isl_basic_{map,set}_free_inequality: return modified result
[isl.git] / isl_coalesce.c
blobdc0a374b73b1a0bd4b24d7fc8949fdba62af963f
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".
351 static void clear(struct isl_coalesce_info *info)
353 info->bmap = isl_basic_map_free(info->bmap);
354 isl_tab_free(info->tab);
355 info->tab = NULL;
358 /* Drop the basic map represented by "info".
359 * That is, clear the memory associated to the entry and
360 * mark it as having been removed.
362 static void drop(struct isl_coalesce_info *info)
364 clear(info);
365 info->removed = 1;
368 /* Exchange the information in "info1" with that in "info2".
370 static void exchange(struct isl_coalesce_info *info1,
371 struct isl_coalesce_info *info2)
373 struct isl_coalesce_info info;
375 info = *info1;
376 *info1 = *info2;
377 *info2 = info;
380 /* This type represents the kind of change that has been performed
381 * while trying to coalesce two basic maps.
383 * isl_change_none: nothing was changed
384 * isl_change_drop_first: the first basic map was removed
385 * isl_change_drop_second: the second basic map was removed
386 * isl_change_fuse: the two basic maps were replaced by a new basic map.
388 enum isl_change {
389 isl_change_error = -1,
390 isl_change_none = 0,
391 isl_change_drop_first,
392 isl_change_drop_second,
393 isl_change_fuse,
396 /* Update "change" based on an interchange of the first and the second
397 * basic map. That is, interchange isl_change_drop_first and
398 * isl_change_drop_second.
400 static enum isl_change invert_change(enum isl_change change)
402 switch (change) {
403 case isl_change_error:
404 return isl_change_error;
405 case isl_change_none:
406 return isl_change_none;
407 case isl_change_drop_first:
408 return isl_change_drop_second;
409 case isl_change_drop_second:
410 return isl_change_drop_first;
411 case isl_change_fuse:
412 return isl_change_fuse;
415 return isl_change_error;
418 /* Add the valid constraints of the basic map represented by "info"
419 * to "bmap". "len" is the size of the constraints.
420 * If only one of the pair of inequalities that make up an equality
421 * is valid, then add that inequality.
423 static __isl_give isl_basic_map *add_valid_constraints(
424 __isl_take isl_basic_map *bmap, struct isl_coalesce_info *info,
425 unsigned len)
427 int k, l;
429 if (!bmap)
430 return NULL;
432 for (k = 0; k < info->bmap->n_eq; ++k) {
433 if (info->eq[2 * k] == STATUS_VALID &&
434 info->eq[2 * k + 1] == STATUS_VALID) {
435 l = isl_basic_map_alloc_equality(bmap);
436 if (l < 0)
437 return isl_basic_map_free(bmap);
438 isl_seq_cpy(bmap->eq[l], info->bmap->eq[k], len);
439 } else if (info->eq[2 * k] == STATUS_VALID) {
440 l = isl_basic_map_alloc_inequality(bmap);
441 if (l < 0)
442 return isl_basic_map_free(bmap);
443 isl_seq_neg(bmap->ineq[l], info->bmap->eq[k], len);
444 } else if (info->eq[2 * k + 1] == STATUS_VALID) {
445 l = isl_basic_map_alloc_inequality(bmap);
446 if (l < 0)
447 return isl_basic_map_free(bmap);
448 isl_seq_cpy(bmap->ineq[l], info->bmap->eq[k], len);
452 for (k = 0; k < info->bmap->n_ineq; ++k) {
453 if (info->ineq[k] != STATUS_VALID)
454 continue;
455 l = isl_basic_map_alloc_inequality(bmap);
456 if (l < 0)
457 return isl_basic_map_free(bmap);
458 isl_seq_cpy(bmap->ineq[l], info->bmap->ineq[k], len);
461 return bmap;
464 /* Is "bmap" defined by a number of (non-redundant) constraints that
465 * is greater than the number of constraints of basic maps i and j combined?
466 * Equalities are counted as two inequalities.
468 static int number_of_constraints_increases(int i, int j,
469 struct isl_coalesce_info *info,
470 __isl_keep isl_basic_map *bmap, struct isl_tab *tab)
472 int k, n_old, n_new;
474 n_old = 2 * info[i].bmap->n_eq + info[i].bmap->n_ineq;
475 n_old += 2 * info[j].bmap->n_eq + info[j].bmap->n_ineq;
477 n_new = 2 * bmap->n_eq;
478 for (k = 0; k < bmap->n_ineq; ++k)
479 if (!isl_tab_is_redundant(tab, bmap->n_eq + k))
480 ++n_new;
482 return n_new > n_old;
485 /* Replace the pair of basic maps i and j by the basic map bounded
486 * by the valid constraints in both basic maps and the constraints
487 * in extra (if not NULL).
488 * Place the fused basic map in the position that is the smallest of i and j.
490 * If "detect_equalities" is set, then look for equalities encoded
491 * as pairs of inequalities.
492 * If "check_number" is set, then the original basic maps are only
493 * replaced if the total number of constraints does not increase.
494 * While the number of integer divisions in the two basic maps
495 * is assumed to be the same, the actual definitions may be different.
496 * We only copy the definition from one of the basic map if it is
497 * the same as that of the other basic map. Otherwise, we mark
498 * the integer division as unknown and simplify the basic map
499 * in an attempt to recover the integer division definition.
501 static enum isl_change fuse(int i, int j, struct isl_coalesce_info *info,
502 __isl_keep isl_mat *extra, int detect_equalities, int check_number)
504 int k, l;
505 struct isl_basic_map *fused = NULL;
506 struct isl_tab *fused_tab = NULL;
507 isl_size total = isl_basic_map_dim(info[i].bmap, isl_dim_all);
508 unsigned extra_rows = extra ? extra->n_row : 0;
509 unsigned n_eq, n_ineq;
510 int simplify = 0;
512 if (total < 0)
513 return isl_change_error;
514 if (j < i)
515 return fuse(j, i, info, extra, detect_equalities, check_number);
517 n_eq = info[i].bmap->n_eq + info[j].bmap->n_eq;
518 n_ineq = info[i].bmap->n_ineq + info[j].bmap->n_ineq;
519 fused = isl_basic_map_alloc_space(isl_space_copy(info[i].bmap->dim),
520 info[i].bmap->n_div, n_eq, n_eq + n_ineq + extra_rows);
521 fused = add_valid_constraints(fused, &info[i], 1 + total);
522 fused = add_valid_constraints(fused, &info[j], 1 + total);
523 if (!fused)
524 goto error;
525 if (ISL_F_ISSET(info[i].bmap, ISL_BASIC_MAP_RATIONAL) &&
526 ISL_F_ISSET(info[j].bmap, ISL_BASIC_MAP_RATIONAL))
527 ISL_F_SET(fused, ISL_BASIC_MAP_RATIONAL);
529 for (k = 0; k < info[i].bmap->n_div; ++k) {
530 int l = isl_basic_map_alloc_div(fused);
531 if (l < 0)
532 goto error;
533 if (isl_seq_eq(info[i].bmap->div[k], info[j].bmap->div[k],
534 1 + 1 + total)) {
535 isl_seq_cpy(fused->div[l], info[i].bmap->div[k],
536 1 + 1 + total);
537 } else {
538 isl_int_set_si(fused->div[l][0], 0);
539 simplify = 1;
543 for (k = 0; k < extra_rows; ++k) {
544 l = isl_basic_map_alloc_inequality(fused);
545 if (l < 0)
546 goto error;
547 isl_seq_cpy(fused->ineq[l], extra->row[k], 1 + total);
550 if (detect_equalities)
551 fused = isl_basic_map_detect_inequality_pairs(fused, NULL);
552 fused = isl_basic_map_gauss(fused, NULL);
553 if (simplify || info[j].simplify) {
554 fused = isl_basic_map_simplify(fused);
555 info[i].simplify = 0;
557 fused = isl_basic_map_finalize(fused);
559 fused_tab = isl_tab_from_basic_map(fused, 0);
560 if (isl_tab_detect_redundant(fused_tab) < 0)
561 goto error;
563 if (check_number &&
564 number_of_constraints_increases(i, j, info, fused, fused_tab)) {
565 isl_tab_free(fused_tab);
566 isl_basic_map_free(fused);
567 return isl_change_none;
570 clear(&info[i]);
571 info[i].bmap = fused;
572 info[i].tab = fused_tab;
573 drop(&info[j]);
575 return isl_change_fuse;
576 error:
577 isl_tab_free(fused_tab);
578 isl_basic_map_free(fused);
579 return isl_change_error;
582 /* Given a pair of basic maps i and j such that all constraints are either
583 * "valid" or "cut", check if the facets corresponding to the "cut"
584 * constraints of i lie entirely within basic map j.
585 * If so, replace the pair by the basic map consisting of the valid
586 * constraints in both basic maps.
587 * Checking whether the facet lies entirely within basic map j
588 * is performed by checking whether the constraints of basic map j
589 * are valid for the facet. These tests are performed on a rational
590 * tableau to avoid the theoretical possibility that a constraint
591 * that was considered to be a cut constraint for the entire basic map i
592 * happens to be considered to be a valid constraint for the facet,
593 * even though it cuts off the same rational points.
595 * To see that we are not introducing any extra points, call the
596 * two basic maps A and B and the resulting map U and let x
597 * be an element of U \setminus ( A \cup B ).
598 * A line connecting x with an element of A \cup B meets a facet F
599 * of either A or B. Assume it is a facet of B and let c_1 be
600 * the corresponding facet constraint. We have c_1(x) < 0 and
601 * so c_1 is a cut constraint. This implies that there is some
602 * (possibly rational) point x' satisfying the constraints of A
603 * and the opposite of c_1 as otherwise c_1 would have been marked
604 * valid for A. The line connecting x and x' meets a facet of A
605 * in a (possibly rational) point that also violates c_1, but this
606 * is impossible since all cut constraints of B are valid for all
607 * cut facets of A.
608 * In case F is a facet of A rather than B, then we can apply the
609 * above reasoning to find a facet of B separating x from A \cup B first.
611 static enum isl_change check_facets(int i, int j,
612 struct isl_coalesce_info *info)
614 int k, l;
615 struct isl_tab_undo *snap, *snap2;
616 unsigned n_eq = info[i].bmap->n_eq;
618 snap = isl_tab_snap(info[i].tab);
619 if (isl_tab_mark_rational(info[i].tab) < 0)
620 return isl_change_error;
621 snap2 = isl_tab_snap(info[i].tab);
623 for (k = 0; k < info[i].bmap->n_ineq; ++k) {
624 if (info[i].ineq[k] != STATUS_CUT)
625 continue;
626 if (isl_tab_select_facet(info[i].tab, n_eq + k) < 0)
627 return isl_change_error;
628 for (l = 0; l < info[j].bmap->n_ineq; ++l) {
629 int stat;
630 if (info[j].ineq[l] != STATUS_CUT)
631 continue;
632 stat = status_in(info[j].bmap->ineq[l], info[i].tab);
633 if (stat < 0)
634 return isl_change_error;
635 if (stat != STATUS_VALID)
636 break;
638 if (isl_tab_rollback(info[i].tab, snap2) < 0)
639 return isl_change_error;
640 if (l < info[j].bmap->n_ineq)
641 break;
644 if (k < info[i].bmap->n_ineq) {
645 if (isl_tab_rollback(info[i].tab, snap) < 0)
646 return isl_change_error;
647 return isl_change_none;
649 return fuse(i, j, info, NULL, 0, 0);
652 /* Check if info->bmap contains the basic map represented
653 * by the tableau "tab".
654 * For each equality, we check both the constraint itself
655 * (as an inequality) and its negation. Make sure the
656 * equality is returned to its original state before returning.
658 static isl_bool contains(struct isl_coalesce_info *info, struct isl_tab *tab)
660 int k;
661 isl_size dim;
662 isl_basic_map *bmap = info->bmap;
664 dim = isl_basic_map_dim(bmap, isl_dim_all);
665 if (dim < 0)
666 return isl_bool_error;
667 for (k = 0; k < bmap->n_eq; ++k) {
668 int stat;
669 isl_seq_neg(bmap->eq[k], bmap->eq[k], 1 + dim);
670 stat = status_in(bmap->eq[k], tab);
671 isl_seq_neg(bmap->eq[k], bmap->eq[k], 1 + dim);
672 if (stat < 0)
673 return isl_bool_error;
674 if (stat != STATUS_VALID)
675 return isl_bool_false;
676 stat = status_in(bmap->eq[k], tab);
677 if (stat < 0)
678 return isl_bool_error;
679 if (stat != STATUS_VALID)
680 return isl_bool_false;
683 for (k = 0; k < bmap->n_ineq; ++k) {
684 int stat;
685 if (info->ineq[k] == STATUS_REDUNDANT)
686 continue;
687 stat = status_in(bmap->ineq[k], tab);
688 if (stat < 0)
689 return isl_bool_error;
690 if (stat != STATUS_VALID)
691 return isl_bool_false;
693 return isl_bool_true;
696 /* Basic map "i" has an inequality (say "k") that is adjacent
697 * to some inequality of basic map "j". All the other inequalities
698 * are valid for "j".
699 * Check if basic map "j" forms an extension of basic map "i".
701 * Note that this function is only called if some of the equalities or
702 * inequalities of basic map "j" do cut basic map "i". The function is
703 * correct even if there are no such cut constraints, but in that case
704 * the additional checks performed by this function are overkill.
706 * In particular, we replace constraint k, say f >= 0, by constraint
707 * f <= -1, add the inequalities of "j" that are valid for "i"
708 * and check if the result is a subset of basic map "j".
709 * To improve the chances of the subset relation being detected,
710 * any variable that only attains a single integer value
711 * in the tableau of "i" is first fixed to that value.
712 * If the result is a subset, then we know that this result is exactly equal
713 * to basic map "j" since all its constraints are valid for basic map "j".
714 * By combining the valid constraints of "i" (all equalities and all
715 * inequalities except "k") and the valid constraints of "j" we therefore
716 * obtain a basic map that is equal to their union.
717 * In this case, there is no need to perform a rollback of the tableau
718 * since it is going to be destroyed in fuse().
721 * |\__ |\__
722 * | \__ | \__
723 * | \_ => | \__
724 * |_______| _ |_________\
727 * |\ |\
728 * | \ | \
729 * | \ | \
730 * | | | \
731 * | ||\ => | \
732 * | || \ | \
733 * | || | | |
734 * |__||_/ |_____/
736 static enum isl_change is_adj_ineq_extension(int i, int j,
737 struct isl_coalesce_info *info)
739 int k;
740 struct isl_tab_undo *snap;
741 unsigned n_eq = info[i].bmap->n_eq;
742 isl_size total = isl_basic_map_dim(info[i].bmap, isl_dim_all);
743 isl_stat r;
744 isl_bool super;
746 if (total < 0)
747 return isl_change_error;
748 if (isl_tab_extend_cons(info[i].tab, 1 + info[j].bmap->n_ineq) < 0)
749 return isl_change_error;
751 k = find_ineq(&info[i], STATUS_ADJ_INEQ);
752 if (k < 0)
753 isl_die(isl_basic_map_get_ctx(info[i].bmap), isl_error_internal,
754 "info[i].ineq should have exactly one STATUS_ADJ_INEQ",
755 return isl_change_error);
757 snap = isl_tab_snap(info[i].tab);
759 if (isl_tab_unrestrict(info[i].tab, n_eq + k) < 0)
760 return isl_change_error;
762 isl_seq_neg(info[i].bmap->ineq[k], info[i].bmap->ineq[k], 1 + total);
763 isl_int_sub_ui(info[i].bmap->ineq[k][0], info[i].bmap->ineq[k][0], 1);
764 r = isl_tab_add_ineq(info[i].tab, info[i].bmap->ineq[k]);
765 isl_seq_neg(info[i].bmap->ineq[k], info[i].bmap->ineq[k], 1 + total);
766 isl_int_sub_ui(info[i].bmap->ineq[k][0], info[i].bmap->ineq[k][0], 1);
767 if (r < 0)
768 return isl_change_error;
770 for (k = 0; k < info[j].bmap->n_ineq; ++k) {
771 if (info[j].ineq[k] != STATUS_VALID)
772 continue;
773 if (isl_tab_add_ineq(info[i].tab, info[j].bmap->ineq[k]) < 0)
774 return isl_change_error;
776 if (isl_tab_detect_constants(info[i].tab) < 0)
777 return isl_change_error;
779 super = contains(&info[j], info[i].tab);
780 if (super < 0)
781 return isl_change_error;
782 if (super)
783 return fuse(i, j, info, NULL, 0, 0);
785 if (isl_tab_rollback(info[i].tab, snap) < 0)
786 return isl_change_error;
788 return isl_change_none;
792 /* Both basic maps have at least one inequality with and adjacent
793 * (but opposite) inequality in the other basic map.
794 * Check that there are no cut constraints and that there is only
795 * a single pair of adjacent inequalities.
796 * If so, we can replace the pair by a single basic map described
797 * by all but the pair of adjacent inequalities.
798 * Any additional points introduced lie strictly between the two
799 * adjacent hyperplanes and can therefore be integral.
801 * ____ _____
802 * / ||\ / \
803 * / || \ / \
804 * \ || \ => \ \
805 * \ || / \ /
806 * \___||_/ \_____/
808 * The test for a single pair of adjancent inequalities is important
809 * for avoiding the combination of two basic maps like the following
811 * /|
812 * / |
813 * /__|
814 * _____
815 * | |
816 * | |
817 * |___|
819 * If there are some cut constraints on one side, then we may
820 * still be able to fuse the two basic maps, but we need to perform
821 * some additional checks in is_adj_ineq_extension.
823 static enum isl_change check_adj_ineq(int i, int j,
824 struct isl_coalesce_info *info)
826 int count_i, count_j;
827 int cut_i, cut_j;
829 count_i = count_ineq(&info[i], STATUS_ADJ_INEQ);
830 count_j = count_ineq(&info[j], STATUS_ADJ_INEQ);
832 if (count_i != 1 && count_j != 1)
833 return isl_change_none;
835 cut_i = any_eq(&info[i], STATUS_CUT) || any_ineq(&info[i], STATUS_CUT);
836 cut_j = any_eq(&info[j], STATUS_CUT) || any_ineq(&info[j], STATUS_CUT);
838 if (!cut_i && !cut_j && count_i == 1 && count_j == 1)
839 return fuse(i, j, info, NULL, 0, 0);
841 if (count_i == 1 && !cut_i)
842 return is_adj_ineq_extension(i, j, info);
844 if (count_j == 1 && !cut_j)
845 return is_adj_ineq_extension(j, i, info);
847 return isl_change_none;
850 /* Given an affine transformation matrix "T", does row "row" represent
851 * anything other than a unit vector (possibly shifted by a constant)
852 * that is not involved in any of the other rows?
854 * That is, if a constraint involves the variable corresponding to
855 * the row, then could its preimage by "T" have any coefficients
856 * that are different from those in the original constraint?
858 static int not_unique_unit_row(__isl_keep isl_mat *T, int row)
860 int i, j;
861 int len = T->n_col - 1;
863 i = isl_seq_first_non_zero(T->row[row] + 1, len);
864 if (i < 0)
865 return 1;
866 if (!isl_int_is_one(T->row[row][1 + i]) &&
867 !isl_int_is_negone(T->row[row][1 + i]))
868 return 1;
870 j = isl_seq_first_non_zero(T->row[row] + 1 + i + 1, len - (i + 1));
871 if (j >= 0)
872 return 1;
874 for (j = 1; j < T->n_row; ++j) {
875 if (j == row)
876 continue;
877 if (!isl_int_is_zero(T->row[j][1 + i]))
878 return 1;
881 return 0;
884 /* Does inequality constraint "ineq" of "bmap" involve any of
885 * the variables marked in "affected"?
886 * "total" is the total number of variables, i.e., the number
887 * of entries in "affected".
889 static isl_bool is_affected(__isl_keep isl_basic_map *bmap, int ineq,
890 int *affected, int total)
892 int i;
894 for (i = 0; i < total; ++i) {
895 if (!affected[i])
896 continue;
897 if (!isl_int_is_zero(bmap->ineq[ineq][1 + i]))
898 return isl_bool_true;
901 return isl_bool_false;
904 /* Given the compressed version of inequality constraint "ineq"
905 * of info->bmap in "v", check if the constraint can be tightened,
906 * where the compression is based on an equality constraint valid
907 * for info->tab.
908 * If so, add the tightened version of the inequality constraint
909 * to info->tab. "v" may be modified by this function.
911 * That is, if the compressed constraint is of the form
913 * m f() + c >= 0
915 * with 0 < c < m, then it is equivalent to
917 * f() >= 0
919 * This means that c can also be subtracted from the original,
920 * uncompressed constraint without affecting the integer points
921 * in info->tab. Add this tightened constraint as an extra row
922 * to info->tab to make this information explicitly available.
924 static __isl_give isl_vec *try_tightening(struct isl_coalesce_info *info,
925 int ineq, __isl_take isl_vec *v)
927 isl_ctx *ctx;
928 isl_stat r;
930 if (!v)
931 return NULL;
933 ctx = isl_vec_get_ctx(v);
934 isl_seq_gcd(v->el + 1, v->size - 1, &ctx->normalize_gcd);
935 if (isl_int_is_zero(ctx->normalize_gcd) ||
936 isl_int_is_one(ctx->normalize_gcd)) {
937 return v;
940 v = isl_vec_cow(v);
941 if (!v)
942 return NULL;
944 isl_int_fdiv_r(v->el[0], v->el[0], ctx->normalize_gcd);
945 if (isl_int_is_zero(v->el[0]))
946 return v;
948 if (isl_tab_extend_cons(info->tab, 1) < 0)
949 return isl_vec_free(v);
951 isl_int_sub(info->bmap->ineq[ineq][0],
952 info->bmap->ineq[ineq][0], v->el[0]);
953 r = isl_tab_add_ineq(info->tab, info->bmap->ineq[ineq]);
954 isl_int_add(info->bmap->ineq[ineq][0],
955 info->bmap->ineq[ineq][0], v->el[0]);
957 if (r < 0)
958 return isl_vec_free(v);
960 return v;
963 /* Tighten the (non-redundant) constraints on the facet represented
964 * by info->tab.
965 * In particular, on input, info->tab represents the result
966 * of relaxing the "n" inequality constraints of info->bmap in "relaxed"
967 * by one, i.e., replacing f_i >= 0 by f_i + 1 >= 0, and then
968 * replacing the one at index "l" by the corresponding equality,
969 * i.e., f_k + 1 = 0, with k = relaxed[l].
971 * Compute a variable compression from the equality constraint f_k + 1 = 0
972 * and use it to tighten the other constraints of info->bmap
973 * (that is, all constraints that have not been relaxed),
974 * updating info->tab (and leaving info->bmap untouched).
975 * The compression handles essentially two cases, one where a variable
976 * is assigned a fixed value and can therefore be eliminated, and one
977 * where one variable is a shifted multiple of some other variable and
978 * can therefore be replaced by that multiple.
979 * Gaussian elimination would also work for the first case, but for
980 * the second case, the effectiveness would depend on the order
981 * of the variables.
982 * After compression, some of the constraints may have coefficients
983 * with a common divisor. If this divisor does not divide the constant
984 * term, then the constraint can be tightened.
985 * The tightening is performed on the tableau info->tab by introducing
986 * extra (temporary) constraints.
988 * Only constraints that are possibly affected by the compression are
989 * considered. In particular, if the constraint only involves variables
990 * that are directly mapped to a distinct set of other variables, then
991 * no common divisor can be introduced and no tightening can occur.
993 * It is important to only consider the non-redundant constraints
994 * since the facet constraint has been relaxed prior to the call
995 * to this function, meaning that the constraints that were redundant
996 * prior to the relaxation may no longer be redundant.
997 * These constraints will be ignored in the fused result, so
998 * the fusion detection should not exploit them.
1000 static isl_stat tighten_on_relaxed_facet(struct isl_coalesce_info *info,
1001 int n, int *relaxed, int l)
1003 isl_size total;
1004 isl_ctx *ctx;
1005 isl_vec *v = NULL;
1006 isl_mat *T;
1007 int i;
1008 int k;
1009 int *affected;
1011 k = relaxed[l];
1012 ctx = isl_basic_map_get_ctx(info->bmap);
1013 total = isl_basic_map_dim(info->bmap, isl_dim_all);
1014 if (total < 0)
1015 return isl_stat_error;
1016 isl_int_add_ui(info->bmap->ineq[k][0], info->bmap->ineq[k][0], 1);
1017 T = isl_mat_sub_alloc6(ctx, info->bmap->ineq, k, 1, 0, 1 + total);
1018 T = isl_mat_variable_compression(T, NULL);
1019 isl_int_sub_ui(info->bmap->ineq[k][0], info->bmap->ineq[k][0], 1);
1020 if (!T)
1021 return isl_stat_error;
1022 if (T->n_col == 0) {
1023 isl_mat_free(T);
1024 return isl_stat_ok;
1027 affected = isl_alloc_array(ctx, int, total);
1028 if (!affected)
1029 goto error;
1031 for (i = 0; i < total; ++i)
1032 affected[i] = not_unique_unit_row(T, 1 + i);
1034 for (i = 0; i < info->bmap->n_ineq; ++i) {
1035 isl_bool handle;
1036 if (any(relaxed, n, i))
1037 continue;
1038 if (info->ineq[i] == STATUS_REDUNDANT)
1039 continue;
1040 handle = is_affected(info->bmap, i, affected, total);
1041 if (handle < 0)
1042 goto error;
1043 if (!handle)
1044 continue;
1045 v = isl_vec_alloc(ctx, 1 + total);
1046 if (!v)
1047 goto error;
1048 isl_seq_cpy(v->el, info->bmap->ineq[i], 1 + total);
1049 v = isl_vec_mat_product(v, isl_mat_copy(T));
1050 v = try_tightening(info, i, v);
1051 isl_vec_free(v);
1052 if (!v)
1053 goto error;
1056 isl_mat_free(T);
1057 free(affected);
1058 return isl_stat_ok;
1059 error:
1060 isl_mat_free(T);
1061 free(affected);
1062 return isl_stat_error;
1065 /* Replace the basic maps "i" and "j" by an extension of "i"
1066 * along the "n" inequality constraints in "relax" by one.
1067 * The tableau info[i].tab has already been extended.
1068 * Extend info[i].bmap accordingly by relaxing all constraints in "relax"
1069 * by one.
1070 * Each integer division that does not have exactly the same
1071 * definition in "i" and "j" is marked unknown and the basic map
1072 * is scheduled to be simplified in an attempt to recover
1073 * the integer division definition.
1074 * Place the extension in the position that is the smallest of i and j.
1076 static enum isl_change extend(int i, int j, int n, int *relax,
1077 struct isl_coalesce_info *info)
1079 int l;
1080 isl_size total;
1082 info[i].bmap = isl_basic_map_cow(info[i].bmap);
1083 total = isl_basic_map_dim(info[i].bmap, isl_dim_all);
1084 if (total < 0)
1085 return isl_change_error;
1086 for (l = 0; l < info[i].bmap->n_div; ++l)
1087 if (!isl_seq_eq(info[i].bmap->div[l],
1088 info[j].bmap->div[l], 1 + 1 + total)) {
1089 isl_int_set_si(info[i].bmap->div[l][0], 0);
1090 info[i].simplify = 1;
1092 for (l = 0; l < n; ++l)
1093 isl_int_add_ui(info[i].bmap->ineq[relax[l]][0],
1094 info[i].bmap->ineq[relax[l]][0], 1);
1095 ISL_F_SET(info[i].bmap, ISL_BASIC_MAP_FINAL);
1096 drop(&info[j]);
1097 if (j < i)
1098 exchange(&info[i], &info[j]);
1099 return isl_change_fuse;
1102 /* Basic map "i" has "n" inequality constraints (collected in "relax")
1103 * that are such that they include basic map "j" if they are relaxed
1104 * by one. All the other inequalities are valid for "j".
1105 * Check if basic map "j" forms an extension of basic map "i".
1107 * In particular, relax the constraints in "relax", compute the corresponding
1108 * facets one by one and check whether each of these is included
1109 * in the other basic map.
1110 * Before testing for inclusion, the constraints on each facet
1111 * are tightened to increase the chance of an inclusion being detected.
1112 * (Adding the valid constraints of "j" to the tableau of "i", as is done
1113 * in is_adj_ineq_extension, may further increase those chances, but this
1114 * is not currently done.)
1115 * If each facet is included, we know that relaxing the constraints extends
1116 * the basic map with exactly the other basic map (we already know that this
1117 * other basic map is included in the extension, because all other
1118 * inequality constraints are valid of "j") and we can replace the
1119 * two basic maps by this extension.
1121 * If any of the relaxed constraints turn out to be redundant, then bail out.
1122 * isl_tab_select_facet refuses to handle such constraints. It may be
1123 * possible to handle them anyway by making a distinction between
1124 * redundant constraints with a corresponding facet that still intersects
1125 * the set (allowing isl_tab_select_facet to handle them) and
1126 * those where the facet does not intersect the set (which can be ignored
1127 * because the empty facet is trivially included in the other disjunct).
1128 * However, relaxed constraints that turn out to be redundant should
1129 * be fairly rare and no such instance has been reported where
1130 * coalescing would be successful.
1131 * ____ _____
1132 * / || / |
1133 * / || / |
1134 * \ || => \ |
1135 * \ || \ |
1136 * \___|| \____|
1139 * \ |\
1140 * |\\ | \
1141 * | \\ | \
1142 * | | => | /
1143 * | / | /
1144 * |/ |/
1146 static enum isl_change is_relaxed_extension(int i, int j, int n, int *relax,
1147 struct isl_coalesce_info *info)
1149 int l;
1150 isl_bool super;
1151 struct isl_tab_undo *snap, *snap2;
1152 unsigned n_eq = info[i].bmap->n_eq;
1154 for (l = 0; l < n; ++l)
1155 if (isl_tab_is_equality(info[i].tab, n_eq + relax[l]))
1156 return isl_change_none;
1158 snap = isl_tab_snap(info[i].tab);
1159 for (l = 0; l < n; ++l)
1160 if (isl_tab_relax(info[i].tab, n_eq + relax[l]) < 0)
1161 return isl_change_error;
1162 for (l = 0; l < n; ++l) {
1163 if (!isl_tab_is_redundant(info[i].tab, n_eq + relax[l]))
1164 continue;
1165 if (isl_tab_rollback(info[i].tab, snap) < 0)
1166 return isl_change_error;
1167 return isl_change_none;
1169 snap2 = isl_tab_snap(info[i].tab);
1170 for (l = 0; l < n; ++l) {
1171 if (isl_tab_rollback(info[i].tab, snap2) < 0)
1172 return isl_change_error;
1173 if (isl_tab_select_facet(info[i].tab, n_eq + relax[l]) < 0)
1174 return isl_change_error;
1175 if (tighten_on_relaxed_facet(&info[i], n, relax, l) < 0)
1176 return isl_change_error;
1177 super = contains(&info[j], info[i].tab);
1178 if (super < 0)
1179 return isl_change_error;
1180 if (super)
1181 continue;
1182 if (isl_tab_rollback(info[i].tab, snap) < 0)
1183 return isl_change_error;
1184 return isl_change_none;
1187 if (isl_tab_rollback(info[i].tab, snap2) < 0)
1188 return isl_change_error;
1189 return extend(i, j, n, relax, info);
1192 /* Data structure that keeps track of the wrapping constraints
1193 * and of information to bound the coefficients of those constraints.
1195 * bound is set if we want to apply a bound on the coefficients
1196 * mat contains the wrapping constraints
1197 * max is the bound on the coefficients (if bound is set)
1199 struct isl_wraps {
1200 int bound;
1201 isl_mat *mat;
1202 isl_int max;
1205 /* Update wraps->max to be greater than or equal to the coefficients
1206 * in the equalities and inequalities of info->bmap that can be removed
1207 * if we end up applying wrapping.
1209 static isl_stat wraps_update_max(struct isl_wraps *wraps,
1210 struct isl_coalesce_info *info)
1212 int k;
1213 isl_int max_k;
1214 isl_size total = isl_basic_map_dim(info->bmap, isl_dim_all);
1216 if (total < 0)
1217 return isl_stat_error;
1218 isl_int_init(max_k);
1220 for (k = 0; k < info->bmap->n_eq; ++k) {
1221 if (info->eq[2 * k] == STATUS_VALID &&
1222 info->eq[2 * k + 1] == STATUS_VALID)
1223 continue;
1224 isl_seq_abs_max(info->bmap->eq[k] + 1, total, &max_k);
1225 if (isl_int_abs_gt(max_k, wraps->max))
1226 isl_int_set(wraps->max, max_k);
1229 for (k = 0; k < info->bmap->n_ineq; ++k) {
1230 if (info->ineq[k] == STATUS_VALID ||
1231 info->ineq[k] == STATUS_REDUNDANT)
1232 continue;
1233 isl_seq_abs_max(info->bmap->ineq[k] + 1, total, &max_k);
1234 if (isl_int_abs_gt(max_k, wraps->max))
1235 isl_int_set(wraps->max, max_k);
1238 isl_int_clear(max_k);
1240 return isl_stat_ok;
1243 /* Initialize the isl_wraps data structure.
1244 * If we want to bound the coefficients of the wrapping constraints,
1245 * we set wraps->max to the largest coefficient
1246 * in the equalities and inequalities that can be removed if we end up
1247 * applying wrapping.
1249 static isl_stat wraps_init(struct isl_wraps *wraps, __isl_take isl_mat *mat,
1250 struct isl_coalesce_info *info, int i, int j)
1252 isl_ctx *ctx;
1254 wraps->bound = 0;
1255 wraps->mat = mat;
1256 if (!mat)
1257 return isl_stat_error;
1258 ctx = isl_mat_get_ctx(mat);
1259 wraps->bound = isl_options_get_coalesce_bounded_wrapping(ctx);
1260 if (!wraps->bound)
1261 return isl_stat_ok;
1262 isl_int_init(wraps->max);
1263 isl_int_set_si(wraps->max, 0);
1264 if (wraps_update_max(wraps, &info[i]) < 0)
1265 return isl_stat_error;
1266 if (wraps_update_max(wraps, &info[j]) < 0)
1267 return isl_stat_error;
1269 return isl_stat_ok;
1272 /* Free the contents of the isl_wraps data structure.
1274 static void wraps_free(struct isl_wraps *wraps)
1276 isl_mat_free(wraps->mat);
1277 if (wraps->bound)
1278 isl_int_clear(wraps->max);
1281 /* Is the wrapping constraint in row "row" allowed?
1283 * If wraps->bound is set, we check that none of the coefficients
1284 * is greater than wraps->max.
1286 static int allow_wrap(struct isl_wraps *wraps, int row)
1288 int i;
1290 if (!wraps->bound)
1291 return 1;
1293 for (i = 1; i < wraps->mat->n_col; ++i)
1294 if (isl_int_abs_gt(wraps->mat->row[row][i], wraps->max))
1295 return 0;
1297 return 1;
1300 /* Wrap "ineq" (or its opposite if "negate" is set) around "bound"
1301 * to include "set" and add the result in position "w" of "wraps".
1302 * "len" is the total number of coefficients in "bound" and "ineq".
1303 * Return 1 on success, 0 on failure and -1 on error.
1304 * Wrapping can fail if the result of wrapping is equal to "bound"
1305 * or if we want to bound the sizes of the coefficients and
1306 * the wrapped constraint does not satisfy this bound.
1308 static int add_wrap(struct isl_wraps *wraps, int w, isl_int *bound,
1309 isl_int *ineq, unsigned len, __isl_keep isl_set *set, int negate)
1311 isl_seq_cpy(wraps->mat->row[w], bound, len);
1312 if (negate) {
1313 isl_seq_neg(wraps->mat->row[w + 1], ineq, len);
1314 ineq = wraps->mat->row[w + 1];
1316 if (!isl_set_wrap_facet(set, wraps->mat->row[w], ineq))
1317 return -1;
1318 if (isl_seq_eq(wraps->mat->row[w], bound, len))
1319 return 0;
1320 if (!allow_wrap(wraps, w))
1321 return 0;
1322 return 1;
1325 /* For each constraint in info->bmap that is not redundant (as determined
1326 * by info->tab) and that is not a valid constraint for the other basic map,
1327 * wrap the constraint around "bound" such that it includes the whole
1328 * set "set" and append the resulting constraint to "wraps".
1329 * Note that the constraints that are valid for the other basic map
1330 * will be added to the combined basic map by default, so there is
1331 * no need to wrap them.
1332 * The caller wrap_in_facets even relies on this function not wrapping
1333 * any constraints that are already valid.
1334 * "wraps" is assumed to have been pre-allocated to the appropriate size.
1335 * wraps->n_row is the number of actual wrapped constraints that have
1336 * been added.
1337 * If any of the wrapping problems results in a constraint that is
1338 * identical to "bound", then this means that "set" is unbounded in such
1339 * way that no wrapping is possible. If this happens then wraps->n_row
1340 * is reset to zero.
1341 * Similarly, if we want to bound the coefficients of the wrapping
1342 * constraints and a newly added wrapping constraint does not
1343 * satisfy the bound, then wraps->n_row is also reset to zero.
1345 static isl_stat add_wraps(struct isl_wraps *wraps,
1346 struct isl_coalesce_info *info, isl_int *bound, __isl_keep isl_set *set)
1348 int l, m;
1349 int w;
1350 int added;
1351 isl_basic_map *bmap = info->bmap;
1352 isl_size total = isl_basic_map_dim(bmap, isl_dim_all);
1353 unsigned len = 1 + total;
1355 if (total < 0)
1356 return isl_stat_error;
1358 w = wraps->mat->n_row;
1360 for (l = 0; l < bmap->n_ineq; ++l) {
1361 if (info->ineq[l] == STATUS_VALID ||
1362 info->ineq[l] == STATUS_REDUNDANT)
1363 continue;
1364 if (isl_seq_is_neg(bound, bmap->ineq[l], len))
1365 continue;
1366 if (isl_seq_eq(bound, bmap->ineq[l], len))
1367 continue;
1368 if (isl_tab_is_redundant(info->tab, bmap->n_eq + l))
1369 continue;
1371 added = add_wrap(wraps, w, bound, bmap->ineq[l], len, set, 0);
1372 if (added < 0)
1373 return isl_stat_error;
1374 if (!added)
1375 goto unbounded;
1376 ++w;
1378 for (l = 0; l < bmap->n_eq; ++l) {
1379 if (isl_seq_is_neg(bound, bmap->eq[l], len))
1380 continue;
1381 if (isl_seq_eq(bound, bmap->eq[l], len))
1382 continue;
1384 for (m = 0; m < 2; ++m) {
1385 if (info->eq[2 * l + m] == STATUS_VALID)
1386 continue;
1387 added = add_wrap(wraps, w, bound, bmap->eq[l], len,
1388 set, !m);
1389 if (added < 0)
1390 return isl_stat_error;
1391 if (!added)
1392 goto unbounded;
1393 ++w;
1397 wraps->mat->n_row = w;
1398 return isl_stat_ok;
1399 unbounded:
1400 wraps->mat->n_row = 0;
1401 return isl_stat_ok;
1404 /* Check if the constraints in "wraps" from "first" until the last
1405 * are all valid for the basic set represented by "tab".
1406 * If not, wraps->n_row is set to zero.
1408 static int check_wraps(__isl_keep isl_mat *wraps, int first,
1409 struct isl_tab *tab)
1411 int i;
1413 for (i = first; i < wraps->n_row; ++i) {
1414 enum isl_ineq_type type;
1415 type = isl_tab_ineq_type(tab, wraps->row[i]);
1416 if (type == isl_ineq_error)
1417 return -1;
1418 if (type == isl_ineq_redundant)
1419 continue;
1420 wraps->n_row = 0;
1421 return 0;
1424 return 0;
1427 /* Return a set that corresponds to the non-redundant constraints
1428 * (as recorded in tab) of bmap.
1430 * It's important to remove the redundant constraints as some
1431 * of the other constraints may have been modified after the
1432 * constraints were marked redundant.
1433 * In particular, a constraint may have been relaxed.
1434 * Redundant constraints are ignored when a constraint is relaxed
1435 * and should therefore continue to be ignored ever after.
1436 * Otherwise, the relaxation might be thwarted by some of
1437 * these constraints.
1439 * Update the underlying set to ensure that the dimension doesn't change.
1440 * Otherwise the integer divisions could get dropped if the tab
1441 * turns out to be empty.
1443 static __isl_give isl_set *set_from_updated_bmap(__isl_keep isl_basic_map *bmap,
1444 struct isl_tab *tab)
1446 isl_basic_set *bset;
1448 bmap = isl_basic_map_copy(bmap);
1449 bset = isl_basic_map_underlying_set(bmap);
1450 bset = isl_basic_set_cow(bset);
1451 bset = isl_basic_set_update_from_tab(bset, tab);
1452 return isl_set_from_basic_set(bset);
1455 /* Wrap the constraints of info->bmap that bound the facet defined
1456 * by inequality "k" around (the opposite of) this inequality to
1457 * include "set". "bound" may be used to store the negated inequality.
1458 * Since the wrapped constraints are not guaranteed to contain the whole
1459 * of info->bmap, we check them in check_wraps.
1460 * If any of the wrapped constraints turn out to be invalid, then
1461 * check_wraps will reset wrap->n_row to zero.
1463 static isl_stat add_wraps_around_facet(struct isl_wraps *wraps,
1464 struct isl_coalesce_info *info, int k, isl_int *bound,
1465 __isl_keep isl_set *set)
1467 struct isl_tab_undo *snap;
1468 int n;
1469 isl_size total = isl_basic_map_dim(info->bmap, isl_dim_all);
1471 if (total < 0)
1472 return isl_stat_error;
1474 snap = isl_tab_snap(info->tab);
1476 if (isl_tab_select_facet(info->tab, info->bmap->n_eq + k) < 0)
1477 return isl_stat_error;
1478 if (isl_tab_detect_redundant(info->tab) < 0)
1479 return isl_stat_error;
1481 isl_seq_neg(bound, info->bmap->ineq[k], 1 + total);
1483 n = wraps->mat->n_row;
1484 if (add_wraps(wraps, info, bound, set) < 0)
1485 return isl_stat_error;
1487 if (isl_tab_rollback(info->tab, snap) < 0)
1488 return isl_stat_error;
1489 if (check_wraps(wraps->mat, n, info->tab) < 0)
1490 return isl_stat_error;
1492 return isl_stat_ok;
1495 /* Given a basic set i with a constraint k that is adjacent to
1496 * basic set j, check if we can wrap
1497 * both the facet corresponding to k (if "wrap_facet" is set) and basic map j
1498 * (always) around their ridges to include the other set.
1499 * If so, replace the pair of basic sets by their union.
1501 * All constraints of i (except k) are assumed to be valid or
1502 * cut constraints for j.
1503 * Wrapping the cut constraints to include basic map j may result
1504 * in constraints that are no longer valid of basic map i
1505 * we have to check that the resulting wrapping constraints are valid for i.
1506 * If "wrap_facet" is not set, then all constraints of i (except k)
1507 * are assumed to be valid for j.
1508 * ____ _____
1509 * / | / \
1510 * / || / |
1511 * \ || => \ |
1512 * \ || \ |
1513 * \___|| \____|
1516 static enum isl_change can_wrap_in_facet(int i, int j, int k,
1517 struct isl_coalesce_info *info, int wrap_facet)
1519 enum isl_change change = isl_change_none;
1520 struct isl_wraps wraps;
1521 isl_ctx *ctx;
1522 isl_mat *mat;
1523 struct isl_set *set_i = NULL;
1524 struct isl_set *set_j = NULL;
1525 struct isl_vec *bound = NULL;
1526 isl_size total = isl_basic_map_dim(info[i].bmap, isl_dim_all);
1528 if (total < 0)
1529 return isl_change_error;
1530 set_i = set_from_updated_bmap(info[i].bmap, info[i].tab);
1531 set_j = set_from_updated_bmap(info[j].bmap, info[j].tab);
1532 ctx = isl_basic_map_get_ctx(info[i].bmap);
1533 mat = isl_mat_alloc(ctx, 2 * (info[i].bmap->n_eq + info[j].bmap->n_eq) +
1534 info[i].bmap->n_ineq + info[j].bmap->n_ineq,
1535 1 + total);
1536 if (wraps_init(&wraps, mat, info, i, j) < 0)
1537 goto error;
1538 bound = isl_vec_alloc(ctx, 1 + total);
1539 if (!set_i || !set_j || !bound)
1540 goto error;
1542 isl_seq_cpy(bound->el, info[i].bmap->ineq[k], 1 + total);
1543 isl_int_add_ui(bound->el[0], bound->el[0], 1);
1544 isl_seq_normalize(ctx, bound->el, 1 + total);
1546 isl_seq_cpy(wraps.mat->row[0], bound->el, 1 + total);
1547 wraps.mat->n_row = 1;
1549 if (add_wraps(&wraps, &info[j], bound->el, set_i) < 0)
1550 goto error;
1551 if (!wraps.mat->n_row)
1552 goto unbounded;
1554 if (wrap_facet) {
1555 if (add_wraps_around_facet(&wraps, &info[i], k,
1556 bound->el, set_j) < 0)
1557 goto error;
1558 if (!wraps.mat->n_row)
1559 goto unbounded;
1562 change = fuse(i, j, info, wraps.mat, 0, 0);
1564 unbounded:
1565 wraps_free(&wraps);
1567 isl_set_free(set_i);
1568 isl_set_free(set_j);
1570 isl_vec_free(bound);
1572 return change;
1573 error:
1574 wraps_free(&wraps);
1575 isl_vec_free(bound);
1576 isl_set_free(set_i);
1577 isl_set_free(set_j);
1578 return isl_change_error;
1581 /* Given a cut constraint t(x) >= 0 of basic map i, stored in row "w"
1582 * of wrap.mat, replace it by its relaxed version t(x) + 1 >= 0, and
1583 * add wrapping constraints to wrap.mat for all constraints
1584 * of basic map j that bound the part of basic map j that sticks out
1585 * of the cut constraint.
1586 * "set_i" is the underlying set of basic map i.
1587 * If any wrapping fails, then wraps->mat.n_row is reset to zero.
1589 * In particular, we first intersect basic map j with t(x) + 1 = 0.
1590 * If the result is empty, then t(x) >= 0 was actually a valid constraint
1591 * (with respect to the integer points), so we add t(x) >= 0 instead.
1592 * Otherwise, we wrap the constraints of basic map j that are not
1593 * redundant in this intersection and that are not already valid
1594 * for basic map i over basic map i.
1595 * Note that it is sufficient to wrap the constraints to include
1596 * basic map i, because we will only wrap the constraints that do
1597 * not include basic map i already. The wrapped constraint will
1598 * therefore be more relaxed compared to the original constraint.
1599 * Since the original constraint is valid for basic map j, so is
1600 * the wrapped constraint.
1602 static isl_stat wrap_in_facet(struct isl_wraps *wraps, int w,
1603 struct isl_coalesce_info *info_j, __isl_keep isl_set *set_i,
1604 struct isl_tab_undo *snap)
1606 isl_int_add_ui(wraps->mat->row[w][0], wraps->mat->row[w][0], 1);
1607 if (isl_tab_add_eq(info_j->tab, wraps->mat->row[w]) < 0)
1608 return isl_stat_error;
1609 if (isl_tab_detect_redundant(info_j->tab) < 0)
1610 return isl_stat_error;
1612 if (info_j->tab->empty)
1613 isl_int_sub_ui(wraps->mat->row[w][0], wraps->mat->row[w][0], 1);
1614 else if (add_wraps(wraps, info_j, wraps->mat->row[w], set_i) < 0)
1615 return isl_stat_error;
1617 if (isl_tab_rollback(info_j->tab, snap) < 0)
1618 return isl_stat_error;
1620 return isl_stat_ok;
1623 /* Given a pair of basic maps i and j such that j sticks out
1624 * of i at n cut constraints, each time by at most one,
1625 * try to compute wrapping constraints and replace the two
1626 * basic maps by a single basic map.
1627 * The other constraints of i are assumed to be valid for j.
1628 * "set_i" is the underlying set of basic map i.
1629 * "wraps" has been initialized to be of the right size.
1631 * For each cut constraint t(x) >= 0 of i, we add the relaxed version
1632 * t(x) + 1 >= 0, along with wrapping constraints for all constraints
1633 * of basic map j that bound the part of basic map j that sticks out
1634 * of the cut constraint.
1636 * If any wrapping fails, i.e., if we cannot wrap to touch
1637 * the union, then we give up.
1638 * Otherwise, the pair of basic maps is replaced by their union.
1640 static enum isl_change try_wrap_in_facets(int i, int j,
1641 struct isl_coalesce_info *info, struct isl_wraps *wraps,
1642 __isl_keep isl_set *set_i)
1644 int k, l, w;
1645 isl_size total;
1646 struct isl_tab_undo *snap;
1648 total = isl_basic_map_dim(info[i].bmap, isl_dim_all);
1649 if (total < 0)
1650 return isl_change_error;
1652 snap = isl_tab_snap(info[j].tab);
1654 wraps->mat->n_row = 0;
1656 for (k = 0; k < info[i].bmap->n_eq; ++k) {
1657 for (l = 0; l < 2; ++l) {
1658 if (info[i].eq[2 * k + l] != STATUS_CUT)
1659 continue;
1660 w = wraps->mat->n_row++;
1661 if (l == 0)
1662 isl_seq_neg(wraps->mat->row[w],
1663 info[i].bmap->eq[k], 1 + total);
1664 else
1665 isl_seq_cpy(wraps->mat->row[w],
1666 info[i].bmap->eq[k], 1 + total);
1667 if (wrap_in_facet(wraps, w, &info[j], set_i, snap) < 0)
1668 return isl_change_error;
1670 if (!wraps->mat->n_row)
1671 return isl_change_none;
1675 for (k = 0; k < info[i].bmap->n_ineq; ++k) {
1676 if (info[i].ineq[k] != STATUS_CUT)
1677 continue;
1678 w = wraps->mat->n_row++;
1679 isl_seq_cpy(wraps->mat->row[w],
1680 info[i].bmap->ineq[k], 1 + total);
1681 if (wrap_in_facet(wraps, w, &info[j], set_i, snap) < 0)
1682 return isl_change_error;
1684 if (!wraps->mat->n_row)
1685 return isl_change_none;
1688 return fuse(i, j, info, wraps->mat, 0, 1);
1691 /* Given a pair of basic maps i and j such that j sticks out
1692 * of i at n cut constraints, each time by at most one,
1693 * try to compute wrapping constraints and replace the two
1694 * basic maps by a single basic map.
1695 * The other constraints of i are assumed to be valid for j.
1697 * The core computation is performed by try_wrap_in_facets.
1698 * This function simply extracts an underlying set representation
1699 * of basic map i and initializes the data structure for keeping
1700 * track of wrapping constraints.
1702 static enum isl_change wrap_in_facets(int i, int j, int n,
1703 struct isl_coalesce_info *info)
1705 enum isl_change change = isl_change_none;
1706 struct isl_wraps wraps;
1707 isl_ctx *ctx;
1708 isl_mat *mat;
1709 isl_set *set_i = NULL;
1710 isl_size total = isl_basic_map_dim(info[i].bmap, isl_dim_all);
1711 int max_wrap;
1713 if (total < 0)
1714 return isl_change_error;
1715 if (isl_tab_extend_cons(info[j].tab, 1) < 0)
1716 return isl_change_error;
1718 max_wrap = 1 + 2 * info[j].bmap->n_eq + info[j].bmap->n_ineq;
1719 max_wrap *= n;
1721 set_i = set_from_updated_bmap(info[i].bmap, info[i].tab);
1722 ctx = isl_basic_map_get_ctx(info[i].bmap);
1723 mat = isl_mat_alloc(ctx, max_wrap, 1 + total);
1724 if (wraps_init(&wraps, mat, info, i, j) < 0)
1725 goto error;
1726 if (!set_i)
1727 goto error;
1729 change = try_wrap_in_facets(i, j, info, &wraps, set_i);
1731 wraps_free(&wraps);
1732 isl_set_free(set_i);
1734 return change;
1735 error:
1736 wraps_free(&wraps);
1737 isl_set_free(set_i);
1738 return isl_change_error;
1741 /* Return the effect of inequality "ineq" on the tableau "tab",
1742 * after relaxing the constant term of "ineq" by one.
1744 static enum isl_ineq_type type_of_relaxed(struct isl_tab *tab, isl_int *ineq)
1746 enum isl_ineq_type type;
1748 isl_int_add_ui(ineq[0], ineq[0], 1);
1749 type = isl_tab_ineq_type(tab, ineq);
1750 isl_int_sub_ui(ineq[0], ineq[0], 1);
1752 return type;
1755 /* Given two basic sets i and j,
1756 * check if relaxing all the cut constraints of i by one turns
1757 * them into valid constraint for j and check if we can wrap in
1758 * the bits that are sticking out.
1759 * If so, replace the pair by their union.
1761 * We first check if all relaxed cut inequalities of i are valid for j
1762 * and then try to wrap in the intersections of the relaxed cut inequalities
1763 * with j.
1765 * During this wrapping, we consider the points of j that lie at a distance
1766 * of exactly 1 from i. In particular, we ignore the points that lie in
1767 * between this lower-dimensional space and the basic map i.
1768 * We can therefore only apply this to integer maps.
1769 * ____ _____
1770 * / ___|_ / \
1771 * / | | / |
1772 * \ | | => \ |
1773 * \|____| \ |
1774 * \___| \____/
1776 * _____ ______
1777 * | ____|_ | \
1778 * | | | | |
1779 * | | | => | |
1780 * |_| | | |
1781 * |_____| \______|
1783 * _______
1784 * | |
1785 * | |\ |
1786 * | | \ |
1787 * | | \ |
1788 * | | \|
1789 * | | \
1790 * | |_____\
1791 * | |
1792 * |_______|
1794 * Wrapping can fail if the result of wrapping one of the facets
1795 * around its edges does not produce any new facet constraint.
1796 * In particular, this happens when we try to wrap in unbounded sets.
1798 * _______________________________________________________________________
1800 * | ___
1801 * | | |
1802 * |_| |_________________________________________________________________
1803 * |___|
1805 * The following is not an acceptable result of coalescing the above two
1806 * sets as it includes extra integer points.
1807 * _______________________________________________________________________
1809 * |
1810 * |
1812 * \______________________________________________________________________
1814 static enum isl_change can_wrap_in_set(int i, int j,
1815 struct isl_coalesce_info *info)
1817 int k, l;
1818 int n;
1819 isl_size total;
1821 if (ISL_F_ISSET(info[i].bmap, ISL_BASIC_MAP_RATIONAL) ||
1822 ISL_F_ISSET(info[j].bmap, ISL_BASIC_MAP_RATIONAL))
1823 return isl_change_none;
1825 n = count_eq(&info[i], STATUS_CUT) + count_ineq(&info[i], STATUS_CUT);
1826 if (n == 0)
1827 return isl_change_none;
1829 total = isl_basic_map_dim(info[i].bmap, isl_dim_all);
1830 if (total < 0)
1831 return isl_change_error;
1832 for (k = 0; k < info[i].bmap->n_eq; ++k) {
1833 for (l = 0; l < 2; ++l) {
1834 enum isl_ineq_type type;
1836 if (info[i].eq[2 * k + l] != STATUS_CUT)
1837 continue;
1839 if (l == 0)
1840 isl_seq_neg(info[i].bmap->eq[k],
1841 info[i].bmap->eq[k], 1 + total);
1842 type = type_of_relaxed(info[j].tab,
1843 info[i].bmap->eq[k]);
1844 if (l == 0)
1845 isl_seq_neg(info[i].bmap->eq[k],
1846 info[i].bmap->eq[k], 1 + total);
1847 if (type == isl_ineq_error)
1848 return isl_change_error;
1849 if (type != isl_ineq_redundant)
1850 return isl_change_none;
1854 for (k = 0; k < info[i].bmap->n_ineq; ++k) {
1855 enum isl_ineq_type type;
1857 if (info[i].ineq[k] != STATUS_CUT)
1858 continue;
1860 type = type_of_relaxed(info[j].tab, info[i].bmap->ineq[k]);
1861 if (type == isl_ineq_error)
1862 return isl_change_error;
1863 if (type != isl_ineq_redundant)
1864 return isl_change_none;
1867 return wrap_in_facets(i, j, n, info);
1870 /* Check if either i or j has only cut constraints that can
1871 * be used to wrap in (a facet of) the other basic set.
1872 * if so, replace the pair by their union.
1874 static enum isl_change check_wrap(int i, int j, struct isl_coalesce_info *info)
1876 enum isl_change change = isl_change_none;
1878 change = can_wrap_in_set(i, j, info);
1879 if (change != isl_change_none)
1880 return change;
1882 change = can_wrap_in_set(j, i, info);
1883 return change;
1886 /* Check if all inequality constraints of "i" that cut "j" cease
1887 * to be cut constraints if they are relaxed by one.
1888 * If so, collect the cut constraints in "list".
1889 * The caller is responsible for allocating "list".
1891 static isl_bool all_cut_by_one(int i, int j, struct isl_coalesce_info *info,
1892 int *list)
1894 int l, n;
1896 n = 0;
1897 for (l = 0; l < info[i].bmap->n_ineq; ++l) {
1898 enum isl_ineq_type type;
1900 if (info[i].ineq[l] != STATUS_CUT)
1901 continue;
1902 type = type_of_relaxed(info[j].tab, info[i].bmap->ineq[l]);
1903 if (type == isl_ineq_error)
1904 return isl_bool_error;
1905 if (type != isl_ineq_redundant)
1906 return isl_bool_false;
1907 list[n++] = l;
1910 return isl_bool_true;
1913 /* Given two basic maps such that "j" has at least one equality constraint
1914 * that is adjacent to an inequality constraint of "i" and such that "i" has
1915 * exactly one inequality constraint that is adjacent to an equality
1916 * constraint of "j", check whether "i" can be extended to include "j" or
1917 * whether "j" can be wrapped into "i".
1918 * All remaining constraints of "i" and "j" are assumed to be valid
1919 * or cut constraints of the other basic map.
1920 * However, none of the equality constraints of "i" are cut constraints.
1922 * If "i" has any "cut" inequality constraints, then check if relaxing
1923 * each of them by one is sufficient for them to become valid.
1924 * If so, check if the inequality constraint adjacent to an equality
1925 * constraint of "j" along with all these cut constraints
1926 * can be relaxed by one to contain exactly "j".
1927 * Otherwise, or if this fails, check if "j" can be wrapped into "i".
1929 static enum isl_change check_single_adj_eq(int i, int j,
1930 struct isl_coalesce_info *info)
1932 enum isl_change change = isl_change_none;
1933 int k;
1934 int n_cut;
1935 int *relax;
1936 isl_ctx *ctx;
1937 isl_bool try_relax;
1939 n_cut = count_ineq(&info[i], STATUS_CUT);
1941 k = find_ineq(&info[i], STATUS_ADJ_EQ);
1943 if (n_cut > 0) {
1944 ctx = isl_basic_map_get_ctx(info[i].bmap);
1945 relax = isl_calloc_array(ctx, int, 1 + n_cut);
1946 if (!relax)
1947 return isl_change_error;
1948 relax[0] = k;
1949 try_relax = all_cut_by_one(i, j, info, relax + 1);
1950 if (try_relax < 0)
1951 change = isl_change_error;
1952 } else {
1953 try_relax = isl_bool_true;
1954 relax = &k;
1956 if (try_relax && change == isl_change_none)
1957 change = is_relaxed_extension(i, j, 1 + n_cut, relax, info);
1958 if (n_cut > 0)
1959 free(relax);
1960 if (change != isl_change_none)
1961 return change;
1963 change = can_wrap_in_facet(i, j, k, info, n_cut > 0);
1965 return change;
1968 /* At least one of the basic maps has an equality that is adjacent
1969 * to an inequality. Make sure that only one of the basic maps has
1970 * such an equality and that the other basic map has exactly one
1971 * inequality adjacent to an equality.
1972 * If the other basic map does not have such an inequality, then
1973 * check if all its constraints are either valid or cut constraints
1974 * and, if so, try wrapping in the first map into the second.
1975 * Otherwise, try to extend one basic map with the other or
1976 * wrap one basic map in the other.
1978 static enum isl_change check_adj_eq(int i, int j,
1979 struct isl_coalesce_info *info)
1981 if (any_eq(&info[i], STATUS_ADJ_INEQ) &&
1982 any_eq(&info[j], STATUS_ADJ_INEQ))
1983 /* ADJ EQ TOO MANY */
1984 return isl_change_none;
1986 if (any_eq(&info[i], STATUS_ADJ_INEQ))
1987 return check_adj_eq(j, i, info);
1989 /* j has an equality adjacent to an inequality in i */
1991 if (count_ineq(&info[i], STATUS_ADJ_EQ) != 1) {
1992 if (all_valid_or_cut(&info[i]))
1993 return can_wrap_in_set(i, j, info);
1994 return isl_change_none;
1996 if (any_eq(&info[i], STATUS_CUT))
1997 return isl_change_none;
1998 if (any_ineq(&info[j], STATUS_ADJ_EQ) ||
1999 any_ineq(&info[i], STATUS_ADJ_INEQ) ||
2000 any_ineq(&info[j], STATUS_ADJ_INEQ))
2001 /* ADJ EQ TOO MANY */
2002 return isl_change_none;
2004 return check_single_adj_eq(i, j, info);
2007 /* Disjunct "j" lies on a hyperplane that is adjacent to disjunct "i".
2008 * In particular, disjunct "i" has an inequality constraint that is adjacent
2009 * to a (combination of) equality constraint(s) of disjunct "j",
2010 * but disjunct "j" has no explicit equality constraint adjacent
2011 * to an inequality constraint of disjunct "i".
2013 * Disjunct "i" is already known not to have any equality constraints
2014 * that are adjacent to an equality or inequality constraint.
2015 * Check that, other than the inequality constraint mentioned above,
2016 * all other constraints of disjunct "i" are valid for disjunct "j".
2017 * If so, try and wrap in disjunct "j".
2019 static enum isl_change check_ineq_adj_eq(int i, int j,
2020 struct isl_coalesce_info *info)
2022 int k;
2024 if (any_eq(&info[i], STATUS_CUT))
2025 return isl_change_none;
2026 if (any_ineq(&info[i], STATUS_CUT))
2027 return isl_change_none;
2028 if (any_ineq(&info[i], STATUS_ADJ_INEQ))
2029 return isl_change_none;
2030 if (count_ineq(&info[i], STATUS_ADJ_EQ) != 1)
2031 return isl_change_none;
2033 k = find_ineq(&info[i], STATUS_ADJ_EQ);
2035 return can_wrap_in_facet(i, j, k, info, 0);
2038 /* The two basic maps lie on adjacent hyperplanes. In particular,
2039 * basic map "i" has an equality that lies parallel to basic map "j".
2040 * Check if we can wrap the facets around the parallel hyperplanes
2041 * to include the other set.
2043 * We perform basically the same operations as can_wrap_in_facet,
2044 * except that we don't need to select a facet of one of the sets.
2046 * \\ \\
2047 * \\ => \\
2048 * \ \|
2050 * If there is more than one equality of "i" adjacent to an equality of "j",
2051 * then the result will satisfy one or more equalities that are a linear
2052 * combination of these equalities. These will be encoded as pairs
2053 * of inequalities in the wrapping constraints and need to be made
2054 * explicit.
2056 static enum isl_change check_eq_adj_eq(int i, int j,
2057 struct isl_coalesce_info *info)
2059 int k;
2060 enum isl_change change = isl_change_none;
2061 int detect_equalities = 0;
2062 struct isl_wraps wraps;
2063 isl_ctx *ctx;
2064 isl_mat *mat;
2065 struct isl_set *set_i = NULL;
2066 struct isl_set *set_j = NULL;
2067 struct isl_vec *bound = NULL;
2068 isl_size total = isl_basic_map_dim(info[i].bmap, isl_dim_all);
2070 if (total < 0)
2071 return isl_change_error;
2072 if (count_eq(&info[i], STATUS_ADJ_EQ) != 1)
2073 detect_equalities = 1;
2075 k = find_eq(&info[i], STATUS_ADJ_EQ);
2077 set_i = set_from_updated_bmap(info[i].bmap, info[i].tab);
2078 set_j = set_from_updated_bmap(info[j].bmap, info[j].tab);
2079 ctx = isl_basic_map_get_ctx(info[i].bmap);
2080 mat = isl_mat_alloc(ctx, 2 * (info[i].bmap->n_eq + info[j].bmap->n_eq) +
2081 info[i].bmap->n_ineq + info[j].bmap->n_ineq,
2082 1 + total);
2083 if (wraps_init(&wraps, mat, info, i, j) < 0)
2084 goto error;
2085 bound = isl_vec_alloc(ctx, 1 + total);
2086 if (!set_i || !set_j || !bound)
2087 goto error;
2089 if (k % 2 == 0)
2090 isl_seq_neg(bound->el, info[i].bmap->eq[k / 2], 1 + total);
2091 else
2092 isl_seq_cpy(bound->el, info[i].bmap->eq[k / 2], 1 + total);
2093 isl_int_add_ui(bound->el[0], bound->el[0], 1);
2095 isl_seq_cpy(wraps.mat->row[0], bound->el, 1 + total);
2096 wraps.mat->n_row = 1;
2098 if (add_wraps(&wraps, &info[j], bound->el, set_i) < 0)
2099 goto error;
2100 if (!wraps.mat->n_row)
2101 goto unbounded;
2103 isl_int_sub_ui(bound->el[0], bound->el[0], 1);
2104 isl_seq_neg(bound->el, bound->el, 1 + total);
2106 isl_seq_cpy(wraps.mat->row[wraps.mat->n_row], bound->el, 1 + total);
2107 wraps.mat->n_row++;
2109 if (add_wraps(&wraps, &info[i], bound->el, set_j) < 0)
2110 goto error;
2111 if (!wraps.mat->n_row)
2112 goto unbounded;
2114 change = fuse(i, j, info, wraps.mat, detect_equalities, 0);
2116 if (0) {
2117 error: change = isl_change_error;
2119 unbounded:
2121 wraps_free(&wraps);
2122 isl_set_free(set_i);
2123 isl_set_free(set_j);
2124 isl_vec_free(bound);
2126 return change;
2129 /* Initialize the "eq" and "ineq" fields of "info".
2131 static void init_status(struct isl_coalesce_info *info)
2133 info->eq = info->ineq = NULL;
2136 /* Set info->eq to the positions of the equalities of info->bmap
2137 * with respect to the basic map represented by "tab".
2138 * If info->eq has already been computed, then do not compute it again.
2140 static void set_eq_status_in(struct isl_coalesce_info *info,
2141 struct isl_tab *tab)
2143 if (info->eq)
2144 return;
2145 info->eq = eq_status_in(info->bmap, tab);
2148 /* Set info->ineq to the positions of the inequalities of info->bmap
2149 * with respect to the basic map represented by "tab".
2150 * If info->ineq has already been computed, then do not compute it again.
2152 static void set_ineq_status_in(struct isl_coalesce_info *info,
2153 struct isl_tab *tab)
2155 if (info->ineq)
2156 return;
2157 info->ineq = ineq_status_in(info->bmap, info->tab, tab);
2160 /* Free the memory allocated by the "eq" and "ineq" fields of "info".
2161 * This function assumes that init_status has been called on "info" first,
2162 * after which the "eq" and "ineq" fields may or may not have been
2163 * assigned a newly allocated array.
2165 static void clear_status(struct isl_coalesce_info *info)
2167 free(info->eq);
2168 free(info->ineq);
2171 /* Are all inequality constraints of the basic map represented by "info"
2172 * valid for the other basic map, except for a single constraint
2173 * that is adjacent to an inequality constraint of the other basic map?
2175 static int all_ineq_valid_or_single_adj_ineq(struct isl_coalesce_info *info)
2177 int i;
2178 int k = -1;
2180 for (i = 0; i < info->bmap->n_ineq; ++i) {
2181 if (info->ineq[i] == STATUS_REDUNDANT)
2182 continue;
2183 if (info->ineq[i] == STATUS_VALID)
2184 continue;
2185 if (info->ineq[i] != STATUS_ADJ_INEQ)
2186 return 0;
2187 if (k != -1)
2188 return 0;
2189 k = i;
2192 return k != -1;
2195 /* Basic map "i" has one or more equality constraints that separate it
2196 * from basic map "j". Check if it happens to be an extension
2197 * of basic map "j".
2198 * In particular, check that all constraints of "j" are valid for "i",
2199 * except for one inequality constraint that is adjacent
2200 * to an inequality constraints of "i".
2201 * If so, check for "i" being an extension of "j" by calling
2202 * is_adj_ineq_extension.
2204 * Clean up the memory allocated for keeping track of the status
2205 * of the constraints before returning.
2207 static enum isl_change separating_equality(int i, int j,
2208 struct isl_coalesce_info *info)
2210 enum isl_change change = isl_change_none;
2212 if (all(info[j].eq, 2 * info[j].bmap->n_eq, STATUS_VALID) &&
2213 all_ineq_valid_or_single_adj_ineq(&info[j]))
2214 change = is_adj_ineq_extension(j, i, info);
2216 clear_status(&info[i]);
2217 clear_status(&info[j]);
2218 return change;
2221 /* Check if the union of the given pair of basic maps
2222 * can be represented by a single basic map.
2223 * If so, replace the pair by the single basic map and return
2224 * isl_change_drop_first, isl_change_drop_second or isl_change_fuse.
2225 * Otherwise, return isl_change_none.
2226 * The two basic maps are assumed to live in the same local space.
2227 * The "eq" and "ineq" fields of info[i] and info[j] are assumed
2228 * to have been initialized by the caller, either to NULL or
2229 * to valid information.
2231 * We first check the effect of each constraint of one basic map
2232 * on the other basic map.
2233 * The constraint may be
2234 * redundant the constraint is redundant in its own
2235 * basic map and should be ignore and removed
2236 * in the end
2237 * valid all (integer) points of the other basic map
2238 * satisfy the constraint
2239 * separate no (integer) point of the other basic map
2240 * satisfies the constraint
2241 * cut some but not all points of the other basic map
2242 * satisfy the constraint
2243 * adj_eq the given constraint is adjacent (on the outside)
2244 * to an equality of the other basic map
2245 * adj_ineq the given constraint is adjacent (on the outside)
2246 * to an inequality of the other basic map
2248 * We consider seven cases in which we can replace the pair by a single
2249 * basic map. We ignore all "redundant" constraints.
2251 * 1. all constraints of one basic map are valid
2252 * => the other basic map is a subset and can be removed
2254 * 2. all constraints of both basic maps are either "valid" or "cut"
2255 * and the facets corresponding to the "cut" constraints
2256 * of one of the basic maps lies entirely inside the other basic map
2257 * => the pair can be replaced by a basic map consisting
2258 * of the valid constraints in both basic maps
2260 * 3. there is a single pair of adjacent inequalities
2261 * (all other constraints are "valid")
2262 * => the pair can be replaced by a basic map consisting
2263 * of the valid constraints in both basic maps
2265 * 4. one basic map has a single adjacent inequality, while the other
2266 * constraints are "valid". The other basic map has some
2267 * "cut" constraints, but replacing the adjacent inequality by
2268 * its opposite and adding the valid constraints of the other
2269 * basic map results in a subset of the other basic map
2270 * => the pair can be replaced by a basic map consisting
2271 * of the valid constraints in both basic maps
2273 * 5. there is a single adjacent pair of an inequality and an equality,
2274 * the other constraints of the basic map containing the inequality are
2275 * "valid". Moreover, if the inequality the basic map is relaxed
2276 * and then turned into an equality, then resulting facet lies
2277 * entirely inside the other basic map
2278 * => the pair can be replaced by the basic map containing
2279 * the inequality, with the inequality relaxed.
2281 * 6. there is a single inequality adjacent to an equality,
2282 * the other constraints of the basic map containing the inequality are
2283 * "valid". Moreover, the facets corresponding to both
2284 * the inequality and the equality can be wrapped around their
2285 * ridges to include the other basic map
2286 * => the pair can be replaced by a basic map consisting
2287 * of the valid constraints in both basic maps together
2288 * with all wrapping constraints
2290 * 7. one of the basic maps extends beyond the other by at most one.
2291 * Moreover, the facets corresponding to the cut constraints and
2292 * the pieces of the other basic map at offset one from these cut
2293 * constraints can be wrapped around their ridges to include
2294 * the union of the two basic maps
2295 * => the pair can be replaced by a basic map consisting
2296 * of the valid constraints in both basic maps together
2297 * with all wrapping constraints
2299 * 8. the two basic maps live in adjacent hyperplanes. In principle
2300 * such sets can always be combined through wrapping, but we impose
2301 * that there is only one such pair, to avoid overeager coalescing.
2303 * Throughout the computation, we maintain a collection of tableaus
2304 * corresponding to the basic maps. When the basic maps are dropped
2305 * or combined, the tableaus are modified accordingly.
2307 static enum isl_change coalesce_local_pair_reuse(int i, int j,
2308 struct isl_coalesce_info *info)
2310 enum isl_change change = isl_change_none;
2312 set_ineq_status_in(&info[i], info[j].tab);
2313 if (info[i].bmap->n_ineq && !info[i].ineq)
2314 goto error;
2315 if (any_ineq(&info[i], STATUS_ERROR))
2316 goto error;
2317 if (any_ineq(&info[i], STATUS_SEPARATE))
2318 goto done;
2320 set_ineq_status_in(&info[j], info[i].tab);
2321 if (info[j].bmap->n_ineq && !info[j].ineq)
2322 goto error;
2323 if (any_ineq(&info[j], STATUS_ERROR))
2324 goto error;
2325 if (any_ineq(&info[j], STATUS_SEPARATE))
2326 goto done;
2328 set_eq_status_in(&info[i], info[j].tab);
2329 if (info[i].bmap->n_eq && !info[i].eq)
2330 goto error;
2331 if (any_eq(&info[i], STATUS_ERROR))
2332 goto error;
2334 set_eq_status_in(&info[j], info[i].tab);
2335 if (info[j].bmap->n_eq && !info[j].eq)
2336 goto error;
2337 if (any_eq(&info[j], STATUS_ERROR))
2338 goto error;
2340 if (any_eq(&info[i], STATUS_SEPARATE))
2341 return separating_equality(i, j, info);
2342 if (any_eq(&info[j], STATUS_SEPARATE))
2343 return separating_equality(j, i, info);
2345 if (all(info[i].eq, 2 * info[i].bmap->n_eq, STATUS_VALID) &&
2346 all(info[i].ineq, info[i].bmap->n_ineq, STATUS_VALID)) {
2347 drop(&info[j]);
2348 change = isl_change_drop_second;
2349 } else if (all(info[j].eq, 2 * info[j].bmap->n_eq, STATUS_VALID) &&
2350 all(info[j].ineq, info[j].bmap->n_ineq, STATUS_VALID)) {
2351 drop(&info[i]);
2352 change = isl_change_drop_first;
2353 } else if (any_eq(&info[i], STATUS_ADJ_EQ)) {
2354 change = check_eq_adj_eq(i, j, info);
2355 } else if (any_eq(&info[j], STATUS_ADJ_EQ)) {
2356 change = check_eq_adj_eq(j, i, info);
2357 } else if (any_eq(&info[i], STATUS_ADJ_INEQ) ||
2358 any_eq(&info[j], STATUS_ADJ_INEQ)) {
2359 change = check_adj_eq(i, j, info);
2360 } else if (any_ineq(&info[i], STATUS_ADJ_EQ)) {
2361 change = check_ineq_adj_eq(i, j, info);
2362 } else if (any_ineq(&info[j], STATUS_ADJ_EQ)) {
2363 change = check_ineq_adj_eq(j, i, info);
2364 } else if (any_ineq(&info[i], STATUS_ADJ_INEQ) ||
2365 any_ineq(&info[j], STATUS_ADJ_INEQ)) {
2366 change = check_adj_ineq(i, j, info);
2367 } else {
2368 if (!any_eq(&info[i], STATUS_CUT) &&
2369 !any_eq(&info[j], STATUS_CUT))
2370 change = check_facets(i, j, info);
2371 if (change == isl_change_none)
2372 change = check_wrap(i, j, info);
2375 done:
2376 clear_status(&info[i]);
2377 clear_status(&info[j]);
2378 return change;
2379 error:
2380 clear_status(&info[i]);
2381 clear_status(&info[j]);
2382 return isl_change_error;
2385 /* Check if the union of the given pair of basic maps
2386 * can be represented by a single basic map.
2387 * If so, replace the pair by the single basic map and return
2388 * isl_change_drop_first, isl_change_drop_second or isl_change_fuse.
2389 * Otherwise, return isl_change_none.
2390 * The two basic maps are assumed to live in the same local space.
2392 static enum isl_change coalesce_local_pair(int i, int j,
2393 struct isl_coalesce_info *info)
2395 init_status(&info[i]);
2396 init_status(&info[j]);
2397 return coalesce_local_pair_reuse(i, j, info);
2400 /* Shift the integer division at position "div" of the basic map
2401 * represented by "info" by "shift".
2403 * That is, if the integer division has the form
2405 * floor(f(x)/d)
2407 * then replace it by
2409 * floor((f(x) + shift * d)/d) - shift
2411 static isl_stat shift_div(struct isl_coalesce_info *info, int div,
2412 isl_int shift)
2414 isl_size total, n_div;
2416 info->bmap = isl_basic_map_shift_div(info->bmap, div, 0, shift);
2417 if (!info->bmap)
2418 return isl_stat_error;
2420 total = isl_basic_map_dim(info->bmap, isl_dim_all);
2421 n_div = isl_basic_map_dim(info->bmap, isl_dim_div);
2422 if (total < 0 || n_div < 0)
2423 return isl_stat_error;
2424 total -= n_div;
2425 if (isl_tab_shift_var(info->tab, total + div, shift) < 0)
2426 return isl_stat_error;
2428 return isl_stat_ok;
2431 /* If the integer division at position "div" is defined by an equality,
2432 * i.e., a stride constraint, then change the integer division expression
2433 * to have a constant term equal to zero.
2435 * Let the equality constraint be
2437 * c + f + m a = 0
2439 * The integer division expression is then typically of the form
2441 * a = floor((-f - c')/m)
2443 * The integer division is first shifted by t = floor(c/m),
2444 * turning the equality constraint into
2446 * c - m floor(c/m) + f + m a' = 0
2448 * i.e.,
2450 * (c mod m) + f + m a' = 0
2452 * That is,
2454 * a' = (-f - (c mod m))/m = floor((-f)/m)
2456 * because a' is an integer and 0 <= (c mod m) < m.
2457 * The constant term of a' can therefore be zeroed out,
2458 * but only if the integer division expression is of the expected form.
2460 static isl_stat normalize_stride_div(struct isl_coalesce_info *info, int div)
2462 isl_bool defined, valid;
2463 isl_stat r;
2464 isl_constraint *c;
2465 isl_int shift, stride;
2467 defined = isl_basic_map_has_defining_equality(info->bmap, isl_dim_div,
2468 div, &c);
2469 if (defined < 0)
2470 return isl_stat_error;
2471 if (!defined)
2472 return isl_stat_ok;
2473 if (!c)
2474 return isl_stat_error;
2475 valid = isl_constraint_is_div_equality(c, div);
2476 isl_int_init(shift);
2477 isl_int_init(stride);
2478 isl_constraint_get_constant(c, &shift);
2479 isl_constraint_get_coefficient(c, isl_dim_div, div, &stride);
2480 isl_int_fdiv_q(shift, shift, stride);
2481 r = shift_div(info, div, shift);
2482 isl_int_clear(stride);
2483 isl_int_clear(shift);
2484 isl_constraint_free(c);
2485 if (r < 0 || valid < 0)
2486 return isl_stat_error;
2487 if (!valid)
2488 return isl_stat_ok;
2489 info->bmap = isl_basic_map_set_div_expr_constant_num_si_inplace(
2490 info->bmap, div, 0);
2491 if (!info->bmap)
2492 return isl_stat_error;
2493 return isl_stat_ok;
2496 /* The basic maps represented by "info1" and "info2" are known
2497 * to have the same number of integer divisions.
2498 * Check if pairs of integer divisions are equal to each other
2499 * despite the fact that they differ by a rational constant.
2501 * In particular, look for any pair of integer divisions that
2502 * only differ in their constant terms.
2503 * If either of these integer divisions is defined
2504 * by stride constraints, then modify it to have a zero constant term.
2505 * If both are defined by stride constraints then in the end they will have
2506 * the same (zero) constant term.
2508 static isl_stat harmonize_stride_divs(struct isl_coalesce_info *info1,
2509 struct isl_coalesce_info *info2)
2511 int i;
2512 isl_size n;
2514 n = isl_basic_map_dim(info1->bmap, isl_dim_div);
2515 if (n < 0)
2516 return isl_stat_error;
2517 for (i = 0; i < n; ++i) {
2518 isl_bool known, harmonize;
2520 known = isl_basic_map_div_is_known(info1->bmap, i);
2521 if (known >= 0 && known)
2522 known = isl_basic_map_div_is_known(info2->bmap, i);
2523 if (known < 0)
2524 return isl_stat_error;
2525 if (!known)
2526 continue;
2527 harmonize = isl_basic_map_equal_div_expr_except_constant(
2528 info1->bmap, i, info2->bmap, i);
2529 if (harmonize < 0)
2530 return isl_stat_error;
2531 if (!harmonize)
2532 continue;
2533 if (normalize_stride_div(info1, i) < 0)
2534 return isl_stat_error;
2535 if (normalize_stride_div(info2, i) < 0)
2536 return isl_stat_error;
2539 return isl_stat_ok;
2542 /* If "shift" is an integer constant, then shift the integer division
2543 * at position "div" of the basic map represented by "info" by "shift".
2544 * If "shift" is not an integer constant, then do nothing.
2545 * If "shift" is equal to zero, then no shift needs to be performed either.
2547 * That is, if the integer division has the form
2549 * floor(f(x)/d)
2551 * then replace it by
2553 * floor((f(x) + shift * d)/d) - shift
2555 static isl_stat shift_if_cst_int(struct isl_coalesce_info *info, int div,
2556 __isl_keep isl_aff *shift)
2558 isl_bool cst;
2559 isl_stat r;
2560 isl_int d;
2561 isl_val *c;
2563 cst = isl_aff_is_cst(shift);
2564 if (cst < 0 || !cst)
2565 return cst < 0 ? isl_stat_error : isl_stat_ok;
2567 c = isl_aff_get_constant_val(shift);
2568 cst = isl_val_is_int(c);
2569 if (cst >= 0 && cst)
2570 cst = isl_bool_not(isl_val_is_zero(c));
2571 if (cst < 0 || !cst) {
2572 isl_val_free(c);
2573 return cst < 0 ? isl_stat_error : isl_stat_ok;
2576 isl_int_init(d);
2577 r = isl_val_get_num_isl_int(c, &d);
2578 if (r >= 0)
2579 r = shift_div(info, div, d);
2580 isl_int_clear(d);
2582 isl_val_free(c);
2584 return r;
2587 /* Check if some of the divs in the basic map represented by "info1"
2588 * are shifts of the corresponding divs in the basic map represented
2589 * by "info2", taking into account the equality constraints "eq1" of "info1"
2590 * and "eq2" of "info2". If so, align them with those of "info2".
2591 * "info1" and "info2" are assumed to have the same number
2592 * of integer divisions.
2594 * An integer division is considered to be a shift of another integer
2595 * division if, after simplification with respect to the equality
2596 * constraints of the other basic map, one is equal to the other
2597 * plus a constant.
2599 * In particular, for each pair of integer divisions, if both are known,
2600 * have the same denominator and are not already equal to each other,
2601 * simplify each with respect to the equality constraints
2602 * of the other basic map. If the difference is an integer constant,
2603 * then move this difference outside.
2604 * That is, if, after simplification, one integer division is of the form
2606 * floor((f(x) + c_1)/d)
2608 * while the other is of the form
2610 * floor((f(x) + c_2)/d)
2612 * and n = (c_2 - c_1)/d is an integer, then replace the first
2613 * integer division by
2615 * floor((f_1(x) + c_1 + n * d)/d) - n,
2617 * where floor((f_1(x) + c_1 + n * d)/d) = floor((f2(x) + c_2)/d)
2618 * after simplification with respect to the equality constraints.
2620 static isl_stat harmonize_divs_with_hulls(struct isl_coalesce_info *info1,
2621 struct isl_coalesce_info *info2, __isl_keep isl_basic_set *eq1,
2622 __isl_keep isl_basic_set *eq2)
2624 int i;
2625 isl_size total;
2626 isl_local_space *ls1, *ls2;
2628 total = isl_basic_map_dim(info1->bmap, isl_dim_all);
2629 if (total < 0)
2630 return isl_stat_error;
2631 ls1 = isl_local_space_wrap(isl_basic_map_get_local_space(info1->bmap));
2632 ls2 = isl_local_space_wrap(isl_basic_map_get_local_space(info2->bmap));
2633 for (i = 0; i < info1->bmap->n_div; ++i) {
2634 isl_stat r;
2635 isl_aff *div1, *div2;
2637 if (!isl_local_space_div_is_known(ls1, i) ||
2638 !isl_local_space_div_is_known(ls2, i))
2639 continue;
2640 if (isl_int_ne(info1->bmap->div[i][0], info2->bmap->div[i][0]))
2641 continue;
2642 if (isl_seq_eq(info1->bmap->div[i] + 1,
2643 info2->bmap->div[i] + 1, 1 + total))
2644 continue;
2645 div1 = isl_local_space_get_div(ls1, i);
2646 div2 = isl_local_space_get_div(ls2, i);
2647 div1 = isl_aff_substitute_equalities(div1,
2648 isl_basic_set_copy(eq2));
2649 div2 = isl_aff_substitute_equalities(div2,
2650 isl_basic_set_copy(eq1));
2651 div2 = isl_aff_sub(div2, div1);
2652 r = shift_if_cst_int(info1, i, div2);
2653 isl_aff_free(div2);
2654 if (r < 0)
2655 break;
2657 isl_local_space_free(ls1);
2658 isl_local_space_free(ls2);
2660 if (i < info1->bmap->n_div)
2661 return isl_stat_error;
2662 return isl_stat_ok;
2665 /* Check if some of the divs in the basic map represented by "info1"
2666 * are shifts of the corresponding divs in the basic map represented
2667 * by "info2". If so, align them with those of "info2".
2668 * Only do this if "info1" and "info2" have the same number
2669 * of integer divisions.
2671 * An integer division is considered to be a shift of another integer
2672 * division if, after simplification with respect to the equality
2673 * constraints of the other basic map, one is equal to the other
2674 * plus a constant.
2676 * First check if pairs of integer divisions are equal to each other
2677 * despite the fact that they differ by a rational constant.
2678 * If so, try and arrange for them to have the same constant term.
2680 * Then, extract the equality constraints and continue with
2681 * harmonize_divs_with_hulls.
2683 * If the equality constraints of both basic maps are the same,
2684 * then there is no need to perform any shifting since
2685 * the coefficients of the integer divisions should have been
2686 * reduced in the same way.
2688 static isl_stat harmonize_divs(struct isl_coalesce_info *info1,
2689 struct isl_coalesce_info *info2)
2691 isl_bool equal;
2692 isl_basic_map *bmap1, *bmap2;
2693 isl_basic_set *eq1, *eq2;
2694 isl_stat r;
2696 if (!info1->bmap || !info2->bmap)
2697 return isl_stat_error;
2699 if (info1->bmap->n_div != info2->bmap->n_div)
2700 return isl_stat_ok;
2701 if (info1->bmap->n_div == 0)
2702 return isl_stat_ok;
2704 if (harmonize_stride_divs(info1, info2) < 0)
2705 return isl_stat_error;
2707 bmap1 = isl_basic_map_copy(info1->bmap);
2708 bmap2 = isl_basic_map_copy(info2->bmap);
2709 eq1 = isl_basic_map_wrap(isl_basic_map_plain_affine_hull(bmap1));
2710 eq2 = isl_basic_map_wrap(isl_basic_map_plain_affine_hull(bmap2));
2711 equal = isl_basic_set_plain_is_equal(eq1, eq2);
2712 if (equal < 0)
2713 r = isl_stat_error;
2714 else if (equal)
2715 r = isl_stat_ok;
2716 else
2717 r = harmonize_divs_with_hulls(info1, info2, eq1, eq2);
2718 isl_basic_set_free(eq1);
2719 isl_basic_set_free(eq2);
2721 return r;
2724 /* Do the two basic maps live in the same local space, i.e.,
2725 * do they have the same (known) divs?
2726 * If either basic map has any unknown divs, then we can only assume
2727 * that they do not live in the same local space.
2729 static isl_bool same_divs(__isl_keep isl_basic_map *bmap1,
2730 __isl_keep isl_basic_map *bmap2)
2732 int i;
2733 isl_bool known;
2734 isl_size total;
2736 if (!bmap1 || !bmap2)
2737 return isl_bool_error;
2738 if (bmap1->n_div != bmap2->n_div)
2739 return isl_bool_false;
2741 if (bmap1->n_div == 0)
2742 return isl_bool_true;
2744 known = isl_basic_map_divs_known(bmap1);
2745 if (known < 0 || !known)
2746 return known;
2747 known = isl_basic_map_divs_known(bmap2);
2748 if (known < 0 || !known)
2749 return known;
2751 total = isl_basic_map_dim(bmap1, isl_dim_all);
2752 if (total < 0)
2753 return isl_bool_error;
2754 for (i = 0; i < bmap1->n_div; ++i)
2755 if (!isl_seq_eq(bmap1->div[i], bmap2->div[i], 2 + total))
2756 return isl_bool_false;
2758 return isl_bool_true;
2761 /* Assuming that "tab" contains the equality constraints and
2762 * the initial inequality constraints of "bmap", copy the remaining
2763 * inequality constraints of "bmap" to "Tab".
2765 static isl_stat copy_ineq(struct isl_tab *tab, __isl_keep isl_basic_map *bmap)
2767 int i, n_ineq;
2769 if (!bmap)
2770 return isl_stat_error;
2772 n_ineq = tab->n_con - tab->n_eq;
2773 for (i = n_ineq; i < bmap->n_ineq; ++i)
2774 if (isl_tab_add_ineq(tab, bmap->ineq[i]) < 0)
2775 return isl_stat_error;
2777 return isl_stat_ok;
2780 /* Description of an integer division that is added
2781 * during an expansion.
2782 * "pos" is the position of the corresponding variable.
2783 * "cst" indicates whether this integer division has a fixed value.
2784 * "val" contains the fixed value, if the value is fixed.
2786 struct isl_expanded {
2787 int pos;
2788 isl_bool cst;
2789 isl_int val;
2792 /* For each of the "n" integer division variables "expanded",
2793 * if the variable has a fixed value, then add two inequality
2794 * constraints expressing the fixed value.
2795 * Otherwise, add the corresponding div constraints.
2796 * The caller is responsible for removing the div constraints
2797 * that it added for all these "n" integer divisions.
2799 * The div constraints and the pair of inequality constraints
2800 * forcing the fixed value cannot both be added for a given variable
2801 * as the combination may render some of the original constraints redundant.
2802 * These would then be ignored during the coalescing detection,
2803 * while they could remain in the fused result.
2805 * The two added inequality constraints are
2807 * -a + v >= 0
2808 * a - v >= 0
2810 * with "a" the variable and "v" its fixed value.
2811 * The facet corresponding to one of these two constraints is selected
2812 * in the tableau to ensure that the pair of inequality constraints
2813 * is treated as an equality constraint.
2815 * The information in info->ineq is thrown away because it was
2816 * computed in terms of div constraints, while some of those
2817 * have now been replaced by these pairs of inequality constraints.
2819 static isl_stat fix_constant_divs(struct isl_coalesce_info *info,
2820 int n, struct isl_expanded *expanded)
2822 unsigned o_div;
2823 int i;
2824 isl_vec *ineq;
2826 o_div = isl_basic_map_offset(info->bmap, isl_dim_div) - 1;
2827 ineq = isl_vec_alloc(isl_tab_get_ctx(info->tab), 1 + info->tab->n_var);
2828 if (!ineq)
2829 return isl_stat_error;
2830 isl_seq_clr(ineq->el + 1, info->tab->n_var);
2832 for (i = 0; i < n; ++i) {
2833 if (!expanded[i].cst) {
2834 info->bmap = isl_basic_map_extend_constraints(
2835 info->bmap, 0, 2);
2836 info->bmap = isl_basic_map_add_div_constraints(
2837 info->bmap, expanded[i].pos - o_div);
2838 } else {
2839 isl_int_set_si(ineq->el[1 + expanded[i].pos], -1);
2840 isl_int_set(ineq->el[0], expanded[i].val);
2841 info->bmap = isl_basic_map_add_ineq(info->bmap,
2842 ineq->el);
2843 isl_int_set_si(ineq->el[1 + expanded[i].pos], 1);
2844 isl_int_neg(ineq->el[0], expanded[i].val);
2845 info->bmap = isl_basic_map_add_ineq(info->bmap,
2846 ineq->el);
2847 isl_int_set_si(ineq->el[1 + expanded[i].pos], 0);
2849 if (copy_ineq(info->tab, info->bmap) < 0)
2850 break;
2851 if (expanded[i].cst &&
2852 isl_tab_select_facet(info->tab, info->tab->n_con - 1) < 0)
2853 break;
2856 isl_vec_free(ineq);
2858 clear_status(info);
2859 init_status(info);
2861 return i < n ? isl_stat_error : isl_stat_ok;
2864 /* Insert the "n" integer division variables "expanded"
2865 * into info->tab and info->bmap and
2866 * update info->ineq with respect to the redundant constraints
2867 * in the resulting tableau.
2868 * "bmap" contains the result of this insertion in info->bmap,
2869 * while info->bmap is the original version
2870 * of "bmap", i.e., the one that corresponds to the current
2871 * state of info->tab. The number of constraints in info->bmap
2872 * is assumed to be the same as the number of constraints
2873 * in info->tab. This is required to be able to detect
2874 * the extra constraints in "bmap".
2876 * In particular, introduce extra variables corresponding
2877 * to the extra integer divisions and add the div constraints
2878 * that were added to "bmap" after info->tab was created
2879 * from info->bmap.
2880 * Furthermore, check if these extra integer divisions happen
2881 * to attain a fixed integer value in info->tab.
2882 * If so, replace the corresponding div constraints by pairs
2883 * of inequality constraints that fix these
2884 * integer divisions to their single integer values.
2885 * Replace info->bmap by "bmap" to match the changes to info->tab.
2886 * info->ineq was computed without a tableau and therefore
2887 * does not take into account the redundant constraints
2888 * in the tableau. Mark them here.
2889 * There is no need to check the newly added div constraints
2890 * since they cannot be redundant.
2891 * The redundancy check is not performed when constants have been discovered
2892 * since info->ineq is completely thrown away in this case.
2894 static isl_stat tab_insert_divs(struct isl_coalesce_info *info,
2895 int n, struct isl_expanded *expanded, __isl_take isl_basic_map *bmap)
2897 int i, n_ineq;
2898 unsigned n_eq;
2899 struct isl_tab_undo *snap;
2900 int any;
2902 if (!bmap)
2903 return isl_stat_error;
2904 if (info->bmap->n_eq + info->bmap->n_ineq != info->tab->n_con)
2905 isl_die(isl_basic_map_get_ctx(bmap), isl_error_internal,
2906 "original tableau does not correspond "
2907 "to original basic map", goto error);
2909 if (isl_tab_extend_vars(info->tab, n) < 0)
2910 goto error;
2911 if (isl_tab_extend_cons(info->tab, 2 * n) < 0)
2912 goto error;
2914 for (i = 0; i < n; ++i) {
2915 if (isl_tab_insert_var(info->tab, expanded[i].pos) < 0)
2916 goto error;
2919 snap = isl_tab_snap(info->tab);
2921 n_ineq = info->tab->n_con - info->tab->n_eq;
2922 if (copy_ineq(info->tab, bmap) < 0)
2923 goto error;
2925 isl_basic_map_free(info->bmap);
2926 info->bmap = bmap;
2928 any = 0;
2929 for (i = 0; i < n; ++i) {
2930 expanded[i].cst = isl_tab_is_constant(info->tab,
2931 expanded[i].pos, &expanded[i].val);
2932 if (expanded[i].cst < 0)
2933 return isl_stat_error;
2934 if (expanded[i].cst)
2935 any = 1;
2938 if (any) {
2939 if (isl_tab_rollback(info->tab, snap) < 0)
2940 return isl_stat_error;
2941 info->bmap = isl_basic_map_cow(info->bmap);
2942 info->bmap = isl_basic_map_free_inequality(info->bmap, 2 * n);
2943 if (info->bmap < 0)
2944 return isl_stat_error;
2946 return fix_constant_divs(info, n, expanded);
2949 n_eq = info->bmap->n_eq;
2950 for (i = 0; i < n_ineq; ++i) {
2951 if (isl_tab_is_redundant(info->tab, n_eq + i))
2952 info->ineq[i] = STATUS_REDUNDANT;
2955 return isl_stat_ok;
2956 error:
2957 isl_basic_map_free(bmap);
2958 return isl_stat_error;
2961 /* Expand info->tab and info->bmap in the same way "bmap" was expanded
2962 * in isl_basic_map_expand_divs using the expansion "exp" and
2963 * update info->ineq with respect to the redundant constraints
2964 * in the resulting tableau. info->bmap is the original version
2965 * of "bmap", i.e., the one that corresponds to the current
2966 * state of info->tab. The number of constraints in info->bmap
2967 * is assumed to be the same as the number of constraints
2968 * in info->tab. This is required to be able to detect
2969 * the extra constraints in "bmap".
2971 * Extract the positions where extra local variables are introduced
2972 * from "exp" and call tab_insert_divs.
2974 static isl_stat expand_tab(struct isl_coalesce_info *info, int *exp,
2975 __isl_take isl_basic_map *bmap)
2977 isl_ctx *ctx;
2978 struct isl_expanded *expanded;
2979 int i, j, k, n;
2980 int extra_var;
2981 isl_size total, n_div;
2982 unsigned pos;
2983 isl_stat r;
2985 total = isl_basic_map_dim(bmap, isl_dim_all);
2986 n_div = isl_basic_map_dim(bmap, isl_dim_div);
2987 if (total < 0 || n_div < 0)
2988 return isl_stat_error;
2989 pos = total - n_div;
2990 extra_var = total - info->tab->n_var;
2991 n = n_div - extra_var;
2993 ctx = isl_basic_map_get_ctx(bmap);
2994 expanded = isl_calloc_array(ctx, struct isl_expanded, extra_var);
2995 if (extra_var && !expanded)
2996 goto error;
2998 i = 0;
2999 k = 0;
3000 for (j = 0; j < n_div; ++j) {
3001 if (i < n && exp[i] == j) {
3002 ++i;
3003 continue;
3005 expanded[k++].pos = pos + j;
3008 for (k = 0; k < extra_var; ++k)
3009 isl_int_init(expanded[k].val);
3011 r = tab_insert_divs(info, extra_var, expanded, bmap);
3013 for (k = 0; k < extra_var; ++k)
3014 isl_int_clear(expanded[k].val);
3015 free(expanded);
3017 return r;
3018 error:
3019 isl_basic_map_free(bmap);
3020 return isl_stat_error;
3023 /* Check if the union of the basic maps represented by info[i] and info[j]
3024 * can be represented by a single basic map,
3025 * after expanding the divs of info[i] to match those of info[j].
3026 * If so, replace the pair by the single basic map and return
3027 * isl_change_drop_first, isl_change_drop_second or isl_change_fuse.
3028 * Otherwise, return isl_change_none.
3030 * The caller has already checked for info[j] being a subset of info[i].
3031 * If some of the divs of info[j] are unknown, then the expanded info[i]
3032 * will not have the corresponding div constraints. The other patterns
3033 * therefore cannot apply. Skip the computation in this case.
3035 * The expansion is performed using the divs "div" and expansion "exp"
3036 * computed by the caller.
3037 * info[i].bmap has already been expanded and the result is passed in
3038 * as "bmap".
3039 * The "eq" and "ineq" fields of info[i] reflect the status of
3040 * the constraints of the expanded "bmap" with respect to info[j].tab.
3041 * However, inequality constraints that are redundant in info[i].tab
3042 * have not yet been marked as such because no tableau was available.
3044 * Replace info[i].bmap by "bmap" and expand info[i].tab as well,
3045 * updating info[i].ineq with respect to the redundant constraints.
3046 * Then try and coalesce the expanded info[i] with info[j],
3047 * reusing the information in info[i].eq and info[i].ineq.
3048 * If this does not result in any coalescing or if it results in info[j]
3049 * getting dropped (which should not happen in practice, since the case
3050 * of info[j] being a subset of info[i] has already been checked by
3051 * the caller), then revert info[i] to its original state.
3053 static enum isl_change coalesce_expand_tab_divs(__isl_take isl_basic_map *bmap,
3054 int i, int j, struct isl_coalesce_info *info, __isl_keep isl_mat *div,
3055 int *exp)
3057 isl_bool known;
3058 isl_basic_map *bmap_i;
3059 struct isl_tab_undo *snap;
3060 enum isl_change change = isl_change_none;
3062 known = isl_basic_map_divs_known(info[j].bmap);
3063 if (known < 0 || !known) {
3064 clear_status(&info[i]);
3065 isl_basic_map_free(bmap);
3066 return known < 0 ? isl_change_error : isl_change_none;
3069 bmap_i = isl_basic_map_copy(info[i].bmap);
3070 snap = isl_tab_snap(info[i].tab);
3071 if (expand_tab(&info[i], exp, bmap) < 0)
3072 change = isl_change_error;
3074 init_status(&info[j]);
3075 if (change == isl_change_none)
3076 change = coalesce_local_pair_reuse(i, j, info);
3077 else
3078 clear_status(&info[i]);
3079 if (change != isl_change_none && change != isl_change_drop_second) {
3080 isl_basic_map_free(bmap_i);
3081 } else {
3082 isl_basic_map_free(info[i].bmap);
3083 info[i].bmap = bmap_i;
3085 if (isl_tab_rollback(info[i].tab, snap) < 0)
3086 change = isl_change_error;
3089 return change;
3092 /* Check if the union of "bmap" and the basic map represented by info[j]
3093 * can be represented by a single basic map,
3094 * after expanding the divs of "bmap" to match those of info[j].
3095 * If so, replace the pair by the single basic map and return
3096 * isl_change_drop_first, isl_change_drop_second or isl_change_fuse.
3097 * Otherwise, return isl_change_none.
3099 * In particular, check if the expanded "bmap" contains the basic map
3100 * represented by the tableau info[j].tab.
3101 * The expansion is performed using the divs "div" and expansion "exp"
3102 * computed by the caller.
3103 * Then we check if all constraints of the expanded "bmap" are valid for
3104 * info[j].tab.
3106 * If "i" is not equal to -1, then "bmap" is equal to info[i].bmap.
3107 * In this case, the positions of the constraints of info[i].bmap
3108 * with respect to the basic map represented by info[j] are stored
3109 * in info[i].
3111 * If the expanded "bmap" does not contain the basic map
3112 * represented by the tableau info[j].tab and if "i" is not -1,
3113 * i.e., if the original "bmap" is info[i].bmap, then expand info[i].tab
3114 * as well and check if that results in coalescing.
3116 static enum isl_change coalesce_with_expanded_divs(
3117 __isl_keep isl_basic_map *bmap, int i, int j,
3118 struct isl_coalesce_info *info, __isl_keep isl_mat *div, int *exp)
3120 enum isl_change change = isl_change_none;
3121 struct isl_coalesce_info info_local, *info_i;
3123 info_i = i >= 0 ? &info[i] : &info_local;
3124 init_status(info_i);
3125 bmap = isl_basic_map_copy(bmap);
3126 bmap = isl_basic_map_expand_divs(bmap, isl_mat_copy(div), exp);
3127 bmap = isl_basic_map_mark_final(bmap);
3129 if (!bmap)
3130 goto error;
3132 info_local.bmap = bmap;
3133 info_i->eq = eq_status_in(bmap, info[j].tab);
3134 if (bmap->n_eq && !info_i->eq)
3135 goto error;
3136 if (any_eq(info_i, STATUS_ERROR))
3137 goto error;
3138 if (any_eq(info_i, STATUS_SEPARATE))
3139 goto done;
3141 info_i->ineq = ineq_status_in(bmap, NULL, info[j].tab);
3142 if (bmap->n_ineq && !info_i->ineq)
3143 goto error;
3144 if (any_ineq(info_i, STATUS_ERROR))
3145 goto error;
3146 if (any_ineq(info_i, STATUS_SEPARATE))
3147 goto done;
3149 if (all(info_i->eq, 2 * bmap->n_eq, STATUS_VALID) &&
3150 all(info_i->ineq, bmap->n_ineq, STATUS_VALID)) {
3151 drop(&info[j]);
3152 change = isl_change_drop_second;
3155 if (change == isl_change_none && i != -1)
3156 return coalesce_expand_tab_divs(bmap, i, j, info, div, exp);
3158 done:
3159 isl_basic_map_free(bmap);
3160 clear_status(info_i);
3161 return change;
3162 error:
3163 isl_basic_map_free(bmap);
3164 clear_status(info_i);
3165 return isl_change_error;
3168 /* Check if the union of "bmap_i" and the basic map represented by info[j]
3169 * can be represented by a single basic map,
3170 * after aligning the divs of "bmap_i" to match those of info[j].
3171 * If so, replace the pair by the single basic map and return
3172 * isl_change_drop_first, isl_change_drop_second or isl_change_fuse.
3173 * Otherwise, return isl_change_none.
3175 * In particular, check if "bmap_i" contains the basic map represented by
3176 * info[j] after aligning the divs of "bmap_i" to those of info[j].
3177 * Note that this can only succeed if the number of divs of "bmap_i"
3178 * is smaller than (or equal to) the number of divs of info[j].
3180 * We first check if the divs of "bmap_i" are all known and form a subset
3181 * of those of info[j].bmap. If so, we pass control over to
3182 * coalesce_with_expanded_divs.
3184 * If "i" is not equal to -1, then "bmap" is equal to info[i].bmap.
3186 static enum isl_change coalesce_after_aligning_divs(
3187 __isl_keep isl_basic_map *bmap_i, int i, int j,
3188 struct isl_coalesce_info *info)
3190 isl_bool known;
3191 isl_mat *div_i, *div_j, *div;
3192 int *exp1 = NULL;
3193 int *exp2 = NULL;
3194 isl_ctx *ctx;
3195 enum isl_change change;
3197 known = isl_basic_map_divs_known(bmap_i);
3198 if (known < 0)
3199 return isl_change_error;
3200 if (!known)
3201 return isl_change_none;
3203 ctx = isl_basic_map_get_ctx(bmap_i);
3205 div_i = isl_basic_map_get_divs(bmap_i);
3206 div_j = isl_basic_map_get_divs(info[j].bmap);
3208 if (!div_i || !div_j)
3209 goto error;
3211 exp1 = isl_alloc_array(ctx, int, div_i->n_row);
3212 exp2 = isl_alloc_array(ctx, int, div_j->n_row);
3213 if ((div_i->n_row && !exp1) || (div_j->n_row && !exp2))
3214 goto error;
3216 div = isl_merge_divs(div_i, div_j, exp1, exp2);
3217 if (!div)
3218 goto error;
3220 if (div->n_row == div_j->n_row)
3221 change = coalesce_with_expanded_divs(bmap_i,
3222 i, j, info, div, exp1);
3223 else
3224 change = isl_change_none;
3226 isl_mat_free(div);
3228 isl_mat_free(div_i);
3229 isl_mat_free(div_j);
3231 free(exp2);
3232 free(exp1);
3234 return change;
3235 error:
3236 isl_mat_free(div_i);
3237 isl_mat_free(div_j);
3238 free(exp1);
3239 free(exp2);
3240 return isl_change_error;
3243 /* Check if basic map "j" is a subset of basic map "i" after
3244 * exploiting the extra equalities of "j" to simplify the divs of "i".
3245 * If so, remove basic map "j" and return isl_change_drop_second.
3247 * If "j" does not have any equalities or if they are the same
3248 * as those of "i", then we cannot exploit them to simplify the divs.
3249 * Similarly, if there are no divs in "i", then they cannot be simplified.
3250 * If, on the other hand, the affine hulls of "i" and "j" do not intersect,
3251 * then "j" cannot be a subset of "i".
3253 * Otherwise, we intersect "i" with the affine hull of "j" and then
3254 * check if "j" is a subset of the result after aligning the divs.
3255 * If so, then "j" is definitely a subset of "i" and can be removed.
3256 * Note that if after intersection with the affine hull of "j".
3257 * "i" still has more divs than "j", then there is no way we can
3258 * align the divs of "i" to those of "j".
3260 static enum isl_change coalesce_subset_with_equalities(int i, int j,
3261 struct isl_coalesce_info *info)
3263 isl_basic_map *hull_i, *hull_j, *bmap_i;
3264 int equal, empty;
3265 enum isl_change change;
3267 if (info[j].bmap->n_eq == 0)
3268 return isl_change_none;
3269 if (info[i].bmap->n_div == 0)
3270 return isl_change_none;
3272 hull_i = isl_basic_map_copy(info[i].bmap);
3273 hull_i = isl_basic_map_plain_affine_hull(hull_i);
3274 hull_j = isl_basic_map_copy(info[j].bmap);
3275 hull_j = isl_basic_map_plain_affine_hull(hull_j);
3277 hull_j = isl_basic_map_intersect(hull_j, isl_basic_map_copy(hull_i));
3278 equal = isl_basic_map_plain_is_equal(hull_i, hull_j);
3279 empty = isl_basic_map_plain_is_empty(hull_j);
3280 isl_basic_map_free(hull_i);
3282 if (equal < 0 || equal || empty < 0 || empty) {
3283 isl_basic_map_free(hull_j);
3284 if (equal < 0 || empty < 0)
3285 return isl_change_error;
3286 return isl_change_none;
3289 bmap_i = isl_basic_map_copy(info[i].bmap);
3290 bmap_i = isl_basic_map_intersect(bmap_i, hull_j);
3291 if (!bmap_i)
3292 return isl_change_error;
3294 if (bmap_i->n_div > info[j].bmap->n_div) {
3295 isl_basic_map_free(bmap_i);
3296 return isl_change_none;
3299 change = coalesce_after_aligning_divs(bmap_i, -1, j, info);
3301 isl_basic_map_free(bmap_i);
3303 return change;
3306 /* Check if the union of and the basic maps represented by info[i] and info[j]
3307 * can be represented by a single basic map, by aligning or equating
3308 * their integer divisions.
3309 * If so, replace the pair by the single basic map and return
3310 * isl_change_drop_first, isl_change_drop_second or isl_change_fuse.
3311 * Otherwise, return isl_change_none.
3313 * Note that we only perform any test if the number of divs is different
3314 * in the two basic maps. In case the number of divs is the same,
3315 * we have already established that the divs are different
3316 * in the two basic maps.
3317 * In particular, if the number of divs of basic map i is smaller than
3318 * the number of divs of basic map j, then we check if j is a subset of i
3319 * and vice versa.
3321 static enum isl_change coalesce_divs(int i, int j,
3322 struct isl_coalesce_info *info)
3324 enum isl_change change = isl_change_none;
3326 if (info[i].bmap->n_div < info[j].bmap->n_div)
3327 change = coalesce_after_aligning_divs(info[i].bmap, i, j, info);
3328 if (change != isl_change_none)
3329 return change;
3331 if (info[j].bmap->n_div < info[i].bmap->n_div)
3332 change = coalesce_after_aligning_divs(info[j].bmap, j, i, info);
3333 if (change != isl_change_none)
3334 return invert_change(change);
3336 change = coalesce_subset_with_equalities(i, j, info);
3337 if (change != isl_change_none)
3338 return change;
3340 change = coalesce_subset_with_equalities(j, i, info);
3341 if (change != isl_change_none)
3342 return invert_change(change);
3344 return isl_change_none;
3347 /* Does "bmap" involve any divs that themselves refer to divs?
3349 static isl_bool has_nested_div(__isl_keep isl_basic_map *bmap)
3351 int i;
3352 isl_size total;
3353 isl_size n_div;
3355 total = isl_basic_map_dim(bmap, isl_dim_all);
3356 n_div = isl_basic_map_dim(bmap, isl_dim_div);
3357 if (total < 0 || n_div < 0)
3358 return isl_bool_error;
3359 total -= n_div;
3361 for (i = 0; i < n_div; ++i)
3362 if (isl_seq_first_non_zero(bmap->div[i] + 2 + total,
3363 n_div) != -1)
3364 return isl_bool_true;
3366 return isl_bool_false;
3369 /* Return a list of affine expressions, one for each integer division
3370 * in "bmap_i". For each integer division that also appears in "bmap_j",
3371 * the affine expression is set to NaN. The number of NaNs in the list
3372 * is equal to the number of integer divisions in "bmap_j".
3373 * For the other integer divisions of "bmap_i", the corresponding
3374 * element in the list is a purely affine expression equal to the integer
3375 * division in "hull".
3376 * If no such list can be constructed, then the number of elements
3377 * in the returned list is smaller than the number of integer divisions
3378 * in "bmap_i".
3380 static __isl_give isl_aff_list *set_up_substitutions(
3381 __isl_keep isl_basic_map *bmap_i, __isl_keep isl_basic_map *bmap_j,
3382 __isl_take isl_basic_map *hull)
3384 isl_size n_div_i, n_div_j, total;
3385 isl_ctx *ctx;
3386 isl_local_space *ls;
3387 isl_basic_set *wrap_hull;
3388 isl_aff *aff_nan;
3389 isl_aff_list *list;
3390 int i, j;
3392 n_div_i = isl_basic_map_dim(bmap_i, isl_dim_div);
3393 n_div_j = isl_basic_map_dim(bmap_j, isl_dim_div);
3394 total = isl_basic_map_dim(bmap_i, isl_dim_all);
3395 if (!hull || n_div_i < 0 || n_div_j < 0 || total < 0)
3396 return NULL;
3398 ctx = isl_basic_map_get_ctx(hull);
3399 total -= n_div_i;
3401 ls = isl_basic_map_get_local_space(bmap_i);
3402 ls = isl_local_space_wrap(ls);
3403 wrap_hull = isl_basic_map_wrap(hull);
3405 aff_nan = isl_aff_nan_on_domain(isl_local_space_copy(ls));
3406 list = isl_aff_list_alloc(ctx, n_div_i);
3408 j = 0;
3409 for (i = 0; i < n_div_i; ++i) {
3410 isl_aff *aff;
3411 isl_size n_div;
3413 if (j < n_div_j &&
3414 isl_basic_map_equal_div_expr_part(bmap_i, i, bmap_j, j,
3415 0, 2 + total)) {
3416 ++j;
3417 list = isl_aff_list_add(list, isl_aff_copy(aff_nan));
3418 continue;
3420 if (n_div_i - i <= n_div_j - j)
3421 break;
3423 aff = isl_local_space_get_div(ls, i);
3424 aff = isl_aff_substitute_equalities(aff,
3425 isl_basic_set_copy(wrap_hull));
3426 aff = isl_aff_floor(aff);
3427 n_div = isl_aff_dim(aff, isl_dim_div);
3428 if (n_div < 0)
3429 goto error;
3430 if (n_div != 0) {
3431 isl_aff_free(aff);
3432 break;
3435 list = isl_aff_list_add(list, aff);
3438 isl_aff_free(aff_nan);
3439 isl_local_space_free(ls);
3440 isl_basic_set_free(wrap_hull);
3442 return list;
3443 error:
3444 isl_aff_free(aff_nan);
3445 isl_local_space_free(ls);
3446 isl_basic_set_free(wrap_hull);
3447 isl_aff_list_free(list);
3448 return NULL;
3451 /* Add variables to info->bmap and info->tab corresponding to the elements
3452 * in "list" that are not set to NaN.
3453 * "extra_var" is the number of these elements.
3454 * "dim" is the offset in the variables of "tab" where we should
3455 * start considering the elements in "list".
3456 * When this function returns, the total number of variables in "tab"
3457 * is equal to "dim" plus the number of elements in "list".
3459 * The newly added existentially quantified variables are not given
3460 * an explicit representation because the corresponding div constraints
3461 * do not appear in info->bmap. These constraints are not added
3462 * to info->bmap because for internal consistency, they would need to
3463 * be added to info->tab as well, where they could combine with the equality
3464 * that is added later to result in constraints that do not hold
3465 * in the original input.
3467 static isl_stat add_sub_vars(struct isl_coalesce_info *info,
3468 __isl_keep isl_aff_list *list, int dim, int extra_var)
3470 int i, j, d;
3471 isl_size n;
3472 isl_space *space;
3474 space = isl_basic_map_get_space(info->bmap);
3475 info->bmap = isl_basic_map_cow(info->bmap);
3476 info->bmap = isl_basic_map_extend_space(info->bmap, space,
3477 extra_var, 0, 0);
3478 n = isl_aff_list_n_aff(list);
3479 if (!info->bmap || n < 0)
3480 return isl_stat_error;
3481 for (i = 0; i < n; ++i) {
3482 int is_nan;
3483 isl_aff *aff;
3485 aff = isl_aff_list_get_aff(list, i);
3486 is_nan = isl_aff_is_nan(aff);
3487 isl_aff_free(aff);
3488 if (is_nan < 0)
3489 return isl_stat_error;
3490 if (is_nan)
3491 continue;
3493 if (isl_tab_insert_var(info->tab, dim + i) < 0)
3494 return isl_stat_error;
3495 d = isl_basic_map_alloc_div(info->bmap);
3496 if (d < 0)
3497 return isl_stat_error;
3498 info->bmap = isl_basic_map_mark_div_unknown(info->bmap, d);
3499 for (j = d; j > i; --j)
3500 info->bmap = isl_basic_map_swap_div(info->bmap,
3501 j - 1, j);
3502 if (!info->bmap)
3503 return isl_stat_error;
3506 return isl_stat_ok;
3509 /* For each element in "list" that is not set to NaN, fix the corresponding
3510 * variable in "tab" to the purely affine expression defined by the element.
3511 * "dim" is the offset in the variables of "tab" where we should
3512 * start considering the elements in "list".
3514 * This function assumes that a sufficient number of rows and
3515 * elements in the constraint array are available in the tableau.
3517 static isl_stat add_sub_equalities(struct isl_tab *tab,
3518 __isl_keep isl_aff_list *list, int dim)
3520 int i;
3521 isl_size n;
3522 isl_ctx *ctx;
3523 isl_vec *sub;
3524 isl_aff *aff;
3526 n = isl_aff_list_n_aff(list);
3527 if (n < 0)
3528 return isl_stat_error;
3530 ctx = isl_tab_get_ctx(tab);
3531 sub = isl_vec_alloc(ctx, 1 + dim + n);
3532 if (!sub)
3533 return isl_stat_error;
3534 isl_seq_clr(sub->el + 1 + dim, n);
3536 for (i = 0; i < n; ++i) {
3537 aff = isl_aff_list_get_aff(list, i);
3538 if (!aff)
3539 goto error;
3540 if (isl_aff_is_nan(aff)) {
3541 isl_aff_free(aff);
3542 continue;
3544 isl_seq_cpy(sub->el, aff->v->el + 1, 1 + dim);
3545 isl_int_neg(sub->el[1 + dim + i], aff->v->el[0]);
3546 if (isl_tab_add_eq(tab, sub->el) < 0)
3547 goto error;
3548 isl_int_set_si(sub->el[1 + dim + i], 0);
3549 isl_aff_free(aff);
3552 isl_vec_free(sub);
3553 return isl_stat_ok;
3554 error:
3555 isl_aff_free(aff);
3556 isl_vec_free(sub);
3557 return isl_stat_error;
3560 /* Add variables to info->tab and info->bmap corresponding to the elements
3561 * in "list" that are not set to NaN. The value of the added variable
3562 * in info->tab is fixed to the purely affine expression defined by the element.
3563 * "dim" is the offset in the variables of info->tab where we should
3564 * start considering the elements in "list".
3565 * When this function returns, the total number of variables in info->tab
3566 * is equal to "dim" plus the number of elements in "list".
3568 static isl_stat add_subs(struct isl_coalesce_info *info,
3569 __isl_keep isl_aff_list *list, int dim)
3571 int extra_var;
3572 isl_size n;
3574 n = isl_aff_list_n_aff(list);
3575 if (n < 0)
3576 return isl_stat_error;
3578 extra_var = n - (info->tab->n_var - dim);
3580 if (isl_tab_extend_vars(info->tab, extra_var) < 0)
3581 return isl_stat_error;
3582 if (isl_tab_extend_cons(info->tab, 2 * extra_var) < 0)
3583 return isl_stat_error;
3584 if (add_sub_vars(info, list, dim, extra_var) < 0)
3585 return isl_stat_error;
3587 return add_sub_equalities(info->tab, list, dim);
3590 /* Coalesce basic map "j" into basic map "i" after adding the extra integer
3591 * divisions in "i" but not in "j" to basic map "j", with values
3592 * specified by "list". The total number of elements in "list"
3593 * is equal to the number of integer divisions in "i", while the number
3594 * of NaN elements in the list is equal to the number of integer divisions
3595 * in "j".
3597 * If no coalescing can be performed, then we need to revert basic map "j"
3598 * to its original state. We do the same if basic map "i" gets dropped
3599 * during the coalescing, even though this should not happen in practice
3600 * since we have already checked for "j" being a subset of "i"
3601 * before we reach this stage.
3603 static enum isl_change coalesce_with_subs(int i, int j,
3604 struct isl_coalesce_info *info, __isl_keep isl_aff_list *list)
3606 isl_basic_map *bmap_j;
3607 struct isl_tab_undo *snap;
3608 isl_size dim, n_div;
3609 enum isl_change change;
3611 bmap_j = isl_basic_map_copy(info[j].bmap);
3612 snap = isl_tab_snap(info[j].tab);
3614 dim = isl_basic_map_dim(bmap_j, isl_dim_all);
3615 n_div = isl_basic_map_dim(bmap_j, isl_dim_div);
3616 if (dim < 0 || n_div < 0)
3617 goto error;
3618 dim -= n_div;
3619 if (add_subs(&info[j], list, dim) < 0)
3620 goto error;
3622 change = coalesce_local_pair(i, j, info);
3623 if (change != isl_change_none && change != isl_change_drop_first) {
3624 isl_basic_map_free(bmap_j);
3625 } else {
3626 isl_basic_map_free(info[j].bmap);
3627 info[j].bmap = bmap_j;
3629 if (isl_tab_rollback(info[j].tab, snap) < 0)
3630 return isl_change_error;
3633 return change;
3634 error:
3635 isl_basic_map_free(bmap_j);
3636 return isl_change_error;
3639 /* Check if we can coalesce basic map "j" into basic map "i" after copying
3640 * those extra integer divisions in "i" that can be simplified away
3641 * using the extra equalities in "j".
3642 * All divs are assumed to be known and not contain any nested divs.
3644 * We first check if there are any extra equalities in "j" that we
3645 * can exploit. Then we check if every integer division in "i"
3646 * either already appears in "j" or can be simplified using the
3647 * extra equalities to a purely affine expression.
3648 * If these tests succeed, then we try to coalesce the two basic maps
3649 * by introducing extra dimensions in "j" corresponding to
3650 * the extra integer divsisions "i" fixed to the corresponding
3651 * purely affine expression.
3653 static enum isl_change check_coalesce_into_eq(int i, int j,
3654 struct isl_coalesce_info *info)
3656 isl_size n_div_i, n_div_j, n;
3657 isl_basic_map *hull_i, *hull_j;
3658 isl_bool equal, empty;
3659 isl_aff_list *list;
3660 enum isl_change change;
3662 n_div_i = isl_basic_map_dim(info[i].bmap, isl_dim_div);
3663 n_div_j = isl_basic_map_dim(info[j].bmap, isl_dim_div);
3664 if (n_div_i < 0 || n_div_j < 0)
3665 return isl_change_error;
3666 if (n_div_i <= n_div_j)
3667 return isl_change_none;
3668 if (info[j].bmap->n_eq == 0)
3669 return isl_change_none;
3671 hull_i = isl_basic_map_copy(info[i].bmap);
3672 hull_i = isl_basic_map_plain_affine_hull(hull_i);
3673 hull_j = isl_basic_map_copy(info[j].bmap);
3674 hull_j = isl_basic_map_plain_affine_hull(hull_j);
3676 hull_j = isl_basic_map_intersect(hull_j, isl_basic_map_copy(hull_i));
3677 equal = isl_basic_map_plain_is_equal(hull_i, hull_j);
3678 empty = isl_basic_map_plain_is_empty(hull_j);
3679 isl_basic_map_free(hull_i);
3681 if (equal < 0 || empty < 0)
3682 goto error;
3683 if (equal || empty) {
3684 isl_basic_map_free(hull_j);
3685 return isl_change_none;
3688 list = set_up_substitutions(info[i].bmap, info[j].bmap, hull_j);
3689 if (!list)
3690 return isl_change_error;
3691 n = isl_aff_list_n_aff(list);
3692 if (n < 0)
3693 change = isl_change_error;
3694 else if (n < n_div_i)
3695 change = isl_change_none;
3696 else
3697 change = coalesce_with_subs(i, j, info, list);
3699 isl_aff_list_free(list);
3701 return change;
3702 error:
3703 isl_basic_map_free(hull_j);
3704 return isl_change_error;
3707 /* Check if we can coalesce basic maps "i" and "j" after copying
3708 * those extra integer divisions in one of the basic maps that can
3709 * be simplified away using the extra equalities in the other basic map.
3710 * We require all divs to be known in both basic maps.
3711 * Furthermore, to simplify the comparison of div expressions,
3712 * we do not allow any nested integer divisions.
3714 static enum isl_change check_coalesce_eq(int i, int j,
3715 struct isl_coalesce_info *info)
3717 isl_bool known, nested;
3718 enum isl_change change;
3720 known = isl_basic_map_divs_known(info[i].bmap);
3721 if (known < 0 || !known)
3722 return known < 0 ? isl_change_error : isl_change_none;
3723 known = isl_basic_map_divs_known(info[j].bmap);
3724 if (known < 0 || !known)
3725 return known < 0 ? isl_change_error : isl_change_none;
3726 nested = has_nested_div(info[i].bmap);
3727 if (nested < 0 || nested)
3728 return nested < 0 ? isl_change_error : isl_change_none;
3729 nested = has_nested_div(info[j].bmap);
3730 if (nested < 0 || nested)
3731 return nested < 0 ? isl_change_error : isl_change_none;
3733 change = check_coalesce_into_eq(i, j, info);
3734 if (change != isl_change_none)
3735 return change;
3736 change = check_coalesce_into_eq(j, i, info);
3737 if (change != isl_change_none)
3738 return invert_change(change);
3740 return isl_change_none;
3743 /* Check if the union of the given pair of basic maps
3744 * can be represented by a single basic map.
3745 * If so, replace the pair by the single basic map and return
3746 * isl_change_drop_first, isl_change_drop_second or isl_change_fuse.
3747 * Otherwise, return isl_change_none.
3749 * We first check if the two basic maps live in the same local space,
3750 * after aligning the divs that differ by only an integer constant.
3751 * If so, we do the complete check. Otherwise, we check if they have
3752 * the same number of integer divisions and can be coalesced, if one is
3753 * an obvious subset of the other or if the extra integer divisions
3754 * of one basic map can be simplified away using the extra equalities
3755 * of the other basic map.
3757 * Note that trying to coalesce pairs of disjuncts with the same
3758 * number, but different local variables may drop the explicit
3759 * representation of some of these local variables.
3760 * This operation is therefore not performed when
3761 * the "coalesce_preserve_locals" option is set.
3763 static enum isl_change coalesce_pair(int i, int j,
3764 struct isl_coalesce_info *info)
3766 int preserve;
3767 isl_bool same;
3768 enum isl_change change;
3769 isl_ctx *ctx;
3771 if (harmonize_divs(&info[i], &info[j]) < 0)
3772 return isl_change_error;
3773 same = same_divs(info[i].bmap, info[j].bmap);
3774 if (same < 0)
3775 return isl_change_error;
3776 if (same)
3777 return coalesce_local_pair(i, j, info);
3779 ctx = isl_basic_map_get_ctx(info[i].bmap);
3780 preserve = isl_options_get_coalesce_preserve_locals(ctx);
3781 if (!preserve && info[i].bmap->n_div == info[j].bmap->n_div) {
3782 change = coalesce_local_pair(i, j, info);
3783 if (change != isl_change_none)
3784 return change;
3787 change = coalesce_divs(i, j, info);
3788 if (change != isl_change_none)
3789 return change;
3791 return check_coalesce_eq(i, j, info);
3794 /* Return the maximum of "a" and "b".
3796 static int isl_max(int a, int b)
3798 return a > b ? a : b;
3801 /* Pairwise coalesce the basic maps in the range [start1, end1[ of "info"
3802 * with those in the range [start2, end2[, skipping basic maps
3803 * that have been removed (either before or within this function).
3805 * For each basic map i in the first range, we check if it can be coalesced
3806 * with respect to any previously considered basic map j in the second range.
3807 * If i gets dropped (because it was a subset of some j), then
3808 * we can move on to the next basic map.
3809 * If j gets dropped, we need to continue checking against the other
3810 * previously considered basic maps.
3811 * If the two basic maps got fused, then we recheck the fused basic map
3812 * against the previously considered basic maps, starting at i + 1
3813 * (even if start2 is greater than i + 1).
3815 static int coalesce_range(isl_ctx *ctx, struct isl_coalesce_info *info,
3816 int start1, int end1, int start2, int end2)
3818 int i, j;
3820 for (i = end1 - 1; i >= start1; --i) {
3821 if (info[i].removed)
3822 continue;
3823 for (j = isl_max(i + 1, start2); j < end2; ++j) {
3824 enum isl_change changed;
3826 if (info[j].removed)
3827 continue;
3828 if (info[i].removed)
3829 isl_die(ctx, isl_error_internal,
3830 "basic map unexpectedly removed",
3831 return -1);
3832 changed = coalesce_pair(i, j, info);
3833 switch (changed) {
3834 case isl_change_error:
3835 return -1;
3836 case isl_change_none:
3837 case isl_change_drop_second:
3838 continue;
3839 case isl_change_drop_first:
3840 j = end2;
3841 break;
3842 case isl_change_fuse:
3843 j = i;
3844 break;
3849 return 0;
3852 /* Pairwise coalesce the basic maps described by the "n" elements of "info".
3854 * We consider groups of basic maps that live in the same apparent
3855 * affine hull and we first coalesce within such a group before we
3856 * coalesce the elements in the group with elements of previously
3857 * considered groups. If a fuse happens during the second phase,
3858 * then we also reconsider the elements within the group.
3860 static int coalesce(isl_ctx *ctx, int n, struct isl_coalesce_info *info)
3862 int start, end;
3864 for (end = n; end > 0; end = start) {
3865 start = end - 1;
3866 while (start >= 1 &&
3867 info[start - 1].hull_hash == info[start].hull_hash)
3868 start--;
3869 if (coalesce_range(ctx, info, start, end, start, end) < 0)
3870 return -1;
3871 if (coalesce_range(ctx, info, start, end, end, n) < 0)
3872 return -1;
3875 return 0;
3878 /* Update the basic maps in "map" based on the information in "info".
3879 * In particular, remove the basic maps that have been marked removed and
3880 * update the others based on the information in the corresponding tableau.
3881 * Since we detected implicit equalities without calling
3882 * isl_basic_map_gauss, we need to do it now.
3883 * Also call isl_basic_map_simplify if we may have lost the definition
3884 * of one or more integer divisions.
3886 static __isl_give isl_map *update_basic_maps(__isl_take isl_map *map,
3887 int n, struct isl_coalesce_info *info)
3889 int i;
3891 if (!map)
3892 return NULL;
3894 for (i = n - 1; i >= 0; --i) {
3895 if (info[i].removed) {
3896 isl_basic_map_free(map->p[i]);
3897 if (i != map->n - 1)
3898 map->p[i] = map->p[map->n - 1];
3899 map->n--;
3900 continue;
3903 info[i].bmap = isl_basic_map_update_from_tab(info[i].bmap,
3904 info[i].tab);
3905 info[i].bmap = isl_basic_map_gauss(info[i].bmap, NULL);
3906 if (info[i].simplify)
3907 info[i].bmap = isl_basic_map_simplify(info[i].bmap);
3908 info[i].bmap = isl_basic_map_finalize(info[i].bmap);
3909 if (!info[i].bmap)
3910 return isl_map_free(map);
3911 ISL_F_SET(info[i].bmap, ISL_BASIC_MAP_NO_IMPLICIT);
3912 ISL_F_SET(info[i].bmap, ISL_BASIC_MAP_NO_REDUNDANT);
3913 isl_basic_map_free(map->p[i]);
3914 map->p[i] = info[i].bmap;
3915 info[i].bmap = NULL;
3918 return map;
3921 /* For each pair of basic maps in the map, check if the union of the two
3922 * can be represented by a single basic map.
3923 * If so, replace the pair by the single basic map and start over.
3925 * We factor out any (hidden) common factor from the constraint
3926 * coefficients to improve the detection of adjacent constraints.
3927 * Note that this function does not call isl_basic_map_gauss,
3928 * but it does make sure that only a single copy of the basic map
3929 * is affected. This means that isl_basic_map_gauss may have
3930 * to be called at the end of the computation (in update_basic_maps)
3931 * on this single copy to ensure that
3932 * the basic maps are not left in an unexpected state.
3934 * Since we are constructing the tableaus of the basic maps anyway,
3935 * we exploit them to detect implicit equalities and redundant constraints.
3936 * This also helps the coalescing as it can ignore the redundant constraints.
3937 * In order to avoid confusion, we make all implicit equalities explicit
3938 * in the basic maps. If the basic map only has a single reference
3939 * (this happens in particular if it was modified by
3940 * isl_basic_map_reduce_coefficients), then isl_basic_map_gauss
3941 * does not get called on the result. The call to
3942 * isl_basic_map_gauss in update_basic_maps resolves this as well.
3943 * For each basic map, we also compute the hash of the apparent affine hull
3944 * for use in coalesce.
3946 __isl_give isl_map *isl_map_coalesce(__isl_take isl_map *map)
3948 int i;
3949 unsigned n;
3950 isl_ctx *ctx;
3951 struct isl_coalesce_info *info = NULL;
3953 map = isl_map_remove_empty_parts(map);
3954 if (!map)
3955 return NULL;
3957 if (map->n <= 1)
3958 return map;
3960 ctx = isl_map_get_ctx(map);
3961 map = isl_map_sort_divs(map);
3962 map = isl_map_cow(map);
3964 if (!map)
3965 return NULL;
3967 n = map->n;
3969 info = isl_calloc_array(map->ctx, struct isl_coalesce_info, n);
3970 if (!info)
3971 goto error;
3973 for (i = 0; i < map->n; ++i) {
3974 map->p[i] = isl_basic_map_reduce_coefficients(map->p[i]);
3975 if (!map->p[i])
3976 goto error;
3977 info[i].bmap = isl_basic_map_copy(map->p[i]);
3978 info[i].tab = isl_tab_from_basic_map(info[i].bmap, 0);
3979 if (!info[i].tab)
3980 goto error;
3981 if (!ISL_F_ISSET(info[i].bmap, ISL_BASIC_MAP_NO_IMPLICIT))
3982 if (isl_tab_detect_implicit_equalities(info[i].tab) < 0)
3983 goto error;
3984 info[i].bmap = isl_tab_make_equalities_explicit(info[i].tab,
3985 info[i].bmap);
3986 if (!info[i].bmap)
3987 goto error;
3988 if (!ISL_F_ISSET(info[i].bmap, ISL_BASIC_MAP_NO_REDUNDANT))
3989 if (isl_tab_detect_redundant(info[i].tab) < 0)
3990 goto error;
3991 if (coalesce_info_set_hull_hash(&info[i]) < 0)
3992 goto error;
3994 for (i = map->n - 1; i >= 0; --i)
3995 if (info[i].tab->empty)
3996 drop(&info[i]);
3998 if (coalesce(ctx, n, info) < 0)
3999 goto error;
4001 map = update_basic_maps(map, n, info);
4003 clear_coalesce_info(n, info);
4005 return map;
4006 error:
4007 clear_coalesce_info(n, info);
4008 isl_map_free(map);
4009 return NULL;
4012 /* For each pair of basic sets in the set, check if the union of the two
4013 * can be represented by a single basic set.
4014 * If so, replace the pair by the single basic set and start over.
4016 struct isl_set *isl_set_coalesce(struct isl_set *set)
4018 return set_from_map(isl_map_coalesce(set_to_map(set)));