isl_union_{map,set}_n_{map,set}: return isl_size
[isl.git] / isl_coalesce.c
blob8f346ef8ea710777897ba6aed1436e1a7637af9e
1 /*
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"
23 #include <isl_seq.h>
24 #include <isl/options.h>
25 #include "isl_tab.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
41 #define STATUS_CUT 4
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);
48 switch (type) {
49 default:
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)
68 int k, l;
69 int *eq;
70 isl_size dim;
72 dim = isl_basic_map_dim(bmap_i, isl_dim_all);
73 if (dim < 0)
74 return NULL;
76 eq = isl_calloc_array(bmap_i->ctx, int, 2 * bmap_i->n_eq);
77 if (!eq)
78 return NULL;
80 for (k = 0; k < bmap_i->n_eq; ++k) {
81 for (l = 0; l < 2; ++l) {
82 isl_seq_neg(bmap_i->eq[k], bmap_i->eq[k], 1+dim);
83 eq[2 * k + l] = status_in(bmap_i->eq[k], tab_j);
84 if (eq[2 * k + l] == STATUS_ERROR)
85 goto error;
89 return eq;
90 error:
91 free(eq);
92 return NULL;
95 /* Compute the position of the inequalities of basic map "bmap_i"
96 * (also represented by "tab_i", if not NULL) with respect to the basic map
97 * represented by "tab_j".
99 static int *ineq_status_in(__isl_keep isl_basic_map *bmap_i,
100 struct isl_tab *tab_i, struct isl_tab *tab_j)
102 int k;
103 unsigned n_eq = bmap_i->n_eq;
104 int *ineq = isl_calloc_array(bmap_i->ctx, int, bmap_i->n_ineq);
106 if (!ineq)
107 return NULL;
109 for (k = 0; k < bmap_i->n_ineq; ++k) {
110 if (tab_i && isl_tab_is_redundant(tab_i, n_eq + k)) {
111 ineq[k] = STATUS_REDUNDANT;
112 continue;
114 ineq[k] = status_in(bmap_i->ineq[k], tab_j);
115 if (ineq[k] == STATUS_ERROR)
116 goto error;
117 if (ineq[k] == STATUS_SEPARATE)
118 break;
121 return ineq;
122 error:
123 free(ineq);
124 return NULL;
127 static int any(int *con, unsigned len, int status)
129 int i;
131 for (i = 0; i < len ; ++i)
132 if (con[i] == status)
133 return 1;
134 return 0;
137 /* Return the first position of "status" in the list "con" of length "len".
138 * Return -1 if there is no such entry.
140 static int find(int *con, unsigned len, int status)
142 int i;
144 for (i = 0; i < len ; ++i)
145 if (con[i] == status)
146 return i;
147 return -1;
150 static int count(int *con, unsigned len, int status)
152 int i;
153 int c = 0;
155 for (i = 0; i < len ; ++i)
156 if (con[i] == status)
157 c++;
158 return c;
161 static int all(int *con, unsigned len, int status)
163 int i;
165 for (i = 0; i < len ; ++i) {
166 if (con[i] == STATUS_REDUNDANT)
167 continue;
168 if (con[i] != status)
169 return 0;
171 return 1;
174 /* Internal information associated to a basic map in a map
175 * that is to be coalesced by isl_map_coalesce.
177 * "bmap" is the basic map itself (or NULL if "removed" is set)
178 * "tab" is the corresponding tableau (or NULL if "removed" is set)
179 * "hull_hash" identifies the affine space in which "bmap" lives.
180 * "removed" is set if this basic map has been removed from the map
181 * "simplify" is set if this basic map may have some unknown integer
182 * divisions that were not present in the input basic maps. The basic
183 * map should then be simplified such that we may be able to find
184 * a definition among the constraints.
186 * "eq" and "ineq" are only set if we are currently trying to coalesce
187 * this basic map with another basic map, in which case they represent
188 * the position of the inequalities of this basic map with respect to
189 * the other basic map. The number of elements in the "eq" array
190 * is twice the number of equalities in the "bmap", corresponding
191 * to the two inequalities that make up each equality.
193 struct isl_coalesce_info {
194 isl_basic_map *bmap;
195 struct isl_tab *tab;
196 uint32_t hull_hash;
197 int removed;
198 int simplify;
199 int *eq;
200 int *ineq;
203 /* Is there any (half of an) equality constraint in the description
204 * of the basic map represented by "info" that
205 * has position "status" with respect to the other basic map?
207 static int any_eq(struct isl_coalesce_info *info, int status)
209 unsigned n_eq;
211 n_eq = isl_basic_map_n_equality(info->bmap);
212 return any(info->eq, 2 * n_eq, status);
215 /* Is there any inequality constraint in the description
216 * of the basic map represented by "info" that
217 * has position "status" with respect to the other basic map?
219 static int any_ineq(struct isl_coalesce_info *info, int status)
221 unsigned n_ineq;
223 n_ineq = isl_basic_map_n_inequality(info->bmap);
224 return any(info->ineq, n_ineq, status);
227 /* Return the position of the first half on an equality constraint
228 * in the description of the basic map represented by "info" that
229 * has position "status" with respect to the other basic map.
230 * The returned value is twice the position of the equality constraint
231 * plus zero for the negative half and plus one for the positive half.
232 * Return -1 if there is no such entry.
234 static int find_eq(struct isl_coalesce_info *info, int status)
236 unsigned n_eq;
238 n_eq = isl_basic_map_n_equality(info->bmap);
239 return find(info->eq, 2 * n_eq, status);
242 /* Return the position of the first inequality constraint in the description
243 * of the basic map represented by "info" that
244 * has position "status" with respect to the other basic map.
245 * Return -1 if there is no such entry.
247 static int find_ineq(struct isl_coalesce_info *info, int status)
249 unsigned n_ineq;
251 n_ineq = isl_basic_map_n_inequality(info->bmap);
252 return find(info->ineq, n_ineq, status);
255 /* Return the number of (halves of) equality constraints in the description
256 * of the basic map represented by "info" that
257 * have position "status" with respect to the other basic map.
259 static int count_eq(struct isl_coalesce_info *info, int status)
261 unsigned n_eq;
263 n_eq = isl_basic_map_n_equality(info->bmap);
264 return count(info->eq, 2 * n_eq, status);
267 /* Return the number of inequality constraints in the description
268 * of the basic map represented by "info" that
269 * have position "status" with respect to the other basic map.
271 static int count_ineq(struct isl_coalesce_info *info, int status)
273 unsigned n_ineq;
275 n_ineq = isl_basic_map_n_inequality(info->bmap);
276 return count(info->ineq, n_ineq, status);
279 /* Are all non-redundant constraints of the basic map represented by "info"
280 * either valid or cut constraints with respect to the other basic map?
282 static int all_valid_or_cut(struct isl_coalesce_info *info)
284 int i;
286 for (i = 0; i < 2 * info->bmap->n_eq; ++i) {
287 if (info->eq[i] == STATUS_REDUNDANT)
288 continue;
289 if (info->eq[i] == STATUS_VALID)
290 continue;
291 if (info->eq[i] == STATUS_CUT)
292 continue;
293 return 0;
296 for (i = 0; i < info->bmap->n_ineq; ++i) {
297 if (info->ineq[i] == STATUS_REDUNDANT)
298 continue;
299 if (info->ineq[i] == STATUS_VALID)
300 continue;
301 if (info->ineq[i] == STATUS_CUT)
302 continue;
303 return 0;
306 return 1;
309 /* Compute the hash of the (apparent) affine hull of info->bmap (with
310 * the existentially quantified variables removed) and store it
311 * in info->hash.
313 static int coalesce_info_set_hull_hash(struct isl_coalesce_info *info)
315 isl_basic_map *hull;
316 isl_size n_div;
318 hull = isl_basic_map_copy(info->bmap);
319 hull = isl_basic_map_plain_affine_hull(hull);
320 n_div = isl_basic_map_dim(hull, isl_dim_div);
321 if (n_div < 0)
322 hull = isl_basic_map_free(hull);
323 hull = isl_basic_map_drop_constraints_involving_dims(hull,
324 isl_dim_div, 0, n_div);
325 info->hull_hash = isl_basic_map_get_hash(hull);
326 isl_basic_map_free(hull);
328 return hull ? 0 : -1;
331 /* Free all the allocated memory in an array
332 * of "n" isl_coalesce_info elements.
334 static void clear_coalesce_info(int n, struct isl_coalesce_info *info)
336 int i;
338 if (!info)
339 return;
341 for (i = 0; i < n; ++i) {
342 isl_basic_map_free(info[i].bmap);
343 isl_tab_free(info[i].tab);
346 free(info);
349 /* Drop the basic map represented by "info".
350 * That is, clear the memory associated to the entry and
351 * mark it as having been removed.
352 * Gaussian elimination needs to be performed on the basic map
353 * before it gets freed because it may have been put
354 * in an inconsistent state in isl_map_coalesce while it may
355 * be shared with other maps.
357 static void drop(struct isl_coalesce_info *info)
359 info->bmap = isl_basic_map_gauss(info->bmap, NULL);
360 info->bmap = isl_basic_map_free(info->bmap);
361 isl_tab_free(info->tab);
362 info->tab = NULL;
363 info->removed = 1;
366 /* Exchange the information in "info1" with that in "info2".
368 static void exchange(struct isl_coalesce_info *info1,
369 struct isl_coalesce_info *info2)
371 struct isl_coalesce_info info;
373 info = *info1;
374 *info1 = *info2;
375 *info2 = info;
378 /* This type represents the kind of change that has been performed
379 * while trying to coalesce two basic maps.
381 * isl_change_none: nothing was changed
382 * isl_change_drop_first: the first basic map was removed
383 * isl_change_drop_second: the second basic map was removed
384 * isl_change_fuse: the two basic maps were replaced by a new basic map.
386 enum isl_change {
387 isl_change_error = -1,
388 isl_change_none = 0,
389 isl_change_drop_first,
390 isl_change_drop_second,
391 isl_change_fuse,
394 /* Update "change" based on an interchange of the first and the second
395 * basic map. That is, interchange isl_change_drop_first and
396 * isl_change_drop_second.
398 static enum isl_change invert_change(enum isl_change change)
400 switch (change) {
401 case isl_change_error:
402 return isl_change_error;
403 case isl_change_none:
404 return isl_change_none;
405 case isl_change_drop_first:
406 return isl_change_drop_second;
407 case isl_change_drop_second:
408 return isl_change_drop_first;
409 case isl_change_fuse:
410 return isl_change_fuse;
413 return isl_change_error;
416 /* Add the valid constraints of the basic map represented by "info"
417 * to "bmap". "len" is the size of the constraints.
418 * If only one of the pair of inequalities that make up an equality
419 * is valid, then add that inequality.
421 static __isl_give isl_basic_map *add_valid_constraints(
422 __isl_take isl_basic_map *bmap, struct isl_coalesce_info *info,
423 unsigned len)
425 int k, l;
427 if (!bmap)
428 return NULL;
430 for (k = 0; k < info->bmap->n_eq; ++k) {
431 if (info->eq[2 * k] == STATUS_VALID &&
432 info->eq[2 * k + 1] == STATUS_VALID) {
433 l = isl_basic_map_alloc_equality(bmap);
434 if (l < 0)
435 return isl_basic_map_free(bmap);
436 isl_seq_cpy(bmap->eq[l], info->bmap->eq[k], len);
437 } else if (info->eq[2 * k] == STATUS_VALID) {
438 l = isl_basic_map_alloc_inequality(bmap);
439 if (l < 0)
440 return isl_basic_map_free(bmap);
441 isl_seq_neg(bmap->ineq[l], info->bmap->eq[k], len);
442 } else if (info->eq[2 * k + 1] == STATUS_VALID) {
443 l = isl_basic_map_alloc_inequality(bmap);
444 if (l < 0)
445 return isl_basic_map_free(bmap);
446 isl_seq_cpy(bmap->ineq[l], info->bmap->eq[k], len);
450 for (k = 0; k < info->bmap->n_ineq; ++k) {
451 if (info->ineq[k] != STATUS_VALID)
452 continue;
453 l = isl_basic_map_alloc_inequality(bmap);
454 if (l < 0)
455 return isl_basic_map_free(bmap);
456 isl_seq_cpy(bmap->ineq[l], info->bmap->ineq[k], len);
459 return bmap;
462 /* Is "bmap" defined by a number of (non-redundant) constraints that
463 * is greater than the number of constraints of basic maps i and j combined?
464 * Equalities are counted as two inequalities.
466 static int number_of_constraints_increases(int i, int j,
467 struct isl_coalesce_info *info,
468 __isl_keep isl_basic_map *bmap, struct isl_tab *tab)
470 int k, n_old, n_new;
472 n_old = 2 * info[i].bmap->n_eq + info[i].bmap->n_ineq;
473 n_old += 2 * info[j].bmap->n_eq + info[j].bmap->n_ineq;
475 n_new = 2 * bmap->n_eq;
476 for (k = 0; k < bmap->n_ineq; ++k)
477 if (!isl_tab_is_redundant(tab, bmap->n_eq + k))
478 ++n_new;
480 return n_new > n_old;
483 /* Replace the pair of basic maps i and j by the basic map bounded
484 * by the valid constraints in both basic maps and the constraints
485 * in extra (if not NULL).
486 * Place the fused basic map in the position that is the smallest of i and j.
488 * If "detect_equalities" is set, then look for equalities encoded
489 * as pairs of inequalities.
490 * If "check_number" is set, then the original basic maps are only
491 * replaced if the total number of constraints does not increase.
492 * While the number of integer divisions in the two basic maps
493 * is assumed to be the same, the actual definitions may be different.
494 * We only copy the definition from one of the basic map if it is
495 * the same as that of the other basic map. Otherwise, we mark
496 * the integer division as unknown and simplify the basic map
497 * in an attempt to recover the integer division definition.
499 static enum isl_change fuse(int i, int j, struct isl_coalesce_info *info,
500 __isl_keep isl_mat *extra, int detect_equalities, int check_number)
502 int k, l;
503 struct isl_basic_map *fused = NULL;
504 struct isl_tab *fused_tab = NULL;
505 isl_size total = isl_basic_map_dim(info[i].bmap, isl_dim_all);
506 unsigned extra_rows = extra ? extra->n_row : 0;
507 unsigned n_eq, n_ineq;
508 int simplify = 0;
510 if (total < 0)
511 return isl_change_error;
512 if (j < i)
513 return fuse(j, i, info, extra, detect_equalities, check_number);
515 n_eq = info[i].bmap->n_eq + info[j].bmap->n_eq;
516 n_ineq = info[i].bmap->n_ineq + info[j].bmap->n_ineq;
517 fused = isl_basic_map_alloc_space(isl_space_copy(info[i].bmap->dim),
518 info[i].bmap->n_div, n_eq, n_eq + n_ineq + extra_rows);
519 fused = add_valid_constraints(fused, &info[i], 1 + total);
520 fused = add_valid_constraints(fused, &info[j], 1 + total);
521 if (!fused)
522 goto error;
523 if (ISL_F_ISSET(info[i].bmap, ISL_BASIC_MAP_RATIONAL) &&
524 ISL_F_ISSET(info[j].bmap, ISL_BASIC_MAP_RATIONAL))
525 ISL_F_SET(fused, ISL_BASIC_MAP_RATIONAL);
527 for (k = 0; k < info[i].bmap->n_div; ++k) {
528 int l = isl_basic_map_alloc_div(fused);
529 if (l < 0)
530 goto error;
531 if (isl_seq_eq(info[i].bmap->div[k], info[j].bmap->div[k],
532 1 + 1 + total)) {
533 isl_seq_cpy(fused->div[l], info[i].bmap->div[k],
534 1 + 1 + total);
535 } else {
536 isl_int_set_si(fused->div[l][0], 0);
537 simplify = 1;
541 for (k = 0; k < extra_rows; ++k) {
542 l = isl_basic_map_alloc_inequality(fused);
543 if (l < 0)
544 goto error;
545 isl_seq_cpy(fused->ineq[l], extra->row[k], 1 + total);
548 if (detect_equalities)
549 fused = isl_basic_map_detect_inequality_pairs(fused, NULL);
550 fused = isl_basic_map_gauss(fused, NULL);
551 if (simplify || info[j].simplify) {
552 fused = isl_basic_map_simplify(fused);
553 info[i].simplify = 0;
555 fused = isl_basic_map_finalize(fused);
557 fused_tab = isl_tab_from_basic_map(fused, 0);
558 if (isl_tab_detect_redundant(fused_tab) < 0)
559 goto error;
561 if (check_number &&
562 number_of_constraints_increases(i, j, info, fused, fused_tab)) {
563 isl_tab_free(fused_tab);
564 isl_basic_map_free(fused);
565 return isl_change_none;
568 isl_basic_map_free(info[i].bmap);
569 info[i].bmap = fused;
570 isl_tab_free(info[i].tab);
571 info[i].tab = fused_tab;
572 drop(&info[j]);
574 return isl_change_fuse;
575 error:
576 isl_tab_free(fused_tab);
577 isl_basic_map_free(fused);
578 return isl_change_error;
581 /* Given a pair of basic maps i and j such that all constraints are either
582 * "valid" or "cut", check if the facets corresponding to the "cut"
583 * constraints of i lie entirely within basic map j.
584 * If so, replace the pair by the basic map consisting of the valid
585 * constraints in both basic maps.
586 * Checking whether the facet lies entirely within basic map j
587 * is performed by checking whether the constraints of basic map j
588 * are valid for the facet. These tests are performed on a rational
589 * tableau to avoid the theoretical possibility that a constraint
590 * that was considered to be a cut constraint for the entire basic map i
591 * happens to be considered to be a valid constraint for the facet,
592 * even though it cuts off the same rational points.
594 * To see that we are not introducing any extra points, call the
595 * two basic maps A and B and the resulting map U and let x
596 * be an element of U \setminus ( A \cup B ).
597 * A line connecting x with an element of A \cup B meets a facet F
598 * of either A or B. Assume it is a facet of B and let c_1 be
599 * the corresponding facet constraint. We have c_1(x) < 0 and
600 * so c_1 is a cut constraint. This implies that there is some
601 * (possibly rational) point x' satisfying the constraints of A
602 * and the opposite of c_1 as otherwise c_1 would have been marked
603 * valid for A. The line connecting x and x' meets a facet of A
604 * in a (possibly rational) point that also violates c_1, but this
605 * is impossible since all cut constraints of B are valid for all
606 * cut facets of A.
607 * In case F is a facet of A rather than B, then we can apply the
608 * above reasoning to find a facet of B separating x from A \cup B first.
610 static enum isl_change check_facets(int i, int j,
611 struct isl_coalesce_info *info)
613 int k, l;
614 struct isl_tab_undo *snap, *snap2;
615 unsigned n_eq = info[i].bmap->n_eq;
617 snap = isl_tab_snap(info[i].tab);
618 if (isl_tab_mark_rational(info[i].tab) < 0)
619 return isl_change_error;
620 snap2 = isl_tab_snap(info[i].tab);
622 for (k = 0; k < info[i].bmap->n_ineq; ++k) {
623 if (info[i].ineq[k] != STATUS_CUT)
624 continue;
625 if (isl_tab_select_facet(info[i].tab, n_eq + k) < 0)
626 return isl_change_error;
627 for (l = 0; l < info[j].bmap->n_ineq; ++l) {
628 int stat;
629 if (info[j].ineq[l] != STATUS_CUT)
630 continue;
631 stat = status_in(info[j].bmap->ineq[l], info[i].tab);
632 if (stat < 0)
633 return isl_change_error;
634 if (stat != STATUS_VALID)
635 break;
637 if (isl_tab_rollback(info[i].tab, snap2) < 0)
638 return isl_change_error;
639 if (l < info[j].bmap->n_ineq)
640 break;
643 if (k < info[i].bmap->n_ineq) {
644 if (isl_tab_rollback(info[i].tab, snap) < 0)
645 return isl_change_error;
646 return isl_change_none;
648 return fuse(i, j, info, NULL, 0, 0);
651 /* Check if info->bmap contains the basic map represented
652 * by the tableau "tab".
653 * For each equality, we check both the constraint itself
654 * (as an inequality) and its negation. Make sure the
655 * equality is returned to its original state before returning.
657 static isl_bool contains(struct isl_coalesce_info *info, struct isl_tab *tab)
659 int k;
660 isl_size dim;
661 isl_basic_map *bmap = info->bmap;
663 dim = isl_basic_map_dim(bmap, isl_dim_all);
664 if (dim < 0)
665 return isl_bool_error;
666 for (k = 0; k < bmap->n_eq; ++k) {
667 int stat;
668 isl_seq_neg(bmap->eq[k], bmap->eq[k], 1 + dim);
669 stat = status_in(bmap->eq[k], tab);
670 isl_seq_neg(bmap->eq[k], bmap->eq[k], 1 + dim);
671 if (stat < 0)
672 return isl_bool_error;
673 if (stat != STATUS_VALID)
674 return isl_bool_false;
675 stat = status_in(bmap->eq[k], tab);
676 if (stat < 0)
677 return isl_bool_error;
678 if (stat != STATUS_VALID)
679 return isl_bool_false;
682 for (k = 0; k < bmap->n_ineq; ++k) {
683 int stat;
684 if (info->ineq[k] == STATUS_REDUNDANT)
685 continue;
686 stat = status_in(bmap->ineq[k], tab);
687 if (stat < 0)
688 return isl_bool_error;
689 if (stat != STATUS_VALID)
690 return isl_bool_false;
692 return isl_bool_true;
695 /* Basic map "i" has an inequality (say "k") that is adjacent
696 * to some inequality of basic map "j". All the other inequalities
697 * are valid for "j".
698 * Check if basic map "j" forms an extension of basic map "i".
700 * Note that this function is only called if some of the equalities or
701 * inequalities of basic map "j" do cut basic map "i". The function is
702 * correct even if there are no such cut constraints, but in that case
703 * the additional checks performed by this function are overkill.
705 * In particular, we replace constraint k, say f >= 0, by constraint
706 * f <= -1, add the inequalities of "j" that are valid for "i"
707 * and check if the result is a subset of basic map "j".
708 * To improve the chances of the subset relation being detected,
709 * any variable that only attains a single integer value
710 * in the tableau of "i" is first fixed to that value.
711 * If the result is a subset, then we know that this result is exactly equal
712 * to basic map "j" since all its constraints are valid for basic map "j".
713 * By combining the valid constraints of "i" (all equalities and all
714 * inequalities except "k") and the valid constraints of "j" we therefore
715 * obtain a basic map that is equal to their union.
716 * In this case, there is no need to perform a rollback of the tableau
717 * since it is going to be destroyed in fuse().
720 * |\__ |\__
721 * | \__ | \__
722 * | \_ => | \__
723 * |_______| _ |_________\
726 * |\ |\
727 * | \ | \
728 * | \ | \
729 * | | | \
730 * | ||\ => | \
731 * | || \ | \
732 * | || | | |
733 * |__||_/ |_____/
735 static enum isl_change is_adj_ineq_extension(int i, int j,
736 struct isl_coalesce_info *info)
738 int k;
739 struct isl_tab_undo *snap;
740 unsigned n_eq = info[i].bmap->n_eq;
741 isl_size total = isl_basic_map_dim(info[i].bmap, isl_dim_all);
742 isl_stat r;
743 isl_bool super;
745 if (total < 0)
746 return isl_change_error;
747 if (isl_tab_extend_cons(info[i].tab, 1 + info[j].bmap->n_ineq) < 0)
748 return isl_change_error;
750 k = find_ineq(&info[i], STATUS_ADJ_INEQ);
751 if (k < 0)
752 isl_die(isl_basic_map_get_ctx(info[i].bmap), isl_error_internal,
753 "info[i].ineq should have exactly one STATUS_ADJ_INEQ",
754 return isl_change_error);
756 snap = isl_tab_snap(info[i].tab);
758 if (isl_tab_unrestrict(info[i].tab, n_eq + k) < 0)
759 return isl_change_error;
761 isl_seq_neg(info[i].bmap->ineq[k], info[i].bmap->ineq[k], 1 + total);
762 isl_int_sub_ui(info[i].bmap->ineq[k][0], info[i].bmap->ineq[k][0], 1);
763 r = isl_tab_add_ineq(info[i].tab, info[i].bmap->ineq[k]);
764 isl_seq_neg(info[i].bmap->ineq[k], info[i].bmap->ineq[k], 1 + total);
765 isl_int_sub_ui(info[i].bmap->ineq[k][0], info[i].bmap->ineq[k][0], 1);
766 if (r < 0)
767 return isl_change_error;
769 for (k = 0; k < info[j].bmap->n_ineq; ++k) {
770 if (info[j].ineq[k] != STATUS_VALID)
771 continue;
772 if (isl_tab_add_ineq(info[i].tab, info[j].bmap->ineq[k]) < 0)
773 return isl_change_error;
775 if (isl_tab_detect_constants(info[i].tab) < 0)
776 return isl_change_error;
778 super = contains(&info[j], info[i].tab);
779 if (super < 0)
780 return isl_change_error;
781 if (super)
782 return fuse(i, j, info, NULL, 0, 0);
784 if (isl_tab_rollback(info[i].tab, snap) < 0)
785 return isl_change_error;
787 return isl_change_none;
791 /* Both basic maps have at least one inequality with and adjacent
792 * (but opposite) inequality in the other basic map.
793 * Check that there are no cut constraints and that there is only
794 * a single pair of adjacent inequalities.
795 * If so, we can replace the pair by a single basic map described
796 * by all but the pair of adjacent inequalities.
797 * Any additional points introduced lie strictly between the two
798 * adjacent hyperplanes and can therefore be integral.
800 * ____ _____
801 * / ||\ / \
802 * / || \ / \
803 * \ || \ => \ \
804 * \ || / \ /
805 * \___||_/ \_____/
807 * The test for a single pair of adjancent inequalities is important
808 * for avoiding the combination of two basic maps like the following
810 * /|
811 * / |
812 * /__|
813 * _____
814 * | |
815 * | |
816 * |___|
818 * If there are some cut constraints on one side, then we may
819 * still be able to fuse the two basic maps, but we need to perform
820 * some additional checks in is_adj_ineq_extension.
822 static enum isl_change check_adj_ineq(int i, int j,
823 struct isl_coalesce_info *info)
825 int count_i, count_j;
826 int cut_i, cut_j;
828 count_i = count_ineq(&info[i], STATUS_ADJ_INEQ);
829 count_j = count_ineq(&info[j], STATUS_ADJ_INEQ);
831 if (count_i != 1 && count_j != 1)
832 return isl_change_none;
834 cut_i = any_eq(&info[i], STATUS_CUT) || any_ineq(&info[i], STATUS_CUT);
835 cut_j = any_eq(&info[j], STATUS_CUT) || any_ineq(&info[j], STATUS_CUT);
837 if (!cut_i && !cut_j && count_i == 1 && count_j == 1)
838 return fuse(i, j, info, NULL, 0, 0);
840 if (count_i == 1 && !cut_i)
841 return is_adj_ineq_extension(i, j, info);
843 if (count_j == 1 && !cut_j)
844 return is_adj_ineq_extension(j, i, info);
846 return isl_change_none;
849 /* Given an affine transformation matrix "T", does row "row" represent
850 * anything other than a unit vector (possibly shifted by a constant)
851 * that is not involved in any of the other rows?
853 * That is, if a constraint involves the variable corresponding to
854 * the row, then could its preimage by "T" have any coefficients
855 * that are different from those in the original constraint?
857 static int not_unique_unit_row(__isl_keep isl_mat *T, int row)
859 int i, j;
860 int len = T->n_col - 1;
862 i = isl_seq_first_non_zero(T->row[row] + 1, len);
863 if (i < 0)
864 return 1;
865 if (!isl_int_is_one(T->row[row][1 + i]) &&
866 !isl_int_is_negone(T->row[row][1 + i]))
867 return 1;
869 j = isl_seq_first_non_zero(T->row[row] + 1 + i + 1, len - (i + 1));
870 if (j >= 0)
871 return 1;
873 for (j = 1; j < T->n_row; ++j) {
874 if (j == row)
875 continue;
876 if (!isl_int_is_zero(T->row[j][1 + i]))
877 return 1;
880 return 0;
883 /* Does inequality constraint "ineq" of "bmap" involve any of
884 * the variables marked in "affected"?
885 * "total" is the total number of variables, i.e., the number
886 * of entries in "affected".
888 static isl_bool is_affected(__isl_keep isl_basic_map *bmap, int ineq,
889 int *affected, int total)
891 int i;
893 for (i = 0; i < total; ++i) {
894 if (!affected[i])
895 continue;
896 if (!isl_int_is_zero(bmap->ineq[ineq][1 + i]))
897 return isl_bool_true;
900 return isl_bool_false;
903 /* Given the compressed version of inequality constraint "ineq"
904 * of info->bmap in "v", check if the constraint can be tightened,
905 * where the compression is based on an equality constraint valid
906 * for info->tab.
907 * If so, add the tightened version of the inequality constraint
908 * to info->tab. "v" may be modified by this function.
910 * That is, if the compressed constraint is of the form
912 * m f() + c >= 0
914 * with 0 < c < m, then it is equivalent to
916 * f() >= 0
918 * This means that c can also be subtracted from the original,
919 * uncompressed constraint without affecting the integer points
920 * in info->tab. Add this tightened constraint as an extra row
921 * to info->tab to make this information explicitly available.
923 static __isl_give isl_vec *try_tightening(struct isl_coalesce_info *info,
924 int ineq, __isl_take isl_vec *v)
926 isl_ctx *ctx;
927 isl_stat r;
929 if (!v)
930 return NULL;
932 ctx = isl_vec_get_ctx(v);
933 isl_seq_gcd(v->el + 1, v->size - 1, &ctx->normalize_gcd);
934 if (isl_int_is_zero(ctx->normalize_gcd) ||
935 isl_int_is_one(ctx->normalize_gcd)) {
936 return v;
939 v = isl_vec_cow(v);
940 if (!v)
941 return NULL;
943 isl_int_fdiv_r(v->el[0], v->el[0], ctx->normalize_gcd);
944 if (isl_int_is_zero(v->el[0]))
945 return v;
947 if (isl_tab_extend_cons(info->tab, 1) < 0)
948 return isl_vec_free(v);
950 isl_int_sub(info->bmap->ineq[ineq][0],
951 info->bmap->ineq[ineq][0], v->el[0]);
952 r = isl_tab_add_ineq(info->tab, info->bmap->ineq[ineq]);
953 isl_int_add(info->bmap->ineq[ineq][0],
954 info->bmap->ineq[ineq][0], v->el[0]);
956 if (r < 0)
957 return isl_vec_free(v);
959 return v;
962 /* Tighten the (non-redundant) constraints on the facet represented
963 * by info->tab.
964 * In particular, on input, info->tab represents the result
965 * of relaxing the "n" inequality constraints of info->bmap in "relaxed"
966 * by one, i.e., replacing f_i >= 0 by f_i + 1 >= 0, and then
967 * replacing the one at index "l" by the corresponding equality,
968 * i.e., f_k + 1 = 0, with k = relaxed[l].
970 * Compute a variable compression from the equality constraint f_k + 1 = 0
971 * and use it to tighten the other constraints of info->bmap
972 * (that is, all constraints that have not been relaxed),
973 * updating info->tab (and leaving info->bmap untouched).
974 * The compression handles essentially two cases, one where a variable
975 * is assigned a fixed value and can therefore be eliminated, and one
976 * where one variable is a shifted multiple of some other variable and
977 * can therefore be replaced by that multiple.
978 * Gaussian elimination would also work for the first case, but for
979 * the second case, the effectiveness would depend on the order
980 * of the variables.
981 * After compression, some of the constraints may have coefficients
982 * with a common divisor. If this divisor does not divide the constant
983 * term, then the constraint can be tightened.
984 * The tightening is performed on the tableau info->tab by introducing
985 * extra (temporary) constraints.
987 * Only constraints that are possibly affected by the compression are
988 * considered. In particular, if the constraint only involves variables
989 * that are directly mapped to a distinct set of other variables, then
990 * no common divisor can be introduced and no tightening can occur.
992 * It is important to only consider the non-redundant constraints
993 * since the facet constraint has been relaxed prior to the call
994 * to this function, meaning that the constraints that were redundant
995 * prior to the relaxation may no longer be redundant.
996 * These constraints will be ignored in the fused result, so
997 * the fusion detection should not exploit them.
999 static isl_stat tighten_on_relaxed_facet(struct isl_coalesce_info *info,
1000 int n, int *relaxed, int l)
1002 isl_size total;
1003 isl_ctx *ctx;
1004 isl_vec *v = NULL;
1005 isl_mat *T;
1006 int i;
1007 int k;
1008 int *affected;
1010 k = relaxed[l];
1011 ctx = isl_basic_map_get_ctx(info->bmap);
1012 total = isl_basic_map_dim(info->bmap, isl_dim_all);
1013 if (total < 0)
1014 return isl_stat_error;
1015 isl_int_add_ui(info->bmap->ineq[k][0], info->bmap->ineq[k][0], 1);
1016 T = isl_mat_sub_alloc6(ctx, info->bmap->ineq, k, 1, 0, 1 + total);
1017 T = isl_mat_variable_compression(T, NULL);
1018 isl_int_sub_ui(info->bmap->ineq[k][0], info->bmap->ineq[k][0], 1);
1019 if (!T)
1020 return isl_stat_error;
1021 if (T->n_col == 0) {
1022 isl_mat_free(T);
1023 return isl_stat_ok;
1026 affected = isl_alloc_array(ctx, int, total);
1027 if (!affected)
1028 goto error;
1030 for (i = 0; i < total; ++i)
1031 affected[i] = not_unique_unit_row(T, 1 + i);
1033 for (i = 0; i < info->bmap->n_ineq; ++i) {
1034 isl_bool handle;
1035 if (any(relaxed, n, i))
1036 continue;
1037 if (info->ineq[i] == STATUS_REDUNDANT)
1038 continue;
1039 handle = is_affected(info->bmap, i, affected, total);
1040 if (handle < 0)
1041 goto error;
1042 if (!handle)
1043 continue;
1044 v = isl_vec_alloc(ctx, 1 + total);
1045 if (!v)
1046 goto error;
1047 isl_seq_cpy(v->el, info->bmap->ineq[i], 1 + total);
1048 v = isl_vec_mat_product(v, isl_mat_copy(T));
1049 v = try_tightening(info, i, v);
1050 isl_vec_free(v);
1051 if (!v)
1052 goto error;
1055 isl_mat_free(T);
1056 free(affected);
1057 return isl_stat_ok;
1058 error:
1059 isl_mat_free(T);
1060 free(affected);
1061 return isl_stat_error;
1064 /* Replace the basic maps "i" and "j" by an extension of "i"
1065 * along the "n" inequality constraints in "relax" by one.
1066 * The tableau info[i].tab has already been extended.
1067 * Extend info[i].bmap accordingly by relaxing all constraints in "relax"
1068 * by one.
1069 * Each integer division that does not have exactly the same
1070 * definition in "i" and "j" is marked unknown and the basic map
1071 * is scheduled to be simplified in an attempt to recover
1072 * the integer division definition.
1073 * Place the extension in the position that is the smallest of i and j.
1075 static enum isl_change extend(int i, int j, int n, int *relax,
1076 struct isl_coalesce_info *info)
1078 int l;
1079 isl_size total;
1081 info[i].bmap = isl_basic_map_cow(info[i].bmap);
1082 total = isl_basic_map_dim(info[i].bmap, isl_dim_all);
1083 if (total < 0)
1084 return isl_change_error;
1085 for (l = 0; l < info[i].bmap->n_div; ++l)
1086 if (!isl_seq_eq(info[i].bmap->div[l],
1087 info[j].bmap->div[l], 1 + 1 + total)) {
1088 isl_int_set_si(info[i].bmap->div[l][0], 0);
1089 info[i].simplify = 1;
1091 for (l = 0; l < n; ++l)
1092 isl_int_add_ui(info[i].bmap->ineq[relax[l]][0],
1093 info[i].bmap->ineq[relax[l]][0], 1);
1094 ISL_F_SET(info[i].bmap, ISL_BASIC_MAP_FINAL);
1095 drop(&info[j]);
1096 if (j < i)
1097 exchange(&info[i], &info[j]);
1098 return isl_change_fuse;
1101 /* Basic map "i" has "n" inequality constraints (collected in "relax")
1102 * that are such that they include basic map "j" if they are relaxed
1103 * by one. All the other inequalities are valid for "j".
1104 * Check if basic map "j" forms an extension of basic map "i".
1106 * In particular, relax the constraints in "relax", compute the corresponding
1107 * facets one by one and check whether each of these is included
1108 * in the other basic map.
1109 * Before testing for inclusion, the constraints on each facet
1110 * are tightened to increase the chance of an inclusion being detected.
1111 * (Adding the valid constraints of "j" to the tableau of "i", as is done
1112 * in is_adj_ineq_extension, may further increase those chances, but this
1113 * is not currently done.)
1114 * If each facet is included, we know that relaxing the constraints extends
1115 * the basic map with exactly the other basic map (we already know that this
1116 * other basic map is included in the extension, because all other
1117 * inequality constraints are valid of "j") and we can replace the
1118 * two basic maps by this extension.
1120 * If any of the relaxed constraints turn out to be redundant, then bail out.
1121 * isl_tab_select_facet refuses to handle such constraints. It may be
1122 * possible to handle them anyway by making a distinction between
1123 * redundant constraints with a corresponding facet that still intersects
1124 * the set (allowing isl_tab_select_facet to handle them) and
1125 * those where the facet does not intersect the set (which can be ignored
1126 * because the empty facet is trivially included in the other disjunct).
1127 * However, relaxed constraints that turn out to be redundant should
1128 * be fairly rare and no such instance has been reported where
1129 * coalescing would be successful.
1130 * ____ _____
1131 * / || / |
1132 * / || / |
1133 * \ || => \ |
1134 * \ || \ |
1135 * \___|| \____|
1138 * \ |\
1139 * |\\ | \
1140 * | \\ | \
1141 * | | => | /
1142 * | / | /
1143 * |/ |/
1145 static enum isl_change is_relaxed_extension(int i, int j, int n, int *relax,
1146 struct isl_coalesce_info *info)
1148 int l;
1149 isl_bool super;
1150 struct isl_tab_undo *snap, *snap2;
1151 unsigned n_eq = info[i].bmap->n_eq;
1153 for (l = 0; l < n; ++l)
1154 if (isl_tab_is_equality(info[i].tab, n_eq + relax[l]))
1155 return isl_change_none;
1157 snap = isl_tab_snap(info[i].tab);
1158 for (l = 0; l < n; ++l)
1159 if (isl_tab_relax(info[i].tab, n_eq + relax[l]) < 0)
1160 return isl_change_error;
1161 for (l = 0; l < n; ++l) {
1162 if (!isl_tab_is_redundant(info[i].tab, n_eq + relax[l]))
1163 continue;
1164 if (isl_tab_rollback(info[i].tab, snap) < 0)
1165 return isl_change_error;
1166 return isl_change_none;
1168 snap2 = isl_tab_snap(info[i].tab);
1169 for (l = 0; l < n; ++l) {
1170 if (isl_tab_rollback(info[i].tab, snap2) < 0)
1171 return isl_change_error;
1172 if (isl_tab_select_facet(info[i].tab, n_eq + relax[l]) < 0)
1173 return isl_change_error;
1174 if (tighten_on_relaxed_facet(&info[i], n, relax, l) < 0)
1175 return isl_change_error;
1176 super = contains(&info[j], info[i].tab);
1177 if (super < 0)
1178 return isl_change_error;
1179 if (super)
1180 continue;
1181 if (isl_tab_rollback(info[i].tab, snap) < 0)
1182 return isl_change_error;
1183 return isl_change_none;
1186 if (isl_tab_rollback(info[i].tab, snap2) < 0)
1187 return isl_change_error;
1188 return extend(i, j, n, relax, info);
1191 /* Data structure that keeps track of the wrapping constraints
1192 * and of information to bound the coefficients of those constraints.
1194 * bound is set if we want to apply a bound on the coefficients
1195 * mat contains the wrapping constraints
1196 * max is the bound on the coefficients (if bound is set)
1198 struct isl_wraps {
1199 int bound;
1200 isl_mat *mat;
1201 isl_int max;
1204 /* Update wraps->max to be greater than or equal to the coefficients
1205 * in the equalities and inequalities of info->bmap that can be removed
1206 * if we end up applying wrapping.
1208 static isl_stat wraps_update_max(struct isl_wraps *wraps,
1209 struct isl_coalesce_info *info)
1211 int k;
1212 isl_int max_k;
1213 isl_size total = isl_basic_map_dim(info->bmap, isl_dim_all);
1215 if (total < 0)
1216 return isl_stat_error;
1217 isl_int_init(max_k);
1219 for (k = 0; k < info->bmap->n_eq; ++k) {
1220 if (info->eq[2 * k] == STATUS_VALID &&
1221 info->eq[2 * k + 1] == STATUS_VALID)
1222 continue;
1223 isl_seq_abs_max(info->bmap->eq[k] + 1, total, &max_k);
1224 if (isl_int_abs_gt(max_k, wraps->max))
1225 isl_int_set(wraps->max, max_k);
1228 for (k = 0; k < info->bmap->n_ineq; ++k) {
1229 if (info->ineq[k] == STATUS_VALID ||
1230 info->ineq[k] == STATUS_REDUNDANT)
1231 continue;
1232 isl_seq_abs_max(info->bmap->ineq[k] + 1, total, &max_k);
1233 if (isl_int_abs_gt(max_k, wraps->max))
1234 isl_int_set(wraps->max, max_k);
1237 isl_int_clear(max_k);
1239 return isl_stat_ok;
1242 /* Initialize the isl_wraps data structure.
1243 * If we want to bound the coefficients of the wrapping constraints,
1244 * we set wraps->max to the largest coefficient
1245 * in the equalities and inequalities that can be removed if we end up
1246 * applying wrapping.
1248 static isl_stat wraps_init(struct isl_wraps *wraps, __isl_take isl_mat *mat,
1249 struct isl_coalesce_info *info, int i, int j)
1251 isl_ctx *ctx;
1253 wraps->bound = 0;
1254 wraps->mat = mat;
1255 if (!mat)
1256 return isl_stat_error;
1257 ctx = isl_mat_get_ctx(mat);
1258 wraps->bound = isl_options_get_coalesce_bounded_wrapping(ctx);
1259 if (!wraps->bound)
1260 return isl_stat_ok;
1261 isl_int_init(wraps->max);
1262 isl_int_set_si(wraps->max, 0);
1263 if (wraps_update_max(wraps, &info[i]) < 0)
1264 return isl_stat_error;
1265 if (wraps_update_max(wraps, &info[j]) < 0)
1266 return isl_stat_error;
1268 return isl_stat_ok;
1271 /* Free the contents of the isl_wraps data structure.
1273 static void wraps_free(struct isl_wraps *wraps)
1275 isl_mat_free(wraps->mat);
1276 if (wraps->bound)
1277 isl_int_clear(wraps->max);
1280 /* Is the wrapping constraint in row "row" allowed?
1282 * If wraps->bound is set, we check that none of the coefficients
1283 * is greater than wraps->max.
1285 static int allow_wrap(struct isl_wraps *wraps, int row)
1287 int i;
1289 if (!wraps->bound)
1290 return 1;
1292 for (i = 1; i < wraps->mat->n_col; ++i)
1293 if (isl_int_abs_gt(wraps->mat->row[row][i], wraps->max))
1294 return 0;
1296 return 1;
1299 /* Wrap "ineq" (or its opposite if "negate" is set) around "bound"
1300 * to include "set" and add the result in position "w" of "wraps".
1301 * "len" is the total number of coefficients in "bound" and "ineq".
1302 * Return 1 on success, 0 on failure and -1 on error.
1303 * Wrapping can fail if the result of wrapping is equal to "bound"
1304 * or if we want to bound the sizes of the coefficients and
1305 * the wrapped constraint does not satisfy this bound.
1307 static int add_wrap(struct isl_wraps *wraps, int w, isl_int *bound,
1308 isl_int *ineq, unsigned len, __isl_keep isl_set *set, int negate)
1310 isl_seq_cpy(wraps->mat->row[w], bound, len);
1311 if (negate) {
1312 isl_seq_neg(wraps->mat->row[w + 1], ineq, len);
1313 ineq = wraps->mat->row[w + 1];
1315 if (!isl_set_wrap_facet(set, wraps->mat->row[w], ineq))
1316 return -1;
1317 if (isl_seq_eq(wraps->mat->row[w], bound, len))
1318 return 0;
1319 if (!allow_wrap(wraps, w))
1320 return 0;
1321 return 1;
1324 /* For each constraint in info->bmap that is not redundant (as determined
1325 * by info->tab) and that is not a valid constraint for the other basic map,
1326 * wrap the constraint around "bound" such that it includes the whole
1327 * set "set" and append the resulting constraint to "wraps".
1328 * Note that the constraints that are valid for the other basic map
1329 * will be added to the combined basic map by default, so there is
1330 * no need to wrap them.
1331 * The caller wrap_in_facets even relies on this function not wrapping
1332 * any constraints that are already valid.
1333 * "wraps" is assumed to have been pre-allocated to the appropriate size.
1334 * wraps->n_row is the number of actual wrapped constraints that have
1335 * been added.
1336 * If any of the wrapping problems results in a constraint that is
1337 * identical to "bound", then this means that "set" is unbounded in such
1338 * way that no wrapping is possible. If this happens then wraps->n_row
1339 * is reset to zero.
1340 * Similarly, if we want to bound the coefficients of the wrapping
1341 * constraints and a newly added wrapping constraint does not
1342 * satisfy the bound, then wraps->n_row is also reset to zero.
1344 static isl_stat add_wraps(struct isl_wraps *wraps,
1345 struct isl_coalesce_info *info, isl_int *bound, __isl_keep isl_set *set)
1347 int l, m;
1348 int w;
1349 int added;
1350 isl_basic_map *bmap = info->bmap;
1351 isl_size total = isl_basic_map_dim(bmap, isl_dim_all);
1352 unsigned len = 1 + total;
1354 if (total < 0)
1355 return isl_stat_error;
1357 w = wraps->mat->n_row;
1359 for (l = 0; l < bmap->n_ineq; ++l) {
1360 if (info->ineq[l] == STATUS_VALID ||
1361 info->ineq[l] == STATUS_REDUNDANT)
1362 continue;
1363 if (isl_seq_is_neg(bound, bmap->ineq[l], len))
1364 continue;
1365 if (isl_seq_eq(bound, bmap->ineq[l], len))
1366 continue;
1367 if (isl_tab_is_redundant(info->tab, bmap->n_eq + l))
1368 continue;
1370 added = add_wrap(wraps, w, bound, bmap->ineq[l], len, set, 0);
1371 if (added < 0)
1372 return isl_stat_error;
1373 if (!added)
1374 goto unbounded;
1375 ++w;
1377 for (l = 0; l < bmap->n_eq; ++l) {
1378 if (isl_seq_is_neg(bound, bmap->eq[l], len))
1379 continue;
1380 if (isl_seq_eq(bound, bmap->eq[l], len))
1381 continue;
1383 for (m = 0; m < 2; ++m) {
1384 if (info->eq[2 * l + m] == STATUS_VALID)
1385 continue;
1386 added = add_wrap(wraps, w, bound, bmap->eq[l], len,
1387 set, !m);
1388 if (added < 0)
1389 return isl_stat_error;
1390 if (!added)
1391 goto unbounded;
1392 ++w;
1396 wraps->mat->n_row = w;
1397 return isl_stat_ok;
1398 unbounded:
1399 wraps->mat->n_row = 0;
1400 return isl_stat_ok;
1403 /* Check if the constraints in "wraps" from "first" until the last
1404 * are all valid for the basic set represented by "tab".
1405 * If not, wraps->n_row is set to zero.
1407 static int check_wraps(__isl_keep isl_mat *wraps, int first,
1408 struct isl_tab *tab)
1410 int i;
1412 for (i = first; i < wraps->n_row; ++i) {
1413 enum isl_ineq_type type;
1414 type = isl_tab_ineq_type(tab, wraps->row[i]);
1415 if (type == isl_ineq_error)
1416 return -1;
1417 if (type == isl_ineq_redundant)
1418 continue;
1419 wraps->n_row = 0;
1420 return 0;
1423 return 0;
1426 /* Return a set that corresponds to the non-redundant constraints
1427 * (as recorded in tab) of bmap.
1429 * It's important to remove the redundant constraints as some
1430 * of the other constraints may have been modified after the
1431 * constraints were marked redundant.
1432 * In particular, a constraint may have been relaxed.
1433 * Redundant constraints are ignored when a constraint is relaxed
1434 * and should therefore continue to be ignored ever after.
1435 * Otherwise, the relaxation might be thwarted by some of
1436 * these constraints.
1438 * Update the underlying set to ensure that the dimension doesn't change.
1439 * Otherwise the integer divisions could get dropped if the tab
1440 * turns out to be empty.
1442 static __isl_give isl_set *set_from_updated_bmap(__isl_keep isl_basic_map *bmap,
1443 struct isl_tab *tab)
1445 isl_basic_set *bset;
1447 bmap = isl_basic_map_copy(bmap);
1448 bset = isl_basic_map_underlying_set(bmap);
1449 bset = isl_basic_set_cow(bset);
1450 bset = isl_basic_set_update_from_tab(bset, tab);
1451 return isl_set_from_basic_set(bset);
1454 /* Wrap the constraints of info->bmap that bound the facet defined
1455 * by inequality "k" around (the opposite of) this inequality to
1456 * include "set". "bound" may be used to store the negated inequality.
1457 * Since the wrapped constraints are not guaranteed to contain the whole
1458 * of info->bmap, we check them in check_wraps.
1459 * If any of the wrapped constraints turn out to be invalid, then
1460 * check_wraps will reset wrap->n_row to zero.
1462 static isl_stat add_wraps_around_facet(struct isl_wraps *wraps,
1463 struct isl_coalesce_info *info, int k, isl_int *bound,
1464 __isl_keep isl_set *set)
1466 struct isl_tab_undo *snap;
1467 int n;
1468 isl_size total = isl_basic_map_dim(info->bmap, isl_dim_all);
1470 if (total < 0)
1471 return isl_stat_error;
1473 snap = isl_tab_snap(info->tab);
1475 if (isl_tab_select_facet(info->tab, info->bmap->n_eq + k) < 0)
1476 return isl_stat_error;
1477 if (isl_tab_detect_redundant(info->tab) < 0)
1478 return isl_stat_error;
1480 isl_seq_neg(bound, info->bmap->ineq[k], 1 + total);
1482 n = wraps->mat->n_row;
1483 if (add_wraps(wraps, info, bound, set) < 0)
1484 return isl_stat_error;
1486 if (isl_tab_rollback(info->tab, snap) < 0)
1487 return isl_stat_error;
1488 if (check_wraps(wraps->mat, n, info->tab) < 0)
1489 return isl_stat_error;
1491 return isl_stat_ok;
1494 /* Given a basic set i with a constraint k that is adjacent to
1495 * basic set j, check if we can wrap
1496 * both the facet corresponding to k (if "wrap_facet" is set) and basic map j
1497 * (always) around their ridges to include the other set.
1498 * If so, replace the pair of basic sets by their union.
1500 * All constraints of i (except k) are assumed to be valid or
1501 * cut constraints for j.
1502 * Wrapping the cut constraints to include basic map j may result
1503 * in constraints that are no longer valid of basic map i
1504 * we have to check that the resulting wrapping constraints are valid for i.
1505 * If "wrap_facet" is not set, then all constraints of i (except k)
1506 * are assumed to be valid for j.
1507 * ____ _____
1508 * / | / \
1509 * / || / |
1510 * \ || => \ |
1511 * \ || \ |
1512 * \___|| \____|
1515 static enum isl_change can_wrap_in_facet(int i, int j, int k,
1516 struct isl_coalesce_info *info, int wrap_facet)
1518 enum isl_change change = isl_change_none;
1519 struct isl_wraps wraps;
1520 isl_ctx *ctx;
1521 isl_mat *mat;
1522 struct isl_set *set_i = NULL;
1523 struct isl_set *set_j = NULL;
1524 struct isl_vec *bound = NULL;
1525 isl_size total = isl_basic_map_dim(info[i].bmap, isl_dim_all);
1527 if (total < 0)
1528 return isl_change_error;
1529 set_i = set_from_updated_bmap(info[i].bmap, info[i].tab);
1530 set_j = set_from_updated_bmap(info[j].bmap, info[j].tab);
1531 ctx = isl_basic_map_get_ctx(info[i].bmap);
1532 mat = isl_mat_alloc(ctx, 2 * (info[i].bmap->n_eq + info[j].bmap->n_eq) +
1533 info[i].bmap->n_ineq + info[j].bmap->n_ineq,
1534 1 + total);
1535 if (wraps_init(&wraps, mat, info, i, j) < 0)
1536 goto error;
1537 bound = isl_vec_alloc(ctx, 1 + total);
1538 if (!set_i || !set_j || !bound)
1539 goto error;
1541 isl_seq_cpy(bound->el, info[i].bmap->ineq[k], 1 + total);
1542 isl_int_add_ui(bound->el[0], bound->el[0], 1);
1543 isl_seq_normalize(ctx, bound->el, 1 + total);
1545 isl_seq_cpy(wraps.mat->row[0], bound->el, 1 + total);
1546 wraps.mat->n_row = 1;
1548 if (add_wraps(&wraps, &info[j], bound->el, set_i) < 0)
1549 goto error;
1550 if (!wraps.mat->n_row)
1551 goto unbounded;
1553 if (wrap_facet) {
1554 if (add_wraps_around_facet(&wraps, &info[i], k,
1555 bound->el, set_j) < 0)
1556 goto error;
1557 if (!wraps.mat->n_row)
1558 goto unbounded;
1561 change = fuse(i, j, info, wraps.mat, 0, 0);
1563 unbounded:
1564 wraps_free(&wraps);
1566 isl_set_free(set_i);
1567 isl_set_free(set_j);
1569 isl_vec_free(bound);
1571 return change;
1572 error:
1573 wraps_free(&wraps);
1574 isl_vec_free(bound);
1575 isl_set_free(set_i);
1576 isl_set_free(set_j);
1577 return isl_change_error;
1580 /* Given a cut constraint t(x) >= 0 of basic map i, stored in row "w"
1581 * of wrap.mat, replace it by its relaxed version t(x) + 1 >= 0, and
1582 * add wrapping constraints to wrap.mat for all constraints
1583 * of basic map j that bound the part of basic map j that sticks out
1584 * of the cut constraint.
1585 * "set_i" is the underlying set of basic map i.
1586 * If any wrapping fails, then wraps->mat.n_row is reset to zero.
1588 * In particular, we first intersect basic map j with t(x) + 1 = 0.
1589 * If the result is empty, then t(x) >= 0 was actually a valid constraint
1590 * (with respect to the integer points), so we add t(x) >= 0 instead.
1591 * Otherwise, we wrap the constraints of basic map j that are not
1592 * redundant in this intersection and that are not already valid
1593 * for basic map i over basic map i.
1594 * Note that it is sufficient to wrap the constraints to include
1595 * basic map i, because we will only wrap the constraints that do
1596 * not include basic map i already. The wrapped constraint will
1597 * therefore be more relaxed compared to the original constraint.
1598 * Since the original constraint is valid for basic map j, so is
1599 * the wrapped constraint.
1601 static isl_stat wrap_in_facet(struct isl_wraps *wraps, int w,
1602 struct isl_coalesce_info *info_j, __isl_keep isl_set *set_i,
1603 struct isl_tab_undo *snap)
1605 isl_int_add_ui(wraps->mat->row[w][0], wraps->mat->row[w][0], 1);
1606 if (isl_tab_add_eq(info_j->tab, wraps->mat->row[w]) < 0)
1607 return isl_stat_error;
1608 if (isl_tab_detect_redundant(info_j->tab) < 0)
1609 return isl_stat_error;
1611 if (info_j->tab->empty)
1612 isl_int_sub_ui(wraps->mat->row[w][0], wraps->mat->row[w][0], 1);
1613 else if (add_wraps(wraps, info_j, wraps->mat->row[w], set_i) < 0)
1614 return isl_stat_error;
1616 if (isl_tab_rollback(info_j->tab, snap) < 0)
1617 return isl_stat_error;
1619 return isl_stat_ok;
1622 /* Given a pair of basic maps i and j such that j sticks out
1623 * of i at n cut constraints, each time by at most one,
1624 * try to compute wrapping constraints and replace the two
1625 * basic maps by a single basic map.
1626 * The other constraints of i are assumed to be valid for j.
1627 * "set_i" is the underlying set of basic map i.
1628 * "wraps" has been initialized to be of the right size.
1630 * For each cut constraint t(x) >= 0 of i, we add the relaxed version
1631 * t(x) + 1 >= 0, along with wrapping constraints for all constraints
1632 * of basic map j that bound the part of basic map j that sticks out
1633 * of the cut constraint.
1635 * If any wrapping fails, i.e., if we cannot wrap to touch
1636 * the union, then we give up.
1637 * Otherwise, the pair of basic maps is replaced by their union.
1639 static enum isl_change try_wrap_in_facets(int i, int j,
1640 struct isl_coalesce_info *info, struct isl_wraps *wraps,
1641 __isl_keep isl_set *set_i)
1643 int k, l, w;
1644 isl_size total;
1645 struct isl_tab_undo *snap;
1647 total = isl_basic_map_dim(info[i].bmap, isl_dim_all);
1648 if (total < 0)
1649 return isl_change_error;
1651 snap = isl_tab_snap(info[j].tab);
1653 wraps->mat->n_row = 0;
1655 for (k = 0; k < info[i].bmap->n_eq; ++k) {
1656 for (l = 0; l < 2; ++l) {
1657 if (info[i].eq[2 * k + l] != STATUS_CUT)
1658 continue;
1659 w = wraps->mat->n_row++;
1660 if (l == 0)
1661 isl_seq_neg(wraps->mat->row[w],
1662 info[i].bmap->eq[k], 1 + total);
1663 else
1664 isl_seq_cpy(wraps->mat->row[w],
1665 info[i].bmap->eq[k], 1 + total);
1666 if (wrap_in_facet(wraps, w, &info[j], set_i, snap) < 0)
1667 return isl_change_error;
1669 if (!wraps->mat->n_row)
1670 return isl_change_none;
1674 for (k = 0; k < info[i].bmap->n_ineq; ++k) {
1675 if (info[i].ineq[k] != STATUS_CUT)
1676 continue;
1677 w = wraps->mat->n_row++;
1678 isl_seq_cpy(wraps->mat->row[w],
1679 info[i].bmap->ineq[k], 1 + total);
1680 if (wrap_in_facet(wraps, w, &info[j], set_i, snap) < 0)
1681 return isl_change_error;
1683 if (!wraps->mat->n_row)
1684 return isl_change_none;
1687 return fuse(i, j, info, wraps->mat, 0, 1);
1690 /* Given a pair of basic maps i and j such that j sticks out
1691 * of i at n cut constraints, each time by at most one,
1692 * try to compute wrapping constraints and replace the two
1693 * basic maps by a single basic map.
1694 * The other constraints of i are assumed to be valid for j.
1696 * The core computation is performed by try_wrap_in_facets.
1697 * This function simply extracts an underlying set representation
1698 * of basic map i and initializes the data structure for keeping
1699 * track of wrapping constraints.
1701 static enum isl_change wrap_in_facets(int i, int j, int n,
1702 struct isl_coalesce_info *info)
1704 enum isl_change change = isl_change_none;
1705 struct isl_wraps wraps;
1706 isl_ctx *ctx;
1707 isl_mat *mat;
1708 isl_set *set_i = NULL;
1709 isl_size total = isl_basic_map_dim(info[i].bmap, isl_dim_all);
1710 int max_wrap;
1712 if (total < 0)
1713 return isl_change_error;
1714 if (isl_tab_extend_cons(info[j].tab, 1) < 0)
1715 return isl_change_error;
1717 max_wrap = 1 + 2 * info[j].bmap->n_eq + info[j].bmap->n_ineq;
1718 max_wrap *= n;
1720 set_i = set_from_updated_bmap(info[i].bmap, info[i].tab);
1721 ctx = isl_basic_map_get_ctx(info[i].bmap);
1722 mat = isl_mat_alloc(ctx, max_wrap, 1 + total);
1723 if (wraps_init(&wraps, mat, info, i, j) < 0)
1724 goto error;
1725 if (!set_i)
1726 goto error;
1728 change = try_wrap_in_facets(i, j, info, &wraps, set_i);
1730 wraps_free(&wraps);
1731 isl_set_free(set_i);
1733 return change;
1734 error:
1735 wraps_free(&wraps);
1736 isl_set_free(set_i);
1737 return isl_change_error;
1740 /* Return the effect of inequality "ineq" on the tableau "tab",
1741 * after relaxing the constant term of "ineq" by one.
1743 static enum isl_ineq_type type_of_relaxed(struct isl_tab *tab, isl_int *ineq)
1745 enum isl_ineq_type type;
1747 isl_int_add_ui(ineq[0], ineq[0], 1);
1748 type = isl_tab_ineq_type(tab, ineq);
1749 isl_int_sub_ui(ineq[0], ineq[0], 1);
1751 return type;
1754 /* Given two basic sets i and j,
1755 * check if relaxing all the cut constraints of i by one turns
1756 * them into valid constraint for j and check if we can wrap in
1757 * the bits that are sticking out.
1758 * If so, replace the pair by their union.
1760 * We first check if all relaxed cut inequalities of i are valid for j
1761 * and then try to wrap in the intersections of the relaxed cut inequalities
1762 * with j.
1764 * During this wrapping, we consider the points of j that lie at a distance
1765 * of exactly 1 from i. In particular, we ignore the points that lie in
1766 * between this lower-dimensional space and the basic map i.
1767 * We can therefore only apply this to integer maps.
1768 * ____ _____
1769 * / ___|_ / \
1770 * / | | / |
1771 * \ | | => \ |
1772 * \|____| \ |
1773 * \___| \____/
1775 * _____ ______
1776 * | ____|_ | \
1777 * | | | | |
1778 * | | | => | |
1779 * |_| | | |
1780 * |_____| \______|
1782 * _______
1783 * | |
1784 * | |\ |
1785 * | | \ |
1786 * | | \ |
1787 * | | \|
1788 * | | \
1789 * | |_____\
1790 * | |
1791 * |_______|
1793 * Wrapping can fail if the result of wrapping one of the facets
1794 * around its edges does not produce any new facet constraint.
1795 * In particular, this happens when we try to wrap in unbounded sets.
1797 * _______________________________________________________________________
1799 * | ___
1800 * | | |
1801 * |_| |_________________________________________________________________
1802 * |___|
1804 * The following is not an acceptable result of coalescing the above two
1805 * sets as it includes extra integer points.
1806 * _______________________________________________________________________
1808 * |
1809 * |
1811 * \______________________________________________________________________
1813 static enum isl_change can_wrap_in_set(int i, int j,
1814 struct isl_coalesce_info *info)
1816 int k, l;
1817 int n;
1818 isl_size total;
1820 if (ISL_F_ISSET(info[i].bmap, ISL_BASIC_MAP_RATIONAL) ||
1821 ISL_F_ISSET(info[j].bmap, ISL_BASIC_MAP_RATIONAL))
1822 return isl_change_none;
1824 n = count_eq(&info[i], STATUS_CUT) + count_ineq(&info[i], STATUS_CUT);
1825 if (n == 0)
1826 return isl_change_none;
1828 total = isl_basic_map_dim(info[i].bmap, isl_dim_all);
1829 if (total < 0)
1830 return isl_change_error;
1831 for (k = 0; k < info[i].bmap->n_eq; ++k) {
1832 for (l = 0; l < 2; ++l) {
1833 enum isl_ineq_type type;
1835 if (info[i].eq[2 * k + l] != STATUS_CUT)
1836 continue;
1838 if (l == 0)
1839 isl_seq_neg(info[i].bmap->eq[k],
1840 info[i].bmap->eq[k], 1 + total);
1841 type = type_of_relaxed(info[j].tab,
1842 info[i].bmap->eq[k]);
1843 if (l == 0)
1844 isl_seq_neg(info[i].bmap->eq[k],
1845 info[i].bmap->eq[k], 1 + total);
1846 if (type == isl_ineq_error)
1847 return isl_change_error;
1848 if (type != isl_ineq_redundant)
1849 return isl_change_none;
1853 for (k = 0; k < info[i].bmap->n_ineq; ++k) {
1854 enum isl_ineq_type type;
1856 if (info[i].ineq[k] != STATUS_CUT)
1857 continue;
1859 type = type_of_relaxed(info[j].tab, info[i].bmap->ineq[k]);
1860 if (type == isl_ineq_error)
1861 return isl_change_error;
1862 if (type != isl_ineq_redundant)
1863 return isl_change_none;
1866 return wrap_in_facets(i, j, n, info);
1869 /* Check if either i or j has only cut constraints that can
1870 * be used to wrap in (a facet of) the other basic set.
1871 * if so, replace the pair by their union.
1873 static enum isl_change check_wrap(int i, int j, struct isl_coalesce_info *info)
1875 enum isl_change change = isl_change_none;
1877 change = can_wrap_in_set(i, j, info);
1878 if (change != isl_change_none)
1879 return change;
1881 change = can_wrap_in_set(j, i, info);
1882 return change;
1885 /* Check if all inequality constraints of "i" that cut "j" cease
1886 * to be cut constraints if they are relaxed by one.
1887 * If so, collect the cut constraints in "list".
1888 * The caller is responsible for allocating "list".
1890 static isl_bool all_cut_by_one(int i, int j, struct isl_coalesce_info *info,
1891 int *list)
1893 int l, n;
1895 n = 0;
1896 for (l = 0; l < info[i].bmap->n_ineq; ++l) {
1897 enum isl_ineq_type type;
1899 if (info[i].ineq[l] != STATUS_CUT)
1900 continue;
1901 type = type_of_relaxed(info[j].tab, info[i].bmap->ineq[l]);
1902 if (type == isl_ineq_error)
1903 return isl_bool_error;
1904 if (type != isl_ineq_redundant)
1905 return isl_bool_false;
1906 list[n++] = l;
1909 return isl_bool_true;
1912 /* Given two basic maps such that "j" has at least one equality constraint
1913 * that is adjacent to an inequality constraint of "i" and such that "i" has
1914 * exactly one inequality constraint that is adjacent to an equality
1915 * constraint of "j", check whether "i" can be extended to include "j" or
1916 * whether "j" can be wrapped into "i".
1917 * All remaining constraints of "i" and "j" are assumed to be valid
1918 * or cut constraints of the other basic map.
1919 * However, none of the equality constraints of "i" are cut constraints.
1921 * If "i" has any "cut" inequality constraints, then check if relaxing
1922 * each of them by one is sufficient for them to become valid.
1923 * If so, check if the inequality constraint adjacent to an equality
1924 * constraint of "j" along with all these cut constraints
1925 * can be relaxed by one to contain exactly "j".
1926 * Otherwise, or if this fails, check if "j" can be wrapped into "i".
1928 static enum isl_change check_single_adj_eq(int i, int j,
1929 struct isl_coalesce_info *info)
1931 enum isl_change change = isl_change_none;
1932 int k;
1933 int n_cut;
1934 int *relax;
1935 isl_ctx *ctx;
1936 isl_bool try_relax;
1938 n_cut = count_ineq(&info[i], STATUS_CUT);
1940 k = find_ineq(&info[i], STATUS_ADJ_EQ);
1942 if (n_cut > 0) {
1943 ctx = isl_basic_map_get_ctx(info[i].bmap);
1944 relax = isl_calloc_array(ctx, int, 1 + n_cut);
1945 if (!relax)
1946 return isl_change_error;
1947 relax[0] = k;
1948 try_relax = all_cut_by_one(i, j, info, relax + 1);
1949 if (try_relax < 0)
1950 change = isl_change_error;
1951 } else {
1952 try_relax = isl_bool_true;
1953 relax = &k;
1955 if (try_relax && change == isl_change_none)
1956 change = is_relaxed_extension(i, j, 1 + n_cut, relax, info);
1957 if (n_cut > 0)
1958 free(relax);
1959 if (change != isl_change_none)
1960 return change;
1962 change = can_wrap_in_facet(i, j, k, info, n_cut > 0);
1964 return change;
1967 /* At least one of the basic maps has an equality that is adjacent
1968 * to an inequality. Make sure that only one of the basic maps has
1969 * such an equality and that the other basic map has exactly one
1970 * inequality adjacent to an equality.
1971 * If the other basic map does not have such an inequality, then
1972 * check if all its constraints are either valid or cut constraints
1973 * and, if so, try wrapping in the first map into the second.
1974 * Otherwise, try to extend one basic map with the other or
1975 * wrap one basic map in the other.
1977 static enum isl_change check_adj_eq(int i, int j,
1978 struct isl_coalesce_info *info)
1980 if (any_eq(&info[i], STATUS_ADJ_INEQ) &&
1981 any_eq(&info[j], STATUS_ADJ_INEQ))
1982 /* ADJ EQ TOO MANY */
1983 return isl_change_none;
1985 if (any_eq(&info[i], STATUS_ADJ_INEQ))
1986 return check_adj_eq(j, i, info);
1988 /* j has an equality adjacent to an inequality in i */
1990 if (count_ineq(&info[i], STATUS_ADJ_EQ) != 1) {
1991 if (all_valid_or_cut(&info[i]))
1992 return can_wrap_in_set(i, j, info);
1993 return isl_change_none;
1995 if (any_eq(&info[i], STATUS_CUT))
1996 return isl_change_none;
1997 if (any_ineq(&info[j], STATUS_ADJ_EQ) ||
1998 any_ineq(&info[i], STATUS_ADJ_INEQ) ||
1999 any_ineq(&info[j], STATUS_ADJ_INEQ))
2000 /* ADJ EQ TOO MANY */
2001 return isl_change_none;
2003 return check_single_adj_eq(i, j, info);
2006 /* Disjunct "j" lies on a hyperplane that is adjacent to disjunct "i".
2007 * In particular, disjunct "i" has an inequality constraint that is adjacent
2008 * to a (combination of) equality constraint(s) of disjunct "j",
2009 * but disjunct "j" has no explicit equality constraint adjacent
2010 * to an inequality constraint of disjunct "i".
2012 * Disjunct "i" is already known not to have any equality constraints
2013 * that are adjacent to an equality or inequality constraint.
2014 * Check that, other than the inequality constraint mentioned above,
2015 * all other constraints of disjunct "i" are valid for disjunct "j".
2016 * If so, try and wrap in disjunct "j".
2018 static enum isl_change check_ineq_adj_eq(int i, int j,
2019 struct isl_coalesce_info *info)
2021 int k;
2023 if (any_eq(&info[i], STATUS_CUT))
2024 return isl_change_none;
2025 if (any_ineq(&info[i], STATUS_CUT))
2026 return isl_change_none;
2027 if (any_ineq(&info[i], STATUS_ADJ_INEQ))
2028 return isl_change_none;
2029 if (count_ineq(&info[i], STATUS_ADJ_EQ) != 1)
2030 return isl_change_none;
2032 k = find_ineq(&info[i], STATUS_ADJ_EQ);
2034 return can_wrap_in_facet(i, j, k, info, 0);
2037 /* The two basic maps lie on adjacent hyperplanes. In particular,
2038 * basic map "i" has an equality that lies parallel to basic map "j".
2039 * Check if we can wrap the facets around the parallel hyperplanes
2040 * to include the other set.
2042 * We perform basically the same operations as can_wrap_in_facet,
2043 * except that we don't need to select a facet of one of the sets.
2045 * \\ \\
2046 * \\ => \\
2047 * \ \|
2049 * If there is more than one equality of "i" adjacent to an equality of "j",
2050 * then the result will satisfy one or more equalities that are a linear
2051 * combination of these equalities. These will be encoded as pairs
2052 * of inequalities in the wrapping constraints and need to be made
2053 * explicit.
2055 static enum isl_change check_eq_adj_eq(int i, int j,
2056 struct isl_coalesce_info *info)
2058 int k;
2059 enum isl_change change = isl_change_none;
2060 int detect_equalities = 0;
2061 struct isl_wraps wraps;
2062 isl_ctx *ctx;
2063 isl_mat *mat;
2064 struct isl_set *set_i = NULL;
2065 struct isl_set *set_j = NULL;
2066 struct isl_vec *bound = NULL;
2067 isl_size total = isl_basic_map_dim(info[i].bmap, isl_dim_all);
2069 if (total < 0)
2070 return isl_change_error;
2071 if (count_eq(&info[i], STATUS_ADJ_EQ) != 1)
2072 detect_equalities = 1;
2074 k = find_eq(&info[i], STATUS_ADJ_EQ);
2076 set_i = set_from_updated_bmap(info[i].bmap, info[i].tab);
2077 set_j = set_from_updated_bmap(info[j].bmap, info[j].tab);
2078 ctx = isl_basic_map_get_ctx(info[i].bmap);
2079 mat = isl_mat_alloc(ctx, 2 * (info[i].bmap->n_eq + info[j].bmap->n_eq) +
2080 info[i].bmap->n_ineq + info[j].bmap->n_ineq,
2081 1 + total);
2082 if (wraps_init(&wraps, mat, info, i, j) < 0)
2083 goto error;
2084 bound = isl_vec_alloc(ctx, 1 + total);
2085 if (!set_i || !set_j || !bound)
2086 goto error;
2088 if (k % 2 == 0)
2089 isl_seq_neg(bound->el, info[i].bmap->eq[k / 2], 1 + total);
2090 else
2091 isl_seq_cpy(bound->el, info[i].bmap->eq[k / 2], 1 + total);
2092 isl_int_add_ui(bound->el[0], bound->el[0], 1);
2094 isl_seq_cpy(wraps.mat->row[0], bound->el, 1 + total);
2095 wraps.mat->n_row = 1;
2097 if (add_wraps(&wraps, &info[j], bound->el, set_i) < 0)
2098 goto error;
2099 if (!wraps.mat->n_row)
2100 goto unbounded;
2102 isl_int_sub_ui(bound->el[0], bound->el[0], 1);
2103 isl_seq_neg(bound->el, bound->el, 1 + total);
2105 isl_seq_cpy(wraps.mat->row[wraps.mat->n_row], bound->el, 1 + total);
2106 wraps.mat->n_row++;
2108 if (add_wraps(&wraps, &info[i], bound->el, set_j) < 0)
2109 goto error;
2110 if (!wraps.mat->n_row)
2111 goto unbounded;
2113 change = fuse(i, j, info, wraps.mat, detect_equalities, 0);
2115 if (0) {
2116 error: change = isl_change_error;
2118 unbounded:
2120 wraps_free(&wraps);
2121 isl_set_free(set_i);
2122 isl_set_free(set_j);
2123 isl_vec_free(bound);
2125 return change;
2128 /* Initialize the "eq" and "ineq" fields of "info".
2130 static void init_status(struct isl_coalesce_info *info)
2132 info->eq = info->ineq = NULL;
2135 /* Set info->eq to the positions of the equalities of info->bmap
2136 * with respect to the basic map represented by "tab".
2137 * If info->eq has already been computed, then do not compute it again.
2139 static void set_eq_status_in(struct isl_coalesce_info *info,
2140 struct isl_tab *tab)
2142 if (info->eq)
2143 return;
2144 info->eq = eq_status_in(info->bmap, tab);
2147 /* Set info->ineq to the positions of the inequalities of info->bmap
2148 * with respect to the basic map represented by "tab".
2149 * If info->ineq has already been computed, then do not compute it again.
2151 static void set_ineq_status_in(struct isl_coalesce_info *info,
2152 struct isl_tab *tab)
2154 if (info->ineq)
2155 return;
2156 info->ineq = ineq_status_in(info->bmap, info->tab, tab);
2159 /* Free the memory allocated by the "eq" and "ineq" fields of "info".
2160 * This function assumes that init_status has been called on "info" first,
2161 * after which the "eq" and "ineq" fields may or may not have been
2162 * assigned a newly allocated array.
2164 static void clear_status(struct isl_coalesce_info *info)
2166 free(info->eq);
2167 free(info->ineq);
2170 /* Are all inequality constraints of the basic map represented by "info"
2171 * valid for the other basic map, except for a single constraint
2172 * that is adjacent to an inequality constraint of the other basic map?
2174 static int all_ineq_valid_or_single_adj_ineq(struct isl_coalesce_info *info)
2176 int i;
2177 int k = -1;
2179 for (i = 0; i < info->bmap->n_ineq; ++i) {
2180 if (info->ineq[i] == STATUS_REDUNDANT)
2181 continue;
2182 if (info->ineq[i] == STATUS_VALID)
2183 continue;
2184 if (info->ineq[i] != STATUS_ADJ_INEQ)
2185 return 0;
2186 if (k != -1)
2187 return 0;
2188 k = i;
2191 return k != -1;
2194 /* Basic map "i" has one or more equality constraints that separate it
2195 * from basic map "j". Check if it happens to be an extension
2196 * of basic map "j".
2197 * In particular, check that all constraints of "j" are valid for "i",
2198 * except for one inequality constraint that is adjacent
2199 * to an inequality constraints of "i".
2200 * If so, check for "i" being an extension of "j" by calling
2201 * is_adj_ineq_extension.
2203 * Clean up the memory allocated for keeping track of the status
2204 * of the constraints before returning.
2206 static enum isl_change separating_equality(int i, int j,
2207 struct isl_coalesce_info *info)
2209 enum isl_change change = isl_change_none;
2211 if (all(info[j].eq, 2 * info[j].bmap->n_eq, STATUS_VALID) &&
2212 all_ineq_valid_or_single_adj_ineq(&info[j]))
2213 change = is_adj_ineq_extension(j, i, info);
2215 clear_status(&info[i]);
2216 clear_status(&info[j]);
2217 return change;
2220 /* Check if the union of the given pair of basic maps
2221 * can be represented by a single basic map.
2222 * If so, replace the pair by the single basic map and return
2223 * isl_change_drop_first, isl_change_drop_second or isl_change_fuse.
2224 * Otherwise, return isl_change_none.
2225 * The two basic maps are assumed to live in the same local space.
2226 * The "eq" and "ineq" fields of info[i] and info[j] are assumed
2227 * to have been initialized by the caller, either to NULL or
2228 * to valid information.
2230 * We first check the effect of each constraint of one basic map
2231 * on the other basic map.
2232 * The constraint may be
2233 * redundant the constraint is redundant in its own
2234 * basic map and should be ignore and removed
2235 * in the end
2236 * valid all (integer) points of the other basic map
2237 * satisfy the constraint
2238 * separate no (integer) point of the other basic map
2239 * satisfies the constraint
2240 * cut some but not all points of the other basic map
2241 * satisfy the constraint
2242 * adj_eq the given constraint is adjacent (on the outside)
2243 * to an equality of the other basic map
2244 * adj_ineq the given constraint is adjacent (on the outside)
2245 * to an inequality of the other basic map
2247 * We consider seven cases in which we can replace the pair by a single
2248 * basic map. We ignore all "redundant" constraints.
2250 * 1. all constraints of one basic map are valid
2251 * => the other basic map is a subset and can be removed
2253 * 2. all constraints of both basic maps are either "valid" or "cut"
2254 * and the facets corresponding to the "cut" constraints
2255 * of one of the basic maps lies entirely inside the other basic map
2256 * => the pair can be replaced by a basic map consisting
2257 * of the valid constraints in both basic maps
2259 * 3. there is a single pair of adjacent inequalities
2260 * (all other constraints are "valid")
2261 * => the pair can be replaced by a basic map consisting
2262 * of the valid constraints in both basic maps
2264 * 4. one basic map has a single adjacent inequality, while the other
2265 * constraints are "valid". The other basic map has some
2266 * "cut" constraints, but replacing the adjacent inequality by
2267 * its opposite and adding the valid constraints of the other
2268 * basic map results in a subset of the other basic map
2269 * => the pair can be replaced by a basic map consisting
2270 * of the valid constraints in both basic maps
2272 * 5. there is a single adjacent pair of an inequality and an equality,
2273 * the other constraints of the basic map containing the inequality are
2274 * "valid". Moreover, if the inequality the basic map is relaxed
2275 * and then turned into an equality, then resulting facet lies
2276 * entirely inside the other basic map
2277 * => the pair can be replaced by the basic map containing
2278 * the inequality, with the inequality relaxed.
2280 * 6. there is a single inequality adjacent to an equality,
2281 * the other constraints of the basic map containing the inequality are
2282 * "valid". Moreover, the facets corresponding to both
2283 * the inequality and the equality can be wrapped around their
2284 * ridges to include the other basic map
2285 * => the pair can be replaced by a basic map consisting
2286 * of the valid constraints in both basic maps together
2287 * with all wrapping constraints
2289 * 7. one of the basic maps extends beyond the other by at most one.
2290 * Moreover, the facets corresponding to the cut constraints and
2291 * the pieces of the other basic map at offset one from these cut
2292 * constraints can be wrapped around their ridges to include
2293 * the union of the two basic maps
2294 * => the pair can be replaced by a basic map consisting
2295 * of the valid constraints in both basic maps together
2296 * with all wrapping constraints
2298 * 8. the two basic maps live in adjacent hyperplanes. In principle
2299 * such sets can always be combined through wrapping, but we impose
2300 * that there is only one such pair, to avoid overeager coalescing.
2302 * Throughout the computation, we maintain a collection of tableaus
2303 * corresponding to the basic maps. When the basic maps are dropped
2304 * or combined, the tableaus are modified accordingly.
2306 static enum isl_change coalesce_local_pair_reuse(int i, int j,
2307 struct isl_coalesce_info *info)
2309 enum isl_change change = isl_change_none;
2311 set_ineq_status_in(&info[i], info[j].tab);
2312 if (info[i].bmap->n_ineq && !info[i].ineq)
2313 goto error;
2314 if (any_ineq(&info[i], STATUS_ERROR))
2315 goto error;
2316 if (any_ineq(&info[i], STATUS_SEPARATE))
2317 goto done;
2319 set_ineq_status_in(&info[j], info[i].tab);
2320 if (info[j].bmap->n_ineq && !info[j].ineq)
2321 goto error;
2322 if (any_ineq(&info[j], STATUS_ERROR))
2323 goto error;
2324 if (any_ineq(&info[j], STATUS_SEPARATE))
2325 goto done;
2327 set_eq_status_in(&info[i], info[j].tab);
2328 if (info[i].bmap->n_eq && !info[i].eq)
2329 goto error;
2330 if (any_eq(&info[i], STATUS_ERROR))
2331 goto error;
2333 set_eq_status_in(&info[j], info[i].tab);
2334 if (info[j].bmap->n_eq && !info[j].eq)
2335 goto error;
2336 if (any_eq(&info[j], STATUS_ERROR))
2337 goto error;
2339 if (any_eq(&info[i], STATUS_SEPARATE))
2340 return separating_equality(i, j, info);
2341 if (any_eq(&info[j], STATUS_SEPARATE))
2342 return separating_equality(j, i, info);
2344 if (all(info[i].eq, 2 * info[i].bmap->n_eq, STATUS_VALID) &&
2345 all(info[i].ineq, info[i].bmap->n_ineq, STATUS_VALID)) {
2346 drop(&info[j]);
2347 change = isl_change_drop_second;
2348 } else if (all(info[j].eq, 2 * info[j].bmap->n_eq, STATUS_VALID) &&
2349 all(info[j].ineq, info[j].bmap->n_ineq, STATUS_VALID)) {
2350 drop(&info[i]);
2351 change = isl_change_drop_first;
2352 } else if (any_eq(&info[i], STATUS_ADJ_EQ)) {
2353 change = check_eq_adj_eq(i, j, info);
2354 } else if (any_eq(&info[j], STATUS_ADJ_EQ)) {
2355 change = check_eq_adj_eq(j, i, info);
2356 } else if (any_eq(&info[i], STATUS_ADJ_INEQ) ||
2357 any_eq(&info[j], STATUS_ADJ_INEQ)) {
2358 change = check_adj_eq(i, j, info);
2359 } else if (any_ineq(&info[i], STATUS_ADJ_EQ)) {
2360 change = check_ineq_adj_eq(i, j, info);
2361 } else if (any_ineq(&info[j], STATUS_ADJ_EQ)) {
2362 change = check_ineq_adj_eq(j, i, info);
2363 } else if (any_ineq(&info[i], STATUS_ADJ_INEQ) ||
2364 any_ineq(&info[j], STATUS_ADJ_INEQ)) {
2365 change = check_adj_ineq(i, j, info);
2366 } else {
2367 if (!any_eq(&info[i], STATUS_CUT) &&
2368 !any_eq(&info[j], STATUS_CUT))
2369 change = check_facets(i, j, info);
2370 if (change == isl_change_none)
2371 change = check_wrap(i, j, info);
2374 done:
2375 clear_status(&info[i]);
2376 clear_status(&info[j]);
2377 return change;
2378 error:
2379 clear_status(&info[i]);
2380 clear_status(&info[j]);
2381 return isl_change_error;
2384 /* Check if the union of the given pair of basic maps
2385 * can be represented by a single basic map.
2386 * If so, replace the pair by the single basic map and return
2387 * isl_change_drop_first, isl_change_drop_second or isl_change_fuse.
2388 * Otherwise, return isl_change_none.
2389 * The two basic maps are assumed to live in the same local space.
2391 static enum isl_change coalesce_local_pair(int i, int j,
2392 struct isl_coalesce_info *info)
2394 init_status(&info[i]);
2395 init_status(&info[j]);
2396 return coalesce_local_pair_reuse(i, j, info);
2399 /* Shift the integer division at position "div" of the basic map
2400 * represented by "info" by "shift".
2402 * That is, if the integer division has the form
2404 * floor(f(x)/d)
2406 * then replace it by
2408 * floor((f(x) + shift * d)/d) - shift
2410 static isl_stat shift_div(struct isl_coalesce_info *info, int div,
2411 isl_int shift)
2413 isl_size total, n_div;
2415 info->bmap = isl_basic_map_shift_div(info->bmap, div, 0, shift);
2416 if (!info->bmap)
2417 return isl_stat_error;
2419 total = isl_basic_map_dim(info->bmap, isl_dim_all);
2420 n_div = isl_basic_map_dim(info->bmap, isl_dim_div);
2421 if (total < 0 || n_div < 0)
2422 return isl_stat_error;
2423 total -= n_div;
2424 if (isl_tab_shift_var(info->tab, total + div, shift) < 0)
2425 return isl_stat_error;
2427 return isl_stat_ok;
2430 /* If the integer division at position "div" is defined by an equality,
2431 * i.e., a stride constraint, then change the integer division expression
2432 * to have a constant term equal to zero.
2434 * Let the equality constraint be
2436 * c + f + m a = 0
2438 * The integer division expression is then typically of the form
2440 * a = floor((-f - c')/m)
2442 * The integer division is first shifted by t = floor(c/m),
2443 * turning the equality constraint into
2445 * c - m floor(c/m) + f + m a' = 0
2447 * i.e.,
2449 * (c mod m) + f + m a' = 0
2451 * That is,
2453 * a' = (-f - (c mod m))/m = floor((-f)/m)
2455 * because a' is an integer and 0 <= (c mod m) < m.
2456 * The constant term of a' can therefore be zeroed out,
2457 * but only if the integer division expression is of the expected form.
2459 static isl_stat normalize_stride_div(struct isl_coalesce_info *info, int div)
2461 isl_bool defined, valid;
2462 isl_stat r;
2463 isl_constraint *c;
2464 isl_int shift, stride;
2466 defined = isl_basic_map_has_defining_equality(info->bmap, isl_dim_div,
2467 div, &c);
2468 if (defined < 0)
2469 return isl_stat_error;
2470 if (!defined)
2471 return isl_stat_ok;
2472 if (!c)
2473 return isl_stat_error;
2474 valid = isl_constraint_is_div_equality(c, div);
2475 isl_int_init(shift);
2476 isl_int_init(stride);
2477 isl_constraint_get_constant(c, &shift);
2478 isl_constraint_get_coefficient(c, isl_dim_div, div, &stride);
2479 isl_int_fdiv_q(shift, shift, stride);
2480 r = shift_div(info, div, shift);
2481 isl_int_clear(stride);
2482 isl_int_clear(shift);
2483 isl_constraint_free(c);
2484 if (r < 0 || valid < 0)
2485 return isl_stat_error;
2486 if (!valid)
2487 return isl_stat_ok;
2488 info->bmap = isl_basic_map_set_div_expr_constant_num_si_inplace(
2489 info->bmap, div, 0);
2490 if (!info->bmap)
2491 return isl_stat_error;
2492 return isl_stat_ok;
2495 /* The basic maps represented by "info1" and "info2" are known
2496 * to have the same number of integer divisions.
2497 * Check if pairs of integer divisions are equal to each other
2498 * despite the fact that they differ by a rational constant.
2500 * In particular, look for any pair of integer divisions that
2501 * only differ in their constant terms.
2502 * If either of these integer divisions is defined
2503 * by stride constraints, then modify it to have a zero constant term.
2504 * If both are defined by stride constraints then in the end they will have
2505 * the same (zero) constant term.
2507 static isl_stat harmonize_stride_divs(struct isl_coalesce_info *info1,
2508 struct isl_coalesce_info *info2)
2510 int i;
2511 isl_size n;
2513 n = isl_basic_map_dim(info1->bmap, isl_dim_div);
2514 if (n < 0)
2515 return isl_stat_error;
2516 for (i = 0; i < n; ++i) {
2517 isl_bool known, harmonize;
2519 known = isl_basic_map_div_is_known(info1->bmap, i);
2520 if (known >= 0 && known)
2521 known = isl_basic_map_div_is_known(info2->bmap, i);
2522 if (known < 0)
2523 return isl_stat_error;
2524 if (!known)
2525 continue;
2526 harmonize = isl_basic_map_equal_div_expr_except_constant(
2527 info1->bmap, i, info2->bmap, i);
2528 if (harmonize < 0)
2529 return isl_stat_error;
2530 if (!harmonize)
2531 continue;
2532 if (normalize_stride_div(info1, i) < 0)
2533 return isl_stat_error;
2534 if (normalize_stride_div(info2, i) < 0)
2535 return isl_stat_error;
2538 return isl_stat_ok;
2541 /* If "shift" is an integer constant, then shift the integer division
2542 * at position "div" of the basic map represented by "info" by "shift".
2543 * If "shift" is not an integer constant, then do nothing.
2544 * If "shift" is equal to zero, then no shift needs to be performed either.
2546 * That is, if the integer division has the form
2548 * floor(f(x)/d)
2550 * then replace it by
2552 * floor((f(x) + shift * d)/d) - shift
2554 static isl_stat shift_if_cst_int(struct isl_coalesce_info *info, int div,
2555 __isl_keep isl_aff *shift)
2557 isl_bool cst;
2558 isl_stat r;
2559 isl_int d;
2560 isl_val *c;
2562 cst = isl_aff_is_cst(shift);
2563 if (cst < 0 || !cst)
2564 return cst < 0 ? isl_stat_error : isl_stat_ok;
2566 c = isl_aff_get_constant_val(shift);
2567 cst = isl_val_is_int(c);
2568 if (cst >= 0 && cst)
2569 cst = isl_bool_not(isl_val_is_zero(c));
2570 if (cst < 0 || !cst) {
2571 isl_val_free(c);
2572 return cst < 0 ? isl_stat_error : isl_stat_ok;
2575 isl_int_init(d);
2576 r = isl_val_get_num_isl_int(c, &d);
2577 if (r >= 0)
2578 r = shift_div(info, div, d);
2579 isl_int_clear(d);
2581 isl_val_free(c);
2583 return r;
2586 /* Check if some of the divs in the basic map represented by "info1"
2587 * are shifts of the corresponding divs in the basic map represented
2588 * by "info2", taking into account the equality constraints "eq1" of "info1"
2589 * and "eq2" of "info2". If so, align them with those of "info2".
2590 * "info1" and "info2" are assumed to have the same number
2591 * of integer divisions.
2593 * An integer division is considered to be a shift of another integer
2594 * division if, after simplification with respect to the equality
2595 * constraints of the other basic map, one is equal to the other
2596 * plus a constant.
2598 * In particular, for each pair of integer divisions, if both are known,
2599 * have the same denominator and are not already equal to each other,
2600 * simplify each with respect to the equality constraints
2601 * of the other basic map. If the difference is an integer constant,
2602 * then move this difference outside.
2603 * That is, if, after simplification, one integer division is of the form
2605 * floor((f(x) + c_1)/d)
2607 * while the other is of the form
2609 * floor((f(x) + c_2)/d)
2611 * and n = (c_2 - c_1)/d is an integer, then replace the first
2612 * integer division by
2614 * floor((f_1(x) + c_1 + n * d)/d) - n,
2616 * where floor((f_1(x) + c_1 + n * d)/d) = floor((f2(x) + c_2)/d)
2617 * after simplification with respect to the equality constraints.
2619 static isl_stat harmonize_divs_with_hulls(struct isl_coalesce_info *info1,
2620 struct isl_coalesce_info *info2, __isl_keep isl_basic_set *eq1,
2621 __isl_keep isl_basic_set *eq2)
2623 int i;
2624 isl_size total;
2625 isl_local_space *ls1, *ls2;
2627 total = isl_basic_map_dim(info1->bmap, isl_dim_all);
2628 if (total < 0)
2629 return isl_stat_error;
2630 ls1 = isl_local_space_wrap(isl_basic_map_get_local_space(info1->bmap));
2631 ls2 = isl_local_space_wrap(isl_basic_map_get_local_space(info2->bmap));
2632 for (i = 0; i < info1->bmap->n_div; ++i) {
2633 isl_stat r;
2634 isl_aff *div1, *div2;
2636 if (!isl_local_space_div_is_known(ls1, i) ||
2637 !isl_local_space_div_is_known(ls2, i))
2638 continue;
2639 if (isl_int_ne(info1->bmap->div[i][0], info2->bmap->div[i][0]))
2640 continue;
2641 if (isl_seq_eq(info1->bmap->div[i] + 1,
2642 info2->bmap->div[i] + 1, 1 + total))
2643 continue;
2644 div1 = isl_local_space_get_div(ls1, i);
2645 div2 = isl_local_space_get_div(ls2, i);
2646 div1 = isl_aff_substitute_equalities(div1,
2647 isl_basic_set_copy(eq2));
2648 div2 = isl_aff_substitute_equalities(div2,
2649 isl_basic_set_copy(eq1));
2650 div2 = isl_aff_sub(div2, div1);
2651 r = shift_if_cst_int(info1, i, div2);
2652 isl_aff_free(div2);
2653 if (r < 0)
2654 break;
2656 isl_local_space_free(ls1);
2657 isl_local_space_free(ls2);
2659 if (i < info1->bmap->n_div)
2660 return isl_stat_error;
2661 return isl_stat_ok;
2664 /* Check if some of the divs in the basic map represented by "info1"
2665 * are shifts of the corresponding divs in the basic map represented
2666 * by "info2". If so, align them with those of "info2".
2667 * Only do this if "info1" and "info2" have the same number
2668 * of integer divisions.
2670 * An integer division is considered to be a shift of another integer
2671 * division if, after simplification with respect to the equality
2672 * constraints of the other basic map, one is equal to the other
2673 * plus a constant.
2675 * First check if pairs of integer divisions are equal to each other
2676 * despite the fact that they differ by a rational constant.
2677 * If so, try and arrange for them to have the same constant term.
2679 * Then, extract the equality constraints and continue with
2680 * harmonize_divs_with_hulls.
2682 * If the equality constraints of both basic maps are the same,
2683 * then there is no need to perform any shifting since
2684 * the coefficients of the integer divisions should have been
2685 * reduced in the same way.
2687 static isl_stat harmonize_divs(struct isl_coalesce_info *info1,
2688 struct isl_coalesce_info *info2)
2690 isl_bool equal;
2691 isl_basic_map *bmap1, *bmap2;
2692 isl_basic_set *eq1, *eq2;
2693 isl_stat r;
2695 if (!info1->bmap || !info2->bmap)
2696 return isl_stat_error;
2698 if (info1->bmap->n_div != info2->bmap->n_div)
2699 return isl_stat_ok;
2700 if (info1->bmap->n_div == 0)
2701 return isl_stat_ok;
2703 if (harmonize_stride_divs(info1, info2) < 0)
2704 return isl_stat_error;
2706 bmap1 = isl_basic_map_copy(info1->bmap);
2707 bmap2 = isl_basic_map_copy(info2->bmap);
2708 eq1 = isl_basic_map_wrap(isl_basic_map_plain_affine_hull(bmap1));
2709 eq2 = isl_basic_map_wrap(isl_basic_map_plain_affine_hull(bmap2));
2710 equal = isl_basic_set_plain_is_equal(eq1, eq2);
2711 if (equal < 0)
2712 r = isl_stat_error;
2713 else if (equal)
2714 r = isl_stat_ok;
2715 else
2716 r = harmonize_divs_with_hulls(info1, info2, eq1, eq2);
2717 isl_basic_set_free(eq1);
2718 isl_basic_set_free(eq2);
2720 return r;
2723 /* Do the two basic maps live in the same local space, i.e.,
2724 * do they have the same (known) divs?
2725 * If either basic map has any unknown divs, then we can only assume
2726 * that they do not live in the same local space.
2728 static isl_bool same_divs(__isl_keep isl_basic_map *bmap1,
2729 __isl_keep isl_basic_map *bmap2)
2731 int i;
2732 isl_bool known;
2733 isl_size total;
2735 if (!bmap1 || !bmap2)
2736 return isl_bool_error;
2737 if (bmap1->n_div != bmap2->n_div)
2738 return isl_bool_false;
2740 if (bmap1->n_div == 0)
2741 return isl_bool_true;
2743 known = isl_basic_map_divs_known(bmap1);
2744 if (known < 0 || !known)
2745 return known;
2746 known = isl_basic_map_divs_known(bmap2);
2747 if (known < 0 || !known)
2748 return known;
2750 total = isl_basic_map_dim(bmap1, isl_dim_all);
2751 if (total < 0)
2752 return isl_bool_error;
2753 for (i = 0; i < bmap1->n_div; ++i)
2754 if (!isl_seq_eq(bmap1->div[i], bmap2->div[i], 2 + total))
2755 return isl_bool_false;
2757 return isl_bool_true;
2760 /* Assuming that "tab" contains the equality constraints and
2761 * the initial inequality constraints of "bmap", copy the remaining
2762 * inequality constraints of "bmap" to "Tab".
2764 static isl_stat copy_ineq(struct isl_tab *tab, __isl_keep isl_basic_map *bmap)
2766 int i, n_ineq;
2768 if (!bmap)
2769 return isl_stat_error;
2771 n_ineq = tab->n_con - tab->n_eq;
2772 for (i = n_ineq; i < bmap->n_ineq; ++i)
2773 if (isl_tab_add_ineq(tab, bmap->ineq[i]) < 0)
2774 return isl_stat_error;
2776 return isl_stat_ok;
2779 /* Description of an integer division that is added
2780 * during an expansion.
2781 * "pos" is the position of the corresponding variable.
2782 * "cst" indicates whether this integer division has a fixed value.
2783 * "val" contains the fixed value, if the value is fixed.
2785 struct isl_expanded {
2786 int pos;
2787 isl_bool cst;
2788 isl_int val;
2791 /* For each of the "n" integer division variables "expanded",
2792 * if the variable has a fixed value, then add two inequality
2793 * constraints expressing the fixed value.
2794 * Otherwise, add the corresponding div constraints.
2795 * The caller is responsible for removing the div constraints
2796 * that it added for all these "n" integer divisions.
2798 * The div constraints and the pair of inequality constraints
2799 * forcing the fixed value cannot both be added for a given variable
2800 * as the combination may render some of the original constraints redundant.
2801 * These would then be ignored during the coalescing detection,
2802 * while they could remain in the fused result.
2804 * The two added inequality constraints are
2806 * -a + v >= 0
2807 * a - v >= 0
2809 * with "a" the variable and "v" its fixed value.
2810 * The facet corresponding to one of these two constraints is selected
2811 * in the tableau to ensure that the pair of inequality constraints
2812 * is treated as an equality constraint.
2814 * The information in info->ineq is thrown away because it was
2815 * computed in terms of div constraints, while some of those
2816 * have now been replaced by these pairs of inequality constraints.
2818 static isl_stat fix_constant_divs(struct isl_coalesce_info *info,
2819 int n, struct isl_expanded *expanded)
2821 unsigned o_div;
2822 int i;
2823 isl_vec *ineq;
2825 o_div = isl_basic_map_offset(info->bmap, isl_dim_div) - 1;
2826 ineq = isl_vec_alloc(isl_tab_get_ctx(info->tab), 1 + info->tab->n_var);
2827 if (!ineq)
2828 return isl_stat_error;
2829 isl_seq_clr(ineq->el + 1, info->tab->n_var);
2831 for (i = 0; i < n; ++i) {
2832 if (!expanded[i].cst) {
2833 info->bmap = isl_basic_map_extend_constraints(
2834 info->bmap, 0, 2);
2835 info->bmap = isl_basic_map_add_div_constraints(
2836 info->bmap, expanded[i].pos - o_div);
2837 } else {
2838 isl_int_set_si(ineq->el[1 + expanded[i].pos], -1);
2839 isl_int_set(ineq->el[0], expanded[i].val);
2840 info->bmap = isl_basic_map_add_ineq(info->bmap,
2841 ineq->el);
2842 isl_int_set_si(ineq->el[1 + expanded[i].pos], 1);
2843 isl_int_neg(ineq->el[0], expanded[i].val);
2844 info->bmap = isl_basic_map_add_ineq(info->bmap,
2845 ineq->el);
2846 isl_int_set_si(ineq->el[1 + expanded[i].pos], 0);
2848 if (copy_ineq(info->tab, info->bmap) < 0)
2849 break;
2850 if (expanded[i].cst &&
2851 isl_tab_select_facet(info->tab, info->tab->n_con - 1) < 0)
2852 break;
2855 isl_vec_free(ineq);
2857 clear_status(info);
2858 init_status(info);
2860 return i < n ? isl_stat_error : isl_stat_ok;
2863 /* Insert the "n" integer division variables "expanded"
2864 * into info->tab and info->bmap and
2865 * update info->ineq with respect to the redundant constraints
2866 * in the resulting tableau.
2867 * "bmap" contains the result of this insertion in info->bmap,
2868 * while info->bmap is the original version
2869 * of "bmap", i.e., the one that corresponds to the current
2870 * state of info->tab. The number of constraints in info->bmap
2871 * is assumed to be the same as the number of constraints
2872 * in info->tab. This is required to be able to detect
2873 * the extra constraints in "bmap".
2875 * In particular, introduce extra variables corresponding
2876 * to the extra integer divisions and add the div constraints
2877 * that were added to "bmap" after info->tab was created
2878 * from info->bmap.
2879 * Furthermore, check if these extra integer divisions happen
2880 * to attain a fixed integer value in info->tab.
2881 * If so, replace the corresponding div constraints by pairs
2882 * of inequality constraints that fix these
2883 * integer divisions to their single integer values.
2884 * Replace info->bmap by "bmap" to match the changes to info->tab.
2885 * info->ineq was computed without a tableau and therefore
2886 * does not take into account the redundant constraints
2887 * in the tableau. Mark them here.
2888 * There is no need to check the newly added div constraints
2889 * since they cannot be redundant.
2890 * The redundancy check is not performed when constants have been discovered
2891 * since info->ineq is completely thrown away in this case.
2893 static isl_stat tab_insert_divs(struct isl_coalesce_info *info,
2894 int n, struct isl_expanded *expanded, __isl_take isl_basic_map *bmap)
2896 int i, n_ineq;
2897 unsigned n_eq;
2898 struct isl_tab_undo *snap;
2899 int any;
2901 if (!bmap)
2902 return isl_stat_error;
2903 if (info->bmap->n_eq + info->bmap->n_ineq != info->tab->n_con)
2904 isl_die(isl_basic_map_get_ctx(bmap), isl_error_internal,
2905 "original tableau does not correspond "
2906 "to original basic map", goto error);
2908 if (isl_tab_extend_vars(info->tab, n) < 0)
2909 goto error;
2910 if (isl_tab_extend_cons(info->tab, 2 * n) < 0)
2911 goto error;
2913 for (i = 0; i < n; ++i) {
2914 if (isl_tab_insert_var(info->tab, expanded[i].pos) < 0)
2915 goto error;
2918 snap = isl_tab_snap(info->tab);
2920 n_ineq = info->tab->n_con - info->tab->n_eq;
2921 if (copy_ineq(info->tab, bmap) < 0)
2922 goto error;
2924 isl_basic_map_free(info->bmap);
2925 info->bmap = bmap;
2927 any = 0;
2928 for (i = 0; i < n; ++i) {
2929 expanded[i].cst = isl_tab_is_constant(info->tab,
2930 expanded[i].pos, &expanded[i].val);
2931 if (expanded[i].cst < 0)
2932 return isl_stat_error;
2933 if (expanded[i].cst)
2934 any = 1;
2937 if (any) {
2938 if (isl_tab_rollback(info->tab, snap) < 0)
2939 return isl_stat_error;
2940 info->bmap = isl_basic_map_cow(info->bmap);
2941 if (isl_basic_map_free_inequality(info->bmap, 2 * n) < 0)
2942 return isl_stat_error;
2944 return fix_constant_divs(info, n, expanded);
2947 n_eq = info->bmap->n_eq;
2948 for (i = 0; i < n_ineq; ++i) {
2949 if (isl_tab_is_redundant(info->tab, n_eq + i))
2950 info->ineq[i] = STATUS_REDUNDANT;
2953 return isl_stat_ok;
2954 error:
2955 isl_basic_map_free(bmap);
2956 return isl_stat_error;
2959 /* Expand info->tab and info->bmap in the same way "bmap" was expanded
2960 * in isl_basic_map_expand_divs using the expansion "exp" and
2961 * update info->ineq with respect to the redundant constraints
2962 * in the resulting tableau. info->bmap is the original version
2963 * of "bmap", i.e., the one that corresponds to the current
2964 * state of info->tab. The number of constraints in info->bmap
2965 * is assumed to be the same as the number of constraints
2966 * in info->tab. This is required to be able to detect
2967 * the extra constraints in "bmap".
2969 * Extract the positions where extra local variables are introduced
2970 * from "exp" and call tab_insert_divs.
2972 static isl_stat expand_tab(struct isl_coalesce_info *info, int *exp,
2973 __isl_take isl_basic_map *bmap)
2975 isl_ctx *ctx;
2976 struct isl_expanded *expanded;
2977 int i, j, k, n;
2978 int extra_var;
2979 isl_size total, n_div;
2980 unsigned pos;
2981 isl_stat r;
2983 total = isl_basic_map_dim(bmap, isl_dim_all);
2984 n_div = isl_basic_map_dim(bmap, isl_dim_div);
2985 if (total < 0 || n_div < 0)
2986 return isl_stat_error;
2987 pos = total - n_div;
2988 extra_var = total - info->tab->n_var;
2989 n = n_div - extra_var;
2991 ctx = isl_basic_map_get_ctx(bmap);
2992 expanded = isl_calloc_array(ctx, struct isl_expanded, extra_var);
2993 if (extra_var && !expanded)
2994 goto error;
2996 i = 0;
2997 k = 0;
2998 for (j = 0; j < n_div; ++j) {
2999 if (i < n && exp[i] == j) {
3000 ++i;
3001 continue;
3003 expanded[k++].pos = pos + j;
3006 for (k = 0; k < extra_var; ++k)
3007 isl_int_init(expanded[k].val);
3009 r = tab_insert_divs(info, extra_var, expanded, bmap);
3011 for (k = 0; k < extra_var; ++k)
3012 isl_int_clear(expanded[k].val);
3013 free(expanded);
3015 return r;
3016 error:
3017 isl_basic_map_free(bmap);
3018 return isl_stat_error;
3021 /* Check if the union of the basic maps represented by info[i] and info[j]
3022 * can be represented by a single basic map,
3023 * after expanding the divs of info[i] to match those of info[j].
3024 * If so, replace the pair by the single basic map and return
3025 * isl_change_drop_first, isl_change_drop_second or isl_change_fuse.
3026 * Otherwise, return isl_change_none.
3028 * The caller has already checked for info[j] being a subset of info[i].
3029 * If some of the divs of info[j] are unknown, then the expanded info[i]
3030 * will not have the corresponding div constraints. The other patterns
3031 * therefore cannot apply. Skip the computation in this case.
3033 * The expansion is performed using the divs "div" and expansion "exp"
3034 * computed by the caller.
3035 * info[i].bmap has already been expanded and the result is passed in
3036 * as "bmap".
3037 * The "eq" and "ineq" fields of info[i] reflect the status of
3038 * the constraints of the expanded "bmap" with respect to info[j].tab.
3039 * However, inequality constraints that are redundant in info[i].tab
3040 * have not yet been marked as such because no tableau was available.
3042 * Replace info[i].bmap by "bmap" and expand info[i].tab as well,
3043 * updating info[i].ineq with respect to the redundant constraints.
3044 * Then try and coalesce the expanded info[i] with info[j],
3045 * reusing the information in info[i].eq and info[i].ineq.
3046 * If this does not result in any coalescing or if it results in info[j]
3047 * getting dropped (which should not happen in practice, since the case
3048 * of info[j] being a subset of info[i] has already been checked by
3049 * the caller), then revert info[i] to its original state.
3051 static enum isl_change coalesce_expand_tab_divs(__isl_take isl_basic_map *bmap,
3052 int i, int j, struct isl_coalesce_info *info, __isl_keep isl_mat *div,
3053 int *exp)
3055 isl_bool known;
3056 isl_basic_map *bmap_i;
3057 struct isl_tab_undo *snap;
3058 enum isl_change change = isl_change_none;
3060 known = isl_basic_map_divs_known(info[j].bmap);
3061 if (known < 0 || !known) {
3062 clear_status(&info[i]);
3063 isl_basic_map_free(bmap);
3064 return known < 0 ? isl_change_error : isl_change_none;
3067 bmap_i = isl_basic_map_copy(info[i].bmap);
3068 snap = isl_tab_snap(info[i].tab);
3069 if (expand_tab(&info[i], exp, bmap) < 0)
3070 change = isl_change_error;
3072 init_status(&info[j]);
3073 if (change == isl_change_none)
3074 change = coalesce_local_pair_reuse(i, j, info);
3075 else
3076 clear_status(&info[i]);
3077 if (change != isl_change_none && change != isl_change_drop_second) {
3078 isl_basic_map_free(bmap_i);
3079 } else {
3080 isl_basic_map_free(info[i].bmap);
3081 info[i].bmap = bmap_i;
3083 if (isl_tab_rollback(info[i].tab, snap) < 0)
3084 change = isl_change_error;
3087 return change;
3090 /* Check if the union of "bmap" and the basic map represented by info[j]
3091 * can be represented by a single basic map,
3092 * after expanding the divs of "bmap" to match those of info[j].
3093 * If so, replace the pair by the single basic map and return
3094 * isl_change_drop_first, isl_change_drop_second or isl_change_fuse.
3095 * Otherwise, return isl_change_none.
3097 * In particular, check if the expanded "bmap" contains the basic map
3098 * represented by the tableau info[j].tab.
3099 * The expansion is performed using the divs "div" and expansion "exp"
3100 * computed by the caller.
3101 * Then we check if all constraints of the expanded "bmap" are valid for
3102 * info[j].tab.
3104 * If "i" is not equal to -1, then "bmap" is equal to info[i].bmap.
3105 * In this case, the positions of the constraints of info[i].bmap
3106 * with respect to the basic map represented by info[j] are stored
3107 * in info[i].
3109 * If the expanded "bmap" does not contain the basic map
3110 * represented by the tableau info[j].tab and if "i" is not -1,
3111 * i.e., if the original "bmap" is info[i].bmap, then expand info[i].tab
3112 * as well and check if that results in coalescing.
3114 static enum isl_change coalesce_with_expanded_divs(
3115 __isl_keep isl_basic_map *bmap, int i, int j,
3116 struct isl_coalesce_info *info, __isl_keep isl_mat *div, int *exp)
3118 enum isl_change change = isl_change_none;
3119 struct isl_coalesce_info info_local, *info_i;
3121 info_i = i >= 0 ? &info[i] : &info_local;
3122 init_status(info_i);
3123 bmap = isl_basic_map_copy(bmap);
3124 bmap = isl_basic_map_expand_divs(bmap, isl_mat_copy(div), exp);
3125 bmap = isl_basic_map_mark_final(bmap);
3127 if (!bmap)
3128 goto error;
3130 info_local.bmap = bmap;
3131 info_i->eq = eq_status_in(bmap, info[j].tab);
3132 if (bmap->n_eq && !info_i->eq)
3133 goto error;
3134 if (any_eq(info_i, STATUS_ERROR))
3135 goto error;
3136 if (any_eq(info_i, STATUS_SEPARATE))
3137 goto done;
3139 info_i->ineq = ineq_status_in(bmap, NULL, info[j].tab);
3140 if (bmap->n_ineq && !info_i->ineq)
3141 goto error;
3142 if (any_ineq(info_i, STATUS_ERROR))
3143 goto error;
3144 if (any_ineq(info_i, STATUS_SEPARATE))
3145 goto done;
3147 if (all(info_i->eq, 2 * bmap->n_eq, STATUS_VALID) &&
3148 all(info_i->ineq, bmap->n_ineq, STATUS_VALID)) {
3149 drop(&info[j]);
3150 change = isl_change_drop_second;
3153 if (change == isl_change_none && i != -1)
3154 return coalesce_expand_tab_divs(bmap, i, j, info, div, exp);
3156 done:
3157 isl_basic_map_free(bmap);
3158 clear_status(info_i);
3159 return change;
3160 error:
3161 isl_basic_map_free(bmap);
3162 clear_status(info_i);
3163 return isl_change_error;
3166 /* Check if the union of "bmap_i" and the basic map represented by info[j]
3167 * can be represented by a single basic map,
3168 * after aligning the divs of "bmap_i" to match those of info[j].
3169 * If so, replace the pair by the single basic map and return
3170 * isl_change_drop_first, isl_change_drop_second or isl_change_fuse.
3171 * Otherwise, return isl_change_none.
3173 * In particular, check if "bmap_i" contains the basic map represented by
3174 * info[j] after aligning the divs of "bmap_i" to those of info[j].
3175 * Note that this can only succeed if the number of divs of "bmap_i"
3176 * is smaller than (or equal to) the number of divs of info[j].
3178 * We first check if the divs of "bmap_i" are all known and form a subset
3179 * of those of info[j].bmap. If so, we pass control over to
3180 * coalesce_with_expanded_divs.
3182 * If "i" is not equal to -1, then "bmap" is equal to info[i].bmap.
3184 static enum isl_change coalesce_after_aligning_divs(
3185 __isl_keep isl_basic_map *bmap_i, int i, int j,
3186 struct isl_coalesce_info *info)
3188 isl_bool known;
3189 isl_mat *div_i, *div_j, *div;
3190 int *exp1 = NULL;
3191 int *exp2 = NULL;
3192 isl_ctx *ctx;
3193 enum isl_change change;
3195 known = isl_basic_map_divs_known(bmap_i);
3196 if (known < 0)
3197 return isl_change_error;
3198 if (!known)
3199 return isl_change_none;
3201 ctx = isl_basic_map_get_ctx(bmap_i);
3203 div_i = isl_basic_map_get_divs(bmap_i);
3204 div_j = isl_basic_map_get_divs(info[j].bmap);
3206 if (!div_i || !div_j)
3207 goto error;
3209 exp1 = isl_alloc_array(ctx, int, div_i->n_row);
3210 exp2 = isl_alloc_array(ctx, int, div_j->n_row);
3211 if ((div_i->n_row && !exp1) || (div_j->n_row && !exp2))
3212 goto error;
3214 div = isl_merge_divs(div_i, div_j, exp1, exp2);
3215 if (!div)
3216 goto error;
3218 if (div->n_row == div_j->n_row)
3219 change = coalesce_with_expanded_divs(bmap_i,
3220 i, j, info, div, exp1);
3221 else
3222 change = isl_change_none;
3224 isl_mat_free(div);
3226 isl_mat_free(div_i);
3227 isl_mat_free(div_j);
3229 free(exp2);
3230 free(exp1);
3232 return change;
3233 error:
3234 isl_mat_free(div_i);
3235 isl_mat_free(div_j);
3236 free(exp1);
3237 free(exp2);
3238 return isl_change_error;
3241 /* Check if basic map "j" is a subset of basic map "i" after
3242 * exploiting the extra equalities of "j" to simplify the divs of "i".
3243 * If so, remove basic map "j" and return isl_change_drop_second.
3245 * If "j" does not have any equalities or if they are the same
3246 * as those of "i", then we cannot exploit them to simplify the divs.
3247 * Similarly, if there are no divs in "i", then they cannot be simplified.
3248 * If, on the other hand, the affine hulls of "i" and "j" do not intersect,
3249 * then "j" cannot be a subset of "i".
3251 * Otherwise, we intersect "i" with the affine hull of "j" and then
3252 * check if "j" is a subset of the result after aligning the divs.
3253 * If so, then "j" is definitely a subset of "i" and can be removed.
3254 * Note that if after intersection with the affine hull of "j".
3255 * "i" still has more divs than "j", then there is no way we can
3256 * align the divs of "i" to those of "j".
3258 static enum isl_change coalesce_subset_with_equalities(int i, int j,
3259 struct isl_coalesce_info *info)
3261 isl_basic_map *hull_i, *hull_j, *bmap_i;
3262 int equal, empty;
3263 enum isl_change change;
3265 if (info[j].bmap->n_eq == 0)
3266 return isl_change_none;
3267 if (info[i].bmap->n_div == 0)
3268 return isl_change_none;
3270 hull_i = isl_basic_map_copy(info[i].bmap);
3271 hull_i = isl_basic_map_plain_affine_hull(hull_i);
3272 hull_j = isl_basic_map_copy(info[j].bmap);
3273 hull_j = isl_basic_map_plain_affine_hull(hull_j);
3275 hull_j = isl_basic_map_intersect(hull_j, isl_basic_map_copy(hull_i));
3276 equal = isl_basic_map_plain_is_equal(hull_i, hull_j);
3277 empty = isl_basic_map_plain_is_empty(hull_j);
3278 isl_basic_map_free(hull_i);
3280 if (equal < 0 || equal || empty < 0 || empty) {
3281 isl_basic_map_free(hull_j);
3282 if (equal < 0 || empty < 0)
3283 return isl_change_error;
3284 return isl_change_none;
3287 bmap_i = isl_basic_map_copy(info[i].bmap);
3288 bmap_i = isl_basic_map_intersect(bmap_i, hull_j);
3289 if (!bmap_i)
3290 return isl_change_error;
3292 if (bmap_i->n_div > info[j].bmap->n_div) {
3293 isl_basic_map_free(bmap_i);
3294 return isl_change_none;
3297 change = coalesce_after_aligning_divs(bmap_i, -1, j, info);
3299 isl_basic_map_free(bmap_i);
3301 return change;
3304 /* Check if the union of and the basic maps represented by info[i] and info[j]
3305 * can be represented by a single basic map, by aligning or equating
3306 * their integer divisions.
3307 * If so, replace the pair by the single basic map and return
3308 * isl_change_drop_first, isl_change_drop_second or isl_change_fuse.
3309 * Otherwise, return isl_change_none.
3311 * Note that we only perform any test if the number of divs is different
3312 * in the two basic maps. In case the number of divs is the same,
3313 * we have already established that the divs are different
3314 * in the two basic maps.
3315 * In particular, if the number of divs of basic map i is smaller than
3316 * the number of divs of basic map j, then we check if j is a subset of i
3317 * and vice versa.
3319 static enum isl_change coalesce_divs(int i, int j,
3320 struct isl_coalesce_info *info)
3322 enum isl_change change = isl_change_none;
3324 if (info[i].bmap->n_div < info[j].bmap->n_div)
3325 change = coalesce_after_aligning_divs(info[i].bmap, i, j, info);
3326 if (change != isl_change_none)
3327 return change;
3329 if (info[j].bmap->n_div < info[i].bmap->n_div)
3330 change = coalesce_after_aligning_divs(info[j].bmap, j, i, info);
3331 if (change != isl_change_none)
3332 return invert_change(change);
3334 change = coalesce_subset_with_equalities(i, j, info);
3335 if (change != isl_change_none)
3336 return change;
3338 change = coalesce_subset_with_equalities(j, i, info);
3339 if (change != isl_change_none)
3340 return invert_change(change);
3342 return isl_change_none;
3345 /* Does "bmap" involve any divs that themselves refer to divs?
3347 static isl_bool has_nested_div(__isl_keep isl_basic_map *bmap)
3349 int i;
3350 isl_size total;
3351 isl_size n_div;
3353 total = isl_basic_map_dim(bmap, isl_dim_all);
3354 n_div = isl_basic_map_dim(bmap, isl_dim_div);
3355 if (total < 0 || n_div < 0)
3356 return isl_bool_error;
3357 total -= n_div;
3359 for (i = 0; i < n_div; ++i)
3360 if (isl_seq_first_non_zero(bmap->div[i] + 2 + total,
3361 n_div) != -1)
3362 return isl_bool_true;
3364 return isl_bool_false;
3367 /* Return a list of affine expressions, one for each integer division
3368 * in "bmap_i". For each integer division that also appears in "bmap_j",
3369 * the affine expression is set to NaN. The number of NaNs in the list
3370 * is equal to the number of integer divisions in "bmap_j".
3371 * For the other integer divisions of "bmap_i", the corresponding
3372 * element in the list is a purely affine expression equal to the integer
3373 * division in "hull".
3374 * If no such list can be constructed, then the number of elements
3375 * in the returned list is smaller than the number of integer divisions
3376 * in "bmap_i".
3378 static __isl_give isl_aff_list *set_up_substitutions(
3379 __isl_keep isl_basic_map *bmap_i, __isl_keep isl_basic_map *bmap_j,
3380 __isl_take isl_basic_map *hull)
3382 isl_size n_div_i, n_div_j, total;
3383 isl_ctx *ctx;
3384 isl_local_space *ls;
3385 isl_basic_set *wrap_hull;
3386 isl_aff *aff_nan;
3387 isl_aff_list *list;
3388 int i, j;
3390 n_div_i = isl_basic_map_dim(bmap_i, isl_dim_div);
3391 n_div_j = isl_basic_map_dim(bmap_j, isl_dim_div);
3392 total = isl_basic_map_dim(bmap_i, isl_dim_all);
3393 if (!hull || n_div_i < 0 || n_div_j < 0 || total < 0)
3394 return NULL;
3396 ctx = isl_basic_map_get_ctx(hull);
3397 total -= n_div_i;
3399 ls = isl_basic_map_get_local_space(bmap_i);
3400 ls = isl_local_space_wrap(ls);
3401 wrap_hull = isl_basic_map_wrap(hull);
3403 aff_nan = isl_aff_nan_on_domain(isl_local_space_copy(ls));
3404 list = isl_aff_list_alloc(ctx, n_div_i);
3406 j = 0;
3407 for (i = 0; i < n_div_i; ++i) {
3408 isl_aff *aff;
3409 isl_size n_div;
3411 if (j < n_div_j &&
3412 isl_basic_map_equal_div_expr_part(bmap_i, i, bmap_j, j,
3413 0, 2 + total)) {
3414 ++j;
3415 list = isl_aff_list_add(list, isl_aff_copy(aff_nan));
3416 continue;
3418 if (n_div_i - i <= n_div_j - j)
3419 break;
3421 aff = isl_local_space_get_div(ls, i);
3422 aff = isl_aff_substitute_equalities(aff,
3423 isl_basic_set_copy(wrap_hull));
3424 aff = isl_aff_floor(aff);
3425 n_div = isl_aff_dim(aff, isl_dim_div);
3426 if (n_div < 0)
3427 goto error;
3428 if (n_div != 0) {
3429 isl_aff_free(aff);
3430 break;
3433 list = isl_aff_list_add(list, aff);
3436 isl_aff_free(aff_nan);
3437 isl_local_space_free(ls);
3438 isl_basic_set_free(wrap_hull);
3440 return list;
3441 error:
3442 isl_aff_free(aff_nan);
3443 isl_local_space_free(ls);
3444 isl_basic_set_free(wrap_hull);
3445 isl_aff_list_free(list);
3446 return NULL;
3449 /* Add variables to info->bmap and info->tab corresponding to the elements
3450 * in "list" that are not set to NaN.
3451 * "extra_var" is the number of these elements.
3452 * "dim" is the offset in the variables of "tab" where we should
3453 * start considering the elements in "list".
3454 * When this function returns, the total number of variables in "tab"
3455 * is equal to "dim" plus the number of elements in "list".
3457 * The newly added existentially quantified variables are not given
3458 * an explicit representation because the corresponding div constraints
3459 * do not appear in info->bmap. These constraints are not added
3460 * to info->bmap because for internal consistency, they would need to
3461 * be added to info->tab as well, where they could combine with the equality
3462 * that is added later to result in constraints that do not hold
3463 * in the original input.
3465 static isl_stat add_sub_vars(struct isl_coalesce_info *info,
3466 __isl_keep isl_aff_list *list, int dim, int extra_var)
3468 int i, j, n, d;
3469 isl_space *space;
3471 space = isl_basic_map_get_space(info->bmap);
3472 info->bmap = isl_basic_map_cow(info->bmap);
3473 info->bmap = isl_basic_map_extend_space(info->bmap, space,
3474 extra_var, 0, 0);
3475 if (!info->bmap)
3476 return isl_stat_error;
3477 n = isl_aff_list_n_aff(list);
3478 for (i = 0; i < n; ++i) {
3479 int is_nan;
3480 isl_aff *aff;
3482 aff = isl_aff_list_get_aff(list, i);
3483 is_nan = isl_aff_is_nan(aff);
3484 isl_aff_free(aff);
3485 if (is_nan < 0)
3486 return isl_stat_error;
3487 if (is_nan)
3488 continue;
3490 if (isl_tab_insert_var(info->tab, dim + i) < 0)
3491 return isl_stat_error;
3492 d = isl_basic_map_alloc_div(info->bmap);
3493 if (d < 0)
3494 return isl_stat_error;
3495 info->bmap = isl_basic_map_mark_div_unknown(info->bmap, d);
3496 for (j = d; j > i; --j)
3497 info->bmap = isl_basic_map_swap_div(info->bmap,
3498 j - 1, j);
3499 if (!info->bmap)
3500 return isl_stat_error;
3503 return isl_stat_ok;
3506 /* For each element in "list" that is not set to NaN, fix the corresponding
3507 * variable in "tab" to the purely affine expression defined by the element.
3508 * "dim" is the offset in the variables of "tab" where we should
3509 * start considering the elements in "list".
3511 * This function assumes that a sufficient number of rows and
3512 * elements in the constraint array are available in the tableau.
3514 static isl_stat add_sub_equalities(struct isl_tab *tab,
3515 __isl_keep isl_aff_list *list, int dim)
3517 int i, n;
3518 isl_ctx *ctx;
3519 isl_vec *sub;
3520 isl_aff *aff;
3522 n = isl_aff_list_n_aff(list);
3524 ctx = isl_tab_get_ctx(tab);
3525 sub = isl_vec_alloc(ctx, 1 + dim + n);
3526 if (!sub)
3527 return isl_stat_error;
3528 isl_seq_clr(sub->el + 1 + dim, n);
3530 for (i = 0; i < n; ++i) {
3531 aff = isl_aff_list_get_aff(list, i);
3532 if (!aff)
3533 goto error;
3534 if (isl_aff_is_nan(aff)) {
3535 isl_aff_free(aff);
3536 continue;
3538 isl_seq_cpy(sub->el, aff->v->el + 1, 1 + dim);
3539 isl_int_neg(sub->el[1 + dim + i], aff->v->el[0]);
3540 if (isl_tab_add_eq(tab, sub->el) < 0)
3541 goto error;
3542 isl_int_set_si(sub->el[1 + dim + i], 0);
3543 isl_aff_free(aff);
3546 isl_vec_free(sub);
3547 return isl_stat_ok;
3548 error:
3549 isl_aff_free(aff);
3550 isl_vec_free(sub);
3551 return isl_stat_error;
3554 /* Add variables to info->tab and info->bmap corresponding to the elements
3555 * in "list" that are not set to NaN. The value of the added variable
3556 * in info->tab is fixed to the purely affine expression defined by the element.
3557 * "dim" is the offset in the variables of info->tab where we should
3558 * start considering the elements in "list".
3559 * When this function returns, the total number of variables in info->tab
3560 * is equal to "dim" plus the number of elements in "list".
3562 static isl_stat add_subs(struct isl_coalesce_info *info,
3563 __isl_keep isl_aff_list *list, int dim)
3565 int extra_var;
3566 int n;
3568 if (!list)
3569 return isl_stat_error;
3571 n = isl_aff_list_n_aff(list);
3572 extra_var = n - (info->tab->n_var - dim);
3574 if (isl_tab_extend_vars(info->tab, extra_var) < 0)
3575 return isl_stat_error;
3576 if (isl_tab_extend_cons(info->tab, 2 * extra_var) < 0)
3577 return isl_stat_error;
3578 if (add_sub_vars(info, list, dim, extra_var) < 0)
3579 return isl_stat_error;
3581 return add_sub_equalities(info->tab, list, dim);
3584 /* Coalesce basic map "j" into basic map "i" after adding the extra integer
3585 * divisions in "i" but not in "j" to basic map "j", with values
3586 * specified by "list". The total number of elements in "list"
3587 * is equal to the number of integer divisions in "i", while the number
3588 * of NaN elements in the list is equal to the number of integer divisions
3589 * in "j".
3591 * If no coalescing can be performed, then we need to revert basic map "j"
3592 * to its original state. We do the same if basic map "i" gets dropped
3593 * during the coalescing, even though this should not happen in practice
3594 * since we have already checked for "j" being a subset of "i"
3595 * before we reach this stage.
3597 static enum isl_change coalesce_with_subs(int i, int j,
3598 struct isl_coalesce_info *info, __isl_keep isl_aff_list *list)
3600 isl_basic_map *bmap_j;
3601 struct isl_tab_undo *snap;
3602 isl_size dim, n_div;
3603 enum isl_change change;
3605 bmap_j = isl_basic_map_copy(info[j].bmap);
3606 snap = isl_tab_snap(info[j].tab);
3608 dim = isl_basic_map_dim(bmap_j, isl_dim_all);
3609 n_div = isl_basic_map_dim(bmap_j, isl_dim_div);
3610 if (dim < 0 || n_div < 0)
3611 goto error;
3612 dim -= n_div;
3613 if (add_subs(&info[j], list, dim) < 0)
3614 goto error;
3616 change = coalesce_local_pair(i, j, info);
3617 if (change != isl_change_none && change != isl_change_drop_first) {
3618 isl_basic_map_free(bmap_j);
3619 } else {
3620 isl_basic_map_free(info[j].bmap);
3621 info[j].bmap = bmap_j;
3623 if (isl_tab_rollback(info[j].tab, snap) < 0)
3624 return isl_change_error;
3627 return change;
3628 error:
3629 isl_basic_map_free(bmap_j);
3630 return isl_change_error;
3633 /* Check if we can coalesce basic map "j" into basic map "i" after copying
3634 * those extra integer divisions in "i" that can be simplified away
3635 * using the extra equalities in "j".
3636 * All divs are assumed to be known and not contain any nested divs.
3638 * We first check if there are any extra equalities in "j" that we
3639 * can exploit. Then we check if every integer division in "i"
3640 * either already appears in "j" or can be simplified using the
3641 * extra equalities to a purely affine expression.
3642 * If these tests succeed, then we try to coalesce the two basic maps
3643 * by introducing extra dimensions in "j" corresponding to
3644 * the extra integer divsisions "i" fixed to the corresponding
3645 * purely affine expression.
3647 static enum isl_change check_coalesce_into_eq(int i, int j,
3648 struct isl_coalesce_info *info)
3650 isl_size n_div_i, n_div_j;
3651 isl_basic_map *hull_i, *hull_j;
3652 isl_bool equal, empty;
3653 isl_aff_list *list;
3654 enum isl_change change;
3656 n_div_i = isl_basic_map_dim(info[i].bmap, isl_dim_div);
3657 n_div_j = isl_basic_map_dim(info[j].bmap, isl_dim_div);
3658 if (n_div_i < 0 || n_div_j < 0)
3659 return isl_change_error;
3660 if (n_div_i <= n_div_j)
3661 return isl_change_none;
3662 if (info[j].bmap->n_eq == 0)
3663 return isl_change_none;
3665 hull_i = isl_basic_map_copy(info[i].bmap);
3666 hull_i = isl_basic_map_plain_affine_hull(hull_i);
3667 hull_j = isl_basic_map_copy(info[j].bmap);
3668 hull_j = isl_basic_map_plain_affine_hull(hull_j);
3670 hull_j = isl_basic_map_intersect(hull_j, isl_basic_map_copy(hull_i));
3671 equal = isl_basic_map_plain_is_equal(hull_i, hull_j);
3672 empty = isl_basic_map_plain_is_empty(hull_j);
3673 isl_basic_map_free(hull_i);
3675 if (equal < 0 || empty < 0)
3676 goto error;
3677 if (equal || empty) {
3678 isl_basic_map_free(hull_j);
3679 return isl_change_none;
3682 list = set_up_substitutions(info[i].bmap, info[j].bmap, hull_j);
3683 if (!list)
3684 return isl_change_error;
3685 if (isl_aff_list_n_aff(list) < n_div_i)
3686 change = isl_change_none;
3687 else
3688 change = coalesce_with_subs(i, j, info, list);
3690 isl_aff_list_free(list);
3692 return change;
3693 error:
3694 isl_basic_map_free(hull_j);
3695 return isl_change_error;
3698 /* Check if we can coalesce basic maps "i" and "j" after copying
3699 * those extra integer divisions in one of the basic maps that can
3700 * be simplified away using the extra equalities in the other basic map.
3701 * We require all divs to be known in both basic maps.
3702 * Furthermore, to simplify the comparison of div expressions,
3703 * we do not allow any nested integer divisions.
3705 static enum isl_change check_coalesce_eq(int i, int j,
3706 struct isl_coalesce_info *info)
3708 isl_bool known, nested;
3709 enum isl_change change;
3711 known = isl_basic_map_divs_known(info[i].bmap);
3712 if (known < 0 || !known)
3713 return known < 0 ? isl_change_error : isl_change_none;
3714 known = isl_basic_map_divs_known(info[j].bmap);
3715 if (known < 0 || !known)
3716 return known < 0 ? isl_change_error : isl_change_none;
3717 nested = has_nested_div(info[i].bmap);
3718 if (nested < 0 || nested)
3719 return nested < 0 ? isl_change_error : isl_change_none;
3720 nested = has_nested_div(info[j].bmap);
3721 if (nested < 0 || nested)
3722 return nested < 0 ? isl_change_error : isl_change_none;
3724 change = check_coalesce_into_eq(i, j, info);
3725 if (change != isl_change_none)
3726 return change;
3727 change = check_coalesce_into_eq(j, i, info);
3728 if (change != isl_change_none)
3729 return invert_change(change);
3731 return isl_change_none;
3734 /* Check if the union of the given pair of basic maps
3735 * can be represented by a single basic map.
3736 * If so, replace the pair by the single basic map and return
3737 * isl_change_drop_first, isl_change_drop_second or isl_change_fuse.
3738 * Otherwise, return isl_change_none.
3740 * We first check if the two basic maps live in the same local space,
3741 * after aligning the divs that differ by only an integer constant.
3742 * If so, we do the complete check. Otherwise, we check if they have
3743 * the same number of integer divisions and can be coalesced, if one is
3744 * an obvious subset of the other or if the extra integer divisions
3745 * of one basic map can be simplified away using the extra equalities
3746 * of the other basic map.
3748 * Note that trying to coalesce pairs of disjuncts with the same
3749 * number, but different local variables may drop the explicit
3750 * representation of some of these local variables.
3751 * This operation is therefore not performed when
3752 * the "coalesce_preserve_locals" option is set.
3754 static enum isl_change coalesce_pair(int i, int j,
3755 struct isl_coalesce_info *info)
3757 int preserve;
3758 isl_bool same;
3759 enum isl_change change;
3760 isl_ctx *ctx;
3762 if (harmonize_divs(&info[i], &info[j]) < 0)
3763 return isl_change_error;
3764 same = same_divs(info[i].bmap, info[j].bmap);
3765 if (same < 0)
3766 return isl_change_error;
3767 if (same)
3768 return coalesce_local_pair(i, j, info);
3770 ctx = isl_basic_map_get_ctx(info[i].bmap);
3771 preserve = isl_options_get_coalesce_preserve_locals(ctx);
3772 if (!preserve && info[i].bmap->n_div == info[j].bmap->n_div) {
3773 change = coalesce_local_pair(i, j, info);
3774 if (change != isl_change_none)
3775 return change;
3778 change = coalesce_divs(i, j, info);
3779 if (change != isl_change_none)
3780 return change;
3782 return check_coalesce_eq(i, j, info);
3785 /* Return the maximum of "a" and "b".
3787 static int isl_max(int a, int b)
3789 return a > b ? a : b;
3792 /* Pairwise coalesce the basic maps in the range [start1, end1[ of "info"
3793 * with those in the range [start2, end2[, skipping basic maps
3794 * that have been removed (either before or within this function).
3796 * For each basic map i in the first range, we check if it can be coalesced
3797 * with respect to any previously considered basic map j in the second range.
3798 * If i gets dropped (because it was a subset of some j), then
3799 * we can move on to the next basic map.
3800 * If j gets dropped, we need to continue checking against the other
3801 * previously considered basic maps.
3802 * If the two basic maps got fused, then we recheck the fused basic map
3803 * against the previously considered basic maps, starting at i + 1
3804 * (even if start2 is greater than i + 1).
3806 static int coalesce_range(isl_ctx *ctx, struct isl_coalesce_info *info,
3807 int start1, int end1, int start2, int end2)
3809 int i, j;
3811 for (i = end1 - 1; i >= start1; --i) {
3812 if (info[i].removed)
3813 continue;
3814 for (j = isl_max(i + 1, start2); j < end2; ++j) {
3815 enum isl_change changed;
3817 if (info[j].removed)
3818 continue;
3819 if (info[i].removed)
3820 isl_die(ctx, isl_error_internal,
3821 "basic map unexpectedly removed",
3822 return -1);
3823 changed = coalesce_pair(i, j, info);
3824 switch (changed) {
3825 case isl_change_error:
3826 return -1;
3827 case isl_change_none:
3828 case isl_change_drop_second:
3829 continue;
3830 case isl_change_drop_first:
3831 j = end2;
3832 break;
3833 case isl_change_fuse:
3834 j = i;
3835 break;
3840 return 0;
3843 /* Pairwise coalesce the basic maps described by the "n" elements of "info".
3845 * We consider groups of basic maps that live in the same apparent
3846 * affine hull and we first coalesce within such a group before we
3847 * coalesce the elements in the group with elements of previously
3848 * considered groups. If a fuse happens during the second phase,
3849 * then we also reconsider the elements within the group.
3851 static int coalesce(isl_ctx *ctx, int n, struct isl_coalesce_info *info)
3853 int start, end;
3855 for (end = n; end > 0; end = start) {
3856 start = end - 1;
3857 while (start >= 1 &&
3858 info[start - 1].hull_hash == info[start].hull_hash)
3859 start--;
3860 if (coalesce_range(ctx, info, start, end, start, end) < 0)
3861 return -1;
3862 if (coalesce_range(ctx, info, start, end, end, n) < 0)
3863 return -1;
3866 return 0;
3869 /* Update the basic maps in "map" based on the information in "info".
3870 * In particular, remove the basic maps that have been marked removed and
3871 * update the others based on the information in the corresponding tableau.
3872 * Since we detected implicit equalities without calling
3873 * isl_basic_map_gauss, we need to do it now.
3874 * Also call isl_basic_map_simplify if we may have lost the definition
3875 * of one or more integer divisions.
3877 static __isl_give isl_map *update_basic_maps(__isl_take isl_map *map,
3878 int n, struct isl_coalesce_info *info)
3880 int i;
3882 if (!map)
3883 return NULL;
3885 for (i = n - 1; i >= 0; --i) {
3886 if (info[i].removed) {
3887 isl_basic_map_free(map->p[i]);
3888 if (i != map->n - 1)
3889 map->p[i] = map->p[map->n - 1];
3890 map->n--;
3891 continue;
3894 info[i].bmap = isl_basic_map_update_from_tab(info[i].bmap,
3895 info[i].tab);
3896 info[i].bmap = isl_basic_map_gauss(info[i].bmap, NULL);
3897 if (info[i].simplify)
3898 info[i].bmap = isl_basic_map_simplify(info[i].bmap);
3899 info[i].bmap = isl_basic_map_finalize(info[i].bmap);
3900 if (!info[i].bmap)
3901 return isl_map_free(map);
3902 ISL_F_SET(info[i].bmap, ISL_BASIC_MAP_NO_IMPLICIT);
3903 ISL_F_SET(info[i].bmap, ISL_BASIC_MAP_NO_REDUNDANT);
3904 isl_basic_map_free(map->p[i]);
3905 map->p[i] = info[i].bmap;
3906 info[i].bmap = NULL;
3909 return map;
3912 /* For each pair of basic maps in the map, check if the union of the two
3913 * can be represented by a single basic map.
3914 * If so, replace the pair by the single basic map and start over.
3916 * We factor out any (hidden) common factor from the constraint
3917 * coefficients to improve the detection of adjacent constraints.
3919 * Since we are constructing the tableaus of the basic maps anyway,
3920 * we exploit them to detect implicit equalities and redundant constraints.
3921 * This also helps the coalescing as it can ignore the redundant constraints.
3922 * In order to avoid confusion, we make all implicit equalities explicit
3923 * in the basic maps. We don't call isl_basic_map_gauss, though,
3924 * as that may affect the number of constraints.
3925 * This means that we have to call isl_basic_map_gauss at the end
3926 * of the computation (in update_basic_maps and in drop) to ensure that
3927 * the basic maps are not left in an unexpected state.
3928 * For each basic map, we also compute the hash of the apparent affine hull
3929 * for use in coalesce.
3931 __isl_give isl_map *isl_map_coalesce(__isl_take isl_map *map)
3933 int i;
3934 unsigned n;
3935 isl_ctx *ctx;
3936 struct isl_coalesce_info *info = NULL;
3938 map = isl_map_remove_empty_parts(map);
3939 if (!map)
3940 return NULL;
3942 if (map->n <= 1)
3943 return map;
3945 ctx = isl_map_get_ctx(map);
3946 map = isl_map_sort_divs(map);
3947 map = isl_map_cow(map);
3949 if (!map)
3950 return NULL;
3952 n = map->n;
3954 info = isl_calloc_array(map->ctx, struct isl_coalesce_info, n);
3955 if (!info)
3956 goto error;
3958 for (i = 0; i < map->n; ++i) {
3959 map->p[i] = isl_basic_map_reduce_coefficients(map->p[i]);
3960 if (!map->p[i])
3961 goto error;
3962 info[i].bmap = isl_basic_map_copy(map->p[i]);
3963 info[i].tab = isl_tab_from_basic_map(info[i].bmap, 0);
3964 if (!info[i].tab)
3965 goto error;
3966 if (!ISL_F_ISSET(info[i].bmap, ISL_BASIC_MAP_NO_IMPLICIT))
3967 if (isl_tab_detect_implicit_equalities(info[i].tab) < 0)
3968 goto error;
3969 info[i].bmap = isl_tab_make_equalities_explicit(info[i].tab,
3970 info[i].bmap);
3971 if (!info[i].bmap)
3972 goto error;
3973 if (!ISL_F_ISSET(info[i].bmap, ISL_BASIC_MAP_NO_REDUNDANT))
3974 if (isl_tab_detect_redundant(info[i].tab) < 0)
3975 goto error;
3976 if (coalesce_info_set_hull_hash(&info[i]) < 0)
3977 goto error;
3979 for (i = map->n - 1; i >= 0; --i)
3980 if (info[i].tab->empty)
3981 drop(&info[i]);
3983 if (coalesce(ctx, n, info) < 0)
3984 goto error;
3986 map = update_basic_maps(map, n, info);
3988 clear_coalesce_info(n, info);
3990 return map;
3991 error:
3992 clear_coalesce_info(n, info);
3993 isl_map_free(map);
3994 return NULL;
3997 /* For each pair of basic sets in the set, check if the union of the two
3998 * can be represented by a single basic set.
3999 * If so, replace the pair by the single basic set and start over.
4001 struct isl_set *isl_set_coalesce(struct isl_set *set)
4003 return set_from_map(isl_map_coalesce(set_to_map(set)));