isl_tab.c: context_tab_insert_div: use isl_bool_ok
[isl.git] / isl_coalesce.c
blobee9a27a973d42b271c269dba01de9d5fbd186ca9
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 /* Clear the memory associated to"info".
350 * Gaussian elimination needs to be performed on the basic map
351 * before it gets freed because it may have been put
352 * in an inconsistent state in isl_map_coalesce while it may
353 * be shared with other maps.
355 static void clear(struct isl_coalesce_info *info)
357 info->bmap = isl_basic_map_gauss(info->bmap, NULL);
358 info->bmap = isl_basic_map_free(info->bmap);
359 isl_tab_free(info->tab);
360 info->tab = NULL;
363 /* Drop the basic map represented by "info".
364 * That is, clear the memory associated to the entry and
365 * mark it as having been removed.
367 static void drop(struct isl_coalesce_info *info)
369 clear(info);
370 info->removed = 1;
373 /* Exchange the information in "info1" with that in "info2".
375 static void exchange(struct isl_coalesce_info *info1,
376 struct isl_coalesce_info *info2)
378 struct isl_coalesce_info info;
380 info = *info1;
381 *info1 = *info2;
382 *info2 = info;
385 /* This type represents the kind of change that has been performed
386 * while trying to coalesce two basic maps.
388 * isl_change_none: nothing was changed
389 * isl_change_drop_first: the first basic map was removed
390 * isl_change_drop_second: the second basic map was removed
391 * isl_change_fuse: the two basic maps were replaced by a new basic map.
393 enum isl_change {
394 isl_change_error = -1,
395 isl_change_none = 0,
396 isl_change_drop_first,
397 isl_change_drop_second,
398 isl_change_fuse,
401 /* Update "change" based on an interchange of the first and the second
402 * basic map. That is, interchange isl_change_drop_first and
403 * isl_change_drop_second.
405 static enum isl_change invert_change(enum isl_change change)
407 switch (change) {
408 case isl_change_error:
409 return isl_change_error;
410 case isl_change_none:
411 return isl_change_none;
412 case isl_change_drop_first:
413 return isl_change_drop_second;
414 case isl_change_drop_second:
415 return isl_change_drop_first;
416 case isl_change_fuse:
417 return isl_change_fuse;
420 return isl_change_error;
423 /* Add the valid constraints of the basic map represented by "info"
424 * to "bmap". "len" is the size of the constraints.
425 * If only one of the pair of inequalities that make up an equality
426 * is valid, then add that inequality.
428 static __isl_give isl_basic_map *add_valid_constraints(
429 __isl_take isl_basic_map *bmap, struct isl_coalesce_info *info,
430 unsigned len)
432 int k, l;
434 if (!bmap)
435 return NULL;
437 for (k = 0; k < info->bmap->n_eq; ++k) {
438 if (info->eq[2 * k] == STATUS_VALID &&
439 info->eq[2 * k + 1] == STATUS_VALID) {
440 l = isl_basic_map_alloc_equality(bmap);
441 if (l < 0)
442 return isl_basic_map_free(bmap);
443 isl_seq_cpy(bmap->eq[l], info->bmap->eq[k], len);
444 } else if (info->eq[2 * k] == STATUS_VALID) {
445 l = isl_basic_map_alloc_inequality(bmap);
446 if (l < 0)
447 return isl_basic_map_free(bmap);
448 isl_seq_neg(bmap->ineq[l], info->bmap->eq[k], len);
449 } else if (info->eq[2 * k + 1] == STATUS_VALID) {
450 l = isl_basic_map_alloc_inequality(bmap);
451 if (l < 0)
452 return isl_basic_map_free(bmap);
453 isl_seq_cpy(bmap->ineq[l], info->bmap->eq[k], len);
457 for (k = 0; k < info->bmap->n_ineq; ++k) {
458 if (info->ineq[k] != STATUS_VALID)
459 continue;
460 l = isl_basic_map_alloc_inequality(bmap);
461 if (l < 0)
462 return isl_basic_map_free(bmap);
463 isl_seq_cpy(bmap->ineq[l], info->bmap->ineq[k], len);
466 return bmap;
469 /* Is "bmap" defined by a number of (non-redundant) constraints that
470 * is greater than the number of constraints of basic maps i and j combined?
471 * Equalities are counted as two inequalities.
473 static int number_of_constraints_increases(int i, int j,
474 struct isl_coalesce_info *info,
475 __isl_keep isl_basic_map *bmap, struct isl_tab *tab)
477 int k, n_old, n_new;
479 n_old = 2 * info[i].bmap->n_eq + info[i].bmap->n_ineq;
480 n_old += 2 * info[j].bmap->n_eq + info[j].bmap->n_ineq;
482 n_new = 2 * bmap->n_eq;
483 for (k = 0; k < bmap->n_ineq; ++k)
484 if (!isl_tab_is_redundant(tab, bmap->n_eq + k))
485 ++n_new;
487 return n_new > n_old;
490 /* Replace the pair of basic maps i and j by the basic map bounded
491 * by the valid constraints in both basic maps and the constraints
492 * in extra (if not NULL).
493 * Place the fused basic map in the position that is the smallest of i and j.
495 * If "detect_equalities" is set, then look for equalities encoded
496 * as pairs of inequalities.
497 * If "check_number" is set, then the original basic maps are only
498 * replaced if the total number of constraints does not increase.
499 * While the number of integer divisions in the two basic maps
500 * is assumed to be the same, the actual definitions may be different.
501 * We only copy the definition from one of the basic map if it is
502 * the same as that of the other basic map. Otherwise, we mark
503 * the integer division as unknown and simplify the basic map
504 * in an attempt to recover the integer division definition.
506 static enum isl_change fuse(int i, int j, struct isl_coalesce_info *info,
507 __isl_keep isl_mat *extra, int detect_equalities, int check_number)
509 int k, l;
510 struct isl_basic_map *fused = NULL;
511 struct isl_tab *fused_tab = NULL;
512 isl_size total = isl_basic_map_dim(info[i].bmap, isl_dim_all);
513 unsigned extra_rows = extra ? extra->n_row : 0;
514 unsigned n_eq, n_ineq;
515 int simplify = 0;
517 if (total < 0)
518 return isl_change_error;
519 if (j < i)
520 return fuse(j, i, info, extra, detect_equalities, check_number);
522 n_eq = info[i].bmap->n_eq + info[j].bmap->n_eq;
523 n_ineq = info[i].bmap->n_ineq + info[j].bmap->n_ineq;
524 fused = isl_basic_map_alloc_space(isl_space_copy(info[i].bmap->dim),
525 info[i].bmap->n_div, n_eq, n_eq + n_ineq + extra_rows);
526 fused = add_valid_constraints(fused, &info[i], 1 + total);
527 fused = add_valid_constraints(fused, &info[j], 1 + total);
528 if (!fused)
529 goto error;
530 if (ISL_F_ISSET(info[i].bmap, ISL_BASIC_MAP_RATIONAL) &&
531 ISL_F_ISSET(info[j].bmap, ISL_BASIC_MAP_RATIONAL))
532 ISL_F_SET(fused, ISL_BASIC_MAP_RATIONAL);
534 for (k = 0; k < info[i].bmap->n_div; ++k) {
535 int l = isl_basic_map_alloc_div(fused);
536 if (l < 0)
537 goto error;
538 if (isl_seq_eq(info[i].bmap->div[k], info[j].bmap->div[k],
539 1 + 1 + total)) {
540 isl_seq_cpy(fused->div[l], info[i].bmap->div[k],
541 1 + 1 + total);
542 } else {
543 isl_int_set_si(fused->div[l][0], 0);
544 simplify = 1;
548 for (k = 0; k < extra_rows; ++k) {
549 l = isl_basic_map_alloc_inequality(fused);
550 if (l < 0)
551 goto error;
552 isl_seq_cpy(fused->ineq[l], extra->row[k], 1 + total);
555 if (detect_equalities)
556 fused = isl_basic_map_detect_inequality_pairs(fused, NULL);
557 fused = isl_basic_map_gauss(fused, NULL);
558 if (simplify || info[j].simplify) {
559 fused = isl_basic_map_simplify(fused);
560 info[i].simplify = 0;
562 fused = isl_basic_map_finalize(fused);
564 fused_tab = isl_tab_from_basic_map(fused, 0);
565 if (isl_tab_detect_redundant(fused_tab) < 0)
566 goto error;
568 if (check_number &&
569 number_of_constraints_increases(i, j, info, fused, fused_tab)) {
570 isl_tab_free(fused_tab);
571 isl_basic_map_free(fused);
572 return isl_change_none;
575 clear(&info[i]);
576 info[i].bmap = fused;
577 info[i].tab = fused_tab;
578 drop(&info[j]);
580 return isl_change_fuse;
581 error:
582 isl_tab_free(fused_tab);
583 isl_basic_map_free(fused);
584 return isl_change_error;
587 /* Given a pair of basic maps i and j such that all constraints are either
588 * "valid" or "cut", check if the facets corresponding to the "cut"
589 * constraints of i lie entirely within basic map j.
590 * If so, replace the pair by the basic map consisting of the valid
591 * constraints in both basic maps.
592 * Checking whether the facet lies entirely within basic map j
593 * is performed by checking whether the constraints of basic map j
594 * are valid for the facet. These tests are performed on a rational
595 * tableau to avoid the theoretical possibility that a constraint
596 * that was considered to be a cut constraint for the entire basic map i
597 * happens to be considered to be a valid constraint for the facet,
598 * even though it cuts off the same rational points.
600 * To see that we are not introducing any extra points, call the
601 * two basic maps A and B and the resulting map U and let x
602 * be an element of U \setminus ( A \cup B ).
603 * A line connecting x with an element of A \cup B meets a facet F
604 * of either A or B. Assume it is a facet of B and let c_1 be
605 * the corresponding facet constraint. We have c_1(x) < 0 and
606 * so c_1 is a cut constraint. This implies that there is some
607 * (possibly rational) point x' satisfying the constraints of A
608 * and the opposite of c_1 as otherwise c_1 would have been marked
609 * valid for A. The line connecting x and x' meets a facet of A
610 * in a (possibly rational) point that also violates c_1, but this
611 * is impossible since all cut constraints of B are valid for all
612 * cut facets of A.
613 * In case F is a facet of A rather than B, then we can apply the
614 * above reasoning to find a facet of B separating x from A \cup B first.
616 static enum isl_change check_facets(int i, int j,
617 struct isl_coalesce_info *info)
619 int k, l;
620 struct isl_tab_undo *snap, *snap2;
621 unsigned n_eq = info[i].bmap->n_eq;
623 snap = isl_tab_snap(info[i].tab);
624 if (isl_tab_mark_rational(info[i].tab) < 0)
625 return isl_change_error;
626 snap2 = isl_tab_snap(info[i].tab);
628 for (k = 0; k < info[i].bmap->n_ineq; ++k) {
629 if (info[i].ineq[k] != STATUS_CUT)
630 continue;
631 if (isl_tab_select_facet(info[i].tab, n_eq + k) < 0)
632 return isl_change_error;
633 for (l = 0; l < info[j].bmap->n_ineq; ++l) {
634 int stat;
635 if (info[j].ineq[l] != STATUS_CUT)
636 continue;
637 stat = status_in(info[j].bmap->ineq[l], info[i].tab);
638 if (stat < 0)
639 return isl_change_error;
640 if (stat != STATUS_VALID)
641 break;
643 if (isl_tab_rollback(info[i].tab, snap2) < 0)
644 return isl_change_error;
645 if (l < info[j].bmap->n_ineq)
646 break;
649 if (k < info[i].bmap->n_ineq) {
650 if (isl_tab_rollback(info[i].tab, snap) < 0)
651 return isl_change_error;
652 return isl_change_none;
654 return fuse(i, j, info, NULL, 0, 0);
657 /* Check if info->bmap contains the basic map represented
658 * by the tableau "tab".
659 * For each equality, we check both the constraint itself
660 * (as an inequality) and its negation. Make sure the
661 * equality is returned to its original state before returning.
663 static isl_bool contains(struct isl_coalesce_info *info, struct isl_tab *tab)
665 int k;
666 isl_size dim;
667 isl_basic_map *bmap = info->bmap;
669 dim = isl_basic_map_dim(bmap, isl_dim_all);
670 if (dim < 0)
671 return isl_bool_error;
672 for (k = 0; k < bmap->n_eq; ++k) {
673 int stat;
674 isl_seq_neg(bmap->eq[k], bmap->eq[k], 1 + dim);
675 stat = status_in(bmap->eq[k], tab);
676 isl_seq_neg(bmap->eq[k], bmap->eq[k], 1 + dim);
677 if (stat < 0)
678 return isl_bool_error;
679 if (stat != STATUS_VALID)
680 return isl_bool_false;
681 stat = status_in(bmap->eq[k], tab);
682 if (stat < 0)
683 return isl_bool_error;
684 if (stat != STATUS_VALID)
685 return isl_bool_false;
688 for (k = 0; k < bmap->n_ineq; ++k) {
689 int stat;
690 if (info->ineq[k] == STATUS_REDUNDANT)
691 continue;
692 stat = status_in(bmap->ineq[k], tab);
693 if (stat < 0)
694 return isl_bool_error;
695 if (stat != STATUS_VALID)
696 return isl_bool_false;
698 return isl_bool_true;
701 /* Basic map "i" has an inequality (say "k") that is adjacent
702 * to some inequality of basic map "j". All the other inequalities
703 * are valid for "j".
704 * Check if basic map "j" forms an extension of basic map "i".
706 * Note that this function is only called if some of the equalities or
707 * inequalities of basic map "j" do cut basic map "i". The function is
708 * correct even if there are no such cut constraints, but in that case
709 * the additional checks performed by this function are overkill.
711 * In particular, we replace constraint k, say f >= 0, by constraint
712 * f <= -1, add the inequalities of "j" that are valid for "i"
713 * and check if the result is a subset of basic map "j".
714 * To improve the chances of the subset relation being detected,
715 * any variable that only attains a single integer value
716 * in the tableau of "i" is first fixed to that value.
717 * If the result is a subset, then we know that this result is exactly equal
718 * to basic map "j" since all its constraints are valid for basic map "j".
719 * By combining the valid constraints of "i" (all equalities and all
720 * inequalities except "k") and the valid constraints of "j" we therefore
721 * obtain a basic map that is equal to their union.
722 * In this case, there is no need to perform a rollback of the tableau
723 * since it is going to be destroyed in fuse().
726 * |\__ |\__
727 * | \__ | \__
728 * | \_ => | \__
729 * |_______| _ |_________\
732 * |\ |\
733 * | \ | \
734 * | \ | \
735 * | | | \
736 * | ||\ => | \
737 * | || \ | \
738 * | || | | |
739 * |__||_/ |_____/
741 static enum isl_change is_adj_ineq_extension(int i, int j,
742 struct isl_coalesce_info *info)
744 int k;
745 struct isl_tab_undo *snap;
746 unsigned n_eq = info[i].bmap->n_eq;
747 isl_size total = isl_basic_map_dim(info[i].bmap, isl_dim_all);
748 isl_stat r;
749 isl_bool super;
751 if (total < 0)
752 return isl_change_error;
753 if (isl_tab_extend_cons(info[i].tab, 1 + info[j].bmap->n_ineq) < 0)
754 return isl_change_error;
756 k = find_ineq(&info[i], STATUS_ADJ_INEQ);
757 if (k < 0)
758 isl_die(isl_basic_map_get_ctx(info[i].bmap), isl_error_internal,
759 "info[i].ineq should have exactly one STATUS_ADJ_INEQ",
760 return isl_change_error);
762 snap = isl_tab_snap(info[i].tab);
764 if (isl_tab_unrestrict(info[i].tab, n_eq + k) < 0)
765 return isl_change_error;
767 isl_seq_neg(info[i].bmap->ineq[k], info[i].bmap->ineq[k], 1 + total);
768 isl_int_sub_ui(info[i].bmap->ineq[k][0], info[i].bmap->ineq[k][0], 1);
769 r = isl_tab_add_ineq(info[i].tab, info[i].bmap->ineq[k]);
770 isl_seq_neg(info[i].bmap->ineq[k], info[i].bmap->ineq[k], 1 + total);
771 isl_int_sub_ui(info[i].bmap->ineq[k][0], info[i].bmap->ineq[k][0], 1);
772 if (r < 0)
773 return isl_change_error;
775 for (k = 0; k < info[j].bmap->n_ineq; ++k) {
776 if (info[j].ineq[k] != STATUS_VALID)
777 continue;
778 if (isl_tab_add_ineq(info[i].tab, info[j].bmap->ineq[k]) < 0)
779 return isl_change_error;
781 if (isl_tab_detect_constants(info[i].tab) < 0)
782 return isl_change_error;
784 super = contains(&info[j], info[i].tab);
785 if (super < 0)
786 return isl_change_error;
787 if (super)
788 return fuse(i, j, info, NULL, 0, 0);
790 if (isl_tab_rollback(info[i].tab, snap) < 0)
791 return isl_change_error;
793 return isl_change_none;
797 /* Both basic maps have at least one inequality with and adjacent
798 * (but opposite) inequality in the other basic map.
799 * Check that there are no cut constraints and that there is only
800 * a single pair of adjacent inequalities.
801 * If so, we can replace the pair by a single basic map described
802 * by all but the pair of adjacent inequalities.
803 * Any additional points introduced lie strictly between the two
804 * adjacent hyperplanes and can therefore be integral.
806 * ____ _____
807 * / ||\ / \
808 * / || \ / \
809 * \ || \ => \ \
810 * \ || / \ /
811 * \___||_/ \_____/
813 * The test for a single pair of adjancent inequalities is important
814 * for avoiding the combination of two basic maps like the following
816 * /|
817 * / |
818 * /__|
819 * _____
820 * | |
821 * | |
822 * |___|
824 * If there are some cut constraints on one side, then we may
825 * still be able to fuse the two basic maps, but we need to perform
826 * some additional checks in is_adj_ineq_extension.
828 static enum isl_change check_adj_ineq(int i, int j,
829 struct isl_coalesce_info *info)
831 int count_i, count_j;
832 int cut_i, cut_j;
834 count_i = count_ineq(&info[i], STATUS_ADJ_INEQ);
835 count_j = count_ineq(&info[j], STATUS_ADJ_INEQ);
837 if (count_i != 1 && count_j != 1)
838 return isl_change_none;
840 cut_i = any_eq(&info[i], STATUS_CUT) || any_ineq(&info[i], STATUS_CUT);
841 cut_j = any_eq(&info[j], STATUS_CUT) || any_ineq(&info[j], STATUS_CUT);
843 if (!cut_i && !cut_j && count_i == 1 && count_j == 1)
844 return fuse(i, j, info, NULL, 0, 0);
846 if (count_i == 1 && !cut_i)
847 return is_adj_ineq_extension(i, j, info);
849 if (count_j == 1 && !cut_j)
850 return is_adj_ineq_extension(j, i, info);
852 return isl_change_none;
855 /* Given an affine transformation matrix "T", does row "row" represent
856 * anything other than a unit vector (possibly shifted by a constant)
857 * that is not involved in any of the other rows?
859 * That is, if a constraint involves the variable corresponding to
860 * the row, then could its preimage by "T" have any coefficients
861 * that are different from those in the original constraint?
863 static int not_unique_unit_row(__isl_keep isl_mat *T, int row)
865 int i, j;
866 int len = T->n_col - 1;
868 i = isl_seq_first_non_zero(T->row[row] + 1, len);
869 if (i < 0)
870 return 1;
871 if (!isl_int_is_one(T->row[row][1 + i]) &&
872 !isl_int_is_negone(T->row[row][1 + i]))
873 return 1;
875 j = isl_seq_first_non_zero(T->row[row] + 1 + i + 1, len - (i + 1));
876 if (j >= 0)
877 return 1;
879 for (j = 1; j < T->n_row; ++j) {
880 if (j == row)
881 continue;
882 if (!isl_int_is_zero(T->row[j][1 + i]))
883 return 1;
886 return 0;
889 /* Does inequality constraint "ineq" of "bmap" involve any of
890 * the variables marked in "affected"?
891 * "total" is the total number of variables, i.e., the number
892 * of entries in "affected".
894 static isl_bool is_affected(__isl_keep isl_basic_map *bmap, int ineq,
895 int *affected, int total)
897 int i;
899 for (i = 0; i < total; ++i) {
900 if (!affected[i])
901 continue;
902 if (!isl_int_is_zero(bmap->ineq[ineq][1 + i]))
903 return isl_bool_true;
906 return isl_bool_false;
909 /* Given the compressed version of inequality constraint "ineq"
910 * of info->bmap in "v", check if the constraint can be tightened,
911 * where the compression is based on an equality constraint valid
912 * for info->tab.
913 * If so, add the tightened version of the inequality constraint
914 * to info->tab. "v" may be modified by this function.
916 * That is, if the compressed constraint is of the form
918 * m f() + c >= 0
920 * with 0 < c < m, then it is equivalent to
922 * f() >= 0
924 * This means that c can also be subtracted from the original,
925 * uncompressed constraint without affecting the integer points
926 * in info->tab. Add this tightened constraint as an extra row
927 * to info->tab to make this information explicitly available.
929 static __isl_give isl_vec *try_tightening(struct isl_coalesce_info *info,
930 int ineq, __isl_take isl_vec *v)
932 isl_ctx *ctx;
933 isl_stat r;
935 if (!v)
936 return NULL;
938 ctx = isl_vec_get_ctx(v);
939 isl_seq_gcd(v->el + 1, v->size - 1, &ctx->normalize_gcd);
940 if (isl_int_is_zero(ctx->normalize_gcd) ||
941 isl_int_is_one(ctx->normalize_gcd)) {
942 return v;
945 v = isl_vec_cow(v);
946 if (!v)
947 return NULL;
949 isl_int_fdiv_r(v->el[0], v->el[0], ctx->normalize_gcd);
950 if (isl_int_is_zero(v->el[0]))
951 return v;
953 if (isl_tab_extend_cons(info->tab, 1) < 0)
954 return isl_vec_free(v);
956 isl_int_sub(info->bmap->ineq[ineq][0],
957 info->bmap->ineq[ineq][0], v->el[0]);
958 r = isl_tab_add_ineq(info->tab, info->bmap->ineq[ineq]);
959 isl_int_add(info->bmap->ineq[ineq][0],
960 info->bmap->ineq[ineq][0], v->el[0]);
962 if (r < 0)
963 return isl_vec_free(v);
965 return v;
968 /* Tighten the (non-redundant) constraints on the facet represented
969 * by info->tab.
970 * In particular, on input, info->tab represents the result
971 * of relaxing the "n" inequality constraints of info->bmap in "relaxed"
972 * by one, i.e., replacing f_i >= 0 by f_i + 1 >= 0, and then
973 * replacing the one at index "l" by the corresponding equality,
974 * i.e., f_k + 1 = 0, with k = relaxed[l].
976 * Compute a variable compression from the equality constraint f_k + 1 = 0
977 * and use it to tighten the other constraints of info->bmap
978 * (that is, all constraints that have not been relaxed),
979 * updating info->tab (and leaving info->bmap untouched).
980 * The compression handles essentially two cases, one where a variable
981 * is assigned a fixed value and can therefore be eliminated, and one
982 * where one variable is a shifted multiple of some other variable and
983 * can therefore be replaced by that multiple.
984 * Gaussian elimination would also work for the first case, but for
985 * the second case, the effectiveness would depend on the order
986 * of the variables.
987 * After compression, some of the constraints may have coefficients
988 * with a common divisor. If this divisor does not divide the constant
989 * term, then the constraint can be tightened.
990 * The tightening is performed on the tableau info->tab by introducing
991 * extra (temporary) constraints.
993 * Only constraints that are possibly affected by the compression are
994 * considered. In particular, if the constraint only involves variables
995 * that are directly mapped to a distinct set of other variables, then
996 * no common divisor can be introduced and no tightening can occur.
998 * It is important to only consider the non-redundant constraints
999 * since the facet constraint has been relaxed prior to the call
1000 * to this function, meaning that the constraints that were redundant
1001 * prior to the relaxation may no longer be redundant.
1002 * These constraints will be ignored in the fused result, so
1003 * the fusion detection should not exploit them.
1005 static isl_stat tighten_on_relaxed_facet(struct isl_coalesce_info *info,
1006 int n, int *relaxed, int l)
1008 isl_size total;
1009 isl_ctx *ctx;
1010 isl_vec *v = NULL;
1011 isl_mat *T;
1012 int i;
1013 int k;
1014 int *affected;
1016 k = relaxed[l];
1017 ctx = isl_basic_map_get_ctx(info->bmap);
1018 total = isl_basic_map_dim(info->bmap, isl_dim_all);
1019 if (total < 0)
1020 return isl_stat_error;
1021 isl_int_add_ui(info->bmap->ineq[k][0], info->bmap->ineq[k][0], 1);
1022 T = isl_mat_sub_alloc6(ctx, info->bmap->ineq, k, 1, 0, 1 + total);
1023 T = isl_mat_variable_compression(T, NULL);
1024 isl_int_sub_ui(info->bmap->ineq[k][0], info->bmap->ineq[k][0], 1);
1025 if (!T)
1026 return isl_stat_error;
1027 if (T->n_col == 0) {
1028 isl_mat_free(T);
1029 return isl_stat_ok;
1032 affected = isl_alloc_array(ctx, int, total);
1033 if (!affected)
1034 goto error;
1036 for (i = 0; i < total; ++i)
1037 affected[i] = not_unique_unit_row(T, 1 + i);
1039 for (i = 0; i < info->bmap->n_ineq; ++i) {
1040 isl_bool handle;
1041 if (any(relaxed, n, i))
1042 continue;
1043 if (info->ineq[i] == STATUS_REDUNDANT)
1044 continue;
1045 handle = is_affected(info->bmap, i, affected, total);
1046 if (handle < 0)
1047 goto error;
1048 if (!handle)
1049 continue;
1050 v = isl_vec_alloc(ctx, 1 + total);
1051 if (!v)
1052 goto error;
1053 isl_seq_cpy(v->el, info->bmap->ineq[i], 1 + total);
1054 v = isl_vec_mat_product(v, isl_mat_copy(T));
1055 v = try_tightening(info, i, v);
1056 isl_vec_free(v);
1057 if (!v)
1058 goto error;
1061 isl_mat_free(T);
1062 free(affected);
1063 return isl_stat_ok;
1064 error:
1065 isl_mat_free(T);
1066 free(affected);
1067 return isl_stat_error;
1070 /* Replace the basic maps "i" and "j" by an extension of "i"
1071 * along the "n" inequality constraints in "relax" by one.
1072 * The tableau info[i].tab has already been extended.
1073 * Extend info[i].bmap accordingly by relaxing all constraints in "relax"
1074 * by one.
1075 * Each integer division that does not have exactly the same
1076 * definition in "i" and "j" is marked unknown and the basic map
1077 * is scheduled to be simplified in an attempt to recover
1078 * the integer division definition.
1079 * Place the extension in the position that is the smallest of i and j.
1081 static enum isl_change extend(int i, int j, int n, int *relax,
1082 struct isl_coalesce_info *info)
1084 int l;
1085 isl_size total;
1087 info[i].bmap = isl_basic_map_cow(info[i].bmap);
1088 total = isl_basic_map_dim(info[i].bmap, isl_dim_all);
1089 if (total < 0)
1090 return isl_change_error;
1091 for (l = 0; l < info[i].bmap->n_div; ++l)
1092 if (!isl_seq_eq(info[i].bmap->div[l],
1093 info[j].bmap->div[l], 1 + 1 + total)) {
1094 isl_int_set_si(info[i].bmap->div[l][0], 0);
1095 info[i].simplify = 1;
1097 for (l = 0; l < n; ++l)
1098 isl_int_add_ui(info[i].bmap->ineq[relax[l]][0],
1099 info[i].bmap->ineq[relax[l]][0], 1);
1100 ISL_F_SET(info[i].bmap, ISL_BASIC_MAP_FINAL);
1101 drop(&info[j]);
1102 if (j < i)
1103 exchange(&info[i], &info[j]);
1104 return isl_change_fuse;
1107 /* Basic map "i" has "n" inequality constraints (collected in "relax")
1108 * that are such that they include basic map "j" if they are relaxed
1109 * by one. All the other inequalities are valid for "j".
1110 * Check if basic map "j" forms an extension of basic map "i".
1112 * In particular, relax the constraints in "relax", compute the corresponding
1113 * facets one by one and check whether each of these is included
1114 * in the other basic map.
1115 * Before testing for inclusion, the constraints on each facet
1116 * are tightened to increase the chance of an inclusion being detected.
1117 * (Adding the valid constraints of "j" to the tableau of "i", as is done
1118 * in is_adj_ineq_extension, may further increase those chances, but this
1119 * is not currently done.)
1120 * If each facet is included, we know that relaxing the constraints extends
1121 * the basic map with exactly the other basic map (we already know that this
1122 * other basic map is included in the extension, because all other
1123 * inequality constraints are valid of "j") and we can replace the
1124 * two basic maps by this extension.
1126 * If any of the relaxed constraints turn out to be redundant, then bail out.
1127 * isl_tab_select_facet refuses to handle such constraints. It may be
1128 * possible to handle them anyway by making a distinction between
1129 * redundant constraints with a corresponding facet that still intersects
1130 * the set (allowing isl_tab_select_facet to handle them) and
1131 * those where the facet does not intersect the set (which can be ignored
1132 * because the empty facet is trivially included in the other disjunct).
1133 * However, relaxed constraints that turn out to be redundant should
1134 * be fairly rare and no such instance has been reported where
1135 * coalescing would be successful.
1136 * ____ _____
1137 * / || / |
1138 * / || / |
1139 * \ || => \ |
1140 * \ || \ |
1141 * \___|| \____|
1144 * \ |\
1145 * |\\ | \
1146 * | \\ | \
1147 * | | => | /
1148 * | / | /
1149 * |/ |/
1151 static enum isl_change is_relaxed_extension(int i, int j, int n, int *relax,
1152 struct isl_coalesce_info *info)
1154 int l;
1155 isl_bool super;
1156 struct isl_tab_undo *snap, *snap2;
1157 unsigned n_eq = info[i].bmap->n_eq;
1159 for (l = 0; l < n; ++l)
1160 if (isl_tab_is_equality(info[i].tab, n_eq + relax[l]))
1161 return isl_change_none;
1163 snap = isl_tab_snap(info[i].tab);
1164 for (l = 0; l < n; ++l)
1165 if (isl_tab_relax(info[i].tab, n_eq + relax[l]) < 0)
1166 return isl_change_error;
1167 for (l = 0; l < n; ++l) {
1168 if (!isl_tab_is_redundant(info[i].tab, n_eq + relax[l]))
1169 continue;
1170 if (isl_tab_rollback(info[i].tab, snap) < 0)
1171 return isl_change_error;
1172 return isl_change_none;
1174 snap2 = isl_tab_snap(info[i].tab);
1175 for (l = 0; l < n; ++l) {
1176 if (isl_tab_rollback(info[i].tab, snap2) < 0)
1177 return isl_change_error;
1178 if (isl_tab_select_facet(info[i].tab, n_eq + relax[l]) < 0)
1179 return isl_change_error;
1180 if (tighten_on_relaxed_facet(&info[i], n, relax, l) < 0)
1181 return isl_change_error;
1182 super = contains(&info[j], info[i].tab);
1183 if (super < 0)
1184 return isl_change_error;
1185 if (super)
1186 continue;
1187 if (isl_tab_rollback(info[i].tab, snap) < 0)
1188 return isl_change_error;
1189 return isl_change_none;
1192 if (isl_tab_rollback(info[i].tab, snap2) < 0)
1193 return isl_change_error;
1194 return extend(i, j, n, relax, info);
1197 /* Data structure that keeps track of the wrapping constraints
1198 * and of information to bound the coefficients of those constraints.
1200 * bound is set if we want to apply a bound on the coefficients
1201 * mat contains the wrapping constraints
1202 * max is the bound on the coefficients (if bound is set)
1204 struct isl_wraps {
1205 int bound;
1206 isl_mat *mat;
1207 isl_int max;
1210 /* Update wraps->max to be greater than or equal to the coefficients
1211 * in the equalities and inequalities of info->bmap that can be removed
1212 * if we end up applying wrapping.
1214 static isl_stat wraps_update_max(struct isl_wraps *wraps,
1215 struct isl_coalesce_info *info)
1217 int k;
1218 isl_int max_k;
1219 isl_size total = isl_basic_map_dim(info->bmap, isl_dim_all);
1221 if (total < 0)
1222 return isl_stat_error;
1223 isl_int_init(max_k);
1225 for (k = 0; k < info->bmap->n_eq; ++k) {
1226 if (info->eq[2 * k] == STATUS_VALID &&
1227 info->eq[2 * k + 1] == STATUS_VALID)
1228 continue;
1229 isl_seq_abs_max(info->bmap->eq[k] + 1, total, &max_k);
1230 if (isl_int_abs_gt(max_k, wraps->max))
1231 isl_int_set(wraps->max, max_k);
1234 for (k = 0; k < info->bmap->n_ineq; ++k) {
1235 if (info->ineq[k] == STATUS_VALID ||
1236 info->ineq[k] == STATUS_REDUNDANT)
1237 continue;
1238 isl_seq_abs_max(info->bmap->ineq[k] + 1, total, &max_k);
1239 if (isl_int_abs_gt(max_k, wraps->max))
1240 isl_int_set(wraps->max, max_k);
1243 isl_int_clear(max_k);
1245 return isl_stat_ok;
1248 /* Initialize the isl_wraps data structure.
1249 * If we want to bound the coefficients of the wrapping constraints,
1250 * we set wraps->max to the largest coefficient
1251 * in the equalities and inequalities that can be removed if we end up
1252 * applying wrapping.
1254 static isl_stat wraps_init(struct isl_wraps *wraps, __isl_take isl_mat *mat,
1255 struct isl_coalesce_info *info, int i, int j)
1257 isl_ctx *ctx;
1259 wraps->bound = 0;
1260 wraps->mat = mat;
1261 if (!mat)
1262 return isl_stat_error;
1263 ctx = isl_mat_get_ctx(mat);
1264 wraps->bound = isl_options_get_coalesce_bounded_wrapping(ctx);
1265 if (!wraps->bound)
1266 return isl_stat_ok;
1267 isl_int_init(wraps->max);
1268 isl_int_set_si(wraps->max, 0);
1269 if (wraps_update_max(wraps, &info[i]) < 0)
1270 return isl_stat_error;
1271 if (wraps_update_max(wraps, &info[j]) < 0)
1272 return isl_stat_error;
1274 return isl_stat_ok;
1277 /* Free the contents of the isl_wraps data structure.
1279 static void wraps_free(struct isl_wraps *wraps)
1281 isl_mat_free(wraps->mat);
1282 if (wraps->bound)
1283 isl_int_clear(wraps->max);
1286 /* Is the wrapping constraint in row "row" allowed?
1288 * If wraps->bound is set, we check that none of the coefficients
1289 * is greater than wraps->max.
1291 static int allow_wrap(struct isl_wraps *wraps, int row)
1293 int i;
1295 if (!wraps->bound)
1296 return 1;
1298 for (i = 1; i < wraps->mat->n_col; ++i)
1299 if (isl_int_abs_gt(wraps->mat->row[row][i], wraps->max))
1300 return 0;
1302 return 1;
1305 /* Wrap "ineq" (or its opposite if "negate" is set) around "bound"
1306 * to include "set" and add the result in position "w" of "wraps".
1307 * "len" is the total number of coefficients in "bound" and "ineq".
1308 * Return 1 on success, 0 on failure and -1 on error.
1309 * Wrapping can fail if the result of wrapping is equal to "bound"
1310 * or if we want to bound the sizes of the coefficients and
1311 * the wrapped constraint does not satisfy this bound.
1313 static int add_wrap(struct isl_wraps *wraps, int w, isl_int *bound,
1314 isl_int *ineq, unsigned len, __isl_keep isl_set *set, int negate)
1316 isl_seq_cpy(wraps->mat->row[w], bound, len);
1317 if (negate) {
1318 isl_seq_neg(wraps->mat->row[w + 1], ineq, len);
1319 ineq = wraps->mat->row[w + 1];
1321 if (!isl_set_wrap_facet(set, wraps->mat->row[w], ineq))
1322 return -1;
1323 if (isl_seq_eq(wraps->mat->row[w], bound, len))
1324 return 0;
1325 if (!allow_wrap(wraps, w))
1326 return 0;
1327 return 1;
1330 /* For each constraint in info->bmap that is not redundant (as determined
1331 * by info->tab) and that is not a valid constraint for the other basic map,
1332 * wrap the constraint around "bound" such that it includes the whole
1333 * set "set" and append the resulting constraint to "wraps".
1334 * Note that the constraints that are valid for the other basic map
1335 * will be added to the combined basic map by default, so there is
1336 * no need to wrap them.
1337 * The caller wrap_in_facets even relies on this function not wrapping
1338 * any constraints that are already valid.
1339 * "wraps" is assumed to have been pre-allocated to the appropriate size.
1340 * wraps->n_row is the number of actual wrapped constraints that have
1341 * been added.
1342 * If any of the wrapping problems results in a constraint that is
1343 * identical to "bound", then this means that "set" is unbounded in such
1344 * way that no wrapping is possible. If this happens then wraps->n_row
1345 * is reset to zero.
1346 * Similarly, if we want to bound the coefficients of the wrapping
1347 * constraints and a newly added wrapping constraint does not
1348 * satisfy the bound, then wraps->n_row is also reset to zero.
1350 static isl_stat add_wraps(struct isl_wraps *wraps,
1351 struct isl_coalesce_info *info, isl_int *bound, __isl_keep isl_set *set)
1353 int l, m;
1354 int w;
1355 int added;
1356 isl_basic_map *bmap = info->bmap;
1357 isl_size total = isl_basic_map_dim(bmap, isl_dim_all);
1358 unsigned len = 1 + total;
1360 if (total < 0)
1361 return isl_stat_error;
1363 w = wraps->mat->n_row;
1365 for (l = 0; l < bmap->n_ineq; ++l) {
1366 if (info->ineq[l] == STATUS_VALID ||
1367 info->ineq[l] == STATUS_REDUNDANT)
1368 continue;
1369 if (isl_seq_is_neg(bound, bmap->ineq[l], len))
1370 continue;
1371 if (isl_seq_eq(bound, bmap->ineq[l], len))
1372 continue;
1373 if (isl_tab_is_redundant(info->tab, bmap->n_eq + l))
1374 continue;
1376 added = add_wrap(wraps, w, bound, bmap->ineq[l], len, set, 0);
1377 if (added < 0)
1378 return isl_stat_error;
1379 if (!added)
1380 goto unbounded;
1381 ++w;
1383 for (l = 0; l < bmap->n_eq; ++l) {
1384 if (isl_seq_is_neg(bound, bmap->eq[l], len))
1385 continue;
1386 if (isl_seq_eq(bound, bmap->eq[l], len))
1387 continue;
1389 for (m = 0; m < 2; ++m) {
1390 if (info->eq[2 * l + m] == STATUS_VALID)
1391 continue;
1392 added = add_wrap(wraps, w, bound, bmap->eq[l], len,
1393 set, !m);
1394 if (added < 0)
1395 return isl_stat_error;
1396 if (!added)
1397 goto unbounded;
1398 ++w;
1402 wraps->mat->n_row = w;
1403 return isl_stat_ok;
1404 unbounded:
1405 wraps->mat->n_row = 0;
1406 return isl_stat_ok;
1409 /* Check if the constraints in "wraps" from "first" until the last
1410 * are all valid for the basic set represented by "tab".
1411 * If not, wraps->n_row is set to zero.
1413 static int check_wraps(__isl_keep isl_mat *wraps, int first,
1414 struct isl_tab *tab)
1416 int i;
1418 for (i = first; i < wraps->n_row; ++i) {
1419 enum isl_ineq_type type;
1420 type = isl_tab_ineq_type(tab, wraps->row[i]);
1421 if (type == isl_ineq_error)
1422 return -1;
1423 if (type == isl_ineq_redundant)
1424 continue;
1425 wraps->n_row = 0;
1426 return 0;
1429 return 0;
1432 /* Return a set that corresponds to the non-redundant constraints
1433 * (as recorded in tab) of bmap.
1435 * It's important to remove the redundant constraints as some
1436 * of the other constraints may have been modified after the
1437 * constraints were marked redundant.
1438 * In particular, a constraint may have been relaxed.
1439 * Redundant constraints are ignored when a constraint is relaxed
1440 * and should therefore continue to be ignored ever after.
1441 * Otherwise, the relaxation might be thwarted by some of
1442 * these constraints.
1444 * Update the underlying set to ensure that the dimension doesn't change.
1445 * Otherwise the integer divisions could get dropped if the tab
1446 * turns out to be empty.
1448 static __isl_give isl_set *set_from_updated_bmap(__isl_keep isl_basic_map *bmap,
1449 struct isl_tab *tab)
1451 isl_basic_set *bset;
1453 bmap = isl_basic_map_copy(bmap);
1454 bset = isl_basic_map_underlying_set(bmap);
1455 bset = isl_basic_set_cow(bset);
1456 bset = isl_basic_set_update_from_tab(bset, tab);
1457 return isl_set_from_basic_set(bset);
1460 /* Wrap the constraints of info->bmap that bound the facet defined
1461 * by inequality "k" around (the opposite of) this inequality to
1462 * include "set". "bound" may be used to store the negated inequality.
1463 * Since the wrapped constraints are not guaranteed to contain the whole
1464 * of info->bmap, we check them in check_wraps.
1465 * If any of the wrapped constraints turn out to be invalid, then
1466 * check_wraps will reset wrap->n_row to zero.
1468 static isl_stat add_wraps_around_facet(struct isl_wraps *wraps,
1469 struct isl_coalesce_info *info, int k, isl_int *bound,
1470 __isl_keep isl_set *set)
1472 struct isl_tab_undo *snap;
1473 int n;
1474 isl_size total = isl_basic_map_dim(info->bmap, isl_dim_all);
1476 if (total < 0)
1477 return isl_stat_error;
1479 snap = isl_tab_snap(info->tab);
1481 if (isl_tab_select_facet(info->tab, info->bmap->n_eq + k) < 0)
1482 return isl_stat_error;
1483 if (isl_tab_detect_redundant(info->tab) < 0)
1484 return isl_stat_error;
1486 isl_seq_neg(bound, info->bmap->ineq[k], 1 + total);
1488 n = wraps->mat->n_row;
1489 if (add_wraps(wraps, info, bound, set) < 0)
1490 return isl_stat_error;
1492 if (isl_tab_rollback(info->tab, snap) < 0)
1493 return isl_stat_error;
1494 if (check_wraps(wraps->mat, n, info->tab) < 0)
1495 return isl_stat_error;
1497 return isl_stat_ok;
1500 /* Given a basic set i with a constraint k that is adjacent to
1501 * basic set j, check if we can wrap
1502 * both the facet corresponding to k (if "wrap_facet" is set) and basic map j
1503 * (always) around their ridges to include the other set.
1504 * If so, replace the pair of basic sets by their union.
1506 * All constraints of i (except k) are assumed to be valid or
1507 * cut constraints for j.
1508 * Wrapping the cut constraints to include basic map j may result
1509 * in constraints that are no longer valid of basic map i
1510 * we have to check that the resulting wrapping constraints are valid for i.
1511 * If "wrap_facet" is not set, then all constraints of i (except k)
1512 * are assumed to be valid for j.
1513 * ____ _____
1514 * / | / \
1515 * / || / |
1516 * \ || => \ |
1517 * \ || \ |
1518 * \___|| \____|
1521 static enum isl_change can_wrap_in_facet(int i, int j, int k,
1522 struct isl_coalesce_info *info, int wrap_facet)
1524 enum isl_change change = isl_change_none;
1525 struct isl_wraps wraps;
1526 isl_ctx *ctx;
1527 isl_mat *mat;
1528 struct isl_set *set_i = NULL;
1529 struct isl_set *set_j = NULL;
1530 struct isl_vec *bound = NULL;
1531 isl_size total = isl_basic_map_dim(info[i].bmap, isl_dim_all);
1533 if (total < 0)
1534 return isl_change_error;
1535 set_i = set_from_updated_bmap(info[i].bmap, info[i].tab);
1536 set_j = set_from_updated_bmap(info[j].bmap, info[j].tab);
1537 ctx = isl_basic_map_get_ctx(info[i].bmap);
1538 mat = isl_mat_alloc(ctx, 2 * (info[i].bmap->n_eq + info[j].bmap->n_eq) +
1539 info[i].bmap->n_ineq + info[j].bmap->n_ineq,
1540 1 + total);
1541 if (wraps_init(&wraps, mat, info, i, j) < 0)
1542 goto error;
1543 bound = isl_vec_alloc(ctx, 1 + total);
1544 if (!set_i || !set_j || !bound)
1545 goto error;
1547 isl_seq_cpy(bound->el, info[i].bmap->ineq[k], 1 + total);
1548 isl_int_add_ui(bound->el[0], bound->el[0], 1);
1549 isl_seq_normalize(ctx, bound->el, 1 + total);
1551 isl_seq_cpy(wraps.mat->row[0], bound->el, 1 + total);
1552 wraps.mat->n_row = 1;
1554 if (add_wraps(&wraps, &info[j], bound->el, set_i) < 0)
1555 goto error;
1556 if (!wraps.mat->n_row)
1557 goto unbounded;
1559 if (wrap_facet) {
1560 if (add_wraps_around_facet(&wraps, &info[i], k,
1561 bound->el, set_j) < 0)
1562 goto error;
1563 if (!wraps.mat->n_row)
1564 goto unbounded;
1567 change = fuse(i, j, info, wraps.mat, 0, 0);
1569 unbounded:
1570 wraps_free(&wraps);
1572 isl_set_free(set_i);
1573 isl_set_free(set_j);
1575 isl_vec_free(bound);
1577 return change;
1578 error:
1579 wraps_free(&wraps);
1580 isl_vec_free(bound);
1581 isl_set_free(set_i);
1582 isl_set_free(set_j);
1583 return isl_change_error;
1586 /* Given a cut constraint t(x) >= 0 of basic map i, stored in row "w"
1587 * of wrap.mat, replace it by its relaxed version t(x) + 1 >= 0, and
1588 * add wrapping constraints to wrap.mat for all constraints
1589 * of basic map j that bound the part of basic map j that sticks out
1590 * of the cut constraint.
1591 * "set_i" is the underlying set of basic map i.
1592 * If any wrapping fails, then wraps->mat.n_row is reset to zero.
1594 * In particular, we first intersect basic map j with t(x) + 1 = 0.
1595 * If the result is empty, then t(x) >= 0 was actually a valid constraint
1596 * (with respect to the integer points), so we add t(x) >= 0 instead.
1597 * Otherwise, we wrap the constraints of basic map j that are not
1598 * redundant in this intersection and that are not already valid
1599 * for basic map i over basic map i.
1600 * Note that it is sufficient to wrap the constraints to include
1601 * basic map i, because we will only wrap the constraints that do
1602 * not include basic map i already. The wrapped constraint will
1603 * therefore be more relaxed compared to the original constraint.
1604 * Since the original constraint is valid for basic map j, so is
1605 * the wrapped constraint.
1607 static isl_stat wrap_in_facet(struct isl_wraps *wraps, int w,
1608 struct isl_coalesce_info *info_j, __isl_keep isl_set *set_i,
1609 struct isl_tab_undo *snap)
1611 isl_int_add_ui(wraps->mat->row[w][0], wraps->mat->row[w][0], 1);
1612 if (isl_tab_add_eq(info_j->tab, wraps->mat->row[w]) < 0)
1613 return isl_stat_error;
1614 if (isl_tab_detect_redundant(info_j->tab) < 0)
1615 return isl_stat_error;
1617 if (info_j->tab->empty)
1618 isl_int_sub_ui(wraps->mat->row[w][0], wraps->mat->row[w][0], 1);
1619 else if (add_wraps(wraps, info_j, wraps->mat->row[w], set_i) < 0)
1620 return isl_stat_error;
1622 if (isl_tab_rollback(info_j->tab, snap) < 0)
1623 return isl_stat_error;
1625 return isl_stat_ok;
1628 /* Given a pair of basic maps i and j such that j sticks out
1629 * of i at n cut constraints, each time by at most one,
1630 * try to compute wrapping constraints and replace the two
1631 * basic maps by a single basic map.
1632 * The other constraints of i are assumed to be valid for j.
1633 * "set_i" is the underlying set of basic map i.
1634 * "wraps" has been initialized to be of the right size.
1636 * For each cut constraint t(x) >= 0 of i, we add the relaxed version
1637 * t(x) + 1 >= 0, along with wrapping constraints for all constraints
1638 * of basic map j that bound the part of basic map j that sticks out
1639 * of the cut constraint.
1641 * If any wrapping fails, i.e., if we cannot wrap to touch
1642 * the union, then we give up.
1643 * Otherwise, the pair of basic maps is replaced by their union.
1645 static enum isl_change try_wrap_in_facets(int i, int j,
1646 struct isl_coalesce_info *info, struct isl_wraps *wraps,
1647 __isl_keep isl_set *set_i)
1649 int k, l, w;
1650 isl_size total;
1651 struct isl_tab_undo *snap;
1653 total = isl_basic_map_dim(info[i].bmap, isl_dim_all);
1654 if (total < 0)
1655 return isl_change_error;
1657 snap = isl_tab_snap(info[j].tab);
1659 wraps->mat->n_row = 0;
1661 for (k = 0; k < info[i].bmap->n_eq; ++k) {
1662 for (l = 0; l < 2; ++l) {
1663 if (info[i].eq[2 * k + l] != STATUS_CUT)
1664 continue;
1665 w = wraps->mat->n_row++;
1666 if (l == 0)
1667 isl_seq_neg(wraps->mat->row[w],
1668 info[i].bmap->eq[k], 1 + total);
1669 else
1670 isl_seq_cpy(wraps->mat->row[w],
1671 info[i].bmap->eq[k], 1 + total);
1672 if (wrap_in_facet(wraps, w, &info[j], set_i, snap) < 0)
1673 return isl_change_error;
1675 if (!wraps->mat->n_row)
1676 return isl_change_none;
1680 for (k = 0; k < info[i].bmap->n_ineq; ++k) {
1681 if (info[i].ineq[k] != STATUS_CUT)
1682 continue;
1683 w = wraps->mat->n_row++;
1684 isl_seq_cpy(wraps->mat->row[w],
1685 info[i].bmap->ineq[k], 1 + total);
1686 if (wrap_in_facet(wraps, w, &info[j], set_i, snap) < 0)
1687 return isl_change_error;
1689 if (!wraps->mat->n_row)
1690 return isl_change_none;
1693 return fuse(i, j, info, wraps->mat, 0, 1);
1696 /* Given a pair of basic maps i and j such that j sticks out
1697 * of i at n cut constraints, each time by at most one,
1698 * try to compute wrapping constraints and replace the two
1699 * basic maps by a single basic map.
1700 * The other constraints of i are assumed to be valid for j.
1702 * The core computation is performed by try_wrap_in_facets.
1703 * This function simply extracts an underlying set representation
1704 * of basic map i and initializes the data structure for keeping
1705 * track of wrapping constraints.
1707 static enum isl_change wrap_in_facets(int i, int j, int n,
1708 struct isl_coalesce_info *info)
1710 enum isl_change change = isl_change_none;
1711 struct isl_wraps wraps;
1712 isl_ctx *ctx;
1713 isl_mat *mat;
1714 isl_set *set_i = NULL;
1715 isl_size total = isl_basic_map_dim(info[i].bmap, isl_dim_all);
1716 int max_wrap;
1718 if (total < 0)
1719 return isl_change_error;
1720 if (isl_tab_extend_cons(info[j].tab, 1) < 0)
1721 return isl_change_error;
1723 max_wrap = 1 + 2 * info[j].bmap->n_eq + info[j].bmap->n_ineq;
1724 max_wrap *= n;
1726 set_i = set_from_updated_bmap(info[i].bmap, info[i].tab);
1727 ctx = isl_basic_map_get_ctx(info[i].bmap);
1728 mat = isl_mat_alloc(ctx, max_wrap, 1 + total);
1729 if (wraps_init(&wraps, mat, info, i, j) < 0)
1730 goto error;
1731 if (!set_i)
1732 goto error;
1734 change = try_wrap_in_facets(i, j, info, &wraps, set_i);
1736 wraps_free(&wraps);
1737 isl_set_free(set_i);
1739 return change;
1740 error:
1741 wraps_free(&wraps);
1742 isl_set_free(set_i);
1743 return isl_change_error;
1746 /* Return the effect of inequality "ineq" on the tableau "tab",
1747 * after relaxing the constant term of "ineq" by one.
1749 static enum isl_ineq_type type_of_relaxed(struct isl_tab *tab, isl_int *ineq)
1751 enum isl_ineq_type type;
1753 isl_int_add_ui(ineq[0], ineq[0], 1);
1754 type = isl_tab_ineq_type(tab, ineq);
1755 isl_int_sub_ui(ineq[0], ineq[0], 1);
1757 return type;
1760 /* Given two basic sets i and j,
1761 * check if relaxing all the cut constraints of i by one turns
1762 * them into valid constraint for j and check if we can wrap in
1763 * the bits that are sticking out.
1764 * If so, replace the pair by their union.
1766 * We first check if all relaxed cut inequalities of i are valid for j
1767 * and then try to wrap in the intersections of the relaxed cut inequalities
1768 * with j.
1770 * During this wrapping, we consider the points of j that lie at a distance
1771 * of exactly 1 from i. In particular, we ignore the points that lie in
1772 * between this lower-dimensional space and the basic map i.
1773 * We can therefore only apply this to integer maps.
1774 * ____ _____
1775 * / ___|_ / \
1776 * / | | / |
1777 * \ | | => \ |
1778 * \|____| \ |
1779 * \___| \____/
1781 * _____ ______
1782 * | ____|_ | \
1783 * | | | | |
1784 * | | | => | |
1785 * |_| | | |
1786 * |_____| \______|
1788 * _______
1789 * | |
1790 * | |\ |
1791 * | | \ |
1792 * | | \ |
1793 * | | \|
1794 * | | \
1795 * | |_____\
1796 * | |
1797 * |_______|
1799 * Wrapping can fail if the result of wrapping one of the facets
1800 * around its edges does not produce any new facet constraint.
1801 * In particular, this happens when we try to wrap in unbounded sets.
1803 * _______________________________________________________________________
1805 * | ___
1806 * | | |
1807 * |_| |_________________________________________________________________
1808 * |___|
1810 * The following is not an acceptable result of coalescing the above two
1811 * sets as it includes extra integer points.
1812 * _______________________________________________________________________
1814 * |
1815 * |
1817 * \______________________________________________________________________
1819 static enum isl_change can_wrap_in_set(int i, int j,
1820 struct isl_coalesce_info *info)
1822 int k, l;
1823 int n;
1824 isl_size total;
1826 if (ISL_F_ISSET(info[i].bmap, ISL_BASIC_MAP_RATIONAL) ||
1827 ISL_F_ISSET(info[j].bmap, ISL_BASIC_MAP_RATIONAL))
1828 return isl_change_none;
1830 n = count_eq(&info[i], STATUS_CUT) + count_ineq(&info[i], STATUS_CUT);
1831 if (n == 0)
1832 return isl_change_none;
1834 total = isl_basic_map_dim(info[i].bmap, isl_dim_all);
1835 if (total < 0)
1836 return isl_change_error;
1837 for (k = 0; k < info[i].bmap->n_eq; ++k) {
1838 for (l = 0; l < 2; ++l) {
1839 enum isl_ineq_type type;
1841 if (info[i].eq[2 * k + l] != STATUS_CUT)
1842 continue;
1844 if (l == 0)
1845 isl_seq_neg(info[i].bmap->eq[k],
1846 info[i].bmap->eq[k], 1 + total);
1847 type = type_of_relaxed(info[j].tab,
1848 info[i].bmap->eq[k]);
1849 if (l == 0)
1850 isl_seq_neg(info[i].bmap->eq[k],
1851 info[i].bmap->eq[k], 1 + total);
1852 if (type == isl_ineq_error)
1853 return isl_change_error;
1854 if (type != isl_ineq_redundant)
1855 return isl_change_none;
1859 for (k = 0; k < info[i].bmap->n_ineq; ++k) {
1860 enum isl_ineq_type type;
1862 if (info[i].ineq[k] != STATUS_CUT)
1863 continue;
1865 type = type_of_relaxed(info[j].tab, info[i].bmap->ineq[k]);
1866 if (type == isl_ineq_error)
1867 return isl_change_error;
1868 if (type != isl_ineq_redundant)
1869 return isl_change_none;
1872 return wrap_in_facets(i, j, n, info);
1875 /* Check if either i or j has only cut constraints that can
1876 * be used to wrap in (a facet of) the other basic set.
1877 * if so, replace the pair by their union.
1879 static enum isl_change check_wrap(int i, int j, struct isl_coalesce_info *info)
1881 enum isl_change change = isl_change_none;
1883 change = can_wrap_in_set(i, j, info);
1884 if (change != isl_change_none)
1885 return change;
1887 change = can_wrap_in_set(j, i, info);
1888 return change;
1891 /* Check if all inequality constraints of "i" that cut "j" cease
1892 * to be cut constraints if they are relaxed by one.
1893 * If so, collect the cut constraints in "list".
1894 * The caller is responsible for allocating "list".
1896 static isl_bool all_cut_by_one(int i, int j, struct isl_coalesce_info *info,
1897 int *list)
1899 int l, n;
1901 n = 0;
1902 for (l = 0; l < info[i].bmap->n_ineq; ++l) {
1903 enum isl_ineq_type type;
1905 if (info[i].ineq[l] != STATUS_CUT)
1906 continue;
1907 type = type_of_relaxed(info[j].tab, info[i].bmap->ineq[l]);
1908 if (type == isl_ineq_error)
1909 return isl_bool_error;
1910 if (type != isl_ineq_redundant)
1911 return isl_bool_false;
1912 list[n++] = l;
1915 return isl_bool_true;
1918 /* Given two basic maps such that "j" has at least one equality constraint
1919 * that is adjacent to an inequality constraint of "i" and such that "i" has
1920 * exactly one inequality constraint that is adjacent to an equality
1921 * constraint of "j", check whether "i" can be extended to include "j" or
1922 * whether "j" can be wrapped into "i".
1923 * All remaining constraints of "i" and "j" are assumed to be valid
1924 * or cut constraints of the other basic map.
1925 * However, none of the equality constraints of "i" are cut constraints.
1927 * If "i" has any "cut" inequality constraints, then check if relaxing
1928 * each of them by one is sufficient for them to become valid.
1929 * If so, check if the inequality constraint adjacent to an equality
1930 * constraint of "j" along with all these cut constraints
1931 * can be relaxed by one to contain exactly "j".
1932 * Otherwise, or if this fails, check if "j" can be wrapped into "i".
1934 static enum isl_change check_single_adj_eq(int i, int j,
1935 struct isl_coalesce_info *info)
1937 enum isl_change change = isl_change_none;
1938 int k;
1939 int n_cut;
1940 int *relax;
1941 isl_ctx *ctx;
1942 isl_bool try_relax;
1944 n_cut = count_ineq(&info[i], STATUS_CUT);
1946 k = find_ineq(&info[i], STATUS_ADJ_EQ);
1948 if (n_cut > 0) {
1949 ctx = isl_basic_map_get_ctx(info[i].bmap);
1950 relax = isl_calloc_array(ctx, int, 1 + n_cut);
1951 if (!relax)
1952 return isl_change_error;
1953 relax[0] = k;
1954 try_relax = all_cut_by_one(i, j, info, relax + 1);
1955 if (try_relax < 0)
1956 change = isl_change_error;
1957 } else {
1958 try_relax = isl_bool_true;
1959 relax = &k;
1961 if (try_relax && change == isl_change_none)
1962 change = is_relaxed_extension(i, j, 1 + n_cut, relax, info);
1963 if (n_cut > 0)
1964 free(relax);
1965 if (change != isl_change_none)
1966 return change;
1968 change = can_wrap_in_facet(i, j, k, info, n_cut > 0);
1970 return change;
1973 /* At least one of the basic maps has an equality that is adjacent
1974 * to an inequality. Make sure that only one of the basic maps has
1975 * such an equality and that the other basic map has exactly one
1976 * inequality adjacent to an equality.
1977 * If the other basic map does not have such an inequality, then
1978 * check if all its constraints are either valid or cut constraints
1979 * and, if so, try wrapping in the first map into the second.
1980 * Otherwise, try to extend one basic map with the other or
1981 * wrap one basic map in the other.
1983 static enum isl_change check_adj_eq(int i, int j,
1984 struct isl_coalesce_info *info)
1986 if (any_eq(&info[i], STATUS_ADJ_INEQ) &&
1987 any_eq(&info[j], STATUS_ADJ_INEQ))
1988 /* ADJ EQ TOO MANY */
1989 return isl_change_none;
1991 if (any_eq(&info[i], STATUS_ADJ_INEQ))
1992 return check_adj_eq(j, i, info);
1994 /* j has an equality adjacent to an inequality in i */
1996 if (count_ineq(&info[i], STATUS_ADJ_EQ) != 1) {
1997 if (all_valid_or_cut(&info[i]))
1998 return can_wrap_in_set(i, j, info);
1999 return isl_change_none;
2001 if (any_eq(&info[i], STATUS_CUT))
2002 return isl_change_none;
2003 if (any_ineq(&info[j], STATUS_ADJ_EQ) ||
2004 any_ineq(&info[i], STATUS_ADJ_INEQ) ||
2005 any_ineq(&info[j], STATUS_ADJ_INEQ))
2006 /* ADJ EQ TOO MANY */
2007 return isl_change_none;
2009 return check_single_adj_eq(i, j, info);
2012 /* Disjunct "j" lies on a hyperplane that is adjacent to disjunct "i".
2013 * In particular, disjunct "i" has an inequality constraint that is adjacent
2014 * to a (combination of) equality constraint(s) of disjunct "j",
2015 * but disjunct "j" has no explicit equality constraint adjacent
2016 * to an inequality constraint of disjunct "i".
2018 * Disjunct "i" is already known not to have any equality constraints
2019 * that are adjacent to an equality or inequality constraint.
2020 * Check that, other than the inequality constraint mentioned above,
2021 * all other constraints of disjunct "i" are valid for disjunct "j".
2022 * If so, try and wrap in disjunct "j".
2024 static enum isl_change check_ineq_adj_eq(int i, int j,
2025 struct isl_coalesce_info *info)
2027 int k;
2029 if (any_eq(&info[i], STATUS_CUT))
2030 return isl_change_none;
2031 if (any_ineq(&info[i], STATUS_CUT))
2032 return isl_change_none;
2033 if (any_ineq(&info[i], STATUS_ADJ_INEQ))
2034 return isl_change_none;
2035 if (count_ineq(&info[i], STATUS_ADJ_EQ) != 1)
2036 return isl_change_none;
2038 k = find_ineq(&info[i], STATUS_ADJ_EQ);
2040 return can_wrap_in_facet(i, j, k, info, 0);
2043 /* The two basic maps lie on adjacent hyperplanes. In particular,
2044 * basic map "i" has an equality that lies parallel to basic map "j".
2045 * Check if we can wrap the facets around the parallel hyperplanes
2046 * to include the other set.
2048 * We perform basically the same operations as can_wrap_in_facet,
2049 * except that we don't need to select a facet of one of the sets.
2051 * \\ \\
2052 * \\ => \\
2053 * \ \|
2055 * If there is more than one equality of "i" adjacent to an equality of "j",
2056 * then the result will satisfy one or more equalities that are a linear
2057 * combination of these equalities. These will be encoded as pairs
2058 * of inequalities in the wrapping constraints and need to be made
2059 * explicit.
2061 static enum isl_change check_eq_adj_eq(int i, int j,
2062 struct isl_coalesce_info *info)
2064 int k;
2065 enum isl_change change = isl_change_none;
2066 int detect_equalities = 0;
2067 struct isl_wraps wraps;
2068 isl_ctx *ctx;
2069 isl_mat *mat;
2070 struct isl_set *set_i = NULL;
2071 struct isl_set *set_j = NULL;
2072 struct isl_vec *bound = NULL;
2073 isl_size total = isl_basic_map_dim(info[i].bmap, isl_dim_all);
2075 if (total < 0)
2076 return isl_change_error;
2077 if (count_eq(&info[i], STATUS_ADJ_EQ) != 1)
2078 detect_equalities = 1;
2080 k = find_eq(&info[i], STATUS_ADJ_EQ);
2082 set_i = set_from_updated_bmap(info[i].bmap, info[i].tab);
2083 set_j = set_from_updated_bmap(info[j].bmap, info[j].tab);
2084 ctx = isl_basic_map_get_ctx(info[i].bmap);
2085 mat = isl_mat_alloc(ctx, 2 * (info[i].bmap->n_eq + info[j].bmap->n_eq) +
2086 info[i].bmap->n_ineq + info[j].bmap->n_ineq,
2087 1 + total);
2088 if (wraps_init(&wraps, mat, info, i, j) < 0)
2089 goto error;
2090 bound = isl_vec_alloc(ctx, 1 + total);
2091 if (!set_i || !set_j || !bound)
2092 goto error;
2094 if (k % 2 == 0)
2095 isl_seq_neg(bound->el, info[i].bmap->eq[k / 2], 1 + total);
2096 else
2097 isl_seq_cpy(bound->el, info[i].bmap->eq[k / 2], 1 + total);
2098 isl_int_add_ui(bound->el[0], bound->el[0], 1);
2100 isl_seq_cpy(wraps.mat->row[0], bound->el, 1 + total);
2101 wraps.mat->n_row = 1;
2103 if (add_wraps(&wraps, &info[j], bound->el, set_i) < 0)
2104 goto error;
2105 if (!wraps.mat->n_row)
2106 goto unbounded;
2108 isl_int_sub_ui(bound->el[0], bound->el[0], 1);
2109 isl_seq_neg(bound->el, bound->el, 1 + total);
2111 isl_seq_cpy(wraps.mat->row[wraps.mat->n_row], bound->el, 1 + total);
2112 wraps.mat->n_row++;
2114 if (add_wraps(&wraps, &info[i], bound->el, set_j) < 0)
2115 goto error;
2116 if (!wraps.mat->n_row)
2117 goto unbounded;
2119 change = fuse(i, j, info, wraps.mat, detect_equalities, 0);
2121 if (0) {
2122 error: change = isl_change_error;
2124 unbounded:
2126 wraps_free(&wraps);
2127 isl_set_free(set_i);
2128 isl_set_free(set_j);
2129 isl_vec_free(bound);
2131 return change;
2134 /* Initialize the "eq" and "ineq" fields of "info".
2136 static void init_status(struct isl_coalesce_info *info)
2138 info->eq = info->ineq = NULL;
2141 /* Set info->eq to the positions of the equalities of info->bmap
2142 * with respect to the basic map represented by "tab".
2143 * If info->eq has already been computed, then do not compute it again.
2145 static void set_eq_status_in(struct isl_coalesce_info *info,
2146 struct isl_tab *tab)
2148 if (info->eq)
2149 return;
2150 info->eq = eq_status_in(info->bmap, tab);
2153 /* Set info->ineq to the positions of the inequalities of info->bmap
2154 * with respect to the basic map represented by "tab".
2155 * If info->ineq has already been computed, then do not compute it again.
2157 static void set_ineq_status_in(struct isl_coalesce_info *info,
2158 struct isl_tab *tab)
2160 if (info->ineq)
2161 return;
2162 info->ineq = ineq_status_in(info->bmap, info->tab, tab);
2165 /* Free the memory allocated by the "eq" and "ineq" fields of "info".
2166 * This function assumes that init_status has been called on "info" first,
2167 * after which the "eq" and "ineq" fields may or may not have been
2168 * assigned a newly allocated array.
2170 static void clear_status(struct isl_coalesce_info *info)
2172 free(info->eq);
2173 free(info->ineq);
2176 /* Are all inequality constraints of the basic map represented by "info"
2177 * valid for the other basic map, except for a single constraint
2178 * that is adjacent to an inequality constraint of the other basic map?
2180 static int all_ineq_valid_or_single_adj_ineq(struct isl_coalesce_info *info)
2182 int i;
2183 int k = -1;
2185 for (i = 0; i < info->bmap->n_ineq; ++i) {
2186 if (info->ineq[i] == STATUS_REDUNDANT)
2187 continue;
2188 if (info->ineq[i] == STATUS_VALID)
2189 continue;
2190 if (info->ineq[i] != STATUS_ADJ_INEQ)
2191 return 0;
2192 if (k != -1)
2193 return 0;
2194 k = i;
2197 return k != -1;
2200 /* Basic map "i" has one or more equality constraints that separate it
2201 * from basic map "j". Check if it happens to be an extension
2202 * of basic map "j".
2203 * In particular, check that all constraints of "j" are valid for "i",
2204 * except for one inequality constraint that is adjacent
2205 * to an inequality constraints of "i".
2206 * If so, check for "i" being an extension of "j" by calling
2207 * is_adj_ineq_extension.
2209 * Clean up the memory allocated for keeping track of the status
2210 * of the constraints before returning.
2212 static enum isl_change separating_equality(int i, int j,
2213 struct isl_coalesce_info *info)
2215 enum isl_change change = isl_change_none;
2217 if (all(info[j].eq, 2 * info[j].bmap->n_eq, STATUS_VALID) &&
2218 all_ineq_valid_or_single_adj_ineq(&info[j]))
2219 change = is_adj_ineq_extension(j, i, info);
2221 clear_status(&info[i]);
2222 clear_status(&info[j]);
2223 return change;
2226 /* Check if the union of the given pair of basic maps
2227 * can be represented by a single basic map.
2228 * If so, replace the pair by the single basic map and return
2229 * isl_change_drop_first, isl_change_drop_second or isl_change_fuse.
2230 * Otherwise, return isl_change_none.
2231 * The two basic maps are assumed to live in the same local space.
2232 * The "eq" and "ineq" fields of info[i] and info[j] are assumed
2233 * to have been initialized by the caller, either to NULL or
2234 * to valid information.
2236 * We first check the effect of each constraint of one basic map
2237 * on the other basic map.
2238 * The constraint may be
2239 * redundant the constraint is redundant in its own
2240 * basic map and should be ignore and removed
2241 * in the end
2242 * valid all (integer) points of the other basic map
2243 * satisfy the constraint
2244 * separate no (integer) point of the other basic map
2245 * satisfies the constraint
2246 * cut some but not all points of the other basic map
2247 * satisfy the constraint
2248 * adj_eq the given constraint is adjacent (on the outside)
2249 * to an equality of the other basic map
2250 * adj_ineq the given constraint is adjacent (on the outside)
2251 * to an inequality of the other basic map
2253 * We consider seven cases in which we can replace the pair by a single
2254 * basic map. We ignore all "redundant" constraints.
2256 * 1. all constraints of one basic map are valid
2257 * => the other basic map is a subset and can be removed
2259 * 2. all constraints of both basic maps are either "valid" or "cut"
2260 * and the facets corresponding to the "cut" constraints
2261 * of one of the basic maps lies entirely inside the other basic map
2262 * => the pair can be replaced by a basic map consisting
2263 * of the valid constraints in both basic maps
2265 * 3. there is a single pair of adjacent inequalities
2266 * (all other constraints are "valid")
2267 * => the pair can be replaced by a basic map consisting
2268 * of the valid constraints in both basic maps
2270 * 4. one basic map has a single adjacent inequality, while the other
2271 * constraints are "valid". The other basic map has some
2272 * "cut" constraints, but replacing the adjacent inequality by
2273 * its opposite and adding the valid constraints of the other
2274 * basic map results in a subset of the other basic map
2275 * => the pair can be replaced by a basic map consisting
2276 * of the valid constraints in both basic maps
2278 * 5. there is a single adjacent pair of an inequality and an equality,
2279 * the other constraints of the basic map containing the inequality are
2280 * "valid". Moreover, if the inequality the basic map is relaxed
2281 * and then turned into an equality, then resulting facet lies
2282 * entirely inside the other basic map
2283 * => the pair can be replaced by the basic map containing
2284 * the inequality, with the inequality relaxed.
2286 * 6. there is a single inequality adjacent to an equality,
2287 * the other constraints of the basic map containing the inequality are
2288 * "valid". Moreover, the facets corresponding to both
2289 * the inequality and the equality can be wrapped around their
2290 * ridges to include the other basic map
2291 * => the pair can be replaced by a basic map consisting
2292 * of the valid constraints in both basic maps together
2293 * with all wrapping constraints
2295 * 7. one of the basic maps extends beyond the other by at most one.
2296 * Moreover, the facets corresponding to the cut constraints and
2297 * the pieces of the other basic map at offset one from these cut
2298 * constraints can be wrapped around their ridges to include
2299 * the union of the two basic maps
2300 * => the pair can be replaced by a basic map consisting
2301 * of the valid constraints in both basic maps together
2302 * with all wrapping constraints
2304 * 8. the two basic maps live in adjacent hyperplanes. In principle
2305 * such sets can always be combined through wrapping, but we impose
2306 * that there is only one such pair, to avoid overeager coalescing.
2308 * Throughout the computation, we maintain a collection of tableaus
2309 * corresponding to the basic maps. When the basic maps are dropped
2310 * or combined, the tableaus are modified accordingly.
2312 static enum isl_change coalesce_local_pair_reuse(int i, int j,
2313 struct isl_coalesce_info *info)
2315 enum isl_change change = isl_change_none;
2317 set_ineq_status_in(&info[i], info[j].tab);
2318 if (info[i].bmap->n_ineq && !info[i].ineq)
2319 goto error;
2320 if (any_ineq(&info[i], STATUS_ERROR))
2321 goto error;
2322 if (any_ineq(&info[i], STATUS_SEPARATE))
2323 goto done;
2325 set_ineq_status_in(&info[j], info[i].tab);
2326 if (info[j].bmap->n_ineq && !info[j].ineq)
2327 goto error;
2328 if (any_ineq(&info[j], STATUS_ERROR))
2329 goto error;
2330 if (any_ineq(&info[j], STATUS_SEPARATE))
2331 goto done;
2333 set_eq_status_in(&info[i], info[j].tab);
2334 if (info[i].bmap->n_eq && !info[i].eq)
2335 goto error;
2336 if (any_eq(&info[i], STATUS_ERROR))
2337 goto error;
2339 set_eq_status_in(&info[j], info[i].tab);
2340 if (info[j].bmap->n_eq && !info[j].eq)
2341 goto error;
2342 if (any_eq(&info[j], STATUS_ERROR))
2343 goto error;
2345 if (any_eq(&info[i], STATUS_SEPARATE))
2346 return separating_equality(i, j, info);
2347 if (any_eq(&info[j], STATUS_SEPARATE))
2348 return separating_equality(j, i, info);
2350 if (all(info[i].eq, 2 * info[i].bmap->n_eq, STATUS_VALID) &&
2351 all(info[i].ineq, info[i].bmap->n_ineq, STATUS_VALID)) {
2352 drop(&info[j]);
2353 change = isl_change_drop_second;
2354 } else if (all(info[j].eq, 2 * info[j].bmap->n_eq, STATUS_VALID) &&
2355 all(info[j].ineq, info[j].bmap->n_ineq, STATUS_VALID)) {
2356 drop(&info[i]);
2357 change = isl_change_drop_first;
2358 } else if (any_eq(&info[i], STATUS_ADJ_EQ)) {
2359 change = check_eq_adj_eq(i, j, info);
2360 } else if (any_eq(&info[j], STATUS_ADJ_EQ)) {
2361 change = check_eq_adj_eq(j, i, info);
2362 } else if (any_eq(&info[i], STATUS_ADJ_INEQ) ||
2363 any_eq(&info[j], STATUS_ADJ_INEQ)) {
2364 change = check_adj_eq(i, j, info);
2365 } else if (any_ineq(&info[i], STATUS_ADJ_EQ)) {
2366 change = check_ineq_adj_eq(i, j, info);
2367 } else if (any_ineq(&info[j], STATUS_ADJ_EQ)) {
2368 change = check_ineq_adj_eq(j, i, info);
2369 } else if (any_ineq(&info[i], STATUS_ADJ_INEQ) ||
2370 any_ineq(&info[j], STATUS_ADJ_INEQ)) {
2371 change = check_adj_ineq(i, j, info);
2372 } else {
2373 if (!any_eq(&info[i], STATUS_CUT) &&
2374 !any_eq(&info[j], STATUS_CUT))
2375 change = check_facets(i, j, info);
2376 if (change == isl_change_none)
2377 change = check_wrap(i, j, info);
2380 done:
2381 clear_status(&info[i]);
2382 clear_status(&info[j]);
2383 return change;
2384 error:
2385 clear_status(&info[i]);
2386 clear_status(&info[j]);
2387 return isl_change_error;
2390 /* Check if the union of the given pair of basic maps
2391 * can be represented by a single basic map.
2392 * If so, replace the pair by the single basic map and return
2393 * isl_change_drop_first, isl_change_drop_second or isl_change_fuse.
2394 * Otherwise, return isl_change_none.
2395 * The two basic maps are assumed to live in the same local space.
2397 static enum isl_change coalesce_local_pair(int i, int j,
2398 struct isl_coalesce_info *info)
2400 init_status(&info[i]);
2401 init_status(&info[j]);
2402 return coalesce_local_pair_reuse(i, j, info);
2405 /* Shift the integer division at position "div" of the basic map
2406 * represented by "info" by "shift".
2408 * That is, if the integer division has the form
2410 * floor(f(x)/d)
2412 * then replace it by
2414 * floor((f(x) + shift * d)/d) - shift
2416 static isl_stat shift_div(struct isl_coalesce_info *info, int div,
2417 isl_int shift)
2419 isl_size total, n_div;
2421 info->bmap = isl_basic_map_shift_div(info->bmap, div, 0, shift);
2422 if (!info->bmap)
2423 return isl_stat_error;
2425 total = isl_basic_map_dim(info->bmap, isl_dim_all);
2426 n_div = isl_basic_map_dim(info->bmap, isl_dim_div);
2427 if (total < 0 || n_div < 0)
2428 return isl_stat_error;
2429 total -= n_div;
2430 if (isl_tab_shift_var(info->tab, total + div, shift) < 0)
2431 return isl_stat_error;
2433 return isl_stat_ok;
2436 /* If the integer division at position "div" is defined by an equality,
2437 * i.e., a stride constraint, then change the integer division expression
2438 * to have a constant term equal to zero.
2440 * Let the equality constraint be
2442 * c + f + m a = 0
2444 * The integer division expression is then typically of the form
2446 * a = floor((-f - c')/m)
2448 * The integer division is first shifted by t = floor(c/m),
2449 * turning the equality constraint into
2451 * c - m floor(c/m) + f + m a' = 0
2453 * i.e.,
2455 * (c mod m) + f + m a' = 0
2457 * That is,
2459 * a' = (-f - (c mod m))/m = floor((-f)/m)
2461 * because a' is an integer and 0 <= (c mod m) < m.
2462 * The constant term of a' can therefore be zeroed out,
2463 * but only if the integer division expression is of the expected form.
2465 static isl_stat normalize_stride_div(struct isl_coalesce_info *info, int div)
2467 isl_bool defined, valid;
2468 isl_stat r;
2469 isl_constraint *c;
2470 isl_int shift, stride;
2472 defined = isl_basic_map_has_defining_equality(info->bmap, isl_dim_div,
2473 div, &c);
2474 if (defined < 0)
2475 return isl_stat_error;
2476 if (!defined)
2477 return isl_stat_ok;
2478 if (!c)
2479 return isl_stat_error;
2480 valid = isl_constraint_is_div_equality(c, div);
2481 isl_int_init(shift);
2482 isl_int_init(stride);
2483 isl_constraint_get_constant(c, &shift);
2484 isl_constraint_get_coefficient(c, isl_dim_div, div, &stride);
2485 isl_int_fdiv_q(shift, shift, stride);
2486 r = shift_div(info, div, shift);
2487 isl_int_clear(stride);
2488 isl_int_clear(shift);
2489 isl_constraint_free(c);
2490 if (r < 0 || valid < 0)
2491 return isl_stat_error;
2492 if (!valid)
2493 return isl_stat_ok;
2494 info->bmap = isl_basic_map_set_div_expr_constant_num_si_inplace(
2495 info->bmap, div, 0);
2496 if (!info->bmap)
2497 return isl_stat_error;
2498 return isl_stat_ok;
2501 /* The basic maps represented by "info1" and "info2" are known
2502 * to have the same number of integer divisions.
2503 * Check if pairs of integer divisions are equal to each other
2504 * despite the fact that they differ by a rational constant.
2506 * In particular, look for any pair of integer divisions that
2507 * only differ in their constant terms.
2508 * If either of these integer divisions is defined
2509 * by stride constraints, then modify it to have a zero constant term.
2510 * If both are defined by stride constraints then in the end they will have
2511 * the same (zero) constant term.
2513 static isl_stat harmonize_stride_divs(struct isl_coalesce_info *info1,
2514 struct isl_coalesce_info *info2)
2516 int i;
2517 isl_size n;
2519 n = isl_basic_map_dim(info1->bmap, isl_dim_div);
2520 if (n < 0)
2521 return isl_stat_error;
2522 for (i = 0; i < n; ++i) {
2523 isl_bool known, harmonize;
2525 known = isl_basic_map_div_is_known(info1->bmap, i);
2526 if (known >= 0 && known)
2527 known = isl_basic_map_div_is_known(info2->bmap, i);
2528 if (known < 0)
2529 return isl_stat_error;
2530 if (!known)
2531 continue;
2532 harmonize = isl_basic_map_equal_div_expr_except_constant(
2533 info1->bmap, i, info2->bmap, i);
2534 if (harmonize < 0)
2535 return isl_stat_error;
2536 if (!harmonize)
2537 continue;
2538 if (normalize_stride_div(info1, i) < 0)
2539 return isl_stat_error;
2540 if (normalize_stride_div(info2, i) < 0)
2541 return isl_stat_error;
2544 return isl_stat_ok;
2547 /* If "shift" is an integer constant, then shift the integer division
2548 * at position "div" of the basic map represented by "info" by "shift".
2549 * If "shift" is not an integer constant, then do nothing.
2550 * If "shift" is equal to zero, then no shift needs to be performed either.
2552 * That is, if the integer division has the form
2554 * floor(f(x)/d)
2556 * then replace it by
2558 * floor((f(x) + shift * d)/d) - shift
2560 static isl_stat shift_if_cst_int(struct isl_coalesce_info *info, int div,
2561 __isl_keep isl_aff *shift)
2563 isl_bool cst;
2564 isl_stat r;
2565 isl_int d;
2566 isl_val *c;
2568 cst = isl_aff_is_cst(shift);
2569 if (cst < 0 || !cst)
2570 return cst < 0 ? isl_stat_error : isl_stat_ok;
2572 c = isl_aff_get_constant_val(shift);
2573 cst = isl_val_is_int(c);
2574 if (cst >= 0 && cst)
2575 cst = isl_bool_not(isl_val_is_zero(c));
2576 if (cst < 0 || !cst) {
2577 isl_val_free(c);
2578 return cst < 0 ? isl_stat_error : isl_stat_ok;
2581 isl_int_init(d);
2582 r = isl_val_get_num_isl_int(c, &d);
2583 if (r >= 0)
2584 r = shift_div(info, div, d);
2585 isl_int_clear(d);
2587 isl_val_free(c);
2589 return r;
2592 /* Check if some of the divs in the basic map represented by "info1"
2593 * are shifts of the corresponding divs in the basic map represented
2594 * by "info2", taking into account the equality constraints "eq1" of "info1"
2595 * and "eq2" of "info2". If so, align them with those of "info2".
2596 * "info1" and "info2" are assumed to have the same number
2597 * of integer divisions.
2599 * An integer division is considered to be a shift of another integer
2600 * division if, after simplification with respect to the equality
2601 * constraints of the other basic map, one is equal to the other
2602 * plus a constant.
2604 * In particular, for each pair of integer divisions, if both are known,
2605 * have the same denominator and are not already equal to each other,
2606 * simplify each with respect to the equality constraints
2607 * of the other basic map. If the difference is an integer constant,
2608 * then move this difference outside.
2609 * That is, if, after simplification, one integer division is of the form
2611 * floor((f(x) + c_1)/d)
2613 * while the other is of the form
2615 * floor((f(x) + c_2)/d)
2617 * and n = (c_2 - c_1)/d is an integer, then replace the first
2618 * integer division by
2620 * floor((f_1(x) + c_1 + n * d)/d) - n,
2622 * where floor((f_1(x) + c_1 + n * d)/d) = floor((f2(x) + c_2)/d)
2623 * after simplification with respect to the equality constraints.
2625 static isl_stat harmonize_divs_with_hulls(struct isl_coalesce_info *info1,
2626 struct isl_coalesce_info *info2, __isl_keep isl_basic_set *eq1,
2627 __isl_keep isl_basic_set *eq2)
2629 int i;
2630 isl_size total;
2631 isl_local_space *ls1, *ls2;
2633 total = isl_basic_map_dim(info1->bmap, isl_dim_all);
2634 if (total < 0)
2635 return isl_stat_error;
2636 ls1 = isl_local_space_wrap(isl_basic_map_get_local_space(info1->bmap));
2637 ls2 = isl_local_space_wrap(isl_basic_map_get_local_space(info2->bmap));
2638 for (i = 0; i < info1->bmap->n_div; ++i) {
2639 isl_stat r;
2640 isl_aff *div1, *div2;
2642 if (!isl_local_space_div_is_known(ls1, i) ||
2643 !isl_local_space_div_is_known(ls2, i))
2644 continue;
2645 if (isl_int_ne(info1->bmap->div[i][0], info2->bmap->div[i][0]))
2646 continue;
2647 if (isl_seq_eq(info1->bmap->div[i] + 1,
2648 info2->bmap->div[i] + 1, 1 + total))
2649 continue;
2650 div1 = isl_local_space_get_div(ls1, i);
2651 div2 = isl_local_space_get_div(ls2, i);
2652 div1 = isl_aff_substitute_equalities(div1,
2653 isl_basic_set_copy(eq2));
2654 div2 = isl_aff_substitute_equalities(div2,
2655 isl_basic_set_copy(eq1));
2656 div2 = isl_aff_sub(div2, div1);
2657 r = shift_if_cst_int(info1, i, div2);
2658 isl_aff_free(div2);
2659 if (r < 0)
2660 break;
2662 isl_local_space_free(ls1);
2663 isl_local_space_free(ls2);
2665 if (i < info1->bmap->n_div)
2666 return isl_stat_error;
2667 return isl_stat_ok;
2670 /* Check if some of the divs in the basic map represented by "info1"
2671 * are shifts of the corresponding divs in the basic map represented
2672 * by "info2". If so, align them with those of "info2".
2673 * Only do this if "info1" and "info2" have the same number
2674 * of integer divisions.
2676 * An integer division is considered to be a shift of another integer
2677 * division if, after simplification with respect to the equality
2678 * constraints of the other basic map, one is equal to the other
2679 * plus a constant.
2681 * First check if pairs of integer divisions are equal to each other
2682 * despite the fact that they differ by a rational constant.
2683 * If so, try and arrange for them to have the same constant term.
2685 * Then, extract the equality constraints and continue with
2686 * harmonize_divs_with_hulls.
2688 * If the equality constraints of both basic maps are the same,
2689 * then there is no need to perform any shifting since
2690 * the coefficients of the integer divisions should have been
2691 * reduced in the same way.
2693 static isl_stat harmonize_divs(struct isl_coalesce_info *info1,
2694 struct isl_coalesce_info *info2)
2696 isl_bool equal;
2697 isl_basic_map *bmap1, *bmap2;
2698 isl_basic_set *eq1, *eq2;
2699 isl_stat r;
2701 if (!info1->bmap || !info2->bmap)
2702 return isl_stat_error;
2704 if (info1->bmap->n_div != info2->bmap->n_div)
2705 return isl_stat_ok;
2706 if (info1->bmap->n_div == 0)
2707 return isl_stat_ok;
2709 if (harmonize_stride_divs(info1, info2) < 0)
2710 return isl_stat_error;
2712 bmap1 = isl_basic_map_copy(info1->bmap);
2713 bmap2 = isl_basic_map_copy(info2->bmap);
2714 eq1 = isl_basic_map_wrap(isl_basic_map_plain_affine_hull(bmap1));
2715 eq2 = isl_basic_map_wrap(isl_basic_map_plain_affine_hull(bmap2));
2716 equal = isl_basic_set_plain_is_equal(eq1, eq2);
2717 if (equal < 0)
2718 r = isl_stat_error;
2719 else if (equal)
2720 r = isl_stat_ok;
2721 else
2722 r = harmonize_divs_with_hulls(info1, info2, eq1, eq2);
2723 isl_basic_set_free(eq1);
2724 isl_basic_set_free(eq2);
2726 return r;
2729 /* Do the two basic maps live in the same local space, i.e.,
2730 * do they have the same (known) divs?
2731 * If either basic map has any unknown divs, then we can only assume
2732 * that they do not live in the same local space.
2734 static isl_bool same_divs(__isl_keep isl_basic_map *bmap1,
2735 __isl_keep isl_basic_map *bmap2)
2737 int i;
2738 isl_bool known;
2739 isl_size total;
2741 if (!bmap1 || !bmap2)
2742 return isl_bool_error;
2743 if (bmap1->n_div != bmap2->n_div)
2744 return isl_bool_false;
2746 if (bmap1->n_div == 0)
2747 return isl_bool_true;
2749 known = isl_basic_map_divs_known(bmap1);
2750 if (known < 0 || !known)
2751 return known;
2752 known = isl_basic_map_divs_known(bmap2);
2753 if (known < 0 || !known)
2754 return known;
2756 total = isl_basic_map_dim(bmap1, isl_dim_all);
2757 if (total < 0)
2758 return isl_bool_error;
2759 for (i = 0; i < bmap1->n_div; ++i)
2760 if (!isl_seq_eq(bmap1->div[i], bmap2->div[i], 2 + total))
2761 return isl_bool_false;
2763 return isl_bool_true;
2766 /* Assuming that "tab" contains the equality constraints and
2767 * the initial inequality constraints of "bmap", copy the remaining
2768 * inequality constraints of "bmap" to "Tab".
2770 static isl_stat copy_ineq(struct isl_tab *tab, __isl_keep isl_basic_map *bmap)
2772 int i, n_ineq;
2774 if (!bmap)
2775 return isl_stat_error;
2777 n_ineq = tab->n_con - tab->n_eq;
2778 for (i = n_ineq; i < bmap->n_ineq; ++i)
2779 if (isl_tab_add_ineq(tab, bmap->ineq[i]) < 0)
2780 return isl_stat_error;
2782 return isl_stat_ok;
2785 /* Description of an integer division that is added
2786 * during an expansion.
2787 * "pos" is the position of the corresponding variable.
2788 * "cst" indicates whether this integer division has a fixed value.
2789 * "val" contains the fixed value, if the value is fixed.
2791 struct isl_expanded {
2792 int pos;
2793 isl_bool cst;
2794 isl_int val;
2797 /* For each of the "n" integer division variables "expanded",
2798 * if the variable has a fixed value, then add two inequality
2799 * constraints expressing the fixed value.
2800 * Otherwise, add the corresponding div constraints.
2801 * The caller is responsible for removing the div constraints
2802 * that it added for all these "n" integer divisions.
2804 * The div constraints and the pair of inequality constraints
2805 * forcing the fixed value cannot both be added for a given variable
2806 * as the combination may render some of the original constraints redundant.
2807 * These would then be ignored during the coalescing detection,
2808 * while they could remain in the fused result.
2810 * The two added inequality constraints are
2812 * -a + v >= 0
2813 * a - v >= 0
2815 * with "a" the variable and "v" its fixed value.
2816 * The facet corresponding to one of these two constraints is selected
2817 * in the tableau to ensure that the pair of inequality constraints
2818 * is treated as an equality constraint.
2820 * The information in info->ineq is thrown away because it was
2821 * computed in terms of div constraints, while some of those
2822 * have now been replaced by these pairs of inequality constraints.
2824 static isl_stat fix_constant_divs(struct isl_coalesce_info *info,
2825 int n, struct isl_expanded *expanded)
2827 unsigned o_div;
2828 int i;
2829 isl_vec *ineq;
2831 o_div = isl_basic_map_offset(info->bmap, isl_dim_div) - 1;
2832 ineq = isl_vec_alloc(isl_tab_get_ctx(info->tab), 1 + info->tab->n_var);
2833 if (!ineq)
2834 return isl_stat_error;
2835 isl_seq_clr(ineq->el + 1, info->tab->n_var);
2837 for (i = 0; i < n; ++i) {
2838 if (!expanded[i].cst) {
2839 info->bmap = isl_basic_map_extend_constraints(
2840 info->bmap, 0, 2);
2841 info->bmap = isl_basic_map_add_div_constraints(
2842 info->bmap, expanded[i].pos - o_div);
2843 } else {
2844 isl_int_set_si(ineq->el[1 + expanded[i].pos], -1);
2845 isl_int_set(ineq->el[0], expanded[i].val);
2846 info->bmap = isl_basic_map_add_ineq(info->bmap,
2847 ineq->el);
2848 isl_int_set_si(ineq->el[1 + expanded[i].pos], 1);
2849 isl_int_neg(ineq->el[0], expanded[i].val);
2850 info->bmap = isl_basic_map_add_ineq(info->bmap,
2851 ineq->el);
2852 isl_int_set_si(ineq->el[1 + expanded[i].pos], 0);
2854 if (copy_ineq(info->tab, info->bmap) < 0)
2855 break;
2856 if (expanded[i].cst &&
2857 isl_tab_select_facet(info->tab, info->tab->n_con - 1) < 0)
2858 break;
2861 isl_vec_free(ineq);
2863 clear_status(info);
2864 init_status(info);
2866 return i < n ? isl_stat_error : isl_stat_ok;
2869 /* Insert the "n" integer division variables "expanded"
2870 * into info->tab and info->bmap and
2871 * update info->ineq with respect to the redundant constraints
2872 * in the resulting tableau.
2873 * "bmap" contains the result of this insertion in info->bmap,
2874 * while info->bmap is the original version
2875 * of "bmap", i.e., the one that corresponds to the current
2876 * state of info->tab. The number of constraints in info->bmap
2877 * is assumed to be the same as the number of constraints
2878 * in info->tab. This is required to be able to detect
2879 * the extra constraints in "bmap".
2881 * In particular, introduce extra variables corresponding
2882 * to the extra integer divisions and add the div constraints
2883 * that were added to "bmap" after info->tab was created
2884 * from info->bmap.
2885 * Furthermore, check if these extra integer divisions happen
2886 * to attain a fixed integer value in info->tab.
2887 * If so, replace the corresponding div constraints by pairs
2888 * of inequality constraints that fix these
2889 * integer divisions to their single integer values.
2890 * Replace info->bmap by "bmap" to match the changes to info->tab.
2891 * info->ineq was computed without a tableau and therefore
2892 * does not take into account the redundant constraints
2893 * in the tableau. Mark them here.
2894 * There is no need to check the newly added div constraints
2895 * since they cannot be redundant.
2896 * The redundancy check is not performed when constants have been discovered
2897 * since info->ineq is completely thrown away in this case.
2899 static isl_stat tab_insert_divs(struct isl_coalesce_info *info,
2900 int n, struct isl_expanded *expanded, __isl_take isl_basic_map *bmap)
2902 int i, n_ineq;
2903 unsigned n_eq;
2904 struct isl_tab_undo *snap;
2905 int any;
2907 if (!bmap)
2908 return isl_stat_error;
2909 if (info->bmap->n_eq + info->bmap->n_ineq != info->tab->n_con)
2910 isl_die(isl_basic_map_get_ctx(bmap), isl_error_internal,
2911 "original tableau does not correspond "
2912 "to original basic map", goto error);
2914 if (isl_tab_extend_vars(info->tab, n) < 0)
2915 goto error;
2916 if (isl_tab_extend_cons(info->tab, 2 * n) < 0)
2917 goto error;
2919 for (i = 0; i < n; ++i) {
2920 if (isl_tab_insert_var(info->tab, expanded[i].pos) < 0)
2921 goto error;
2924 snap = isl_tab_snap(info->tab);
2926 n_ineq = info->tab->n_con - info->tab->n_eq;
2927 if (copy_ineq(info->tab, bmap) < 0)
2928 goto error;
2930 isl_basic_map_free(info->bmap);
2931 info->bmap = bmap;
2933 any = 0;
2934 for (i = 0; i < n; ++i) {
2935 expanded[i].cst = isl_tab_is_constant(info->tab,
2936 expanded[i].pos, &expanded[i].val);
2937 if (expanded[i].cst < 0)
2938 return isl_stat_error;
2939 if (expanded[i].cst)
2940 any = 1;
2943 if (any) {
2944 if (isl_tab_rollback(info->tab, snap) < 0)
2945 return isl_stat_error;
2946 info->bmap = isl_basic_map_cow(info->bmap);
2947 if (isl_basic_map_free_inequality(info->bmap, 2 * n) < 0)
2948 return isl_stat_error;
2950 return fix_constant_divs(info, n, expanded);
2953 n_eq = info->bmap->n_eq;
2954 for (i = 0; i < n_ineq; ++i) {
2955 if (isl_tab_is_redundant(info->tab, n_eq + i))
2956 info->ineq[i] = STATUS_REDUNDANT;
2959 return isl_stat_ok;
2960 error:
2961 isl_basic_map_free(bmap);
2962 return isl_stat_error;
2965 /* Expand info->tab and info->bmap in the same way "bmap" was expanded
2966 * in isl_basic_map_expand_divs using the expansion "exp" and
2967 * update info->ineq with respect to the redundant constraints
2968 * in the resulting tableau. info->bmap is the original version
2969 * of "bmap", i.e., the one that corresponds to the current
2970 * state of info->tab. The number of constraints in info->bmap
2971 * is assumed to be the same as the number of constraints
2972 * in info->tab. This is required to be able to detect
2973 * the extra constraints in "bmap".
2975 * Extract the positions where extra local variables are introduced
2976 * from "exp" and call tab_insert_divs.
2978 static isl_stat expand_tab(struct isl_coalesce_info *info, int *exp,
2979 __isl_take isl_basic_map *bmap)
2981 isl_ctx *ctx;
2982 struct isl_expanded *expanded;
2983 int i, j, k, n;
2984 int extra_var;
2985 isl_size total, n_div;
2986 unsigned pos;
2987 isl_stat r;
2989 total = isl_basic_map_dim(bmap, isl_dim_all);
2990 n_div = isl_basic_map_dim(bmap, isl_dim_div);
2991 if (total < 0 || n_div < 0)
2992 return isl_stat_error;
2993 pos = total - n_div;
2994 extra_var = total - info->tab->n_var;
2995 n = n_div - extra_var;
2997 ctx = isl_basic_map_get_ctx(bmap);
2998 expanded = isl_calloc_array(ctx, struct isl_expanded, extra_var);
2999 if (extra_var && !expanded)
3000 goto error;
3002 i = 0;
3003 k = 0;
3004 for (j = 0; j < n_div; ++j) {
3005 if (i < n && exp[i] == j) {
3006 ++i;
3007 continue;
3009 expanded[k++].pos = pos + j;
3012 for (k = 0; k < extra_var; ++k)
3013 isl_int_init(expanded[k].val);
3015 r = tab_insert_divs(info, extra_var, expanded, bmap);
3017 for (k = 0; k < extra_var; ++k)
3018 isl_int_clear(expanded[k].val);
3019 free(expanded);
3021 return r;
3022 error:
3023 isl_basic_map_free(bmap);
3024 return isl_stat_error;
3027 /* Check if the union of the basic maps represented by info[i] and info[j]
3028 * can be represented by a single basic map,
3029 * after expanding the divs of info[i] to match those of info[j].
3030 * If so, replace the pair by the single basic map and return
3031 * isl_change_drop_first, isl_change_drop_second or isl_change_fuse.
3032 * Otherwise, return isl_change_none.
3034 * The caller has already checked for info[j] being a subset of info[i].
3035 * If some of the divs of info[j] are unknown, then the expanded info[i]
3036 * will not have the corresponding div constraints. The other patterns
3037 * therefore cannot apply. Skip the computation in this case.
3039 * The expansion is performed using the divs "div" and expansion "exp"
3040 * computed by the caller.
3041 * info[i].bmap has already been expanded and the result is passed in
3042 * as "bmap".
3043 * The "eq" and "ineq" fields of info[i] reflect the status of
3044 * the constraints of the expanded "bmap" with respect to info[j].tab.
3045 * However, inequality constraints that are redundant in info[i].tab
3046 * have not yet been marked as such because no tableau was available.
3048 * Replace info[i].bmap by "bmap" and expand info[i].tab as well,
3049 * updating info[i].ineq with respect to the redundant constraints.
3050 * Then try and coalesce the expanded info[i] with info[j],
3051 * reusing the information in info[i].eq and info[i].ineq.
3052 * If this does not result in any coalescing or if it results in info[j]
3053 * getting dropped (which should not happen in practice, since the case
3054 * of info[j] being a subset of info[i] has already been checked by
3055 * the caller), then revert info[i] to its original state.
3057 static enum isl_change coalesce_expand_tab_divs(__isl_take isl_basic_map *bmap,
3058 int i, int j, struct isl_coalesce_info *info, __isl_keep isl_mat *div,
3059 int *exp)
3061 isl_bool known;
3062 isl_basic_map *bmap_i;
3063 struct isl_tab_undo *snap;
3064 enum isl_change change = isl_change_none;
3066 known = isl_basic_map_divs_known(info[j].bmap);
3067 if (known < 0 || !known) {
3068 clear_status(&info[i]);
3069 isl_basic_map_free(bmap);
3070 return known < 0 ? isl_change_error : isl_change_none;
3073 bmap_i = isl_basic_map_copy(info[i].bmap);
3074 snap = isl_tab_snap(info[i].tab);
3075 if (expand_tab(&info[i], exp, bmap) < 0)
3076 change = isl_change_error;
3078 init_status(&info[j]);
3079 if (change == isl_change_none)
3080 change = coalesce_local_pair_reuse(i, j, info);
3081 else
3082 clear_status(&info[i]);
3083 if (change != isl_change_none && change != isl_change_drop_second) {
3084 isl_basic_map_free(bmap_i);
3085 } else {
3086 isl_basic_map_free(info[i].bmap);
3087 info[i].bmap = bmap_i;
3089 if (isl_tab_rollback(info[i].tab, snap) < 0)
3090 change = isl_change_error;
3093 return change;
3096 /* Check if the union of "bmap" and the basic map represented by info[j]
3097 * can be represented by a single basic map,
3098 * after expanding the divs of "bmap" to match those of info[j].
3099 * If so, replace the pair by the single basic map and return
3100 * isl_change_drop_first, isl_change_drop_second or isl_change_fuse.
3101 * Otherwise, return isl_change_none.
3103 * In particular, check if the expanded "bmap" contains the basic map
3104 * represented by the tableau info[j].tab.
3105 * The expansion is performed using the divs "div" and expansion "exp"
3106 * computed by the caller.
3107 * Then we check if all constraints of the expanded "bmap" are valid for
3108 * info[j].tab.
3110 * If "i" is not equal to -1, then "bmap" is equal to info[i].bmap.
3111 * In this case, the positions of the constraints of info[i].bmap
3112 * with respect to the basic map represented by info[j] are stored
3113 * in info[i].
3115 * If the expanded "bmap" does not contain the basic map
3116 * represented by the tableau info[j].tab and if "i" is not -1,
3117 * i.e., if the original "bmap" is info[i].bmap, then expand info[i].tab
3118 * as well and check if that results in coalescing.
3120 static enum isl_change coalesce_with_expanded_divs(
3121 __isl_keep isl_basic_map *bmap, int i, int j,
3122 struct isl_coalesce_info *info, __isl_keep isl_mat *div, int *exp)
3124 enum isl_change change = isl_change_none;
3125 struct isl_coalesce_info info_local, *info_i;
3127 info_i = i >= 0 ? &info[i] : &info_local;
3128 init_status(info_i);
3129 bmap = isl_basic_map_copy(bmap);
3130 bmap = isl_basic_map_expand_divs(bmap, isl_mat_copy(div), exp);
3131 bmap = isl_basic_map_mark_final(bmap);
3133 if (!bmap)
3134 goto error;
3136 info_local.bmap = bmap;
3137 info_i->eq = eq_status_in(bmap, info[j].tab);
3138 if (bmap->n_eq && !info_i->eq)
3139 goto error;
3140 if (any_eq(info_i, STATUS_ERROR))
3141 goto error;
3142 if (any_eq(info_i, STATUS_SEPARATE))
3143 goto done;
3145 info_i->ineq = ineq_status_in(bmap, NULL, info[j].tab);
3146 if (bmap->n_ineq && !info_i->ineq)
3147 goto error;
3148 if (any_ineq(info_i, STATUS_ERROR))
3149 goto error;
3150 if (any_ineq(info_i, STATUS_SEPARATE))
3151 goto done;
3153 if (all(info_i->eq, 2 * bmap->n_eq, STATUS_VALID) &&
3154 all(info_i->ineq, bmap->n_ineq, STATUS_VALID)) {
3155 drop(&info[j]);
3156 change = isl_change_drop_second;
3159 if (change == isl_change_none && i != -1)
3160 return coalesce_expand_tab_divs(bmap, i, j, info, div, exp);
3162 done:
3163 isl_basic_map_free(bmap);
3164 clear_status(info_i);
3165 return change;
3166 error:
3167 isl_basic_map_free(bmap);
3168 clear_status(info_i);
3169 return isl_change_error;
3172 /* Check if the union of "bmap_i" and the basic map represented by info[j]
3173 * can be represented by a single basic map,
3174 * after aligning the divs of "bmap_i" to match those of info[j].
3175 * If so, replace the pair by the single basic map and return
3176 * isl_change_drop_first, isl_change_drop_second or isl_change_fuse.
3177 * Otherwise, return isl_change_none.
3179 * In particular, check if "bmap_i" contains the basic map represented by
3180 * info[j] after aligning the divs of "bmap_i" to those of info[j].
3181 * Note that this can only succeed if the number of divs of "bmap_i"
3182 * is smaller than (or equal to) the number of divs of info[j].
3184 * We first check if the divs of "bmap_i" are all known and form a subset
3185 * of those of info[j].bmap. If so, we pass control over to
3186 * coalesce_with_expanded_divs.
3188 * If "i" is not equal to -1, then "bmap" is equal to info[i].bmap.
3190 static enum isl_change coalesce_after_aligning_divs(
3191 __isl_keep isl_basic_map *bmap_i, int i, int j,
3192 struct isl_coalesce_info *info)
3194 isl_bool known;
3195 isl_mat *div_i, *div_j, *div;
3196 int *exp1 = NULL;
3197 int *exp2 = NULL;
3198 isl_ctx *ctx;
3199 enum isl_change change;
3201 known = isl_basic_map_divs_known(bmap_i);
3202 if (known < 0)
3203 return isl_change_error;
3204 if (!known)
3205 return isl_change_none;
3207 ctx = isl_basic_map_get_ctx(bmap_i);
3209 div_i = isl_basic_map_get_divs(bmap_i);
3210 div_j = isl_basic_map_get_divs(info[j].bmap);
3212 if (!div_i || !div_j)
3213 goto error;
3215 exp1 = isl_alloc_array(ctx, int, div_i->n_row);
3216 exp2 = isl_alloc_array(ctx, int, div_j->n_row);
3217 if ((div_i->n_row && !exp1) || (div_j->n_row && !exp2))
3218 goto error;
3220 div = isl_merge_divs(div_i, div_j, exp1, exp2);
3221 if (!div)
3222 goto error;
3224 if (div->n_row == div_j->n_row)
3225 change = coalesce_with_expanded_divs(bmap_i,
3226 i, j, info, div, exp1);
3227 else
3228 change = isl_change_none;
3230 isl_mat_free(div);
3232 isl_mat_free(div_i);
3233 isl_mat_free(div_j);
3235 free(exp2);
3236 free(exp1);
3238 return change;
3239 error:
3240 isl_mat_free(div_i);
3241 isl_mat_free(div_j);
3242 free(exp1);
3243 free(exp2);
3244 return isl_change_error;
3247 /* Check if basic map "j" is a subset of basic map "i" after
3248 * exploiting the extra equalities of "j" to simplify the divs of "i".
3249 * If so, remove basic map "j" and return isl_change_drop_second.
3251 * If "j" does not have any equalities or if they are the same
3252 * as those of "i", then we cannot exploit them to simplify the divs.
3253 * Similarly, if there are no divs in "i", then they cannot be simplified.
3254 * If, on the other hand, the affine hulls of "i" and "j" do not intersect,
3255 * then "j" cannot be a subset of "i".
3257 * Otherwise, we intersect "i" with the affine hull of "j" and then
3258 * check if "j" is a subset of the result after aligning the divs.
3259 * If so, then "j" is definitely a subset of "i" and can be removed.
3260 * Note that if after intersection with the affine hull of "j".
3261 * "i" still has more divs than "j", then there is no way we can
3262 * align the divs of "i" to those of "j".
3264 static enum isl_change coalesce_subset_with_equalities(int i, int j,
3265 struct isl_coalesce_info *info)
3267 isl_basic_map *hull_i, *hull_j, *bmap_i;
3268 int equal, empty;
3269 enum isl_change change;
3271 if (info[j].bmap->n_eq == 0)
3272 return isl_change_none;
3273 if (info[i].bmap->n_div == 0)
3274 return isl_change_none;
3276 hull_i = isl_basic_map_copy(info[i].bmap);
3277 hull_i = isl_basic_map_plain_affine_hull(hull_i);
3278 hull_j = isl_basic_map_copy(info[j].bmap);
3279 hull_j = isl_basic_map_plain_affine_hull(hull_j);
3281 hull_j = isl_basic_map_intersect(hull_j, isl_basic_map_copy(hull_i));
3282 equal = isl_basic_map_plain_is_equal(hull_i, hull_j);
3283 empty = isl_basic_map_plain_is_empty(hull_j);
3284 isl_basic_map_free(hull_i);
3286 if (equal < 0 || equal || empty < 0 || empty) {
3287 isl_basic_map_free(hull_j);
3288 if (equal < 0 || empty < 0)
3289 return isl_change_error;
3290 return isl_change_none;
3293 bmap_i = isl_basic_map_copy(info[i].bmap);
3294 bmap_i = isl_basic_map_intersect(bmap_i, hull_j);
3295 if (!bmap_i)
3296 return isl_change_error;
3298 if (bmap_i->n_div > info[j].bmap->n_div) {
3299 isl_basic_map_free(bmap_i);
3300 return isl_change_none;
3303 change = coalesce_after_aligning_divs(bmap_i, -1, j, info);
3305 isl_basic_map_free(bmap_i);
3307 return change;
3310 /* Check if the union of and the basic maps represented by info[i] and info[j]
3311 * can be represented by a single basic map, by aligning or equating
3312 * their integer divisions.
3313 * If so, replace the pair by the single basic map and return
3314 * isl_change_drop_first, isl_change_drop_second or isl_change_fuse.
3315 * Otherwise, return isl_change_none.
3317 * Note that we only perform any test if the number of divs is different
3318 * in the two basic maps. In case the number of divs is the same,
3319 * we have already established that the divs are different
3320 * in the two basic maps.
3321 * In particular, if the number of divs of basic map i is smaller than
3322 * the number of divs of basic map j, then we check if j is a subset of i
3323 * and vice versa.
3325 static enum isl_change coalesce_divs(int i, int j,
3326 struct isl_coalesce_info *info)
3328 enum isl_change change = isl_change_none;
3330 if (info[i].bmap->n_div < info[j].bmap->n_div)
3331 change = coalesce_after_aligning_divs(info[i].bmap, i, j, info);
3332 if (change != isl_change_none)
3333 return change;
3335 if (info[j].bmap->n_div < info[i].bmap->n_div)
3336 change = coalesce_after_aligning_divs(info[j].bmap, j, i, info);
3337 if (change != isl_change_none)
3338 return invert_change(change);
3340 change = coalesce_subset_with_equalities(i, j, info);
3341 if (change != isl_change_none)
3342 return change;
3344 change = coalesce_subset_with_equalities(j, i, info);
3345 if (change != isl_change_none)
3346 return invert_change(change);
3348 return isl_change_none;
3351 /* Does "bmap" involve any divs that themselves refer to divs?
3353 static isl_bool has_nested_div(__isl_keep isl_basic_map *bmap)
3355 int i;
3356 isl_size total;
3357 isl_size n_div;
3359 total = isl_basic_map_dim(bmap, isl_dim_all);
3360 n_div = isl_basic_map_dim(bmap, isl_dim_div);
3361 if (total < 0 || n_div < 0)
3362 return isl_bool_error;
3363 total -= n_div;
3365 for (i = 0; i < n_div; ++i)
3366 if (isl_seq_first_non_zero(bmap->div[i] + 2 + total,
3367 n_div) != -1)
3368 return isl_bool_true;
3370 return isl_bool_false;
3373 /* Return a list of affine expressions, one for each integer division
3374 * in "bmap_i". For each integer division that also appears in "bmap_j",
3375 * the affine expression is set to NaN. The number of NaNs in the list
3376 * is equal to the number of integer divisions in "bmap_j".
3377 * For the other integer divisions of "bmap_i", the corresponding
3378 * element in the list is a purely affine expression equal to the integer
3379 * division in "hull".
3380 * If no such list can be constructed, then the number of elements
3381 * in the returned list is smaller than the number of integer divisions
3382 * in "bmap_i".
3384 static __isl_give isl_aff_list *set_up_substitutions(
3385 __isl_keep isl_basic_map *bmap_i, __isl_keep isl_basic_map *bmap_j,
3386 __isl_take isl_basic_map *hull)
3388 isl_size n_div_i, n_div_j, total;
3389 isl_ctx *ctx;
3390 isl_local_space *ls;
3391 isl_basic_set *wrap_hull;
3392 isl_aff *aff_nan;
3393 isl_aff_list *list;
3394 int i, j;
3396 n_div_i = isl_basic_map_dim(bmap_i, isl_dim_div);
3397 n_div_j = isl_basic_map_dim(bmap_j, isl_dim_div);
3398 total = isl_basic_map_dim(bmap_i, isl_dim_all);
3399 if (!hull || n_div_i < 0 || n_div_j < 0 || total < 0)
3400 return NULL;
3402 ctx = isl_basic_map_get_ctx(hull);
3403 total -= n_div_i;
3405 ls = isl_basic_map_get_local_space(bmap_i);
3406 ls = isl_local_space_wrap(ls);
3407 wrap_hull = isl_basic_map_wrap(hull);
3409 aff_nan = isl_aff_nan_on_domain(isl_local_space_copy(ls));
3410 list = isl_aff_list_alloc(ctx, n_div_i);
3412 j = 0;
3413 for (i = 0; i < n_div_i; ++i) {
3414 isl_aff *aff;
3415 isl_size n_div;
3417 if (j < n_div_j &&
3418 isl_basic_map_equal_div_expr_part(bmap_i, i, bmap_j, j,
3419 0, 2 + total)) {
3420 ++j;
3421 list = isl_aff_list_add(list, isl_aff_copy(aff_nan));
3422 continue;
3424 if (n_div_i - i <= n_div_j - j)
3425 break;
3427 aff = isl_local_space_get_div(ls, i);
3428 aff = isl_aff_substitute_equalities(aff,
3429 isl_basic_set_copy(wrap_hull));
3430 aff = isl_aff_floor(aff);
3431 n_div = isl_aff_dim(aff, isl_dim_div);
3432 if (n_div < 0)
3433 goto error;
3434 if (n_div != 0) {
3435 isl_aff_free(aff);
3436 break;
3439 list = isl_aff_list_add(list, aff);
3442 isl_aff_free(aff_nan);
3443 isl_local_space_free(ls);
3444 isl_basic_set_free(wrap_hull);
3446 return list;
3447 error:
3448 isl_aff_free(aff_nan);
3449 isl_local_space_free(ls);
3450 isl_basic_set_free(wrap_hull);
3451 isl_aff_list_free(list);
3452 return NULL;
3455 /* Add variables to info->bmap and info->tab corresponding to the elements
3456 * in "list" that are not set to NaN.
3457 * "extra_var" is the number of these elements.
3458 * "dim" is the offset in the variables of "tab" where we should
3459 * start considering the elements in "list".
3460 * When this function returns, the total number of variables in "tab"
3461 * is equal to "dim" plus the number of elements in "list".
3463 * The newly added existentially quantified variables are not given
3464 * an explicit representation because the corresponding div constraints
3465 * do not appear in info->bmap. These constraints are not added
3466 * to info->bmap because for internal consistency, they would need to
3467 * be added to info->tab as well, where they could combine with the equality
3468 * that is added later to result in constraints that do not hold
3469 * in the original input.
3471 static isl_stat add_sub_vars(struct isl_coalesce_info *info,
3472 __isl_keep isl_aff_list *list, int dim, int extra_var)
3474 int i, j, d;
3475 isl_size n;
3476 isl_space *space;
3478 space = isl_basic_map_get_space(info->bmap);
3479 info->bmap = isl_basic_map_cow(info->bmap);
3480 info->bmap = isl_basic_map_extend_space(info->bmap, space,
3481 extra_var, 0, 0);
3482 n = isl_aff_list_n_aff(list);
3483 if (!info->bmap || n < 0)
3484 return isl_stat_error;
3485 for (i = 0; i < n; ++i) {
3486 int is_nan;
3487 isl_aff *aff;
3489 aff = isl_aff_list_get_aff(list, i);
3490 is_nan = isl_aff_is_nan(aff);
3491 isl_aff_free(aff);
3492 if (is_nan < 0)
3493 return isl_stat_error;
3494 if (is_nan)
3495 continue;
3497 if (isl_tab_insert_var(info->tab, dim + i) < 0)
3498 return isl_stat_error;
3499 d = isl_basic_map_alloc_div(info->bmap);
3500 if (d < 0)
3501 return isl_stat_error;
3502 info->bmap = isl_basic_map_mark_div_unknown(info->bmap, d);
3503 for (j = d; j > i; --j)
3504 info->bmap = isl_basic_map_swap_div(info->bmap,
3505 j - 1, j);
3506 if (!info->bmap)
3507 return isl_stat_error;
3510 return isl_stat_ok;
3513 /* For each element in "list" that is not set to NaN, fix the corresponding
3514 * variable in "tab" to the purely affine expression defined by the element.
3515 * "dim" is the offset in the variables of "tab" where we should
3516 * start considering the elements in "list".
3518 * This function assumes that a sufficient number of rows and
3519 * elements in the constraint array are available in the tableau.
3521 static isl_stat add_sub_equalities(struct isl_tab *tab,
3522 __isl_keep isl_aff_list *list, int dim)
3524 int i;
3525 isl_size n;
3526 isl_ctx *ctx;
3527 isl_vec *sub;
3528 isl_aff *aff;
3530 n = isl_aff_list_n_aff(list);
3531 if (n < 0)
3532 return isl_stat_error;
3534 ctx = isl_tab_get_ctx(tab);
3535 sub = isl_vec_alloc(ctx, 1 + dim + n);
3536 if (!sub)
3537 return isl_stat_error;
3538 isl_seq_clr(sub->el + 1 + dim, n);
3540 for (i = 0; i < n; ++i) {
3541 aff = isl_aff_list_get_aff(list, i);
3542 if (!aff)
3543 goto error;
3544 if (isl_aff_is_nan(aff)) {
3545 isl_aff_free(aff);
3546 continue;
3548 isl_seq_cpy(sub->el, aff->v->el + 1, 1 + dim);
3549 isl_int_neg(sub->el[1 + dim + i], aff->v->el[0]);
3550 if (isl_tab_add_eq(tab, sub->el) < 0)
3551 goto error;
3552 isl_int_set_si(sub->el[1 + dim + i], 0);
3553 isl_aff_free(aff);
3556 isl_vec_free(sub);
3557 return isl_stat_ok;
3558 error:
3559 isl_aff_free(aff);
3560 isl_vec_free(sub);
3561 return isl_stat_error;
3564 /* Add variables to info->tab and info->bmap corresponding to the elements
3565 * in "list" that are not set to NaN. The value of the added variable
3566 * in info->tab is fixed to the purely affine expression defined by the element.
3567 * "dim" is the offset in the variables of info->tab where we should
3568 * start considering the elements in "list".
3569 * When this function returns, the total number of variables in info->tab
3570 * is equal to "dim" plus the number of elements in "list".
3572 static isl_stat add_subs(struct isl_coalesce_info *info,
3573 __isl_keep isl_aff_list *list, int dim)
3575 int extra_var;
3576 isl_size n;
3578 n = isl_aff_list_n_aff(list);
3579 if (n < 0)
3580 return isl_stat_error;
3582 extra_var = n - (info->tab->n_var - dim);
3584 if (isl_tab_extend_vars(info->tab, extra_var) < 0)
3585 return isl_stat_error;
3586 if (isl_tab_extend_cons(info->tab, 2 * extra_var) < 0)
3587 return isl_stat_error;
3588 if (add_sub_vars(info, list, dim, extra_var) < 0)
3589 return isl_stat_error;
3591 return add_sub_equalities(info->tab, list, dim);
3594 /* Coalesce basic map "j" into basic map "i" after adding the extra integer
3595 * divisions in "i" but not in "j" to basic map "j", with values
3596 * specified by "list". The total number of elements in "list"
3597 * is equal to the number of integer divisions in "i", while the number
3598 * of NaN elements in the list is equal to the number of integer divisions
3599 * in "j".
3601 * If no coalescing can be performed, then we need to revert basic map "j"
3602 * to its original state. We do the same if basic map "i" gets dropped
3603 * during the coalescing, even though this should not happen in practice
3604 * since we have already checked for "j" being a subset of "i"
3605 * before we reach this stage.
3607 static enum isl_change coalesce_with_subs(int i, int j,
3608 struct isl_coalesce_info *info, __isl_keep isl_aff_list *list)
3610 isl_basic_map *bmap_j;
3611 struct isl_tab_undo *snap;
3612 isl_size dim, n_div;
3613 enum isl_change change;
3615 bmap_j = isl_basic_map_copy(info[j].bmap);
3616 snap = isl_tab_snap(info[j].tab);
3618 dim = isl_basic_map_dim(bmap_j, isl_dim_all);
3619 n_div = isl_basic_map_dim(bmap_j, isl_dim_div);
3620 if (dim < 0 || n_div < 0)
3621 goto error;
3622 dim -= n_div;
3623 if (add_subs(&info[j], list, dim) < 0)
3624 goto error;
3626 change = coalesce_local_pair(i, j, info);
3627 if (change != isl_change_none && change != isl_change_drop_first) {
3628 isl_basic_map_free(bmap_j);
3629 } else {
3630 isl_basic_map_free(info[j].bmap);
3631 info[j].bmap = bmap_j;
3633 if (isl_tab_rollback(info[j].tab, snap) < 0)
3634 return isl_change_error;
3637 return change;
3638 error:
3639 isl_basic_map_free(bmap_j);
3640 return isl_change_error;
3643 /* Check if we can coalesce basic map "j" into basic map "i" after copying
3644 * those extra integer divisions in "i" that can be simplified away
3645 * using the extra equalities in "j".
3646 * All divs are assumed to be known and not contain any nested divs.
3648 * We first check if there are any extra equalities in "j" that we
3649 * can exploit. Then we check if every integer division in "i"
3650 * either already appears in "j" or can be simplified using the
3651 * extra equalities to a purely affine expression.
3652 * If these tests succeed, then we try to coalesce the two basic maps
3653 * by introducing extra dimensions in "j" corresponding to
3654 * the extra integer divsisions "i" fixed to the corresponding
3655 * purely affine expression.
3657 static enum isl_change check_coalesce_into_eq(int i, int j,
3658 struct isl_coalesce_info *info)
3660 isl_size n_div_i, n_div_j, n;
3661 isl_basic_map *hull_i, *hull_j;
3662 isl_bool equal, empty;
3663 isl_aff_list *list;
3664 enum isl_change change;
3666 n_div_i = isl_basic_map_dim(info[i].bmap, isl_dim_div);
3667 n_div_j = isl_basic_map_dim(info[j].bmap, isl_dim_div);
3668 if (n_div_i < 0 || n_div_j < 0)
3669 return isl_change_error;
3670 if (n_div_i <= n_div_j)
3671 return isl_change_none;
3672 if (info[j].bmap->n_eq == 0)
3673 return isl_change_none;
3675 hull_i = isl_basic_map_copy(info[i].bmap);
3676 hull_i = isl_basic_map_plain_affine_hull(hull_i);
3677 hull_j = isl_basic_map_copy(info[j].bmap);
3678 hull_j = isl_basic_map_plain_affine_hull(hull_j);
3680 hull_j = isl_basic_map_intersect(hull_j, isl_basic_map_copy(hull_i));
3681 equal = isl_basic_map_plain_is_equal(hull_i, hull_j);
3682 empty = isl_basic_map_plain_is_empty(hull_j);
3683 isl_basic_map_free(hull_i);
3685 if (equal < 0 || empty < 0)
3686 goto error;
3687 if (equal || empty) {
3688 isl_basic_map_free(hull_j);
3689 return isl_change_none;
3692 list = set_up_substitutions(info[i].bmap, info[j].bmap, hull_j);
3693 if (!list)
3694 return isl_change_error;
3695 n = isl_aff_list_n_aff(list);
3696 if (n < 0)
3697 change = isl_change_error;
3698 else if (n < n_div_i)
3699 change = isl_change_none;
3700 else
3701 change = coalesce_with_subs(i, j, info, list);
3703 isl_aff_list_free(list);
3705 return change;
3706 error:
3707 isl_basic_map_free(hull_j);
3708 return isl_change_error;
3711 /* Check if we can coalesce basic maps "i" and "j" after copying
3712 * those extra integer divisions in one of the basic maps that can
3713 * be simplified away using the extra equalities in the other basic map.
3714 * We require all divs to be known in both basic maps.
3715 * Furthermore, to simplify the comparison of div expressions,
3716 * we do not allow any nested integer divisions.
3718 static enum isl_change check_coalesce_eq(int i, int j,
3719 struct isl_coalesce_info *info)
3721 isl_bool known, nested;
3722 enum isl_change change;
3724 known = isl_basic_map_divs_known(info[i].bmap);
3725 if (known < 0 || !known)
3726 return known < 0 ? isl_change_error : isl_change_none;
3727 known = isl_basic_map_divs_known(info[j].bmap);
3728 if (known < 0 || !known)
3729 return known < 0 ? isl_change_error : isl_change_none;
3730 nested = has_nested_div(info[i].bmap);
3731 if (nested < 0 || nested)
3732 return nested < 0 ? isl_change_error : isl_change_none;
3733 nested = has_nested_div(info[j].bmap);
3734 if (nested < 0 || nested)
3735 return nested < 0 ? isl_change_error : isl_change_none;
3737 change = check_coalesce_into_eq(i, j, info);
3738 if (change != isl_change_none)
3739 return change;
3740 change = check_coalesce_into_eq(j, i, info);
3741 if (change != isl_change_none)
3742 return invert_change(change);
3744 return isl_change_none;
3747 /* Check if the union of the given pair of basic maps
3748 * can be represented by a single basic map.
3749 * If so, replace the pair by the single basic map and return
3750 * isl_change_drop_first, isl_change_drop_second or isl_change_fuse.
3751 * Otherwise, return isl_change_none.
3753 * We first check if the two basic maps live in the same local space,
3754 * after aligning the divs that differ by only an integer constant.
3755 * If so, we do the complete check. Otherwise, we check if they have
3756 * the same number of integer divisions and can be coalesced, if one is
3757 * an obvious subset of the other or if the extra integer divisions
3758 * of one basic map can be simplified away using the extra equalities
3759 * of the other basic map.
3761 * Note that trying to coalesce pairs of disjuncts with the same
3762 * number, but different local variables may drop the explicit
3763 * representation of some of these local variables.
3764 * This operation is therefore not performed when
3765 * the "coalesce_preserve_locals" option is set.
3767 static enum isl_change coalesce_pair(int i, int j,
3768 struct isl_coalesce_info *info)
3770 int preserve;
3771 isl_bool same;
3772 enum isl_change change;
3773 isl_ctx *ctx;
3775 if (harmonize_divs(&info[i], &info[j]) < 0)
3776 return isl_change_error;
3777 same = same_divs(info[i].bmap, info[j].bmap);
3778 if (same < 0)
3779 return isl_change_error;
3780 if (same)
3781 return coalesce_local_pair(i, j, info);
3783 ctx = isl_basic_map_get_ctx(info[i].bmap);
3784 preserve = isl_options_get_coalesce_preserve_locals(ctx);
3785 if (!preserve && info[i].bmap->n_div == info[j].bmap->n_div) {
3786 change = coalesce_local_pair(i, j, info);
3787 if (change != isl_change_none)
3788 return change;
3791 change = coalesce_divs(i, j, info);
3792 if (change != isl_change_none)
3793 return change;
3795 return check_coalesce_eq(i, j, info);
3798 /* Return the maximum of "a" and "b".
3800 static int isl_max(int a, int b)
3802 return a > b ? a : b;
3805 /* Pairwise coalesce the basic maps in the range [start1, end1[ of "info"
3806 * with those in the range [start2, end2[, skipping basic maps
3807 * that have been removed (either before or within this function).
3809 * For each basic map i in the first range, we check if it can be coalesced
3810 * with respect to any previously considered basic map j in the second range.
3811 * If i gets dropped (because it was a subset of some j), then
3812 * we can move on to the next basic map.
3813 * If j gets dropped, we need to continue checking against the other
3814 * previously considered basic maps.
3815 * If the two basic maps got fused, then we recheck the fused basic map
3816 * against the previously considered basic maps, starting at i + 1
3817 * (even if start2 is greater than i + 1).
3819 static int coalesce_range(isl_ctx *ctx, struct isl_coalesce_info *info,
3820 int start1, int end1, int start2, int end2)
3822 int i, j;
3824 for (i = end1 - 1; i >= start1; --i) {
3825 if (info[i].removed)
3826 continue;
3827 for (j = isl_max(i + 1, start2); j < end2; ++j) {
3828 enum isl_change changed;
3830 if (info[j].removed)
3831 continue;
3832 if (info[i].removed)
3833 isl_die(ctx, isl_error_internal,
3834 "basic map unexpectedly removed",
3835 return -1);
3836 changed = coalesce_pair(i, j, info);
3837 switch (changed) {
3838 case isl_change_error:
3839 return -1;
3840 case isl_change_none:
3841 case isl_change_drop_second:
3842 continue;
3843 case isl_change_drop_first:
3844 j = end2;
3845 break;
3846 case isl_change_fuse:
3847 j = i;
3848 break;
3853 return 0;
3856 /* Pairwise coalesce the basic maps described by the "n" elements of "info".
3858 * We consider groups of basic maps that live in the same apparent
3859 * affine hull and we first coalesce within such a group before we
3860 * coalesce the elements in the group with elements of previously
3861 * considered groups. If a fuse happens during the second phase,
3862 * then we also reconsider the elements within the group.
3864 static int coalesce(isl_ctx *ctx, int n, struct isl_coalesce_info *info)
3866 int start, end;
3868 for (end = n; end > 0; end = start) {
3869 start = end - 1;
3870 while (start >= 1 &&
3871 info[start - 1].hull_hash == info[start].hull_hash)
3872 start--;
3873 if (coalesce_range(ctx, info, start, end, start, end) < 0)
3874 return -1;
3875 if (coalesce_range(ctx, info, start, end, end, n) < 0)
3876 return -1;
3879 return 0;
3882 /* Update the basic maps in "map" based on the information in "info".
3883 * In particular, remove the basic maps that have been marked removed and
3884 * update the others based on the information in the corresponding tableau.
3885 * Since we detected implicit equalities without calling
3886 * isl_basic_map_gauss, we need to do it now.
3887 * Also call isl_basic_map_simplify if we may have lost the definition
3888 * of one or more integer divisions.
3890 static __isl_give isl_map *update_basic_maps(__isl_take isl_map *map,
3891 int n, struct isl_coalesce_info *info)
3893 int i;
3895 if (!map)
3896 return NULL;
3898 for (i = n - 1; i >= 0; --i) {
3899 if (info[i].removed) {
3900 isl_basic_map_free(map->p[i]);
3901 if (i != map->n - 1)
3902 map->p[i] = map->p[map->n - 1];
3903 map->n--;
3904 continue;
3907 info[i].bmap = isl_basic_map_update_from_tab(info[i].bmap,
3908 info[i].tab);
3909 info[i].bmap = isl_basic_map_gauss(info[i].bmap, NULL);
3910 if (info[i].simplify)
3911 info[i].bmap = isl_basic_map_simplify(info[i].bmap);
3912 info[i].bmap = isl_basic_map_finalize(info[i].bmap);
3913 if (!info[i].bmap)
3914 return isl_map_free(map);
3915 ISL_F_SET(info[i].bmap, ISL_BASIC_MAP_NO_IMPLICIT);
3916 ISL_F_SET(info[i].bmap, ISL_BASIC_MAP_NO_REDUNDANT);
3917 isl_basic_map_free(map->p[i]);
3918 map->p[i] = info[i].bmap;
3919 info[i].bmap = NULL;
3922 return map;
3925 /* For each pair of basic maps in the map, check if the union of the two
3926 * can be represented by a single basic map.
3927 * If so, replace the pair by the single basic map and start over.
3929 * We factor out any (hidden) common factor from the constraint
3930 * coefficients to improve the detection of adjacent constraints.
3932 * Since we are constructing the tableaus of the basic maps anyway,
3933 * we exploit them to detect implicit equalities and redundant constraints.
3934 * This also helps the coalescing as it can ignore the redundant constraints.
3935 * In order to avoid confusion, we make all implicit equalities explicit
3936 * in the basic maps. We don't call isl_basic_map_gauss, though,
3937 * as that may affect the number of constraints.
3938 * This means that we have to call isl_basic_map_gauss at the end
3939 * of the computation (in update_basic_maps and in clear) to ensure that
3940 * the basic maps are not left in an unexpected state.
3941 * For each basic map, we also compute the hash of the apparent affine hull
3942 * for use in coalesce.
3944 __isl_give isl_map *isl_map_coalesce(__isl_take isl_map *map)
3946 int i;
3947 unsigned n;
3948 isl_ctx *ctx;
3949 struct isl_coalesce_info *info = NULL;
3951 map = isl_map_remove_empty_parts(map);
3952 if (!map)
3953 return NULL;
3955 if (map->n <= 1)
3956 return map;
3958 ctx = isl_map_get_ctx(map);
3959 map = isl_map_sort_divs(map);
3960 map = isl_map_cow(map);
3962 if (!map)
3963 return NULL;
3965 n = map->n;
3967 info = isl_calloc_array(map->ctx, struct isl_coalesce_info, n);
3968 if (!info)
3969 goto error;
3971 for (i = 0; i < map->n; ++i) {
3972 map->p[i] = isl_basic_map_reduce_coefficients(map->p[i]);
3973 if (!map->p[i])
3974 goto error;
3975 info[i].bmap = isl_basic_map_copy(map->p[i]);
3976 info[i].tab = isl_tab_from_basic_map(info[i].bmap, 0);
3977 if (!info[i].tab)
3978 goto error;
3979 if (!ISL_F_ISSET(info[i].bmap, ISL_BASIC_MAP_NO_IMPLICIT))
3980 if (isl_tab_detect_implicit_equalities(info[i].tab) < 0)
3981 goto error;
3982 info[i].bmap = isl_tab_make_equalities_explicit(info[i].tab,
3983 info[i].bmap);
3984 if (!info[i].bmap)
3985 goto error;
3986 if (!ISL_F_ISSET(info[i].bmap, ISL_BASIC_MAP_NO_REDUNDANT))
3987 if (isl_tab_detect_redundant(info[i].tab) < 0)
3988 goto error;
3989 if (coalesce_info_set_hull_hash(&info[i]) < 0)
3990 goto error;
3992 for (i = map->n - 1; i >= 0; --i)
3993 if (info[i].tab->empty)
3994 drop(&info[i]);
3996 if (coalesce(ctx, n, info) < 0)
3997 goto error;
3999 map = update_basic_maps(map, n, info);
4001 clear_coalesce_info(n, info);
4003 return map;
4004 error:
4005 clear_coalesce_info(n, info);
4006 isl_map_free(map);
4007 return NULL;
4010 /* For each pair of basic sets in the set, check if the union of the two
4011 * can be represented by a single basic set.
4012 * If so, replace the pair by the single basic set and start over.
4014 struct isl_set *isl_set_coalesce(struct isl_set *set)
4016 return set_from_map(isl_map_coalesce(set_to_map(set)));