isl_qpolynomial_eval: reenable evaluation of polynomials at rational points
[isl.git] / isl_coalesce.c
blob2aaf0e984767d1a6accd5b0ca5e9b807c123a99f
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 = isl_calloc_array(bmap_i->ctx, int, 2 * bmap_i->n_eq);
70 unsigned dim;
72 if (!eq)
73 return NULL;
75 dim = isl_basic_map_total_dim(bmap_i);
76 for (k = 0; k < bmap_i->n_eq; ++k) {
77 for (l = 0; l < 2; ++l) {
78 isl_seq_neg(bmap_i->eq[k], bmap_i->eq[k], 1+dim);
79 eq[2 * k + l] = status_in(bmap_i->eq[k], tab_j);
80 if (eq[2 * k + l] == STATUS_ERROR)
81 goto error;
85 return eq;
86 error:
87 free(eq);
88 return NULL;
91 /* Compute the position of the inequalities of basic map "bmap_i"
92 * (also represented by "tab_i", if not NULL) with respect to the basic map
93 * represented by "tab_j".
95 static int *ineq_status_in(__isl_keep isl_basic_map *bmap_i,
96 struct isl_tab *tab_i, struct isl_tab *tab_j)
98 int k;
99 unsigned n_eq = bmap_i->n_eq;
100 int *ineq = isl_calloc_array(bmap_i->ctx, int, bmap_i->n_ineq);
102 if (!ineq)
103 return NULL;
105 for (k = 0; k < bmap_i->n_ineq; ++k) {
106 if (tab_i && isl_tab_is_redundant(tab_i, n_eq + k)) {
107 ineq[k] = STATUS_REDUNDANT;
108 continue;
110 ineq[k] = status_in(bmap_i->ineq[k], tab_j);
111 if (ineq[k] == STATUS_ERROR)
112 goto error;
113 if (ineq[k] == STATUS_SEPARATE)
114 break;
117 return ineq;
118 error:
119 free(ineq);
120 return NULL;
123 static int any(int *con, unsigned len, int status)
125 int i;
127 for (i = 0; i < len ; ++i)
128 if (con[i] == status)
129 return 1;
130 return 0;
133 /* Return the first position of "status" in the list "con" of length "len".
134 * Return -1 if there is no such entry.
136 static int find(int *con, unsigned len, int status)
138 int i;
140 for (i = 0; i < len ; ++i)
141 if (con[i] == status)
142 return i;
143 return -1;
146 static int count(int *con, unsigned len, int status)
148 int i;
149 int c = 0;
151 for (i = 0; i < len ; ++i)
152 if (con[i] == status)
153 c++;
154 return c;
157 static int all(int *con, unsigned len, int status)
159 int i;
161 for (i = 0; i < len ; ++i) {
162 if (con[i] == STATUS_REDUNDANT)
163 continue;
164 if (con[i] != status)
165 return 0;
167 return 1;
170 /* Internal information associated to a basic map in a map
171 * that is to be coalesced by isl_map_coalesce.
173 * "bmap" is the basic map itself (or NULL if "removed" is set)
174 * "tab" is the corresponding tableau (or NULL if "removed" is set)
175 * "hull_hash" identifies the affine space in which "bmap" lives.
176 * "removed" is set if this basic map has been removed from the map
177 * "simplify" is set if this basic map may have some unknown integer
178 * divisions that were not present in the input basic maps. The basic
179 * map should then be simplified such that we may be able to find
180 * a definition among the constraints.
182 * "eq" and "ineq" are only set if we are currently trying to coalesce
183 * this basic map with another basic map, in which case they represent
184 * the position of the inequalities of this basic map with respect to
185 * the other basic map. The number of elements in the "eq" array
186 * is twice the number of equalities in the "bmap", corresponding
187 * to the two inequalities that make up each equality.
189 struct isl_coalesce_info {
190 isl_basic_map *bmap;
191 struct isl_tab *tab;
192 uint32_t hull_hash;
193 int removed;
194 int simplify;
195 int *eq;
196 int *ineq;
199 /* Is there any (half of an) equality constraint in the description
200 * of the basic map represented by "info" that
201 * has position "status" with respect to the other basic map?
203 static int any_eq(struct isl_coalesce_info *info, int status)
205 unsigned n_eq;
207 n_eq = isl_basic_map_n_equality(info->bmap);
208 return any(info->eq, 2 * n_eq, status);
211 /* Is there any inequality constraint in the description
212 * of the basic map represented by "info" that
213 * has position "status" with respect to the other basic map?
215 static int any_ineq(struct isl_coalesce_info *info, int status)
217 unsigned n_ineq;
219 n_ineq = isl_basic_map_n_inequality(info->bmap);
220 return any(info->ineq, n_ineq, status);
223 /* Return the position of the first half on an equality constraint
224 * in the description of the basic map represented by "info" that
225 * has position "status" with respect to the other basic map.
226 * The returned value is twice the position of the equality constraint
227 * plus zero for the negative half and plus one for the positive half.
228 * Return -1 if there is no such entry.
230 static int find_eq(struct isl_coalesce_info *info, int status)
232 unsigned n_eq;
234 n_eq = isl_basic_map_n_equality(info->bmap);
235 return find(info->eq, 2 * n_eq, status);
238 /* Return the position of the first inequality constraint in the description
239 * of the basic map represented by "info" that
240 * has position "status" with respect to the other basic map.
241 * Return -1 if there is no such entry.
243 static int find_ineq(struct isl_coalesce_info *info, int status)
245 unsigned n_ineq;
247 n_ineq = isl_basic_map_n_inequality(info->bmap);
248 return find(info->ineq, n_ineq, status);
251 /* Return the number of (halves of) equality constraints in the description
252 * of the basic map represented by "info" that
253 * have position "status" with respect to the other basic map.
255 static int count_eq(struct isl_coalesce_info *info, int status)
257 unsigned n_eq;
259 n_eq = isl_basic_map_n_equality(info->bmap);
260 return count(info->eq, 2 * n_eq, status);
263 /* Return the number of inequality constraints in the description
264 * of the basic map represented by "info" that
265 * have position "status" with respect to the other basic map.
267 static int count_ineq(struct isl_coalesce_info *info, int status)
269 unsigned n_ineq;
271 n_ineq = isl_basic_map_n_inequality(info->bmap);
272 return count(info->ineq, n_ineq, status);
275 /* Are all non-redundant constraints of the basic map represented by "info"
276 * either valid or cut constraints with respect to the other basic map?
278 static int all_valid_or_cut(struct isl_coalesce_info *info)
280 int i;
282 for (i = 0; i < 2 * info->bmap->n_eq; ++i) {
283 if (info->eq[i] == STATUS_REDUNDANT)
284 continue;
285 if (info->eq[i] == STATUS_VALID)
286 continue;
287 if (info->eq[i] == STATUS_CUT)
288 continue;
289 return 0;
292 for (i = 0; i < info->bmap->n_ineq; ++i) {
293 if (info->ineq[i] == STATUS_REDUNDANT)
294 continue;
295 if (info->ineq[i] == STATUS_VALID)
296 continue;
297 if (info->ineq[i] == STATUS_CUT)
298 continue;
299 return 0;
302 return 1;
305 /* Compute the hash of the (apparent) affine hull of info->bmap (with
306 * the existentially quantified variables removed) and store it
307 * in info->hash.
309 static int coalesce_info_set_hull_hash(struct isl_coalesce_info *info)
311 isl_basic_map *hull;
312 unsigned n_div;
314 hull = isl_basic_map_copy(info->bmap);
315 hull = isl_basic_map_plain_affine_hull(hull);
316 n_div = isl_basic_map_dim(hull, isl_dim_div);
317 hull = isl_basic_map_drop_constraints_involving_dims(hull,
318 isl_dim_div, 0, n_div);
319 info->hull_hash = isl_basic_map_get_hash(hull);
320 isl_basic_map_free(hull);
322 return hull ? 0 : -1;
325 /* Free all the allocated memory in an array
326 * of "n" isl_coalesce_info elements.
328 static void clear_coalesce_info(int n, struct isl_coalesce_info *info)
330 int i;
332 if (!info)
333 return;
335 for (i = 0; i < n; ++i) {
336 isl_basic_map_free(info[i].bmap);
337 isl_tab_free(info[i].tab);
340 free(info);
343 /* Clear the memory associated to"info".
344 * Gaussian elimination needs to be performed on the basic map
345 * before it gets freed because it may have been put
346 * in an inconsistent state in isl_map_coalesce while it may
347 * be shared with other maps.
349 static void clear(struct isl_coalesce_info *info)
351 info->bmap = isl_basic_map_gauss(info->bmap, NULL);
352 info->bmap = isl_basic_map_free(info->bmap);
353 isl_tab_free(info->tab);
354 info->tab = NULL;
357 /* Drop the basic map represented by "info".
358 * That is, clear the memory associated to the entry and
359 * mark it as having been removed.
361 static void drop(struct isl_coalesce_info *info)
363 clear(info);
364 info->removed = 1;
367 /* Exchange the information in "info1" with that in "info2".
369 static void exchange(struct isl_coalesce_info *info1,
370 struct isl_coalesce_info *info2)
372 struct isl_coalesce_info info;
374 info = *info1;
375 *info1 = *info2;
376 *info2 = info;
379 /* This type represents the kind of change that has been performed
380 * while trying to coalesce two basic maps.
382 * isl_change_none: nothing was changed
383 * isl_change_drop_first: the first basic map was removed
384 * isl_change_drop_second: the second basic map was removed
385 * isl_change_fuse: the two basic maps were replaced by a new basic map.
387 enum isl_change {
388 isl_change_error = -1,
389 isl_change_none = 0,
390 isl_change_drop_first,
391 isl_change_drop_second,
392 isl_change_fuse,
395 /* Update "change" based on an interchange of the first and the second
396 * basic map. That is, interchange isl_change_drop_first and
397 * isl_change_drop_second.
399 static enum isl_change invert_change(enum isl_change change)
401 switch (change) {
402 case isl_change_error:
403 return isl_change_error;
404 case isl_change_none:
405 return isl_change_none;
406 case isl_change_drop_first:
407 return isl_change_drop_second;
408 case isl_change_drop_second:
409 return isl_change_drop_first;
410 case isl_change_fuse:
411 return isl_change_fuse;
414 return isl_change_error;
417 /* Add the valid constraints of the basic map represented by "info"
418 * to "bmap". "len" is the size of the constraints.
419 * If only one of the pair of inequalities that make up an equality
420 * is valid, then add that inequality.
422 static __isl_give isl_basic_map *add_valid_constraints(
423 __isl_take isl_basic_map *bmap, struct isl_coalesce_info *info,
424 unsigned len)
426 int k, l;
428 if (!bmap)
429 return NULL;
431 for (k = 0; k < info->bmap->n_eq; ++k) {
432 if (info->eq[2 * k] == STATUS_VALID &&
433 info->eq[2 * k + 1] == STATUS_VALID) {
434 l = isl_basic_map_alloc_equality(bmap);
435 if (l < 0)
436 return isl_basic_map_free(bmap);
437 isl_seq_cpy(bmap->eq[l], info->bmap->eq[k], len);
438 } else if (info->eq[2 * k] == STATUS_VALID) {
439 l = isl_basic_map_alloc_inequality(bmap);
440 if (l < 0)
441 return isl_basic_map_free(bmap);
442 isl_seq_neg(bmap->ineq[l], info->bmap->eq[k], len);
443 } else if (info->eq[2 * k + 1] == STATUS_VALID) {
444 l = isl_basic_map_alloc_inequality(bmap);
445 if (l < 0)
446 return isl_basic_map_free(bmap);
447 isl_seq_cpy(bmap->ineq[l], info->bmap->eq[k], len);
451 for (k = 0; k < info->bmap->n_ineq; ++k) {
452 if (info->ineq[k] != STATUS_VALID)
453 continue;
454 l = isl_basic_map_alloc_inequality(bmap);
455 if (l < 0)
456 return isl_basic_map_free(bmap);
457 isl_seq_cpy(bmap->ineq[l], info->bmap->ineq[k], len);
460 return bmap;
463 /* Is "bmap" defined by a number of (non-redundant) constraints that
464 * is greater than the number of constraints of basic maps i and j combined?
465 * Equalities are counted as two inequalities.
467 static int number_of_constraints_increases(int i, int j,
468 struct isl_coalesce_info *info,
469 __isl_keep isl_basic_map *bmap, struct isl_tab *tab)
471 int k, n_old, n_new;
473 n_old = 2 * info[i].bmap->n_eq + info[i].bmap->n_ineq;
474 n_old += 2 * info[j].bmap->n_eq + info[j].bmap->n_ineq;
476 n_new = 2 * bmap->n_eq;
477 for (k = 0; k < bmap->n_ineq; ++k)
478 if (!isl_tab_is_redundant(tab, bmap->n_eq + k))
479 ++n_new;
481 return n_new > n_old;
484 /* Replace the pair of basic maps i and j by the basic map bounded
485 * by the valid constraints in both basic maps and the constraints
486 * in extra (if not NULL).
487 * Place the fused basic map in the position that is the smallest of i and j.
489 * If "detect_equalities" is set, then look for equalities encoded
490 * as pairs of inequalities.
491 * If "check_number" is set, then the original basic maps are only
492 * replaced if the total number of constraints does not increase.
493 * While the number of integer divisions in the two basic maps
494 * is assumed to be the same, the actual definitions may be different.
495 * We only copy the definition from one of the basic map if it is
496 * the same as that of the other basic map. Otherwise, we mark
497 * the integer division as unknown and simplify the basic map
498 * in an attempt to recover the integer division definition.
500 static enum isl_change fuse(int i, int j, struct isl_coalesce_info *info,
501 __isl_keep isl_mat *extra, int detect_equalities, int check_number)
503 int k, l;
504 struct isl_basic_map *fused = NULL;
505 struct isl_tab *fused_tab = NULL;
506 unsigned total = isl_basic_map_total_dim(info[i].bmap);
507 unsigned extra_rows = extra ? extra->n_row : 0;
508 unsigned n_eq, n_ineq;
509 int simplify = 0;
511 if (j < i)
512 return fuse(j, i, info, extra, detect_equalities, check_number);
514 n_eq = info[i].bmap->n_eq + info[j].bmap->n_eq;
515 n_ineq = info[i].bmap->n_ineq + info[j].bmap->n_ineq;
516 fused = isl_basic_map_alloc_space(isl_space_copy(info[i].bmap->dim),
517 info[i].bmap->n_div, n_eq, n_eq + n_ineq + extra_rows);
518 fused = add_valid_constraints(fused, &info[i], 1 + total);
519 fused = add_valid_constraints(fused, &info[j], 1 + total);
520 if (!fused)
521 goto error;
522 if (ISL_F_ISSET(info[i].bmap, ISL_BASIC_MAP_RATIONAL) &&
523 ISL_F_ISSET(info[j].bmap, ISL_BASIC_MAP_RATIONAL))
524 ISL_F_SET(fused, ISL_BASIC_MAP_RATIONAL);
526 for (k = 0; k < info[i].bmap->n_div; ++k) {
527 int l = isl_basic_map_alloc_div(fused);
528 if (l < 0)
529 goto error;
530 if (isl_seq_eq(info[i].bmap->div[k], info[j].bmap->div[k],
531 1 + 1 + total)) {
532 isl_seq_cpy(fused->div[l], info[i].bmap->div[k],
533 1 + 1 + total);
534 } else {
535 isl_int_set_si(fused->div[l][0], 0);
536 simplify = 1;
540 for (k = 0; k < extra_rows; ++k) {
541 l = isl_basic_map_alloc_inequality(fused);
542 if (l < 0)
543 goto error;
544 isl_seq_cpy(fused->ineq[l], extra->row[k], 1 + total);
547 if (detect_equalities)
548 fused = isl_basic_map_detect_inequality_pairs(fused, NULL);
549 fused = isl_basic_map_gauss(fused, NULL);
550 if (simplify || info[j].simplify) {
551 fused = isl_basic_map_simplify(fused);
552 info[i].simplify = 0;
554 fused = isl_basic_map_finalize(fused);
556 fused_tab = isl_tab_from_basic_map(fused, 0);
557 if (isl_tab_detect_redundant(fused_tab) < 0)
558 goto error;
560 if (check_number &&
561 number_of_constraints_increases(i, j, info, fused, fused_tab)) {
562 isl_tab_free(fused_tab);
563 isl_basic_map_free(fused);
564 return isl_change_none;
567 clear(&info[i]);
568 info[i].bmap = fused;
569 info[i].tab = fused_tab;
570 drop(&info[j]);
572 return isl_change_fuse;
573 error:
574 isl_tab_free(fused_tab);
575 isl_basic_map_free(fused);
576 return isl_change_error;
579 /* Given a pair of basic maps i and j such that all constraints are either
580 * "valid" or "cut", check if the facets corresponding to the "cut"
581 * constraints of i lie entirely within basic map j.
582 * If so, replace the pair by the basic map consisting of the valid
583 * constraints in both basic maps.
584 * Checking whether the facet lies entirely within basic map j
585 * is performed by checking whether the constraints of basic map j
586 * are valid for the facet. These tests are performed on a rational
587 * tableau to avoid the theoretical possibility that a constraint
588 * that was considered to be a cut constraint for the entire basic map i
589 * happens to be considered to be a valid constraint for the facet,
590 * even though it cuts off the same rational points.
592 * To see that we are not introducing any extra points, call the
593 * two basic maps A and B and the resulting map U and let x
594 * be an element of U \setminus ( A \cup B ).
595 * A line connecting x with an element of A \cup B meets a facet F
596 * of either A or B. Assume it is a facet of B and let c_1 be
597 * the corresponding facet constraint. We have c_1(x) < 0 and
598 * so c_1 is a cut constraint. This implies that there is some
599 * (possibly rational) point x' satisfying the constraints of A
600 * and the opposite of c_1 as otherwise c_1 would have been marked
601 * valid for A. The line connecting x and x' meets a facet of A
602 * in a (possibly rational) point that also violates c_1, but this
603 * is impossible since all cut constraints of B are valid for all
604 * cut facets of A.
605 * In case F is a facet of A rather than B, then we can apply the
606 * above reasoning to find a facet of B separating x from A \cup B first.
608 static enum isl_change check_facets(int i, int j,
609 struct isl_coalesce_info *info)
611 int k, l;
612 struct isl_tab_undo *snap, *snap2;
613 unsigned n_eq = info[i].bmap->n_eq;
615 snap = isl_tab_snap(info[i].tab);
616 if (isl_tab_mark_rational(info[i].tab) < 0)
617 return isl_change_error;
618 snap2 = isl_tab_snap(info[i].tab);
620 for (k = 0; k < info[i].bmap->n_ineq; ++k) {
621 if (info[i].ineq[k] != STATUS_CUT)
622 continue;
623 if (isl_tab_select_facet(info[i].tab, n_eq + k) < 0)
624 return isl_change_error;
625 for (l = 0; l < info[j].bmap->n_ineq; ++l) {
626 int stat;
627 if (info[j].ineq[l] != STATUS_CUT)
628 continue;
629 stat = status_in(info[j].bmap->ineq[l], info[i].tab);
630 if (stat < 0)
631 return isl_change_error;
632 if (stat != STATUS_VALID)
633 break;
635 if (isl_tab_rollback(info[i].tab, snap2) < 0)
636 return isl_change_error;
637 if (l < info[j].bmap->n_ineq)
638 break;
641 if (k < info[i].bmap->n_ineq) {
642 if (isl_tab_rollback(info[i].tab, snap) < 0)
643 return isl_change_error;
644 return isl_change_none;
646 return fuse(i, j, info, NULL, 0, 0);
649 /* Check if info->bmap contains the basic map represented
650 * by the tableau "tab".
651 * For each equality, we check both the constraint itself
652 * (as an inequality) and its negation. Make sure the
653 * equality is returned to its original state before returning.
655 static isl_bool contains(struct isl_coalesce_info *info, struct isl_tab *tab)
657 int k;
658 unsigned dim;
659 isl_basic_map *bmap = info->bmap;
661 dim = isl_basic_map_total_dim(bmap);
662 for (k = 0; k < bmap->n_eq; ++k) {
663 int stat;
664 isl_seq_neg(bmap->eq[k], bmap->eq[k], 1 + dim);
665 stat = status_in(bmap->eq[k], tab);
666 isl_seq_neg(bmap->eq[k], bmap->eq[k], 1 + dim);
667 if (stat < 0)
668 return isl_bool_error;
669 if (stat != STATUS_VALID)
670 return isl_bool_false;
671 stat = status_in(bmap->eq[k], tab);
672 if (stat < 0)
673 return isl_bool_error;
674 if (stat != STATUS_VALID)
675 return isl_bool_false;
678 for (k = 0; k < bmap->n_ineq; ++k) {
679 int stat;
680 if (info->ineq[k] == STATUS_REDUNDANT)
681 continue;
682 stat = status_in(bmap->ineq[k], tab);
683 if (stat < 0)
684 return isl_bool_error;
685 if (stat != STATUS_VALID)
686 return isl_bool_false;
688 return isl_bool_true;
691 /* Basic map "i" has an inequality (say "k") that is adjacent
692 * to some inequality of basic map "j". All the other inequalities
693 * are valid for "j".
694 * Check if basic map "j" forms an extension of basic map "i".
696 * Note that this function is only called if some of the equalities or
697 * inequalities of basic map "j" do cut basic map "i". The function is
698 * correct even if there are no such cut constraints, but in that case
699 * the additional checks performed by this function are overkill.
701 * In particular, we replace constraint k, say f >= 0, by constraint
702 * f <= -1, add the inequalities of "j" that are valid for "i"
703 * and check if the result is a subset of basic map "j".
704 * To improve the chances of the subset relation being detected,
705 * any variable that only attains a single integer value
706 * in the tableau of "i" is first fixed to that value.
707 * If the result is a subset, then we know that this result is exactly equal
708 * to basic map "j" since all its constraints are valid for basic map "j".
709 * By combining the valid constraints of "i" (all equalities and all
710 * inequalities except "k") and the valid constraints of "j" we therefore
711 * obtain a basic map that is equal to their union.
712 * In this case, there is no need to perform a rollback of the tableau
713 * since it is going to be destroyed in fuse().
716 * |\__ |\__
717 * | \__ | \__
718 * | \_ => | \__
719 * |_______| _ |_________\
722 * |\ |\
723 * | \ | \
724 * | \ | \
725 * | | | \
726 * | ||\ => | \
727 * | || \ | \
728 * | || | | |
729 * |__||_/ |_____/
731 static enum isl_change is_adj_ineq_extension(int i, int j,
732 struct isl_coalesce_info *info)
734 int k;
735 struct isl_tab_undo *snap;
736 unsigned n_eq = info[i].bmap->n_eq;
737 unsigned total = isl_basic_map_total_dim(info[i].bmap);
738 isl_stat r;
739 isl_bool super;
741 if (isl_tab_extend_cons(info[i].tab, 1 + info[j].bmap->n_ineq) < 0)
742 return isl_change_error;
744 k = find_ineq(&info[i], STATUS_ADJ_INEQ);
745 if (k < 0)
746 isl_die(isl_basic_map_get_ctx(info[i].bmap), isl_error_internal,
747 "info[i].ineq should have exactly one STATUS_ADJ_INEQ",
748 return isl_change_error);
750 snap = isl_tab_snap(info[i].tab);
752 if (isl_tab_unrestrict(info[i].tab, n_eq + k) < 0)
753 return isl_change_error;
755 isl_seq_neg(info[i].bmap->ineq[k], info[i].bmap->ineq[k], 1 + total);
756 isl_int_sub_ui(info[i].bmap->ineq[k][0], info[i].bmap->ineq[k][0], 1);
757 r = isl_tab_add_ineq(info[i].tab, info[i].bmap->ineq[k]);
758 isl_seq_neg(info[i].bmap->ineq[k], info[i].bmap->ineq[k], 1 + total);
759 isl_int_sub_ui(info[i].bmap->ineq[k][0], info[i].bmap->ineq[k][0], 1);
760 if (r < 0)
761 return isl_change_error;
763 for (k = 0; k < info[j].bmap->n_ineq; ++k) {
764 if (info[j].ineq[k] != STATUS_VALID)
765 continue;
766 if (isl_tab_add_ineq(info[i].tab, info[j].bmap->ineq[k]) < 0)
767 return isl_change_error;
769 if (isl_tab_detect_constants(info[i].tab) < 0)
770 return isl_change_error;
772 super = contains(&info[j], info[i].tab);
773 if (super < 0)
774 return isl_change_error;
775 if (super)
776 return fuse(i, j, info, NULL, 0, 0);
778 if (isl_tab_rollback(info[i].tab, snap) < 0)
779 return isl_change_error;
781 return isl_change_none;
785 /* Both basic maps have at least one inequality with and adjacent
786 * (but opposite) inequality in the other basic map.
787 * Check that there are no cut constraints and that there is only
788 * a single pair of adjacent inequalities.
789 * If so, we can replace the pair by a single basic map described
790 * by all but the pair of adjacent inequalities.
791 * Any additional points introduced lie strictly between the two
792 * adjacent hyperplanes and can therefore be integral.
794 * ____ _____
795 * / ||\ / \
796 * / || \ / \
797 * \ || \ => \ \
798 * \ || / \ /
799 * \___||_/ \_____/
801 * The test for a single pair of adjancent inequalities is important
802 * for avoiding the combination of two basic maps like the following
804 * /|
805 * / |
806 * /__|
807 * _____
808 * | |
809 * | |
810 * |___|
812 * If there are some cut constraints on one side, then we may
813 * still be able to fuse the two basic maps, but we need to perform
814 * some additional checks in is_adj_ineq_extension.
816 static enum isl_change check_adj_ineq(int i, int j,
817 struct isl_coalesce_info *info)
819 int count_i, count_j;
820 int cut_i, cut_j;
822 count_i = count_ineq(&info[i], STATUS_ADJ_INEQ);
823 count_j = count_ineq(&info[j], STATUS_ADJ_INEQ);
825 if (count_i != 1 && count_j != 1)
826 return isl_change_none;
828 cut_i = any_eq(&info[i], STATUS_CUT) || any_ineq(&info[i], STATUS_CUT);
829 cut_j = any_eq(&info[j], STATUS_CUT) || any_ineq(&info[j], STATUS_CUT);
831 if (!cut_i && !cut_j && count_i == 1 && count_j == 1)
832 return fuse(i, j, info, NULL, 0, 0);
834 if (count_i == 1 && !cut_i)
835 return is_adj_ineq_extension(i, j, info);
837 if (count_j == 1 && !cut_j)
838 return is_adj_ineq_extension(j, i, info);
840 return isl_change_none;
843 /* Given an affine transformation matrix "T", does row "row" represent
844 * anything other than a unit vector (possibly shifted by a constant)
845 * that is not involved in any of the other rows?
847 * That is, if a constraint involves the variable corresponding to
848 * the row, then could its preimage by "T" have any coefficients
849 * that are different from those in the original constraint?
851 static int not_unique_unit_row(__isl_keep isl_mat *T, int row)
853 int i, j;
854 int len = T->n_col - 1;
856 i = isl_seq_first_non_zero(T->row[row] + 1, len);
857 if (i < 0)
858 return 1;
859 if (!isl_int_is_one(T->row[row][1 + i]) &&
860 !isl_int_is_negone(T->row[row][1 + i]))
861 return 1;
863 j = isl_seq_first_non_zero(T->row[row] + 1 + i + 1, len - (i + 1));
864 if (j >= 0)
865 return 1;
867 for (j = 1; j < T->n_row; ++j) {
868 if (j == row)
869 continue;
870 if (!isl_int_is_zero(T->row[j][1 + i]))
871 return 1;
874 return 0;
877 /* Does inequality constraint "ineq" of "bmap" involve any of
878 * the variables marked in "affected"?
879 * "total" is the total number of variables, i.e., the number
880 * of entries in "affected".
882 static isl_bool is_affected(__isl_keep isl_basic_map *bmap, int ineq,
883 int *affected, int total)
885 int i;
887 for (i = 0; i < total; ++i) {
888 if (!affected[i])
889 continue;
890 if (!isl_int_is_zero(bmap->ineq[ineq][1 + i]))
891 return isl_bool_true;
894 return isl_bool_false;
897 /* Given the compressed version of inequality constraint "ineq"
898 * of info->bmap in "v", check if the constraint can be tightened,
899 * where the compression is based on an equality constraint valid
900 * for info->tab.
901 * If so, add the tightened version of the inequality constraint
902 * to info->tab. "v" may be modified by this function.
904 * That is, if the compressed constraint is of the form
906 * m f() + c >= 0
908 * with 0 < c < m, then it is equivalent to
910 * f() >= 0
912 * This means that c can also be subtracted from the original,
913 * uncompressed constraint without affecting the integer points
914 * in info->tab. Add this tightened constraint as an extra row
915 * to info->tab to make this information explicitly available.
917 static __isl_give isl_vec *try_tightening(struct isl_coalesce_info *info,
918 int ineq, __isl_take isl_vec *v)
920 isl_ctx *ctx;
921 isl_stat r;
923 if (!v)
924 return NULL;
926 ctx = isl_vec_get_ctx(v);
927 isl_seq_gcd(v->el + 1, v->size - 1, &ctx->normalize_gcd);
928 if (isl_int_is_zero(ctx->normalize_gcd) ||
929 isl_int_is_one(ctx->normalize_gcd)) {
930 return v;
933 v = isl_vec_cow(v);
934 if (!v)
935 return NULL;
937 isl_int_fdiv_r(v->el[0], v->el[0], ctx->normalize_gcd);
938 if (isl_int_is_zero(v->el[0]))
939 return v;
941 if (isl_tab_extend_cons(info->tab, 1) < 0)
942 return isl_vec_free(v);
944 isl_int_sub(info->bmap->ineq[ineq][0],
945 info->bmap->ineq[ineq][0], v->el[0]);
946 r = isl_tab_add_ineq(info->tab, info->bmap->ineq[ineq]);
947 isl_int_add(info->bmap->ineq[ineq][0],
948 info->bmap->ineq[ineq][0], v->el[0]);
950 if (r < 0)
951 return isl_vec_free(v);
953 return v;
956 /* Tighten the (non-redundant) constraints on the facet represented
957 * by info->tab.
958 * In particular, on input, info->tab represents the result
959 * of relaxing the "n" inequality constraints of info->bmap in "relaxed"
960 * by one, i.e., replacing f_i >= 0 by f_i + 1 >= 0, and then
961 * replacing the one at index "l" by the corresponding equality,
962 * i.e., f_k + 1 = 0, with k = relaxed[l].
964 * Compute a variable compression from the equality constraint f_k + 1 = 0
965 * and use it to tighten the other constraints of info->bmap
966 * (that is, all constraints that have not been relaxed),
967 * updating info->tab (and leaving info->bmap untouched).
968 * The compression handles essentially two cases, one where a variable
969 * is assigned a fixed value and can therefore be eliminated, and one
970 * where one variable is a shifted multiple of some other variable and
971 * can therefore be replaced by that multiple.
972 * Gaussian elimination would also work for the first case, but for
973 * the second case, the effectiveness would depend on the order
974 * of the variables.
975 * After compression, some of the constraints may have coefficients
976 * with a common divisor. If this divisor does not divide the constant
977 * term, then the constraint can be tightened.
978 * The tightening is performed on the tableau info->tab by introducing
979 * extra (temporary) constraints.
981 * Only constraints that are possibly affected by the compression are
982 * considered. In particular, if the constraint only involves variables
983 * that are directly mapped to a distinct set of other variables, then
984 * no common divisor can be introduced and no tightening can occur.
986 * It is important to only consider the non-redundant constraints
987 * since the facet constraint has been relaxed prior to the call
988 * to this function, meaning that the constraints that were redundant
989 * prior to the relaxation may no longer be redundant.
990 * These constraints will be ignored in the fused result, so
991 * the fusion detection should not exploit them.
993 static isl_stat tighten_on_relaxed_facet(struct isl_coalesce_info *info,
994 int n, int *relaxed, int l)
996 unsigned total;
997 isl_ctx *ctx;
998 isl_vec *v = NULL;
999 isl_mat *T;
1000 int i;
1001 int k;
1002 int *affected;
1004 k = relaxed[l];
1005 ctx = isl_basic_map_get_ctx(info->bmap);
1006 total = isl_basic_map_total_dim(info->bmap);
1007 isl_int_add_ui(info->bmap->ineq[k][0], info->bmap->ineq[k][0], 1);
1008 T = isl_mat_sub_alloc6(ctx, info->bmap->ineq, k, 1, 0, 1 + total);
1009 T = isl_mat_variable_compression(T, NULL);
1010 isl_int_sub_ui(info->bmap->ineq[k][0], info->bmap->ineq[k][0], 1);
1011 if (!T)
1012 return isl_stat_error;
1013 if (T->n_col == 0) {
1014 isl_mat_free(T);
1015 return isl_stat_ok;
1018 affected = isl_alloc_array(ctx, int, total);
1019 if (!affected)
1020 goto error;
1022 for (i = 0; i < total; ++i)
1023 affected[i] = not_unique_unit_row(T, 1 + i);
1025 for (i = 0; i < info->bmap->n_ineq; ++i) {
1026 isl_bool handle;
1027 if (any(relaxed, n, i))
1028 continue;
1029 if (info->ineq[i] == STATUS_REDUNDANT)
1030 continue;
1031 handle = is_affected(info->bmap, i, affected, total);
1032 if (handle < 0)
1033 goto error;
1034 if (!handle)
1035 continue;
1036 v = isl_vec_alloc(ctx, 1 + total);
1037 if (!v)
1038 goto error;
1039 isl_seq_cpy(v->el, info->bmap->ineq[i], 1 + total);
1040 v = isl_vec_mat_product(v, isl_mat_copy(T));
1041 v = try_tightening(info, i, v);
1042 isl_vec_free(v);
1043 if (!v)
1044 goto error;
1047 isl_mat_free(T);
1048 free(affected);
1049 return isl_stat_ok;
1050 error:
1051 isl_mat_free(T);
1052 free(affected);
1053 return isl_stat_error;
1056 /* Replace the basic maps "i" and "j" by an extension of "i"
1057 * along the "n" inequality constraints in "relax" by one.
1058 * The tableau info[i].tab has already been extended.
1059 * Extend info[i].bmap accordingly by relaxing all constraints in "relax"
1060 * by one.
1061 * Each integer division that does not have exactly the same
1062 * definition in "i" and "j" is marked unknown and the basic map
1063 * is scheduled to be simplified in an attempt to recover
1064 * the integer division definition.
1065 * Place the extension in the position that is the smallest of i and j.
1067 static enum isl_change extend(int i, int j, int n, int *relax,
1068 struct isl_coalesce_info *info)
1070 int l;
1071 unsigned total;
1073 info[i].bmap = isl_basic_map_cow(info[i].bmap);
1074 if (!info[i].bmap)
1075 return isl_change_error;
1076 total = isl_basic_map_total_dim(info[i].bmap);
1077 for (l = 0; l < info[i].bmap->n_div; ++l)
1078 if (!isl_seq_eq(info[i].bmap->div[l],
1079 info[j].bmap->div[l], 1 + 1 + total)) {
1080 isl_int_set_si(info[i].bmap->div[l][0], 0);
1081 info[i].simplify = 1;
1083 for (l = 0; l < n; ++l)
1084 isl_int_add_ui(info[i].bmap->ineq[relax[l]][0],
1085 info[i].bmap->ineq[relax[l]][0], 1);
1086 ISL_F_SET(info[i].bmap, ISL_BASIC_MAP_FINAL);
1087 drop(&info[j]);
1088 if (j < i)
1089 exchange(&info[i], &info[j]);
1090 return isl_change_fuse;
1093 /* Basic map "i" has "n" inequality constraints (collected in "relax")
1094 * that are such that they include basic map "j" if they are relaxed
1095 * by one. All the other inequalities are valid for "j".
1096 * Check if basic map "j" forms an extension of basic map "i".
1098 * In particular, relax the constraints in "relax", compute the corresponding
1099 * facets one by one and check whether each of these is included
1100 * in the other basic map.
1101 * Before testing for inclusion, the constraints on each facet
1102 * are tightened to increase the chance of an inclusion being detected.
1103 * (Adding the valid constraints of "j" to the tableau of "i", as is done
1104 * in is_adj_ineq_extension, may further increase those chances, but this
1105 * is not currently done.)
1106 * If each facet is included, we know that relaxing the constraints extends
1107 * the basic map with exactly the other basic map (we already know that this
1108 * other basic map is included in the extension, because all other
1109 * inequality constraints are valid of "j") and we can replace the
1110 * two basic maps by this extension.
1112 * If any of the relaxed constraints turn out to be redundant, then bail out.
1113 * isl_tab_select_facet refuses to handle such constraints. It may be
1114 * possible to handle them anyway by making a distinction between
1115 * redundant constraints with a corresponding facet that still intersects
1116 * the set (allowing isl_tab_select_facet to handle them) and
1117 * those where the facet does not intersect the set (which can be ignored
1118 * because the empty facet is trivially included in the other disjunct).
1119 * However, relaxed constraints that turn out to be redundant should
1120 * be fairly rare and no such instance has been reported where
1121 * coalescing would be successful.
1122 * ____ _____
1123 * / || / |
1124 * / || / |
1125 * \ || => \ |
1126 * \ || \ |
1127 * \___|| \____|
1130 * \ |\
1131 * |\\ | \
1132 * | \\ | \
1133 * | | => | /
1134 * | / | /
1135 * |/ |/
1137 static enum isl_change is_relaxed_extension(int i, int j, int n, int *relax,
1138 struct isl_coalesce_info *info)
1140 int l;
1141 isl_bool super;
1142 struct isl_tab_undo *snap, *snap2;
1143 unsigned n_eq = info[i].bmap->n_eq;
1145 for (l = 0; l < n; ++l)
1146 if (isl_tab_is_equality(info[i].tab, n_eq + relax[l]))
1147 return isl_change_none;
1149 snap = isl_tab_snap(info[i].tab);
1150 for (l = 0; l < n; ++l)
1151 if (isl_tab_relax(info[i].tab, n_eq + relax[l]) < 0)
1152 return isl_change_error;
1153 for (l = 0; l < n; ++l) {
1154 if (!isl_tab_is_redundant(info[i].tab, n_eq + relax[l]))
1155 continue;
1156 if (isl_tab_rollback(info[i].tab, snap) < 0)
1157 return isl_change_error;
1158 return isl_change_none;
1160 snap2 = isl_tab_snap(info[i].tab);
1161 for (l = 0; l < n; ++l) {
1162 if (isl_tab_rollback(info[i].tab, snap2) < 0)
1163 return isl_change_error;
1164 if (isl_tab_select_facet(info[i].tab, n_eq + relax[l]) < 0)
1165 return isl_change_error;
1166 if (tighten_on_relaxed_facet(&info[i], n, relax, l) < 0)
1167 return isl_change_error;
1168 super = contains(&info[j], info[i].tab);
1169 if (super < 0)
1170 return isl_change_error;
1171 if (super)
1172 continue;
1173 if (isl_tab_rollback(info[i].tab, snap) < 0)
1174 return isl_change_error;
1175 return isl_change_none;
1178 if (isl_tab_rollback(info[i].tab, snap2) < 0)
1179 return isl_change_error;
1180 return extend(i, j, n, relax, info);
1183 /* Data structure that keeps track of the wrapping constraints
1184 * and of information to bound the coefficients of those constraints.
1186 * bound is set if we want to apply a bound on the coefficients
1187 * mat contains the wrapping constraints
1188 * max is the bound on the coefficients (if bound is set)
1190 struct isl_wraps {
1191 int bound;
1192 isl_mat *mat;
1193 isl_int max;
1196 /* Update wraps->max to be greater than or equal to the coefficients
1197 * in the equalities and inequalities of info->bmap that can be removed
1198 * if we end up applying wrapping.
1200 static isl_stat wraps_update_max(struct isl_wraps *wraps,
1201 struct isl_coalesce_info *info)
1203 int k;
1204 isl_int max_k;
1205 unsigned total = isl_basic_map_total_dim(info->bmap);
1207 isl_int_init(max_k);
1209 for (k = 0; k < info->bmap->n_eq; ++k) {
1210 if (info->eq[2 * k] == STATUS_VALID &&
1211 info->eq[2 * k + 1] == STATUS_VALID)
1212 continue;
1213 isl_seq_abs_max(info->bmap->eq[k] + 1, total, &max_k);
1214 if (isl_int_abs_gt(max_k, wraps->max))
1215 isl_int_set(wraps->max, max_k);
1218 for (k = 0; k < info->bmap->n_ineq; ++k) {
1219 if (info->ineq[k] == STATUS_VALID ||
1220 info->ineq[k] == STATUS_REDUNDANT)
1221 continue;
1222 isl_seq_abs_max(info->bmap->ineq[k] + 1, total, &max_k);
1223 if (isl_int_abs_gt(max_k, wraps->max))
1224 isl_int_set(wraps->max, max_k);
1227 isl_int_clear(max_k);
1229 return isl_stat_ok;
1232 /* Initialize the isl_wraps data structure.
1233 * If we want to bound the coefficients of the wrapping constraints,
1234 * we set wraps->max to the largest coefficient
1235 * in the equalities and inequalities that can be removed if we end up
1236 * applying wrapping.
1238 static isl_stat wraps_init(struct isl_wraps *wraps, __isl_take isl_mat *mat,
1239 struct isl_coalesce_info *info, int i, int j)
1241 isl_ctx *ctx;
1243 wraps->bound = 0;
1244 wraps->mat = mat;
1245 if (!mat)
1246 return isl_stat_error;
1247 ctx = isl_mat_get_ctx(mat);
1248 wraps->bound = isl_options_get_coalesce_bounded_wrapping(ctx);
1249 if (!wraps->bound)
1250 return isl_stat_ok;
1251 isl_int_init(wraps->max);
1252 isl_int_set_si(wraps->max, 0);
1253 if (wraps_update_max(wraps, &info[i]) < 0)
1254 return isl_stat_error;
1255 if (wraps_update_max(wraps, &info[j]) < 0)
1256 return isl_stat_error;
1258 return isl_stat_ok;
1261 /* Free the contents of the isl_wraps data structure.
1263 static void wraps_free(struct isl_wraps *wraps)
1265 isl_mat_free(wraps->mat);
1266 if (wraps->bound)
1267 isl_int_clear(wraps->max);
1270 /* Is the wrapping constraint in row "row" allowed?
1272 * If wraps->bound is set, we check that none of the coefficients
1273 * is greater than wraps->max.
1275 static int allow_wrap(struct isl_wraps *wraps, int row)
1277 int i;
1279 if (!wraps->bound)
1280 return 1;
1282 for (i = 1; i < wraps->mat->n_col; ++i)
1283 if (isl_int_abs_gt(wraps->mat->row[row][i], wraps->max))
1284 return 0;
1286 return 1;
1289 /* Wrap "ineq" (or its opposite if "negate" is set) around "bound"
1290 * to include "set" and add the result in position "w" of "wraps".
1291 * "len" is the total number of coefficients in "bound" and "ineq".
1292 * Return 1 on success, 0 on failure and -1 on error.
1293 * Wrapping can fail if the result of wrapping is equal to "bound"
1294 * or if we want to bound the sizes of the coefficients and
1295 * the wrapped constraint does not satisfy this bound.
1297 static int add_wrap(struct isl_wraps *wraps, int w, isl_int *bound,
1298 isl_int *ineq, unsigned len, __isl_keep isl_set *set, int negate)
1300 isl_seq_cpy(wraps->mat->row[w], bound, len);
1301 if (negate) {
1302 isl_seq_neg(wraps->mat->row[w + 1], ineq, len);
1303 ineq = wraps->mat->row[w + 1];
1305 if (!isl_set_wrap_facet(set, wraps->mat->row[w], ineq))
1306 return -1;
1307 if (isl_seq_eq(wraps->mat->row[w], bound, len))
1308 return 0;
1309 if (!allow_wrap(wraps, w))
1310 return 0;
1311 return 1;
1314 /* For each constraint in info->bmap that is not redundant (as determined
1315 * by info->tab) and that is not a valid constraint for the other basic map,
1316 * wrap the constraint around "bound" such that it includes the whole
1317 * set "set" and append the resulting constraint to "wraps".
1318 * Note that the constraints that are valid for the other basic map
1319 * will be added to the combined basic map by default, so there is
1320 * no need to wrap them.
1321 * The caller wrap_in_facets even relies on this function not wrapping
1322 * any constraints that are already valid.
1323 * "wraps" is assumed to have been pre-allocated to the appropriate size.
1324 * wraps->n_row is the number of actual wrapped constraints that have
1325 * been added.
1326 * If any of the wrapping problems results in a constraint that is
1327 * identical to "bound", then this means that "set" is unbounded in such
1328 * way that no wrapping is possible. If this happens then wraps->n_row
1329 * is reset to zero.
1330 * Similarly, if we want to bound the coefficients of the wrapping
1331 * constraints and a newly added wrapping constraint does not
1332 * satisfy the bound, then wraps->n_row is also reset to zero.
1334 static isl_stat add_wraps(struct isl_wraps *wraps,
1335 struct isl_coalesce_info *info, isl_int *bound, __isl_keep isl_set *set)
1337 int l, m;
1338 int w;
1339 int added;
1340 isl_basic_map *bmap = info->bmap;
1341 unsigned len = 1 + isl_basic_map_total_dim(bmap);
1343 w = wraps->mat->n_row;
1345 for (l = 0; l < bmap->n_ineq; ++l) {
1346 if (info->ineq[l] == STATUS_VALID ||
1347 info->ineq[l] == STATUS_REDUNDANT)
1348 continue;
1349 if (isl_seq_is_neg(bound, bmap->ineq[l], len))
1350 continue;
1351 if (isl_seq_eq(bound, bmap->ineq[l], len))
1352 continue;
1353 if (isl_tab_is_redundant(info->tab, bmap->n_eq + l))
1354 continue;
1356 added = add_wrap(wraps, w, bound, bmap->ineq[l], len, set, 0);
1357 if (added < 0)
1358 return isl_stat_error;
1359 if (!added)
1360 goto unbounded;
1361 ++w;
1363 for (l = 0; l < bmap->n_eq; ++l) {
1364 if (isl_seq_is_neg(bound, bmap->eq[l], len))
1365 continue;
1366 if (isl_seq_eq(bound, bmap->eq[l], len))
1367 continue;
1369 for (m = 0; m < 2; ++m) {
1370 if (info->eq[2 * l + m] == STATUS_VALID)
1371 continue;
1372 added = add_wrap(wraps, w, bound, bmap->eq[l], len,
1373 set, !m);
1374 if (added < 0)
1375 return isl_stat_error;
1376 if (!added)
1377 goto unbounded;
1378 ++w;
1382 wraps->mat->n_row = w;
1383 return isl_stat_ok;
1384 unbounded:
1385 wraps->mat->n_row = 0;
1386 return isl_stat_ok;
1389 /* Check if the constraints in "wraps" from "first" until the last
1390 * are all valid for the basic set represented by "tab".
1391 * If not, wraps->n_row is set to zero.
1393 static int check_wraps(__isl_keep isl_mat *wraps, int first,
1394 struct isl_tab *tab)
1396 int i;
1398 for (i = first; i < wraps->n_row; ++i) {
1399 enum isl_ineq_type type;
1400 type = isl_tab_ineq_type(tab, wraps->row[i]);
1401 if (type == isl_ineq_error)
1402 return -1;
1403 if (type == isl_ineq_redundant)
1404 continue;
1405 wraps->n_row = 0;
1406 return 0;
1409 return 0;
1412 /* Return a set that corresponds to the non-redundant constraints
1413 * (as recorded in tab) of bmap.
1415 * It's important to remove the redundant constraints as some
1416 * of the other constraints may have been modified after the
1417 * constraints were marked redundant.
1418 * In particular, a constraint may have been relaxed.
1419 * Redundant constraints are ignored when a constraint is relaxed
1420 * and should therefore continue to be ignored ever after.
1421 * Otherwise, the relaxation might be thwarted by some of
1422 * these constraints.
1424 * Update the underlying set to ensure that the dimension doesn't change.
1425 * Otherwise the integer divisions could get dropped if the tab
1426 * turns out to be empty.
1428 static __isl_give isl_set *set_from_updated_bmap(__isl_keep isl_basic_map *bmap,
1429 struct isl_tab *tab)
1431 isl_basic_set *bset;
1433 bmap = isl_basic_map_copy(bmap);
1434 bset = isl_basic_map_underlying_set(bmap);
1435 bset = isl_basic_set_cow(bset);
1436 bset = isl_basic_set_update_from_tab(bset, tab);
1437 return isl_set_from_basic_set(bset);
1440 /* Wrap the constraints of info->bmap that bound the facet defined
1441 * by inequality "k" around (the opposite of) this inequality to
1442 * include "set". "bound" may be used to store the negated inequality.
1443 * Since the wrapped constraints are not guaranteed to contain the whole
1444 * of info->bmap, we check them in check_wraps.
1445 * If any of the wrapped constraints turn out to be invalid, then
1446 * check_wraps will reset wrap->n_row to zero.
1448 static isl_stat add_wraps_around_facet(struct isl_wraps *wraps,
1449 struct isl_coalesce_info *info, int k, isl_int *bound,
1450 __isl_keep isl_set *set)
1452 struct isl_tab_undo *snap;
1453 int n;
1454 unsigned total = isl_basic_map_total_dim(info->bmap);
1456 snap = isl_tab_snap(info->tab);
1458 if (isl_tab_select_facet(info->tab, info->bmap->n_eq + k) < 0)
1459 return isl_stat_error;
1460 if (isl_tab_detect_redundant(info->tab) < 0)
1461 return isl_stat_error;
1463 isl_seq_neg(bound, info->bmap->ineq[k], 1 + total);
1465 n = wraps->mat->n_row;
1466 if (add_wraps(wraps, info, bound, set) < 0)
1467 return isl_stat_error;
1469 if (isl_tab_rollback(info->tab, snap) < 0)
1470 return isl_stat_error;
1471 if (check_wraps(wraps->mat, n, info->tab) < 0)
1472 return isl_stat_error;
1474 return isl_stat_ok;
1477 /* Given a basic set i with a constraint k that is adjacent to
1478 * basic set j, check if we can wrap
1479 * both the facet corresponding to k (if "wrap_facet" is set) and basic map j
1480 * (always) around their ridges to include the other set.
1481 * If so, replace the pair of basic sets by their union.
1483 * All constraints of i (except k) are assumed to be valid or
1484 * cut constraints for j.
1485 * Wrapping the cut constraints to include basic map j may result
1486 * in constraints that are no longer valid of basic map i
1487 * we have to check that the resulting wrapping constraints are valid for i.
1488 * If "wrap_facet" is not set, then all constraints of i (except k)
1489 * are assumed to be valid for j.
1490 * ____ _____
1491 * / | / \
1492 * / || / |
1493 * \ || => \ |
1494 * \ || \ |
1495 * \___|| \____|
1498 static enum isl_change can_wrap_in_facet(int i, int j, int k,
1499 struct isl_coalesce_info *info, int wrap_facet)
1501 enum isl_change change = isl_change_none;
1502 struct isl_wraps wraps;
1503 isl_ctx *ctx;
1504 isl_mat *mat;
1505 struct isl_set *set_i = NULL;
1506 struct isl_set *set_j = NULL;
1507 struct isl_vec *bound = NULL;
1508 unsigned total = isl_basic_map_total_dim(info[i].bmap);
1510 set_i = set_from_updated_bmap(info[i].bmap, info[i].tab);
1511 set_j = set_from_updated_bmap(info[j].bmap, info[j].tab);
1512 ctx = isl_basic_map_get_ctx(info[i].bmap);
1513 mat = isl_mat_alloc(ctx, 2 * (info[i].bmap->n_eq + info[j].bmap->n_eq) +
1514 info[i].bmap->n_ineq + info[j].bmap->n_ineq,
1515 1 + total);
1516 if (wraps_init(&wraps, mat, info, i, j) < 0)
1517 goto error;
1518 bound = isl_vec_alloc(ctx, 1 + total);
1519 if (!set_i || !set_j || !bound)
1520 goto error;
1522 isl_seq_cpy(bound->el, info[i].bmap->ineq[k], 1 + total);
1523 isl_int_add_ui(bound->el[0], bound->el[0], 1);
1524 isl_seq_normalize(ctx, bound->el, 1 + total);
1526 isl_seq_cpy(wraps.mat->row[0], bound->el, 1 + total);
1527 wraps.mat->n_row = 1;
1529 if (add_wraps(&wraps, &info[j], bound->el, set_i) < 0)
1530 goto error;
1531 if (!wraps.mat->n_row)
1532 goto unbounded;
1534 if (wrap_facet) {
1535 if (add_wraps_around_facet(&wraps, &info[i], k,
1536 bound->el, set_j) < 0)
1537 goto error;
1538 if (!wraps.mat->n_row)
1539 goto unbounded;
1542 change = fuse(i, j, info, wraps.mat, 0, 0);
1544 unbounded:
1545 wraps_free(&wraps);
1547 isl_set_free(set_i);
1548 isl_set_free(set_j);
1550 isl_vec_free(bound);
1552 return change;
1553 error:
1554 wraps_free(&wraps);
1555 isl_vec_free(bound);
1556 isl_set_free(set_i);
1557 isl_set_free(set_j);
1558 return isl_change_error;
1561 /* Given a cut constraint t(x) >= 0 of basic map i, stored in row "w"
1562 * of wrap.mat, replace it by its relaxed version t(x) + 1 >= 0, and
1563 * add wrapping constraints to wrap.mat for all constraints
1564 * of basic map j that bound the part of basic map j that sticks out
1565 * of the cut constraint.
1566 * "set_i" is the underlying set of basic map i.
1567 * If any wrapping fails, then wraps->mat.n_row is reset to zero.
1569 * In particular, we first intersect basic map j with t(x) + 1 = 0.
1570 * If the result is empty, then t(x) >= 0 was actually a valid constraint
1571 * (with respect to the integer points), so we add t(x) >= 0 instead.
1572 * Otherwise, we wrap the constraints of basic map j that are not
1573 * redundant in this intersection and that are not already valid
1574 * for basic map i over basic map i.
1575 * Note that it is sufficient to wrap the constraints to include
1576 * basic map i, because we will only wrap the constraints that do
1577 * not include basic map i already. The wrapped constraint will
1578 * therefore be more relaxed compared to the original constraint.
1579 * Since the original constraint is valid for basic map j, so is
1580 * the wrapped constraint.
1582 static isl_stat wrap_in_facet(struct isl_wraps *wraps, int w,
1583 struct isl_coalesce_info *info_j, __isl_keep isl_set *set_i,
1584 struct isl_tab_undo *snap)
1586 isl_int_add_ui(wraps->mat->row[w][0], wraps->mat->row[w][0], 1);
1587 if (isl_tab_add_eq(info_j->tab, wraps->mat->row[w]) < 0)
1588 return isl_stat_error;
1589 if (isl_tab_detect_redundant(info_j->tab) < 0)
1590 return isl_stat_error;
1592 if (info_j->tab->empty)
1593 isl_int_sub_ui(wraps->mat->row[w][0], wraps->mat->row[w][0], 1);
1594 else if (add_wraps(wraps, info_j, wraps->mat->row[w], set_i) < 0)
1595 return isl_stat_error;
1597 if (isl_tab_rollback(info_j->tab, snap) < 0)
1598 return isl_stat_error;
1600 return isl_stat_ok;
1603 /* Given a pair of basic maps i and j such that j sticks out
1604 * of i at n cut constraints, each time by at most one,
1605 * try to compute wrapping constraints and replace the two
1606 * basic maps by a single basic map.
1607 * The other constraints of i are assumed to be valid for j.
1608 * "set_i" is the underlying set of basic map i.
1609 * "wraps" has been initialized to be of the right size.
1611 * For each cut constraint t(x) >= 0 of i, we add the relaxed version
1612 * t(x) + 1 >= 0, along with wrapping constraints for all constraints
1613 * of basic map j that bound the part of basic map j that sticks out
1614 * of the cut constraint.
1616 * If any wrapping fails, i.e., if we cannot wrap to touch
1617 * the union, then we give up.
1618 * Otherwise, the pair of basic maps is replaced by their union.
1620 static enum isl_change try_wrap_in_facets(int i, int j,
1621 struct isl_coalesce_info *info, struct isl_wraps *wraps,
1622 __isl_keep isl_set *set_i)
1624 int k, l, w;
1625 unsigned total;
1626 struct isl_tab_undo *snap;
1628 total = isl_basic_map_total_dim(info[i].bmap);
1630 snap = isl_tab_snap(info[j].tab);
1632 wraps->mat->n_row = 0;
1634 for (k = 0; k < info[i].bmap->n_eq; ++k) {
1635 for (l = 0; l < 2; ++l) {
1636 if (info[i].eq[2 * k + l] != STATUS_CUT)
1637 continue;
1638 w = wraps->mat->n_row++;
1639 if (l == 0)
1640 isl_seq_neg(wraps->mat->row[w],
1641 info[i].bmap->eq[k], 1 + total);
1642 else
1643 isl_seq_cpy(wraps->mat->row[w],
1644 info[i].bmap->eq[k], 1 + total);
1645 if (wrap_in_facet(wraps, w, &info[j], set_i, snap) < 0)
1646 return isl_change_error;
1648 if (!wraps->mat->n_row)
1649 return isl_change_none;
1653 for (k = 0; k < info[i].bmap->n_ineq; ++k) {
1654 if (info[i].ineq[k] != STATUS_CUT)
1655 continue;
1656 w = wraps->mat->n_row++;
1657 isl_seq_cpy(wraps->mat->row[w],
1658 info[i].bmap->ineq[k], 1 + total);
1659 if (wrap_in_facet(wraps, w, &info[j], set_i, snap) < 0)
1660 return isl_change_error;
1662 if (!wraps->mat->n_row)
1663 return isl_change_none;
1666 return fuse(i, j, info, wraps->mat, 0, 1);
1669 /* Given a pair of basic maps i and j such that j sticks out
1670 * of i at n cut constraints, each time by at most one,
1671 * try to compute wrapping constraints and replace the two
1672 * basic maps by a single basic map.
1673 * The other constraints of i are assumed to be valid for j.
1675 * The core computation is performed by try_wrap_in_facets.
1676 * This function simply extracts an underlying set representation
1677 * of basic map i and initializes the data structure for keeping
1678 * track of wrapping constraints.
1680 static enum isl_change wrap_in_facets(int i, int j, int n,
1681 struct isl_coalesce_info *info)
1683 enum isl_change change = isl_change_none;
1684 struct isl_wraps wraps;
1685 isl_ctx *ctx;
1686 isl_mat *mat;
1687 isl_set *set_i = NULL;
1688 unsigned total = isl_basic_map_total_dim(info[i].bmap);
1689 int max_wrap;
1691 if (isl_tab_extend_cons(info[j].tab, 1) < 0)
1692 return isl_change_error;
1694 max_wrap = 1 + 2 * info[j].bmap->n_eq + info[j].bmap->n_ineq;
1695 max_wrap *= n;
1697 set_i = set_from_updated_bmap(info[i].bmap, info[i].tab);
1698 ctx = isl_basic_map_get_ctx(info[i].bmap);
1699 mat = isl_mat_alloc(ctx, max_wrap, 1 + total);
1700 if (wraps_init(&wraps, mat, info, i, j) < 0)
1701 goto error;
1702 if (!set_i)
1703 goto error;
1705 change = try_wrap_in_facets(i, j, info, &wraps, set_i);
1707 wraps_free(&wraps);
1708 isl_set_free(set_i);
1710 return change;
1711 error:
1712 wraps_free(&wraps);
1713 isl_set_free(set_i);
1714 return isl_change_error;
1717 /* Return the effect of inequality "ineq" on the tableau "tab",
1718 * after relaxing the constant term of "ineq" by one.
1720 static enum isl_ineq_type type_of_relaxed(struct isl_tab *tab, isl_int *ineq)
1722 enum isl_ineq_type type;
1724 isl_int_add_ui(ineq[0], ineq[0], 1);
1725 type = isl_tab_ineq_type(tab, ineq);
1726 isl_int_sub_ui(ineq[0], ineq[0], 1);
1728 return type;
1731 /* Given two basic sets i and j,
1732 * check if relaxing all the cut constraints of i by one turns
1733 * them into valid constraint for j and check if we can wrap in
1734 * the bits that are sticking out.
1735 * If so, replace the pair by their union.
1737 * We first check if all relaxed cut inequalities of i are valid for j
1738 * and then try to wrap in the intersections of the relaxed cut inequalities
1739 * with j.
1741 * During this wrapping, we consider the points of j that lie at a distance
1742 * of exactly 1 from i. In particular, we ignore the points that lie in
1743 * between this lower-dimensional space and the basic map i.
1744 * We can therefore only apply this to integer maps.
1745 * ____ _____
1746 * / ___|_ / \
1747 * / | | / |
1748 * \ | | => \ |
1749 * \|____| \ |
1750 * \___| \____/
1752 * _____ ______
1753 * | ____|_ | \
1754 * | | | | |
1755 * | | | => | |
1756 * |_| | | |
1757 * |_____| \______|
1759 * _______
1760 * | |
1761 * | |\ |
1762 * | | \ |
1763 * | | \ |
1764 * | | \|
1765 * | | \
1766 * | |_____\
1767 * | |
1768 * |_______|
1770 * Wrapping can fail if the result of wrapping one of the facets
1771 * around its edges does not produce any new facet constraint.
1772 * In particular, this happens when we try to wrap in unbounded sets.
1774 * _______________________________________________________________________
1776 * | ___
1777 * | | |
1778 * |_| |_________________________________________________________________
1779 * |___|
1781 * The following is not an acceptable result of coalescing the above two
1782 * sets as it includes extra integer points.
1783 * _______________________________________________________________________
1785 * |
1786 * |
1788 * \______________________________________________________________________
1790 static enum isl_change can_wrap_in_set(int i, int j,
1791 struct isl_coalesce_info *info)
1793 int k, l;
1794 int n;
1795 unsigned total;
1797 if (ISL_F_ISSET(info[i].bmap, ISL_BASIC_MAP_RATIONAL) ||
1798 ISL_F_ISSET(info[j].bmap, ISL_BASIC_MAP_RATIONAL))
1799 return isl_change_none;
1801 n = count_eq(&info[i], STATUS_CUT) + count_ineq(&info[i], STATUS_CUT);
1802 if (n == 0)
1803 return isl_change_none;
1805 total = isl_basic_map_total_dim(info[i].bmap);
1806 for (k = 0; k < info[i].bmap->n_eq; ++k) {
1807 for (l = 0; l < 2; ++l) {
1808 enum isl_ineq_type type;
1810 if (info[i].eq[2 * k + l] != STATUS_CUT)
1811 continue;
1813 if (l == 0)
1814 isl_seq_neg(info[i].bmap->eq[k],
1815 info[i].bmap->eq[k], 1 + total);
1816 type = type_of_relaxed(info[j].tab,
1817 info[i].bmap->eq[k]);
1818 if (l == 0)
1819 isl_seq_neg(info[i].bmap->eq[k],
1820 info[i].bmap->eq[k], 1 + total);
1821 if (type == isl_ineq_error)
1822 return isl_change_error;
1823 if (type != isl_ineq_redundant)
1824 return isl_change_none;
1828 for (k = 0; k < info[i].bmap->n_ineq; ++k) {
1829 enum isl_ineq_type type;
1831 if (info[i].ineq[k] != STATUS_CUT)
1832 continue;
1834 type = type_of_relaxed(info[j].tab, info[i].bmap->ineq[k]);
1835 if (type == isl_ineq_error)
1836 return isl_change_error;
1837 if (type != isl_ineq_redundant)
1838 return isl_change_none;
1841 return wrap_in_facets(i, j, n, info);
1844 /* Check if either i or j has only cut constraints that can
1845 * be used to wrap in (a facet of) the other basic set.
1846 * if so, replace the pair by their union.
1848 static enum isl_change check_wrap(int i, int j, struct isl_coalesce_info *info)
1850 enum isl_change change = isl_change_none;
1852 change = can_wrap_in_set(i, j, info);
1853 if (change != isl_change_none)
1854 return change;
1856 change = can_wrap_in_set(j, i, info);
1857 return change;
1860 /* Check if all inequality constraints of "i" that cut "j" cease
1861 * to be cut constraints if they are relaxed by one.
1862 * If so, collect the cut constraints in "list".
1863 * The caller is responsible for allocating "list".
1865 static isl_bool all_cut_by_one(int i, int j, struct isl_coalesce_info *info,
1866 int *list)
1868 int l, n;
1870 n = 0;
1871 for (l = 0; l < info[i].bmap->n_ineq; ++l) {
1872 enum isl_ineq_type type;
1874 if (info[i].ineq[l] != STATUS_CUT)
1875 continue;
1876 type = type_of_relaxed(info[j].tab, info[i].bmap->ineq[l]);
1877 if (type == isl_ineq_error)
1878 return isl_bool_error;
1879 if (type != isl_ineq_redundant)
1880 return isl_bool_false;
1881 list[n++] = l;
1884 return isl_bool_true;
1887 /* Given two basic maps such that "j" has at least one equality constraint
1888 * that is adjacent to an inequality constraint of "i" and such that "i" has
1889 * exactly one inequality constraint that is adjacent to an equality
1890 * constraint of "j", check whether "i" can be extended to include "j" or
1891 * whether "j" can be wrapped into "i".
1892 * All remaining constraints of "i" and "j" are assumed to be valid
1893 * or cut constraints of the other basic map.
1894 * However, none of the equality constraints of "i" are cut constraints.
1896 * If "i" has any "cut" inequality constraints, then check if relaxing
1897 * each of them by one is sufficient for them to become valid.
1898 * If so, check if the inequality constraint adjacent to an equality
1899 * constraint of "j" along with all these cut constraints
1900 * can be relaxed by one to contain exactly "j".
1901 * Otherwise, or if this fails, check if "j" can be wrapped into "i".
1903 static enum isl_change check_single_adj_eq(int i, int j,
1904 struct isl_coalesce_info *info)
1906 enum isl_change change = isl_change_none;
1907 int k;
1908 int n_cut;
1909 int *relax;
1910 isl_ctx *ctx;
1911 isl_bool try_relax;
1913 n_cut = count_ineq(&info[i], STATUS_CUT);
1915 k = find_ineq(&info[i], STATUS_ADJ_EQ);
1917 if (n_cut > 0) {
1918 ctx = isl_basic_map_get_ctx(info[i].bmap);
1919 relax = isl_calloc_array(ctx, int, 1 + n_cut);
1920 if (!relax)
1921 return isl_change_error;
1922 relax[0] = k;
1923 try_relax = all_cut_by_one(i, j, info, relax + 1);
1924 if (try_relax < 0)
1925 change = isl_change_error;
1926 } else {
1927 try_relax = isl_bool_true;
1928 relax = &k;
1930 if (try_relax && change == isl_change_none)
1931 change = is_relaxed_extension(i, j, 1 + n_cut, relax, info);
1932 if (n_cut > 0)
1933 free(relax);
1934 if (change != isl_change_none)
1935 return change;
1937 change = can_wrap_in_facet(i, j, k, info, n_cut > 0);
1939 return change;
1942 /* At least one of the basic maps has an equality that is adjacent
1943 * to an inequality. Make sure that only one of the basic maps has
1944 * such an equality and that the other basic map has exactly one
1945 * inequality adjacent to an equality.
1946 * If the other basic map does not have such an inequality, then
1947 * check if all its constraints are either valid or cut constraints
1948 * and, if so, try wrapping in the first map into the second.
1949 * Otherwise, try to extend one basic map with the other or
1950 * wrap one basic map in the other.
1952 static enum isl_change check_adj_eq(int i, int j,
1953 struct isl_coalesce_info *info)
1955 if (any_eq(&info[i], STATUS_ADJ_INEQ) &&
1956 any_eq(&info[j], STATUS_ADJ_INEQ))
1957 /* ADJ EQ TOO MANY */
1958 return isl_change_none;
1960 if (any_eq(&info[i], STATUS_ADJ_INEQ))
1961 return check_adj_eq(j, i, info);
1963 /* j has an equality adjacent to an inequality in i */
1965 if (count_ineq(&info[i], STATUS_ADJ_EQ) != 1) {
1966 if (all_valid_or_cut(&info[i]))
1967 return can_wrap_in_set(i, j, info);
1968 return isl_change_none;
1970 if (any_eq(&info[i], STATUS_CUT))
1971 return isl_change_none;
1972 if (any_ineq(&info[j], STATUS_ADJ_EQ) ||
1973 any_ineq(&info[i], STATUS_ADJ_INEQ) ||
1974 any_ineq(&info[j], STATUS_ADJ_INEQ))
1975 /* ADJ EQ TOO MANY */
1976 return isl_change_none;
1978 return check_single_adj_eq(i, j, info);
1981 /* Disjunct "j" lies on a hyperplane that is adjacent to disjunct "i".
1982 * In particular, disjunct "i" has an inequality constraint that is adjacent
1983 * to a (combination of) equality constraint(s) of disjunct "j",
1984 * but disjunct "j" has no explicit equality constraint adjacent
1985 * to an inequality constraint of disjunct "i".
1987 * Disjunct "i" is already known not to have any equality constraints
1988 * that are adjacent to an equality or inequality constraint.
1989 * Check that, other than the inequality constraint mentioned above,
1990 * all other constraints of disjunct "i" are valid for disjunct "j".
1991 * If so, try and wrap in disjunct "j".
1993 static enum isl_change check_ineq_adj_eq(int i, int j,
1994 struct isl_coalesce_info *info)
1996 int k;
1998 if (any_eq(&info[i], STATUS_CUT))
1999 return isl_change_none;
2000 if (any_ineq(&info[i], STATUS_CUT))
2001 return isl_change_none;
2002 if (any_ineq(&info[i], STATUS_ADJ_INEQ))
2003 return isl_change_none;
2004 if (count_ineq(&info[i], STATUS_ADJ_EQ) != 1)
2005 return isl_change_none;
2007 k = find_ineq(&info[i], STATUS_ADJ_EQ);
2009 return can_wrap_in_facet(i, j, k, info, 0);
2012 /* The two basic maps lie on adjacent hyperplanes. In particular,
2013 * basic map "i" has an equality that lies parallel to basic map "j".
2014 * Check if we can wrap the facets around the parallel hyperplanes
2015 * to include the other set.
2017 * We perform basically the same operations as can_wrap_in_facet,
2018 * except that we don't need to select a facet of one of the sets.
2020 * \\ \\
2021 * \\ => \\
2022 * \ \|
2024 * If there is more than one equality of "i" adjacent to an equality of "j",
2025 * then the result will satisfy one or more equalities that are a linear
2026 * combination of these equalities. These will be encoded as pairs
2027 * of inequalities in the wrapping constraints and need to be made
2028 * explicit.
2030 static enum isl_change check_eq_adj_eq(int i, int j,
2031 struct isl_coalesce_info *info)
2033 int k;
2034 enum isl_change change = isl_change_none;
2035 int detect_equalities = 0;
2036 struct isl_wraps wraps;
2037 isl_ctx *ctx;
2038 isl_mat *mat;
2039 struct isl_set *set_i = NULL;
2040 struct isl_set *set_j = NULL;
2041 struct isl_vec *bound = NULL;
2042 unsigned total = isl_basic_map_total_dim(info[i].bmap);
2044 if (count_eq(&info[i], STATUS_ADJ_EQ) != 1)
2045 detect_equalities = 1;
2047 k = find_eq(&info[i], STATUS_ADJ_EQ);
2049 set_i = set_from_updated_bmap(info[i].bmap, info[i].tab);
2050 set_j = set_from_updated_bmap(info[j].bmap, info[j].tab);
2051 ctx = isl_basic_map_get_ctx(info[i].bmap);
2052 mat = isl_mat_alloc(ctx, 2 * (info[i].bmap->n_eq + info[j].bmap->n_eq) +
2053 info[i].bmap->n_ineq + info[j].bmap->n_ineq,
2054 1 + total);
2055 if (wraps_init(&wraps, mat, info, i, j) < 0)
2056 goto error;
2057 bound = isl_vec_alloc(ctx, 1 + total);
2058 if (!set_i || !set_j || !bound)
2059 goto error;
2061 if (k % 2 == 0)
2062 isl_seq_neg(bound->el, info[i].bmap->eq[k / 2], 1 + total);
2063 else
2064 isl_seq_cpy(bound->el, info[i].bmap->eq[k / 2], 1 + total);
2065 isl_int_add_ui(bound->el[0], bound->el[0], 1);
2067 isl_seq_cpy(wraps.mat->row[0], bound->el, 1 + total);
2068 wraps.mat->n_row = 1;
2070 if (add_wraps(&wraps, &info[j], bound->el, set_i) < 0)
2071 goto error;
2072 if (!wraps.mat->n_row)
2073 goto unbounded;
2075 isl_int_sub_ui(bound->el[0], bound->el[0], 1);
2076 isl_seq_neg(bound->el, bound->el, 1 + total);
2078 isl_seq_cpy(wraps.mat->row[wraps.mat->n_row], bound->el, 1 + total);
2079 wraps.mat->n_row++;
2081 if (add_wraps(&wraps, &info[i], bound->el, set_j) < 0)
2082 goto error;
2083 if (!wraps.mat->n_row)
2084 goto unbounded;
2086 change = fuse(i, j, info, wraps.mat, detect_equalities, 0);
2088 if (0) {
2089 error: change = isl_change_error;
2091 unbounded:
2093 wraps_free(&wraps);
2094 isl_set_free(set_i);
2095 isl_set_free(set_j);
2096 isl_vec_free(bound);
2098 return change;
2101 /* Initialize the "eq" and "ineq" fields of "info".
2103 static void init_status(struct isl_coalesce_info *info)
2105 info->eq = info->ineq = NULL;
2108 /* Set info->eq to the positions of the equalities of info->bmap
2109 * with respect to the basic map represented by "tab".
2110 * If info->eq has already been computed, then do not compute it again.
2112 static void set_eq_status_in(struct isl_coalesce_info *info,
2113 struct isl_tab *tab)
2115 if (info->eq)
2116 return;
2117 info->eq = eq_status_in(info->bmap, tab);
2120 /* Set info->ineq to the positions of the inequalities of info->bmap
2121 * with respect to the basic map represented by "tab".
2122 * If info->ineq has already been computed, then do not compute it again.
2124 static void set_ineq_status_in(struct isl_coalesce_info *info,
2125 struct isl_tab *tab)
2127 if (info->ineq)
2128 return;
2129 info->ineq = ineq_status_in(info->bmap, info->tab, tab);
2132 /* Free the memory allocated by the "eq" and "ineq" fields of "info".
2133 * This function assumes that init_status has been called on "info" first,
2134 * after which the "eq" and "ineq" fields may or may not have been
2135 * assigned a newly allocated array.
2137 static void clear_status(struct isl_coalesce_info *info)
2139 free(info->eq);
2140 free(info->ineq);
2143 /* Are all inequality constraints of the basic map represented by "info"
2144 * valid for the other basic map, except for a single constraint
2145 * that is adjacent to an inequality constraint of the other basic map?
2147 static int all_ineq_valid_or_single_adj_ineq(struct isl_coalesce_info *info)
2149 int i;
2150 int k = -1;
2152 for (i = 0; i < info->bmap->n_ineq; ++i) {
2153 if (info->ineq[i] == STATUS_REDUNDANT)
2154 continue;
2155 if (info->ineq[i] == STATUS_VALID)
2156 continue;
2157 if (info->ineq[i] != STATUS_ADJ_INEQ)
2158 return 0;
2159 if (k != -1)
2160 return 0;
2161 k = i;
2164 return k != -1;
2167 /* Basic map "i" has one or more equality constraints that separate it
2168 * from basic map "j". Check if it happens to be an extension
2169 * of basic map "j".
2170 * In particular, check that all constraints of "j" are valid for "i",
2171 * except for one inequality constraint that is adjacent
2172 * to an inequality constraints of "i".
2173 * If so, check for "i" being an extension of "j" by calling
2174 * is_adj_ineq_extension.
2176 * Clean up the memory allocated for keeping track of the status
2177 * of the constraints before returning.
2179 static enum isl_change separating_equality(int i, int j,
2180 struct isl_coalesce_info *info)
2182 enum isl_change change = isl_change_none;
2184 if (all(info[j].eq, 2 * info[j].bmap->n_eq, STATUS_VALID) &&
2185 all_ineq_valid_or_single_adj_ineq(&info[j]))
2186 change = is_adj_ineq_extension(j, i, info);
2188 clear_status(&info[i]);
2189 clear_status(&info[j]);
2190 return change;
2193 /* Check if the union of the given pair of basic maps
2194 * can be represented by a single basic map.
2195 * If so, replace the pair by the single basic map and return
2196 * isl_change_drop_first, isl_change_drop_second or isl_change_fuse.
2197 * Otherwise, return isl_change_none.
2198 * The two basic maps are assumed to live in the same local space.
2199 * The "eq" and "ineq" fields of info[i] and info[j] are assumed
2200 * to have been initialized by the caller, either to NULL or
2201 * to valid information.
2203 * We first check the effect of each constraint of one basic map
2204 * on the other basic map.
2205 * The constraint may be
2206 * redundant the constraint is redundant in its own
2207 * basic map and should be ignore and removed
2208 * in the end
2209 * valid all (integer) points of the other basic map
2210 * satisfy the constraint
2211 * separate no (integer) point of the other basic map
2212 * satisfies the constraint
2213 * cut some but not all points of the other basic map
2214 * satisfy the constraint
2215 * adj_eq the given constraint is adjacent (on the outside)
2216 * to an equality of the other basic map
2217 * adj_ineq the given constraint is adjacent (on the outside)
2218 * to an inequality of the other basic map
2220 * We consider seven cases in which we can replace the pair by a single
2221 * basic map. We ignore all "redundant" constraints.
2223 * 1. all constraints of one basic map are valid
2224 * => the other basic map is a subset and can be removed
2226 * 2. all constraints of both basic maps are either "valid" or "cut"
2227 * and the facets corresponding to the "cut" constraints
2228 * of one of the basic maps lies entirely inside the other basic map
2229 * => the pair can be replaced by a basic map consisting
2230 * of the valid constraints in both basic maps
2232 * 3. there is a single pair of adjacent inequalities
2233 * (all other constraints are "valid")
2234 * => the pair can be replaced by a basic map consisting
2235 * of the valid constraints in both basic maps
2237 * 4. one basic map has a single adjacent inequality, while the other
2238 * constraints are "valid". The other basic map has some
2239 * "cut" constraints, but replacing the adjacent inequality by
2240 * its opposite and adding the valid constraints of the other
2241 * basic map results in a subset of the other basic map
2242 * => the pair can be replaced by a basic map consisting
2243 * of the valid constraints in both basic maps
2245 * 5. there is a single adjacent pair of an inequality and an equality,
2246 * the other constraints of the basic map containing the inequality are
2247 * "valid". Moreover, if the inequality the basic map is relaxed
2248 * and then turned into an equality, then resulting facet lies
2249 * entirely inside the other basic map
2250 * => the pair can be replaced by the basic map containing
2251 * the inequality, with the inequality relaxed.
2253 * 6. there is a single inequality adjacent to an equality,
2254 * the other constraints of the basic map containing the inequality are
2255 * "valid". Moreover, the facets corresponding to both
2256 * the inequality and the equality can be wrapped around their
2257 * ridges to include the other basic map
2258 * => the pair can be replaced by a basic map consisting
2259 * of the valid constraints in both basic maps together
2260 * with all wrapping constraints
2262 * 7. one of the basic maps extends beyond the other by at most one.
2263 * Moreover, the facets corresponding to the cut constraints and
2264 * the pieces of the other basic map at offset one from these cut
2265 * constraints can be wrapped around their ridges to include
2266 * the union of the two basic maps
2267 * => the pair can be replaced by a basic map consisting
2268 * of the valid constraints in both basic maps together
2269 * with all wrapping constraints
2271 * 8. the two basic maps live in adjacent hyperplanes. In principle
2272 * such sets can always be combined through wrapping, but we impose
2273 * that there is only one such pair, to avoid overeager coalescing.
2275 * Throughout the computation, we maintain a collection of tableaus
2276 * corresponding to the basic maps. When the basic maps are dropped
2277 * or combined, the tableaus are modified accordingly.
2279 static enum isl_change coalesce_local_pair_reuse(int i, int j,
2280 struct isl_coalesce_info *info)
2282 enum isl_change change = isl_change_none;
2284 set_ineq_status_in(&info[i], info[j].tab);
2285 if (info[i].bmap->n_ineq && !info[i].ineq)
2286 goto error;
2287 if (any_ineq(&info[i], STATUS_ERROR))
2288 goto error;
2289 if (any_ineq(&info[i], STATUS_SEPARATE))
2290 goto done;
2292 set_ineq_status_in(&info[j], info[i].tab);
2293 if (info[j].bmap->n_ineq && !info[j].ineq)
2294 goto error;
2295 if (any_ineq(&info[j], STATUS_ERROR))
2296 goto error;
2297 if (any_ineq(&info[j], STATUS_SEPARATE))
2298 goto done;
2300 set_eq_status_in(&info[i], info[j].tab);
2301 if (info[i].bmap->n_eq && !info[i].eq)
2302 goto error;
2303 if (any_eq(&info[i], STATUS_ERROR))
2304 goto error;
2306 set_eq_status_in(&info[j], info[i].tab);
2307 if (info[j].bmap->n_eq && !info[j].eq)
2308 goto error;
2309 if (any_eq(&info[j], STATUS_ERROR))
2310 goto error;
2312 if (any_eq(&info[i], STATUS_SEPARATE))
2313 return separating_equality(i, j, info);
2314 if (any_eq(&info[j], STATUS_SEPARATE))
2315 return separating_equality(j, i, info);
2317 if (all(info[i].eq, 2 * info[i].bmap->n_eq, STATUS_VALID) &&
2318 all(info[i].ineq, info[i].bmap->n_ineq, STATUS_VALID)) {
2319 drop(&info[j]);
2320 change = isl_change_drop_second;
2321 } else if (all(info[j].eq, 2 * info[j].bmap->n_eq, STATUS_VALID) &&
2322 all(info[j].ineq, info[j].bmap->n_ineq, STATUS_VALID)) {
2323 drop(&info[i]);
2324 change = isl_change_drop_first;
2325 } else if (any_eq(&info[i], STATUS_ADJ_EQ)) {
2326 change = check_eq_adj_eq(i, j, info);
2327 } else if (any_eq(&info[j], STATUS_ADJ_EQ)) {
2328 change = check_eq_adj_eq(j, i, info);
2329 } else if (any_eq(&info[i], STATUS_ADJ_INEQ) ||
2330 any_eq(&info[j], STATUS_ADJ_INEQ)) {
2331 change = check_adj_eq(i, j, info);
2332 } else if (any_ineq(&info[i], STATUS_ADJ_EQ)) {
2333 change = check_ineq_adj_eq(i, j, info);
2334 } else if (any_ineq(&info[j], STATUS_ADJ_EQ)) {
2335 change = check_ineq_adj_eq(j, i, info);
2336 } else if (any_ineq(&info[i], STATUS_ADJ_INEQ) ||
2337 any_ineq(&info[j], STATUS_ADJ_INEQ)) {
2338 change = check_adj_ineq(i, j, info);
2339 } else {
2340 if (!any_eq(&info[i], STATUS_CUT) &&
2341 !any_eq(&info[j], STATUS_CUT))
2342 change = check_facets(i, j, info);
2343 if (change == isl_change_none)
2344 change = check_wrap(i, j, info);
2347 done:
2348 clear_status(&info[i]);
2349 clear_status(&info[j]);
2350 return change;
2351 error:
2352 clear_status(&info[i]);
2353 clear_status(&info[j]);
2354 return isl_change_error;
2357 /* Check if the union of the given pair of basic maps
2358 * can be represented by a single basic map.
2359 * If so, replace the pair by the single basic map and return
2360 * isl_change_drop_first, isl_change_drop_second or isl_change_fuse.
2361 * Otherwise, return isl_change_none.
2362 * The two basic maps are assumed to live in the same local space.
2364 static enum isl_change coalesce_local_pair(int i, int j,
2365 struct isl_coalesce_info *info)
2367 init_status(&info[i]);
2368 init_status(&info[j]);
2369 return coalesce_local_pair_reuse(i, j, info);
2372 /* Shift the integer division at position "div" of the basic map
2373 * represented by "info" by "shift".
2375 * That is, if the integer division has the form
2377 * floor(f(x)/d)
2379 * then replace it by
2381 * floor((f(x) + shift * d)/d) - shift
2383 static isl_stat shift_div(struct isl_coalesce_info *info, int div,
2384 isl_int shift)
2386 unsigned total;
2388 info->bmap = isl_basic_map_shift_div(info->bmap, div, 0, shift);
2389 if (!info->bmap)
2390 return isl_stat_error;
2392 total = isl_basic_map_dim(info->bmap, isl_dim_all);
2393 total -= isl_basic_map_dim(info->bmap, isl_dim_div);
2394 if (isl_tab_shift_var(info->tab, total + div, shift) < 0)
2395 return isl_stat_error;
2397 return isl_stat_ok;
2400 /* If the integer division at position "div" is defined by an equality,
2401 * i.e., a stride constraint, then change the integer division expression
2402 * to have a constant term equal to zero.
2404 * Let the equality constraint be
2406 * c + f + m a = 0
2408 * The integer division expression is then typically of the form
2410 * a = floor((-f - c')/m)
2412 * The integer division is first shifted by t = floor(c/m),
2413 * turning the equality constraint into
2415 * c - m floor(c/m) + f + m a' = 0
2417 * i.e.,
2419 * (c mod m) + f + m a' = 0
2421 * That is,
2423 * a' = (-f - (c mod m))/m = floor((-f)/m)
2425 * because a' is an integer and 0 <= (c mod m) < m.
2426 * The constant term of a' can therefore be zeroed out,
2427 * but only if the integer division expression is of the expected form.
2429 static isl_stat normalize_stride_div(struct isl_coalesce_info *info, int div)
2431 isl_bool defined, valid;
2432 isl_stat r;
2433 isl_constraint *c;
2434 isl_int shift, stride;
2436 defined = isl_basic_map_has_defining_equality(info->bmap, isl_dim_div,
2437 div, &c);
2438 if (defined < 0)
2439 return isl_stat_error;
2440 if (!defined)
2441 return isl_stat_ok;
2442 if (!c)
2443 return isl_stat_error;
2444 valid = isl_constraint_is_div_equality(c, div);
2445 isl_int_init(shift);
2446 isl_int_init(stride);
2447 isl_constraint_get_constant(c, &shift);
2448 isl_constraint_get_coefficient(c, isl_dim_div, div, &stride);
2449 isl_int_fdiv_q(shift, shift, stride);
2450 r = shift_div(info, div, shift);
2451 isl_int_clear(stride);
2452 isl_int_clear(shift);
2453 isl_constraint_free(c);
2454 if (r < 0 || valid < 0)
2455 return isl_stat_error;
2456 if (!valid)
2457 return isl_stat_ok;
2458 info->bmap = isl_basic_map_set_div_expr_constant_num_si_inplace(
2459 info->bmap, div, 0);
2460 if (!info->bmap)
2461 return isl_stat_error;
2462 return isl_stat_ok;
2465 /* The basic maps represented by "info1" and "info2" are known
2466 * to have the same number of integer divisions.
2467 * Check if pairs of integer divisions are equal to each other
2468 * despite the fact that they differ by a rational constant.
2470 * In particular, look for any pair of integer divisions that
2471 * only differ in their constant terms.
2472 * If either of these integer divisions is defined
2473 * by stride constraints, then modify it to have a zero constant term.
2474 * If both are defined by stride constraints then in the end they will have
2475 * the same (zero) constant term.
2477 static isl_stat harmonize_stride_divs(struct isl_coalesce_info *info1,
2478 struct isl_coalesce_info *info2)
2480 int i, n;
2482 n = isl_basic_map_dim(info1->bmap, isl_dim_div);
2483 for (i = 0; i < n; ++i) {
2484 isl_bool known, harmonize;
2486 known = isl_basic_map_div_is_known(info1->bmap, i);
2487 if (known >= 0 && known)
2488 known = isl_basic_map_div_is_known(info2->bmap, i);
2489 if (known < 0)
2490 return isl_stat_error;
2491 if (!known)
2492 continue;
2493 harmonize = isl_basic_map_equal_div_expr_except_constant(
2494 info1->bmap, i, info2->bmap, i);
2495 if (harmonize < 0)
2496 return isl_stat_error;
2497 if (!harmonize)
2498 continue;
2499 if (normalize_stride_div(info1, i) < 0)
2500 return isl_stat_error;
2501 if (normalize_stride_div(info2, i) < 0)
2502 return isl_stat_error;
2505 return isl_stat_ok;
2508 /* If "shift" is an integer constant, then shift the integer division
2509 * at position "div" of the basic map represented by "info" by "shift".
2510 * If "shift" is not an integer constant, then do nothing.
2511 * If "shift" is equal to zero, then no shift needs to be performed either.
2513 * That is, if the integer division has the form
2515 * floor(f(x)/d)
2517 * then replace it by
2519 * floor((f(x) + shift * d)/d) - shift
2521 static isl_stat shift_if_cst_int(struct isl_coalesce_info *info, int div,
2522 __isl_keep isl_aff *shift)
2524 isl_bool cst;
2525 isl_stat r;
2526 isl_int d;
2527 isl_val *c;
2529 cst = isl_aff_is_cst(shift);
2530 if (cst < 0 || !cst)
2531 return cst < 0 ? isl_stat_error : isl_stat_ok;
2533 c = isl_aff_get_constant_val(shift);
2534 cst = isl_val_is_int(c);
2535 if (cst >= 0 && cst)
2536 cst = isl_bool_not(isl_val_is_zero(c));
2537 if (cst < 0 || !cst) {
2538 isl_val_free(c);
2539 return cst < 0 ? isl_stat_error : isl_stat_ok;
2542 isl_int_init(d);
2543 r = isl_val_get_num_isl_int(c, &d);
2544 if (r >= 0)
2545 r = shift_div(info, div, d);
2546 isl_int_clear(d);
2548 isl_val_free(c);
2550 return r;
2553 /* Check if some of the divs in the basic map represented by "info1"
2554 * are shifts of the corresponding divs in the basic map represented
2555 * by "info2", taking into account the equality constraints "eq1" of "info1"
2556 * and "eq2" of "info2". If so, align them with those of "info2".
2557 * "info1" and "info2" are assumed to have the same number
2558 * of integer divisions.
2560 * An integer division is considered to be a shift of another integer
2561 * division if, after simplification with respect to the equality
2562 * constraints of the other basic map, one is equal to the other
2563 * plus a constant.
2565 * In particular, for each pair of integer divisions, if both are known,
2566 * have the same denominator and are not already equal to each other,
2567 * simplify each with respect to the equality constraints
2568 * of the other basic map. If the difference is an integer constant,
2569 * then move this difference outside.
2570 * That is, if, after simplification, one integer division is of the form
2572 * floor((f(x) + c_1)/d)
2574 * while the other is of the form
2576 * floor((f(x) + c_2)/d)
2578 * and n = (c_2 - c_1)/d is an integer, then replace the first
2579 * integer division by
2581 * floor((f_1(x) + c_1 + n * d)/d) - n,
2583 * where floor((f_1(x) + c_1 + n * d)/d) = floor((f2(x) + c_2)/d)
2584 * after simplification with respect to the equality constraints.
2586 static isl_stat harmonize_divs_with_hulls(struct isl_coalesce_info *info1,
2587 struct isl_coalesce_info *info2, __isl_keep isl_basic_set *eq1,
2588 __isl_keep isl_basic_set *eq2)
2590 int i;
2591 int total;
2592 isl_local_space *ls1, *ls2;
2594 total = isl_basic_map_total_dim(info1->bmap);
2595 ls1 = isl_local_space_wrap(isl_basic_map_get_local_space(info1->bmap));
2596 ls2 = isl_local_space_wrap(isl_basic_map_get_local_space(info2->bmap));
2597 for (i = 0; i < info1->bmap->n_div; ++i) {
2598 isl_stat r;
2599 isl_aff *div1, *div2;
2601 if (!isl_local_space_div_is_known(ls1, i) ||
2602 !isl_local_space_div_is_known(ls2, i))
2603 continue;
2604 if (isl_int_ne(info1->bmap->div[i][0], info2->bmap->div[i][0]))
2605 continue;
2606 if (isl_seq_eq(info1->bmap->div[i] + 1,
2607 info2->bmap->div[i] + 1, 1 + total))
2608 continue;
2609 div1 = isl_local_space_get_div(ls1, i);
2610 div2 = isl_local_space_get_div(ls2, i);
2611 div1 = isl_aff_substitute_equalities(div1,
2612 isl_basic_set_copy(eq2));
2613 div2 = isl_aff_substitute_equalities(div2,
2614 isl_basic_set_copy(eq1));
2615 div2 = isl_aff_sub(div2, div1);
2616 r = shift_if_cst_int(info1, i, div2);
2617 isl_aff_free(div2);
2618 if (r < 0)
2619 break;
2621 isl_local_space_free(ls1);
2622 isl_local_space_free(ls2);
2624 if (i < info1->bmap->n_div)
2625 return isl_stat_error;
2626 return isl_stat_ok;
2629 /* Check if some of the divs in the basic map represented by "info1"
2630 * are shifts of the corresponding divs in the basic map represented
2631 * by "info2". If so, align them with those of "info2".
2632 * Only do this if "info1" and "info2" have the same number
2633 * of integer divisions.
2635 * An integer division is considered to be a shift of another integer
2636 * division if, after simplification with respect to the equality
2637 * constraints of the other basic map, one is equal to the other
2638 * plus a constant.
2640 * First check if pairs of integer divisions are equal to each other
2641 * despite the fact that they differ by a rational constant.
2642 * If so, try and arrange for them to have the same constant term.
2644 * Then, extract the equality constraints and continue with
2645 * harmonize_divs_with_hulls.
2647 * If the equality constraints of both basic maps are the same,
2648 * then there is no need to perform any shifting since
2649 * the coefficients of the integer divisions should have been
2650 * reduced in the same way.
2652 static isl_stat harmonize_divs(struct isl_coalesce_info *info1,
2653 struct isl_coalesce_info *info2)
2655 isl_bool equal;
2656 isl_basic_map *bmap1, *bmap2;
2657 isl_basic_set *eq1, *eq2;
2658 isl_stat r;
2660 if (!info1->bmap || !info2->bmap)
2661 return isl_stat_error;
2663 if (info1->bmap->n_div != info2->bmap->n_div)
2664 return isl_stat_ok;
2665 if (info1->bmap->n_div == 0)
2666 return isl_stat_ok;
2668 if (harmonize_stride_divs(info1, info2) < 0)
2669 return isl_stat_error;
2671 bmap1 = isl_basic_map_copy(info1->bmap);
2672 bmap2 = isl_basic_map_copy(info2->bmap);
2673 eq1 = isl_basic_map_wrap(isl_basic_map_plain_affine_hull(bmap1));
2674 eq2 = isl_basic_map_wrap(isl_basic_map_plain_affine_hull(bmap2));
2675 equal = isl_basic_set_plain_is_equal(eq1, eq2);
2676 if (equal < 0)
2677 r = isl_stat_error;
2678 else if (equal)
2679 r = isl_stat_ok;
2680 else
2681 r = harmonize_divs_with_hulls(info1, info2, eq1, eq2);
2682 isl_basic_set_free(eq1);
2683 isl_basic_set_free(eq2);
2685 return r;
2688 /* Do the two basic maps live in the same local space, i.e.,
2689 * do they have the same (known) divs?
2690 * If either basic map has any unknown divs, then we can only assume
2691 * that they do not live in the same local space.
2693 static isl_bool same_divs(__isl_keep isl_basic_map *bmap1,
2694 __isl_keep isl_basic_map *bmap2)
2696 int i;
2697 isl_bool known;
2698 int total;
2700 if (!bmap1 || !bmap2)
2701 return isl_bool_error;
2702 if (bmap1->n_div != bmap2->n_div)
2703 return isl_bool_false;
2705 if (bmap1->n_div == 0)
2706 return isl_bool_true;
2708 known = isl_basic_map_divs_known(bmap1);
2709 if (known < 0 || !known)
2710 return known;
2711 known = isl_basic_map_divs_known(bmap2);
2712 if (known < 0 || !known)
2713 return known;
2715 total = isl_basic_map_total_dim(bmap1);
2716 for (i = 0; i < bmap1->n_div; ++i)
2717 if (!isl_seq_eq(bmap1->div[i], bmap2->div[i], 2 + total))
2718 return isl_bool_false;
2720 return isl_bool_true;
2723 /* Assuming that "tab" contains the equality constraints and
2724 * the initial inequality constraints of "bmap", copy the remaining
2725 * inequality constraints of "bmap" to "Tab".
2727 static isl_stat copy_ineq(struct isl_tab *tab, __isl_keep isl_basic_map *bmap)
2729 int i, n_ineq;
2731 if (!bmap)
2732 return isl_stat_error;
2734 n_ineq = tab->n_con - tab->n_eq;
2735 for (i = n_ineq; i < bmap->n_ineq; ++i)
2736 if (isl_tab_add_ineq(tab, bmap->ineq[i]) < 0)
2737 return isl_stat_error;
2739 return isl_stat_ok;
2742 /* Description of an integer division that is added
2743 * during an expansion.
2744 * "pos" is the position of the corresponding variable.
2745 * "cst" indicates whether this integer division has a fixed value.
2746 * "val" contains the fixed value, if the value is fixed.
2748 struct isl_expanded {
2749 int pos;
2750 isl_bool cst;
2751 isl_int val;
2754 /* For each of the "n" integer division variables "expanded",
2755 * if the variable has a fixed value, then add two inequality
2756 * constraints expressing the fixed value.
2757 * Otherwise, add the corresponding div constraints.
2758 * The caller is responsible for removing the div constraints
2759 * that it added for all these "n" integer divisions.
2761 * The div constraints and the pair of inequality constraints
2762 * forcing the fixed value cannot both be added for a given variable
2763 * as the combination may render some of the original constraints redundant.
2764 * These would then be ignored during the coalescing detection,
2765 * while they could remain in the fused result.
2767 * The two added inequality constraints are
2769 * -a + v >= 0
2770 * a - v >= 0
2772 * with "a" the variable and "v" its fixed value.
2773 * The facet corresponding to one of these two constraints is selected
2774 * in the tableau to ensure that the pair of inequality constraints
2775 * is treated as an equality constraint.
2777 * The information in info->ineq is thrown away because it was
2778 * computed in terms of div constraints, while some of those
2779 * have now been replaced by these pairs of inequality constraints.
2781 static isl_stat fix_constant_divs(struct isl_coalesce_info *info,
2782 int n, struct isl_expanded *expanded)
2784 unsigned o_div;
2785 int i;
2786 isl_vec *ineq;
2788 o_div = isl_basic_map_offset(info->bmap, isl_dim_div) - 1;
2789 ineq = isl_vec_alloc(isl_tab_get_ctx(info->tab), 1 + info->tab->n_var);
2790 if (!ineq)
2791 return isl_stat_error;
2792 isl_seq_clr(ineq->el + 1, info->tab->n_var);
2794 for (i = 0; i < n; ++i) {
2795 if (!expanded[i].cst) {
2796 info->bmap = isl_basic_map_extend_constraints(
2797 info->bmap, 0, 2);
2798 if (isl_basic_map_add_div_constraints(info->bmap,
2799 expanded[i].pos - o_div) < 0)
2800 break;
2801 } else {
2802 isl_int_set_si(ineq->el[1 + expanded[i].pos], -1);
2803 isl_int_set(ineq->el[0], expanded[i].val);
2804 info->bmap = isl_basic_map_add_ineq(info->bmap,
2805 ineq->el);
2806 isl_int_set_si(ineq->el[1 + expanded[i].pos], 1);
2807 isl_int_neg(ineq->el[0], expanded[i].val);
2808 info->bmap = isl_basic_map_add_ineq(info->bmap,
2809 ineq->el);
2810 isl_int_set_si(ineq->el[1 + expanded[i].pos], 0);
2812 if (copy_ineq(info->tab, info->bmap) < 0)
2813 break;
2814 if (expanded[i].cst &&
2815 isl_tab_select_facet(info->tab, info->tab->n_con - 1) < 0)
2816 break;
2819 isl_vec_free(ineq);
2821 clear_status(info);
2822 init_status(info);
2824 return i < n ? isl_stat_error : isl_stat_ok;
2827 /* Insert the "n" integer division variables "expanded"
2828 * into info->tab and info->bmap and
2829 * update info->ineq with respect to the redundant constraints
2830 * in the resulting tableau.
2831 * "bmap" contains the result of this insertion in info->bmap,
2832 * while info->bmap is the original version
2833 * of "bmap", i.e., the one that corresponds to the current
2834 * state of info->tab. The number of constraints in info->bmap
2835 * is assumed to be the same as the number of constraints
2836 * in info->tab. This is required to be able to detect
2837 * the extra constraints in "bmap".
2839 * In particular, introduce extra variables corresponding
2840 * to the extra integer divisions and add the div constraints
2841 * that were added to "bmap" after info->tab was created
2842 * from info->bmap.
2843 * Furthermore, check if these extra integer divisions happen
2844 * to attain a fixed integer value in info->tab.
2845 * If so, replace the corresponding div constraints by pairs
2846 * of inequality constraints that fix these
2847 * integer divisions to their single integer values.
2848 * Replace info->bmap by "bmap" to match the changes to info->tab.
2849 * info->ineq was computed without a tableau and therefore
2850 * does not take into account the redundant constraints
2851 * in the tableau. Mark them here.
2852 * There is no need to check the newly added div constraints
2853 * since they cannot be redundant.
2854 * The redundancy check is not performed when constants have been discovered
2855 * since info->ineq is completely thrown away in this case.
2857 static isl_stat tab_insert_divs(struct isl_coalesce_info *info,
2858 int n, struct isl_expanded *expanded, __isl_take isl_basic_map *bmap)
2860 int i, n_ineq;
2861 unsigned n_eq;
2862 struct isl_tab_undo *snap;
2863 int any;
2865 if (!bmap)
2866 return isl_stat_error;
2867 if (info->bmap->n_eq + info->bmap->n_ineq != info->tab->n_con)
2868 isl_die(isl_basic_map_get_ctx(bmap), isl_error_internal,
2869 "original tableau does not correspond "
2870 "to original basic map", goto error);
2872 if (isl_tab_extend_vars(info->tab, n) < 0)
2873 goto error;
2874 if (isl_tab_extend_cons(info->tab, 2 * n) < 0)
2875 goto error;
2877 for (i = 0; i < n; ++i) {
2878 if (isl_tab_insert_var(info->tab, expanded[i].pos) < 0)
2879 goto error;
2882 snap = isl_tab_snap(info->tab);
2884 n_ineq = info->tab->n_con - info->tab->n_eq;
2885 if (copy_ineq(info->tab, bmap) < 0)
2886 goto error;
2888 isl_basic_map_free(info->bmap);
2889 info->bmap = bmap;
2891 any = 0;
2892 for (i = 0; i < n; ++i) {
2893 expanded[i].cst = isl_tab_is_constant(info->tab,
2894 expanded[i].pos, &expanded[i].val);
2895 if (expanded[i].cst < 0)
2896 return isl_stat_error;
2897 if (expanded[i].cst)
2898 any = 1;
2901 if (any) {
2902 if (isl_tab_rollback(info->tab, snap) < 0)
2903 return isl_stat_error;
2904 info->bmap = isl_basic_map_cow(info->bmap);
2905 if (isl_basic_map_free_inequality(info->bmap, 2 * n) < 0)
2906 return isl_stat_error;
2908 return fix_constant_divs(info, n, expanded);
2911 n_eq = info->bmap->n_eq;
2912 for (i = 0; i < n_ineq; ++i) {
2913 if (isl_tab_is_redundant(info->tab, n_eq + i))
2914 info->ineq[i] = STATUS_REDUNDANT;
2917 return isl_stat_ok;
2918 error:
2919 isl_basic_map_free(bmap);
2920 return isl_stat_error;
2923 /* Expand info->tab and info->bmap in the same way "bmap" was expanded
2924 * in isl_basic_map_expand_divs using the expansion "exp" and
2925 * update info->ineq with respect to the redundant constraints
2926 * in the resulting tableau. info->bmap is the original version
2927 * of "bmap", i.e., the one that corresponds to the current
2928 * state of info->tab. The number of constraints in info->bmap
2929 * is assumed to be the same as the number of constraints
2930 * in info->tab. This is required to be able to detect
2931 * the extra constraints in "bmap".
2933 * Extract the positions where extra local variables are introduced
2934 * from "exp" and call tab_insert_divs.
2936 static isl_stat expand_tab(struct isl_coalesce_info *info, int *exp,
2937 __isl_take isl_basic_map *bmap)
2939 isl_ctx *ctx;
2940 struct isl_expanded *expanded;
2941 int i, j, k, n;
2942 int extra_var;
2943 unsigned total, pos, n_div;
2944 isl_stat r;
2946 total = isl_basic_map_dim(bmap, isl_dim_all);
2947 n_div = isl_basic_map_dim(bmap, isl_dim_div);
2948 pos = total - n_div;
2949 extra_var = total - info->tab->n_var;
2950 n = n_div - extra_var;
2952 ctx = isl_basic_map_get_ctx(bmap);
2953 expanded = isl_calloc_array(ctx, struct isl_expanded, extra_var);
2954 if (extra_var && !expanded)
2955 goto error;
2957 i = 0;
2958 k = 0;
2959 for (j = 0; j < n_div; ++j) {
2960 if (i < n && exp[i] == j) {
2961 ++i;
2962 continue;
2964 expanded[k++].pos = pos + j;
2967 for (k = 0; k < extra_var; ++k)
2968 isl_int_init(expanded[k].val);
2970 r = tab_insert_divs(info, extra_var, expanded, bmap);
2972 for (k = 0; k < extra_var; ++k)
2973 isl_int_clear(expanded[k].val);
2974 free(expanded);
2976 return r;
2977 error:
2978 isl_basic_map_free(bmap);
2979 return isl_stat_error;
2982 /* Check if the union of the basic maps represented by info[i] and info[j]
2983 * can be represented by a single basic map,
2984 * after expanding the divs of info[i] to match those of info[j].
2985 * If so, replace the pair by the single basic map and return
2986 * isl_change_drop_first, isl_change_drop_second or isl_change_fuse.
2987 * Otherwise, return isl_change_none.
2989 * The caller has already checked for info[j] being a subset of info[i].
2990 * If some of the divs of info[j] are unknown, then the expanded info[i]
2991 * will not have the corresponding div constraints. The other patterns
2992 * therefore cannot apply. Skip the computation in this case.
2994 * The expansion is performed using the divs "div" and expansion "exp"
2995 * computed by the caller.
2996 * info[i].bmap has already been expanded and the result is passed in
2997 * as "bmap".
2998 * The "eq" and "ineq" fields of info[i] reflect the status of
2999 * the constraints of the expanded "bmap" with respect to info[j].tab.
3000 * However, inequality constraints that are redundant in info[i].tab
3001 * have not yet been marked as such because no tableau was available.
3003 * Replace info[i].bmap by "bmap" and expand info[i].tab as well,
3004 * updating info[i].ineq with respect to the redundant constraints.
3005 * Then try and coalesce the expanded info[i] with info[j],
3006 * reusing the information in info[i].eq and info[i].ineq.
3007 * If this does not result in any coalescing or if it results in info[j]
3008 * getting dropped (which should not happen in practice, since the case
3009 * of info[j] being a subset of info[i] has already been checked by
3010 * the caller), then revert info[i] to its original state.
3012 static enum isl_change coalesce_expand_tab_divs(__isl_take isl_basic_map *bmap,
3013 int i, int j, struct isl_coalesce_info *info, __isl_keep isl_mat *div,
3014 int *exp)
3016 isl_bool known;
3017 isl_basic_map *bmap_i;
3018 struct isl_tab_undo *snap;
3019 enum isl_change change = isl_change_none;
3021 known = isl_basic_map_divs_known(info[j].bmap);
3022 if (known < 0 || !known) {
3023 clear_status(&info[i]);
3024 isl_basic_map_free(bmap);
3025 return known < 0 ? isl_change_error : isl_change_none;
3028 bmap_i = isl_basic_map_copy(info[i].bmap);
3029 snap = isl_tab_snap(info[i].tab);
3030 if (expand_tab(&info[i], exp, bmap) < 0)
3031 change = isl_change_error;
3033 init_status(&info[j]);
3034 if (change == isl_change_none)
3035 change = coalesce_local_pair_reuse(i, j, info);
3036 else
3037 clear_status(&info[i]);
3038 if (change != isl_change_none && change != isl_change_drop_second) {
3039 isl_basic_map_free(bmap_i);
3040 } else {
3041 isl_basic_map_free(info[i].bmap);
3042 info[i].bmap = bmap_i;
3044 if (isl_tab_rollback(info[i].tab, snap) < 0)
3045 change = isl_change_error;
3048 return change;
3051 /* Check if the union of "bmap" and the basic map represented by info[j]
3052 * can be represented by a single basic map,
3053 * after expanding the divs of "bmap" to match those of info[j].
3054 * If so, replace the pair by the single basic map and return
3055 * isl_change_drop_first, isl_change_drop_second or isl_change_fuse.
3056 * Otherwise, return isl_change_none.
3058 * In particular, check if the expanded "bmap" contains the basic map
3059 * represented by the tableau info[j].tab.
3060 * The expansion is performed using the divs "div" and expansion "exp"
3061 * computed by the caller.
3062 * Then we check if all constraints of the expanded "bmap" are valid for
3063 * info[j].tab.
3065 * If "i" is not equal to -1, then "bmap" is equal to info[i].bmap.
3066 * In this case, the positions of the constraints of info[i].bmap
3067 * with respect to the basic map represented by info[j] are stored
3068 * in info[i].
3070 * If the expanded "bmap" does not contain the basic map
3071 * represented by the tableau info[j].tab and if "i" is not -1,
3072 * i.e., if the original "bmap" is info[i].bmap, then expand info[i].tab
3073 * as well and check if that results in coalescing.
3075 static enum isl_change coalesce_with_expanded_divs(
3076 __isl_keep isl_basic_map *bmap, int i, int j,
3077 struct isl_coalesce_info *info, __isl_keep isl_mat *div, int *exp)
3079 enum isl_change change = isl_change_none;
3080 struct isl_coalesce_info info_local, *info_i;
3082 info_i = i >= 0 ? &info[i] : &info_local;
3083 init_status(info_i);
3084 bmap = isl_basic_map_copy(bmap);
3085 bmap = isl_basic_map_expand_divs(bmap, isl_mat_copy(div), exp);
3086 bmap = isl_basic_map_mark_final(bmap);
3088 if (!bmap)
3089 goto error;
3091 info_local.bmap = bmap;
3092 info_i->eq = eq_status_in(bmap, info[j].tab);
3093 if (bmap->n_eq && !info_i->eq)
3094 goto error;
3095 if (any_eq(info_i, STATUS_ERROR))
3096 goto error;
3097 if (any_eq(info_i, STATUS_SEPARATE))
3098 goto done;
3100 info_i->ineq = ineq_status_in(bmap, NULL, info[j].tab);
3101 if (bmap->n_ineq && !info_i->ineq)
3102 goto error;
3103 if (any_ineq(info_i, STATUS_ERROR))
3104 goto error;
3105 if (any_ineq(info_i, STATUS_SEPARATE))
3106 goto done;
3108 if (all(info_i->eq, 2 * bmap->n_eq, STATUS_VALID) &&
3109 all(info_i->ineq, bmap->n_ineq, STATUS_VALID)) {
3110 drop(&info[j]);
3111 change = isl_change_drop_second;
3114 if (change == isl_change_none && i != -1)
3115 return coalesce_expand_tab_divs(bmap, i, j, info, div, exp);
3117 done:
3118 isl_basic_map_free(bmap);
3119 clear_status(info_i);
3120 return change;
3121 error:
3122 isl_basic_map_free(bmap);
3123 clear_status(info_i);
3124 return isl_change_error;
3127 /* Check if the union of "bmap_i" and the basic map represented by info[j]
3128 * can be represented by a single basic map,
3129 * after aligning the divs of "bmap_i" to match those of info[j].
3130 * If so, replace the pair by the single basic map and return
3131 * isl_change_drop_first, isl_change_drop_second or isl_change_fuse.
3132 * Otherwise, return isl_change_none.
3134 * In particular, check if "bmap_i" contains the basic map represented by
3135 * info[j] after aligning the divs of "bmap_i" to those of info[j].
3136 * Note that this can only succeed if the number of divs of "bmap_i"
3137 * is smaller than (or equal to) the number of divs of info[j].
3139 * We first check if the divs of "bmap_i" are all known and form a subset
3140 * of those of info[j].bmap. If so, we pass control over to
3141 * coalesce_with_expanded_divs.
3143 * If "i" is not equal to -1, then "bmap" is equal to info[i].bmap.
3145 static enum isl_change coalesce_after_aligning_divs(
3146 __isl_keep isl_basic_map *bmap_i, int i, int j,
3147 struct isl_coalesce_info *info)
3149 isl_bool known;
3150 isl_mat *div_i, *div_j, *div;
3151 int *exp1 = NULL;
3152 int *exp2 = NULL;
3153 isl_ctx *ctx;
3154 enum isl_change change;
3156 known = isl_basic_map_divs_known(bmap_i);
3157 if (known < 0)
3158 return isl_change_error;
3159 if (!known)
3160 return isl_change_none;
3162 ctx = isl_basic_map_get_ctx(bmap_i);
3164 div_i = isl_basic_map_get_divs(bmap_i);
3165 div_j = isl_basic_map_get_divs(info[j].bmap);
3167 if (!div_i || !div_j)
3168 goto error;
3170 exp1 = isl_alloc_array(ctx, int, div_i->n_row);
3171 exp2 = isl_alloc_array(ctx, int, div_j->n_row);
3172 if ((div_i->n_row && !exp1) || (div_j->n_row && !exp2))
3173 goto error;
3175 div = isl_merge_divs(div_i, div_j, exp1, exp2);
3176 if (!div)
3177 goto error;
3179 if (div->n_row == div_j->n_row)
3180 change = coalesce_with_expanded_divs(bmap_i,
3181 i, j, info, div, exp1);
3182 else
3183 change = isl_change_none;
3185 isl_mat_free(div);
3187 isl_mat_free(div_i);
3188 isl_mat_free(div_j);
3190 free(exp2);
3191 free(exp1);
3193 return change;
3194 error:
3195 isl_mat_free(div_i);
3196 isl_mat_free(div_j);
3197 free(exp1);
3198 free(exp2);
3199 return isl_change_error;
3202 /* Check if basic map "j" is a subset of basic map "i" after
3203 * exploiting the extra equalities of "j" to simplify the divs of "i".
3204 * If so, remove basic map "j" and return isl_change_drop_second.
3206 * If "j" does not have any equalities or if they are the same
3207 * as those of "i", then we cannot exploit them to simplify the divs.
3208 * Similarly, if there are no divs in "i", then they cannot be simplified.
3209 * If, on the other hand, the affine hulls of "i" and "j" do not intersect,
3210 * then "j" cannot be a subset of "i".
3212 * Otherwise, we intersect "i" with the affine hull of "j" and then
3213 * check if "j" is a subset of the result after aligning the divs.
3214 * If so, then "j" is definitely a subset of "i" and can be removed.
3215 * Note that if after intersection with the affine hull of "j".
3216 * "i" still has more divs than "j", then there is no way we can
3217 * align the divs of "i" to those of "j".
3219 static enum isl_change coalesce_subset_with_equalities(int i, int j,
3220 struct isl_coalesce_info *info)
3222 isl_basic_map *hull_i, *hull_j, *bmap_i;
3223 int equal, empty;
3224 enum isl_change change;
3226 if (info[j].bmap->n_eq == 0)
3227 return isl_change_none;
3228 if (info[i].bmap->n_div == 0)
3229 return isl_change_none;
3231 hull_i = isl_basic_map_copy(info[i].bmap);
3232 hull_i = isl_basic_map_plain_affine_hull(hull_i);
3233 hull_j = isl_basic_map_copy(info[j].bmap);
3234 hull_j = isl_basic_map_plain_affine_hull(hull_j);
3236 hull_j = isl_basic_map_intersect(hull_j, isl_basic_map_copy(hull_i));
3237 equal = isl_basic_map_plain_is_equal(hull_i, hull_j);
3238 empty = isl_basic_map_plain_is_empty(hull_j);
3239 isl_basic_map_free(hull_i);
3241 if (equal < 0 || equal || empty < 0 || empty) {
3242 isl_basic_map_free(hull_j);
3243 if (equal < 0 || empty < 0)
3244 return isl_change_error;
3245 return isl_change_none;
3248 bmap_i = isl_basic_map_copy(info[i].bmap);
3249 bmap_i = isl_basic_map_intersect(bmap_i, hull_j);
3250 if (!bmap_i)
3251 return isl_change_error;
3253 if (bmap_i->n_div > info[j].bmap->n_div) {
3254 isl_basic_map_free(bmap_i);
3255 return isl_change_none;
3258 change = coalesce_after_aligning_divs(bmap_i, -1, j, info);
3260 isl_basic_map_free(bmap_i);
3262 return change;
3265 /* Check if the union of and the basic maps represented by info[i] and info[j]
3266 * can be represented by a single basic map, by aligning or equating
3267 * their integer divisions.
3268 * If so, replace the pair by the single basic map and return
3269 * isl_change_drop_first, isl_change_drop_second or isl_change_fuse.
3270 * Otherwise, return isl_change_none.
3272 * Note that we only perform any test if the number of divs is different
3273 * in the two basic maps. In case the number of divs is the same,
3274 * we have already established that the divs are different
3275 * in the two basic maps.
3276 * In particular, if the number of divs of basic map i is smaller than
3277 * the number of divs of basic map j, then we check if j is a subset of i
3278 * and vice versa.
3280 static enum isl_change coalesce_divs(int i, int j,
3281 struct isl_coalesce_info *info)
3283 enum isl_change change = isl_change_none;
3285 if (info[i].bmap->n_div < info[j].bmap->n_div)
3286 change = coalesce_after_aligning_divs(info[i].bmap, i, j, info);
3287 if (change != isl_change_none)
3288 return change;
3290 if (info[j].bmap->n_div < info[i].bmap->n_div)
3291 change = coalesce_after_aligning_divs(info[j].bmap, j, i, info);
3292 if (change != isl_change_none)
3293 return invert_change(change);
3295 change = coalesce_subset_with_equalities(i, j, info);
3296 if (change != isl_change_none)
3297 return change;
3299 change = coalesce_subset_with_equalities(j, i, info);
3300 if (change != isl_change_none)
3301 return invert_change(change);
3303 return isl_change_none;
3306 /* Does "bmap" involve any divs that themselves refer to divs?
3308 static isl_bool has_nested_div(__isl_keep isl_basic_map *bmap)
3310 int i;
3311 unsigned total;
3312 unsigned n_div;
3314 total = isl_basic_map_dim(bmap, isl_dim_all);
3315 n_div = isl_basic_map_dim(bmap, isl_dim_div);
3316 total -= n_div;
3318 for (i = 0; i < n_div; ++i)
3319 if (isl_seq_first_non_zero(bmap->div[i] + 2 + total,
3320 n_div) != -1)
3321 return isl_bool_true;
3323 return isl_bool_false;
3326 /* Return a list of affine expressions, one for each integer division
3327 * in "bmap_i". For each integer division that also appears in "bmap_j",
3328 * the affine expression is set to NaN. The number of NaNs in the list
3329 * is equal to the number of integer divisions in "bmap_j".
3330 * For the other integer divisions of "bmap_i", the corresponding
3331 * element in the list is a purely affine expression equal to the integer
3332 * division in "hull".
3333 * If no such list can be constructed, then the number of elements
3334 * in the returned list is smaller than the number of integer divisions
3335 * in "bmap_i".
3337 static __isl_give isl_aff_list *set_up_substitutions(
3338 __isl_keep isl_basic_map *bmap_i, __isl_keep isl_basic_map *bmap_j,
3339 __isl_take isl_basic_map *hull)
3341 unsigned n_div_i, n_div_j, total;
3342 isl_ctx *ctx;
3343 isl_local_space *ls;
3344 isl_basic_set *wrap_hull;
3345 isl_aff *aff_nan;
3346 isl_aff_list *list;
3347 int i, j;
3349 if (!hull)
3350 return NULL;
3352 ctx = isl_basic_map_get_ctx(hull);
3354 n_div_i = isl_basic_map_dim(bmap_i, isl_dim_div);
3355 n_div_j = isl_basic_map_dim(bmap_j, isl_dim_div);
3356 total = isl_basic_map_total_dim(bmap_i) - n_div_i;
3358 ls = isl_basic_map_get_local_space(bmap_i);
3359 ls = isl_local_space_wrap(ls);
3360 wrap_hull = isl_basic_map_wrap(hull);
3362 aff_nan = isl_aff_nan_on_domain(isl_local_space_copy(ls));
3363 list = isl_aff_list_alloc(ctx, n_div_i);
3365 j = 0;
3366 for (i = 0; i < n_div_i; ++i) {
3367 isl_aff *aff;
3369 if (j < n_div_j &&
3370 isl_basic_map_equal_div_expr_part(bmap_i, i, bmap_j, j,
3371 0, 2 + total)) {
3372 ++j;
3373 list = isl_aff_list_add(list, isl_aff_copy(aff_nan));
3374 continue;
3376 if (n_div_i - i <= n_div_j - j)
3377 break;
3379 aff = isl_local_space_get_div(ls, i);
3380 aff = isl_aff_substitute_equalities(aff,
3381 isl_basic_set_copy(wrap_hull));
3382 aff = isl_aff_floor(aff);
3383 if (!aff)
3384 goto error;
3385 if (isl_aff_dim(aff, isl_dim_div) != 0) {
3386 isl_aff_free(aff);
3387 break;
3390 list = isl_aff_list_add(list, aff);
3393 isl_aff_free(aff_nan);
3394 isl_local_space_free(ls);
3395 isl_basic_set_free(wrap_hull);
3397 return list;
3398 error:
3399 isl_aff_free(aff_nan);
3400 isl_local_space_free(ls);
3401 isl_basic_set_free(wrap_hull);
3402 isl_aff_list_free(list);
3403 return NULL;
3406 /* Add variables to info->bmap and info->tab corresponding to the elements
3407 * in "list" that are not set to NaN.
3408 * "extra_var" is the number of these elements.
3409 * "dim" is the offset in the variables of "tab" where we should
3410 * start considering the elements in "list".
3411 * When this function returns, the total number of variables in "tab"
3412 * is equal to "dim" plus the number of elements in "list".
3414 * The newly added existentially quantified variables are not given
3415 * an explicit representation because the corresponding div constraints
3416 * do not appear in info->bmap. These constraints are not added
3417 * to info->bmap because for internal consistency, they would need to
3418 * be added to info->tab as well, where they could combine with the equality
3419 * that is added later to result in constraints that do not hold
3420 * in the original input.
3422 static isl_stat add_sub_vars(struct isl_coalesce_info *info,
3423 __isl_keep isl_aff_list *list, int dim, int extra_var)
3425 int i, j, n, d;
3426 isl_space *space;
3428 space = isl_basic_map_get_space(info->bmap);
3429 info->bmap = isl_basic_map_cow(info->bmap);
3430 info->bmap = isl_basic_map_extend_space(info->bmap, space,
3431 extra_var, 0, 0);
3432 if (!info->bmap)
3433 return isl_stat_error;
3434 n = isl_aff_list_n_aff(list);
3435 for (i = 0; i < n; ++i) {
3436 int is_nan;
3437 isl_aff *aff;
3439 aff = isl_aff_list_get_aff(list, i);
3440 is_nan = isl_aff_is_nan(aff);
3441 isl_aff_free(aff);
3442 if (is_nan < 0)
3443 return isl_stat_error;
3444 if (is_nan)
3445 continue;
3447 if (isl_tab_insert_var(info->tab, dim + i) < 0)
3448 return isl_stat_error;
3449 d = isl_basic_map_alloc_div(info->bmap);
3450 if (d < 0)
3451 return isl_stat_error;
3452 info->bmap = isl_basic_map_mark_div_unknown(info->bmap, d);
3453 if (!info->bmap)
3454 return isl_stat_error;
3455 for (j = d; j > i; --j)
3456 isl_basic_map_swap_div(info->bmap, j - 1, j);
3459 return isl_stat_ok;
3462 /* For each element in "list" that is not set to NaN, fix the corresponding
3463 * variable in "tab" to the purely affine expression defined by the element.
3464 * "dim" is the offset in the variables of "tab" where we should
3465 * start considering the elements in "list".
3467 * This function assumes that a sufficient number of rows and
3468 * elements in the constraint array are available in the tableau.
3470 static int add_sub_equalities(struct isl_tab *tab,
3471 __isl_keep isl_aff_list *list, int dim)
3473 int i, n;
3474 isl_ctx *ctx;
3475 isl_vec *sub;
3476 isl_aff *aff;
3478 n = isl_aff_list_n_aff(list);
3480 ctx = isl_tab_get_ctx(tab);
3481 sub = isl_vec_alloc(ctx, 1 + dim + n);
3482 if (!sub)
3483 return -1;
3484 isl_seq_clr(sub->el + 1 + dim, n);
3486 for (i = 0; i < n; ++i) {
3487 aff = isl_aff_list_get_aff(list, i);
3488 if (!aff)
3489 goto error;
3490 if (isl_aff_is_nan(aff)) {
3491 isl_aff_free(aff);
3492 continue;
3494 isl_seq_cpy(sub->el, aff->v->el + 1, 1 + dim);
3495 isl_int_neg(sub->el[1 + dim + i], aff->v->el[0]);
3496 if (isl_tab_add_eq(tab, sub->el) < 0)
3497 goto error;
3498 isl_int_set_si(sub->el[1 + dim + i], 0);
3499 isl_aff_free(aff);
3502 isl_vec_free(sub);
3503 return 0;
3504 error:
3505 isl_aff_free(aff);
3506 isl_vec_free(sub);
3507 return -1;
3510 /* Add variables to info->tab and info->bmap corresponding to the elements
3511 * in "list" that are not set to NaN. The value of the added variable
3512 * in info->tab is fixed to the purely affine expression defined by the element.
3513 * "dim" is the offset in the variables of info->tab where we should
3514 * start considering the elements in "list".
3515 * When this function returns, the total number of variables in info->tab
3516 * is equal to "dim" plus the number of elements in "list".
3518 static int add_subs(struct isl_coalesce_info *info,
3519 __isl_keep isl_aff_list *list, int dim)
3521 int extra_var;
3522 int n;
3524 if (!list)
3525 return -1;
3527 n = isl_aff_list_n_aff(list);
3528 extra_var = n - (info->tab->n_var - dim);
3530 if (isl_tab_extend_vars(info->tab, extra_var) < 0)
3531 return -1;
3532 if (isl_tab_extend_cons(info->tab, 2 * extra_var) < 0)
3533 return -1;
3534 if (add_sub_vars(info, list, dim, extra_var) < 0)
3535 return -1;
3537 return add_sub_equalities(info->tab, list, dim);
3540 /* Coalesce basic map "j" into basic map "i" after adding the extra integer
3541 * divisions in "i" but not in "j" to basic map "j", with values
3542 * specified by "list". The total number of elements in "list"
3543 * is equal to the number of integer divisions in "i", while the number
3544 * of NaN elements in the list is equal to the number of integer divisions
3545 * in "j".
3547 * If no coalescing can be performed, then we need to revert basic map "j"
3548 * to its original state. We do the same if basic map "i" gets dropped
3549 * during the coalescing, even though this should not happen in practice
3550 * since we have already checked for "j" being a subset of "i"
3551 * before we reach this stage.
3553 static enum isl_change coalesce_with_subs(int i, int j,
3554 struct isl_coalesce_info *info, __isl_keep isl_aff_list *list)
3556 isl_basic_map *bmap_j;
3557 struct isl_tab_undo *snap;
3558 unsigned dim;
3559 enum isl_change change;
3561 bmap_j = isl_basic_map_copy(info[j].bmap);
3562 snap = isl_tab_snap(info[j].tab);
3564 dim = isl_basic_map_dim(bmap_j, isl_dim_all);
3565 dim -= isl_basic_map_dim(bmap_j, isl_dim_div);
3566 if (add_subs(&info[j], list, dim) < 0)
3567 goto error;
3569 change = coalesce_local_pair(i, j, info);
3570 if (change != isl_change_none && change != isl_change_drop_first) {
3571 isl_basic_map_free(bmap_j);
3572 } else {
3573 isl_basic_map_free(info[j].bmap);
3574 info[j].bmap = bmap_j;
3576 if (isl_tab_rollback(info[j].tab, snap) < 0)
3577 return isl_change_error;
3580 return change;
3581 error:
3582 isl_basic_map_free(bmap_j);
3583 return isl_change_error;
3586 /* Check if we can coalesce basic map "j" into basic map "i" after copying
3587 * those extra integer divisions in "i" that can be simplified away
3588 * using the extra equalities in "j".
3589 * All divs are assumed to be known and not contain any nested divs.
3591 * We first check if there are any extra equalities in "j" that we
3592 * can exploit. Then we check if every integer division in "i"
3593 * either already appears in "j" or can be simplified using the
3594 * extra equalities to a purely affine expression.
3595 * If these tests succeed, then we try to coalesce the two basic maps
3596 * by introducing extra dimensions in "j" corresponding to
3597 * the extra integer divsisions "i" fixed to the corresponding
3598 * purely affine expression.
3600 static enum isl_change check_coalesce_into_eq(int i, int j,
3601 struct isl_coalesce_info *info)
3603 unsigned n_div_i, n_div_j;
3604 isl_basic_map *hull_i, *hull_j;
3605 int equal, empty;
3606 isl_aff_list *list;
3607 enum isl_change change;
3609 n_div_i = isl_basic_map_dim(info[i].bmap, isl_dim_div);
3610 n_div_j = isl_basic_map_dim(info[j].bmap, isl_dim_div);
3611 if (n_div_i <= n_div_j)
3612 return isl_change_none;
3613 if (info[j].bmap->n_eq == 0)
3614 return isl_change_none;
3616 hull_i = isl_basic_map_copy(info[i].bmap);
3617 hull_i = isl_basic_map_plain_affine_hull(hull_i);
3618 hull_j = isl_basic_map_copy(info[j].bmap);
3619 hull_j = isl_basic_map_plain_affine_hull(hull_j);
3621 hull_j = isl_basic_map_intersect(hull_j, isl_basic_map_copy(hull_i));
3622 equal = isl_basic_map_plain_is_equal(hull_i, hull_j);
3623 empty = isl_basic_map_plain_is_empty(hull_j);
3624 isl_basic_map_free(hull_i);
3626 if (equal < 0 || empty < 0)
3627 goto error;
3628 if (equal || empty) {
3629 isl_basic_map_free(hull_j);
3630 return isl_change_none;
3633 list = set_up_substitutions(info[i].bmap, info[j].bmap, hull_j);
3634 if (!list)
3635 return isl_change_error;
3636 if (isl_aff_list_n_aff(list) < n_div_i)
3637 change = isl_change_none;
3638 else
3639 change = coalesce_with_subs(i, j, info, list);
3641 isl_aff_list_free(list);
3643 return change;
3644 error:
3645 isl_basic_map_free(hull_j);
3646 return isl_change_error;
3649 /* Check if we can coalesce basic maps "i" and "j" after copying
3650 * those extra integer divisions in one of the basic maps that can
3651 * be simplified away using the extra equalities in the other basic map.
3652 * We require all divs to be known in both basic maps.
3653 * Furthermore, to simplify the comparison of div expressions,
3654 * we do not allow any nested integer divisions.
3656 static enum isl_change check_coalesce_eq(int i, int j,
3657 struct isl_coalesce_info *info)
3659 isl_bool known, nested;
3660 enum isl_change change;
3662 known = isl_basic_map_divs_known(info[i].bmap);
3663 if (known < 0 || !known)
3664 return known < 0 ? isl_change_error : isl_change_none;
3665 known = isl_basic_map_divs_known(info[j].bmap);
3666 if (known < 0 || !known)
3667 return known < 0 ? isl_change_error : isl_change_none;
3668 nested = has_nested_div(info[i].bmap);
3669 if (nested < 0 || nested)
3670 return nested < 0 ? isl_change_error : isl_change_none;
3671 nested = has_nested_div(info[j].bmap);
3672 if (nested < 0 || nested)
3673 return nested < 0 ? isl_change_error : isl_change_none;
3675 change = check_coalesce_into_eq(i, j, info);
3676 if (change != isl_change_none)
3677 return change;
3678 change = check_coalesce_into_eq(j, i, info);
3679 if (change != isl_change_none)
3680 return invert_change(change);
3682 return isl_change_none;
3685 /* Check if the union of the given pair of basic maps
3686 * can be represented by a single basic map.
3687 * If so, replace the pair by the single basic map and return
3688 * isl_change_drop_first, isl_change_drop_second or isl_change_fuse.
3689 * Otherwise, return isl_change_none.
3691 * We first check if the two basic maps live in the same local space,
3692 * after aligning the divs that differ by only an integer constant.
3693 * If so, we do the complete check. Otherwise, we check if they have
3694 * the same number of integer divisions and can be coalesced, if one is
3695 * an obvious subset of the other or if the extra integer divisions
3696 * of one basic map can be simplified away using the extra equalities
3697 * of the other basic map.
3699 * Note that trying to coalesce pairs of disjuncts with the same
3700 * number, but different local variables may drop the explicit
3701 * representation of some of these local variables.
3702 * This operation is therefore not performed when
3703 * the "coalesce_preserve_locals" option is set.
3705 static enum isl_change coalesce_pair(int i, int j,
3706 struct isl_coalesce_info *info)
3708 int preserve;
3709 isl_bool same;
3710 enum isl_change change;
3711 isl_ctx *ctx;
3713 if (harmonize_divs(&info[i], &info[j]) < 0)
3714 return isl_change_error;
3715 same = same_divs(info[i].bmap, info[j].bmap);
3716 if (same < 0)
3717 return isl_change_error;
3718 if (same)
3719 return coalesce_local_pair(i, j, info);
3721 ctx = isl_basic_map_get_ctx(info[i].bmap);
3722 preserve = isl_options_get_coalesce_preserve_locals(ctx);
3723 if (!preserve && info[i].bmap->n_div == info[j].bmap->n_div) {
3724 change = coalesce_local_pair(i, j, info);
3725 if (change != isl_change_none)
3726 return change;
3729 change = coalesce_divs(i, j, info);
3730 if (change != isl_change_none)
3731 return change;
3733 return check_coalesce_eq(i, j, info);
3736 /* Return the maximum of "a" and "b".
3738 static int isl_max(int a, int b)
3740 return a > b ? a : b;
3743 /* Pairwise coalesce the basic maps in the range [start1, end1[ of "info"
3744 * with those in the range [start2, end2[, skipping basic maps
3745 * that have been removed (either before or within this function).
3747 * For each basic map i in the first range, we check if it can be coalesced
3748 * with respect to any previously considered basic map j in the second range.
3749 * If i gets dropped (because it was a subset of some j), then
3750 * we can move on to the next basic map.
3751 * If j gets dropped, we need to continue checking against the other
3752 * previously considered basic maps.
3753 * If the two basic maps got fused, then we recheck the fused basic map
3754 * against the previously considered basic maps, starting at i + 1
3755 * (even if start2 is greater than i + 1).
3757 static int coalesce_range(isl_ctx *ctx, struct isl_coalesce_info *info,
3758 int start1, int end1, int start2, int end2)
3760 int i, j;
3762 for (i = end1 - 1; i >= start1; --i) {
3763 if (info[i].removed)
3764 continue;
3765 for (j = isl_max(i + 1, start2); j < end2; ++j) {
3766 enum isl_change changed;
3768 if (info[j].removed)
3769 continue;
3770 if (info[i].removed)
3771 isl_die(ctx, isl_error_internal,
3772 "basic map unexpectedly removed",
3773 return -1);
3774 changed = coalesce_pair(i, j, info);
3775 switch (changed) {
3776 case isl_change_error:
3777 return -1;
3778 case isl_change_none:
3779 case isl_change_drop_second:
3780 continue;
3781 case isl_change_drop_first:
3782 j = end2;
3783 break;
3784 case isl_change_fuse:
3785 j = i;
3786 break;
3791 return 0;
3794 /* Pairwise coalesce the basic maps described by the "n" elements of "info".
3796 * We consider groups of basic maps that live in the same apparent
3797 * affine hull and we first coalesce within such a group before we
3798 * coalesce the elements in the group with elements of previously
3799 * considered groups. If a fuse happens during the second phase,
3800 * then we also reconsider the elements within the group.
3802 static int coalesce(isl_ctx *ctx, int n, struct isl_coalesce_info *info)
3804 int start, end;
3806 for (end = n; end > 0; end = start) {
3807 start = end - 1;
3808 while (start >= 1 &&
3809 info[start - 1].hull_hash == info[start].hull_hash)
3810 start--;
3811 if (coalesce_range(ctx, info, start, end, start, end) < 0)
3812 return -1;
3813 if (coalesce_range(ctx, info, start, end, end, n) < 0)
3814 return -1;
3817 return 0;
3820 /* Update the basic maps in "map" based on the information in "info".
3821 * In particular, remove the basic maps that have been marked removed and
3822 * update the others based on the information in the corresponding tableau.
3823 * Since we detected implicit equalities without calling
3824 * isl_basic_map_gauss, we need to do it now.
3825 * Also call isl_basic_map_simplify if we may have lost the definition
3826 * of one or more integer divisions.
3828 static __isl_give isl_map *update_basic_maps(__isl_take isl_map *map,
3829 int n, struct isl_coalesce_info *info)
3831 int i;
3833 if (!map)
3834 return NULL;
3836 for (i = n - 1; i >= 0; --i) {
3837 if (info[i].removed) {
3838 isl_basic_map_free(map->p[i]);
3839 if (i != map->n - 1)
3840 map->p[i] = map->p[map->n - 1];
3841 map->n--;
3842 continue;
3845 info[i].bmap = isl_basic_map_update_from_tab(info[i].bmap,
3846 info[i].tab);
3847 info[i].bmap = isl_basic_map_gauss(info[i].bmap, NULL);
3848 if (info[i].simplify)
3849 info[i].bmap = isl_basic_map_simplify(info[i].bmap);
3850 info[i].bmap = isl_basic_map_finalize(info[i].bmap);
3851 if (!info[i].bmap)
3852 return isl_map_free(map);
3853 ISL_F_SET(info[i].bmap, ISL_BASIC_MAP_NO_IMPLICIT);
3854 ISL_F_SET(info[i].bmap, ISL_BASIC_MAP_NO_REDUNDANT);
3855 isl_basic_map_free(map->p[i]);
3856 map->p[i] = info[i].bmap;
3857 info[i].bmap = NULL;
3860 return map;
3863 /* For each pair of basic maps in the map, check if the union of the two
3864 * can be represented by a single basic map.
3865 * If so, replace the pair by the single basic map and start over.
3867 * We factor out any (hidden) common factor from the constraint
3868 * coefficients to improve the detection of adjacent constraints.
3870 * Since we are constructing the tableaus of the basic maps anyway,
3871 * we exploit them to detect implicit equalities and redundant constraints.
3872 * This also helps the coalescing as it can ignore the redundant constraints.
3873 * In order to avoid confusion, we make all implicit equalities explicit
3874 * in the basic maps. We don't call isl_basic_map_gauss, though,
3875 * as that may affect the number of constraints.
3876 * This means that we have to call isl_basic_map_gauss at the end
3877 * of the computation (in update_basic_maps and in clear) to ensure that
3878 * the basic maps are not left in an unexpected state.
3879 * For each basic map, we also compute the hash of the apparent affine hull
3880 * for use in coalesce.
3882 __isl_give isl_map *isl_map_coalesce(__isl_take isl_map *map)
3884 int i;
3885 unsigned n;
3886 isl_ctx *ctx;
3887 struct isl_coalesce_info *info = NULL;
3889 map = isl_map_remove_empty_parts(map);
3890 if (!map)
3891 return NULL;
3893 if (map->n <= 1)
3894 return map;
3896 ctx = isl_map_get_ctx(map);
3897 map = isl_map_sort_divs(map);
3898 map = isl_map_cow(map);
3900 if (!map)
3901 return NULL;
3903 n = map->n;
3905 info = isl_calloc_array(map->ctx, struct isl_coalesce_info, n);
3906 if (!info)
3907 goto error;
3909 for (i = 0; i < map->n; ++i) {
3910 map->p[i] = isl_basic_map_reduce_coefficients(map->p[i]);
3911 if (!map->p[i])
3912 goto error;
3913 info[i].bmap = isl_basic_map_copy(map->p[i]);
3914 info[i].tab = isl_tab_from_basic_map(info[i].bmap, 0);
3915 if (!info[i].tab)
3916 goto error;
3917 if (!ISL_F_ISSET(info[i].bmap, ISL_BASIC_MAP_NO_IMPLICIT))
3918 if (isl_tab_detect_implicit_equalities(info[i].tab) < 0)
3919 goto error;
3920 info[i].bmap = isl_tab_make_equalities_explicit(info[i].tab,
3921 info[i].bmap);
3922 if (!info[i].bmap)
3923 goto error;
3924 if (!ISL_F_ISSET(info[i].bmap, ISL_BASIC_MAP_NO_REDUNDANT))
3925 if (isl_tab_detect_redundant(info[i].tab) < 0)
3926 goto error;
3927 if (coalesce_info_set_hull_hash(&info[i]) < 0)
3928 goto error;
3930 for (i = map->n - 1; i >= 0; --i)
3931 if (info[i].tab->empty)
3932 drop(&info[i]);
3934 if (coalesce(ctx, n, info) < 0)
3935 goto error;
3937 map = update_basic_maps(map, n, info);
3939 clear_coalesce_info(n, info);
3941 return map;
3942 error:
3943 clear_coalesce_info(n, info);
3944 isl_map_free(map);
3945 return NULL;
3948 /* For each pair of basic sets in the set, check if the union of the two
3949 * can be represented by a single basic set.
3950 * If so, replace the pair by the single basic set and start over.
3952 struct isl_set *isl_set_coalesce(struct isl_set *set)
3954 return set_from_map(isl_map_coalesce(set_to_map(set)));