2 * Copyright 2008-2009 Katholieke Universiteit Leuven
3 * Copyright 2010 INRIA Saclay
4 * Copyright 2012-2013 Ecole Normale Superieure
5 * Copyright 2014 INRIA Rocquencourt
6 * Copyright 2016 INRIA Paris
8 * Use of this software is governed by the MIT license
10 * Written by Sven Verdoolaege, K.U.Leuven, Departement
11 * Computerwetenschappen, Celestijnenlaan 200A, B-3001 Leuven, Belgium
12 * and INRIA Saclay - Ile-de-France, Parc Club Orsay Universite,
13 * ZAC des vignes, 4 rue Jacques Monod, 91893 Orsay, France
14 * and Ecole Normale Superieure, 45 rue d’Ulm, 75230 Paris, France
15 * and Inria Paris - Rocquencourt, Domaine de Voluceau - Rocquencourt,
16 * B.P. 105 - 78153 Le Chesnay, France
17 * and Centre de Recherche Inria de Paris, 2 rue Simone Iff - Voie DQ12,
18 * CS 42112, 75589 Paris Cedex 12, France
21 #include <isl_ctx_private.h>
22 #include "isl_map_private.h"
24 #include <isl/options.h>
26 #include <isl_mat_private.h>
27 #include <isl_local_space_private.h>
28 #include <isl_val_private.h>
29 #include <isl_vec_private.h>
30 #include <isl_aff_private.h>
31 #include <isl_equalities.h>
32 #include <isl_constraint_private.h>
34 #include <set_to_map.c>
35 #include <set_from_map.c>
37 #define STATUS_ERROR -1
38 #define STATUS_REDUNDANT 1
39 #define STATUS_VALID 2
40 #define STATUS_SEPARATE 3
42 #define STATUS_ADJ_EQ 5
43 #define STATUS_ADJ_INEQ 6
45 static int status_in(isl_int
*ineq
, struct isl_tab
*tab
)
47 enum isl_ineq_type type
= isl_tab_ineq_type(tab
, ineq
);
50 case isl_ineq_error
: return STATUS_ERROR
;
51 case isl_ineq_redundant
: return STATUS_VALID
;
52 case isl_ineq_separate
: return STATUS_SEPARATE
;
53 case isl_ineq_cut
: return STATUS_CUT
;
54 case isl_ineq_adj_eq
: return STATUS_ADJ_EQ
;
55 case isl_ineq_adj_ineq
: return STATUS_ADJ_INEQ
;
59 /* Compute the position of the equalities of basic map "bmap_i"
60 * with respect to the basic map represented by "tab_j".
61 * The resulting array has twice as many entries as the number
62 * of equalities corresponding to the two inequalities to which
63 * each equality corresponds.
65 static int *eq_status_in(__isl_keep isl_basic_map
*bmap_i
,
66 struct isl_tab
*tab_j
)
72 dim
= isl_basic_map_dim(bmap_i
, isl_dim_all
);
76 eq
= isl_calloc_array(bmap_i
->ctx
, int, 2 * bmap_i
->n_eq
);
80 for (k
= 0; k
< bmap_i
->n_eq
; ++k
) {
81 for (l
= 0; l
< 2; ++l
) {
82 isl_seq_neg(bmap_i
->eq
[k
], bmap_i
->eq
[k
], 1+dim
);
83 eq
[2 * k
+ l
] = status_in(bmap_i
->eq
[k
], tab_j
);
84 if (eq
[2 * k
+ l
] == STATUS_ERROR
)
95 /* Compute the position of the inequalities of basic map "bmap_i"
96 * (also represented by "tab_i", if not NULL) with respect to the basic map
97 * represented by "tab_j".
99 static int *ineq_status_in(__isl_keep isl_basic_map
*bmap_i
,
100 struct isl_tab
*tab_i
, struct isl_tab
*tab_j
)
103 unsigned n_eq
= bmap_i
->n_eq
;
104 int *ineq
= isl_calloc_array(bmap_i
->ctx
, int, bmap_i
->n_ineq
);
109 for (k
= 0; k
< bmap_i
->n_ineq
; ++k
) {
110 if (tab_i
&& isl_tab_is_redundant(tab_i
, n_eq
+ k
)) {
111 ineq
[k
] = STATUS_REDUNDANT
;
114 ineq
[k
] = status_in(bmap_i
->ineq
[k
], tab_j
);
115 if (ineq
[k
] == STATUS_ERROR
)
117 if (ineq
[k
] == STATUS_SEPARATE
)
127 static int any(int *con
, unsigned len
, int status
)
131 for (i
= 0; i
< len
; ++i
)
132 if (con
[i
] == status
)
137 /* Return the first position of "status" in the list "con" of length "len".
138 * Return -1 if there is no such entry.
140 static int find(int *con
, unsigned len
, int status
)
144 for (i
= 0; i
< len
; ++i
)
145 if (con
[i
] == status
)
150 static int count(int *con
, unsigned len
, int status
)
155 for (i
= 0; i
< len
; ++i
)
156 if (con
[i
] == status
)
161 static int all(int *con
, unsigned len
, int status
)
165 for (i
= 0; i
< len
; ++i
) {
166 if (con
[i
] == STATUS_REDUNDANT
)
168 if (con
[i
] != status
)
174 /* Internal information associated to a basic map in a map
175 * that is to be coalesced by isl_map_coalesce.
177 * "bmap" is the basic map itself (or NULL if "removed" is set)
178 * "tab" is the corresponding tableau (or NULL if "removed" is set)
179 * "hull_hash" identifies the affine space in which "bmap" lives.
180 * "modified" is set if this basic map may not be identical
181 * to any of the basic maps in the input.
182 * "removed" is set if this basic map has been removed from the map
183 * "simplify" is set if this basic map may have some unknown integer
184 * divisions that were not present in the input basic maps. The basic
185 * map should then be simplified such that we may be able to find
186 * a definition among the constraints.
188 * "eq" and "ineq" are only set if we are currently trying to coalesce
189 * this basic map with another basic map, in which case they represent
190 * the position of the inequalities of this basic map with respect to
191 * the other basic map. The number of elements in the "eq" array
192 * is twice the number of equalities in the "bmap", corresponding
193 * to the two inequalities that make up each equality.
195 struct isl_coalesce_info
{
206 /* Is there any (half of an) equality constraint in the description
207 * of the basic map represented by "info" that
208 * has position "status" with respect to the other basic map?
210 static int any_eq(struct isl_coalesce_info
*info
, int status
)
214 n_eq
= isl_basic_map_n_equality(info
->bmap
);
215 return any(info
->eq
, 2 * n_eq
, status
);
218 /* Is there any inequality constraint in the description
219 * of the basic map represented by "info" that
220 * has position "status" with respect to the other basic map?
222 static int any_ineq(struct isl_coalesce_info
*info
, int status
)
226 n_ineq
= isl_basic_map_n_inequality(info
->bmap
);
227 return any(info
->ineq
, n_ineq
, status
);
230 /* Return the position of the first half on an equality constraint
231 * in the description of the basic map represented by "info" that
232 * has position "status" with respect to the other basic map.
233 * The returned value is twice the position of the equality constraint
234 * plus zero for the negative half and plus one for the positive half.
235 * Return -1 if there is no such entry.
237 static int find_eq(struct isl_coalesce_info
*info
, int status
)
241 n_eq
= isl_basic_map_n_equality(info
->bmap
);
242 return find(info
->eq
, 2 * n_eq
, status
);
245 /* Return the position of the first inequality constraint in the description
246 * of the basic map represented by "info" that
247 * has position "status" with respect to the other basic map.
248 * Return -1 if there is no such entry.
250 static int find_ineq(struct isl_coalesce_info
*info
, int status
)
254 n_ineq
= isl_basic_map_n_inequality(info
->bmap
);
255 return find(info
->ineq
, n_ineq
, status
);
258 /* Return the number of (halves of) equality constraints in the description
259 * of the basic map represented by "info" that
260 * have position "status" with respect to the other basic map.
262 static int count_eq(struct isl_coalesce_info
*info
, int status
)
266 n_eq
= isl_basic_map_n_equality(info
->bmap
);
267 return count(info
->eq
, 2 * n_eq
, status
);
270 /* Return the number of inequality constraints in the description
271 * of the basic map represented by "info" that
272 * have position "status" with respect to the other basic map.
274 static int count_ineq(struct isl_coalesce_info
*info
, int status
)
278 n_ineq
= isl_basic_map_n_inequality(info
->bmap
);
279 return count(info
->ineq
, n_ineq
, status
);
282 /* Are all non-redundant constraints of the basic map represented by "info"
283 * either valid or cut constraints with respect to the other basic map?
285 static int all_valid_or_cut(struct isl_coalesce_info
*info
)
289 for (i
= 0; i
< 2 * info
->bmap
->n_eq
; ++i
) {
290 if (info
->eq
[i
] == STATUS_REDUNDANT
)
292 if (info
->eq
[i
] == STATUS_VALID
)
294 if (info
->eq
[i
] == STATUS_CUT
)
299 for (i
= 0; i
< info
->bmap
->n_ineq
; ++i
) {
300 if (info
->ineq
[i
] == STATUS_REDUNDANT
)
302 if (info
->ineq
[i
] == STATUS_VALID
)
304 if (info
->ineq
[i
] == STATUS_CUT
)
312 /* Compute the hash of the (apparent) affine hull of info->bmap (with
313 * the existentially quantified variables removed) and store it
316 static int coalesce_info_set_hull_hash(struct isl_coalesce_info
*info
)
321 hull
= isl_basic_map_copy(info
->bmap
);
322 hull
= isl_basic_map_plain_affine_hull(hull
);
323 n_div
= isl_basic_map_dim(hull
, isl_dim_div
);
325 hull
= isl_basic_map_free(hull
);
326 hull
= isl_basic_map_drop_constraints_involving_dims(hull
,
327 isl_dim_div
, 0, n_div
);
328 info
->hull_hash
= isl_basic_map_get_hash(hull
);
329 isl_basic_map_free(hull
);
331 return hull
? 0 : -1;
334 /* Free all the allocated memory in an array
335 * of "n" isl_coalesce_info elements.
337 static void clear_coalesce_info(int n
, struct isl_coalesce_info
*info
)
344 for (i
= 0; i
< n
; ++i
) {
345 isl_basic_map_free(info
[i
].bmap
);
346 isl_tab_free(info
[i
].tab
);
352 /* Clear the memory associated to "info".
354 static void clear(struct isl_coalesce_info
*info
)
356 info
->bmap
= isl_basic_map_free(info
->bmap
);
357 isl_tab_free(info
->tab
);
361 /* Drop the basic map represented by "info".
362 * That is, clear the memory associated to the entry and
363 * mark it as having been removed.
365 static void drop(struct isl_coalesce_info
*info
)
371 /* Exchange the information in "info1" with that in "info2".
373 static void exchange(struct isl_coalesce_info
*info1
,
374 struct isl_coalesce_info
*info2
)
376 struct isl_coalesce_info info
;
383 /* This type represents the kind of change that has been performed
384 * while trying to coalesce two basic maps.
386 * isl_change_none: nothing was changed
387 * isl_change_drop_first: the first basic map was removed
388 * isl_change_drop_second: the second basic map was removed
389 * isl_change_fuse: the two basic maps were replaced by a new basic map.
392 isl_change_error
= -1,
394 isl_change_drop_first
,
395 isl_change_drop_second
,
399 /* Update "change" based on an interchange of the first and the second
400 * basic map. That is, interchange isl_change_drop_first and
401 * isl_change_drop_second.
403 static enum isl_change
invert_change(enum isl_change change
)
406 case isl_change_error
:
407 return isl_change_error
;
408 case isl_change_none
:
409 return isl_change_none
;
410 case isl_change_drop_first
:
411 return isl_change_drop_second
;
412 case isl_change_drop_second
:
413 return isl_change_drop_first
;
414 case isl_change_fuse
:
415 return isl_change_fuse
;
418 return isl_change_error
;
421 /* Add the valid constraints of the basic map represented by "info"
422 * to "bmap". "len" is the size of the constraints.
423 * If only one of the pair of inequalities that make up an equality
424 * is valid, then add that inequality.
426 static __isl_give isl_basic_map
*add_valid_constraints(
427 __isl_take isl_basic_map
*bmap
, struct isl_coalesce_info
*info
,
435 for (k
= 0; k
< info
->bmap
->n_eq
; ++k
) {
436 if (info
->eq
[2 * k
] == STATUS_VALID
&&
437 info
->eq
[2 * k
+ 1] == STATUS_VALID
) {
438 l
= isl_basic_map_alloc_equality(bmap
);
440 return isl_basic_map_free(bmap
);
441 isl_seq_cpy(bmap
->eq
[l
], info
->bmap
->eq
[k
], len
);
442 } else if (info
->eq
[2 * k
] == STATUS_VALID
) {
443 l
= isl_basic_map_alloc_inequality(bmap
);
445 return isl_basic_map_free(bmap
);
446 isl_seq_neg(bmap
->ineq
[l
], info
->bmap
->eq
[k
], len
);
447 } else if (info
->eq
[2 * k
+ 1] == STATUS_VALID
) {
448 l
= isl_basic_map_alloc_inequality(bmap
);
450 return isl_basic_map_free(bmap
);
451 isl_seq_cpy(bmap
->ineq
[l
], info
->bmap
->eq
[k
], len
);
455 for (k
= 0; k
< info
->bmap
->n_ineq
; ++k
) {
456 if (info
->ineq
[k
] != STATUS_VALID
)
458 l
= isl_basic_map_alloc_inequality(bmap
);
460 return isl_basic_map_free(bmap
);
461 isl_seq_cpy(bmap
->ineq
[l
], info
->bmap
->ineq
[k
], len
);
467 /* Is "bmap" defined by a number of (non-redundant) constraints that
468 * is greater than the number of constraints of basic maps i and j combined?
469 * Equalities are counted as two inequalities.
471 static int number_of_constraints_increases(int i
, int j
,
472 struct isl_coalesce_info
*info
,
473 __isl_keep isl_basic_map
*bmap
, struct isl_tab
*tab
)
477 n_old
= 2 * info
[i
].bmap
->n_eq
+ info
[i
].bmap
->n_ineq
;
478 n_old
+= 2 * info
[j
].bmap
->n_eq
+ info
[j
].bmap
->n_ineq
;
480 n_new
= 2 * bmap
->n_eq
;
481 for (k
= 0; k
< bmap
->n_ineq
; ++k
)
482 if (!isl_tab_is_redundant(tab
, bmap
->n_eq
+ k
))
485 return n_new
> n_old
;
488 /* Replace the pair of basic maps i and j by the basic map bounded
489 * by the valid constraints in both basic maps and the constraints
490 * in extra (if not NULL).
491 * Place the fused basic map in the position that is the smallest of i and j.
493 * If "detect_equalities" is set, then look for equalities encoded
494 * as pairs of inequalities.
495 * If "check_number" is set, then the original basic maps are only
496 * replaced if the total number of constraints does not increase.
497 * While the number of integer divisions in the two basic maps
498 * is assumed to be the same, the actual definitions may be different.
499 * We only copy the definition from one of the basic map if it is
500 * the same as that of the other basic map. Otherwise, we mark
501 * the integer division as unknown and simplify the basic map
502 * in an attempt to recover the integer division definition.
504 static enum isl_change
fuse(int i
, int j
, struct isl_coalesce_info
*info
,
505 __isl_keep isl_mat
*extra
, int detect_equalities
, int check_number
)
508 struct isl_basic_map
*fused
= NULL
;
509 struct isl_tab
*fused_tab
= NULL
;
510 isl_size total
= isl_basic_map_dim(info
[i
].bmap
, isl_dim_all
);
511 unsigned extra_rows
= extra
? extra
->n_row
: 0;
512 unsigned n_eq
, n_ineq
;
516 return isl_change_error
;
518 return fuse(j
, i
, info
, extra
, detect_equalities
, check_number
);
520 n_eq
= info
[i
].bmap
->n_eq
+ info
[j
].bmap
->n_eq
;
521 n_ineq
= info
[i
].bmap
->n_ineq
+ info
[j
].bmap
->n_ineq
;
522 fused
= isl_basic_map_alloc_space(isl_space_copy(info
[i
].bmap
->dim
),
523 info
[i
].bmap
->n_div
, n_eq
, n_eq
+ n_ineq
+ extra_rows
);
524 fused
= add_valid_constraints(fused
, &info
[i
], 1 + total
);
525 fused
= add_valid_constraints(fused
, &info
[j
], 1 + total
);
528 if (ISL_F_ISSET(info
[i
].bmap
, ISL_BASIC_MAP_RATIONAL
) &&
529 ISL_F_ISSET(info
[j
].bmap
, ISL_BASIC_MAP_RATIONAL
))
530 ISL_F_SET(fused
, ISL_BASIC_MAP_RATIONAL
);
532 for (k
= 0; k
< info
[i
].bmap
->n_div
; ++k
) {
533 int l
= isl_basic_map_alloc_div(fused
);
536 if (isl_seq_eq(info
[i
].bmap
->div
[k
], info
[j
].bmap
->div
[k
],
538 isl_seq_cpy(fused
->div
[l
], info
[i
].bmap
->div
[k
],
541 isl_int_set_si(fused
->div
[l
][0], 0);
546 for (k
= 0; k
< extra_rows
; ++k
) {
547 l
= isl_basic_map_alloc_inequality(fused
);
550 isl_seq_cpy(fused
->ineq
[l
], extra
->row
[k
], 1 + total
);
553 if (detect_equalities
)
554 fused
= isl_basic_map_detect_inequality_pairs(fused
, NULL
);
555 fused
= isl_basic_map_gauss(fused
, NULL
);
556 if (simplify
|| info
[j
].simplify
) {
557 fused
= isl_basic_map_simplify(fused
);
558 info
[i
].simplify
= 0;
560 fused
= isl_basic_map_finalize(fused
);
562 fused_tab
= isl_tab_from_basic_map(fused
, 0);
563 if (isl_tab_detect_redundant(fused_tab
) < 0)
567 number_of_constraints_increases(i
, j
, info
, fused
, fused_tab
)) {
568 isl_tab_free(fused_tab
);
569 isl_basic_map_free(fused
);
570 return isl_change_none
;
574 info
[i
].bmap
= fused
;
575 info
[i
].tab
= fused_tab
;
576 info
[i
].modified
= 1;
579 return isl_change_fuse
;
581 isl_tab_free(fused_tab
);
582 isl_basic_map_free(fused
);
583 return isl_change_error
;
586 /* Given a pair of basic maps i and j such that all constraints are either
587 * "valid" or "cut", check if the facets corresponding to the "cut"
588 * constraints of i lie entirely within basic map j.
589 * If so, replace the pair by the basic map consisting of the valid
590 * constraints in both basic maps.
591 * Checking whether the facet lies entirely within basic map j
592 * is performed by checking whether the constraints of basic map j
593 * are valid for the facet. These tests are performed on a rational
594 * tableau to avoid the theoretical possibility that a constraint
595 * that was considered to be a cut constraint for the entire basic map i
596 * happens to be considered to be a valid constraint for the facet,
597 * even though it cuts off the same rational points.
599 * To see that we are not introducing any extra points, call the
600 * two basic maps A and B and the resulting map U and let x
601 * be an element of U \setminus ( A \cup B ).
602 * A line connecting x with an element of A \cup B meets a facet F
603 * of either A or B. Assume it is a facet of B and let c_1 be
604 * the corresponding facet constraint. We have c_1(x) < 0 and
605 * so c_1 is a cut constraint. This implies that there is some
606 * (possibly rational) point x' satisfying the constraints of A
607 * and the opposite of c_1 as otherwise c_1 would have been marked
608 * valid for A. The line connecting x and x' meets a facet of A
609 * in a (possibly rational) point that also violates c_1, but this
610 * is impossible since all cut constraints of B are valid for all
612 * In case F is a facet of A rather than B, then we can apply the
613 * above reasoning to find a facet of B separating x from A \cup B first.
615 static enum isl_change
check_facets(int i
, int j
,
616 struct isl_coalesce_info
*info
)
619 struct isl_tab_undo
*snap
, *snap2
;
620 unsigned n_eq
= info
[i
].bmap
->n_eq
;
622 snap
= isl_tab_snap(info
[i
].tab
);
623 if (isl_tab_mark_rational(info
[i
].tab
) < 0)
624 return isl_change_error
;
625 snap2
= isl_tab_snap(info
[i
].tab
);
627 for (k
= 0; k
< info
[i
].bmap
->n_ineq
; ++k
) {
628 if (info
[i
].ineq
[k
] != STATUS_CUT
)
630 if (isl_tab_select_facet(info
[i
].tab
, n_eq
+ k
) < 0)
631 return isl_change_error
;
632 for (l
= 0; l
< info
[j
].bmap
->n_ineq
; ++l
) {
634 if (info
[j
].ineq
[l
] != STATUS_CUT
)
636 stat
= status_in(info
[j
].bmap
->ineq
[l
], info
[i
].tab
);
638 return isl_change_error
;
639 if (stat
!= STATUS_VALID
)
642 if (isl_tab_rollback(info
[i
].tab
, snap2
) < 0)
643 return isl_change_error
;
644 if (l
< info
[j
].bmap
->n_ineq
)
648 if (k
< info
[i
].bmap
->n_ineq
) {
649 if (isl_tab_rollback(info
[i
].tab
, snap
) < 0)
650 return isl_change_error
;
651 return isl_change_none
;
653 return fuse(i
, j
, info
, NULL
, 0, 0);
656 /* Check if info->bmap contains the basic map represented
657 * by the tableau "tab".
658 * For each equality, we check both the constraint itself
659 * (as an inequality) and its negation. Make sure the
660 * equality is returned to its original state before returning.
662 static isl_bool
contains(struct isl_coalesce_info
*info
, struct isl_tab
*tab
)
666 isl_basic_map
*bmap
= info
->bmap
;
668 dim
= isl_basic_map_dim(bmap
, isl_dim_all
);
670 return isl_bool_error
;
671 for (k
= 0; k
< bmap
->n_eq
; ++k
) {
673 isl_seq_neg(bmap
->eq
[k
], bmap
->eq
[k
], 1 + dim
);
674 stat
= status_in(bmap
->eq
[k
], tab
);
675 isl_seq_neg(bmap
->eq
[k
], bmap
->eq
[k
], 1 + dim
);
677 return isl_bool_error
;
678 if (stat
!= STATUS_VALID
)
679 return isl_bool_false
;
680 stat
= status_in(bmap
->eq
[k
], tab
);
682 return isl_bool_error
;
683 if (stat
!= STATUS_VALID
)
684 return isl_bool_false
;
687 for (k
= 0; k
< bmap
->n_ineq
; ++k
) {
689 if (info
->ineq
[k
] == STATUS_REDUNDANT
)
691 stat
= status_in(bmap
->ineq
[k
], tab
);
693 return isl_bool_error
;
694 if (stat
!= STATUS_VALID
)
695 return isl_bool_false
;
697 return isl_bool_true
;
700 /* Basic map "i" has an inequality (say "k") that is adjacent
701 * to some inequality of basic map "j". All the other inequalities
703 * Check if basic map "j" forms an extension of basic map "i".
705 * Note that this function is only called if some of the equalities or
706 * inequalities of basic map "j" do cut basic map "i". The function is
707 * correct even if there are no such cut constraints, but in that case
708 * the additional checks performed by this function are overkill.
710 * In particular, we replace constraint k, say f >= 0, by constraint
711 * f <= -1, add the inequalities of "j" that are valid for "i"
712 * and check if the result is a subset of basic map "j".
713 * To improve the chances of the subset relation being detected,
714 * any variable that only attains a single integer value
715 * in the tableau of "i" is first fixed to that value.
716 * If the result is a subset, then we know that this result is exactly equal
717 * to basic map "j" since all its constraints are valid for basic map "j".
718 * By combining the valid constraints of "i" (all equalities and all
719 * inequalities except "k") and the valid constraints of "j" we therefore
720 * obtain a basic map that is equal to their union.
721 * In this case, there is no need to perform a rollback of the tableau
722 * since it is going to be destroyed in fuse().
728 * |_______| _ |_________\
740 static enum isl_change
is_adj_ineq_extension(int i
, int j
,
741 struct isl_coalesce_info
*info
)
744 struct isl_tab_undo
*snap
;
745 unsigned n_eq
= info
[i
].bmap
->n_eq
;
746 isl_size total
= isl_basic_map_dim(info
[i
].bmap
, isl_dim_all
);
751 return isl_change_error
;
752 if (isl_tab_extend_cons(info
[i
].tab
, 1 + info
[j
].bmap
->n_ineq
) < 0)
753 return isl_change_error
;
755 k
= find_ineq(&info
[i
], STATUS_ADJ_INEQ
);
757 isl_die(isl_basic_map_get_ctx(info
[i
].bmap
), isl_error_internal
,
758 "info[i].ineq should have exactly one STATUS_ADJ_INEQ",
759 return isl_change_error
);
761 snap
= isl_tab_snap(info
[i
].tab
);
763 if (isl_tab_unrestrict(info
[i
].tab
, n_eq
+ k
) < 0)
764 return isl_change_error
;
766 isl_seq_neg(info
[i
].bmap
->ineq
[k
], info
[i
].bmap
->ineq
[k
], 1 + total
);
767 isl_int_sub_ui(info
[i
].bmap
->ineq
[k
][0], info
[i
].bmap
->ineq
[k
][0], 1);
768 r
= isl_tab_add_ineq(info
[i
].tab
, info
[i
].bmap
->ineq
[k
]);
769 isl_seq_neg(info
[i
].bmap
->ineq
[k
], info
[i
].bmap
->ineq
[k
], 1 + total
);
770 isl_int_sub_ui(info
[i
].bmap
->ineq
[k
][0], info
[i
].bmap
->ineq
[k
][0], 1);
772 return isl_change_error
;
774 for (k
= 0; k
< info
[j
].bmap
->n_ineq
; ++k
) {
775 if (info
[j
].ineq
[k
] != STATUS_VALID
)
777 if (isl_tab_add_ineq(info
[i
].tab
, info
[j
].bmap
->ineq
[k
]) < 0)
778 return isl_change_error
;
780 if (isl_tab_detect_constants(info
[i
].tab
) < 0)
781 return isl_change_error
;
783 super
= contains(&info
[j
], info
[i
].tab
);
785 return isl_change_error
;
787 return fuse(i
, j
, info
, NULL
, 0, 0);
789 if (isl_tab_rollback(info
[i
].tab
, snap
) < 0)
790 return isl_change_error
;
792 return isl_change_none
;
796 /* Both basic maps have at least one inequality with and adjacent
797 * (but opposite) inequality in the other basic map.
798 * Check that there are no cut constraints and that there is only
799 * a single pair of adjacent inequalities.
800 * If so, we can replace the pair by a single basic map described
801 * by all but the pair of adjacent inequalities.
802 * Any additional points introduced lie strictly between the two
803 * adjacent hyperplanes and can therefore be integral.
812 * The test for a single pair of adjancent inequalities is important
813 * for avoiding the combination of two basic maps like the following
823 * If there are some cut constraints on one side, then we may
824 * still be able to fuse the two basic maps, but we need to perform
825 * some additional checks in is_adj_ineq_extension.
827 static enum isl_change
check_adj_ineq(int i
, int j
,
828 struct isl_coalesce_info
*info
)
830 int count_i
, count_j
;
833 count_i
= count_ineq(&info
[i
], STATUS_ADJ_INEQ
);
834 count_j
= count_ineq(&info
[j
], STATUS_ADJ_INEQ
);
836 if (count_i
!= 1 && count_j
!= 1)
837 return isl_change_none
;
839 cut_i
= any_eq(&info
[i
], STATUS_CUT
) || any_ineq(&info
[i
], STATUS_CUT
);
840 cut_j
= any_eq(&info
[j
], STATUS_CUT
) || any_ineq(&info
[j
], STATUS_CUT
);
842 if (!cut_i
&& !cut_j
&& count_i
== 1 && count_j
== 1)
843 return fuse(i
, j
, info
, NULL
, 0, 0);
845 if (count_i
== 1 && !cut_i
)
846 return is_adj_ineq_extension(i
, j
, info
);
848 if (count_j
== 1 && !cut_j
)
849 return is_adj_ineq_extension(j
, i
, info
);
851 return isl_change_none
;
854 /* Given an affine transformation matrix "T", does row "row" represent
855 * anything other than a unit vector (possibly shifted by a constant)
856 * that is not involved in any of the other rows?
858 * That is, if a constraint involves the variable corresponding to
859 * the row, then could its preimage by "T" have any coefficients
860 * that are different from those in the original constraint?
862 static int not_unique_unit_row(__isl_keep isl_mat
*T
, int row
)
865 int len
= T
->n_col
- 1;
867 i
= isl_seq_first_non_zero(T
->row
[row
] + 1, len
);
870 if (!isl_int_is_one(T
->row
[row
][1 + i
]) &&
871 !isl_int_is_negone(T
->row
[row
][1 + i
]))
874 j
= isl_seq_first_non_zero(T
->row
[row
] + 1 + i
+ 1, len
- (i
+ 1));
878 for (j
= 1; j
< T
->n_row
; ++j
) {
881 if (!isl_int_is_zero(T
->row
[j
][1 + i
]))
888 /* Does inequality constraint "ineq" of "bmap" involve any of
889 * the variables marked in "affected"?
890 * "total" is the total number of variables, i.e., the number
891 * of entries in "affected".
893 static isl_bool
is_affected(__isl_keep isl_basic_map
*bmap
, int ineq
,
894 int *affected
, int total
)
898 for (i
= 0; i
< total
; ++i
) {
901 if (!isl_int_is_zero(bmap
->ineq
[ineq
][1 + i
]))
902 return isl_bool_true
;
905 return isl_bool_false
;
908 /* Given the compressed version of inequality constraint "ineq"
909 * of info->bmap in "v", check if the constraint can be tightened,
910 * where the compression is based on an equality constraint valid
912 * If so, add the tightened version of the inequality constraint
913 * to info->tab. "v" may be modified by this function.
915 * That is, if the compressed constraint is of the form
919 * with 0 < c < m, then it is equivalent to
923 * This means that c can also be subtracted from the original,
924 * uncompressed constraint without affecting the integer points
925 * in info->tab. Add this tightened constraint as an extra row
926 * to info->tab to make this information explicitly available.
928 static __isl_give isl_vec
*try_tightening(struct isl_coalesce_info
*info
,
929 int ineq
, __isl_take isl_vec
*v
)
937 ctx
= isl_vec_get_ctx(v
);
938 isl_seq_gcd(v
->el
+ 1, v
->size
- 1, &ctx
->normalize_gcd
);
939 if (isl_int_is_zero(ctx
->normalize_gcd
) ||
940 isl_int_is_one(ctx
->normalize_gcd
)) {
948 isl_int_fdiv_r(v
->el
[0], v
->el
[0], ctx
->normalize_gcd
);
949 if (isl_int_is_zero(v
->el
[0]))
952 if (isl_tab_extend_cons(info
->tab
, 1) < 0)
953 return isl_vec_free(v
);
955 isl_int_sub(info
->bmap
->ineq
[ineq
][0],
956 info
->bmap
->ineq
[ineq
][0], v
->el
[0]);
957 r
= isl_tab_add_ineq(info
->tab
, info
->bmap
->ineq
[ineq
]);
958 isl_int_add(info
->bmap
->ineq
[ineq
][0],
959 info
->bmap
->ineq
[ineq
][0], v
->el
[0]);
962 return isl_vec_free(v
);
967 /* Tighten the (non-redundant) constraints on the facet represented
969 * In particular, on input, info->tab represents the result
970 * of relaxing the "n" inequality constraints of info->bmap in "relaxed"
971 * by one, i.e., replacing f_i >= 0 by f_i + 1 >= 0, and then
972 * replacing the one at index "l" by the corresponding equality,
973 * i.e., f_k + 1 = 0, with k = relaxed[l].
975 * Compute a variable compression from the equality constraint f_k + 1 = 0
976 * and use it to tighten the other constraints of info->bmap
977 * (that is, all constraints that have not been relaxed),
978 * updating info->tab (and leaving info->bmap untouched).
979 * The compression handles essentially two cases, one where a variable
980 * is assigned a fixed value and can therefore be eliminated, and one
981 * where one variable is a shifted multiple of some other variable and
982 * can therefore be replaced by that multiple.
983 * Gaussian elimination would also work for the first case, but for
984 * the second case, the effectiveness would depend on the order
986 * After compression, some of the constraints may have coefficients
987 * with a common divisor. If this divisor does not divide the constant
988 * term, then the constraint can be tightened.
989 * The tightening is performed on the tableau info->tab by introducing
990 * extra (temporary) constraints.
992 * Only constraints that are possibly affected by the compression are
993 * considered. In particular, if the constraint only involves variables
994 * that are directly mapped to a distinct set of other variables, then
995 * no common divisor can be introduced and no tightening can occur.
997 * It is important to only consider the non-redundant constraints
998 * since the facet constraint has been relaxed prior to the call
999 * to this function, meaning that the constraints that were redundant
1000 * prior to the relaxation may no longer be redundant.
1001 * These constraints will be ignored in the fused result, so
1002 * the fusion detection should not exploit them.
1004 static isl_stat
tighten_on_relaxed_facet(struct isl_coalesce_info
*info
,
1005 int n
, int *relaxed
, int l
)
1016 ctx
= isl_basic_map_get_ctx(info
->bmap
);
1017 total
= isl_basic_map_dim(info
->bmap
, isl_dim_all
);
1019 return isl_stat_error
;
1020 isl_int_add_ui(info
->bmap
->ineq
[k
][0], info
->bmap
->ineq
[k
][0], 1);
1021 T
= isl_mat_sub_alloc6(ctx
, info
->bmap
->ineq
, k
, 1, 0, 1 + total
);
1022 T
= isl_mat_variable_compression(T
, NULL
);
1023 isl_int_sub_ui(info
->bmap
->ineq
[k
][0], info
->bmap
->ineq
[k
][0], 1);
1025 return isl_stat_error
;
1026 if (T
->n_col
== 0) {
1031 affected
= isl_alloc_array(ctx
, int, total
);
1035 for (i
= 0; i
< total
; ++i
)
1036 affected
[i
] = not_unique_unit_row(T
, 1 + i
);
1038 for (i
= 0; i
< info
->bmap
->n_ineq
; ++i
) {
1040 if (any(relaxed
, n
, i
))
1042 if (info
->ineq
[i
] == STATUS_REDUNDANT
)
1044 handle
= is_affected(info
->bmap
, i
, affected
, total
);
1049 v
= isl_vec_alloc(ctx
, 1 + total
);
1052 isl_seq_cpy(v
->el
, info
->bmap
->ineq
[i
], 1 + total
);
1053 v
= isl_vec_mat_product(v
, isl_mat_copy(T
));
1054 v
= try_tightening(info
, i
, v
);
1066 return isl_stat_error
;
1069 /* Replace the basic maps "i" and "j" by an extension of "i"
1070 * along the "n" inequality constraints in "relax" by one.
1071 * The tableau info[i].tab has already been extended.
1072 * Extend info[i].bmap accordingly by relaxing all constraints in "relax"
1074 * Each integer division that does not have exactly the same
1075 * definition in "i" and "j" is marked unknown and the basic map
1076 * is scheduled to be simplified in an attempt to recover
1077 * the integer division definition.
1078 * Place the extension in the position that is the smallest of i and j.
1080 static enum isl_change
extend(int i
, int j
, int n
, int *relax
,
1081 struct isl_coalesce_info
*info
)
1086 info
[i
].bmap
= isl_basic_map_cow(info
[i
].bmap
);
1087 total
= isl_basic_map_dim(info
[i
].bmap
, isl_dim_all
);
1089 return isl_change_error
;
1090 for (l
= 0; l
< info
[i
].bmap
->n_div
; ++l
)
1091 if (!isl_seq_eq(info
[i
].bmap
->div
[l
],
1092 info
[j
].bmap
->div
[l
], 1 + 1 + total
)) {
1093 isl_int_set_si(info
[i
].bmap
->div
[l
][0], 0);
1094 info
[i
].simplify
= 1;
1096 for (l
= 0; l
< n
; ++l
)
1097 isl_int_add_ui(info
[i
].bmap
->ineq
[relax
[l
]][0],
1098 info
[i
].bmap
->ineq
[relax
[l
]][0], 1);
1099 ISL_F_CLR(info
[i
].bmap
, ISL_BASIC_MAP_NO_REDUNDANT
);
1100 ISL_F_SET(info
[i
].bmap
, ISL_BASIC_MAP_FINAL
);
1102 info
[i
].modified
= 1;
1104 exchange(&info
[i
], &info
[j
]);
1105 return isl_change_fuse
;
1108 /* Basic map "i" has "n" inequality constraints (collected in "relax")
1109 * that are such that they include basic map "j" if they are relaxed
1110 * by one. All the other inequalities are valid for "j".
1111 * Check if basic map "j" forms an extension of basic map "i".
1113 * In particular, relax the constraints in "relax", compute the corresponding
1114 * facets one by one and check whether each of these is included
1115 * in the other basic map.
1116 * Before testing for inclusion, the constraints on each facet
1117 * are tightened to increase the chance of an inclusion being detected.
1118 * (Adding the valid constraints of "j" to the tableau of "i", as is done
1119 * in is_adj_ineq_extension, may further increase those chances, but this
1120 * is not currently done.)
1121 * If each facet is included, we know that relaxing the constraints extends
1122 * the basic map with exactly the other basic map (we already know that this
1123 * other basic map is included in the extension, because all other
1124 * inequality constraints are valid of "j") and we can replace the
1125 * two basic maps by this extension.
1127 * If any of the relaxed constraints turn out to be redundant, then bail out.
1128 * isl_tab_select_facet refuses to handle such constraints. It may be
1129 * possible to handle them anyway by making a distinction between
1130 * redundant constraints with a corresponding facet that still intersects
1131 * the set (allowing isl_tab_select_facet to handle them) and
1132 * those where the facet does not intersect the set (which can be ignored
1133 * because the empty facet is trivially included in the other disjunct).
1134 * However, relaxed constraints that turn out to be redundant should
1135 * be fairly rare and no such instance has been reported where
1136 * coalescing would be successful.
1152 static enum isl_change
is_relaxed_extension(int i
, int j
, int n
, int *relax
,
1153 struct isl_coalesce_info
*info
)
1157 struct isl_tab_undo
*snap
, *snap2
;
1158 unsigned n_eq
= info
[i
].bmap
->n_eq
;
1160 for (l
= 0; l
< n
; ++l
)
1161 if (isl_tab_is_equality(info
[i
].tab
, n_eq
+ relax
[l
]))
1162 return isl_change_none
;
1164 snap
= isl_tab_snap(info
[i
].tab
);
1165 for (l
= 0; l
< n
; ++l
)
1166 if (isl_tab_relax(info
[i
].tab
, n_eq
+ relax
[l
]) < 0)
1167 return isl_change_error
;
1168 for (l
= 0; l
< n
; ++l
) {
1169 if (!isl_tab_is_redundant(info
[i
].tab
, n_eq
+ relax
[l
]))
1171 if (isl_tab_rollback(info
[i
].tab
, snap
) < 0)
1172 return isl_change_error
;
1173 return isl_change_none
;
1175 snap2
= isl_tab_snap(info
[i
].tab
);
1176 for (l
= 0; l
< n
; ++l
) {
1177 if (isl_tab_rollback(info
[i
].tab
, snap2
) < 0)
1178 return isl_change_error
;
1179 if (isl_tab_select_facet(info
[i
].tab
, n_eq
+ relax
[l
]) < 0)
1180 return isl_change_error
;
1181 if (tighten_on_relaxed_facet(&info
[i
], n
, relax
, l
) < 0)
1182 return isl_change_error
;
1183 super
= contains(&info
[j
], info
[i
].tab
);
1185 return isl_change_error
;
1188 if (isl_tab_rollback(info
[i
].tab
, snap
) < 0)
1189 return isl_change_error
;
1190 return isl_change_none
;
1193 if (isl_tab_rollback(info
[i
].tab
, snap2
) < 0)
1194 return isl_change_error
;
1195 return extend(i
, j
, n
, relax
, info
);
1198 /* Data structure that keeps track of the wrapping constraints
1199 * and of information to bound the coefficients of those constraints.
1201 * bound is set if we want to apply a bound on the coefficients
1202 * mat contains the wrapping constraints
1203 * max is the bound on the coefficients (if bound is set)
1211 /* Update wraps->max to be greater than or equal to the coefficients
1212 * in the equalities and inequalities of info->bmap that can be removed
1213 * if we end up applying wrapping.
1215 static isl_stat
wraps_update_max(struct isl_wraps
*wraps
,
1216 struct isl_coalesce_info
*info
)
1220 isl_size total
= isl_basic_map_dim(info
->bmap
, isl_dim_all
);
1223 return isl_stat_error
;
1224 isl_int_init(max_k
);
1226 for (k
= 0; k
< info
->bmap
->n_eq
; ++k
) {
1227 if (info
->eq
[2 * k
] == STATUS_VALID
&&
1228 info
->eq
[2 * k
+ 1] == STATUS_VALID
)
1230 isl_seq_abs_max(info
->bmap
->eq
[k
] + 1, total
, &max_k
);
1231 if (isl_int_abs_gt(max_k
, wraps
->max
))
1232 isl_int_set(wraps
->max
, max_k
);
1235 for (k
= 0; k
< info
->bmap
->n_ineq
; ++k
) {
1236 if (info
->ineq
[k
] == STATUS_VALID
||
1237 info
->ineq
[k
] == STATUS_REDUNDANT
)
1239 isl_seq_abs_max(info
->bmap
->ineq
[k
] + 1, total
, &max_k
);
1240 if (isl_int_abs_gt(max_k
, wraps
->max
))
1241 isl_int_set(wraps
->max
, max_k
);
1244 isl_int_clear(max_k
);
1249 /* Initialize the isl_wraps data structure.
1250 * If we want to bound the coefficients of the wrapping constraints,
1251 * we set wraps->max to the largest coefficient
1252 * in the equalities and inequalities that can be removed if we end up
1253 * applying wrapping.
1255 static isl_stat
wraps_init(struct isl_wraps
*wraps
, __isl_take isl_mat
*mat
,
1256 struct isl_coalesce_info
*info
, int i
, int j
)
1263 return isl_stat_error
;
1264 ctx
= isl_mat_get_ctx(mat
);
1265 wraps
->bound
= isl_options_get_coalesce_bounded_wrapping(ctx
);
1268 isl_int_init(wraps
->max
);
1269 isl_int_set_si(wraps
->max
, 0);
1270 if (wraps_update_max(wraps
, &info
[i
]) < 0)
1271 return isl_stat_error
;
1272 if (wraps_update_max(wraps
, &info
[j
]) < 0)
1273 return isl_stat_error
;
1278 /* Free the contents of the isl_wraps data structure.
1280 static void wraps_free(struct isl_wraps
*wraps
)
1282 isl_mat_free(wraps
->mat
);
1284 isl_int_clear(wraps
->max
);
1287 /* Mark the wrapping as failed by resetting wraps->mat->n_row to zero.
1289 static isl_stat
wraps_mark_failed(struct isl_wraps
*wraps
)
1291 wraps
->mat
->n_row
= 0;
1295 /* Is the wrapping constraint in row "row" allowed?
1297 * If wraps->bound is set, we check that none of the coefficients
1298 * is greater than wraps->max.
1300 static int allow_wrap(struct isl_wraps
*wraps
, int row
)
1307 for (i
= 1; i
< wraps
->mat
->n_col
; ++i
)
1308 if (isl_int_abs_gt(wraps
->mat
->row
[row
][i
], wraps
->max
))
1314 /* Wrap "ineq" (or its opposite if "negate" is set) around "bound"
1315 * to include "set" and add the result in position "w" of "wraps".
1316 * "len" is the total number of coefficients in "bound" and "ineq".
1317 * Return 1 on success, 0 on failure and -1 on error.
1318 * Wrapping can fail if the result of wrapping is equal to "bound"
1319 * or if we want to bound the sizes of the coefficients and
1320 * the wrapped constraint does not satisfy this bound.
1322 static int add_wrap(struct isl_wraps
*wraps
, int w
, isl_int
*bound
,
1323 isl_int
*ineq
, unsigned len
, __isl_keep isl_set
*set
, int negate
)
1325 isl_seq_cpy(wraps
->mat
->row
[w
], bound
, len
);
1327 isl_seq_neg(wraps
->mat
->row
[w
+ 1], ineq
, len
);
1328 ineq
= wraps
->mat
->row
[w
+ 1];
1330 if (!isl_set_wrap_facet(set
, wraps
->mat
->row
[w
], ineq
))
1332 if (isl_seq_eq(wraps
->mat
->row
[w
], bound
, len
))
1334 if (!allow_wrap(wraps
, w
))
1339 /* For each constraint in info->bmap that is not redundant (as determined
1340 * by info->tab) and that is not a valid constraint for the other basic map,
1341 * wrap the constraint around "bound" such that it includes the whole
1342 * set "set" and append the resulting constraint to "wraps".
1343 * Note that the constraints that are valid for the other basic map
1344 * will be added to the combined basic map by default, so there is
1345 * no need to wrap them.
1346 * The caller wrap_in_facets even relies on this function not wrapping
1347 * any constraints that are already valid.
1348 * "wraps" is assumed to have been pre-allocated to the appropriate size.
1349 * wraps->n_row is the number of actual wrapped constraints that have
1351 * If any of the wrapping problems results in a constraint that is
1352 * identical to "bound", then this means that "set" is unbounded in such
1353 * way that no wrapping is possible. If this happens then wraps->n_row
1355 * Similarly, if we want to bound the coefficients of the wrapping
1356 * constraints and a newly added wrapping constraint does not
1357 * satisfy the bound, then wraps->n_row is also reset to zero.
1359 static isl_stat
add_wraps(struct isl_wraps
*wraps
,
1360 struct isl_coalesce_info
*info
, isl_int
*bound
, __isl_keep isl_set
*set
)
1365 isl_basic_map
*bmap
= info
->bmap
;
1366 isl_size total
= isl_basic_map_dim(bmap
, isl_dim_all
);
1367 unsigned len
= 1 + total
;
1370 return isl_stat_error
;
1372 w
= wraps
->mat
->n_row
;
1374 for (l
= 0; l
< bmap
->n_ineq
; ++l
) {
1375 if (info
->ineq
[l
] == STATUS_VALID
||
1376 info
->ineq
[l
] == STATUS_REDUNDANT
)
1378 if (isl_seq_is_neg(bound
, bmap
->ineq
[l
], len
))
1380 if (isl_seq_eq(bound
, bmap
->ineq
[l
], len
))
1382 if (isl_tab_is_redundant(info
->tab
, bmap
->n_eq
+ l
))
1385 added
= add_wrap(wraps
, w
, bound
, bmap
->ineq
[l
], len
, set
, 0);
1387 return isl_stat_error
;
1392 for (l
= 0; l
< bmap
->n_eq
; ++l
) {
1393 if (isl_seq_is_neg(bound
, bmap
->eq
[l
], len
))
1395 if (isl_seq_eq(bound
, bmap
->eq
[l
], len
))
1398 for (m
= 0; m
< 2; ++m
) {
1399 if (info
->eq
[2 * l
+ m
] == STATUS_VALID
)
1401 added
= add_wrap(wraps
, w
, bound
, bmap
->eq
[l
], len
,
1404 return isl_stat_error
;
1411 wraps
->mat
->n_row
= w
;
1414 return wraps_mark_failed(wraps
);
1417 /* Check if the constraints in "wraps" from "first" until the last
1418 * are all valid for the basic set represented by "tab".
1419 * If not, wraps->n_row is set to zero.
1421 static int check_wraps(__isl_keep isl_mat
*wraps
, int first
,
1422 struct isl_tab
*tab
)
1426 for (i
= first
; i
< wraps
->n_row
; ++i
) {
1427 enum isl_ineq_type type
;
1428 type
= isl_tab_ineq_type(tab
, wraps
->row
[i
]);
1429 if (type
== isl_ineq_error
)
1431 if (type
== isl_ineq_redundant
)
1440 /* Return a set that corresponds to the non-redundant constraints
1441 * (as recorded in tab) of bmap.
1443 * It's important to remove the redundant constraints as some
1444 * of the other constraints may have been modified after the
1445 * constraints were marked redundant.
1446 * In particular, a constraint may have been relaxed.
1447 * Redundant constraints are ignored when a constraint is relaxed
1448 * and should therefore continue to be ignored ever after.
1449 * Otherwise, the relaxation might be thwarted by some of
1450 * these constraints.
1452 * Update the underlying set to ensure that the dimension doesn't change.
1453 * Otherwise the integer divisions could get dropped if the tab
1454 * turns out to be empty.
1456 static __isl_give isl_set
*set_from_updated_bmap(__isl_keep isl_basic_map
*bmap
,
1457 struct isl_tab
*tab
)
1459 isl_basic_set
*bset
;
1461 bmap
= isl_basic_map_copy(bmap
);
1462 bset
= isl_basic_map_underlying_set(bmap
);
1463 bset
= isl_basic_set_cow(bset
);
1464 bset
= isl_basic_set_update_from_tab(bset
, tab
);
1465 return isl_set_from_basic_set(bset
);
1468 /* Does "info" have any cut constraints that are redundant?
1470 static isl_bool
has_redundant_cuts(struct isl_coalesce_info
*info
)
1475 n_eq
= isl_basic_map_n_equality(info
->bmap
);
1476 n_ineq
= isl_basic_map_n_inequality(info
->bmap
);
1477 if (n_eq
< 0 || n_ineq
< 0)
1478 return isl_bool_error
;
1479 for (l
= 0; l
< n_ineq
; ++l
) {
1482 if (info
->ineq
[l
] != STATUS_CUT
)
1484 red
= isl_tab_is_redundant(info
->tab
, n_eq
+ l
);
1486 return isl_bool_error
;
1488 return isl_bool_true
;
1491 return isl_bool_false
;
1494 /* Wrap the constraints of info->bmap that bound the facet defined
1495 * by inequality "k" around (the opposite of) this inequality to
1496 * include "set". "bound" may be used to store the negated inequality.
1497 * Since the wrapped constraints are not guaranteed to contain the whole
1498 * of info->bmap, we check them in check_wraps.
1499 * If any of the wrapped constraints turn out to be invalid, then
1500 * check_wraps will reset wrap->n_row to zero.
1502 * If any of the cut constraints of info->bmap turn out
1503 * to be redundant with respect to other constraints
1504 * then these will neither be wrapped nor added directly to the result.
1505 * The result may therefore not be correct.
1506 * Skip wrapping and reset wrap->mat->n_row to zero in this case.
1508 static isl_stat
add_wraps_around_facet(struct isl_wraps
*wraps
,
1509 struct isl_coalesce_info
*info
, int k
, isl_int
*bound
,
1510 __isl_keep isl_set
*set
)
1513 struct isl_tab_undo
*snap
;
1515 isl_size total
= isl_basic_map_dim(info
->bmap
, isl_dim_all
);
1518 return isl_stat_error
;
1520 snap
= isl_tab_snap(info
->tab
);
1522 if (isl_tab_select_facet(info
->tab
, info
->bmap
->n_eq
+ k
) < 0)
1523 return isl_stat_error
;
1524 if (isl_tab_detect_redundant(info
->tab
) < 0)
1525 return isl_stat_error
;
1526 nowrap
= has_redundant_cuts(info
);
1528 return isl_stat_error
;
1530 n
= wraps
->mat
->n_row
;
1532 isl_seq_neg(bound
, info
->bmap
->ineq
[k
], 1 + total
);
1534 if (add_wraps(wraps
, info
, bound
, set
) < 0)
1535 return isl_stat_error
;
1538 if (isl_tab_rollback(info
->tab
, snap
) < 0)
1539 return isl_stat_error
;
1541 return wraps_mark_failed(wraps
);
1542 if (check_wraps(wraps
->mat
, n
, info
->tab
) < 0)
1543 return isl_stat_error
;
1548 /* Given a basic set i with a constraint k that is adjacent to
1549 * basic set j, check if we can wrap
1550 * both the facet corresponding to k (if "wrap_facet" is set) and basic map j
1551 * (always) around their ridges to include the other set.
1552 * If so, replace the pair of basic sets by their union.
1554 * All constraints of i (except k) are assumed to be valid or
1555 * cut constraints for j.
1556 * Wrapping the cut constraints to include basic map j may result
1557 * in constraints that are no longer valid of basic map i
1558 * we have to check that the resulting wrapping constraints are valid for i.
1559 * If "wrap_facet" is not set, then all constraints of i (except k)
1560 * are assumed to be valid for j.
1569 static enum isl_change
can_wrap_in_facet(int i
, int j
, int k
,
1570 struct isl_coalesce_info
*info
, int wrap_facet
)
1572 enum isl_change change
= isl_change_none
;
1573 struct isl_wraps wraps
;
1576 struct isl_set
*set_i
= NULL
;
1577 struct isl_set
*set_j
= NULL
;
1578 struct isl_vec
*bound
= NULL
;
1579 isl_size total
= isl_basic_map_dim(info
[i
].bmap
, isl_dim_all
);
1582 return isl_change_error
;
1583 set_i
= set_from_updated_bmap(info
[i
].bmap
, info
[i
].tab
);
1584 set_j
= set_from_updated_bmap(info
[j
].bmap
, info
[j
].tab
);
1585 ctx
= isl_basic_map_get_ctx(info
[i
].bmap
);
1586 mat
= isl_mat_alloc(ctx
, 2 * (info
[i
].bmap
->n_eq
+ info
[j
].bmap
->n_eq
) +
1587 info
[i
].bmap
->n_ineq
+ info
[j
].bmap
->n_ineq
,
1589 if (wraps_init(&wraps
, mat
, info
, i
, j
) < 0)
1591 bound
= isl_vec_alloc(ctx
, 1 + total
);
1592 if (!set_i
|| !set_j
|| !bound
)
1595 isl_seq_cpy(bound
->el
, info
[i
].bmap
->ineq
[k
], 1 + total
);
1596 isl_int_add_ui(bound
->el
[0], bound
->el
[0], 1);
1597 isl_seq_normalize(ctx
, bound
->el
, 1 + total
);
1599 isl_seq_cpy(wraps
.mat
->row
[0], bound
->el
, 1 + total
);
1600 wraps
.mat
->n_row
= 1;
1602 if (add_wraps(&wraps
, &info
[j
], bound
->el
, set_i
) < 0)
1604 if (!wraps
.mat
->n_row
)
1608 if (add_wraps_around_facet(&wraps
, &info
[i
], k
,
1609 bound
->el
, set_j
) < 0)
1611 if (!wraps
.mat
->n_row
)
1615 change
= fuse(i
, j
, info
, wraps
.mat
, 0, 0);
1620 isl_set_free(set_i
);
1621 isl_set_free(set_j
);
1623 isl_vec_free(bound
);
1628 isl_vec_free(bound
);
1629 isl_set_free(set_i
);
1630 isl_set_free(set_j
);
1631 return isl_change_error
;
1634 /* Given a cut constraint t(x) >= 0 of basic map i, stored in row "w"
1635 * of wrap.mat, replace it by its relaxed version t(x) + 1 >= 0, and
1636 * add wrapping constraints to wrap.mat for all constraints
1637 * of basic map j that bound the part of basic map j that sticks out
1638 * of the cut constraint.
1639 * "set_i" is the underlying set of basic map i.
1640 * If any wrapping fails, then wraps->mat.n_row is reset to zero.
1642 * In particular, we first intersect basic map j with t(x) + 1 = 0.
1643 * If the result is empty, then t(x) >= 0 was actually a valid constraint
1644 * (with respect to the integer points), so we add t(x) >= 0 instead.
1645 * Otherwise, we wrap the constraints of basic map j that are not
1646 * redundant in this intersection and that are not already valid
1647 * for basic map i over basic map i.
1648 * Note that it is sufficient to wrap the constraints to include
1649 * basic map i, because we will only wrap the constraints that do
1650 * not include basic map i already. The wrapped constraint will
1651 * therefore be more relaxed compared to the original constraint.
1652 * Since the original constraint is valid for basic map j, so is
1653 * the wrapped constraint.
1655 static isl_stat
wrap_in_facet(struct isl_wraps
*wraps
, int w
,
1656 struct isl_coalesce_info
*info_j
, __isl_keep isl_set
*set_i
,
1657 struct isl_tab_undo
*snap
)
1659 isl_int_add_ui(wraps
->mat
->row
[w
][0], wraps
->mat
->row
[w
][0], 1);
1660 if (isl_tab_add_eq(info_j
->tab
, wraps
->mat
->row
[w
]) < 0)
1661 return isl_stat_error
;
1662 if (isl_tab_detect_redundant(info_j
->tab
) < 0)
1663 return isl_stat_error
;
1665 if (info_j
->tab
->empty
)
1666 isl_int_sub_ui(wraps
->mat
->row
[w
][0], wraps
->mat
->row
[w
][0], 1);
1667 else if (add_wraps(wraps
, info_j
, wraps
->mat
->row
[w
], set_i
) < 0)
1668 return isl_stat_error
;
1670 if (isl_tab_rollback(info_j
->tab
, snap
) < 0)
1671 return isl_stat_error
;
1676 /* Given a pair of basic maps i and j such that j sticks out
1677 * of i at n cut constraints, each time by at most one,
1678 * try to compute wrapping constraints and replace the two
1679 * basic maps by a single basic map.
1680 * The other constraints of i are assumed to be valid for j.
1681 * "set_i" is the underlying set of basic map i.
1682 * "wraps" has been initialized to be of the right size.
1684 * For each cut constraint t(x) >= 0 of i, we add the relaxed version
1685 * t(x) + 1 >= 0, along with wrapping constraints for all constraints
1686 * of basic map j that bound the part of basic map j that sticks out
1687 * of the cut constraint.
1689 * If any wrapping fails, i.e., if we cannot wrap to touch
1690 * the union, then we give up.
1691 * Otherwise, the pair of basic maps is replaced by their union.
1693 static enum isl_change
try_wrap_in_facets(int i
, int j
,
1694 struct isl_coalesce_info
*info
, struct isl_wraps
*wraps
,
1695 __isl_keep isl_set
*set_i
)
1699 struct isl_tab_undo
*snap
;
1701 total
= isl_basic_map_dim(info
[i
].bmap
, isl_dim_all
);
1703 return isl_change_error
;
1705 snap
= isl_tab_snap(info
[j
].tab
);
1707 wraps
->mat
->n_row
= 0;
1709 for (k
= 0; k
< info
[i
].bmap
->n_eq
; ++k
) {
1710 for (l
= 0; l
< 2; ++l
) {
1711 if (info
[i
].eq
[2 * k
+ l
] != STATUS_CUT
)
1713 w
= wraps
->mat
->n_row
++;
1715 isl_seq_neg(wraps
->mat
->row
[w
],
1716 info
[i
].bmap
->eq
[k
], 1 + total
);
1718 isl_seq_cpy(wraps
->mat
->row
[w
],
1719 info
[i
].bmap
->eq
[k
], 1 + total
);
1720 if (wrap_in_facet(wraps
, w
, &info
[j
], set_i
, snap
) < 0)
1721 return isl_change_error
;
1723 if (!wraps
->mat
->n_row
)
1724 return isl_change_none
;
1728 for (k
= 0; k
< info
[i
].bmap
->n_ineq
; ++k
) {
1729 if (info
[i
].ineq
[k
] != STATUS_CUT
)
1731 w
= wraps
->mat
->n_row
++;
1732 isl_seq_cpy(wraps
->mat
->row
[w
],
1733 info
[i
].bmap
->ineq
[k
], 1 + total
);
1734 if (wrap_in_facet(wraps
, w
, &info
[j
], set_i
, snap
) < 0)
1735 return isl_change_error
;
1737 if (!wraps
->mat
->n_row
)
1738 return isl_change_none
;
1741 return fuse(i
, j
, info
, wraps
->mat
, 0, 1);
1744 /* Given a pair of basic maps i and j such that j sticks out
1745 * of i at n cut constraints, each time by at most one,
1746 * try to compute wrapping constraints and replace the two
1747 * basic maps by a single basic map.
1748 * The other constraints of i are assumed to be valid for j.
1750 * The core computation is performed by try_wrap_in_facets.
1751 * This function simply extracts an underlying set representation
1752 * of basic map i and initializes the data structure for keeping
1753 * track of wrapping constraints.
1755 static enum isl_change
wrap_in_facets(int i
, int j
, int n
,
1756 struct isl_coalesce_info
*info
)
1758 enum isl_change change
= isl_change_none
;
1759 struct isl_wraps wraps
;
1762 isl_set
*set_i
= NULL
;
1763 isl_size total
= isl_basic_map_dim(info
[i
].bmap
, isl_dim_all
);
1767 return isl_change_error
;
1768 if (isl_tab_extend_cons(info
[j
].tab
, 1) < 0)
1769 return isl_change_error
;
1771 max_wrap
= 1 + 2 * info
[j
].bmap
->n_eq
+ info
[j
].bmap
->n_ineq
;
1774 set_i
= set_from_updated_bmap(info
[i
].bmap
, info
[i
].tab
);
1775 ctx
= isl_basic_map_get_ctx(info
[i
].bmap
);
1776 mat
= isl_mat_alloc(ctx
, max_wrap
, 1 + total
);
1777 if (wraps_init(&wraps
, mat
, info
, i
, j
) < 0)
1782 change
= try_wrap_in_facets(i
, j
, info
, &wraps
, set_i
);
1785 isl_set_free(set_i
);
1790 isl_set_free(set_i
);
1791 return isl_change_error
;
1794 /* Return the effect of inequality "ineq" on the tableau "tab",
1795 * after relaxing the constant term of "ineq" by one.
1797 static enum isl_ineq_type
type_of_relaxed(struct isl_tab
*tab
, isl_int
*ineq
)
1799 enum isl_ineq_type type
;
1801 isl_int_add_ui(ineq
[0], ineq
[0], 1);
1802 type
= isl_tab_ineq_type(tab
, ineq
);
1803 isl_int_sub_ui(ineq
[0], ineq
[0], 1);
1808 /* Given two basic sets i and j,
1809 * check if relaxing all the cut constraints of i by one turns
1810 * them into valid constraint for j and check if we can wrap in
1811 * the bits that are sticking out.
1812 * If so, replace the pair by their union.
1814 * We first check if all relaxed cut inequalities of i are valid for j
1815 * and then try to wrap in the intersections of the relaxed cut inequalities
1818 * During this wrapping, we consider the points of j that lie at a distance
1819 * of exactly 1 from i. In particular, we ignore the points that lie in
1820 * between this lower-dimensional space and the basic map i.
1821 * We can therefore only apply this to integer maps.
1847 * Wrapping can fail if the result of wrapping one of the facets
1848 * around its edges does not produce any new facet constraint.
1849 * In particular, this happens when we try to wrap in unbounded sets.
1851 * _______________________________________________________________________
1855 * |_| |_________________________________________________________________
1858 * The following is not an acceptable result of coalescing the above two
1859 * sets as it includes extra integer points.
1860 * _______________________________________________________________________
1865 * \______________________________________________________________________
1867 static enum isl_change
can_wrap_in_set(int i
, int j
,
1868 struct isl_coalesce_info
*info
)
1874 if (ISL_F_ISSET(info
[i
].bmap
, ISL_BASIC_MAP_RATIONAL
) ||
1875 ISL_F_ISSET(info
[j
].bmap
, ISL_BASIC_MAP_RATIONAL
))
1876 return isl_change_none
;
1878 n
= count_eq(&info
[i
], STATUS_CUT
) + count_ineq(&info
[i
], STATUS_CUT
);
1880 return isl_change_none
;
1882 total
= isl_basic_map_dim(info
[i
].bmap
, isl_dim_all
);
1884 return isl_change_error
;
1885 for (k
= 0; k
< info
[i
].bmap
->n_eq
; ++k
) {
1886 for (l
= 0; l
< 2; ++l
) {
1887 enum isl_ineq_type type
;
1889 if (info
[i
].eq
[2 * k
+ l
] != STATUS_CUT
)
1893 isl_seq_neg(info
[i
].bmap
->eq
[k
],
1894 info
[i
].bmap
->eq
[k
], 1 + total
);
1895 type
= type_of_relaxed(info
[j
].tab
,
1896 info
[i
].bmap
->eq
[k
]);
1898 isl_seq_neg(info
[i
].bmap
->eq
[k
],
1899 info
[i
].bmap
->eq
[k
], 1 + total
);
1900 if (type
== isl_ineq_error
)
1901 return isl_change_error
;
1902 if (type
!= isl_ineq_redundant
)
1903 return isl_change_none
;
1907 for (k
= 0; k
< info
[i
].bmap
->n_ineq
; ++k
) {
1908 enum isl_ineq_type type
;
1910 if (info
[i
].ineq
[k
] != STATUS_CUT
)
1913 type
= type_of_relaxed(info
[j
].tab
, info
[i
].bmap
->ineq
[k
]);
1914 if (type
== isl_ineq_error
)
1915 return isl_change_error
;
1916 if (type
!= isl_ineq_redundant
)
1917 return isl_change_none
;
1920 return wrap_in_facets(i
, j
, n
, info
);
1923 /* Check if either i or j has only cut constraints that can
1924 * be used to wrap in (a facet of) the other basic set.
1925 * if so, replace the pair by their union.
1927 static enum isl_change
check_wrap(int i
, int j
, struct isl_coalesce_info
*info
)
1929 enum isl_change change
= isl_change_none
;
1931 change
= can_wrap_in_set(i
, j
, info
);
1932 if (change
!= isl_change_none
)
1935 change
= can_wrap_in_set(j
, i
, info
);
1939 /* Check if all inequality constraints of "i" that cut "j" cease
1940 * to be cut constraints if they are relaxed by one.
1941 * If so, collect the cut constraints in "list".
1942 * The caller is responsible for allocating "list".
1944 static isl_bool
all_cut_by_one(int i
, int j
, struct isl_coalesce_info
*info
,
1950 for (l
= 0; l
< info
[i
].bmap
->n_ineq
; ++l
) {
1951 enum isl_ineq_type type
;
1953 if (info
[i
].ineq
[l
] != STATUS_CUT
)
1955 type
= type_of_relaxed(info
[j
].tab
, info
[i
].bmap
->ineq
[l
]);
1956 if (type
== isl_ineq_error
)
1957 return isl_bool_error
;
1958 if (type
!= isl_ineq_redundant
)
1959 return isl_bool_false
;
1963 return isl_bool_true
;
1966 /* Given two basic maps such that "j" has at least one equality constraint
1967 * that is adjacent to an inequality constraint of "i" and such that "i" has
1968 * exactly one inequality constraint that is adjacent to an equality
1969 * constraint of "j", check whether "i" can be extended to include "j" or
1970 * whether "j" can be wrapped into "i".
1971 * All remaining constraints of "i" and "j" are assumed to be valid
1972 * or cut constraints of the other basic map.
1973 * However, none of the equality constraints of "i" are cut constraints.
1975 * If "i" has any "cut" inequality constraints, then check if relaxing
1976 * each of them by one is sufficient for them to become valid.
1977 * If so, check if the inequality constraint adjacent to an equality
1978 * constraint of "j" along with all these cut constraints
1979 * can be relaxed by one to contain exactly "j".
1980 * Otherwise, or if this fails, check if "j" can be wrapped into "i".
1982 static enum isl_change
check_single_adj_eq(int i
, int j
,
1983 struct isl_coalesce_info
*info
)
1985 enum isl_change change
= isl_change_none
;
1992 n_cut
= count_ineq(&info
[i
], STATUS_CUT
);
1994 k
= find_ineq(&info
[i
], STATUS_ADJ_EQ
);
1997 ctx
= isl_basic_map_get_ctx(info
[i
].bmap
);
1998 relax
= isl_calloc_array(ctx
, int, 1 + n_cut
);
2000 return isl_change_error
;
2002 try_relax
= all_cut_by_one(i
, j
, info
, relax
+ 1);
2004 change
= isl_change_error
;
2006 try_relax
= isl_bool_true
;
2009 if (try_relax
&& change
== isl_change_none
)
2010 change
= is_relaxed_extension(i
, j
, 1 + n_cut
, relax
, info
);
2013 if (change
!= isl_change_none
)
2016 change
= can_wrap_in_facet(i
, j
, k
, info
, n_cut
> 0);
2021 /* At least one of the basic maps has an equality that is adjacent
2022 * to an inequality. Make sure that only one of the basic maps has
2023 * such an equality and that the other basic map has exactly one
2024 * inequality adjacent to an equality.
2025 * If the other basic map does not have such an inequality, then
2026 * check if all its constraints are either valid or cut constraints
2027 * and, if so, try wrapping in the first map into the second.
2028 * Otherwise, try to extend one basic map with the other or
2029 * wrap one basic map in the other.
2031 static enum isl_change
check_adj_eq(int i
, int j
,
2032 struct isl_coalesce_info
*info
)
2034 if (any_eq(&info
[i
], STATUS_ADJ_INEQ
) &&
2035 any_eq(&info
[j
], STATUS_ADJ_INEQ
))
2036 /* ADJ EQ TOO MANY */
2037 return isl_change_none
;
2039 if (any_eq(&info
[i
], STATUS_ADJ_INEQ
))
2040 return check_adj_eq(j
, i
, info
);
2042 /* j has an equality adjacent to an inequality in i */
2044 if (count_ineq(&info
[i
], STATUS_ADJ_EQ
) != 1) {
2045 if (all_valid_or_cut(&info
[i
]))
2046 return can_wrap_in_set(i
, j
, info
);
2047 return isl_change_none
;
2049 if (any_eq(&info
[i
], STATUS_CUT
))
2050 return isl_change_none
;
2051 if (any_ineq(&info
[j
], STATUS_ADJ_EQ
) ||
2052 any_ineq(&info
[i
], STATUS_ADJ_INEQ
) ||
2053 any_ineq(&info
[j
], STATUS_ADJ_INEQ
))
2054 /* ADJ EQ TOO MANY */
2055 return isl_change_none
;
2057 return check_single_adj_eq(i
, j
, info
);
2060 /* Disjunct "j" lies on a hyperplane that is adjacent to disjunct "i".
2061 * In particular, disjunct "i" has an inequality constraint that is adjacent
2062 * to a (combination of) equality constraint(s) of disjunct "j",
2063 * but disjunct "j" has no explicit equality constraint adjacent
2064 * to an inequality constraint of disjunct "i".
2066 * Disjunct "i" is already known not to have any equality constraints
2067 * that are adjacent to an equality or inequality constraint.
2068 * Check that, other than the inequality constraint mentioned above,
2069 * all other constraints of disjunct "i" are valid for disjunct "j".
2070 * If so, try and wrap in disjunct "j".
2072 static enum isl_change
check_ineq_adj_eq(int i
, int j
,
2073 struct isl_coalesce_info
*info
)
2077 if (any_eq(&info
[i
], STATUS_CUT
))
2078 return isl_change_none
;
2079 if (any_ineq(&info
[i
], STATUS_CUT
))
2080 return isl_change_none
;
2081 if (any_ineq(&info
[i
], STATUS_ADJ_INEQ
))
2082 return isl_change_none
;
2083 if (count_ineq(&info
[i
], STATUS_ADJ_EQ
) != 1)
2084 return isl_change_none
;
2086 k
= find_ineq(&info
[i
], STATUS_ADJ_EQ
);
2088 return can_wrap_in_facet(i
, j
, k
, info
, 0);
2091 /* The two basic maps lie on adjacent hyperplanes. In particular,
2092 * basic map "i" has an equality that lies parallel to basic map "j".
2093 * Check if we can wrap the facets around the parallel hyperplanes
2094 * to include the other set.
2096 * We perform basically the same operations as can_wrap_in_facet,
2097 * except that we don't need to select a facet of one of the sets.
2103 * If there is more than one equality of "i" adjacent to an equality of "j",
2104 * then the result will satisfy one or more equalities that are a linear
2105 * combination of these equalities. These will be encoded as pairs
2106 * of inequalities in the wrapping constraints and need to be made
2109 static enum isl_change
check_eq_adj_eq(int i
, int j
,
2110 struct isl_coalesce_info
*info
)
2113 enum isl_change change
= isl_change_none
;
2114 int detect_equalities
= 0;
2115 struct isl_wraps wraps
;
2118 struct isl_set
*set_i
= NULL
;
2119 struct isl_set
*set_j
= NULL
;
2120 struct isl_vec
*bound
= NULL
;
2121 isl_size total
= isl_basic_map_dim(info
[i
].bmap
, isl_dim_all
);
2124 return isl_change_error
;
2125 if (count_eq(&info
[i
], STATUS_ADJ_EQ
) != 1)
2126 detect_equalities
= 1;
2128 k
= find_eq(&info
[i
], STATUS_ADJ_EQ
);
2130 set_i
= set_from_updated_bmap(info
[i
].bmap
, info
[i
].tab
);
2131 set_j
= set_from_updated_bmap(info
[j
].bmap
, info
[j
].tab
);
2132 ctx
= isl_basic_map_get_ctx(info
[i
].bmap
);
2133 mat
= isl_mat_alloc(ctx
, 2 * (info
[i
].bmap
->n_eq
+ info
[j
].bmap
->n_eq
) +
2134 info
[i
].bmap
->n_ineq
+ info
[j
].bmap
->n_ineq
,
2136 if (wraps_init(&wraps
, mat
, info
, i
, j
) < 0)
2138 bound
= isl_vec_alloc(ctx
, 1 + total
);
2139 if (!set_i
|| !set_j
|| !bound
)
2143 isl_seq_neg(bound
->el
, info
[i
].bmap
->eq
[k
/ 2], 1 + total
);
2145 isl_seq_cpy(bound
->el
, info
[i
].bmap
->eq
[k
/ 2], 1 + total
);
2146 isl_int_add_ui(bound
->el
[0], bound
->el
[0], 1);
2148 isl_seq_cpy(wraps
.mat
->row
[0], bound
->el
, 1 + total
);
2149 wraps
.mat
->n_row
= 1;
2151 if (add_wraps(&wraps
, &info
[j
], bound
->el
, set_i
) < 0)
2153 if (!wraps
.mat
->n_row
)
2156 isl_int_sub_ui(bound
->el
[0], bound
->el
[0], 1);
2157 isl_seq_neg(bound
->el
, bound
->el
, 1 + total
);
2159 isl_seq_cpy(wraps
.mat
->row
[wraps
.mat
->n_row
], bound
->el
, 1 + total
);
2162 if (add_wraps(&wraps
, &info
[i
], bound
->el
, set_j
) < 0)
2164 if (!wraps
.mat
->n_row
)
2167 change
= fuse(i
, j
, info
, wraps
.mat
, detect_equalities
, 0);
2170 error
: change
= isl_change_error
;
2175 isl_set_free(set_i
);
2176 isl_set_free(set_j
);
2177 isl_vec_free(bound
);
2182 /* Initialize the "eq" and "ineq" fields of "info".
2184 static void init_status(struct isl_coalesce_info
*info
)
2186 info
->eq
= info
->ineq
= NULL
;
2189 /* Set info->eq to the positions of the equalities of info->bmap
2190 * with respect to the basic map represented by "tab".
2191 * If info->eq has already been computed, then do not compute it again.
2193 static void set_eq_status_in(struct isl_coalesce_info
*info
,
2194 struct isl_tab
*tab
)
2198 info
->eq
= eq_status_in(info
->bmap
, tab
);
2201 /* Set info->ineq to the positions of the inequalities of info->bmap
2202 * with respect to the basic map represented by "tab".
2203 * If info->ineq has already been computed, then do not compute it again.
2205 static void set_ineq_status_in(struct isl_coalesce_info
*info
,
2206 struct isl_tab
*tab
)
2210 info
->ineq
= ineq_status_in(info
->bmap
, info
->tab
, tab
);
2213 /* Free the memory allocated by the "eq" and "ineq" fields of "info".
2214 * This function assumes that init_status has been called on "info" first,
2215 * after which the "eq" and "ineq" fields may or may not have been
2216 * assigned a newly allocated array.
2218 static void clear_status(struct isl_coalesce_info
*info
)
2224 /* Are all inequality constraints of the basic map represented by "info"
2225 * valid for the other basic map, except for a single constraint
2226 * that is adjacent to an inequality constraint of the other basic map?
2228 static int all_ineq_valid_or_single_adj_ineq(struct isl_coalesce_info
*info
)
2233 for (i
= 0; i
< info
->bmap
->n_ineq
; ++i
) {
2234 if (info
->ineq
[i
] == STATUS_REDUNDANT
)
2236 if (info
->ineq
[i
] == STATUS_VALID
)
2238 if (info
->ineq
[i
] != STATUS_ADJ_INEQ
)
2248 /* Basic map "i" has one or more equality constraints that separate it
2249 * from basic map "j". Check if it happens to be an extension
2251 * In particular, check that all constraints of "j" are valid for "i",
2252 * except for one inequality constraint that is adjacent
2253 * to an inequality constraints of "i".
2254 * If so, check for "i" being an extension of "j" by calling
2255 * is_adj_ineq_extension.
2257 * Clean up the memory allocated for keeping track of the status
2258 * of the constraints before returning.
2260 static enum isl_change
separating_equality(int i
, int j
,
2261 struct isl_coalesce_info
*info
)
2263 enum isl_change change
= isl_change_none
;
2265 if (all(info
[j
].eq
, 2 * info
[j
].bmap
->n_eq
, STATUS_VALID
) &&
2266 all_ineq_valid_or_single_adj_ineq(&info
[j
]))
2267 change
= is_adj_ineq_extension(j
, i
, info
);
2269 clear_status(&info
[i
]);
2270 clear_status(&info
[j
]);
2274 /* Check if the union of the given pair of basic maps
2275 * can be represented by a single basic map.
2276 * If so, replace the pair by the single basic map and return
2277 * isl_change_drop_first, isl_change_drop_second or isl_change_fuse.
2278 * Otherwise, return isl_change_none.
2279 * The two basic maps are assumed to live in the same local space.
2280 * The "eq" and "ineq" fields of info[i] and info[j] are assumed
2281 * to have been initialized by the caller, either to NULL or
2282 * to valid information.
2284 * We first check the effect of each constraint of one basic map
2285 * on the other basic map.
2286 * The constraint may be
2287 * redundant the constraint is redundant in its own
2288 * basic map and should be ignore and removed
2290 * valid all (integer) points of the other basic map
2291 * satisfy the constraint
2292 * separate no (integer) point of the other basic map
2293 * satisfies the constraint
2294 * cut some but not all points of the other basic map
2295 * satisfy the constraint
2296 * adj_eq the given constraint is adjacent (on the outside)
2297 * to an equality of the other basic map
2298 * adj_ineq the given constraint is adjacent (on the outside)
2299 * to an inequality of the other basic map
2301 * We consider seven cases in which we can replace the pair by a single
2302 * basic map. We ignore all "redundant" constraints.
2304 * 1. all constraints of one basic map are valid
2305 * => the other basic map is a subset and can be removed
2307 * 2. all constraints of both basic maps are either "valid" or "cut"
2308 * and the facets corresponding to the "cut" constraints
2309 * of one of the basic maps lies entirely inside the other basic map
2310 * => the pair can be replaced by a basic map consisting
2311 * of the valid constraints in both basic maps
2313 * 3. there is a single pair of adjacent inequalities
2314 * (all other constraints are "valid")
2315 * => the pair can be replaced by a basic map consisting
2316 * of the valid constraints in both basic maps
2318 * 4. one basic map has a single adjacent inequality, while the other
2319 * constraints are "valid". The other basic map has some
2320 * "cut" constraints, but replacing the adjacent inequality by
2321 * its opposite and adding the valid constraints of the other
2322 * basic map results in a subset of the other basic map
2323 * => the pair can be replaced by a basic map consisting
2324 * of the valid constraints in both basic maps
2326 * 5. there is a single adjacent pair of an inequality and an equality,
2327 * the other constraints of the basic map containing the inequality are
2328 * "valid". Moreover, if the inequality the basic map is relaxed
2329 * and then turned into an equality, then resulting facet lies
2330 * entirely inside the other basic map
2331 * => the pair can be replaced by the basic map containing
2332 * the inequality, with the inequality relaxed.
2334 * 6. there is a single inequality adjacent to an equality,
2335 * the other constraints of the basic map containing the inequality are
2336 * "valid". Moreover, the facets corresponding to both
2337 * the inequality and the equality can be wrapped around their
2338 * ridges to include the other basic map
2339 * => the pair can be replaced by a basic map consisting
2340 * of the valid constraints in both basic maps together
2341 * with all wrapping constraints
2343 * 7. one of the basic maps extends beyond the other by at most one.
2344 * Moreover, the facets corresponding to the cut constraints and
2345 * the pieces of the other basic map at offset one from these cut
2346 * constraints can be wrapped around their ridges to include
2347 * the union of the two basic maps
2348 * => the pair can be replaced by a basic map consisting
2349 * of the valid constraints in both basic maps together
2350 * with all wrapping constraints
2352 * 8. the two basic maps live in adjacent hyperplanes. In principle
2353 * such sets can always be combined through wrapping, but we impose
2354 * that there is only one such pair, to avoid overeager coalescing.
2356 * Throughout the computation, we maintain a collection of tableaus
2357 * corresponding to the basic maps. When the basic maps are dropped
2358 * or combined, the tableaus are modified accordingly.
2360 static enum isl_change
coalesce_local_pair_reuse(int i
, int j
,
2361 struct isl_coalesce_info
*info
)
2363 enum isl_change change
= isl_change_none
;
2365 set_ineq_status_in(&info
[i
], info
[j
].tab
);
2366 if (info
[i
].bmap
->n_ineq
&& !info
[i
].ineq
)
2368 if (any_ineq(&info
[i
], STATUS_ERROR
))
2370 if (any_ineq(&info
[i
], STATUS_SEPARATE
))
2373 set_ineq_status_in(&info
[j
], info
[i
].tab
);
2374 if (info
[j
].bmap
->n_ineq
&& !info
[j
].ineq
)
2376 if (any_ineq(&info
[j
], STATUS_ERROR
))
2378 if (any_ineq(&info
[j
], STATUS_SEPARATE
))
2381 set_eq_status_in(&info
[i
], info
[j
].tab
);
2382 if (info
[i
].bmap
->n_eq
&& !info
[i
].eq
)
2384 if (any_eq(&info
[i
], STATUS_ERROR
))
2387 set_eq_status_in(&info
[j
], info
[i
].tab
);
2388 if (info
[j
].bmap
->n_eq
&& !info
[j
].eq
)
2390 if (any_eq(&info
[j
], STATUS_ERROR
))
2393 if (any_eq(&info
[i
], STATUS_SEPARATE
))
2394 return separating_equality(i
, j
, info
);
2395 if (any_eq(&info
[j
], STATUS_SEPARATE
))
2396 return separating_equality(j
, i
, info
);
2398 if (all(info
[i
].eq
, 2 * info
[i
].bmap
->n_eq
, STATUS_VALID
) &&
2399 all(info
[i
].ineq
, info
[i
].bmap
->n_ineq
, STATUS_VALID
)) {
2401 change
= isl_change_drop_second
;
2402 } else if (all(info
[j
].eq
, 2 * info
[j
].bmap
->n_eq
, STATUS_VALID
) &&
2403 all(info
[j
].ineq
, info
[j
].bmap
->n_ineq
, STATUS_VALID
)) {
2405 change
= isl_change_drop_first
;
2406 } else if (any_eq(&info
[i
], STATUS_ADJ_EQ
)) {
2407 change
= check_eq_adj_eq(i
, j
, info
);
2408 } else if (any_eq(&info
[j
], STATUS_ADJ_EQ
)) {
2409 change
= check_eq_adj_eq(j
, i
, info
);
2410 } else if (any_eq(&info
[i
], STATUS_ADJ_INEQ
) ||
2411 any_eq(&info
[j
], STATUS_ADJ_INEQ
)) {
2412 change
= check_adj_eq(i
, j
, info
);
2413 } else if (any_ineq(&info
[i
], STATUS_ADJ_EQ
)) {
2414 change
= check_ineq_adj_eq(i
, j
, info
);
2415 } else if (any_ineq(&info
[j
], STATUS_ADJ_EQ
)) {
2416 change
= check_ineq_adj_eq(j
, i
, info
);
2417 } else if (any_ineq(&info
[i
], STATUS_ADJ_INEQ
) ||
2418 any_ineq(&info
[j
], STATUS_ADJ_INEQ
)) {
2419 change
= check_adj_ineq(i
, j
, info
);
2421 if (!any_eq(&info
[i
], STATUS_CUT
) &&
2422 !any_eq(&info
[j
], STATUS_CUT
))
2423 change
= check_facets(i
, j
, info
);
2424 if (change
== isl_change_none
)
2425 change
= check_wrap(i
, j
, info
);
2429 clear_status(&info
[i
]);
2430 clear_status(&info
[j
]);
2433 clear_status(&info
[i
]);
2434 clear_status(&info
[j
]);
2435 return isl_change_error
;
2438 /* Check if the union of the given pair of basic maps
2439 * can be represented by a single basic map.
2440 * If so, replace the pair by the single basic map and return
2441 * isl_change_drop_first, isl_change_drop_second or isl_change_fuse.
2442 * Otherwise, return isl_change_none.
2443 * The two basic maps are assumed to live in the same local space.
2445 static enum isl_change
coalesce_local_pair(int i
, int j
,
2446 struct isl_coalesce_info
*info
)
2448 init_status(&info
[i
]);
2449 init_status(&info
[j
]);
2450 return coalesce_local_pair_reuse(i
, j
, info
);
2453 /* Shift the integer division at position "div" of the basic map
2454 * represented by "info" by "shift".
2456 * That is, if the integer division has the form
2460 * then replace it by
2462 * floor((f(x) + shift * d)/d) - shift
2464 static isl_stat
shift_div(struct isl_coalesce_info
*info
, int div
,
2467 isl_size total
, n_div
;
2469 info
->bmap
= isl_basic_map_shift_div(info
->bmap
, div
, 0, shift
);
2471 return isl_stat_error
;
2473 total
= isl_basic_map_dim(info
->bmap
, isl_dim_all
);
2474 n_div
= isl_basic_map_dim(info
->bmap
, isl_dim_div
);
2475 if (total
< 0 || n_div
< 0)
2476 return isl_stat_error
;
2478 if (isl_tab_shift_var(info
->tab
, total
+ div
, shift
) < 0)
2479 return isl_stat_error
;
2484 /* If the integer division at position "div" is defined by an equality,
2485 * i.e., a stride constraint, then change the integer division expression
2486 * to have a constant term equal to zero.
2488 * Let the equality constraint be
2492 * The integer division expression is then typically of the form
2494 * a = floor((-f - c')/m)
2496 * The integer division is first shifted by t = floor(c/m),
2497 * turning the equality constraint into
2499 * c - m floor(c/m) + f + m a' = 0
2503 * (c mod m) + f + m a' = 0
2507 * a' = (-f - (c mod m))/m = floor((-f)/m)
2509 * because a' is an integer and 0 <= (c mod m) < m.
2510 * The constant term of a' can therefore be zeroed out,
2511 * but only if the integer division expression is of the expected form.
2513 static isl_stat
normalize_stride_div(struct isl_coalesce_info
*info
, int div
)
2515 isl_bool defined
, valid
;
2518 isl_int shift
, stride
;
2520 defined
= isl_basic_map_has_defining_equality(info
->bmap
, isl_dim_div
,
2523 return isl_stat_error
;
2527 return isl_stat_error
;
2528 valid
= isl_constraint_is_div_equality(c
, div
);
2529 isl_int_init(shift
);
2530 isl_int_init(stride
);
2531 isl_constraint_get_constant(c
, &shift
);
2532 isl_constraint_get_coefficient(c
, isl_dim_div
, div
, &stride
);
2533 isl_int_fdiv_q(shift
, shift
, stride
);
2534 r
= shift_div(info
, div
, shift
);
2535 isl_int_clear(stride
);
2536 isl_int_clear(shift
);
2537 isl_constraint_free(c
);
2538 if (r
< 0 || valid
< 0)
2539 return isl_stat_error
;
2542 info
->bmap
= isl_basic_map_set_div_expr_constant_num_si_inplace(
2543 info
->bmap
, div
, 0);
2545 return isl_stat_error
;
2549 /* The basic maps represented by "info1" and "info2" are known
2550 * to have the same number of integer divisions.
2551 * Check if pairs of integer divisions are equal to each other
2552 * despite the fact that they differ by a rational constant.
2554 * In particular, look for any pair of integer divisions that
2555 * only differ in their constant terms.
2556 * If either of these integer divisions is defined
2557 * by stride constraints, then modify it to have a zero constant term.
2558 * If both are defined by stride constraints then in the end they will have
2559 * the same (zero) constant term.
2561 static isl_stat
harmonize_stride_divs(struct isl_coalesce_info
*info1
,
2562 struct isl_coalesce_info
*info2
)
2567 n
= isl_basic_map_dim(info1
->bmap
, isl_dim_div
);
2569 return isl_stat_error
;
2570 for (i
= 0; i
< n
; ++i
) {
2571 isl_bool known
, harmonize
;
2573 known
= isl_basic_map_div_is_known(info1
->bmap
, i
);
2574 if (known
>= 0 && known
)
2575 known
= isl_basic_map_div_is_known(info2
->bmap
, i
);
2577 return isl_stat_error
;
2580 harmonize
= isl_basic_map_equal_div_expr_except_constant(
2581 info1
->bmap
, i
, info2
->bmap
, i
);
2583 return isl_stat_error
;
2586 if (normalize_stride_div(info1
, i
) < 0)
2587 return isl_stat_error
;
2588 if (normalize_stride_div(info2
, i
) < 0)
2589 return isl_stat_error
;
2595 /* If "shift" is an integer constant, then shift the integer division
2596 * at position "div" of the basic map represented by "info" by "shift".
2597 * If "shift" is not an integer constant, then do nothing.
2598 * If "shift" is equal to zero, then no shift needs to be performed either.
2600 * That is, if the integer division has the form
2604 * then replace it by
2606 * floor((f(x) + shift * d)/d) - shift
2608 static isl_stat
shift_if_cst_int(struct isl_coalesce_info
*info
, int div
,
2609 __isl_keep isl_aff
*shift
)
2616 cst
= isl_aff_is_cst(shift
);
2617 if (cst
< 0 || !cst
)
2618 return cst
< 0 ? isl_stat_error
: isl_stat_ok
;
2620 c
= isl_aff_get_constant_val(shift
);
2621 cst
= isl_val_is_int(c
);
2622 if (cst
>= 0 && cst
)
2623 cst
= isl_bool_not(isl_val_is_zero(c
));
2624 if (cst
< 0 || !cst
) {
2626 return cst
< 0 ? isl_stat_error
: isl_stat_ok
;
2630 r
= isl_val_get_num_isl_int(c
, &d
);
2632 r
= shift_div(info
, div
, d
);
2640 /* Check if some of the divs in the basic map represented by "info1"
2641 * are shifts of the corresponding divs in the basic map represented
2642 * by "info2", taking into account the equality constraints "eq1" of "info1"
2643 * and "eq2" of "info2". If so, align them with those of "info2".
2644 * "info1" and "info2" are assumed to have the same number
2645 * of integer divisions.
2647 * An integer division is considered to be a shift of another integer
2648 * division if, after simplification with respect to the equality
2649 * constraints of the other basic map, one is equal to the other
2652 * In particular, for each pair of integer divisions, if both are known,
2653 * have the same denominator and are not already equal to each other,
2654 * simplify each with respect to the equality constraints
2655 * of the other basic map. If the difference is an integer constant,
2656 * then move this difference outside.
2657 * That is, if, after simplification, one integer division is of the form
2659 * floor((f(x) + c_1)/d)
2661 * while the other is of the form
2663 * floor((f(x) + c_2)/d)
2665 * and n = (c_2 - c_1)/d is an integer, then replace the first
2666 * integer division by
2668 * floor((f_1(x) + c_1 + n * d)/d) - n,
2670 * where floor((f_1(x) + c_1 + n * d)/d) = floor((f2(x) + c_2)/d)
2671 * after simplification with respect to the equality constraints.
2673 static isl_stat
harmonize_divs_with_hulls(struct isl_coalesce_info
*info1
,
2674 struct isl_coalesce_info
*info2
, __isl_keep isl_basic_set
*eq1
,
2675 __isl_keep isl_basic_set
*eq2
)
2679 isl_local_space
*ls1
, *ls2
;
2681 total
= isl_basic_map_dim(info1
->bmap
, isl_dim_all
);
2683 return isl_stat_error
;
2684 ls1
= isl_local_space_wrap(isl_basic_map_get_local_space(info1
->bmap
));
2685 ls2
= isl_local_space_wrap(isl_basic_map_get_local_space(info2
->bmap
));
2686 for (i
= 0; i
< info1
->bmap
->n_div
; ++i
) {
2688 isl_aff
*div1
, *div2
;
2690 if (!isl_local_space_div_is_known(ls1
, i
) ||
2691 !isl_local_space_div_is_known(ls2
, i
))
2693 if (isl_int_ne(info1
->bmap
->div
[i
][0], info2
->bmap
->div
[i
][0]))
2695 if (isl_seq_eq(info1
->bmap
->div
[i
] + 1,
2696 info2
->bmap
->div
[i
] + 1, 1 + total
))
2698 div1
= isl_local_space_get_div(ls1
, i
);
2699 div2
= isl_local_space_get_div(ls2
, i
);
2700 div1
= isl_aff_substitute_equalities(div1
,
2701 isl_basic_set_copy(eq2
));
2702 div2
= isl_aff_substitute_equalities(div2
,
2703 isl_basic_set_copy(eq1
));
2704 div2
= isl_aff_sub(div2
, div1
);
2705 r
= shift_if_cst_int(info1
, i
, div2
);
2710 isl_local_space_free(ls1
);
2711 isl_local_space_free(ls2
);
2713 if (i
< info1
->bmap
->n_div
)
2714 return isl_stat_error
;
2718 /* Check if some of the divs in the basic map represented by "info1"
2719 * are shifts of the corresponding divs in the basic map represented
2720 * by "info2". If so, align them with those of "info2".
2721 * Only do this if "info1" and "info2" have the same number
2722 * of integer divisions.
2724 * An integer division is considered to be a shift of another integer
2725 * division if, after simplification with respect to the equality
2726 * constraints of the other basic map, one is equal to the other
2729 * First check if pairs of integer divisions are equal to each other
2730 * despite the fact that they differ by a rational constant.
2731 * If so, try and arrange for them to have the same constant term.
2733 * Then, extract the equality constraints and continue with
2734 * harmonize_divs_with_hulls.
2736 * If the equality constraints of both basic maps are the same,
2737 * then there is no need to perform any shifting since
2738 * the coefficients of the integer divisions should have been
2739 * reduced in the same way.
2741 static isl_stat
harmonize_divs(struct isl_coalesce_info
*info1
,
2742 struct isl_coalesce_info
*info2
)
2745 isl_basic_map
*bmap1
, *bmap2
;
2746 isl_basic_set
*eq1
, *eq2
;
2749 if (!info1
->bmap
|| !info2
->bmap
)
2750 return isl_stat_error
;
2752 if (info1
->bmap
->n_div
!= info2
->bmap
->n_div
)
2754 if (info1
->bmap
->n_div
== 0)
2757 if (harmonize_stride_divs(info1
, info2
) < 0)
2758 return isl_stat_error
;
2760 bmap1
= isl_basic_map_copy(info1
->bmap
);
2761 bmap2
= isl_basic_map_copy(info2
->bmap
);
2762 eq1
= isl_basic_map_wrap(isl_basic_map_plain_affine_hull(bmap1
));
2763 eq2
= isl_basic_map_wrap(isl_basic_map_plain_affine_hull(bmap2
));
2764 equal
= isl_basic_set_plain_is_equal(eq1
, eq2
);
2770 r
= harmonize_divs_with_hulls(info1
, info2
, eq1
, eq2
);
2771 isl_basic_set_free(eq1
);
2772 isl_basic_set_free(eq2
);
2777 /* Do the two basic maps live in the same local space, i.e.,
2778 * do they have the same (known) divs?
2779 * If either basic map has any unknown divs, then we can only assume
2780 * that they do not live in the same local space.
2782 static isl_bool
same_divs(__isl_keep isl_basic_map
*bmap1
,
2783 __isl_keep isl_basic_map
*bmap2
)
2789 if (!bmap1
|| !bmap2
)
2790 return isl_bool_error
;
2791 if (bmap1
->n_div
!= bmap2
->n_div
)
2792 return isl_bool_false
;
2794 if (bmap1
->n_div
== 0)
2795 return isl_bool_true
;
2797 known
= isl_basic_map_divs_known(bmap1
);
2798 if (known
< 0 || !known
)
2800 known
= isl_basic_map_divs_known(bmap2
);
2801 if (known
< 0 || !known
)
2804 total
= isl_basic_map_dim(bmap1
, isl_dim_all
);
2806 return isl_bool_error
;
2807 for (i
= 0; i
< bmap1
->n_div
; ++i
)
2808 if (!isl_seq_eq(bmap1
->div
[i
], bmap2
->div
[i
], 2 + total
))
2809 return isl_bool_false
;
2811 return isl_bool_true
;
2814 /* Assuming that "tab" contains the equality constraints and
2815 * the initial inequality constraints of "bmap", copy the remaining
2816 * inequality constraints of "bmap" to "Tab".
2818 static isl_stat
copy_ineq(struct isl_tab
*tab
, __isl_keep isl_basic_map
*bmap
)
2823 return isl_stat_error
;
2825 n_ineq
= tab
->n_con
- tab
->n_eq
;
2826 for (i
= n_ineq
; i
< bmap
->n_ineq
; ++i
)
2827 if (isl_tab_add_ineq(tab
, bmap
->ineq
[i
]) < 0)
2828 return isl_stat_error
;
2833 /* Description of an integer division that is added
2834 * during an expansion.
2835 * "pos" is the position of the corresponding variable.
2836 * "cst" indicates whether this integer division has a fixed value.
2837 * "val" contains the fixed value, if the value is fixed.
2839 struct isl_expanded
{
2845 /* For each of the "n" integer division variables "expanded",
2846 * if the variable has a fixed value, then add two inequality
2847 * constraints expressing the fixed value.
2848 * Otherwise, add the corresponding div constraints.
2849 * The caller is responsible for removing the div constraints
2850 * that it added for all these "n" integer divisions.
2852 * The div constraints and the pair of inequality constraints
2853 * forcing the fixed value cannot both be added for a given variable
2854 * as the combination may render some of the original constraints redundant.
2855 * These would then be ignored during the coalescing detection,
2856 * while they could remain in the fused result.
2858 * The two added inequality constraints are
2863 * with "a" the variable and "v" its fixed value.
2864 * The facet corresponding to one of these two constraints is selected
2865 * in the tableau to ensure that the pair of inequality constraints
2866 * is treated as an equality constraint.
2868 * The information in info->ineq is thrown away because it was
2869 * computed in terms of div constraints, while some of those
2870 * have now been replaced by these pairs of inequality constraints.
2872 static isl_stat
fix_constant_divs(struct isl_coalesce_info
*info
,
2873 int n
, struct isl_expanded
*expanded
)
2879 o_div
= isl_basic_map_offset(info
->bmap
, isl_dim_div
) - 1;
2880 ineq
= isl_vec_alloc(isl_tab_get_ctx(info
->tab
), 1 + info
->tab
->n_var
);
2882 return isl_stat_error
;
2883 isl_seq_clr(ineq
->el
+ 1, info
->tab
->n_var
);
2885 for (i
= 0; i
< n
; ++i
) {
2886 if (!expanded
[i
].cst
) {
2887 info
->bmap
= isl_basic_map_extend_constraints(
2889 info
->bmap
= isl_basic_map_add_div_constraints(
2890 info
->bmap
, expanded
[i
].pos
- o_div
);
2892 isl_int_set_si(ineq
->el
[1 + expanded
[i
].pos
], -1);
2893 isl_int_set(ineq
->el
[0], expanded
[i
].val
);
2894 info
->bmap
= isl_basic_map_add_ineq(info
->bmap
,
2896 isl_int_set_si(ineq
->el
[1 + expanded
[i
].pos
], 1);
2897 isl_int_neg(ineq
->el
[0], expanded
[i
].val
);
2898 info
->bmap
= isl_basic_map_add_ineq(info
->bmap
,
2900 isl_int_set_si(ineq
->el
[1 + expanded
[i
].pos
], 0);
2902 if (copy_ineq(info
->tab
, info
->bmap
) < 0)
2904 if (expanded
[i
].cst
&&
2905 isl_tab_select_facet(info
->tab
, info
->tab
->n_con
- 1) < 0)
2914 return i
< n
? isl_stat_error
: isl_stat_ok
;
2917 /* Insert the "n" integer division variables "expanded"
2918 * into info->tab and info->bmap and
2919 * update info->ineq with respect to the redundant constraints
2920 * in the resulting tableau.
2921 * "bmap" contains the result of this insertion in info->bmap,
2922 * while info->bmap is the original version
2923 * of "bmap", i.e., the one that corresponds to the current
2924 * state of info->tab. The number of constraints in info->bmap
2925 * is assumed to be the same as the number of constraints
2926 * in info->tab. This is required to be able to detect
2927 * the extra constraints in "bmap".
2929 * In particular, introduce extra variables corresponding
2930 * to the extra integer divisions and add the div constraints
2931 * that were added to "bmap" after info->tab was created
2933 * Furthermore, check if these extra integer divisions happen
2934 * to attain a fixed integer value in info->tab.
2935 * If so, replace the corresponding div constraints by pairs
2936 * of inequality constraints that fix these
2937 * integer divisions to their single integer values.
2938 * Replace info->bmap by "bmap" to match the changes to info->tab.
2939 * info->ineq was computed without a tableau and therefore
2940 * does not take into account the redundant constraints
2941 * in the tableau. Mark them here.
2942 * There is no need to check the newly added div constraints
2943 * since they cannot be redundant.
2944 * The redundancy check is not performed when constants have been discovered
2945 * since info->ineq is completely thrown away in this case.
2947 static isl_stat
tab_insert_divs(struct isl_coalesce_info
*info
,
2948 int n
, struct isl_expanded
*expanded
, __isl_take isl_basic_map
*bmap
)
2952 struct isl_tab_undo
*snap
;
2956 return isl_stat_error
;
2957 if (info
->bmap
->n_eq
+ info
->bmap
->n_ineq
!= info
->tab
->n_con
)
2958 isl_die(isl_basic_map_get_ctx(bmap
), isl_error_internal
,
2959 "original tableau does not correspond "
2960 "to original basic map", goto error
);
2962 if (isl_tab_extend_vars(info
->tab
, n
) < 0)
2964 if (isl_tab_extend_cons(info
->tab
, 2 * n
) < 0)
2967 for (i
= 0; i
< n
; ++i
) {
2968 if (isl_tab_insert_var(info
->tab
, expanded
[i
].pos
) < 0)
2972 snap
= isl_tab_snap(info
->tab
);
2974 n_ineq
= info
->tab
->n_con
- info
->tab
->n_eq
;
2975 if (copy_ineq(info
->tab
, bmap
) < 0)
2978 isl_basic_map_free(info
->bmap
);
2982 for (i
= 0; i
< n
; ++i
) {
2983 expanded
[i
].cst
= isl_tab_is_constant(info
->tab
,
2984 expanded
[i
].pos
, &expanded
[i
].val
);
2985 if (expanded
[i
].cst
< 0)
2986 return isl_stat_error
;
2987 if (expanded
[i
].cst
)
2992 if (isl_tab_rollback(info
->tab
, snap
) < 0)
2993 return isl_stat_error
;
2994 info
->bmap
= isl_basic_map_cow(info
->bmap
);
2995 info
->bmap
= isl_basic_map_free_inequality(info
->bmap
, 2 * n
);
2997 return isl_stat_error
;
2999 return fix_constant_divs(info
, n
, expanded
);
3002 n_eq
= info
->bmap
->n_eq
;
3003 for (i
= 0; i
< n_ineq
; ++i
) {
3004 if (isl_tab_is_redundant(info
->tab
, n_eq
+ i
))
3005 info
->ineq
[i
] = STATUS_REDUNDANT
;
3010 isl_basic_map_free(bmap
);
3011 return isl_stat_error
;
3014 /* Expand info->tab and info->bmap in the same way "bmap" was expanded
3015 * in isl_basic_map_expand_divs using the expansion "exp" and
3016 * update info->ineq with respect to the redundant constraints
3017 * in the resulting tableau. info->bmap is the original version
3018 * of "bmap", i.e., the one that corresponds to the current
3019 * state of info->tab. The number of constraints in info->bmap
3020 * is assumed to be the same as the number of constraints
3021 * in info->tab. This is required to be able to detect
3022 * the extra constraints in "bmap".
3024 * Extract the positions where extra local variables are introduced
3025 * from "exp" and call tab_insert_divs.
3027 static isl_stat
expand_tab(struct isl_coalesce_info
*info
, int *exp
,
3028 __isl_take isl_basic_map
*bmap
)
3031 struct isl_expanded
*expanded
;
3034 isl_size total
, n_div
;
3038 total
= isl_basic_map_dim(bmap
, isl_dim_all
);
3039 n_div
= isl_basic_map_dim(bmap
, isl_dim_div
);
3040 if (total
< 0 || n_div
< 0)
3041 return isl_stat_error
;
3042 pos
= total
- n_div
;
3043 extra_var
= total
- info
->tab
->n_var
;
3044 n
= n_div
- extra_var
;
3046 ctx
= isl_basic_map_get_ctx(bmap
);
3047 expanded
= isl_calloc_array(ctx
, struct isl_expanded
, extra_var
);
3048 if (extra_var
&& !expanded
)
3053 for (j
= 0; j
< n_div
; ++j
) {
3054 if (i
< n
&& exp
[i
] == j
) {
3058 expanded
[k
++].pos
= pos
+ j
;
3061 for (k
= 0; k
< extra_var
; ++k
)
3062 isl_int_init(expanded
[k
].val
);
3064 r
= tab_insert_divs(info
, extra_var
, expanded
, bmap
);
3066 for (k
= 0; k
< extra_var
; ++k
)
3067 isl_int_clear(expanded
[k
].val
);
3072 isl_basic_map_free(bmap
);
3073 return isl_stat_error
;
3076 /* Check if the union of the basic maps represented by info[i] and info[j]
3077 * can be represented by a single basic map,
3078 * after expanding the divs of info[i] to match those of info[j].
3079 * If so, replace the pair by the single basic map and return
3080 * isl_change_drop_first, isl_change_drop_second or isl_change_fuse.
3081 * Otherwise, return isl_change_none.
3083 * The caller has already checked for info[j] being a subset of info[i].
3084 * If some of the divs of info[j] are unknown, then the expanded info[i]
3085 * will not have the corresponding div constraints. The other patterns
3086 * therefore cannot apply. Skip the computation in this case.
3088 * The expansion is performed using the divs "div" and expansion "exp"
3089 * computed by the caller.
3090 * info[i].bmap has already been expanded and the result is passed in
3092 * The "eq" and "ineq" fields of info[i] reflect the status of
3093 * the constraints of the expanded "bmap" with respect to info[j].tab.
3094 * However, inequality constraints that are redundant in info[i].tab
3095 * have not yet been marked as such because no tableau was available.
3097 * Replace info[i].bmap by "bmap" and expand info[i].tab as well,
3098 * updating info[i].ineq with respect to the redundant constraints.
3099 * Then try and coalesce the expanded info[i] with info[j],
3100 * reusing the information in info[i].eq and info[i].ineq.
3101 * If this does not result in any coalescing or if it results in info[j]
3102 * getting dropped (which should not happen in practice, since the case
3103 * of info[j] being a subset of info[i] has already been checked by
3104 * the caller), then revert info[i] to its original state.
3106 static enum isl_change
coalesce_expand_tab_divs(__isl_take isl_basic_map
*bmap
,
3107 int i
, int j
, struct isl_coalesce_info
*info
, __isl_keep isl_mat
*div
,
3111 isl_basic_map
*bmap_i
;
3112 struct isl_tab_undo
*snap
;
3113 enum isl_change change
= isl_change_none
;
3115 known
= isl_basic_map_divs_known(info
[j
].bmap
);
3116 if (known
< 0 || !known
) {
3117 clear_status(&info
[i
]);
3118 isl_basic_map_free(bmap
);
3119 return known
< 0 ? isl_change_error
: isl_change_none
;
3122 bmap_i
= isl_basic_map_copy(info
[i
].bmap
);
3123 snap
= isl_tab_snap(info
[i
].tab
);
3124 if (expand_tab(&info
[i
], exp
, bmap
) < 0)
3125 change
= isl_change_error
;
3127 init_status(&info
[j
]);
3128 if (change
== isl_change_none
)
3129 change
= coalesce_local_pair_reuse(i
, j
, info
);
3131 clear_status(&info
[i
]);
3132 if (change
!= isl_change_none
&& change
!= isl_change_drop_second
) {
3133 isl_basic_map_free(bmap_i
);
3135 isl_basic_map_free(info
[i
].bmap
);
3136 info
[i
].bmap
= bmap_i
;
3138 if (isl_tab_rollback(info
[i
].tab
, snap
) < 0)
3139 change
= isl_change_error
;
3145 /* Check if the union of "bmap" and the basic map represented by info[j]
3146 * can be represented by a single basic map,
3147 * after expanding the divs of "bmap" to match those of info[j].
3148 * If so, replace the pair by the single basic map and return
3149 * isl_change_drop_first, isl_change_drop_second or isl_change_fuse.
3150 * Otherwise, return isl_change_none.
3152 * In particular, check if the expanded "bmap" contains the basic map
3153 * represented by the tableau info[j].tab.
3154 * The expansion is performed using the divs "div" and expansion "exp"
3155 * computed by the caller.
3156 * Then we check if all constraints of the expanded "bmap" are valid for
3159 * If "i" is not equal to -1, then "bmap" is equal to info[i].bmap.
3160 * In this case, the positions of the constraints of info[i].bmap
3161 * with respect to the basic map represented by info[j] are stored
3164 * If the expanded "bmap" does not contain the basic map
3165 * represented by the tableau info[j].tab and if "i" is not -1,
3166 * i.e., if the original "bmap" is info[i].bmap, then expand info[i].tab
3167 * as well and check if that results in coalescing.
3169 static enum isl_change
coalesce_with_expanded_divs(
3170 __isl_keep isl_basic_map
*bmap
, int i
, int j
,
3171 struct isl_coalesce_info
*info
, __isl_keep isl_mat
*div
, int *exp
)
3173 enum isl_change change
= isl_change_none
;
3174 struct isl_coalesce_info info_local
, *info_i
;
3176 info_i
= i
>= 0 ? &info
[i
] : &info_local
;
3177 init_status(info_i
);
3178 bmap
= isl_basic_map_copy(bmap
);
3179 bmap
= isl_basic_map_expand_divs(bmap
, isl_mat_copy(div
), exp
);
3180 bmap
= isl_basic_map_mark_final(bmap
);
3185 info_local
.bmap
= bmap
;
3186 info_i
->eq
= eq_status_in(bmap
, info
[j
].tab
);
3187 if (bmap
->n_eq
&& !info_i
->eq
)
3189 if (any_eq(info_i
, STATUS_ERROR
))
3191 if (any_eq(info_i
, STATUS_SEPARATE
))
3194 info_i
->ineq
= ineq_status_in(bmap
, NULL
, info
[j
].tab
);
3195 if (bmap
->n_ineq
&& !info_i
->ineq
)
3197 if (any_ineq(info_i
, STATUS_ERROR
))
3199 if (any_ineq(info_i
, STATUS_SEPARATE
))
3202 if (all(info_i
->eq
, 2 * bmap
->n_eq
, STATUS_VALID
) &&
3203 all(info_i
->ineq
, bmap
->n_ineq
, STATUS_VALID
)) {
3205 change
= isl_change_drop_second
;
3208 if (change
== isl_change_none
&& i
!= -1)
3209 return coalesce_expand_tab_divs(bmap
, i
, j
, info
, div
, exp
);
3212 isl_basic_map_free(bmap
);
3213 clear_status(info_i
);
3216 isl_basic_map_free(bmap
);
3217 clear_status(info_i
);
3218 return isl_change_error
;
3221 /* Check if the union of "bmap_i" and the basic map represented by info[j]
3222 * can be represented by a single basic map,
3223 * after aligning the divs of "bmap_i" to match those of info[j].
3224 * If so, replace the pair by the single basic map and return
3225 * isl_change_drop_first, isl_change_drop_second or isl_change_fuse.
3226 * Otherwise, return isl_change_none.
3228 * In particular, check if "bmap_i" contains the basic map represented by
3229 * info[j] after aligning the divs of "bmap_i" to those of info[j].
3230 * Note that this can only succeed if the number of divs of "bmap_i"
3231 * is smaller than (or equal to) the number of divs of info[j].
3233 * We first check if the divs of "bmap_i" are all known and form a subset
3234 * of those of info[j].bmap. If so, we pass control over to
3235 * coalesce_with_expanded_divs.
3237 * If "i" is not equal to -1, then "bmap" is equal to info[i].bmap.
3239 static enum isl_change
coalesce_after_aligning_divs(
3240 __isl_keep isl_basic_map
*bmap_i
, int i
, int j
,
3241 struct isl_coalesce_info
*info
)
3244 isl_mat
*div_i
, *div_j
, *div
;
3248 enum isl_change change
;
3250 known
= isl_basic_map_divs_known(bmap_i
);
3252 return isl_change_error
;
3254 return isl_change_none
;
3256 ctx
= isl_basic_map_get_ctx(bmap_i
);
3258 div_i
= isl_basic_map_get_divs(bmap_i
);
3259 div_j
= isl_basic_map_get_divs(info
[j
].bmap
);
3261 if (!div_i
|| !div_j
)
3264 exp1
= isl_alloc_array(ctx
, int, div_i
->n_row
);
3265 exp2
= isl_alloc_array(ctx
, int, div_j
->n_row
);
3266 if ((div_i
->n_row
&& !exp1
) || (div_j
->n_row
&& !exp2
))
3269 div
= isl_merge_divs(div_i
, div_j
, exp1
, exp2
);
3273 if (div
->n_row
== div_j
->n_row
)
3274 change
= coalesce_with_expanded_divs(bmap_i
,
3275 i
, j
, info
, div
, exp1
);
3277 change
= isl_change_none
;
3281 isl_mat_free(div_i
);
3282 isl_mat_free(div_j
);
3289 isl_mat_free(div_i
);
3290 isl_mat_free(div_j
);
3293 return isl_change_error
;
3296 /* Check if basic map "j" is a subset of basic map "i" after
3297 * exploiting the extra equalities of "j" to simplify the divs of "i".
3298 * If so, remove basic map "j" and return isl_change_drop_second.
3300 * If "j" does not have any equalities or if they are the same
3301 * as those of "i", then we cannot exploit them to simplify the divs.
3302 * Similarly, if there are no divs in "i", then they cannot be simplified.
3303 * If, on the other hand, the affine hulls of "i" and "j" do not intersect,
3304 * then "j" cannot be a subset of "i".
3306 * Otherwise, we intersect "i" with the affine hull of "j" and then
3307 * check if "j" is a subset of the result after aligning the divs.
3308 * If so, then "j" is definitely a subset of "i" and can be removed.
3309 * Note that if after intersection with the affine hull of "j".
3310 * "i" still has more divs than "j", then there is no way we can
3311 * align the divs of "i" to those of "j".
3313 static enum isl_change
coalesce_subset_with_equalities(int i
, int j
,
3314 struct isl_coalesce_info
*info
)
3316 isl_basic_map
*hull_i
, *hull_j
, *bmap_i
;
3318 enum isl_change change
;
3320 if (info
[j
].bmap
->n_eq
== 0)
3321 return isl_change_none
;
3322 if (info
[i
].bmap
->n_div
== 0)
3323 return isl_change_none
;
3325 hull_i
= isl_basic_map_copy(info
[i
].bmap
);
3326 hull_i
= isl_basic_map_plain_affine_hull(hull_i
);
3327 hull_j
= isl_basic_map_copy(info
[j
].bmap
);
3328 hull_j
= isl_basic_map_plain_affine_hull(hull_j
);
3330 hull_j
= isl_basic_map_intersect(hull_j
, isl_basic_map_copy(hull_i
));
3331 equal
= isl_basic_map_plain_is_equal(hull_i
, hull_j
);
3332 empty
= isl_basic_map_plain_is_empty(hull_j
);
3333 isl_basic_map_free(hull_i
);
3335 if (equal
< 0 || equal
|| empty
< 0 || empty
) {
3336 isl_basic_map_free(hull_j
);
3337 if (equal
< 0 || empty
< 0)
3338 return isl_change_error
;
3339 return isl_change_none
;
3342 bmap_i
= isl_basic_map_copy(info
[i
].bmap
);
3343 bmap_i
= isl_basic_map_intersect(bmap_i
, hull_j
);
3345 return isl_change_error
;
3347 if (bmap_i
->n_div
> info
[j
].bmap
->n_div
) {
3348 isl_basic_map_free(bmap_i
);
3349 return isl_change_none
;
3352 change
= coalesce_after_aligning_divs(bmap_i
, -1, j
, info
);
3354 isl_basic_map_free(bmap_i
);
3359 /* Check if the union of and the basic maps represented by info[i] and info[j]
3360 * can be represented by a single basic map, by aligning or equating
3361 * their integer divisions.
3362 * If so, replace the pair by the single basic map and return
3363 * isl_change_drop_first, isl_change_drop_second or isl_change_fuse.
3364 * Otherwise, return isl_change_none.
3366 * Note that we only perform any test if the number of divs is different
3367 * in the two basic maps. In case the number of divs is the same,
3368 * we have already established that the divs are different
3369 * in the two basic maps.
3370 * In particular, if the number of divs of basic map i is smaller than
3371 * the number of divs of basic map j, then we check if j is a subset of i
3374 static enum isl_change
coalesce_divs(int i
, int j
,
3375 struct isl_coalesce_info
*info
)
3377 enum isl_change change
= isl_change_none
;
3379 if (info
[i
].bmap
->n_div
< info
[j
].bmap
->n_div
)
3380 change
= coalesce_after_aligning_divs(info
[i
].bmap
, i
, j
, info
);
3381 if (change
!= isl_change_none
)
3384 if (info
[j
].bmap
->n_div
< info
[i
].bmap
->n_div
)
3385 change
= coalesce_after_aligning_divs(info
[j
].bmap
, j
, i
, info
);
3386 if (change
!= isl_change_none
)
3387 return invert_change(change
);
3389 change
= coalesce_subset_with_equalities(i
, j
, info
);
3390 if (change
!= isl_change_none
)
3393 change
= coalesce_subset_with_equalities(j
, i
, info
);
3394 if (change
!= isl_change_none
)
3395 return invert_change(change
);
3397 return isl_change_none
;
3400 /* Does "bmap" involve any divs that themselves refer to divs?
3402 static isl_bool
has_nested_div(__isl_keep isl_basic_map
*bmap
)
3408 total
= isl_basic_map_dim(bmap
, isl_dim_all
);
3409 n_div
= isl_basic_map_dim(bmap
, isl_dim_div
);
3410 if (total
< 0 || n_div
< 0)
3411 return isl_bool_error
;
3414 for (i
= 0; i
< n_div
; ++i
)
3415 if (isl_seq_first_non_zero(bmap
->div
[i
] + 2 + total
,
3417 return isl_bool_true
;
3419 return isl_bool_false
;
3422 /* Return a list of affine expressions, one for each integer division
3423 * in "bmap_i". For each integer division that also appears in "bmap_j",
3424 * the affine expression is set to NaN. The number of NaNs in the list
3425 * is equal to the number of integer divisions in "bmap_j".
3426 * For the other integer divisions of "bmap_i", the corresponding
3427 * element in the list is a purely affine expression equal to the integer
3428 * division in "hull".
3429 * If no such list can be constructed, then the number of elements
3430 * in the returned list is smaller than the number of integer divisions
3433 static __isl_give isl_aff_list
*set_up_substitutions(
3434 __isl_keep isl_basic_map
*bmap_i
, __isl_keep isl_basic_map
*bmap_j
,
3435 __isl_take isl_basic_map
*hull
)
3437 isl_size n_div_i
, n_div_j
, total
;
3439 isl_local_space
*ls
;
3440 isl_basic_set
*wrap_hull
;
3445 n_div_i
= isl_basic_map_dim(bmap_i
, isl_dim_div
);
3446 n_div_j
= isl_basic_map_dim(bmap_j
, isl_dim_div
);
3447 total
= isl_basic_map_dim(bmap_i
, isl_dim_all
);
3448 if (!hull
|| n_div_i
< 0 || n_div_j
< 0 || total
< 0)
3451 ctx
= isl_basic_map_get_ctx(hull
);
3454 ls
= isl_basic_map_get_local_space(bmap_i
);
3455 ls
= isl_local_space_wrap(ls
);
3456 wrap_hull
= isl_basic_map_wrap(hull
);
3458 aff_nan
= isl_aff_nan_on_domain(isl_local_space_copy(ls
));
3459 list
= isl_aff_list_alloc(ctx
, n_div_i
);
3462 for (i
= 0; i
< n_div_i
; ++i
) {
3467 isl_basic_map_equal_div_expr_part(bmap_i
, i
, bmap_j
, j
,
3470 list
= isl_aff_list_add(list
, isl_aff_copy(aff_nan
));
3473 if (n_div_i
- i
<= n_div_j
- j
)
3476 aff
= isl_local_space_get_div(ls
, i
);
3477 aff
= isl_aff_substitute_equalities(aff
,
3478 isl_basic_set_copy(wrap_hull
));
3479 aff
= isl_aff_floor(aff
);
3480 n_div
= isl_aff_dim(aff
, isl_dim_div
);
3488 list
= isl_aff_list_add(list
, aff
);
3491 isl_aff_free(aff_nan
);
3492 isl_local_space_free(ls
);
3493 isl_basic_set_free(wrap_hull
);
3497 isl_aff_free(aff_nan
);
3498 isl_local_space_free(ls
);
3499 isl_basic_set_free(wrap_hull
);
3500 isl_aff_list_free(list
);
3504 /* Add variables to info->bmap and info->tab corresponding to the elements
3505 * in "list" that are not set to NaN.
3506 * "extra_var" is the number of these elements.
3507 * "dim" is the offset in the variables of "tab" where we should
3508 * start considering the elements in "list".
3509 * When this function returns, the total number of variables in "tab"
3510 * is equal to "dim" plus the number of elements in "list".
3512 * The newly added existentially quantified variables are not given
3513 * an explicit representation because the corresponding div constraints
3514 * do not appear in info->bmap. These constraints are not added
3515 * to info->bmap because for internal consistency, they would need to
3516 * be added to info->tab as well, where they could combine with the equality
3517 * that is added later to result in constraints that do not hold
3518 * in the original input.
3520 static isl_stat
add_sub_vars(struct isl_coalesce_info
*info
,
3521 __isl_keep isl_aff_list
*list
, int dim
, int extra_var
)
3527 space
= isl_basic_map_get_space(info
->bmap
);
3528 info
->bmap
= isl_basic_map_cow(info
->bmap
);
3529 info
->bmap
= isl_basic_map_extend_space(info
->bmap
, space
,
3531 n
= isl_aff_list_n_aff(list
);
3532 if (!info
->bmap
|| n
< 0)
3533 return isl_stat_error
;
3534 for (i
= 0; i
< n
; ++i
) {
3538 aff
= isl_aff_list_get_aff(list
, i
);
3539 is_nan
= isl_aff_is_nan(aff
);
3542 return isl_stat_error
;
3546 if (isl_tab_insert_var(info
->tab
, dim
+ i
) < 0)
3547 return isl_stat_error
;
3548 d
= isl_basic_map_alloc_div(info
->bmap
);
3550 return isl_stat_error
;
3551 info
->bmap
= isl_basic_map_mark_div_unknown(info
->bmap
, d
);
3552 for (j
= d
; j
> i
; --j
)
3553 info
->bmap
= isl_basic_map_swap_div(info
->bmap
,
3556 return isl_stat_error
;
3562 /* For each element in "list" that is not set to NaN, fix the corresponding
3563 * variable in "tab" to the purely affine expression defined by the element.
3564 * "dim" is the offset in the variables of "tab" where we should
3565 * start considering the elements in "list".
3567 * This function assumes that a sufficient number of rows and
3568 * elements in the constraint array are available in the tableau.
3570 static isl_stat
add_sub_equalities(struct isl_tab
*tab
,
3571 __isl_keep isl_aff_list
*list
, int dim
)
3579 n
= isl_aff_list_n_aff(list
);
3581 return isl_stat_error
;
3583 ctx
= isl_tab_get_ctx(tab
);
3584 sub
= isl_vec_alloc(ctx
, 1 + dim
+ n
);
3586 return isl_stat_error
;
3587 isl_seq_clr(sub
->el
+ 1 + dim
, n
);
3589 for (i
= 0; i
< n
; ++i
) {
3590 aff
= isl_aff_list_get_aff(list
, i
);
3593 if (isl_aff_is_nan(aff
)) {
3597 isl_seq_cpy(sub
->el
, aff
->v
->el
+ 1, 1 + dim
);
3598 isl_int_neg(sub
->el
[1 + dim
+ i
], aff
->v
->el
[0]);
3599 if (isl_tab_add_eq(tab
, sub
->el
) < 0)
3601 isl_int_set_si(sub
->el
[1 + dim
+ i
], 0);
3610 return isl_stat_error
;
3613 /* Add variables to info->tab and info->bmap corresponding to the elements
3614 * in "list" that are not set to NaN. The value of the added variable
3615 * in info->tab is fixed to the purely affine expression defined by the element.
3616 * "dim" is the offset in the variables of info->tab where we should
3617 * start considering the elements in "list".
3618 * When this function returns, the total number of variables in info->tab
3619 * is equal to "dim" plus the number of elements in "list".
3621 static isl_stat
add_subs(struct isl_coalesce_info
*info
,
3622 __isl_keep isl_aff_list
*list
, int dim
)
3627 n
= isl_aff_list_n_aff(list
);
3629 return isl_stat_error
;
3631 extra_var
= n
- (info
->tab
->n_var
- dim
);
3633 if (isl_tab_extend_vars(info
->tab
, extra_var
) < 0)
3634 return isl_stat_error
;
3635 if (isl_tab_extend_cons(info
->tab
, 2 * extra_var
) < 0)
3636 return isl_stat_error
;
3637 if (add_sub_vars(info
, list
, dim
, extra_var
) < 0)
3638 return isl_stat_error
;
3640 return add_sub_equalities(info
->tab
, list
, dim
);
3643 /* Coalesce basic map "j" into basic map "i" after adding the extra integer
3644 * divisions in "i" but not in "j" to basic map "j", with values
3645 * specified by "list". The total number of elements in "list"
3646 * is equal to the number of integer divisions in "i", while the number
3647 * of NaN elements in the list is equal to the number of integer divisions
3650 * If no coalescing can be performed, then we need to revert basic map "j"
3651 * to its original state. We do the same if basic map "i" gets dropped
3652 * during the coalescing, even though this should not happen in practice
3653 * since we have already checked for "j" being a subset of "i"
3654 * before we reach this stage.
3656 static enum isl_change
coalesce_with_subs(int i
, int j
,
3657 struct isl_coalesce_info
*info
, __isl_keep isl_aff_list
*list
)
3659 isl_basic_map
*bmap_j
;
3660 struct isl_tab_undo
*snap
;
3661 isl_size dim
, n_div
;
3662 enum isl_change change
;
3664 bmap_j
= isl_basic_map_copy(info
[j
].bmap
);
3665 snap
= isl_tab_snap(info
[j
].tab
);
3667 dim
= isl_basic_map_dim(bmap_j
, isl_dim_all
);
3668 n_div
= isl_basic_map_dim(bmap_j
, isl_dim_div
);
3669 if (dim
< 0 || n_div
< 0)
3672 if (add_subs(&info
[j
], list
, dim
) < 0)
3675 change
= coalesce_local_pair(i
, j
, info
);
3676 if (change
!= isl_change_none
&& change
!= isl_change_drop_first
) {
3677 isl_basic_map_free(bmap_j
);
3679 isl_basic_map_free(info
[j
].bmap
);
3680 info
[j
].bmap
= bmap_j
;
3682 if (isl_tab_rollback(info
[j
].tab
, snap
) < 0)
3683 return isl_change_error
;
3688 isl_basic_map_free(bmap_j
);
3689 return isl_change_error
;
3692 /* Check if we can coalesce basic map "j" into basic map "i" after copying
3693 * those extra integer divisions in "i" that can be simplified away
3694 * using the extra equalities in "j".
3695 * All divs are assumed to be known and not contain any nested divs.
3697 * We first check if there are any extra equalities in "j" that we
3698 * can exploit. Then we check if every integer division in "i"
3699 * either already appears in "j" or can be simplified using the
3700 * extra equalities to a purely affine expression.
3701 * If these tests succeed, then we try to coalesce the two basic maps
3702 * by introducing extra dimensions in "j" corresponding to
3703 * the extra integer divisions "i" fixed to the corresponding
3704 * purely affine expression.
3706 static enum isl_change
check_coalesce_into_eq(int i
, int j
,
3707 struct isl_coalesce_info
*info
)
3709 isl_size n_div_i
, n_div_j
, n
;
3710 isl_basic_map
*hull_i
, *hull_j
;
3711 isl_bool equal
, empty
;
3713 enum isl_change change
;
3715 n_div_i
= isl_basic_map_dim(info
[i
].bmap
, isl_dim_div
);
3716 n_div_j
= isl_basic_map_dim(info
[j
].bmap
, isl_dim_div
);
3717 if (n_div_i
< 0 || n_div_j
< 0)
3718 return isl_change_error
;
3719 if (n_div_i
<= n_div_j
)
3720 return isl_change_none
;
3721 if (info
[j
].bmap
->n_eq
== 0)
3722 return isl_change_none
;
3724 hull_i
= isl_basic_map_copy(info
[i
].bmap
);
3725 hull_i
= isl_basic_map_plain_affine_hull(hull_i
);
3726 hull_j
= isl_basic_map_copy(info
[j
].bmap
);
3727 hull_j
= isl_basic_map_plain_affine_hull(hull_j
);
3729 hull_j
= isl_basic_map_intersect(hull_j
, isl_basic_map_copy(hull_i
));
3730 equal
= isl_basic_map_plain_is_equal(hull_i
, hull_j
);
3731 empty
= isl_basic_map_plain_is_empty(hull_j
);
3732 isl_basic_map_free(hull_i
);
3734 if (equal
< 0 || empty
< 0)
3736 if (equal
|| empty
) {
3737 isl_basic_map_free(hull_j
);
3738 return isl_change_none
;
3741 list
= set_up_substitutions(info
[i
].bmap
, info
[j
].bmap
, hull_j
);
3743 return isl_change_error
;
3744 n
= isl_aff_list_n_aff(list
);
3746 change
= isl_change_error
;
3747 else if (n
< n_div_i
)
3748 change
= isl_change_none
;
3750 change
= coalesce_with_subs(i
, j
, info
, list
);
3752 isl_aff_list_free(list
);
3756 isl_basic_map_free(hull_j
);
3757 return isl_change_error
;
3760 /* Check if we can coalesce basic maps "i" and "j" after copying
3761 * those extra integer divisions in one of the basic maps that can
3762 * be simplified away using the extra equalities in the other basic map.
3763 * We require all divs to be known in both basic maps.
3764 * Furthermore, to simplify the comparison of div expressions,
3765 * we do not allow any nested integer divisions.
3767 static enum isl_change
check_coalesce_eq(int i
, int j
,
3768 struct isl_coalesce_info
*info
)
3770 isl_bool known
, nested
;
3771 enum isl_change change
;
3773 known
= isl_basic_map_divs_known(info
[i
].bmap
);
3774 if (known
< 0 || !known
)
3775 return known
< 0 ? isl_change_error
: isl_change_none
;
3776 known
= isl_basic_map_divs_known(info
[j
].bmap
);
3777 if (known
< 0 || !known
)
3778 return known
< 0 ? isl_change_error
: isl_change_none
;
3779 nested
= has_nested_div(info
[i
].bmap
);
3780 if (nested
< 0 || nested
)
3781 return nested
< 0 ? isl_change_error
: isl_change_none
;
3782 nested
= has_nested_div(info
[j
].bmap
);
3783 if (nested
< 0 || nested
)
3784 return nested
< 0 ? isl_change_error
: isl_change_none
;
3786 change
= check_coalesce_into_eq(i
, j
, info
);
3787 if (change
!= isl_change_none
)
3789 change
= check_coalesce_into_eq(j
, i
, info
);
3790 if (change
!= isl_change_none
)
3791 return invert_change(change
);
3793 return isl_change_none
;
3796 /* Check if the union of the given pair of basic maps
3797 * can be represented by a single basic map.
3798 * If so, replace the pair by the single basic map and return
3799 * isl_change_drop_first, isl_change_drop_second or isl_change_fuse.
3800 * Otherwise, return isl_change_none.
3802 * We first check if the two basic maps live in the same local space,
3803 * after aligning the divs that differ by only an integer constant.
3804 * If so, we do the complete check. Otherwise, we check if they have
3805 * the same number of integer divisions and can be coalesced, if one is
3806 * an obvious subset of the other or if the extra integer divisions
3807 * of one basic map can be simplified away using the extra equalities
3808 * of the other basic map.
3810 * Note that trying to coalesce pairs of disjuncts with the same
3811 * number, but different local variables may drop the explicit
3812 * representation of some of these local variables.
3813 * This operation is therefore not performed when
3814 * the "coalesce_preserve_locals" option is set.
3816 static enum isl_change
coalesce_pair(int i
, int j
,
3817 struct isl_coalesce_info
*info
)
3821 enum isl_change change
;
3824 if (harmonize_divs(&info
[i
], &info
[j
]) < 0)
3825 return isl_change_error
;
3826 same
= same_divs(info
[i
].bmap
, info
[j
].bmap
);
3828 return isl_change_error
;
3830 return coalesce_local_pair(i
, j
, info
);
3832 ctx
= isl_basic_map_get_ctx(info
[i
].bmap
);
3833 preserve
= isl_options_get_coalesce_preserve_locals(ctx
);
3834 if (!preserve
&& info
[i
].bmap
->n_div
== info
[j
].bmap
->n_div
) {
3835 change
= coalesce_local_pair(i
, j
, info
);
3836 if (change
!= isl_change_none
)
3840 change
= coalesce_divs(i
, j
, info
);
3841 if (change
!= isl_change_none
)
3844 return check_coalesce_eq(i
, j
, info
);
3847 /* Return the maximum of "a" and "b".
3849 static int isl_max(int a
, int b
)
3851 return a
> b
? a
: b
;
3854 /* Pairwise coalesce the basic maps in the range [start1, end1[ of "info"
3855 * with those in the range [start2, end2[, skipping basic maps
3856 * that have been removed (either before or within this function).
3858 * For each basic map i in the first range, we check if it can be coalesced
3859 * with respect to any previously considered basic map j in the second range.
3860 * If i gets dropped (because it was a subset of some j), then
3861 * we can move on to the next basic map.
3862 * If j gets dropped, we need to continue checking against the other
3863 * previously considered basic maps.
3864 * If the two basic maps got fused, then we recheck the fused basic map
3865 * against the previously considered basic maps, starting at i + 1
3866 * (even if start2 is greater than i + 1).
3868 static int coalesce_range(isl_ctx
*ctx
, struct isl_coalesce_info
*info
,
3869 int start1
, int end1
, int start2
, int end2
)
3873 for (i
= end1
- 1; i
>= start1
; --i
) {
3874 if (info
[i
].removed
)
3876 for (j
= isl_max(i
+ 1, start2
); j
< end2
; ++j
) {
3877 enum isl_change changed
;
3879 if (info
[j
].removed
)
3881 if (info
[i
].removed
)
3882 isl_die(ctx
, isl_error_internal
,
3883 "basic map unexpectedly removed",
3885 changed
= coalesce_pair(i
, j
, info
);
3887 case isl_change_error
:
3889 case isl_change_none
:
3890 case isl_change_drop_second
:
3892 case isl_change_drop_first
:
3895 case isl_change_fuse
:
3905 /* Pairwise coalesce the basic maps described by the "n" elements of "info".
3907 * We consider groups of basic maps that live in the same apparent
3908 * affine hull and we first coalesce within such a group before we
3909 * coalesce the elements in the group with elements of previously
3910 * considered groups. If a fuse happens during the second phase,
3911 * then we also reconsider the elements within the group.
3913 static int coalesce(isl_ctx
*ctx
, int n
, struct isl_coalesce_info
*info
)
3917 for (end
= n
; end
> 0; end
= start
) {
3919 while (start
>= 1 &&
3920 info
[start
- 1].hull_hash
== info
[start
].hull_hash
)
3922 if (coalesce_range(ctx
, info
, start
, end
, start
, end
) < 0)
3924 if (coalesce_range(ctx
, info
, start
, end
, end
, n
) < 0)
3931 /* Update the basic maps in "map" based on the information in "info".
3932 * In particular, remove the basic maps that have been marked removed and
3933 * update the others based on the information in the corresponding tableau.
3934 * Since we detected implicit equalities without calling
3935 * isl_basic_map_gauss, we need to do it now.
3936 * Also call isl_basic_map_simplify if we may have lost the definition
3937 * of one or more integer divisions.
3938 * If a basic map is still equal to the one from which the corresponding "info"
3939 * entry was created, then redundant constraint and
3940 * implicit equality constraint detection have been performed
3941 * on the corresponding tableau and the basic map can be marked as such.
3943 static __isl_give isl_map
*update_basic_maps(__isl_take isl_map
*map
,
3944 int n
, struct isl_coalesce_info
*info
)
3951 for (i
= n
- 1; i
>= 0; --i
) {
3952 if (info
[i
].removed
) {
3953 isl_basic_map_free(map
->p
[i
]);
3954 if (i
!= map
->n
- 1)
3955 map
->p
[i
] = map
->p
[map
->n
- 1];
3960 info
[i
].bmap
= isl_basic_map_update_from_tab(info
[i
].bmap
,
3962 info
[i
].bmap
= isl_basic_map_gauss(info
[i
].bmap
, NULL
);
3963 if (info
[i
].simplify
)
3964 info
[i
].bmap
= isl_basic_map_simplify(info
[i
].bmap
);
3965 info
[i
].bmap
= isl_basic_map_finalize(info
[i
].bmap
);
3967 return isl_map_free(map
);
3968 if (!info
[i
].modified
) {
3969 ISL_F_SET(info
[i
].bmap
, ISL_BASIC_MAP_NO_IMPLICIT
);
3970 ISL_F_SET(info
[i
].bmap
, ISL_BASIC_MAP_NO_REDUNDANT
);
3972 isl_basic_map_free(map
->p
[i
]);
3973 map
->p
[i
] = info
[i
].bmap
;
3974 info
[i
].bmap
= NULL
;
3980 /* For each pair of basic maps in the map, check if the union of the two
3981 * can be represented by a single basic map.
3982 * If so, replace the pair by the single basic map and start over.
3984 * We factor out any (hidden) common factor from the constraint
3985 * coefficients to improve the detection of adjacent constraints.
3986 * Note that this function does not call isl_basic_map_gauss,
3987 * but it does make sure that only a single copy of the basic map
3988 * is affected. This means that isl_basic_map_gauss may have
3989 * to be called at the end of the computation (in update_basic_maps)
3990 * on this single copy to ensure that
3991 * the basic maps are not left in an unexpected state.
3993 * Since we are constructing the tableaus of the basic maps anyway,
3994 * we exploit them to detect implicit equalities and redundant constraints.
3995 * This also helps the coalescing as it can ignore the redundant constraints.
3996 * In order to avoid confusion, we make all implicit equalities explicit
3997 * in the basic maps. If the basic map only has a single reference
3998 * (this happens in particular if it was modified by
3999 * isl_basic_map_reduce_coefficients), then isl_basic_map_gauss
4000 * does not get called on the result. The call to
4001 * isl_basic_map_gauss in update_basic_maps resolves this as well.
4002 * For each basic map, we also compute the hash of the apparent affine hull
4003 * for use in coalesce.
4005 __isl_give isl_map
*isl_map_coalesce(__isl_take isl_map
*map
)
4010 struct isl_coalesce_info
*info
= NULL
;
4012 map
= isl_map_remove_empty_parts(map
);
4019 ctx
= isl_map_get_ctx(map
);
4020 map
= isl_map_sort_divs(map
);
4021 map
= isl_map_cow(map
);
4028 info
= isl_calloc_array(map
->ctx
, struct isl_coalesce_info
, n
);
4032 for (i
= 0; i
< map
->n
; ++i
) {
4033 map
->p
[i
] = isl_basic_map_reduce_coefficients(map
->p
[i
]);
4036 info
[i
].bmap
= isl_basic_map_copy(map
->p
[i
]);
4037 info
[i
].tab
= isl_tab_from_basic_map(info
[i
].bmap
, 0);
4040 if (!ISL_F_ISSET(info
[i
].bmap
, ISL_BASIC_MAP_NO_IMPLICIT
))
4041 if (isl_tab_detect_implicit_equalities(info
[i
].tab
) < 0)
4043 info
[i
].bmap
= isl_tab_make_equalities_explicit(info
[i
].tab
,
4047 if (!ISL_F_ISSET(info
[i
].bmap
, ISL_BASIC_MAP_NO_REDUNDANT
))
4048 if (isl_tab_detect_redundant(info
[i
].tab
) < 0)
4050 if (coalesce_info_set_hull_hash(&info
[i
]) < 0)
4053 for (i
= map
->n
- 1; i
>= 0; --i
)
4054 if (info
[i
].tab
->empty
)
4057 if (coalesce(ctx
, n
, info
) < 0)
4060 map
= update_basic_maps(map
, n
, info
);
4062 clear_coalesce_info(n
, info
);
4066 clear_coalesce_info(n
, info
);
4071 /* For each pair of basic sets in the set, check if the union of the two
4072 * can be represented by a single basic set.
4073 * If so, replace the pair by the single basic set and start over.
4075 struct isl_set
*isl_set_coalesce(struct isl_set
*set
)
4077 return set_from_map(isl_map_coalesce(set_to_map(set
)));