+ i
], 1);
2068 res
= isl_set_solve_lp(delta
, 0, obj
->el
, map
->ctx
->one
, &opt
,
2070 if (res
== isl_lp_error
)
2072 if (res
== isl_lp_ok
) {
2073 k
= isl_basic_map_alloc_inequality(bmap
);
2076 isl_seq_clr(bmap
->ineq
[k
],
2077 1 + nparam
+ 2 * d
+ bmap
->n_div
);
2078 isl_int_set_si(bmap
->ineq
[k
][1 + nparam
+ i
], -1);
2079 isl_int_set_si(bmap
->ineq
[k
][1 + nparam
+ d
+ i
], 1);
2080 isl_int_neg(bmap
->ineq
[k
][1 + nparam
+ 2 * d
+ aff
->n_div
], opt
);
2083 res
= isl_set_solve_lp(delta
, 1, obj
->el
, map
->ctx
->one
, &opt
,
2085 if (res
== isl_lp_error
)
2087 if (res
== isl_lp_ok
) {
2088 k
= isl_basic_map_alloc_inequality(bmap
);
2091 isl_seq_clr(bmap
->ineq
[k
],
2092 1 + nparam
+ 2 * d
+ bmap
->n_div
);
2093 isl_int_set_si(bmap
->ineq
[k
][1 + nparam
+ i
], 1);
2094 isl_int_set_si(bmap
->ineq
[k
][1 + nparam
+ d
+ i
], -1);
2095 isl_int_set(bmap
->ineq
[k
][1 + nparam
+ 2 * d
+ aff
->n_div
], opt
);
2098 isl_int_set_si(obj
->el
[1 + nparam
+ i
], 0);
2100 k
= isl_basic_map_alloc_inequality(bmap
);
2103 isl_seq_clr(bmap
->ineq
[k
],
2104 1 + nparam
+ 2 * d
+ bmap
->n_div
);
2106 isl_int_set_si(bmap
->ineq
[k
][0], -1);
2107 isl_int_set_si(bmap
->ineq
[k
][1 + nparam
+ 2 * d
+ aff
->n_div
], 1);
2109 app
= isl_map_from_domain_and_range(dom
, ran
);
2112 isl_basic_set_free(aff
);
2114 bmap
= isl_basic_map_finalize(bmap
);
2115 isl_set_free(delta
);
2118 map
= isl_map_from_basic_map(bmap
);
2119 map
= isl_map_intersect(map
, app
);
2124 isl_basic_map_free(bmap
);
2125 isl_basic_set_free(aff
);
2129 isl_set_free(delta
);
2134 /* Given a map, compute the smallest superset of this map that is of the form
2136 * { i -> j : L <= j - i <= U and exists a_p: j_p - i_p = M_p a_p }
2138 * (where p ranges over the (non-parametric) dimensions),
2139 * compute the transitive closure of this map, i.e.,
2141 * { i -> j : exists k > 0:
2142 * k L <= j - i <= k U and exists a: j_p - i_p = M_p a_p }
2144 * and intersect domain and range of this transitive closure with
2145 * domain and range of the original map.
2147 static __isl_give isl_map
*box_closure(__isl_take isl_map
*map
)
2152 domain
= isl_map_domain(isl_map_copy(map
));
2153 domain
= isl_set_coalesce(domain
);
2154 range
= isl_map_range(isl_map_copy(map
));
2155 range
= isl_set_coalesce(range
);
2157 return box_closure_on_domain(map
, domain
, range
, 0);
2160 /* Given a map, compute the smallest superset of this map that is of the form
2162 * { i -> j : L <= j - i <= U and exists a_p: j_p - i_p = M_p a_p }
2164 * (where p ranges over the (non-parametric) dimensions),
2165 * compute the transitive and partially reflexive closure of this map, i.e.,
2167 * { i -> j : exists k >= 0:
2168 * k L <= j - i <= k U and exists a: j_p - i_p = M_p a_p }
2170 * and intersect domain and range of this transitive closure with
2173 static __isl_give isl_map
*box_closure_with_identity(__isl_take isl_map
*map
,
2174 __isl_take isl_set
*dom
)
2176 return box_closure_on_domain(map
, dom
, isl_set_copy(dom
), 1);
2179 /* Check whether app is the transitive closure of map.
2180 * In particular, check that app is acyclic and, if so,
2183 * app \subset (map \cup (map \circ app))
2185 static int check_exactness_omega(__isl_keep isl_map
*map
,
2186 __isl_keep isl_map
*app
)
2190 int is_empty
, is_exact
;
2194 delta
= isl_map_deltas(isl_map_copy(app
));
2195 d
= isl_set_dim(delta
, isl_dim_set
);
2196 for (i
= 0; i
< d
; ++i
)
2197 delta
= isl_set_fix_si(delta
, isl_dim_set
, i
, 0);
2198 is_empty
= isl_set_is_empty(delta
);
2199 isl_set_free(delta
);
2205 test
= isl_map_apply_range(isl_map_copy(app
), isl_map_copy(map
));
2206 test
= isl_map_union(test
, isl_map_copy(map
));
2207 is_exact
= isl_map_is_subset(app
, test
);
2213 /* Check if basic map M_i can be combined with all the other
2214 * basic maps such that
2218 * can be computed as
2220 * M_i \cup (\cup_{j \ne i} M_i^* \circ M_j \circ M_i^*)^+
2222 * In particular, check if we can compute a compact representation
2225 * M_i^* \circ M_j \circ M_i^*
2228 * Let M_i^? be an extension of M_i^+ that allows paths
2229 * of length zero, i.e., the result of box_closure(., 1).
2230 * The criterion, as proposed by Kelly et al., is that
2231 * id = M_i^? - M_i^+ can be represented as a basic map
2234 * id \circ M_j \circ id = M_j
2238 * If this function returns 1, then tc and qc are set to
2239 * M_i^+ and M_i^?, respectively.
2241 static int can_be_split_off(__isl_keep isl_map
*map
, int i
,
2242 __isl_give isl_map
**tc
, __isl_give isl_map
**qc
)
2244 isl_map
*map_i
, *id
= NULL
;
2251 C
= isl_set_union(isl_map_domain(isl_map_copy(map
)),
2252 isl_map_range(isl_map_copy(map
)));
2253 C
= isl_set_from_basic_set(isl_set_simple_hull(C
));
2257 map_i
= isl_map_from_basic_map(isl_basic_map_copy(map
->p
[i
]));
2258 *tc
= box_closure(isl_map_copy(map_i
));
2259 *qc
= box_closure_with_identity(map_i
, C
);
2260 id
= isl_map_subtract(isl_map_copy(*qc
), isl_map_copy(*tc
));
2264 if (id
->n
!= 1 || (*qc
)->n
!= 1)
2267 for (j
= 0; j
< map
->n
; ++j
) {
2268 isl_map
*map_j
, *test
;
2273 map_j
= isl_map_from_basic_map(
2274 isl_basic_map_copy(map
->p
[j
]));
2275 test
= isl_map_apply_range(isl_map_copy(id
),
2276 isl_map_copy(map_j
));
2277 test
= isl_map_apply_range(test
, isl_map_copy(id
));
2278 is_ok
= isl_map_is_equal(test
, map_j
);
2279 isl_map_free(map_j
);
2307 static __isl_give isl_map
*box_closure_with_check(__isl_take isl_map
*map
,
2312 app
= box_closure(isl_map_copy(map
));
2314 *exact
= check_exactness_omega(map
, app
);
2320 /* Compute an overapproximation of the transitive closure of "map"
2321 * using a variation of the algorithm from
2322 * "Transitive Closure of Infinite Graphs and its Applications"
2325 * We first check whether we can can split of any basic map M_i and
2332 * M_i \cup (\cup_{j \ne i} M_i^* \circ M_j \circ M_i^*)^+
2334 * using a recursive call on the remaining map.
2336 * If not, we simply call box_closure on the whole map.
2338 static __isl_give isl_map
*transitive_closure_omega(__isl_take isl_map
*map
,
2348 return box_closure_with_check(map
, exact
);
2350 for (i
= 0; i
< map
->n
; ++i
) {
2353 ok
= can_be_split_off(map
, i
, &tc
, &qc
);
2359 app
= isl_map_alloc_dim(isl_map_get_dim(map
), map
->n
- 1, 0);
2361 for (j
= 0; j
< map
->n
; ++j
) {
2364 app
= isl_map_add_basic_map(app
,
2365 isl_basic_map_copy(map
->p
[j
]));
2368 app
= isl_map_apply_range(isl_map_copy(qc
), app
);
2369 app
= isl_map_apply_range(app
, qc
);
2371 app
= isl_map_union(tc
, transitive_closure_omega(app
, NULL
));
2372 exact_i
= check_exactness_omega(map
, app
);
2384 return box_closure_with_check(map
, exact
);
2390 int isl_map_is_transitively_closed(__isl_keep isl_map
*map
)
2395 map2
= isl_map_apply_range(isl_map_copy(map
), isl_map_copy(map
));
2396 closed
= isl_map_is_subset(map2
, map
);
2402 /* Compute the transitive closure of "map", or an overapproximation.
2403 * If the result is exact, then *exact is set to 1.
2404 * Simply use map_power to compute the powers of map, but tell
2405 * it to project out the lengths of the paths instead of equating
2406 * the length to a parameter.
2408 __isl_give isl_map
*isl_map_transitive_closure(__isl_take isl_map
*map
,
2417 if (map
->ctx
->opt
->closure
== ISL_CLOSURE_OMEGA
)
2418 return transitive_closure_omega(map
, exact
);
2420 map
= isl_map_compute_divs(map
);
2421 map
= isl_map_coalesce(map
);
2422 closed
= isl_map_is_transitively_closed(map
);
2431 param
= isl_map_dim(map
, isl_dim_param
);
2432 map
= map_power(map
, param
, exact
, 1);