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
)
69 int *eq
= isl_calloc_array(bmap_i
->ctx
, int, 2 * bmap_i
->n_eq
);
75 dim
= isl_basic_map_total_dim(bmap_i
);
76 for (k
= 0; k
< bmap_i
->n_eq
; ++k
) {
77 for (l
= 0; l
< 2; ++l
) {
78 isl_seq_neg(bmap_i
->eq
[k
], bmap_i
->eq
[k
], 1+dim
);
79 eq
[2 * k
+ l
] = status_in(bmap_i
->eq
[k
], tab_j
);
80 if (eq
[2 * k
+ l
] == STATUS_ERROR
)
91 /* Compute the position of the inequalities of basic map "bmap_i"
92 * (also represented by "tab_i", if not NULL) with respect to the basic map
93 * represented by "tab_j".
95 static int *ineq_status_in(__isl_keep isl_basic_map
*bmap_i
,
96 struct isl_tab
*tab_i
, struct isl_tab
*tab_j
)
99 unsigned n_eq
= bmap_i
->n_eq
;
100 int *ineq
= isl_calloc_array(bmap_i
->ctx
, int, bmap_i
->n_ineq
);
105 for (k
= 0; k
< bmap_i
->n_ineq
; ++k
) {
106 if (tab_i
&& isl_tab_is_redundant(tab_i
, n_eq
+ k
)) {
107 ineq
[k
] = STATUS_REDUNDANT
;
110 ineq
[k
] = status_in(bmap_i
->ineq
[k
], tab_j
);
111 if (ineq
[k
] == STATUS_ERROR
)
113 if (ineq
[k
] == STATUS_SEPARATE
)
123 static int any(int *con
, unsigned len
, int status
)
127 for (i
= 0; i
< len
; ++i
)
128 if (con
[i
] == status
)
133 /* Return the first position of "status" in the list "con" of length "len".
134 * Return -1 if there is no such entry.
136 static int find(int *con
, unsigned len
, int status
)
140 for (i
= 0; i
< len
; ++i
)
141 if (con
[i
] == status
)
146 static int count(int *con
, unsigned len
, int status
)
151 for (i
= 0; i
< len
; ++i
)
152 if (con
[i
] == status
)
157 static int all(int *con
, unsigned len
, int status
)
161 for (i
= 0; i
< len
; ++i
) {
162 if (con
[i
] == STATUS_REDUNDANT
)
164 if (con
[i
] != status
)
170 /* Internal information associated to a basic map in a map
171 * that is to be coalesced by isl_map_coalesce.
173 * "bmap" is the basic map itself (or NULL if "removed" is set)
174 * "tab" is the corresponding tableau (or NULL if "removed" is set)
175 * "hull_hash" identifies the affine space in which "bmap" lives.
176 * "removed" is set if this basic map has been removed from the map
177 * "simplify" is set if this basic map may have some unknown integer
178 * divisions that were not present in the input basic maps. The basic
179 * map should then be simplified such that we may be able to find
180 * a definition among the constraints.
182 * "eq" and "ineq" are only set if we are currently trying to coalesce
183 * this basic map with another basic map, in which case they represent
184 * the position of the inequalities of this basic map with respect to
185 * the other basic map. The number of elements in the "eq" array
186 * is twice the number of equalities in the "bmap", corresponding
187 * to the two inequalities that make up each equality.
189 struct isl_coalesce_info
{
199 /* Is there any (half of an) equality constraint in the description
200 * of the basic map represented by "info" that
201 * has position "status" with respect to the other basic map?
203 static int any_eq(struct isl_coalesce_info
*info
, int status
)
207 n_eq
= isl_basic_map_n_equality(info
->bmap
);
208 return any(info
->eq
, 2 * n_eq
, status
);
211 /* Is there any inequality constraint in the description
212 * of the basic map represented by "info" that
213 * has position "status" with respect to the other basic map?
215 static int any_ineq(struct isl_coalesce_info
*info
, int status
)
219 n_ineq
= isl_basic_map_n_inequality(info
->bmap
);
220 return any(info
->ineq
, n_ineq
, status
);
223 /* Return the position of the first half on an equality constraint
224 * in the description of the basic map represented by "info" that
225 * has position "status" with respect to the other basic map.
226 * The returned value is twice the position of the equality constraint
227 * plus zero for the negative half and plus one for the positive half.
228 * Return -1 if there is no such entry.
230 static int find_eq(struct isl_coalesce_info
*info
, int status
)
234 n_eq
= isl_basic_map_n_equality(info
->bmap
);
235 return find(info
->eq
, 2 * n_eq
, status
);
238 /* Return the position of the first inequality constraint in the description
239 * of the basic map represented by "info" that
240 * has position "status" with respect to the other basic map.
241 * Return -1 if there is no such entry.
243 static int find_ineq(struct isl_coalesce_info
*info
, int status
)
247 n_ineq
= isl_basic_map_n_inequality(info
->bmap
);
248 return find(info
->ineq
, n_ineq
, status
);
251 /* Return the number of (halves of) equality constraints in the description
252 * of the basic map represented by "info" that
253 * have position "status" with respect to the other basic map.
255 static int count_eq(struct isl_coalesce_info
*info
, int status
)
259 n_eq
= isl_basic_map_n_equality(info
->bmap
);
260 return count(info
->eq
, 2 * n_eq
, status
);
263 /* Return the number of inequality constraints in the description
264 * of the basic map represented by "info" that
265 * have position "status" with respect to the other basic map.
267 static int count_ineq(struct isl_coalesce_info
*info
, int status
)
271 n_ineq
= isl_basic_map_n_inequality(info
->bmap
);
272 return count(info
->ineq
, n_ineq
, status
);
275 /* Are all non-redundant constraints of the basic map represented by "info"
276 * either valid or cut constraints with respect to the other basic map?
278 static int all_valid_or_cut(struct isl_coalesce_info
*info
)
282 for (i
= 0; i
< 2 * info
->bmap
->n_eq
; ++i
) {
283 if (info
->eq
[i
] == STATUS_REDUNDANT
)
285 if (info
->eq
[i
] == STATUS_VALID
)
287 if (info
->eq
[i
] == STATUS_CUT
)
292 for (i
= 0; i
< info
->bmap
->n_ineq
; ++i
) {
293 if (info
->ineq
[i
] == STATUS_REDUNDANT
)
295 if (info
->ineq
[i
] == STATUS_VALID
)
297 if (info
->ineq
[i
] == STATUS_CUT
)
305 /* Compute the hash of the (apparent) affine hull of info->bmap (with
306 * the existentially quantified variables removed) and store it
309 static int coalesce_info_set_hull_hash(struct isl_coalesce_info
*info
)
314 hull
= isl_basic_map_copy(info
->bmap
);
315 hull
= isl_basic_map_plain_affine_hull(hull
);
316 n_div
= isl_basic_map_dim(hull
, isl_dim_div
);
317 hull
= isl_basic_map_drop_constraints_involving_dims(hull
,
318 isl_dim_div
, 0, n_div
);
319 info
->hull_hash
= isl_basic_map_get_hash(hull
);
320 isl_basic_map_free(hull
);
322 return hull
? 0 : -1;
325 /* Free all the allocated memory in an array
326 * of "n" isl_coalesce_info elements.
328 static void clear_coalesce_info(int n
, struct isl_coalesce_info
*info
)
335 for (i
= 0; i
< n
; ++i
) {
336 isl_basic_map_free(info
[i
].bmap
);
337 isl_tab_free(info
[i
].tab
);
343 /* Drop the basic map represented by "info".
344 * That is, clear the memory associated to the entry and
345 * mark it as having been removed.
347 static void drop(struct isl_coalesce_info
*info
)
349 info
->bmap
= isl_basic_map_free(info
->bmap
);
350 isl_tab_free(info
->tab
);
355 /* Exchange the information in "info1" with that in "info2".
357 static void exchange(struct isl_coalesce_info
*info1
,
358 struct isl_coalesce_info
*info2
)
360 struct isl_coalesce_info info
;
367 /* This type represents the kind of change that has been performed
368 * while trying to coalesce two basic maps.
370 * isl_change_none: nothing was changed
371 * isl_change_drop_first: the first basic map was removed
372 * isl_change_drop_second: the second basic map was removed
373 * isl_change_fuse: the two basic maps were replaced by a new basic map.
376 isl_change_error
= -1,
378 isl_change_drop_first
,
379 isl_change_drop_second
,
383 /* Update "change" based on an interchange of the first and the second
384 * basic map. That is, interchange isl_change_drop_first and
385 * isl_change_drop_second.
387 static enum isl_change
invert_change(enum isl_change change
)
390 case isl_change_error
:
391 return isl_change_error
;
392 case isl_change_none
:
393 return isl_change_none
;
394 case isl_change_drop_first
:
395 return isl_change_drop_second
;
396 case isl_change_drop_second
:
397 return isl_change_drop_first
;
398 case isl_change_fuse
:
399 return isl_change_fuse
;
402 return isl_change_error
;
405 /* Add the valid constraints of the basic map represented by "info"
406 * to "bmap". "len" is the size of the constraints.
407 * If only one of the pair of inequalities that make up an equality
408 * is valid, then add that inequality.
410 static __isl_give isl_basic_map
*add_valid_constraints(
411 __isl_take isl_basic_map
*bmap
, struct isl_coalesce_info
*info
,
419 for (k
= 0; k
< info
->bmap
->n_eq
; ++k
) {
420 if (info
->eq
[2 * k
] == STATUS_VALID
&&
421 info
->eq
[2 * k
+ 1] == STATUS_VALID
) {
422 l
= isl_basic_map_alloc_equality(bmap
);
424 return isl_basic_map_free(bmap
);
425 isl_seq_cpy(bmap
->eq
[l
], info
->bmap
->eq
[k
], len
);
426 } else if (info
->eq
[2 * k
] == STATUS_VALID
) {
427 l
= isl_basic_map_alloc_inequality(bmap
);
429 return isl_basic_map_free(bmap
);
430 isl_seq_neg(bmap
->ineq
[l
], info
->bmap
->eq
[k
], len
);
431 } else if (info
->eq
[2 * k
+ 1] == STATUS_VALID
) {
432 l
= isl_basic_map_alloc_inequality(bmap
);
434 return isl_basic_map_free(bmap
);
435 isl_seq_cpy(bmap
->ineq
[l
], info
->bmap
->eq
[k
], len
);
439 for (k
= 0; k
< info
->bmap
->n_ineq
; ++k
) {
440 if (info
->ineq
[k
] != STATUS_VALID
)
442 l
= isl_basic_map_alloc_inequality(bmap
);
444 return isl_basic_map_free(bmap
);
445 isl_seq_cpy(bmap
->ineq
[l
], info
->bmap
->ineq
[k
], len
);
451 /* Is "bmap" defined by a number of (non-redundant) constraints that
452 * is greater than the number of constraints of basic maps i and j combined?
453 * Equalities are counted as two inequalities.
455 static int number_of_constraints_increases(int i
, int j
,
456 struct isl_coalesce_info
*info
,
457 __isl_keep isl_basic_map
*bmap
, struct isl_tab
*tab
)
461 n_old
= 2 * info
[i
].bmap
->n_eq
+ info
[i
].bmap
->n_ineq
;
462 n_old
+= 2 * info
[j
].bmap
->n_eq
+ info
[j
].bmap
->n_ineq
;
464 n_new
= 2 * bmap
->n_eq
;
465 for (k
= 0; k
< bmap
->n_ineq
; ++k
)
466 if (!isl_tab_is_redundant(tab
, bmap
->n_eq
+ k
))
469 return n_new
> n_old
;
472 /* Replace the pair of basic maps i and j by the basic map bounded
473 * by the valid constraints in both basic maps and the constraints
474 * in extra (if not NULL).
475 * Place the fused basic map in the position that is the smallest of i and j.
477 * If "detect_equalities" is set, then look for equalities encoded
478 * as pairs of inequalities.
479 * If "check_number" is set, then the original basic maps are only
480 * replaced if the total number of constraints does not increase.
481 * While the number of integer divisions in the two basic maps
482 * is assumed to be the same, the actual definitions may be different.
483 * We only copy the definition from one of the basic map if it is
484 * the same as that of the other basic map. Otherwise, we mark
485 * the integer division as unknown and simplify the basic map
486 * in an attempt to recover the integer division definition.
488 static enum isl_change
fuse(int i
, int j
, struct isl_coalesce_info
*info
,
489 __isl_keep isl_mat
*extra
, int detect_equalities
, int check_number
)
492 struct isl_basic_map
*fused
= NULL
;
493 struct isl_tab
*fused_tab
= NULL
;
494 unsigned total
= isl_basic_map_total_dim(info
[i
].bmap
);
495 unsigned extra_rows
= extra
? extra
->n_row
: 0;
496 unsigned n_eq
, n_ineq
;
500 return fuse(j
, i
, info
, extra
, detect_equalities
, check_number
);
502 n_eq
= info
[i
].bmap
->n_eq
+ info
[j
].bmap
->n_eq
;
503 n_ineq
= info
[i
].bmap
->n_ineq
+ info
[j
].bmap
->n_ineq
;
504 fused
= isl_basic_map_alloc_space(isl_space_copy(info
[i
].bmap
->dim
),
505 info
[i
].bmap
->n_div
, n_eq
, n_eq
+ n_ineq
+ extra_rows
);
506 fused
= add_valid_constraints(fused
, &info
[i
], 1 + total
);
507 fused
= add_valid_constraints(fused
, &info
[j
], 1 + total
);
510 if (ISL_F_ISSET(info
[i
].bmap
, ISL_BASIC_MAP_RATIONAL
) &&
511 ISL_F_ISSET(info
[j
].bmap
, ISL_BASIC_MAP_RATIONAL
))
512 ISL_F_SET(fused
, ISL_BASIC_MAP_RATIONAL
);
514 for (k
= 0; k
< info
[i
].bmap
->n_div
; ++k
) {
515 int l
= isl_basic_map_alloc_div(fused
);
518 if (isl_seq_eq(info
[i
].bmap
->div
[k
], info
[j
].bmap
->div
[k
],
520 isl_seq_cpy(fused
->div
[l
], info
[i
].bmap
->div
[k
],
523 isl_int_set_si(fused
->div
[l
][0], 0);
528 for (k
= 0; k
< extra_rows
; ++k
) {
529 l
= isl_basic_map_alloc_inequality(fused
);
532 isl_seq_cpy(fused
->ineq
[l
], extra
->row
[k
], 1 + total
);
535 if (detect_equalities
)
536 fused
= isl_basic_map_detect_inequality_pairs(fused
, NULL
);
537 fused
= isl_basic_map_gauss(fused
, NULL
);
538 if (simplify
|| info
[j
].simplify
) {
539 fused
= isl_basic_map_simplify(fused
);
540 info
[i
].simplify
= 0;
542 fused
= isl_basic_map_finalize(fused
);
544 fused_tab
= isl_tab_from_basic_map(fused
, 0);
545 if (isl_tab_detect_redundant(fused_tab
) < 0)
549 number_of_constraints_increases(i
, j
, info
, fused
, fused_tab
)) {
550 isl_tab_free(fused_tab
);
551 isl_basic_map_free(fused
);
552 return isl_change_none
;
555 isl_basic_map_free(info
[i
].bmap
);
556 info
[i
].bmap
= fused
;
557 isl_tab_free(info
[i
].tab
);
558 info
[i
].tab
= fused_tab
;
561 return isl_change_fuse
;
563 isl_tab_free(fused_tab
);
564 isl_basic_map_free(fused
);
565 return isl_change_error
;
568 /* Given a pair of basic maps i and j such that all constraints are either
569 * "valid" or "cut", check if the facets corresponding to the "cut"
570 * constraints of i lie entirely within basic map j.
571 * If so, replace the pair by the basic map consisting of the valid
572 * constraints in both basic maps.
573 * Checking whether the facet lies entirely within basic map j
574 * is performed by checking whether the constraints of basic map j
575 * are valid for the facet. These tests are performed on a rational
576 * tableau to avoid the theoretical possibility that a constraint
577 * that was considered to be a cut constraint for the entire basic map i
578 * happens to be considered to be a valid constraint for the facet,
579 * even though it cuts off the same rational points.
581 * To see that we are not introducing any extra points, call the
582 * two basic maps A and B and the resulting map U and let x
583 * be an element of U \setminus ( A \cup B ).
584 * A line connecting x with an element of A \cup B meets a facet F
585 * of either A or B. Assume it is a facet of B and let c_1 be
586 * the corresponding facet constraint. We have c_1(x) < 0 and
587 * so c_1 is a cut constraint. This implies that there is some
588 * (possibly rational) point x' satisfying the constraints of A
589 * and the opposite of c_1 as otherwise c_1 would have been marked
590 * valid for A. The line connecting x and x' meets a facet of A
591 * in a (possibly rational) point that also violates c_1, but this
592 * is impossible since all cut constraints of B are valid for all
594 * In case F is a facet of A rather than B, then we can apply the
595 * above reasoning to find a facet of B separating x from A \cup B first.
597 static enum isl_change
check_facets(int i
, int j
,
598 struct isl_coalesce_info
*info
)
601 struct isl_tab_undo
*snap
, *snap2
;
602 unsigned n_eq
= info
[i
].bmap
->n_eq
;
604 snap
= isl_tab_snap(info
[i
].tab
);
605 if (isl_tab_mark_rational(info
[i
].tab
) < 0)
606 return isl_change_error
;
607 snap2
= isl_tab_snap(info
[i
].tab
);
609 for (k
= 0; k
< info
[i
].bmap
->n_ineq
; ++k
) {
610 if (info
[i
].ineq
[k
] != STATUS_CUT
)
612 if (isl_tab_select_facet(info
[i
].tab
, n_eq
+ k
) < 0)
613 return isl_change_error
;
614 for (l
= 0; l
< info
[j
].bmap
->n_ineq
; ++l
) {
616 if (info
[j
].ineq
[l
] != STATUS_CUT
)
618 stat
= status_in(info
[j
].bmap
->ineq
[l
], info
[i
].tab
);
620 return isl_change_error
;
621 if (stat
!= STATUS_VALID
)
624 if (isl_tab_rollback(info
[i
].tab
, snap2
) < 0)
625 return isl_change_error
;
626 if (l
< info
[j
].bmap
->n_ineq
)
630 if (k
< info
[i
].bmap
->n_ineq
) {
631 if (isl_tab_rollback(info
[i
].tab
, snap
) < 0)
632 return isl_change_error
;
633 return isl_change_none
;
635 return fuse(i
, j
, info
, NULL
, 0, 0);
638 /* Check if info->bmap contains the basic map represented
639 * by the tableau "tab".
640 * For each equality, we check both the constraint itself
641 * (as an inequality) and its negation. Make sure the
642 * equality is returned to its original state before returning.
644 static isl_bool
contains(struct isl_coalesce_info
*info
, struct isl_tab
*tab
)
648 isl_basic_map
*bmap
= info
->bmap
;
650 dim
= isl_basic_map_total_dim(bmap
);
651 for (k
= 0; k
< bmap
->n_eq
; ++k
) {
653 isl_seq_neg(bmap
->eq
[k
], bmap
->eq
[k
], 1 + dim
);
654 stat
= status_in(bmap
->eq
[k
], tab
);
655 isl_seq_neg(bmap
->eq
[k
], bmap
->eq
[k
], 1 + dim
);
657 return isl_bool_error
;
658 if (stat
!= STATUS_VALID
)
659 return isl_bool_false
;
660 stat
= status_in(bmap
->eq
[k
], tab
);
662 return isl_bool_error
;
663 if (stat
!= STATUS_VALID
)
664 return isl_bool_false
;
667 for (k
= 0; k
< bmap
->n_ineq
; ++k
) {
669 if (info
->ineq
[k
] == STATUS_REDUNDANT
)
671 stat
= status_in(bmap
->ineq
[k
], tab
);
673 return isl_bool_error
;
674 if (stat
!= STATUS_VALID
)
675 return isl_bool_false
;
677 return isl_bool_true
;
680 /* Basic map "i" has an inequality (say "k") that is adjacent
681 * to some inequality of basic map "j". All the other inequalities
683 * Check if basic map "j" forms an extension of basic map "i".
685 * Note that this function is only called if some of the equalities or
686 * inequalities of basic map "j" do cut basic map "i". The function is
687 * correct even if there are no such cut constraints, but in that case
688 * the additional checks performed by this function are overkill.
690 * In particular, we replace constraint k, say f >= 0, by constraint
691 * f <= -1, add the inequalities of "j" that are valid for "i"
692 * and check if the result is a subset of basic map "j".
693 * To improve the chances of the subset relation being detected,
694 * any variable that only attains a single integer value
695 * in the tableau of "i" is first fixed to that value.
696 * If the result is a subset, then we know that this result is exactly equal
697 * to basic map "j" since all its constraints are valid for basic map "j".
698 * By combining the valid constraints of "i" (all equalities and all
699 * inequalities except "k") and the valid constraints of "j" we therefore
700 * obtain a basic map that is equal to their union.
701 * In this case, there is no need to perform a rollback of the tableau
702 * since it is going to be destroyed in fuse().
708 * |_______| _ |_________\
720 static enum isl_change
is_adj_ineq_extension(int i
, int j
,
721 struct isl_coalesce_info
*info
)
724 struct isl_tab_undo
*snap
;
725 unsigned n_eq
= info
[i
].bmap
->n_eq
;
726 unsigned total
= isl_basic_map_total_dim(info
[i
].bmap
);
730 if (isl_tab_extend_cons(info
[i
].tab
, 1 + info
[j
].bmap
->n_ineq
) < 0)
731 return isl_change_error
;
733 k
= find_ineq(&info
[i
], STATUS_ADJ_INEQ
);
735 isl_die(isl_basic_map_get_ctx(info
[i
].bmap
), isl_error_internal
,
736 "info[i].ineq should have exactly one STATUS_ADJ_INEQ",
737 return isl_change_error
);
739 snap
= isl_tab_snap(info
[i
].tab
);
741 if (isl_tab_unrestrict(info
[i
].tab
, n_eq
+ k
) < 0)
742 return isl_change_error
;
744 isl_seq_neg(info
[i
].bmap
->ineq
[k
], info
[i
].bmap
->ineq
[k
], 1 + total
);
745 isl_int_sub_ui(info
[i
].bmap
->ineq
[k
][0], info
[i
].bmap
->ineq
[k
][0], 1);
746 r
= isl_tab_add_ineq(info
[i
].tab
, info
[i
].bmap
->ineq
[k
]);
747 isl_seq_neg(info
[i
].bmap
->ineq
[k
], info
[i
].bmap
->ineq
[k
], 1 + total
);
748 isl_int_sub_ui(info
[i
].bmap
->ineq
[k
][0], info
[i
].bmap
->ineq
[k
][0], 1);
750 return isl_change_error
;
752 for (k
= 0; k
< info
[j
].bmap
->n_ineq
; ++k
) {
753 if (info
[j
].ineq
[k
] != STATUS_VALID
)
755 if (isl_tab_add_ineq(info
[i
].tab
, info
[j
].bmap
->ineq
[k
]) < 0)
756 return isl_change_error
;
758 if (isl_tab_detect_constants(info
[i
].tab
) < 0)
759 return isl_change_error
;
761 super
= contains(&info
[j
], info
[i
].tab
);
763 return isl_change_error
;
765 return fuse(i
, j
, info
, NULL
, 0, 0);
767 if (isl_tab_rollback(info
[i
].tab
, snap
) < 0)
768 return isl_change_error
;
770 return isl_change_none
;
774 /* Both basic maps have at least one inequality with and adjacent
775 * (but opposite) inequality in the other basic map.
776 * Check that there are no cut constraints and that there is only
777 * a single pair of adjacent inequalities.
778 * If so, we can replace the pair by a single basic map described
779 * by all but the pair of adjacent inequalities.
780 * Any additional points introduced lie strictly between the two
781 * adjacent hyperplanes and can therefore be integral.
790 * The test for a single pair of adjancent inequalities is important
791 * for avoiding the combination of two basic maps like the following
801 * If there are some cut constraints on one side, then we may
802 * still be able to fuse the two basic maps, but we need to perform
803 * some additional checks in is_adj_ineq_extension.
805 static enum isl_change
check_adj_ineq(int i
, int j
,
806 struct isl_coalesce_info
*info
)
808 int count_i
, count_j
;
811 count_i
= count_ineq(&info
[i
], STATUS_ADJ_INEQ
);
812 count_j
= count_ineq(&info
[j
], STATUS_ADJ_INEQ
);
814 if (count_i
!= 1 && count_j
!= 1)
815 return isl_change_none
;
817 cut_i
= any_eq(&info
[i
], STATUS_CUT
) || any_ineq(&info
[i
], STATUS_CUT
);
818 cut_j
= any_eq(&info
[j
], STATUS_CUT
) || any_ineq(&info
[j
], STATUS_CUT
);
820 if (!cut_i
&& !cut_j
&& count_i
== 1 && count_j
== 1)
821 return fuse(i
, j
, info
, NULL
, 0, 0);
823 if (count_i
== 1 && !cut_i
)
824 return is_adj_ineq_extension(i
, j
, info
);
826 if (count_j
== 1 && !cut_j
)
827 return is_adj_ineq_extension(j
, i
, info
);
829 return isl_change_none
;
832 /* Given an affine transformation matrix "T", does row "row" represent
833 * anything other than a unit vector (possibly shifted by a constant)
834 * that is not involved in any of the other rows?
836 * That is, if a constraint involves the variable corresponding to
837 * the row, then could its preimage by "T" have any coefficients
838 * that are different from those in the original constraint?
840 static int not_unique_unit_row(__isl_keep isl_mat
*T
, int row
)
843 int len
= T
->n_col
- 1;
845 i
= isl_seq_first_non_zero(T
->row
[row
] + 1, len
);
848 if (!isl_int_is_one(T
->row
[row
][1 + i
]) &&
849 !isl_int_is_negone(T
->row
[row
][1 + i
]))
852 j
= isl_seq_first_non_zero(T
->row
[row
] + 1 + i
+ 1, len
- (i
+ 1));
856 for (j
= 1; j
< T
->n_row
; ++j
) {
859 if (!isl_int_is_zero(T
->row
[j
][1 + i
]))
866 /* Does inequality constraint "ineq" of "bmap" involve any of
867 * the variables marked in "affected"?
868 * "total" is the total number of variables, i.e., the number
869 * of entries in "affected".
871 static isl_bool
is_affected(__isl_keep isl_basic_map
*bmap
, int ineq
,
872 int *affected
, int total
)
876 for (i
= 0; i
< total
; ++i
) {
879 if (!isl_int_is_zero(bmap
->ineq
[ineq
][1 + i
]))
880 return isl_bool_true
;
883 return isl_bool_false
;
886 /* Given the compressed version of inequality constraint "ineq"
887 * of info->bmap in "v", check if the constraint can be tightened,
888 * where the compression is based on an equality constraint valid
890 * If so, add the tightened version of the inequality constraint
891 * to info->tab. "v" may be modified by this function.
893 * That is, if the compressed constraint is of the form
897 * with 0 < c < m, then it is equivalent to
901 * This means that c can also be subtracted from the original,
902 * uncompressed constraint without affecting the integer points
903 * in info->tab. Add this tightened constraint as an extra row
904 * to info->tab to make this information explicitly available.
906 static __isl_give isl_vec
*try_tightening(struct isl_coalesce_info
*info
,
907 int ineq
, __isl_take isl_vec
*v
)
915 ctx
= isl_vec_get_ctx(v
);
916 isl_seq_gcd(v
->el
+ 1, v
->size
- 1, &ctx
->normalize_gcd
);
917 if (isl_int_is_zero(ctx
->normalize_gcd
) ||
918 isl_int_is_one(ctx
->normalize_gcd
)) {
926 isl_int_fdiv_r(v
->el
[0], v
->el
[0], ctx
->normalize_gcd
);
927 if (isl_int_is_zero(v
->el
[0]))
930 if (isl_tab_extend_cons(info
->tab
, 1) < 0)
931 return isl_vec_free(v
);
933 isl_int_sub(info
->bmap
->ineq
[ineq
][0],
934 info
->bmap
->ineq
[ineq
][0], v
->el
[0]);
935 r
= isl_tab_add_ineq(info
->tab
, info
->bmap
->ineq
[ineq
]);
936 isl_int_add(info
->bmap
->ineq
[ineq
][0],
937 info
->bmap
->ineq
[ineq
][0], v
->el
[0]);
940 return isl_vec_free(v
);
945 /* Tighten the (non-redundant) constraints on the facet represented
947 * In particular, on input, info->tab represents the result
948 * of relaxing the "n" inequality constraints of info->bmap in "relaxed"
949 * by one, i.e., replacing f_i >= 0 by f_i + 1 >= 0, and then
950 * replacing the one at index "l" by the corresponding equality,
951 * i.e., f_k + 1 = 0, with k = relaxed[l].
953 * Compute a variable compression from the equality constraint f_k + 1 = 0
954 * and use it to tighten the other constraints of info->bmap
955 * (that is, all constraints that have not been relaxed),
956 * updating info->tab (and leaving info->bmap untouched).
957 * The compression handles essentially two cases, one where a variable
958 * is assigned a fixed value and can therefore be eliminated, and one
959 * where one variable is a shifted multiple of some other variable and
960 * can therefore be replaced by that multiple.
961 * Gaussian elimination would also work for the first case, but for
962 * the second case, the effectiveness would depend on the order
964 * After compression, some of the constraints may have coefficients
965 * with a common divisor. If this divisor does not divide the constant
966 * term, then the constraint can be tightened.
967 * The tightening is performed on the tableau info->tab by introducing
968 * extra (temporary) constraints.
970 * Only constraints that are possibly affected by the compression are
971 * considered. In particular, if the constraint only involves variables
972 * that are directly mapped to a distinct set of other variables, then
973 * no common divisor can be introduced and no tightening can occur.
975 * It is important to only consider the non-redundant constraints
976 * since the facet constraint has been relaxed prior to the call
977 * to this function, meaning that the constraints that were redundant
978 * prior to the relaxation may no longer be redundant.
979 * These constraints will be ignored in the fused result, so
980 * the fusion detection should not exploit them.
982 static isl_stat
tighten_on_relaxed_facet(struct isl_coalesce_info
*info
,
983 int n
, int *relaxed
, int l
)
994 ctx
= isl_basic_map_get_ctx(info
->bmap
);
995 total
= isl_basic_map_total_dim(info
->bmap
);
996 isl_int_add_ui(info
->bmap
->ineq
[k
][0], info
->bmap
->ineq
[k
][0], 1);
997 T
= isl_mat_sub_alloc6(ctx
, info
->bmap
->ineq
, k
, 1, 0, 1 + total
);
998 T
= isl_mat_variable_compression(T
, NULL
);
999 isl_int_sub_ui(info
->bmap
->ineq
[k
][0], info
->bmap
->ineq
[k
][0], 1);
1001 return isl_stat_error
;
1002 if (T
->n_col
== 0) {
1007 affected
= isl_alloc_array(ctx
, int, total
);
1011 for (i
= 0; i
< total
; ++i
)
1012 affected
[i
] = not_unique_unit_row(T
, 1 + i
);
1014 for (i
= 0; i
< info
->bmap
->n_ineq
; ++i
) {
1016 if (any(relaxed
, n
, i
))
1018 if (info
->ineq
[i
] == STATUS_REDUNDANT
)
1020 handle
= is_affected(info
->bmap
, i
, affected
, total
);
1025 v
= isl_vec_alloc(ctx
, 1 + total
);
1028 isl_seq_cpy(v
->el
, info
->bmap
->ineq
[i
], 1 + total
);
1029 v
= isl_vec_mat_product(v
, isl_mat_copy(T
));
1030 v
= try_tightening(info
, i
, v
);
1042 return isl_stat_error
;
1045 /* Replace the basic maps "i" and "j" by an extension of "i"
1046 * along the "n" inequality constraints in "relax" by one.
1047 * The tableau info[i].tab has already been extended.
1048 * Extend info[i].bmap accordingly by relaxing all constraints in "relax"
1050 * Each integer division that does not have exactly the same
1051 * definition in "i" and "j" is marked unknown and the basic map
1052 * is scheduled to be simplified in an attempt to recover
1053 * the integer division definition.
1054 * Place the extension in the position that is the smallest of i and j.
1056 static enum isl_change
extend(int i
, int j
, int n
, int *relax
,
1057 struct isl_coalesce_info
*info
)
1062 info
[i
].bmap
= isl_basic_map_cow(info
[i
].bmap
);
1064 return isl_change_error
;
1065 total
= isl_basic_map_total_dim(info
[i
].bmap
);
1066 for (l
= 0; l
< info
[i
].bmap
->n_div
; ++l
)
1067 if (!isl_seq_eq(info
[i
].bmap
->div
[l
],
1068 info
[j
].bmap
->div
[l
], 1 + 1 + total
)) {
1069 isl_int_set_si(info
[i
].bmap
->div
[l
][0], 0);
1070 info
[i
].simplify
= 1;
1072 for (l
= 0; l
< n
; ++l
)
1073 isl_int_add_ui(info
[i
].bmap
->ineq
[relax
[l
]][0],
1074 info
[i
].bmap
->ineq
[relax
[l
]][0], 1);
1075 ISL_F_SET(info
[i
].bmap
, ISL_BASIC_MAP_FINAL
);
1078 exchange(&info
[i
], &info
[j
]);
1079 return isl_change_fuse
;
1082 /* Basic map "i" has "n" inequality constraints (collected in "relax")
1083 * that are such that they include basic map "j" if they are relaxed
1084 * by one. All the other inequalities are valid for "j".
1085 * Check if basic map "j" forms an extension of basic map "i".
1087 * In particular, relax the constraints in "relax", compute the corresponding
1088 * facets one by one and check whether each of these is included
1089 * in the other basic map.
1090 * Before testing for inclusion, the constraints on each facet
1091 * are tightened to increase the chance of an inclusion being detected.
1092 * (Adding the valid constraints of "j" to the tableau of "i", as is done
1093 * in is_adj_ineq_extension, may further increase those chances, but this
1094 * is not currently done.)
1095 * If each facet is included, we know that relaxing the constraints extends
1096 * the basic map with exactly the other basic map (we already know that this
1097 * other basic map is included in the extension, because all other
1098 * inequality constraints are valid of "j") and we can replace the
1099 * two basic maps by this extension.
1101 * If any of the relaxed constraints turn out to be redundant, then bail out.
1102 * isl_tab_select_facet refuses to handle such constraints. It may be
1103 * possible to handle them anyway by making a distinction between
1104 * redundant constraints with a corresponding facet that still intersects
1105 * the set (allowing isl_tab_select_facet to handle them) and
1106 * those where the facet does not intersect the set (which can be ignored
1107 * because the empty facet is trivially included in the other disjunct).
1108 * However, relaxed constraints that turn out to be redundant should
1109 * be fairly rare and no such instance has been reported where
1110 * coalescing would be successful.
1126 static enum isl_change
is_relaxed_extension(int i
, int j
, int n
, int *relax
,
1127 struct isl_coalesce_info
*info
)
1131 struct isl_tab_undo
*snap
, *snap2
;
1132 unsigned n_eq
= info
[i
].bmap
->n_eq
;
1134 for (l
= 0; l
< n
; ++l
)
1135 if (isl_tab_is_equality(info
[i
].tab
, n_eq
+ relax
[l
]))
1136 return isl_change_none
;
1138 snap
= isl_tab_snap(info
[i
].tab
);
1139 for (l
= 0; l
< n
; ++l
)
1140 if (isl_tab_relax(info
[i
].tab
, n_eq
+ relax
[l
]) < 0)
1141 return isl_change_error
;
1142 for (l
= 0; l
< n
; ++l
) {
1143 if (!isl_tab_is_redundant(info
[i
].tab
, n_eq
+ relax
[l
]))
1145 if (isl_tab_rollback(info
[i
].tab
, snap
) < 0)
1146 return isl_change_error
;
1147 return isl_change_none
;
1149 snap2
= isl_tab_snap(info
[i
].tab
);
1150 for (l
= 0; l
< n
; ++l
) {
1151 if (isl_tab_rollback(info
[i
].tab
, snap2
) < 0)
1152 return isl_change_error
;
1153 if (isl_tab_select_facet(info
[i
].tab
, n_eq
+ relax
[l
]) < 0)
1154 return isl_change_error
;
1155 if (tighten_on_relaxed_facet(&info
[i
], n
, relax
, l
) < 0)
1156 return isl_change_error
;
1157 super
= contains(&info
[j
], info
[i
].tab
);
1159 return isl_change_error
;
1162 if (isl_tab_rollback(info
[i
].tab
, snap
) < 0)
1163 return isl_change_error
;
1164 return isl_change_none
;
1167 if (isl_tab_rollback(info
[i
].tab
, snap2
) < 0)
1168 return isl_change_error
;
1169 return extend(i
, j
, n
, relax
, info
);
1172 /* Data structure that keeps track of the wrapping constraints
1173 * and of information to bound the coefficients of those constraints.
1175 * bound is set if we want to apply a bound on the coefficients
1176 * mat contains the wrapping constraints
1177 * max is the bound on the coefficients (if bound is set)
1185 /* Update wraps->max to be greater than or equal to the coefficients
1186 * in the equalities and inequalities of info->bmap that can be removed
1187 * if we end up applying wrapping.
1189 static isl_stat
wraps_update_max(struct isl_wraps
*wraps
,
1190 struct isl_coalesce_info
*info
)
1194 unsigned total
= isl_basic_map_total_dim(info
->bmap
);
1196 isl_int_init(max_k
);
1198 for (k
= 0; k
< info
->bmap
->n_eq
; ++k
) {
1199 if (info
->eq
[2 * k
] == STATUS_VALID
&&
1200 info
->eq
[2 * k
+ 1] == STATUS_VALID
)
1202 isl_seq_abs_max(info
->bmap
->eq
[k
] + 1, total
, &max_k
);
1203 if (isl_int_abs_gt(max_k
, wraps
->max
))
1204 isl_int_set(wraps
->max
, max_k
);
1207 for (k
= 0; k
< info
->bmap
->n_ineq
; ++k
) {
1208 if (info
->ineq
[k
] == STATUS_VALID
||
1209 info
->ineq
[k
] == STATUS_REDUNDANT
)
1211 isl_seq_abs_max(info
->bmap
->ineq
[k
] + 1, total
, &max_k
);
1212 if (isl_int_abs_gt(max_k
, wraps
->max
))
1213 isl_int_set(wraps
->max
, max_k
);
1216 isl_int_clear(max_k
);
1221 /* Initialize the isl_wraps data structure.
1222 * If we want to bound the coefficients of the wrapping constraints,
1223 * we set wraps->max to the largest coefficient
1224 * in the equalities and inequalities that can be removed if we end up
1225 * applying wrapping.
1227 static isl_stat
wraps_init(struct isl_wraps
*wraps
, __isl_take isl_mat
*mat
,
1228 struct isl_coalesce_info
*info
, int i
, int j
)
1235 return isl_stat_error
;
1236 ctx
= isl_mat_get_ctx(mat
);
1237 wraps
->bound
= isl_options_get_coalesce_bounded_wrapping(ctx
);
1240 isl_int_init(wraps
->max
);
1241 isl_int_set_si(wraps
->max
, 0);
1242 if (wraps_update_max(wraps
, &info
[i
]) < 0)
1243 return isl_stat_error
;
1244 if (wraps_update_max(wraps
, &info
[j
]) < 0)
1245 return isl_stat_error
;
1250 /* Free the contents of the isl_wraps data structure.
1252 static void wraps_free(struct isl_wraps
*wraps
)
1254 isl_mat_free(wraps
->mat
);
1256 isl_int_clear(wraps
->max
);
1259 /* Is the wrapping constraint in row "row" allowed?
1261 * If wraps->bound is set, we check that none of the coefficients
1262 * is greater than wraps->max.
1264 static int allow_wrap(struct isl_wraps
*wraps
, int row
)
1271 for (i
= 1; i
< wraps
->mat
->n_col
; ++i
)
1272 if (isl_int_abs_gt(wraps
->mat
->row
[row
][i
], wraps
->max
))
1278 /* Wrap "ineq" (or its opposite if "negate" is set) around "bound"
1279 * to include "set" and add the result in position "w" of "wraps".
1280 * "len" is the total number of coefficients in "bound" and "ineq".
1281 * Return 1 on success, 0 on failure and -1 on error.
1282 * Wrapping can fail if the result of wrapping is equal to "bound"
1283 * or if we want to bound the sizes of the coefficients and
1284 * the wrapped constraint does not satisfy this bound.
1286 static int add_wrap(struct isl_wraps
*wraps
, int w
, isl_int
*bound
,
1287 isl_int
*ineq
, unsigned len
, __isl_keep isl_set
*set
, int negate
)
1289 isl_seq_cpy(wraps
->mat
->row
[w
], bound
, len
);
1291 isl_seq_neg(wraps
->mat
->row
[w
+ 1], ineq
, len
);
1292 ineq
= wraps
->mat
->row
[w
+ 1];
1294 if (!isl_set_wrap_facet(set
, wraps
->mat
->row
[w
], ineq
))
1296 if (isl_seq_eq(wraps
->mat
->row
[w
], bound
, len
))
1298 if (!allow_wrap(wraps
, w
))
1303 /* For each constraint in info->bmap that is not redundant (as determined
1304 * by info->tab) and that is not a valid constraint for the other basic map,
1305 * wrap the constraint around "bound" such that it includes the whole
1306 * set "set" and append the resulting constraint to "wraps".
1307 * Note that the constraints that are valid for the other basic map
1308 * will be added to the combined basic map by default, so there is
1309 * no need to wrap them.
1310 * The caller wrap_in_facets even relies on this function not wrapping
1311 * any constraints that are already valid.
1312 * "wraps" is assumed to have been pre-allocated to the appropriate size.
1313 * wraps->n_row is the number of actual wrapped constraints that have
1315 * If any of the wrapping problems results in a constraint that is
1316 * identical to "bound", then this means that "set" is unbounded in such
1317 * way that no wrapping is possible. If this happens then wraps->n_row
1319 * Similarly, if we want to bound the coefficients of the wrapping
1320 * constraints and a newly added wrapping constraint does not
1321 * satisfy the bound, then wraps->n_row is also reset to zero.
1323 static isl_stat
add_wraps(struct isl_wraps
*wraps
,
1324 struct isl_coalesce_info
*info
, isl_int
*bound
, __isl_keep isl_set
*set
)
1329 isl_basic_map
*bmap
= info
->bmap
;
1330 unsigned len
= 1 + isl_basic_map_total_dim(bmap
);
1332 w
= wraps
->mat
->n_row
;
1334 for (l
= 0; l
< bmap
->n_ineq
; ++l
) {
1335 if (info
->ineq
[l
] == STATUS_VALID
||
1336 info
->ineq
[l
] == STATUS_REDUNDANT
)
1338 if (isl_seq_is_neg(bound
, bmap
->ineq
[l
], len
))
1340 if (isl_seq_eq(bound
, bmap
->ineq
[l
], len
))
1342 if (isl_tab_is_redundant(info
->tab
, bmap
->n_eq
+ l
))
1345 added
= add_wrap(wraps
, w
, bound
, bmap
->ineq
[l
], len
, set
, 0);
1347 return isl_stat_error
;
1352 for (l
= 0; l
< bmap
->n_eq
; ++l
) {
1353 if (isl_seq_is_neg(bound
, bmap
->eq
[l
], len
))
1355 if (isl_seq_eq(bound
, bmap
->eq
[l
], len
))
1358 for (m
= 0; m
< 2; ++m
) {
1359 if (info
->eq
[2 * l
+ m
] == STATUS_VALID
)
1361 added
= add_wrap(wraps
, w
, bound
, bmap
->eq
[l
], len
,
1364 return isl_stat_error
;
1371 wraps
->mat
->n_row
= w
;
1374 wraps
->mat
->n_row
= 0;
1378 /* Check if the constraints in "wraps" from "first" until the last
1379 * are all valid for the basic set represented by "tab".
1380 * If not, wraps->n_row is set to zero.
1382 static int check_wraps(__isl_keep isl_mat
*wraps
, int first
,
1383 struct isl_tab
*tab
)
1387 for (i
= first
; i
< wraps
->n_row
; ++i
) {
1388 enum isl_ineq_type type
;
1389 type
= isl_tab_ineq_type(tab
, wraps
->row
[i
]);
1390 if (type
== isl_ineq_error
)
1392 if (type
== isl_ineq_redundant
)
1401 /* Return a set that corresponds to the non-redundant constraints
1402 * (as recorded in tab) of bmap.
1404 * It's important to remove the redundant constraints as some
1405 * of the other constraints may have been modified after the
1406 * constraints were marked redundant.
1407 * In particular, a constraint may have been relaxed.
1408 * Redundant constraints are ignored when a constraint is relaxed
1409 * and should therefore continue to be ignored ever after.
1410 * Otherwise, the relaxation might be thwarted by some of
1411 * these constraints.
1413 * Update the underlying set to ensure that the dimension doesn't change.
1414 * Otherwise the integer divisions could get dropped if the tab
1415 * turns out to be empty.
1417 static __isl_give isl_set
*set_from_updated_bmap(__isl_keep isl_basic_map
*bmap
,
1418 struct isl_tab
*tab
)
1420 isl_basic_set
*bset
;
1422 bmap
= isl_basic_map_copy(bmap
);
1423 bset
= isl_basic_map_underlying_set(bmap
);
1424 bset
= isl_basic_set_cow(bset
);
1425 bset
= isl_basic_set_update_from_tab(bset
, tab
);
1426 return isl_set_from_basic_set(bset
);
1429 /* Wrap the constraints of info->bmap that bound the facet defined
1430 * by inequality "k" around (the opposite of) this inequality to
1431 * include "set". "bound" may be used to store the negated inequality.
1432 * Since the wrapped constraints are not guaranteed to contain the whole
1433 * of info->bmap, we check them in check_wraps.
1434 * If any of the wrapped constraints turn out to be invalid, then
1435 * check_wraps will reset wrap->n_row to zero.
1437 static isl_stat
add_wraps_around_facet(struct isl_wraps
*wraps
,
1438 struct isl_coalesce_info
*info
, int k
, isl_int
*bound
,
1439 __isl_keep isl_set
*set
)
1441 struct isl_tab_undo
*snap
;
1443 unsigned total
= isl_basic_map_total_dim(info
->bmap
);
1445 snap
= isl_tab_snap(info
->tab
);
1447 if (isl_tab_select_facet(info
->tab
, info
->bmap
->n_eq
+ k
) < 0)
1448 return isl_stat_error
;
1449 if (isl_tab_detect_redundant(info
->tab
) < 0)
1450 return isl_stat_error
;
1452 isl_seq_neg(bound
, info
->bmap
->ineq
[k
], 1 + total
);
1454 n
= wraps
->mat
->n_row
;
1455 if (add_wraps(wraps
, info
, bound
, set
) < 0)
1456 return isl_stat_error
;
1458 if (isl_tab_rollback(info
->tab
, snap
) < 0)
1459 return isl_stat_error
;
1460 if (check_wraps(wraps
->mat
, n
, info
->tab
) < 0)
1461 return isl_stat_error
;
1466 /* Given a basic set i with a constraint k that is adjacent to
1467 * basic set j, check if we can wrap
1468 * both the facet corresponding to k (if "wrap_facet" is set) and basic map j
1469 * (always) around their ridges to include the other set.
1470 * If so, replace the pair of basic sets by their union.
1472 * All constraints of i (except k) are assumed to be valid or
1473 * cut constraints for j.
1474 * Wrapping the cut constraints to include basic map j may result
1475 * in constraints that are no longer valid of basic map i
1476 * we have to check that the resulting wrapping constraints are valid for i.
1477 * If "wrap_facet" is not set, then all constraints of i (except k)
1478 * are assumed to be valid for j.
1487 static enum isl_change
can_wrap_in_facet(int i
, int j
, int k
,
1488 struct isl_coalesce_info
*info
, int wrap_facet
)
1490 enum isl_change change
= isl_change_none
;
1491 struct isl_wraps wraps
;
1494 struct isl_set
*set_i
= NULL
;
1495 struct isl_set
*set_j
= NULL
;
1496 struct isl_vec
*bound
= NULL
;
1497 unsigned total
= isl_basic_map_total_dim(info
[i
].bmap
);
1499 set_i
= set_from_updated_bmap(info
[i
].bmap
, info
[i
].tab
);
1500 set_j
= set_from_updated_bmap(info
[j
].bmap
, info
[j
].tab
);
1501 ctx
= isl_basic_map_get_ctx(info
[i
].bmap
);
1502 mat
= isl_mat_alloc(ctx
, 2 * (info
[i
].bmap
->n_eq
+ info
[j
].bmap
->n_eq
) +
1503 info
[i
].bmap
->n_ineq
+ info
[j
].bmap
->n_ineq
,
1505 if (wraps_init(&wraps
, mat
, info
, i
, j
) < 0)
1507 bound
= isl_vec_alloc(ctx
, 1 + total
);
1508 if (!set_i
|| !set_j
|| !bound
)
1511 isl_seq_cpy(bound
->el
, info
[i
].bmap
->ineq
[k
], 1 + total
);
1512 isl_int_add_ui(bound
->el
[0], bound
->el
[0], 1);
1514 isl_seq_cpy(wraps
.mat
->row
[0], bound
->el
, 1 + total
);
1515 wraps
.mat
->n_row
= 1;
1517 if (add_wraps(&wraps
, &info
[j
], bound
->el
, set_i
) < 0)
1519 if (!wraps
.mat
->n_row
)
1523 if (add_wraps_around_facet(&wraps
, &info
[i
], k
,
1524 bound
->el
, set_j
) < 0)
1526 if (!wraps
.mat
->n_row
)
1530 change
= fuse(i
, j
, info
, wraps
.mat
, 0, 0);
1535 isl_set_free(set_i
);
1536 isl_set_free(set_j
);
1538 isl_vec_free(bound
);
1543 isl_vec_free(bound
);
1544 isl_set_free(set_i
);
1545 isl_set_free(set_j
);
1546 return isl_change_error
;
1549 /* Given a cut constraint t(x) >= 0 of basic map i, stored in row "w"
1550 * of wrap.mat, replace it by its relaxed version t(x) + 1 >= 0, and
1551 * add wrapping constraints to wrap.mat for all constraints
1552 * of basic map j that bound the part of basic map j that sticks out
1553 * of the cut constraint.
1554 * "set_i" is the underlying set of basic map i.
1555 * If any wrapping fails, then wraps->mat.n_row is reset to zero.
1557 * In particular, we first intersect basic map j with t(x) + 1 = 0.
1558 * If the result is empty, then t(x) >= 0 was actually a valid constraint
1559 * (with respect to the integer points), so we add t(x) >= 0 instead.
1560 * Otherwise, we wrap the constraints of basic map j that are not
1561 * redundant in this intersection and that are not already valid
1562 * for basic map i over basic map i.
1563 * Note that it is sufficient to wrap the constraints to include
1564 * basic map i, because we will only wrap the constraints that do
1565 * not include basic map i already. The wrapped constraint will
1566 * therefore be more relaxed compared to the original constraint.
1567 * Since the original constraint is valid for basic map j, so is
1568 * the wrapped constraint.
1570 static isl_stat
wrap_in_facet(struct isl_wraps
*wraps
, int w
,
1571 struct isl_coalesce_info
*info_j
, __isl_keep isl_set
*set_i
,
1572 struct isl_tab_undo
*snap
)
1574 isl_int_add_ui(wraps
->mat
->row
[w
][0], wraps
->mat
->row
[w
][0], 1);
1575 if (isl_tab_add_eq(info_j
->tab
, wraps
->mat
->row
[w
]) < 0)
1576 return isl_stat_error
;
1577 if (isl_tab_detect_redundant(info_j
->tab
) < 0)
1578 return isl_stat_error
;
1580 if (info_j
->tab
->empty
)
1581 isl_int_sub_ui(wraps
->mat
->row
[w
][0], wraps
->mat
->row
[w
][0], 1);
1582 else if (add_wraps(wraps
, info_j
, wraps
->mat
->row
[w
], set_i
) < 0)
1583 return isl_stat_error
;
1585 if (isl_tab_rollback(info_j
->tab
, snap
) < 0)
1586 return isl_stat_error
;
1591 /* Given a pair of basic maps i and j such that j sticks out
1592 * of i at n cut constraints, each time by at most one,
1593 * try to compute wrapping constraints and replace the two
1594 * basic maps by a single basic map.
1595 * The other constraints of i are assumed to be valid for j.
1596 * "set_i" is the underlying set of basic map i.
1597 * "wraps" has been initialized to be of the right size.
1599 * For each cut constraint t(x) >= 0 of i, we add the relaxed version
1600 * t(x) + 1 >= 0, along with wrapping constraints for all constraints
1601 * of basic map j that bound the part of basic map j that sticks out
1602 * of the cut constraint.
1604 * If any wrapping fails, i.e., if we cannot wrap to touch
1605 * the union, then we give up.
1606 * Otherwise, the pair of basic maps is replaced by their union.
1608 static enum isl_change
try_wrap_in_facets(int i
, int j
,
1609 struct isl_coalesce_info
*info
, struct isl_wraps
*wraps
,
1610 __isl_keep isl_set
*set_i
)
1614 struct isl_tab_undo
*snap
;
1616 total
= isl_basic_map_total_dim(info
[i
].bmap
);
1618 snap
= isl_tab_snap(info
[j
].tab
);
1620 wraps
->mat
->n_row
= 0;
1622 for (k
= 0; k
< info
[i
].bmap
->n_eq
; ++k
) {
1623 for (l
= 0; l
< 2; ++l
) {
1624 if (info
[i
].eq
[2 * k
+ l
] != STATUS_CUT
)
1626 w
= wraps
->mat
->n_row
++;
1628 isl_seq_neg(wraps
->mat
->row
[w
],
1629 info
[i
].bmap
->eq
[k
], 1 + total
);
1631 isl_seq_cpy(wraps
->mat
->row
[w
],
1632 info
[i
].bmap
->eq
[k
], 1 + total
);
1633 if (wrap_in_facet(wraps
, w
, &info
[j
], set_i
, snap
) < 0)
1634 return isl_change_error
;
1636 if (!wraps
->mat
->n_row
)
1637 return isl_change_none
;
1641 for (k
= 0; k
< info
[i
].bmap
->n_ineq
; ++k
) {
1642 if (info
[i
].ineq
[k
] != STATUS_CUT
)
1644 w
= wraps
->mat
->n_row
++;
1645 isl_seq_cpy(wraps
->mat
->row
[w
],
1646 info
[i
].bmap
->ineq
[k
], 1 + total
);
1647 if (wrap_in_facet(wraps
, w
, &info
[j
], set_i
, snap
) < 0)
1648 return isl_change_error
;
1650 if (!wraps
->mat
->n_row
)
1651 return isl_change_none
;
1654 return fuse(i
, j
, info
, wraps
->mat
, 0, 1);
1657 /* Given a pair of basic maps i and j such that j sticks out
1658 * of i at n cut constraints, each time by at most one,
1659 * try to compute wrapping constraints and replace the two
1660 * basic maps by a single basic map.
1661 * The other constraints of i are assumed to be valid for j.
1663 * The core computation is performed by try_wrap_in_facets.
1664 * This function simply extracts an underlying set representation
1665 * of basic map i and initializes the data structure for keeping
1666 * track of wrapping constraints.
1668 static enum isl_change
wrap_in_facets(int i
, int j
, int n
,
1669 struct isl_coalesce_info
*info
)
1671 enum isl_change change
= isl_change_none
;
1672 struct isl_wraps wraps
;
1675 isl_set
*set_i
= NULL
;
1676 unsigned total
= isl_basic_map_total_dim(info
[i
].bmap
);
1679 if (isl_tab_extend_cons(info
[j
].tab
, 1) < 0)
1680 return isl_change_error
;
1682 max_wrap
= 1 + 2 * info
[j
].bmap
->n_eq
+ info
[j
].bmap
->n_ineq
;
1685 set_i
= set_from_updated_bmap(info
[i
].bmap
, info
[i
].tab
);
1686 ctx
= isl_basic_map_get_ctx(info
[i
].bmap
);
1687 mat
= isl_mat_alloc(ctx
, max_wrap
, 1 + total
);
1688 if (wraps_init(&wraps
, mat
, info
, i
, j
) < 0)
1693 change
= try_wrap_in_facets(i
, j
, info
, &wraps
, set_i
);
1696 isl_set_free(set_i
);
1701 isl_set_free(set_i
);
1702 return isl_change_error
;
1705 /* Return the effect of inequality "ineq" on the tableau "tab",
1706 * after relaxing the constant term of "ineq" by one.
1708 static enum isl_ineq_type
type_of_relaxed(struct isl_tab
*tab
, isl_int
*ineq
)
1710 enum isl_ineq_type type
;
1712 isl_int_add_ui(ineq
[0], ineq
[0], 1);
1713 type
= isl_tab_ineq_type(tab
, ineq
);
1714 isl_int_sub_ui(ineq
[0], ineq
[0], 1);
1719 /* Given two basic sets i and j,
1720 * check if relaxing all the cut constraints of i by one turns
1721 * them into valid constraint for j and check if we can wrap in
1722 * the bits that are sticking out.
1723 * If so, replace the pair by their union.
1725 * We first check if all relaxed cut inequalities of i are valid for j
1726 * and then try to wrap in the intersections of the relaxed cut inequalities
1729 * During this wrapping, we consider the points of j that lie at a distance
1730 * of exactly 1 from i. In particular, we ignore the points that lie in
1731 * between this lower-dimensional space and the basic map i.
1732 * We can therefore only apply this to integer maps.
1758 * Wrapping can fail if the result of wrapping one of the facets
1759 * around its edges does not produce any new facet constraint.
1760 * In particular, this happens when we try to wrap in unbounded sets.
1762 * _______________________________________________________________________
1766 * |_| |_________________________________________________________________
1769 * The following is not an acceptable result of coalescing the above two
1770 * sets as it includes extra integer points.
1771 * _______________________________________________________________________
1776 * \______________________________________________________________________
1778 static enum isl_change
can_wrap_in_set(int i
, int j
,
1779 struct isl_coalesce_info
*info
)
1785 if (ISL_F_ISSET(info
[i
].bmap
, ISL_BASIC_MAP_RATIONAL
) ||
1786 ISL_F_ISSET(info
[j
].bmap
, ISL_BASIC_MAP_RATIONAL
))
1787 return isl_change_none
;
1789 n
= count_eq(&info
[i
], STATUS_CUT
) + count_ineq(&info
[i
], STATUS_CUT
);
1791 return isl_change_none
;
1793 total
= isl_basic_map_total_dim(info
[i
].bmap
);
1794 for (k
= 0; k
< info
[i
].bmap
->n_eq
; ++k
) {
1795 for (l
= 0; l
< 2; ++l
) {
1796 enum isl_ineq_type type
;
1798 if (info
[i
].eq
[2 * k
+ l
] != STATUS_CUT
)
1802 isl_seq_neg(info
[i
].bmap
->eq
[k
],
1803 info
[i
].bmap
->eq
[k
], 1 + total
);
1804 type
= type_of_relaxed(info
[j
].tab
,
1805 info
[i
].bmap
->eq
[k
]);
1807 isl_seq_neg(info
[i
].bmap
->eq
[k
],
1808 info
[i
].bmap
->eq
[k
], 1 + total
);
1809 if (type
== isl_ineq_error
)
1810 return isl_change_error
;
1811 if (type
!= isl_ineq_redundant
)
1812 return isl_change_none
;
1816 for (k
= 0; k
< info
[i
].bmap
->n_ineq
; ++k
) {
1817 enum isl_ineq_type type
;
1819 if (info
[i
].ineq
[k
] != STATUS_CUT
)
1822 type
= type_of_relaxed(info
[j
].tab
, info
[i
].bmap
->ineq
[k
]);
1823 if (type
== isl_ineq_error
)
1824 return isl_change_error
;
1825 if (type
!= isl_ineq_redundant
)
1826 return isl_change_none
;
1829 return wrap_in_facets(i
, j
, n
, info
);
1832 /* Check if either i or j has only cut constraints that can
1833 * be used to wrap in (a facet of) the other basic set.
1834 * if so, replace the pair by their union.
1836 static enum isl_change
check_wrap(int i
, int j
, struct isl_coalesce_info
*info
)
1838 enum isl_change change
= isl_change_none
;
1840 change
= can_wrap_in_set(i
, j
, info
);
1841 if (change
!= isl_change_none
)
1844 change
= can_wrap_in_set(j
, i
, info
);
1848 /* Check if all inequality constraints of "i" that cut "j" cease
1849 * to be cut constraints if they are relaxed by one.
1850 * If so, collect the cut constraints in "list".
1851 * The caller is responsible for allocating "list".
1853 static isl_bool
all_cut_by_one(int i
, int j
, struct isl_coalesce_info
*info
,
1859 for (l
= 0; l
< info
[i
].bmap
->n_ineq
; ++l
) {
1860 enum isl_ineq_type type
;
1862 if (info
[i
].ineq
[l
] != STATUS_CUT
)
1864 type
= type_of_relaxed(info
[j
].tab
, info
[i
].bmap
->ineq
[l
]);
1865 if (type
== isl_ineq_error
)
1866 return isl_bool_error
;
1867 if (type
!= isl_ineq_redundant
)
1868 return isl_bool_false
;
1872 return isl_bool_true
;
1875 /* Given two basic maps such that "j" has at least one equality constraint
1876 * that is adjacent to an inequality constraint of "i" and such that "i" has
1877 * exactly one inequality constraint that is adjacent to an equality
1878 * constraint of "j", check whether "i" can be extended to include "j" or
1879 * whether "j" can be wrapped into "i".
1880 * All remaining constraints of "i" and "j" are assumed to be valid
1881 * or cut constraints of the other basic map.
1882 * However, none of the equality constraints of "i" are cut constraints.
1884 * If "i" has any "cut" inequality constraints, then check if relaxing
1885 * each of them by one is sufficient for them to become valid.
1886 * If so, check if the inequality constraint adjacent to an equality
1887 * constraint of "j" along with all these cut constraints
1888 * can be relaxed by one to contain exactly "j".
1889 * Otherwise, or if this fails, check if "j" can be wrapped into "i".
1891 static enum isl_change
check_single_adj_eq(int i
, int j
,
1892 struct isl_coalesce_info
*info
)
1894 enum isl_change change
= isl_change_none
;
1901 n_cut
= count_ineq(&info
[i
], STATUS_CUT
);
1903 k
= find_ineq(&info
[i
], STATUS_ADJ_EQ
);
1906 ctx
= isl_basic_map_get_ctx(info
[i
].bmap
);
1907 relax
= isl_calloc_array(ctx
, int, 1 + n_cut
);
1909 return isl_change_error
;
1911 try_relax
= all_cut_by_one(i
, j
, info
, relax
+ 1);
1913 change
= isl_change_error
;
1915 try_relax
= isl_bool_true
;
1918 if (try_relax
&& change
== isl_change_none
)
1919 change
= is_relaxed_extension(i
, j
, 1 + n_cut
, relax
, info
);
1922 if (change
!= isl_change_none
)
1925 change
= can_wrap_in_facet(i
, j
, k
, info
, n_cut
> 0);
1930 /* At least one of the basic maps has an equality that is adjacent
1931 * to an inequality. Make sure that only one of the basic maps has
1932 * such an equality and that the other basic map has exactly one
1933 * inequality adjacent to an equality.
1934 * If the other basic map does not have such an inequality, then
1935 * check if all its constraints are either valid or cut constraints
1936 * and, if so, try wrapping in the first map into the second.
1937 * Otherwise, try to extend one basic map with the other or
1938 * wrap one basic map in the other.
1940 static enum isl_change
check_adj_eq(int i
, int j
,
1941 struct isl_coalesce_info
*info
)
1943 if (any_eq(&info
[i
], STATUS_ADJ_INEQ
) &&
1944 any_eq(&info
[j
], STATUS_ADJ_INEQ
))
1945 /* ADJ EQ TOO MANY */
1946 return isl_change_none
;
1948 if (any_eq(&info
[i
], STATUS_ADJ_INEQ
))
1949 return check_adj_eq(j
, i
, info
);
1951 /* j has an equality adjacent to an inequality in i */
1953 if (count_ineq(&info
[i
], STATUS_ADJ_EQ
) != 1) {
1954 if (all_valid_or_cut(&info
[i
]))
1955 return can_wrap_in_set(i
, j
, info
);
1956 return isl_change_none
;
1958 if (any_eq(&info
[i
], STATUS_CUT
))
1959 return isl_change_none
;
1960 if (any_ineq(&info
[j
], STATUS_ADJ_EQ
) ||
1961 any_ineq(&info
[i
], STATUS_ADJ_INEQ
) ||
1962 any_ineq(&info
[j
], STATUS_ADJ_INEQ
))
1963 /* ADJ EQ TOO MANY */
1964 return isl_change_none
;
1966 return check_single_adj_eq(i
, j
, info
);
1969 /* Disjunct "j" lies on a hyperplane that is adjacent to disjunct "i".
1970 * In particular, disjunct "i" has an inequality constraint that is adjacent
1971 * to a (combination of) equality constraint(s) of disjunct "j",
1972 * but disjunct "j" has no explicit equality constraint adjacent
1973 * to an inequality constraint of disjunct "i".
1975 * Disjunct "i" is already known not to have any equality constraints
1976 * that are adjacent to an equality or inequality constraint.
1977 * Check that, other than the inequality constraint mentioned above,
1978 * all other constraints of disjunct "i" are valid for disjunct "j".
1979 * If so, try and wrap in disjunct "j".
1981 static enum isl_change
check_ineq_adj_eq(int i
, int j
,
1982 struct isl_coalesce_info
*info
)
1986 if (any_eq(&info
[i
], STATUS_CUT
))
1987 return isl_change_none
;
1988 if (any_ineq(&info
[i
], STATUS_CUT
))
1989 return isl_change_none
;
1990 if (any_ineq(&info
[i
], STATUS_ADJ_INEQ
))
1991 return isl_change_none
;
1992 if (count_ineq(&info
[i
], STATUS_ADJ_EQ
) != 1)
1993 return isl_change_none
;
1995 k
= find_ineq(&info
[i
], STATUS_ADJ_EQ
);
1997 return can_wrap_in_facet(i
, j
, k
, info
, 0);
2000 /* The two basic maps lie on adjacent hyperplanes. In particular,
2001 * basic map "i" has an equality that lies parallel to basic map "j".
2002 * Check if we can wrap the facets around the parallel hyperplanes
2003 * to include the other set.
2005 * We perform basically the same operations as can_wrap_in_facet,
2006 * except that we don't need to select a facet of one of the sets.
2012 * If there is more than one equality of "i" adjacent to an equality of "j",
2013 * then the result will satisfy one or more equalities that are a linear
2014 * combination of these equalities. These will be encoded as pairs
2015 * of inequalities in the wrapping constraints and need to be made
2018 static enum isl_change
check_eq_adj_eq(int i
, int j
,
2019 struct isl_coalesce_info
*info
)
2022 enum isl_change change
= isl_change_none
;
2023 int detect_equalities
= 0;
2024 struct isl_wraps wraps
;
2027 struct isl_set
*set_i
= NULL
;
2028 struct isl_set
*set_j
= NULL
;
2029 struct isl_vec
*bound
= NULL
;
2030 unsigned total
= isl_basic_map_total_dim(info
[i
].bmap
);
2032 if (count_eq(&info
[i
], STATUS_ADJ_EQ
) != 1)
2033 detect_equalities
= 1;
2035 k
= find_eq(&info
[i
], STATUS_ADJ_EQ
);
2037 set_i
= set_from_updated_bmap(info
[i
].bmap
, info
[i
].tab
);
2038 set_j
= set_from_updated_bmap(info
[j
].bmap
, info
[j
].tab
);
2039 ctx
= isl_basic_map_get_ctx(info
[i
].bmap
);
2040 mat
= isl_mat_alloc(ctx
, 2 * (info
[i
].bmap
->n_eq
+ info
[j
].bmap
->n_eq
) +
2041 info
[i
].bmap
->n_ineq
+ info
[j
].bmap
->n_ineq
,
2043 if (wraps_init(&wraps
, mat
, info
, i
, j
) < 0)
2045 bound
= isl_vec_alloc(ctx
, 1 + total
);
2046 if (!set_i
|| !set_j
|| !bound
)
2050 isl_seq_neg(bound
->el
, info
[i
].bmap
->eq
[k
/ 2], 1 + total
);
2052 isl_seq_cpy(bound
->el
, info
[i
].bmap
->eq
[k
/ 2], 1 + total
);
2053 isl_int_add_ui(bound
->el
[0], bound
->el
[0], 1);
2055 isl_seq_cpy(wraps
.mat
->row
[0], bound
->el
, 1 + total
);
2056 wraps
.mat
->n_row
= 1;
2058 if (add_wraps(&wraps
, &info
[j
], bound
->el
, set_i
) < 0)
2060 if (!wraps
.mat
->n_row
)
2063 isl_int_sub_ui(bound
->el
[0], bound
->el
[0], 1);
2064 isl_seq_neg(bound
->el
, bound
->el
, 1 + total
);
2066 isl_seq_cpy(wraps
.mat
->row
[wraps
.mat
->n_row
], bound
->el
, 1 + total
);
2069 if (add_wraps(&wraps
, &info
[i
], bound
->el
, set_j
) < 0)
2071 if (!wraps
.mat
->n_row
)
2074 change
= fuse(i
, j
, info
, wraps
.mat
, detect_equalities
, 0);
2077 error
: change
= isl_change_error
;
2082 isl_set_free(set_i
);
2083 isl_set_free(set_j
);
2084 isl_vec_free(bound
);
2089 /* Initialize the "eq" and "ineq" fields of "info".
2091 static void init_status(struct isl_coalesce_info
*info
)
2093 info
->eq
= info
->ineq
= NULL
;
2096 /* Set info->eq to the positions of the equalities of info->bmap
2097 * with respect to the basic map represented by "tab".
2098 * If info->eq has already been computed, then do not compute it again.
2100 static void set_eq_status_in(struct isl_coalesce_info
*info
,
2101 struct isl_tab
*tab
)
2105 info
->eq
= eq_status_in(info
->bmap
, tab
);
2108 /* Set info->ineq to the positions of the inequalities of info->bmap
2109 * with respect to the basic map represented by "tab".
2110 * If info->ineq has already been computed, then do not compute it again.
2112 static void set_ineq_status_in(struct isl_coalesce_info
*info
,
2113 struct isl_tab
*tab
)
2117 info
->ineq
= ineq_status_in(info
->bmap
, info
->tab
, tab
);
2120 /* Free the memory allocated by the "eq" and "ineq" fields of "info".
2121 * This function assumes that init_status has been called on "info" first,
2122 * after which the "eq" and "ineq" fields may or may not have been
2123 * assigned a newly allocated array.
2125 static void clear_status(struct isl_coalesce_info
*info
)
2131 /* Are all inequality constraints of the basic map represented by "info"
2132 * valid for the other basic map, except for a single constraint
2133 * that is adjacent to an inequality constraint of the other basic map?
2135 static int all_ineq_valid_or_single_adj_ineq(struct isl_coalesce_info
*info
)
2140 for (i
= 0; i
< info
->bmap
->n_ineq
; ++i
) {
2141 if (info
->ineq
[i
] == STATUS_REDUNDANT
)
2143 if (info
->ineq
[i
] == STATUS_VALID
)
2145 if (info
->ineq
[i
] != STATUS_ADJ_INEQ
)
2155 /* Basic map "i" has one or more equality constraints that separate it
2156 * from basic map "j". Check if it happens to be an extension
2158 * In particular, check that all constraints of "j" are valid for "i",
2159 * except for one inequality constraint that is adjacent
2160 * to an inequality constraints of "i".
2161 * If so, check for "i" being an extension of "j" by calling
2162 * is_adj_ineq_extension.
2164 * Clean up the memory allocated for keeping track of the status
2165 * of the constraints before returning.
2167 static enum isl_change
separating_equality(int i
, int j
,
2168 struct isl_coalesce_info
*info
)
2170 enum isl_change change
= isl_change_none
;
2172 if (all(info
[j
].eq
, 2 * info
[j
].bmap
->n_eq
, STATUS_VALID
) &&
2173 all_ineq_valid_or_single_adj_ineq(&info
[j
]))
2174 change
= is_adj_ineq_extension(j
, i
, info
);
2176 clear_status(&info
[i
]);
2177 clear_status(&info
[j
]);
2181 /* Check if the union of the given pair of basic maps
2182 * can be represented by a single basic map.
2183 * If so, replace the pair by the single basic map and return
2184 * isl_change_drop_first, isl_change_drop_second or isl_change_fuse.
2185 * Otherwise, return isl_change_none.
2186 * The two basic maps are assumed to live in the same local space.
2187 * The "eq" and "ineq" fields of info[i] and info[j] are assumed
2188 * to have been initialized by the caller, either to NULL or
2189 * to valid information.
2191 * We first check the effect of each constraint of one basic map
2192 * on the other basic map.
2193 * The constraint may be
2194 * redundant the constraint is redundant in its own
2195 * basic map and should be ignore and removed
2197 * valid all (integer) points of the other basic map
2198 * satisfy the constraint
2199 * separate no (integer) point of the other basic map
2200 * satisfies the constraint
2201 * cut some but not all points of the other basic map
2202 * satisfy the constraint
2203 * adj_eq the given constraint is adjacent (on the outside)
2204 * to an equality of the other basic map
2205 * adj_ineq the given constraint is adjacent (on the outside)
2206 * to an inequality of the other basic map
2208 * We consider seven cases in which we can replace the pair by a single
2209 * basic map. We ignore all "redundant" constraints.
2211 * 1. all constraints of one basic map are valid
2212 * => the other basic map is a subset and can be removed
2214 * 2. all constraints of both basic maps are either "valid" or "cut"
2215 * and the facets corresponding to the "cut" constraints
2216 * of one of the basic maps lies entirely inside the other basic map
2217 * => the pair can be replaced by a basic map consisting
2218 * of the valid constraints in both basic maps
2220 * 3. there is a single pair of adjacent inequalities
2221 * (all other constraints are "valid")
2222 * => the pair can be replaced by a basic map consisting
2223 * of the valid constraints in both basic maps
2225 * 4. one basic map has a single adjacent inequality, while the other
2226 * constraints are "valid". The other basic map has some
2227 * "cut" constraints, but replacing the adjacent inequality by
2228 * its opposite and adding the valid constraints of the other
2229 * basic map results in a subset of the other basic map
2230 * => the pair can be replaced by a basic map consisting
2231 * of the valid constraints in both basic maps
2233 * 5. there is a single adjacent pair of an inequality and an equality,
2234 * the other constraints of the basic map containing the inequality are
2235 * "valid". Moreover, if the inequality the basic map is relaxed
2236 * and then turned into an equality, then resulting facet lies
2237 * entirely inside the other basic map
2238 * => the pair can be replaced by the basic map containing
2239 * the inequality, with the inequality relaxed.
2241 * 6. there is a single inequality adjacent to an equality,
2242 * the other constraints of the basic map containing the inequality are
2243 * "valid". Moreover, the facets corresponding to both
2244 * the inequality and the equality can be wrapped around their
2245 * ridges to include the other basic map
2246 * => the pair can be replaced by a basic map consisting
2247 * of the valid constraints in both basic maps together
2248 * with all wrapping constraints
2250 * 7. one of the basic maps extends beyond the other by at most one.
2251 * Moreover, the facets corresponding to the cut constraints and
2252 * the pieces of the other basic map at offset one from these cut
2253 * constraints can be wrapped around their ridges to include
2254 * the union of the two basic maps
2255 * => the pair can be replaced by a basic map consisting
2256 * of the valid constraints in both basic maps together
2257 * with all wrapping constraints
2259 * 8. the two basic maps live in adjacent hyperplanes. In principle
2260 * such sets can always be combined through wrapping, but we impose
2261 * that there is only one such pair, to avoid overeager coalescing.
2263 * Throughout the computation, we maintain a collection of tableaus
2264 * corresponding to the basic maps. When the basic maps are dropped
2265 * or combined, the tableaus are modified accordingly.
2267 static enum isl_change
coalesce_local_pair_reuse(int i
, int j
,
2268 struct isl_coalesce_info
*info
)
2270 enum isl_change change
= isl_change_none
;
2272 set_ineq_status_in(&info
[i
], info
[j
].tab
);
2273 if (info
[i
].bmap
->n_ineq
&& !info
[i
].ineq
)
2275 if (any_ineq(&info
[i
], STATUS_ERROR
))
2277 if (any_ineq(&info
[i
], STATUS_SEPARATE
))
2280 set_ineq_status_in(&info
[j
], info
[i
].tab
);
2281 if (info
[j
].bmap
->n_ineq
&& !info
[j
].ineq
)
2283 if (any_ineq(&info
[j
], STATUS_ERROR
))
2285 if (any_ineq(&info
[j
], STATUS_SEPARATE
))
2288 set_eq_status_in(&info
[i
], info
[j
].tab
);
2289 if (info
[i
].bmap
->n_eq
&& !info
[i
].eq
)
2291 if (any_eq(&info
[i
], STATUS_ERROR
))
2294 set_eq_status_in(&info
[j
], info
[i
].tab
);
2295 if (info
[j
].bmap
->n_eq
&& !info
[j
].eq
)
2297 if (any_eq(&info
[j
], STATUS_ERROR
))
2300 if (any_eq(&info
[i
], STATUS_SEPARATE
))
2301 return separating_equality(i
, j
, info
);
2302 if (any_eq(&info
[j
], STATUS_SEPARATE
))
2303 return separating_equality(j
, i
, info
);
2305 if (all(info
[i
].eq
, 2 * info
[i
].bmap
->n_eq
, STATUS_VALID
) &&
2306 all(info
[i
].ineq
, info
[i
].bmap
->n_ineq
, STATUS_VALID
)) {
2308 change
= isl_change_drop_second
;
2309 } else if (all(info
[j
].eq
, 2 * info
[j
].bmap
->n_eq
, STATUS_VALID
) &&
2310 all(info
[j
].ineq
, info
[j
].bmap
->n_ineq
, STATUS_VALID
)) {
2312 change
= isl_change_drop_first
;
2313 } else if (any_eq(&info
[i
], STATUS_ADJ_EQ
)) {
2314 change
= check_eq_adj_eq(i
, j
, info
);
2315 } else if (any_eq(&info
[j
], STATUS_ADJ_EQ
)) {
2316 change
= check_eq_adj_eq(j
, i
, info
);
2317 } else if (any_eq(&info
[i
], STATUS_ADJ_INEQ
) ||
2318 any_eq(&info
[j
], STATUS_ADJ_INEQ
)) {
2319 change
= check_adj_eq(i
, j
, info
);
2320 } else if (any_ineq(&info
[i
], STATUS_ADJ_EQ
)) {
2321 change
= check_ineq_adj_eq(i
, j
, info
);
2322 } else if (any_ineq(&info
[j
], STATUS_ADJ_EQ
)) {
2323 change
= check_ineq_adj_eq(j
, i
, info
);
2324 } else if (any_ineq(&info
[i
], STATUS_ADJ_INEQ
) ||
2325 any_ineq(&info
[j
], STATUS_ADJ_INEQ
)) {
2326 change
= check_adj_ineq(i
, j
, info
);
2328 if (!any_eq(&info
[i
], STATUS_CUT
) &&
2329 !any_eq(&info
[j
], STATUS_CUT
))
2330 change
= check_facets(i
, j
, info
);
2331 if (change
== isl_change_none
)
2332 change
= check_wrap(i
, j
, info
);
2336 clear_status(&info
[i
]);
2337 clear_status(&info
[j
]);
2340 clear_status(&info
[i
]);
2341 clear_status(&info
[j
]);
2342 return isl_change_error
;
2345 /* Check if the union of the given pair of basic maps
2346 * can be represented by a single basic map.
2347 * If so, replace the pair by the single basic map and return
2348 * isl_change_drop_first, isl_change_drop_second or isl_change_fuse.
2349 * Otherwise, return isl_change_none.
2350 * The two basic maps are assumed to live in the same local space.
2352 static enum isl_change
coalesce_local_pair(int i
, int j
,
2353 struct isl_coalesce_info
*info
)
2355 init_status(&info
[i
]);
2356 init_status(&info
[j
]);
2357 return coalesce_local_pair_reuse(i
, j
, info
);
2360 /* Shift the integer division at position "div" of the basic map
2361 * represented by "info" by "shift".
2363 * That is, if the integer division has the form
2367 * then replace it by
2369 * floor((f(x) + shift * d)/d) - shift
2371 static isl_stat
shift_div(struct isl_coalesce_info
*info
, int div
,
2376 info
->bmap
= isl_basic_map_shift_div(info
->bmap
, div
, 0, shift
);
2378 return isl_stat_error
;
2380 total
= isl_basic_map_dim(info
->bmap
, isl_dim_all
);
2381 total
-= isl_basic_map_dim(info
->bmap
, isl_dim_div
);
2382 if (isl_tab_shift_var(info
->tab
, total
+ div
, shift
) < 0)
2383 return isl_stat_error
;
2388 /* If the integer division at position "div" is defined by an equality,
2389 * i.e., a stride constraint, then change the integer division expression
2390 * to have a constant term equal to zero.
2392 * Let the equality constraint be
2396 * The integer division expression is then typically of the form
2398 * a = floor((-f - c')/m)
2400 * The integer division is first shifted by t = floor(c/m),
2401 * turning the equality constraint into
2403 * c - m floor(c/m) + f + m a' = 0
2407 * (c mod m) + f + m a' = 0
2411 * a' = (-f - (c mod m))/m = floor((-f)/m)
2413 * because a' is an integer and 0 <= (c mod m) < m.
2414 * The constant term of a' can therefore be zeroed out,
2415 * but only if the integer division expression is of the expected form.
2417 static isl_stat
normalize_stride_div(struct isl_coalesce_info
*info
, int div
)
2419 isl_bool defined
, valid
;
2422 isl_int shift
, stride
;
2424 defined
= isl_basic_map_has_defining_equality(info
->bmap
, isl_dim_div
,
2427 return isl_stat_error
;
2431 return isl_stat_error
;
2432 valid
= isl_constraint_is_div_equality(c
, div
);
2433 isl_int_init(shift
);
2434 isl_int_init(stride
);
2435 isl_constraint_get_constant(c
, &shift
);
2436 isl_constraint_get_coefficient(c
, isl_dim_div
, div
, &stride
);
2437 isl_int_fdiv_q(shift
, shift
, stride
);
2438 r
= shift_div(info
, div
, shift
);
2439 isl_int_clear(stride
);
2440 isl_int_clear(shift
);
2441 isl_constraint_free(c
);
2442 if (r
< 0 || valid
< 0)
2443 return isl_stat_error
;
2446 info
->bmap
= isl_basic_map_set_div_expr_constant_num_si_inplace(
2447 info
->bmap
, div
, 0);
2449 return isl_stat_error
;
2453 /* The basic maps represented by "info1" and "info2" are known
2454 * to have the same number of integer divisions.
2455 * Check if pairs of integer divisions are equal to each other
2456 * despite the fact that they differ by a rational constant.
2458 * In particular, look for any pair of integer divisions that
2459 * only differ in their constant terms.
2460 * If either of these integer divisions is defined
2461 * by stride constraints, then modify it to have a zero constant term.
2462 * If both are defined by stride constraints then in the end they will have
2463 * the same (zero) constant term.
2465 static isl_stat
harmonize_stride_divs(struct isl_coalesce_info
*info1
,
2466 struct isl_coalesce_info
*info2
)
2470 n
= isl_basic_map_dim(info1
->bmap
, isl_dim_div
);
2471 for (i
= 0; i
< n
; ++i
) {
2472 isl_bool known
, harmonize
;
2474 known
= isl_basic_map_div_is_known(info1
->bmap
, i
);
2475 if (known
>= 0 && known
)
2476 known
= isl_basic_map_div_is_known(info2
->bmap
, i
);
2478 return isl_stat_error
;
2481 harmonize
= isl_basic_map_equal_div_expr_except_constant(
2482 info1
->bmap
, i
, info2
->bmap
, i
);
2484 return isl_stat_error
;
2487 if (normalize_stride_div(info1
, i
) < 0)
2488 return isl_stat_error
;
2489 if (normalize_stride_div(info2
, i
) < 0)
2490 return isl_stat_error
;
2496 /* If "shift" is an integer constant, then shift the integer division
2497 * at position "div" of the basic map represented by "info" by "shift".
2498 * If "shift" is not an integer constant, then do nothing.
2499 * If "shift" is equal to zero, then no shift needs to be performed either.
2501 * That is, if the integer division has the form
2505 * then replace it by
2507 * floor((f(x) + shift * d)/d) - shift
2509 static isl_stat
shift_if_cst_int(struct isl_coalesce_info
*info
, int div
,
2510 __isl_keep isl_aff
*shift
)
2517 cst
= isl_aff_is_cst(shift
);
2518 if (cst
< 0 || !cst
)
2519 return cst
< 0 ? isl_stat_error
: isl_stat_ok
;
2521 c
= isl_aff_get_constant_val(shift
);
2522 cst
= isl_val_is_int(c
);
2523 if (cst
>= 0 && cst
)
2524 cst
= isl_bool_not(isl_val_is_zero(c
));
2525 if (cst
< 0 || !cst
) {
2527 return cst
< 0 ? isl_stat_error
: isl_stat_ok
;
2531 r
= isl_val_get_num_isl_int(c
, &d
);
2533 r
= shift_div(info
, div
, d
);
2541 /* Check if some of the divs in the basic map represented by "info1"
2542 * are shifts of the corresponding divs in the basic map represented
2543 * by "info2", taking into account the equality constraints "eq1" of "info1"
2544 * and "eq2" of "info2". If so, align them with those of "info2".
2545 * "info1" and "info2" are assumed to have the same number
2546 * of integer divisions.
2548 * An integer division is considered to be a shift of another integer
2549 * division if, after simplification with respect to the equality
2550 * constraints of the other basic map, one is equal to the other
2553 * In particular, for each pair of integer divisions, if both are known,
2554 * have the same denominator and are not already equal to each other,
2555 * simplify each with respect to the equality constraints
2556 * of the other basic map. If the difference is an integer constant,
2557 * then move this difference outside.
2558 * That is, if, after simplification, one integer division is of the form
2560 * floor((f(x) + c_1)/d)
2562 * while the other is of the form
2564 * floor((f(x) + c_2)/d)
2566 * and n = (c_2 - c_1)/d is an integer, then replace the first
2567 * integer division by
2569 * floor((f_1(x) + c_1 + n * d)/d) - n,
2571 * where floor((f_1(x) + c_1 + n * d)/d) = floor((f2(x) + c_2)/d)
2572 * after simplification with respect to the equality constraints.
2574 static isl_stat
harmonize_divs_with_hulls(struct isl_coalesce_info
*info1
,
2575 struct isl_coalesce_info
*info2
, __isl_keep isl_basic_set
*eq1
,
2576 __isl_keep isl_basic_set
*eq2
)
2580 isl_local_space
*ls1
, *ls2
;
2582 total
= isl_basic_map_total_dim(info1
->bmap
);
2583 ls1
= isl_local_space_wrap(isl_basic_map_get_local_space(info1
->bmap
));
2584 ls2
= isl_local_space_wrap(isl_basic_map_get_local_space(info2
->bmap
));
2585 for (i
= 0; i
< info1
->bmap
->n_div
; ++i
) {
2587 isl_aff
*div1
, *div2
;
2589 if (!isl_local_space_div_is_known(ls1
, i
) ||
2590 !isl_local_space_div_is_known(ls2
, i
))
2592 if (isl_int_ne(info1
->bmap
->div
[i
][0], info2
->bmap
->div
[i
][0]))
2594 if (isl_seq_eq(info1
->bmap
->div
[i
] + 1,
2595 info2
->bmap
->div
[i
] + 1, 1 + total
))
2597 div1
= isl_local_space_get_div(ls1
, i
);
2598 div2
= isl_local_space_get_div(ls2
, i
);
2599 div1
= isl_aff_substitute_equalities(div1
,
2600 isl_basic_set_copy(eq2
));
2601 div2
= isl_aff_substitute_equalities(div2
,
2602 isl_basic_set_copy(eq1
));
2603 div2
= isl_aff_sub(div2
, div1
);
2604 r
= shift_if_cst_int(info1
, i
, div2
);
2609 isl_local_space_free(ls1
);
2610 isl_local_space_free(ls2
);
2612 if (i
< info1
->bmap
->n_div
)
2613 return isl_stat_error
;
2617 /* Check if some of the divs in the basic map represented by "info1"
2618 * are shifts of the corresponding divs in the basic map represented
2619 * by "info2". If so, align them with those of "info2".
2620 * Only do this if "info1" and "info2" have the same number
2621 * of integer divisions.
2623 * An integer division is considered to be a shift of another integer
2624 * division if, after simplification with respect to the equality
2625 * constraints of the other basic map, one is equal to the other
2628 * First check if pairs of integer divisions are equal to each other
2629 * despite the fact that they differ by a rational constant.
2630 * If so, try and arrange for them to have the same constant term.
2632 * Then, extract the equality constraints and continue with
2633 * harmonize_divs_with_hulls.
2635 * If the equality constraints of both basic maps are the same,
2636 * then there is no need to perform any shifting since
2637 * the coefficients of the integer divisions should have been
2638 * reduced in the same way.
2640 static isl_stat
harmonize_divs(struct isl_coalesce_info
*info1
,
2641 struct isl_coalesce_info
*info2
)
2644 isl_basic_map
*bmap1
, *bmap2
;
2645 isl_basic_set
*eq1
, *eq2
;
2648 if (!info1
->bmap
|| !info2
->bmap
)
2649 return isl_stat_error
;
2651 if (info1
->bmap
->n_div
!= info2
->bmap
->n_div
)
2653 if (info1
->bmap
->n_div
== 0)
2656 if (harmonize_stride_divs(info1
, info2
) < 0)
2657 return isl_stat_error
;
2659 bmap1
= isl_basic_map_copy(info1
->bmap
);
2660 bmap2
= isl_basic_map_copy(info2
->bmap
);
2661 eq1
= isl_basic_map_wrap(isl_basic_map_plain_affine_hull(bmap1
));
2662 eq2
= isl_basic_map_wrap(isl_basic_map_plain_affine_hull(bmap2
));
2663 equal
= isl_basic_set_plain_is_equal(eq1
, eq2
);
2669 r
= harmonize_divs_with_hulls(info1
, info2
, eq1
, eq2
);
2670 isl_basic_set_free(eq1
);
2671 isl_basic_set_free(eq2
);
2676 /* Do the two basic maps live in the same local space, i.e.,
2677 * do they have the same (known) divs?
2678 * If either basic map has any unknown divs, then we can only assume
2679 * that they do not live in the same local space.
2681 static isl_bool
same_divs(__isl_keep isl_basic_map
*bmap1
,
2682 __isl_keep isl_basic_map
*bmap2
)
2688 if (!bmap1
|| !bmap2
)
2689 return isl_bool_error
;
2690 if (bmap1
->n_div
!= bmap2
->n_div
)
2691 return isl_bool_false
;
2693 if (bmap1
->n_div
== 0)
2694 return isl_bool_true
;
2696 known
= isl_basic_map_divs_known(bmap1
);
2697 if (known
< 0 || !known
)
2699 known
= isl_basic_map_divs_known(bmap2
);
2700 if (known
< 0 || !known
)
2703 total
= isl_basic_map_total_dim(bmap1
);
2704 for (i
= 0; i
< bmap1
->n_div
; ++i
)
2705 if (!isl_seq_eq(bmap1
->div
[i
], bmap2
->div
[i
], 2 + total
))
2706 return isl_bool_false
;
2708 return isl_bool_true
;
2711 /* Assuming that "tab" contains the equality constraints and
2712 * the initial inequality constraints of "bmap", copy the remaining
2713 * inequality constraints of "bmap" to "Tab".
2715 static isl_stat
copy_ineq(struct isl_tab
*tab
, __isl_keep isl_basic_map
*bmap
)
2720 return isl_stat_error
;
2722 n_ineq
= tab
->n_con
- tab
->n_eq
;
2723 for (i
= n_ineq
; i
< bmap
->n_ineq
; ++i
)
2724 if (isl_tab_add_ineq(tab
, bmap
->ineq
[i
]) < 0)
2725 return isl_stat_error
;
2730 /* Description of an integer division that is added
2731 * during an expansion.
2732 * "pos" is the position of the corresponding variable.
2733 * "cst" indicates whether this integer division has a fixed value.
2734 * "val" contains the fixed value, if the value is fixed.
2736 struct isl_expanded
{
2742 /* For each of the "n" integer division variables "expanded",
2743 * if the variable has a fixed value, then add two inequality
2744 * constraints expressing the fixed value.
2745 * Otherwise, add the corresponding div constraints.
2746 * The caller is responsible for removing the div constraints
2747 * that it added for all these "n" integer divisions.
2749 * The div constraints and the pair of inequality constraints
2750 * forcing the fixed value cannot both be added for a given variable
2751 * as the combination may render some of the original constraints redundant.
2752 * These would then be ignored during the coalescing detection,
2753 * while they could remain in the fused result.
2755 * The two added inequality constraints are
2760 * with "a" the variable and "v" its fixed value.
2761 * The facet corresponding to one of these two constraints is selected
2762 * in the tableau to ensure that the pair of inequality constraints
2763 * is treated as an equality constraint.
2765 * The information in info->ineq is thrown away because it was
2766 * computed in terms of div constraints, while some of those
2767 * have now been replaced by these pairs of inequality constraints.
2769 static isl_stat
fix_constant_divs(struct isl_coalesce_info
*info
,
2770 int n
, struct isl_expanded
*expanded
)
2776 o_div
= isl_basic_map_offset(info
->bmap
, isl_dim_div
) - 1;
2777 ineq
= isl_vec_alloc(isl_tab_get_ctx(info
->tab
), 1 + info
->tab
->n_var
);
2779 return isl_stat_error
;
2780 isl_seq_clr(ineq
->el
+ 1, info
->tab
->n_var
);
2782 for (i
= 0; i
< n
; ++i
) {
2783 if (!expanded
[i
].cst
) {
2784 info
->bmap
= isl_basic_map_extend_constraints(
2786 if (isl_basic_map_add_div_constraints(info
->bmap
,
2787 expanded
[i
].pos
- o_div
) < 0)
2790 isl_int_set_si(ineq
->el
[1 + expanded
[i
].pos
], -1);
2791 isl_int_set(ineq
->el
[0], expanded
[i
].val
);
2792 info
->bmap
= isl_basic_map_add_ineq(info
->bmap
,
2794 isl_int_set_si(ineq
->el
[1 + expanded
[i
].pos
], 1);
2795 isl_int_neg(ineq
->el
[0], expanded
[i
].val
);
2796 info
->bmap
= isl_basic_map_add_ineq(info
->bmap
,
2798 isl_int_set_si(ineq
->el
[1 + expanded
[i
].pos
], 0);
2800 if (copy_ineq(info
->tab
, info
->bmap
) < 0)
2802 if (expanded
[i
].cst
&&
2803 isl_tab_select_facet(info
->tab
, info
->tab
->n_con
- 1) < 0)
2812 return i
< n
? isl_stat_error
: isl_stat_ok
;
2815 /* Insert the "n" integer division variables "expanded"
2816 * into info->tab and info->bmap and
2817 * update info->ineq with respect to the redundant constraints
2818 * in the resulting tableau.
2819 * "bmap" contains the result of this insertion in info->bmap,
2820 * while info->bmap is the original version
2821 * of "bmap", i.e., the one that corresponds to the current
2822 * state of info->tab. The number of constraints in info->bmap
2823 * is assumed to be the same as the number of constraints
2824 * in info->tab. This is required to be able to detect
2825 * the extra constraints in "bmap".
2827 * In particular, introduce extra variables corresponding
2828 * to the extra integer divisions and add the div constraints
2829 * that were added to "bmap" after info->tab was created
2831 * Furthermore, check if these extra integer divisions happen
2832 * to attain a fixed integer value in info->tab.
2833 * If so, replace the corresponding div constraints by pairs
2834 * of inequality constraints that fix these
2835 * integer divisions to their single integer values.
2836 * Replace info->bmap by "bmap" to match the changes to info->tab.
2837 * info->ineq was computed without a tableau and therefore
2838 * does not take into account the redundant constraints
2839 * in the tableau. Mark them here.
2840 * There is no need to check the newly added div constraints
2841 * since they cannot be redundant.
2842 * The redundancy check is not performed when constants have been discovered
2843 * since info->ineq is completely thrown away in this case.
2845 static isl_stat
tab_insert_divs(struct isl_coalesce_info
*info
,
2846 int n
, struct isl_expanded
*expanded
, __isl_take isl_basic_map
*bmap
)
2850 struct isl_tab_undo
*snap
;
2854 return isl_stat_error
;
2855 if (info
->bmap
->n_eq
+ info
->bmap
->n_ineq
!= info
->tab
->n_con
)
2856 isl_die(isl_basic_map_get_ctx(bmap
), isl_error_internal
,
2857 "original tableau does not correspond "
2858 "to original basic map", goto error
);
2860 if (isl_tab_extend_vars(info
->tab
, n
) < 0)
2862 if (isl_tab_extend_cons(info
->tab
, 2 * n
) < 0)
2865 for (i
= 0; i
< n
; ++i
) {
2866 if (isl_tab_insert_var(info
->tab
, expanded
[i
].pos
) < 0)
2870 snap
= isl_tab_snap(info
->tab
);
2872 n_ineq
= info
->tab
->n_con
- info
->tab
->n_eq
;
2873 if (copy_ineq(info
->tab
, bmap
) < 0)
2876 isl_basic_map_free(info
->bmap
);
2880 for (i
= 0; i
< n
; ++i
) {
2881 expanded
[i
].cst
= isl_tab_is_constant(info
->tab
,
2882 expanded
[i
].pos
, &expanded
[i
].val
);
2883 if (expanded
[i
].cst
< 0)
2884 return isl_stat_error
;
2885 if (expanded
[i
].cst
)
2890 if (isl_tab_rollback(info
->tab
, snap
) < 0)
2891 return isl_stat_error
;
2892 info
->bmap
= isl_basic_map_cow(info
->bmap
);
2893 if (isl_basic_map_free_inequality(info
->bmap
, 2 * n
) < 0)
2894 return isl_stat_error
;
2896 return fix_constant_divs(info
, n
, expanded
);
2899 n_eq
= info
->bmap
->n_eq
;
2900 for (i
= 0; i
< n_ineq
; ++i
) {
2901 if (isl_tab_is_redundant(info
->tab
, n_eq
+ i
))
2902 info
->ineq
[i
] = STATUS_REDUNDANT
;
2907 isl_basic_map_free(bmap
);
2908 return isl_stat_error
;
2911 /* Expand info->tab and info->bmap in the same way "bmap" was expanded
2912 * in isl_basic_map_expand_divs using the expansion "exp" and
2913 * update info->ineq with respect to the redundant constraints
2914 * in the resulting tableau. info->bmap is the original version
2915 * of "bmap", i.e., the one that corresponds to the current
2916 * state of info->tab. The number of constraints in info->bmap
2917 * is assumed to be the same as the number of constraints
2918 * in info->tab. This is required to be able to detect
2919 * the extra constraints in "bmap".
2921 * Extract the positions where extra local variables are introduced
2922 * from "exp" and call tab_insert_divs.
2924 static isl_stat
expand_tab(struct isl_coalesce_info
*info
, int *exp
,
2925 __isl_take isl_basic_map
*bmap
)
2928 struct isl_expanded
*expanded
;
2931 unsigned total
, pos
, n_div
;
2934 total
= isl_basic_map_dim(bmap
, isl_dim_all
);
2935 n_div
= isl_basic_map_dim(bmap
, isl_dim_div
);
2936 pos
= total
- n_div
;
2937 extra_var
= total
- info
->tab
->n_var
;
2938 n
= n_div
- extra_var
;
2940 ctx
= isl_basic_map_get_ctx(bmap
);
2941 expanded
= isl_calloc_array(ctx
, struct isl_expanded
, extra_var
);
2942 if (extra_var
&& !expanded
)
2947 for (j
= 0; j
< n_div
; ++j
) {
2948 if (i
< n
&& exp
[i
] == j
) {
2952 expanded
[k
++].pos
= pos
+ j
;
2955 for (k
= 0; k
< extra_var
; ++k
)
2956 isl_int_init(expanded
[k
].val
);
2958 r
= tab_insert_divs(info
, extra_var
, expanded
, bmap
);
2960 for (k
= 0; k
< extra_var
; ++k
)
2961 isl_int_clear(expanded
[k
].val
);
2966 isl_basic_map_free(bmap
);
2967 return isl_stat_error
;
2970 /* Check if the union of the basic maps represented by info[i] and info[j]
2971 * can be represented by a single basic map,
2972 * after expanding the divs of info[i] to match those of info[j].
2973 * If so, replace the pair by the single basic map and return
2974 * isl_change_drop_first, isl_change_drop_second or isl_change_fuse.
2975 * Otherwise, return isl_change_none.
2977 * The caller has already checked for info[j] being a subset of info[i].
2978 * If some of the divs of info[j] are unknown, then the expanded info[i]
2979 * will not have the corresponding div constraints. The other patterns
2980 * therefore cannot apply. Skip the computation in this case.
2982 * The expansion is performed using the divs "div" and expansion "exp"
2983 * computed by the caller.
2984 * info[i].bmap has already been expanded and the result is passed in
2986 * The "eq" and "ineq" fields of info[i] reflect the status of
2987 * the constraints of the expanded "bmap" with respect to info[j].tab.
2988 * However, inequality constraints that are redundant in info[i].tab
2989 * have not yet been marked as such because no tableau was available.
2991 * Replace info[i].bmap by "bmap" and expand info[i].tab as well,
2992 * updating info[i].ineq with respect to the redundant constraints.
2993 * Then try and coalesce the expanded info[i] with info[j],
2994 * reusing the information in info[i].eq and info[i].ineq.
2995 * If this does not result in any coalescing or if it results in info[j]
2996 * getting dropped (which should not happen in practice, since the case
2997 * of info[j] being a subset of info[i] has already been checked by
2998 * the caller), then revert info[i] to its original state.
3000 static enum isl_change
coalesce_expand_tab_divs(__isl_take isl_basic_map
*bmap
,
3001 int i
, int j
, struct isl_coalesce_info
*info
, __isl_keep isl_mat
*div
,
3005 isl_basic_map
*bmap_i
;
3006 struct isl_tab_undo
*snap
;
3007 enum isl_change change
= isl_change_none
;
3009 known
= isl_basic_map_divs_known(info
[j
].bmap
);
3010 if (known
< 0 || !known
) {
3011 clear_status(&info
[i
]);
3012 isl_basic_map_free(bmap
);
3013 return known
< 0 ? isl_change_error
: isl_change_none
;
3016 bmap_i
= isl_basic_map_copy(info
[i
].bmap
);
3017 snap
= isl_tab_snap(info
[i
].tab
);
3018 if (expand_tab(&info
[i
], exp
, bmap
) < 0)
3019 change
= isl_change_error
;
3021 init_status(&info
[j
]);
3022 if (change
== isl_change_none
)
3023 change
= coalesce_local_pair_reuse(i
, j
, info
);
3025 clear_status(&info
[i
]);
3026 if (change
!= isl_change_none
&& change
!= isl_change_drop_second
) {
3027 isl_basic_map_free(bmap_i
);
3029 isl_basic_map_free(info
[i
].bmap
);
3030 info
[i
].bmap
= bmap_i
;
3032 if (isl_tab_rollback(info
[i
].tab
, snap
) < 0)
3033 change
= isl_change_error
;
3039 /* Check if the union of "bmap" and the basic map represented by info[j]
3040 * can be represented by a single basic map,
3041 * after expanding the divs of "bmap" to match those of info[j].
3042 * If so, replace the pair by the single basic map and return
3043 * isl_change_drop_first, isl_change_drop_second or isl_change_fuse.
3044 * Otherwise, return isl_change_none.
3046 * In particular, check if the expanded "bmap" contains the basic map
3047 * represented by the tableau info[j].tab.
3048 * The expansion is performed using the divs "div" and expansion "exp"
3049 * computed by the caller.
3050 * Then we check if all constraints of the expanded "bmap" are valid for
3053 * If "i" is not equal to -1, then "bmap" is equal to info[i].bmap.
3054 * In this case, the positions of the constraints of info[i].bmap
3055 * with respect to the basic map represented by info[j] are stored
3058 * If the expanded "bmap" does not contain the basic map
3059 * represented by the tableau info[j].tab and if "i" is not -1,
3060 * i.e., if the original "bmap" is info[i].bmap, then expand info[i].tab
3061 * as well and check if that results in coalescing.
3063 static enum isl_change
coalesce_with_expanded_divs(
3064 __isl_keep isl_basic_map
*bmap
, int i
, int j
,
3065 struct isl_coalesce_info
*info
, __isl_keep isl_mat
*div
, int *exp
)
3067 enum isl_change change
= isl_change_none
;
3068 struct isl_coalesce_info info_local
, *info_i
;
3070 info_i
= i
>= 0 ? &info
[i
] : &info_local
;
3071 init_status(info_i
);
3072 bmap
= isl_basic_map_copy(bmap
);
3073 bmap
= isl_basic_map_expand_divs(bmap
, isl_mat_copy(div
), exp
);
3074 bmap
= isl_basic_map_mark_final(bmap
);
3079 info_local
.bmap
= bmap
;
3080 info_i
->eq
= eq_status_in(bmap
, info
[j
].tab
);
3081 if (bmap
->n_eq
&& !info_i
->eq
)
3083 if (any_eq(info_i
, STATUS_ERROR
))
3085 if (any_eq(info_i
, STATUS_SEPARATE
))
3088 info_i
->ineq
= ineq_status_in(bmap
, NULL
, info
[j
].tab
);
3089 if (bmap
->n_ineq
&& !info_i
->ineq
)
3091 if (any_ineq(info_i
, STATUS_ERROR
))
3093 if (any_ineq(info_i
, STATUS_SEPARATE
))
3096 if (all(info_i
->eq
, 2 * bmap
->n_eq
, STATUS_VALID
) &&
3097 all(info_i
->ineq
, bmap
->n_ineq
, STATUS_VALID
)) {
3099 change
= isl_change_drop_second
;
3102 if (change
== isl_change_none
&& i
!= -1)
3103 return coalesce_expand_tab_divs(bmap
, i
, j
, info
, div
, exp
);
3106 isl_basic_map_free(bmap
);
3107 clear_status(info_i
);
3110 isl_basic_map_free(bmap
);
3111 clear_status(info_i
);
3112 return isl_change_error
;
3115 /* Check if the union of "bmap_i" and the basic map represented by info[j]
3116 * can be represented by a single basic map,
3117 * after aligning the divs of "bmap_i" to match those of info[j].
3118 * If so, replace the pair by the single basic map and return
3119 * isl_change_drop_first, isl_change_drop_second or isl_change_fuse.
3120 * Otherwise, return isl_change_none.
3122 * In particular, check if "bmap_i" contains the basic map represented by
3123 * info[j] after aligning the divs of "bmap_i" to those of info[j].
3124 * Note that this can only succeed if the number of divs of "bmap_i"
3125 * is smaller than (or equal to) the number of divs of info[j].
3127 * We first check if the divs of "bmap_i" are all known and form a subset
3128 * of those of info[j].bmap. If so, we pass control over to
3129 * coalesce_with_expanded_divs.
3131 * If "i" is not equal to -1, then "bmap" is equal to info[i].bmap.
3133 static enum isl_change
coalesce_after_aligning_divs(
3134 __isl_keep isl_basic_map
*bmap_i
, int i
, int j
,
3135 struct isl_coalesce_info
*info
)
3138 isl_mat
*div_i
, *div_j
, *div
;
3142 enum isl_change change
;
3144 known
= isl_basic_map_divs_known(bmap_i
);
3145 if (known
< 0 || !known
)
3148 ctx
= isl_basic_map_get_ctx(bmap_i
);
3150 div_i
= isl_basic_map_get_divs(bmap_i
);
3151 div_j
= isl_basic_map_get_divs(info
[j
].bmap
);
3153 if (!div_i
|| !div_j
)
3156 exp1
= isl_alloc_array(ctx
, int, div_i
->n_row
);
3157 exp2
= isl_alloc_array(ctx
, int, div_j
->n_row
);
3158 if ((div_i
->n_row
&& !exp1
) || (div_j
->n_row
&& !exp2
))
3161 div
= isl_merge_divs(div_i
, div_j
, exp1
, exp2
);
3165 if (div
->n_row
== div_j
->n_row
)
3166 change
= coalesce_with_expanded_divs(bmap_i
,
3167 i
, j
, info
, div
, exp1
);
3169 change
= isl_change_none
;
3173 isl_mat_free(div_i
);
3174 isl_mat_free(div_j
);
3181 isl_mat_free(div_i
);
3182 isl_mat_free(div_j
);
3185 return isl_change_error
;
3188 /* Check if basic map "j" is a subset of basic map "i" after
3189 * exploiting the extra equalities of "j" to simplify the divs of "i".
3190 * If so, remove basic map "j" and return isl_change_drop_second.
3192 * If "j" does not have any equalities or if they are the same
3193 * as those of "i", then we cannot exploit them to simplify the divs.
3194 * Similarly, if there are no divs in "i", then they cannot be simplified.
3195 * If, on the other hand, the affine hulls of "i" and "j" do not intersect,
3196 * then "j" cannot be a subset of "i".
3198 * Otherwise, we intersect "i" with the affine hull of "j" and then
3199 * check if "j" is a subset of the result after aligning the divs.
3200 * If so, then "j" is definitely a subset of "i" and can be removed.
3201 * Note that if after intersection with the affine hull of "j".
3202 * "i" still has more divs than "j", then there is no way we can
3203 * align the divs of "i" to those of "j".
3205 static enum isl_change
coalesce_subset_with_equalities(int i
, int j
,
3206 struct isl_coalesce_info
*info
)
3208 isl_basic_map
*hull_i
, *hull_j
, *bmap_i
;
3210 enum isl_change change
;
3212 if (info
[j
].bmap
->n_eq
== 0)
3213 return isl_change_none
;
3214 if (info
[i
].bmap
->n_div
== 0)
3215 return isl_change_none
;
3217 hull_i
= isl_basic_map_copy(info
[i
].bmap
);
3218 hull_i
= isl_basic_map_plain_affine_hull(hull_i
);
3219 hull_j
= isl_basic_map_copy(info
[j
].bmap
);
3220 hull_j
= isl_basic_map_plain_affine_hull(hull_j
);
3222 hull_j
= isl_basic_map_intersect(hull_j
, isl_basic_map_copy(hull_i
));
3223 equal
= isl_basic_map_plain_is_equal(hull_i
, hull_j
);
3224 empty
= isl_basic_map_plain_is_empty(hull_j
);
3225 isl_basic_map_free(hull_i
);
3227 if (equal
< 0 || equal
|| empty
< 0 || empty
) {
3228 isl_basic_map_free(hull_j
);
3229 if (equal
< 0 || empty
< 0)
3230 return isl_change_error
;
3231 return isl_change_none
;
3234 bmap_i
= isl_basic_map_copy(info
[i
].bmap
);
3235 bmap_i
= isl_basic_map_intersect(bmap_i
, hull_j
);
3237 return isl_change_error
;
3239 if (bmap_i
->n_div
> info
[j
].bmap
->n_div
) {
3240 isl_basic_map_free(bmap_i
);
3241 return isl_change_none
;
3244 change
= coalesce_after_aligning_divs(bmap_i
, -1, j
, info
);
3246 isl_basic_map_free(bmap_i
);
3251 /* Check if the union of and the basic maps represented by info[i] and info[j]
3252 * can be represented by a single basic map, by aligning or equating
3253 * their integer divisions.
3254 * If so, replace the pair by the single basic map and return
3255 * isl_change_drop_first, isl_change_drop_second or isl_change_fuse.
3256 * Otherwise, return isl_change_none.
3258 * Note that we only perform any test if the number of divs is different
3259 * in the two basic maps. In case the number of divs is the same,
3260 * we have already established that the divs are different
3261 * in the two basic maps.
3262 * In particular, if the number of divs of basic map i is smaller than
3263 * the number of divs of basic map j, then we check if j is a subset of i
3266 static enum isl_change
coalesce_divs(int i
, int j
,
3267 struct isl_coalesce_info
*info
)
3269 enum isl_change change
= isl_change_none
;
3271 if (info
[i
].bmap
->n_div
< info
[j
].bmap
->n_div
)
3272 change
= coalesce_after_aligning_divs(info
[i
].bmap
, i
, j
, info
);
3273 if (change
!= isl_change_none
)
3276 if (info
[j
].bmap
->n_div
< info
[i
].bmap
->n_div
)
3277 change
= coalesce_after_aligning_divs(info
[j
].bmap
, j
, i
, info
);
3278 if (change
!= isl_change_none
)
3279 return invert_change(change
);
3281 change
= coalesce_subset_with_equalities(i
, j
, info
);
3282 if (change
!= isl_change_none
)
3285 change
= coalesce_subset_with_equalities(j
, i
, info
);
3286 if (change
!= isl_change_none
)
3287 return invert_change(change
);
3289 return isl_change_none
;
3292 /* Does "bmap" involve any divs that themselves refer to divs?
3294 static isl_bool
has_nested_div(__isl_keep isl_basic_map
*bmap
)
3300 total
= isl_basic_map_dim(bmap
, isl_dim_all
);
3301 n_div
= isl_basic_map_dim(bmap
, isl_dim_div
);
3304 for (i
= 0; i
< n_div
; ++i
)
3305 if (isl_seq_first_non_zero(bmap
->div
[i
] + 2 + total
,
3307 return isl_bool_true
;
3309 return isl_bool_false
;
3312 /* Return a list of affine expressions, one for each integer division
3313 * in "bmap_i". For each integer division that also appears in "bmap_j",
3314 * the affine expression is set to NaN. The number of NaNs in the list
3315 * is equal to the number of integer divisions in "bmap_j".
3316 * For the other integer divisions of "bmap_i", the corresponding
3317 * element in the list is a purely affine expression equal to the integer
3318 * division in "hull".
3319 * If no such list can be constructed, then the number of elements
3320 * in the returned list is smaller than the number of integer divisions
3323 static __isl_give isl_aff_list
*set_up_substitutions(
3324 __isl_keep isl_basic_map
*bmap_i
, __isl_keep isl_basic_map
*bmap_j
,
3325 __isl_take isl_basic_map
*hull
)
3327 unsigned n_div_i
, n_div_j
, total
;
3329 isl_local_space
*ls
;
3330 isl_basic_set
*wrap_hull
;
3338 ctx
= isl_basic_map_get_ctx(hull
);
3340 n_div_i
= isl_basic_map_dim(bmap_i
, isl_dim_div
);
3341 n_div_j
= isl_basic_map_dim(bmap_j
, isl_dim_div
);
3342 total
= isl_basic_map_total_dim(bmap_i
) - n_div_i
;
3344 ls
= isl_basic_map_get_local_space(bmap_i
);
3345 ls
= isl_local_space_wrap(ls
);
3346 wrap_hull
= isl_basic_map_wrap(hull
);
3348 aff_nan
= isl_aff_nan_on_domain(isl_local_space_copy(ls
));
3349 list
= isl_aff_list_alloc(ctx
, n_div_i
);
3352 for (i
= 0; i
< n_div_i
; ++i
) {
3356 isl_basic_map_equal_div_expr_part(bmap_i
, i
, bmap_j
, j
,
3359 list
= isl_aff_list_add(list
, isl_aff_copy(aff_nan
));
3362 if (n_div_i
- i
<= n_div_j
- j
)
3365 aff
= isl_local_space_get_div(ls
, i
);
3366 aff
= isl_aff_substitute_equalities(aff
,
3367 isl_basic_set_copy(wrap_hull
));
3368 aff
= isl_aff_floor(aff
);
3371 if (isl_aff_dim(aff
, isl_dim_div
) != 0) {
3376 list
= isl_aff_list_add(list
, aff
);
3379 isl_aff_free(aff_nan
);
3380 isl_local_space_free(ls
);
3381 isl_basic_set_free(wrap_hull
);
3385 isl_aff_free(aff_nan
);
3386 isl_local_space_free(ls
);
3387 isl_basic_set_free(wrap_hull
);
3388 isl_aff_list_free(list
);
3392 /* Add variables to info->bmap and info->tab corresponding to the elements
3393 * in "list" that are not set to NaN.
3394 * "extra_var" is the number of these elements.
3395 * "dim" is the offset in the variables of "tab" where we should
3396 * start considering the elements in "list".
3397 * When this function returns, the total number of variables in "tab"
3398 * is equal to "dim" plus the number of elements in "list".
3400 * The newly added existentially quantified variables are not given
3401 * an explicit representation because the corresponding div constraints
3402 * do not appear in info->bmap. These constraints are not added
3403 * to info->bmap because for internal consistency, they would need to
3404 * be added to info->tab as well, where they could combine with the equality
3405 * that is added later to result in constraints that do not hold
3406 * in the original input.
3408 static isl_stat
add_sub_vars(struct isl_coalesce_info
*info
,
3409 __isl_keep isl_aff_list
*list
, int dim
, int extra_var
)
3414 space
= isl_basic_map_get_space(info
->bmap
);
3415 info
->bmap
= isl_basic_map_cow(info
->bmap
);
3416 info
->bmap
= isl_basic_map_extend_space(info
->bmap
, space
,
3419 return isl_stat_error
;
3420 n
= isl_aff_list_n_aff(list
);
3421 for (i
= 0; i
< n
; ++i
) {
3425 aff
= isl_aff_list_get_aff(list
, i
);
3426 is_nan
= isl_aff_is_nan(aff
);
3429 return isl_stat_error
;
3433 if (isl_tab_insert_var(info
->tab
, dim
+ i
) < 0)
3434 return isl_stat_error
;
3435 d
= isl_basic_map_alloc_div(info
->bmap
);
3437 return isl_stat_error
;
3438 info
->bmap
= isl_basic_map_mark_div_unknown(info
->bmap
, d
);
3440 return isl_stat_error
;
3441 for (j
= d
; j
> i
; --j
)
3442 isl_basic_map_swap_div(info
->bmap
, j
- 1, j
);
3448 /* For each element in "list" that is not set to NaN, fix the corresponding
3449 * variable in "tab" to the purely affine expression defined by the element.
3450 * "dim" is the offset in the variables of "tab" where we should
3451 * start considering the elements in "list".
3453 * This function assumes that a sufficient number of rows and
3454 * elements in the constraint array are available in the tableau.
3456 static int add_sub_equalities(struct isl_tab
*tab
,
3457 __isl_keep isl_aff_list
*list
, int dim
)
3464 n
= isl_aff_list_n_aff(list
);
3466 ctx
= isl_tab_get_ctx(tab
);
3467 sub
= isl_vec_alloc(ctx
, 1 + dim
+ n
);
3470 isl_seq_clr(sub
->el
+ 1 + dim
, n
);
3472 for (i
= 0; i
< n
; ++i
) {
3473 aff
= isl_aff_list_get_aff(list
, i
);
3476 if (isl_aff_is_nan(aff
)) {
3480 isl_seq_cpy(sub
->el
, aff
->v
->el
+ 1, 1 + dim
);
3481 isl_int_neg(sub
->el
[1 + dim
+ i
], aff
->v
->el
[0]);
3482 if (isl_tab_add_eq(tab
, sub
->el
) < 0)
3484 isl_int_set_si(sub
->el
[1 + dim
+ i
], 0);
3496 /* Add variables to info->tab and info->bmap corresponding to the elements
3497 * in "list" that are not set to NaN. The value of the added variable
3498 * in info->tab is fixed to the purely affine expression defined by the element.
3499 * "dim" is the offset in the variables of info->tab where we should
3500 * start considering the elements in "list".
3501 * When this function returns, the total number of variables in info->tab
3502 * is equal to "dim" plus the number of elements in "list".
3504 static int add_subs(struct isl_coalesce_info
*info
,
3505 __isl_keep isl_aff_list
*list
, int dim
)
3513 n
= isl_aff_list_n_aff(list
);
3514 extra_var
= n
- (info
->tab
->n_var
- dim
);
3516 if (isl_tab_extend_vars(info
->tab
, extra_var
) < 0)
3518 if (isl_tab_extend_cons(info
->tab
, 2 * extra_var
) < 0)
3520 if (add_sub_vars(info
, list
, dim
, extra_var
) < 0)
3523 return add_sub_equalities(info
->tab
, list
, dim
);
3526 /* Coalesce basic map "j" into basic map "i" after adding the extra integer
3527 * divisions in "i" but not in "j" to basic map "j", with values
3528 * specified by "list". The total number of elements in "list"
3529 * is equal to the number of integer divisions in "i", while the number
3530 * of NaN elements in the list is equal to the number of integer divisions
3533 * If no coalescing can be performed, then we need to revert basic map "j"
3534 * to its original state. We do the same if basic map "i" gets dropped
3535 * during the coalescing, even though this should not happen in practice
3536 * since we have already checked for "j" being a subset of "i"
3537 * before we reach this stage.
3539 static enum isl_change
coalesce_with_subs(int i
, int j
,
3540 struct isl_coalesce_info
*info
, __isl_keep isl_aff_list
*list
)
3542 isl_basic_map
*bmap_j
;
3543 struct isl_tab_undo
*snap
;
3545 enum isl_change change
;
3547 bmap_j
= isl_basic_map_copy(info
[j
].bmap
);
3548 snap
= isl_tab_snap(info
[j
].tab
);
3550 dim
= isl_basic_map_dim(bmap_j
, isl_dim_all
);
3551 dim
-= isl_basic_map_dim(bmap_j
, isl_dim_div
);
3552 if (add_subs(&info
[j
], list
, dim
) < 0)
3555 change
= coalesce_local_pair(i
, j
, info
);
3556 if (change
!= isl_change_none
&& change
!= isl_change_drop_first
) {
3557 isl_basic_map_free(bmap_j
);
3559 isl_basic_map_free(info
[j
].bmap
);
3560 info
[j
].bmap
= bmap_j
;
3562 if (isl_tab_rollback(info
[j
].tab
, snap
) < 0)
3563 return isl_change_error
;
3568 isl_basic_map_free(bmap_j
);
3569 return isl_change_error
;
3572 /* Check if we can coalesce basic map "j" into basic map "i" after copying
3573 * those extra integer divisions in "i" that can be simplified away
3574 * using the extra equalities in "j".
3575 * All divs are assumed to be known and not contain any nested divs.
3577 * We first check if there are any extra equalities in "j" that we
3578 * can exploit. Then we check if every integer division in "i"
3579 * either already appears in "j" or can be simplified using the
3580 * extra equalities to a purely affine expression.
3581 * If these tests succeed, then we try to coalesce the two basic maps
3582 * by introducing extra dimensions in "j" corresponding to
3583 * the extra integer divsisions "i" fixed to the corresponding
3584 * purely affine expression.
3586 static enum isl_change
check_coalesce_into_eq(int i
, int j
,
3587 struct isl_coalesce_info
*info
)
3589 unsigned n_div_i
, n_div_j
;
3590 isl_basic_map
*hull_i
, *hull_j
;
3593 enum isl_change change
;
3595 n_div_i
= isl_basic_map_dim(info
[i
].bmap
, isl_dim_div
);
3596 n_div_j
= isl_basic_map_dim(info
[j
].bmap
, isl_dim_div
);
3597 if (n_div_i
<= n_div_j
)
3598 return isl_change_none
;
3599 if (info
[j
].bmap
->n_eq
== 0)
3600 return isl_change_none
;
3602 hull_i
= isl_basic_map_copy(info
[i
].bmap
);
3603 hull_i
= isl_basic_map_plain_affine_hull(hull_i
);
3604 hull_j
= isl_basic_map_copy(info
[j
].bmap
);
3605 hull_j
= isl_basic_map_plain_affine_hull(hull_j
);
3607 hull_j
= isl_basic_map_intersect(hull_j
, isl_basic_map_copy(hull_i
));
3608 equal
= isl_basic_map_plain_is_equal(hull_i
, hull_j
);
3609 empty
= isl_basic_map_plain_is_empty(hull_j
);
3610 isl_basic_map_free(hull_i
);
3612 if (equal
< 0 || empty
< 0)
3614 if (equal
|| empty
) {
3615 isl_basic_map_free(hull_j
);
3616 return isl_change_none
;
3619 list
= set_up_substitutions(info
[i
].bmap
, info
[j
].bmap
, hull_j
);
3621 return isl_change_error
;
3622 if (isl_aff_list_n_aff(list
) < n_div_i
)
3623 change
= isl_change_none
;
3625 change
= coalesce_with_subs(i
, j
, info
, list
);
3627 isl_aff_list_free(list
);
3631 isl_basic_map_free(hull_j
);
3632 return isl_change_error
;
3635 /* Check if we can coalesce basic maps "i" and "j" after copying
3636 * those extra integer divisions in one of the basic maps that can
3637 * be simplified away using the extra equalities in the other basic map.
3638 * We require all divs to be known in both basic maps.
3639 * Furthermore, to simplify the comparison of div expressions,
3640 * we do not allow any nested integer divisions.
3642 static enum isl_change
check_coalesce_eq(int i
, int j
,
3643 struct isl_coalesce_info
*info
)
3645 isl_bool known
, nested
;
3646 enum isl_change change
;
3648 known
= isl_basic_map_divs_known(info
[i
].bmap
);
3649 if (known
< 0 || !known
)
3650 return known
< 0 ? isl_change_error
: isl_change_none
;
3651 known
= isl_basic_map_divs_known(info
[j
].bmap
);
3652 if (known
< 0 || !known
)
3653 return known
< 0 ? isl_change_error
: isl_change_none
;
3654 nested
= has_nested_div(info
[i
].bmap
);
3655 if (nested
< 0 || nested
)
3656 return nested
< 0 ? isl_change_error
: isl_change_none
;
3657 nested
= has_nested_div(info
[j
].bmap
);
3658 if (nested
< 0 || nested
)
3659 return nested
< 0 ? isl_change_error
: isl_change_none
;
3661 change
= check_coalesce_into_eq(i
, j
, info
);
3662 if (change
!= isl_change_none
)
3664 change
= check_coalesce_into_eq(j
, i
, info
);
3665 if (change
!= isl_change_none
)
3666 return invert_change(change
);
3668 return isl_change_none
;
3671 /* Check if the union of the given pair of basic maps
3672 * can be represented by a single basic map.
3673 * If so, replace the pair by the single basic map and return
3674 * isl_change_drop_first, isl_change_drop_second or isl_change_fuse.
3675 * Otherwise, return isl_change_none.
3677 * We first check if the two basic maps live in the same local space,
3678 * after aligning the divs that differ by only an integer constant.
3679 * If so, we do the complete check. Otherwise, we check if they have
3680 * the same number of integer divisions and can be coalesced, if one is
3681 * an obvious subset of the other or if the extra integer divisions
3682 * of one basic map can be simplified away using the extra equalities
3683 * of the other basic map.
3685 static enum isl_change
coalesce_pair(int i
, int j
,
3686 struct isl_coalesce_info
*info
)
3689 enum isl_change change
;
3691 if (harmonize_divs(&info
[i
], &info
[j
]) < 0)
3692 return isl_change_error
;
3693 same
= same_divs(info
[i
].bmap
, info
[j
].bmap
);
3695 return isl_change_error
;
3697 return coalesce_local_pair(i
, j
, info
);
3699 if (info
[i
].bmap
->n_div
== info
[j
].bmap
->n_div
) {
3700 change
= coalesce_local_pair(i
, j
, info
);
3701 if (change
!= isl_change_none
)
3705 change
= coalesce_divs(i
, j
, info
);
3706 if (change
!= isl_change_none
)
3709 return check_coalesce_eq(i
, j
, info
);
3712 /* Return the maximum of "a" and "b".
3714 static int isl_max(int a
, int b
)
3716 return a
> b
? a
: b
;
3719 /* Pairwise coalesce the basic maps in the range [start1, end1[ of "info"
3720 * with those in the range [start2, end2[, skipping basic maps
3721 * that have been removed (either before or within this function).
3723 * For each basic map i in the first range, we check if it can be coalesced
3724 * with respect to any previously considered basic map j in the second range.
3725 * If i gets dropped (because it was a subset of some j), then
3726 * we can move on to the next basic map.
3727 * If j gets dropped, we need to continue checking against the other
3728 * previously considered basic maps.
3729 * If the two basic maps got fused, then we recheck the fused basic map
3730 * against the previously considered basic maps, starting at i + 1
3731 * (even if start2 is greater than i + 1).
3733 static int coalesce_range(isl_ctx
*ctx
, struct isl_coalesce_info
*info
,
3734 int start1
, int end1
, int start2
, int end2
)
3738 for (i
= end1
- 1; i
>= start1
; --i
) {
3739 if (info
[i
].removed
)
3741 for (j
= isl_max(i
+ 1, start2
); j
< end2
; ++j
) {
3742 enum isl_change changed
;
3744 if (info
[j
].removed
)
3746 if (info
[i
].removed
)
3747 isl_die(ctx
, isl_error_internal
,
3748 "basic map unexpectedly removed",
3750 changed
= coalesce_pair(i
, j
, info
);
3752 case isl_change_error
:
3754 case isl_change_none
:
3755 case isl_change_drop_second
:
3757 case isl_change_drop_first
:
3760 case isl_change_fuse
:
3770 /* Pairwise coalesce the basic maps described by the "n" elements of "info".
3772 * We consider groups of basic maps that live in the same apparent
3773 * affine hull and we first coalesce within such a group before we
3774 * coalesce the elements in the group with elements of previously
3775 * considered groups. If a fuse happens during the second phase,
3776 * then we also reconsider the elements within the group.
3778 static int coalesce(isl_ctx
*ctx
, int n
, struct isl_coalesce_info
*info
)
3782 for (end
= n
; end
> 0; end
= start
) {
3784 while (start
>= 1 &&
3785 info
[start
- 1].hull_hash
== info
[start
].hull_hash
)
3787 if (coalesce_range(ctx
, info
, start
, end
, start
, end
) < 0)
3789 if (coalesce_range(ctx
, info
, start
, end
, end
, n
) < 0)
3796 /* Update the basic maps in "map" based on the information in "info".
3797 * In particular, remove the basic maps that have been marked removed and
3798 * update the others based on the information in the corresponding tableau.
3799 * Since we detected implicit equalities without calling
3800 * isl_basic_map_gauss, we need to do it now.
3801 * Also call isl_basic_map_simplify if we may have lost the definition
3802 * of one or more integer divisions.
3804 static __isl_give isl_map
*update_basic_maps(__isl_take isl_map
*map
,
3805 int n
, struct isl_coalesce_info
*info
)
3812 for (i
= n
- 1; i
>= 0; --i
) {
3813 if (info
[i
].removed
) {
3814 isl_basic_map_free(map
->p
[i
]);
3815 if (i
!= map
->n
- 1)
3816 map
->p
[i
] = map
->p
[map
->n
- 1];
3821 info
[i
].bmap
= isl_basic_map_update_from_tab(info
[i
].bmap
,
3823 info
[i
].bmap
= isl_basic_map_gauss(info
[i
].bmap
, NULL
);
3824 if (info
[i
].simplify
)
3825 info
[i
].bmap
= isl_basic_map_simplify(info
[i
].bmap
);
3826 info
[i
].bmap
= isl_basic_map_finalize(info
[i
].bmap
);
3828 return isl_map_free(map
);
3829 ISL_F_SET(info
[i
].bmap
, ISL_BASIC_MAP_NO_IMPLICIT
);
3830 ISL_F_SET(info
[i
].bmap
, ISL_BASIC_MAP_NO_REDUNDANT
);
3831 isl_basic_map_free(map
->p
[i
]);
3832 map
->p
[i
] = info
[i
].bmap
;
3833 info
[i
].bmap
= NULL
;
3839 /* For each pair of basic maps in the map, check if the union of the two
3840 * can be represented by a single basic map.
3841 * If so, replace the pair by the single basic map and start over.
3843 * We factor out any (hidden) common factor from the constraint
3844 * coefficients to improve the detection of adjacent constraints.
3846 * Since we are constructing the tableaus of the basic maps anyway,
3847 * we exploit them to detect implicit equalities and redundant constraints.
3848 * This also helps the coalescing as it can ignore the redundant constraints.
3849 * In order to avoid confusion, we make all implicit equalities explicit
3850 * in the basic maps. We don't call isl_basic_map_gauss, though,
3851 * as that may affect the number of constraints.
3852 * This means that we have to call isl_basic_map_gauss at the end
3853 * of the computation (in update_basic_maps) to ensure that
3854 * the basic maps are not left in an unexpected state.
3855 * For each basic map, we also compute the hash of the apparent affine hull
3856 * for use in coalesce.
3858 __isl_give isl_map
*isl_map_coalesce(__isl_take isl_map
*map
)
3863 struct isl_coalesce_info
*info
= NULL
;
3865 map
= isl_map_remove_empty_parts(map
);
3872 ctx
= isl_map_get_ctx(map
);
3873 map
= isl_map_sort_divs(map
);
3874 map
= isl_map_cow(map
);
3881 info
= isl_calloc_array(map
->ctx
, struct isl_coalesce_info
, n
);
3885 for (i
= 0; i
< map
->n
; ++i
) {
3886 map
->p
[i
] = isl_basic_map_reduce_coefficients(map
->p
[i
]);
3889 info
[i
].bmap
= isl_basic_map_copy(map
->p
[i
]);
3890 info
[i
].tab
= isl_tab_from_basic_map(info
[i
].bmap
, 0);
3893 if (!ISL_F_ISSET(info
[i
].bmap
, ISL_BASIC_MAP_NO_IMPLICIT
))
3894 if (isl_tab_detect_implicit_equalities(info
[i
].tab
) < 0)
3896 info
[i
].bmap
= isl_tab_make_equalities_explicit(info
[i
].tab
,
3900 if (!ISL_F_ISSET(info
[i
].bmap
, ISL_BASIC_MAP_NO_REDUNDANT
))
3901 if (isl_tab_detect_redundant(info
[i
].tab
) < 0)
3903 if (coalesce_info_set_hull_hash(&info
[i
]) < 0)
3906 for (i
= map
->n
- 1; i
>= 0; --i
)
3907 if (info
[i
].tab
->empty
)
3910 if (coalesce(ctx
, n
, info
) < 0)
3913 map
= update_basic_maps(map
, n
, info
);
3915 clear_coalesce_info(n
, info
);
3919 clear_coalesce_info(n
, info
);
3924 /* For each pair of basic sets in the set, check if the union of the two
3925 * can be represented by a single basic set.
3926 * If so, replace the pair by the single basic set and start over.
3928 struct isl_set
*isl_set_coalesce(struct isl_set
*set
)
3930 return set_from_map(isl_map_coalesce(set_to_map(set
)));