isl_mat_right_kernel: add memory management annotations
[isl.git] / isl_coalesce.c
blob598eac23d7c1311bef95f0f504f9164e2646a8ec
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 /* Drop the basic map represented by "info".
344 * That is, clear the memory associated to the entry and
345 * mark it as having been removed.
347 static void drop(struct isl_coalesce_info *info)
349 info->bmap = isl_basic_map_free(info->bmap);
350 isl_tab_free(info->tab);
351 info->tab = NULL;
352 info->removed = 1;
355 /* Exchange the information in "info1" with that in "info2".
357 static void exchange(struct isl_coalesce_info *info1,
358 struct isl_coalesce_info *info2)
360 struct isl_coalesce_info info;
362 info = *info1;
363 *info1 = *info2;
364 *info2 = info;
367 /* This type represents the kind of change that has been performed
368 * while trying to coalesce two basic maps.
370 * isl_change_none: nothing was changed
371 * isl_change_drop_first: the first basic map was removed
372 * isl_change_drop_second: the second basic map was removed
373 * isl_change_fuse: the two basic maps were replaced by a new basic map.
375 enum isl_change {
376 isl_change_error = -1,
377 isl_change_none = 0,
378 isl_change_drop_first,
379 isl_change_drop_second,
380 isl_change_fuse,
383 /* Update "change" based on an interchange of the first and the second
384 * basic map. That is, interchange isl_change_drop_first and
385 * isl_change_drop_second.
387 static enum isl_change invert_change(enum isl_change change)
389 switch (change) {
390 case isl_change_error:
391 return isl_change_error;
392 case isl_change_none:
393 return isl_change_none;
394 case isl_change_drop_first:
395 return isl_change_drop_second;
396 case isl_change_drop_second:
397 return isl_change_drop_first;
398 case isl_change_fuse:
399 return isl_change_fuse;
402 return isl_change_error;
405 /* Add the valid constraints of the basic map represented by "info"
406 * to "bmap". "len" is the size of the constraints.
407 * If only one of the pair of inequalities that make up an equality
408 * is valid, then add that inequality.
410 static __isl_give isl_basic_map *add_valid_constraints(
411 __isl_take isl_basic_map *bmap, struct isl_coalesce_info *info,
412 unsigned len)
414 int k, l;
416 if (!bmap)
417 return NULL;
419 for (k = 0; k < info->bmap->n_eq; ++k) {
420 if (info->eq[2 * k] == STATUS_VALID &&
421 info->eq[2 * k + 1] == STATUS_VALID) {
422 l = isl_basic_map_alloc_equality(bmap);
423 if (l < 0)
424 return isl_basic_map_free(bmap);
425 isl_seq_cpy(bmap->eq[l], info->bmap->eq[k], len);
426 } else if (info->eq[2 * k] == STATUS_VALID) {
427 l = isl_basic_map_alloc_inequality(bmap);
428 if (l < 0)
429 return isl_basic_map_free(bmap);
430 isl_seq_neg(bmap->ineq[l], info->bmap->eq[k], len);
431 } else if (info->eq[2 * k + 1] == STATUS_VALID) {
432 l = isl_basic_map_alloc_inequality(bmap);
433 if (l < 0)
434 return isl_basic_map_free(bmap);
435 isl_seq_cpy(bmap->ineq[l], info->bmap->eq[k], len);
439 for (k = 0; k < info->bmap->n_ineq; ++k) {
440 if (info->ineq[k] != STATUS_VALID)
441 continue;
442 l = isl_basic_map_alloc_inequality(bmap);
443 if (l < 0)
444 return isl_basic_map_free(bmap);
445 isl_seq_cpy(bmap->ineq[l], info->bmap->ineq[k], len);
448 return bmap;
451 /* Is "bmap" defined by a number of (non-redundant) constraints that
452 * is greater than the number of constraints of basic maps i and j combined?
453 * Equalities are counted as two inequalities.
455 static int number_of_constraints_increases(int i, int j,
456 struct isl_coalesce_info *info,
457 __isl_keep isl_basic_map *bmap, struct isl_tab *tab)
459 int k, n_old, n_new;
461 n_old = 2 * info[i].bmap->n_eq + info[i].bmap->n_ineq;
462 n_old += 2 * info[j].bmap->n_eq + info[j].bmap->n_ineq;
464 n_new = 2 * bmap->n_eq;
465 for (k = 0; k < bmap->n_ineq; ++k)
466 if (!isl_tab_is_redundant(tab, bmap->n_eq + k))
467 ++n_new;
469 return n_new > n_old;
472 /* Replace the pair of basic maps i and j by the basic map bounded
473 * by the valid constraints in both basic maps and the constraints
474 * in extra (if not NULL).
475 * Place the fused basic map in the position that is the smallest of i and j.
477 * If "detect_equalities" is set, then look for equalities encoded
478 * as pairs of inequalities.
479 * If "check_number" is set, then the original basic maps are only
480 * replaced if the total number of constraints does not increase.
481 * While the number of integer divisions in the two basic maps
482 * is assumed to be the same, the actual definitions may be different.
483 * We only copy the definition from one of the basic map if it is
484 * the same as that of the other basic map. Otherwise, we mark
485 * the integer division as unknown and simplify the basic map
486 * in an attempt to recover the integer division definition.
488 static enum isl_change fuse(int i, int j, struct isl_coalesce_info *info,
489 __isl_keep isl_mat *extra, int detect_equalities, int check_number)
491 int k, l;
492 struct isl_basic_map *fused = NULL;
493 struct isl_tab *fused_tab = NULL;
494 unsigned total = isl_basic_map_total_dim(info[i].bmap);
495 unsigned extra_rows = extra ? extra->n_row : 0;
496 unsigned n_eq, n_ineq;
497 int simplify = 0;
499 if (j < i)
500 return fuse(j, i, info, extra, detect_equalities, check_number);
502 n_eq = info[i].bmap->n_eq + info[j].bmap->n_eq;
503 n_ineq = info[i].bmap->n_ineq + info[j].bmap->n_ineq;
504 fused = isl_basic_map_alloc_space(isl_space_copy(info[i].bmap->dim),
505 info[i].bmap->n_div, n_eq, n_eq + n_ineq + extra_rows);
506 fused = add_valid_constraints(fused, &info[i], 1 + total);
507 fused = add_valid_constraints(fused, &info[j], 1 + total);
508 if (!fused)
509 goto error;
510 if (ISL_F_ISSET(info[i].bmap, ISL_BASIC_MAP_RATIONAL) &&
511 ISL_F_ISSET(info[j].bmap, ISL_BASIC_MAP_RATIONAL))
512 ISL_F_SET(fused, ISL_BASIC_MAP_RATIONAL);
514 for (k = 0; k < info[i].bmap->n_div; ++k) {
515 int l = isl_basic_map_alloc_div(fused);
516 if (l < 0)
517 goto error;
518 if (isl_seq_eq(info[i].bmap->div[k], info[j].bmap->div[k],
519 1 + 1 + total)) {
520 isl_seq_cpy(fused->div[l], info[i].bmap->div[k],
521 1 + 1 + total);
522 } else {
523 isl_int_set_si(fused->div[l][0], 0);
524 simplify = 1;
528 for (k = 0; k < extra_rows; ++k) {
529 l = isl_basic_map_alloc_inequality(fused);
530 if (l < 0)
531 goto error;
532 isl_seq_cpy(fused->ineq[l], extra->row[k], 1 + total);
535 if (detect_equalities)
536 fused = isl_basic_map_detect_inequality_pairs(fused, NULL);
537 fused = isl_basic_map_gauss(fused, NULL);
538 if (simplify || info[j].simplify) {
539 fused = isl_basic_map_simplify(fused);
540 info[i].simplify = 0;
542 fused = isl_basic_map_finalize(fused);
544 fused_tab = isl_tab_from_basic_map(fused, 0);
545 if (isl_tab_detect_redundant(fused_tab) < 0)
546 goto error;
548 if (check_number &&
549 number_of_constraints_increases(i, j, info, fused, fused_tab)) {
550 isl_tab_free(fused_tab);
551 isl_basic_map_free(fused);
552 return isl_change_none;
555 isl_basic_map_free(info[i].bmap);
556 info[i].bmap = fused;
557 isl_tab_free(info[i].tab);
558 info[i].tab = fused_tab;
559 drop(&info[j]);
561 return isl_change_fuse;
562 error:
563 isl_tab_free(fused_tab);
564 isl_basic_map_free(fused);
565 return isl_change_error;
568 /* Given a pair of basic maps i and j such that all constraints are either
569 * "valid" or "cut", check if the facets corresponding to the "cut"
570 * constraints of i lie entirely within basic map j.
571 * If so, replace the pair by the basic map consisting of the valid
572 * constraints in both basic maps.
573 * Checking whether the facet lies entirely within basic map j
574 * is performed by checking whether the constraints of basic map j
575 * are valid for the facet. These tests are performed on a rational
576 * tableau to avoid the theoretical possibility that a constraint
577 * that was considered to be a cut constraint for the entire basic map i
578 * happens to be considered to be a valid constraint for the facet,
579 * even though it cuts off the same rational points.
581 * To see that we are not introducing any extra points, call the
582 * two basic maps A and B and the resulting map U and let x
583 * be an element of U \setminus ( A \cup B ).
584 * A line connecting x with an element of A \cup B meets a facet F
585 * of either A or B. Assume it is a facet of B and let c_1 be
586 * the corresponding facet constraint. We have c_1(x) < 0 and
587 * so c_1 is a cut constraint. This implies that there is some
588 * (possibly rational) point x' satisfying the constraints of A
589 * and the opposite of c_1 as otherwise c_1 would have been marked
590 * valid for A. The line connecting x and x' meets a facet of A
591 * in a (possibly rational) point that also violates c_1, but this
592 * is impossible since all cut constraints of B are valid for all
593 * cut facets of A.
594 * In case F is a facet of A rather than B, then we can apply the
595 * above reasoning to find a facet of B separating x from A \cup B first.
597 static enum isl_change check_facets(int i, int j,
598 struct isl_coalesce_info *info)
600 int k, l;
601 struct isl_tab_undo *snap, *snap2;
602 unsigned n_eq = info[i].bmap->n_eq;
604 snap = isl_tab_snap(info[i].tab);
605 if (isl_tab_mark_rational(info[i].tab) < 0)
606 return isl_change_error;
607 snap2 = isl_tab_snap(info[i].tab);
609 for (k = 0; k < info[i].bmap->n_ineq; ++k) {
610 if (info[i].ineq[k] != STATUS_CUT)
611 continue;
612 if (isl_tab_select_facet(info[i].tab, n_eq + k) < 0)
613 return isl_change_error;
614 for (l = 0; l < info[j].bmap->n_ineq; ++l) {
615 int stat;
616 if (info[j].ineq[l] != STATUS_CUT)
617 continue;
618 stat = status_in(info[j].bmap->ineq[l], info[i].tab);
619 if (stat < 0)
620 return isl_change_error;
621 if (stat != STATUS_VALID)
622 break;
624 if (isl_tab_rollback(info[i].tab, snap2) < 0)
625 return isl_change_error;
626 if (l < info[j].bmap->n_ineq)
627 break;
630 if (k < info[i].bmap->n_ineq) {
631 if (isl_tab_rollback(info[i].tab, snap) < 0)
632 return isl_change_error;
633 return isl_change_none;
635 return fuse(i, j, info, NULL, 0, 0);
638 /* Check if info->bmap contains the basic map represented
639 * by the tableau "tab".
640 * For each equality, we check both the constraint itself
641 * (as an inequality) and its negation. Make sure the
642 * equality is returned to its original state before returning.
644 static isl_bool contains(struct isl_coalesce_info *info, struct isl_tab *tab)
646 int k;
647 unsigned dim;
648 isl_basic_map *bmap = info->bmap;
650 dim = isl_basic_map_total_dim(bmap);
651 for (k = 0; k < bmap->n_eq; ++k) {
652 int stat;
653 isl_seq_neg(bmap->eq[k], bmap->eq[k], 1 + dim);
654 stat = status_in(bmap->eq[k], tab);
655 isl_seq_neg(bmap->eq[k], bmap->eq[k], 1 + dim);
656 if (stat < 0)
657 return isl_bool_error;
658 if (stat != STATUS_VALID)
659 return isl_bool_false;
660 stat = status_in(bmap->eq[k], tab);
661 if (stat < 0)
662 return isl_bool_error;
663 if (stat != STATUS_VALID)
664 return isl_bool_false;
667 for (k = 0; k < bmap->n_ineq; ++k) {
668 int stat;
669 if (info->ineq[k] == STATUS_REDUNDANT)
670 continue;
671 stat = status_in(bmap->ineq[k], tab);
672 if (stat < 0)
673 return isl_bool_error;
674 if (stat != STATUS_VALID)
675 return isl_bool_false;
677 return isl_bool_true;
680 /* Basic map "i" has an inequality (say "k") that is adjacent
681 * to some inequality of basic map "j". All the other inequalities
682 * are valid for "j".
683 * Check if basic map "j" forms an extension of basic map "i".
685 * Note that this function is only called if some of the equalities or
686 * inequalities of basic map "j" do cut basic map "i". The function is
687 * correct even if there are no such cut constraints, but in that case
688 * the additional checks performed by this function are overkill.
690 * In particular, we replace constraint k, say f >= 0, by constraint
691 * f <= -1, add the inequalities of "j" that are valid for "i"
692 * and check if the result is a subset of basic map "j".
693 * To improve the chances of the subset relation being detected,
694 * any variable that only attains a single integer value
695 * in the tableau of "i" is first fixed to that value.
696 * If the result is a subset, then we know that this result is exactly equal
697 * to basic map "j" since all its constraints are valid for basic map "j".
698 * By combining the valid constraints of "i" (all equalities and all
699 * inequalities except "k") and the valid constraints of "j" we therefore
700 * obtain a basic map that is equal to their union.
701 * In this case, there is no need to perform a rollback of the tableau
702 * since it is going to be destroyed in fuse().
705 * |\__ |\__
706 * | \__ | \__
707 * | \_ => | \__
708 * |_______| _ |_________\
711 * |\ |\
712 * | \ | \
713 * | \ | \
714 * | | | \
715 * | ||\ => | \
716 * | || \ | \
717 * | || | | |
718 * |__||_/ |_____/
720 static enum isl_change is_adj_ineq_extension(int i, int j,
721 struct isl_coalesce_info *info)
723 int k;
724 struct isl_tab_undo *snap;
725 unsigned n_eq = info[i].bmap->n_eq;
726 unsigned total = isl_basic_map_total_dim(info[i].bmap);
727 isl_stat r;
728 isl_bool super;
730 if (isl_tab_extend_cons(info[i].tab, 1 + info[j].bmap->n_ineq) < 0)
731 return isl_change_error;
733 k = find_ineq(&info[i], STATUS_ADJ_INEQ);
734 if (k < 0)
735 isl_die(isl_basic_map_get_ctx(info[i].bmap), isl_error_internal,
736 "info[i].ineq should have exactly one STATUS_ADJ_INEQ",
737 return isl_change_error);
739 snap = isl_tab_snap(info[i].tab);
741 if (isl_tab_unrestrict(info[i].tab, n_eq + k) < 0)
742 return isl_change_error;
744 isl_seq_neg(info[i].bmap->ineq[k], info[i].bmap->ineq[k], 1 + total);
745 isl_int_sub_ui(info[i].bmap->ineq[k][0], info[i].bmap->ineq[k][0], 1);
746 r = isl_tab_add_ineq(info[i].tab, info[i].bmap->ineq[k]);
747 isl_seq_neg(info[i].bmap->ineq[k], info[i].bmap->ineq[k], 1 + total);
748 isl_int_sub_ui(info[i].bmap->ineq[k][0], info[i].bmap->ineq[k][0], 1);
749 if (r < 0)
750 return isl_change_error;
752 for (k = 0; k < info[j].bmap->n_ineq; ++k) {
753 if (info[j].ineq[k] != STATUS_VALID)
754 continue;
755 if (isl_tab_add_ineq(info[i].tab, info[j].bmap->ineq[k]) < 0)
756 return isl_change_error;
758 if (isl_tab_detect_constants(info[i].tab) < 0)
759 return isl_change_error;
761 super = contains(&info[j], info[i].tab);
762 if (super < 0)
763 return isl_change_error;
764 if (super)
765 return fuse(i, j, info, NULL, 0, 0);
767 if (isl_tab_rollback(info[i].tab, snap) < 0)
768 return isl_change_error;
770 return isl_change_none;
774 /* Both basic maps have at least one inequality with and adjacent
775 * (but opposite) inequality in the other basic map.
776 * Check that there are no cut constraints and that there is only
777 * a single pair of adjacent inequalities.
778 * If so, we can replace the pair by a single basic map described
779 * by all but the pair of adjacent inequalities.
780 * Any additional points introduced lie strictly between the two
781 * adjacent hyperplanes and can therefore be integral.
783 * ____ _____
784 * / ||\ / \
785 * / || \ / \
786 * \ || \ => \ \
787 * \ || / \ /
788 * \___||_/ \_____/
790 * The test for a single pair of adjancent inequalities is important
791 * for avoiding the combination of two basic maps like the following
793 * /|
794 * / |
795 * /__|
796 * _____
797 * | |
798 * | |
799 * |___|
801 * If there are some cut constraints on one side, then we may
802 * still be able to fuse the two basic maps, but we need to perform
803 * some additional checks in is_adj_ineq_extension.
805 static enum isl_change check_adj_ineq(int i, int j,
806 struct isl_coalesce_info *info)
808 int count_i, count_j;
809 int cut_i, cut_j;
811 count_i = count_ineq(&info[i], STATUS_ADJ_INEQ);
812 count_j = count_ineq(&info[j], STATUS_ADJ_INEQ);
814 if (count_i != 1 && count_j != 1)
815 return isl_change_none;
817 cut_i = any_eq(&info[i], STATUS_CUT) || any_ineq(&info[i], STATUS_CUT);
818 cut_j = any_eq(&info[j], STATUS_CUT) || any_ineq(&info[j], STATUS_CUT);
820 if (!cut_i && !cut_j && count_i == 1 && count_j == 1)
821 return fuse(i, j, info, NULL, 0, 0);
823 if (count_i == 1 && !cut_i)
824 return is_adj_ineq_extension(i, j, info);
826 if (count_j == 1 && !cut_j)
827 return is_adj_ineq_extension(j, i, info);
829 return isl_change_none;
832 /* Given an affine transformation matrix "T", does row "row" represent
833 * anything other than a unit vector (possibly shifted by a constant)
834 * that is not involved in any of the other rows?
836 * That is, if a constraint involves the variable corresponding to
837 * the row, then could its preimage by "T" have any coefficients
838 * that are different from those in the original constraint?
840 static int not_unique_unit_row(__isl_keep isl_mat *T, int row)
842 int i, j;
843 int len = T->n_col - 1;
845 i = isl_seq_first_non_zero(T->row[row] + 1, len);
846 if (i < 0)
847 return 1;
848 if (!isl_int_is_one(T->row[row][1 + i]) &&
849 !isl_int_is_negone(T->row[row][1 + i]))
850 return 1;
852 j = isl_seq_first_non_zero(T->row[row] + 1 + i + 1, len - (i + 1));
853 if (j >= 0)
854 return 1;
856 for (j = 1; j < T->n_row; ++j) {
857 if (j == row)
858 continue;
859 if (!isl_int_is_zero(T->row[j][1 + i]))
860 return 1;
863 return 0;
866 /* Does inequality constraint "ineq" of "bmap" involve any of
867 * the variables marked in "affected"?
868 * "total" is the total number of variables, i.e., the number
869 * of entries in "affected".
871 static isl_bool is_affected(__isl_keep isl_basic_map *bmap, int ineq,
872 int *affected, int total)
874 int i;
876 for (i = 0; i < total; ++i) {
877 if (!affected[i])
878 continue;
879 if (!isl_int_is_zero(bmap->ineq[ineq][1 + i]))
880 return isl_bool_true;
883 return isl_bool_false;
886 /* Given the compressed version of inequality constraint "ineq"
887 * of info->bmap in "v", check if the constraint can be tightened,
888 * where the compression is based on an equality constraint valid
889 * for info->tab.
890 * If so, add the tightened version of the inequality constraint
891 * to info->tab. "v" may be modified by this function.
893 * That is, if the compressed constraint is of the form
895 * m f() + c >= 0
897 * with 0 < c < m, then it is equivalent to
899 * f() >= 0
901 * This means that c can also be subtracted from the original,
902 * uncompressed constraint without affecting the integer points
903 * in info->tab. Add this tightened constraint as an extra row
904 * to info->tab to make this information explicitly available.
906 static __isl_give isl_vec *try_tightening(struct isl_coalesce_info *info,
907 int ineq, __isl_take isl_vec *v)
909 isl_ctx *ctx;
910 isl_stat r;
912 if (!v)
913 return NULL;
915 ctx = isl_vec_get_ctx(v);
916 isl_seq_gcd(v->el + 1, v->size - 1, &ctx->normalize_gcd);
917 if (isl_int_is_zero(ctx->normalize_gcd) ||
918 isl_int_is_one(ctx->normalize_gcd)) {
919 return v;
922 v = isl_vec_cow(v);
923 if (!v)
924 return NULL;
926 isl_int_fdiv_r(v->el[0], v->el[0], ctx->normalize_gcd);
927 if (isl_int_is_zero(v->el[0]))
928 return v;
930 if (isl_tab_extend_cons(info->tab, 1) < 0)
931 return isl_vec_free(v);
933 isl_int_sub(info->bmap->ineq[ineq][0],
934 info->bmap->ineq[ineq][0], v->el[0]);
935 r = isl_tab_add_ineq(info->tab, info->bmap->ineq[ineq]);
936 isl_int_add(info->bmap->ineq[ineq][0],
937 info->bmap->ineq[ineq][0], v->el[0]);
939 if (r < 0)
940 return isl_vec_free(v);
942 return v;
945 /* Tighten the (non-redundant) constraints on the facet represented
946 * by info->tab.
947 * In particular, on input, info->tab represents the result
948 * of relaxing the "n" inequality constraints of info->bmap in "relaxed"
949 * by one, i.e., replacing f_i >= 0 by f_i + 1 >= 0, and then
950 * replacing the one at index "l" by the corresponding equality,
951 * i.e., f_k + 1 = 0, with k = relaxed[l].
953 * Compute a variable compression from the equality constraint f_k + 1 = 0
954 * and use it to tighten the other constraints of info->bmap
955 * (that is, all constraints that have not been relaxed),
956 * updating info->tab (and leaving info->bmap untouched).
957 * The compression handles essentially two cases, one where a variable
958 * is assigned a fixed value and can therefore be eliminated, and one
959 * where one variable is a shifted multiple of some other variable and
960 * can therefore be replaced by that multiple.
961 * Gaussian elimination would also work for the first case, but for
962 * the second case, the effectiveness would depend on the order
963 * of the variables.
964 * After compression, some of the constraints may have coefficients
965 * with a common divisor. If this divisor does not divide the constant
966 * term, then the constraint can be tightened.
967 * The tightening is performed on the tableau info->tab by introducing
968 * extra (temporary) constraints.
970 * Only constraints that are possibly affected by the compression are
971 * considered. In particular, if the constraint only involves variables
972 * that are directly mapped to a distinct set of other variables, then
973 * no common divisor can be introduced and no tightening can occur.
975 * It is important to only consider the non-redundant constraints
976 * since the facet constraint has been relaxed prior to the call
977 * to this function, meaning that the constraints that were redundant
978 * prior to the relaxation may no longer be redundant.
979 * These constraints will be ignored in the fused result, so
980 * the fusion detection should not exploit them.
982 static isl_stat tighten_on_relaxed_facet(struct isl_coalesce_info *info,
983 int n, int *relaxed, int l)
985 unsigned total;
986 isl_ctx *ctx;
987 isl_vec *v = NULL;
988 isl_mat *T;
989 int i;
990 int k;
991 int *affected;
993 k = relaxed[l];
994 ctx = isl_basic_map_get_ctx(info->bmap);
995 total = isl_basic_map_total_dim(info->bmap);
996 isl_int_add_ui(info->bmap->ineq[k][0], info->bmap->ineq[k][0], 1);
997 T = isl_mat_sub_alloc6(ctx, info->bmap->ineq, k, 1, 0, 1 + total);
998 T = isl_mat_variable_compression(T, NULL);
999 isl_int_sub_ui(info->bmap->ineq[k][0], info->bmap->ineq[k][0], 1);
1000 if (!T)
1001 return isl_stat_error;
1002 if (T->n_col == 0) {
1003 isl_mat_free(T);
1004 return isl_stat_ok;
1007 affected = isl_alloc_array(ctx, int, total);
1008 if (!affected)
1009 goto error;
1011 for (i = 0; i < total; ++i)
1012 affected[i] = not_unique_unit_row(T, 1 + i);
1014 for (i = 0; i < info->bmap->n_ineq; ++i) {
1015 isl_bool handle;
1016 if (any(relaxed, n, i))
1017 continue;
1018 if (info->ineq[i] == STATUS_REDUNDANT)
1019 continue;
1020 handle = is_affected(info->bmap, i, affected, total);
1021 if (handle < 0)
1022 goto error;
1023 if (!handle)
1024 continue;
1025 v = isl_vec_alloc(ctx, 1 + total);
1026 if (!v)
1027 goto error;
1028 isl_seq_cpy(v->el, info->bmap->ineq[i], 1 + total);
1029 v = isl_vec_mat_product(v, isl_mat_copy(T));
1030 v = try_tightening(info, i, v);
1031 isl_vec_free(v);
1032 if (!v)
1033 goto error;
1036 isl_mat_free(T);
1037 free(affected);
1038 return isl_stat_ok;
1039 error:
1040 isl_mat_free(T);
1041 free(affected);
1042 return isl_stat_error;
1045 /* Replace the basic maps "i" and "j" by an extension of "i"
1046 * along the "n" inequality constraints in "relax" by one.
1047 * The tableau info[i].tab has already been extended.
1048 * Extend info[i].bmap accordingly by relaxing all constraints in "relax"
1049 * by one.
1050 * Each integer division that does not have exactly the same
1051 * definition in "i" and "j" is marked unknown and the basic map
1052 * is scheduled to be simplified in an attempt to recover
1053 * the integer division definition.
1054 * Place the extension in the position that is the smallest of i and j.
1056 static enum isl_change extend(int i, int j, int n, int *relax,
1057 struct isl_coalesce_info *info)
1059 int l;
1060 unsigned total;
1062 info[i].bmap = isl_basic_map_cow(info[i].bmap);
1063 if (!info[i].bmap)
1064 return isl_change_error;
1065 total = isl_basic_map_total_dim(info[i].bmap);
1066 for (l = 0; l < info[i].bmap->n_div; ++l)
1067 if (!isl_seq_eq(info[i].bmap->div[l],
1068 info[j].bmap->div[l], 1 + 1 + total)) {
1069 isl_int_set_si(info[i].bmap->div[l][0], 0);
1070 info[i].simplify = 1;
1072 for (l = 0; l < n; ++l)
1073 isl_int_add_ui(info[i].bmap->ineq[relax[l]][0],
1074 info[i].bmap->ineq[relax[l]][0], 1);
1075 ISL_F_SET(info[i].bmap, ISL_BASIC_MAP_FINAL);
1076 drop(&info[j]);
1077 if (j < i)
1078 exchange(&info[i], &info[j]);
1079 return isl_change_fuse;
1082 /* Basic map "i" has "n" inequality constraints (collected in "relax")
1083 * that are such that they include basic map "j" if they are relaxed
1084 * by one. All the other inequalities are valid for "j".
1085 * Check if basic map "j" forms an extension of basic map "i".
1087 * In particular, relax the constraints in "relax", compute the corresponding
1088 * facets one by one and check whether each of these is included
1089 * in the other basic map.
1090 * Before testing for inclusion, the constraints on each facet
1091 * are tightened to increase the chance of an inclusion being detected.
1092 * (Adding the valid constraints of "j" to the tableau of "i", as is done
1093 * in is_adj_ineq_extension, may further increase those chances, but this
1094 * is not currently done.)
1095 * If each facet is included, we know that relaxing the constraints extends
1096 * the basic map with exactly the other basic map (we already know that this
1097 * other basic map is included in the extension, because all other
1098 * inequality constraints are valid of "j") and we can replace the
1099 * two basic maps by this extension.
1101 * If any of the relaxed constraints turn out to be redundant, then bail out.
1102 * isl_tab_select_facet refuses to handle such constraints. It may be
1103 * possible to handle them anyway by making a distinction between
1104 * redundant constraints with a corresponding facet that still intersects
1105 * the set (allowing isl_tab_select_facet to handle them) and
1106 * those where the facet does not intersect the set (which can be ignored
1107 * because the empty facet is trivially included in the other disjunct).
1108 * However, relaxed constraints that turn out to be redundant should
1109 * be fairly rare and no such instance has been reported where
1110 * coalescing would be successful.
1111 * ____ _____
1112 * / || / |
1113 * / || / |
1114 * \ || => \ |
1115 * \ || \ |
1116 * \___|| \____|
1119 * \ |\
1120 * |\\ | \
1121 * | \\ | \
1122 * | | => | /
1123 * | / | /
1124 * |/ |/
1126 static enum isl_change is_relaxed_extension(int i, int j, int n, int *relax,
1127 struct isl_coalesce_info *info)
1129 int l;
1130 isl_bool super;
1131 struct isl_tab_undo *snap, *snap2;
1132 unsigned n_eq = info[i].bmap->n_eq;
1134 for (l = 0; l < n; ++l)
1135 if (isl_tab_is_equality(info[i].tab, n_eq + relax[l]))
1136 return isl_change_none;
1138 snap = isl_tab_snap(info[i].tab);
1139 for (l = 0; l < n; ++l)
1140 if (isl_tab_relax(info[i].tab, n_eq + relax[l]) < 0)
1141 return isl_change_error;
1142 for (l = 0; l < n; ++l) {
1143 if (!isl_tab_is_redundant(info[i].tab, n_eq + relax[l]))
1144 continue;
1145 if (isl_tab_rollback(info[i].tab, snap) < 0)
1146 return isl_change_error;
1147 return isl_change_none;
1149 snap2 = isl_tab_snap(info[i].tab);
1150 for (l = 0; l < n; ++l) {
1151 if (isl_tab_rollback(info[i].tab, snap2) < 0)
1152 return isl_change_error;
1153 if (isl_tab_select_facet(info[i].tab, n_eq + relax[l]) < 0)
1154 return isl_change_error;
1155 if (tighten_on_relaxed_facet(&info[i], n, relax, l) < 0)
1156 return isl_change_error;
1157 super = contains(&info[j], info[i].tab);
1158 if (super < 0)
1159 return isl_change_error;
1160 if (super)
1161 continue;
1162 if (isl_tab_rollback(info[i].tab, snap) < 0)
1163 return isl_change_error;
1164 return isl_change_none;
1167 if (isl_tab_rollback(info[i].tab, snap2) < 0)
1168 return isl_change_error;
1169 return extend(i, j, n, relax, info);
1172 /* Data structure that keeps track of the wrapping constraints
1173 * and of information to bound the coefficients of those constraints.
1175 * bound is set if we want to apply a bound on the coefficients
1176 * mat contains the wrapping constraints
1177 * max is the bound on the coefficients (if bound is set)
1179 struct isl_wraps {
1180 int bound;
1181 isl_mat *mat;
1182 isl_int max;
1185 /* Update wraps->max to be greater than or equal to the coefficients
1186 * in the equalities and inequalities of info->bmap that can be removed
1187 * if we end up applying wrapping.
1189 static isl_stat wraps_update_max(struct isl_wraps *wraps,
1190 struct isl_coalesce_info *info)
1192 int k;
1193 isl_int max_k;
1194 unsigned total = isl_basic_map_total_dim(info->bmap);
1196 isl_int_init(max_k);
1198 for (k = 0; k < info->bmap->n_eq; ++k) {
1199 if (info->eq[2 * k] == STATUS_VALID &&
1200 info->eq[2 * k + 1] == STATUS_VALID)
1201 continue;
1202 isl_seq_abs_max(info->bmap->eq[k] + 1, total, &max_k);
1203 if (isl_int_abs_gt(max_k, wraps->max))
1204 isl_int_set(wraps->max, max_k);
1207 for (k = 0; k < info->bmap->n_ineq; ++k) {
1208 if (info->ineq[k] == STATUS_VALID ||
1209 info->ineq[k] == STATUS_REDUNDANT)
1210 continue;
1211 isl_seq_abs_max(info->bmap->ineq[k] + 1, total, &max_k);
1212 if (isl_int_abs_gt(max_k, wraps->max))
1213 isl_int_set(wraps->max, max_k);
1216 isl_int_clear(max_k);
1218 return isl_stat_ok;
1221 /* Initialize the isl_wraps data structure.
1222 * If we want to bound the coefficients of the wrapping constraints,
1223 * we set wraps->max to the largest coefficient
1224 * in the equalities and inequalities that can be removed if we end up
1225 * applying wrapping.
1227 static isl_stat wraps_init(struct isl_wraps *wraps, __isl_take isl_mat *mat,
1228 struct isl_coalesce_info *info, int i, int j)
1230 isl_ctx *ctx;
1232 wraps->bound = 0;
1233 wraps->mat = mat;
1234 if (!mat)
1235 return isl_stat_error;
1236 ctx = isl_mat_get_ctx(mat);
1237 wraps->bound = isl_options_get_coalesce_bounded_wrapping(ctx);
1238 if (!wraps->bound)
1239 return isl_stat_ok;
1240 isl_int_init(wraps->max);
1241 isl_int_set_si(wraps->max, 0);
1242 if (wraps_update_max(wraps, &info[i]) < 0)
1243 return isl_stat_error;
1244 if (wraps_update_max(wraps, &info[j]) < 0)
1245 return isl_stat_error;
1247 return isl_stat_ok;
1250 /* Free the contents of the isl_wraps data structure.
1252 static void wraps_free(struct isl_wraps *wraps)
1254 isl_mat_free(wraps->mat);
1255 if (wraps->bound)
1256 isl_int_clear(wraps->max);
1259 /* Is the wrapping constraint in row "row" allowed?
1261 * If wraps->bound is set, we check that none of the coefficients
1262 * is greater than wraps->max.
1264 static int allow_wrap(struct isl_wraps *wraps, int row)
1266 int i;
1268 if (!wraps->bound)
1269 return 1;
1271 for (i = 1; i < wraps->mat->n_col; ++i)
1272 if (isl_int_abs_gt(wraps->mat->row[row][i], wraps->max))
1273 return 0;
1275 return 1;
1278 /* Wrap "ineq" (or its opposite if "negate" is set) around "bound"
1279 * to include "set" and add the result in position "w" of "wraps".
1280 * "len" is the total number of coefficients in "bound" and "ineq".
1281 * Return 1 on success, 0 on failure and -1 on error.
1282 * Wrapping can fail if the result of wrapping is equal to "bound"
1283 * or if we want to bound the sizes of the coefficients and
1284 * the wrapped constraint does not satisfy this bound.
1286 static int add_wrap(struct isl_wraps *wraps, int w, isl_int *bound,
1287 isl_int *ineq, unsigned len, __isl_keep isl_set *set, int negate)
1289 isl_seq_cpy(wraps->mat->row[w], bound, len);
1290 if (negate) {
1291 isl_seq_neg(wraps->mat->row[w + 1], ineq, len);
1292 ineq = wraps->mat->row[w + 1];
1294 if (!isl_set_wrap_facet(set, wraps->mat->row[w], ineq))
1295 return -1;
1296 if (isl_seq_eq(wraps->mat->row[w], bound, len))
1297 return 0;
1298 if (!allow_wrap(wraps, w))
1299 return 0;
1300 return 1;
1303 /* For each constraint in info->bmap that is not redundant (as determined
1304 * by info->tab) and that is not a valid constraint for the other basic map,
1305 * wrap the constraint around "bound" such that it includes the whole
1306 * set "set" and append the resulting constraint to "wraps".
1307 * Note that the constraints that are valid for the other basic map
1308 * will be added to the combined basic map by default, so there is
1309 * no need to wrap them.
1310 * The caller wrap_in_facets even relies on this function not wrapping
1311 * any constraints that are already valid.
1312 * "wraps" is assumed to have been pre-allocated to the appropriate size.
1313 * wraps->n_row is the number of actual wrapped constraints that have
1314 * been added.
1315 * If any of the wrapping problems results in a constraint that is
1316 * identical to "bound", then this means that "set" is unbounded in such
1317 * way that no wrapping is possible. If this happens then wraps->n_row
1318 * is reset to zero.
1319 * Similarly, if we want to bound the coefficients of the wrapping
1320 * constraints and a newly added wrapping constraint does not
1321 * satisfy the bound, then wraps->n_row is also reset to zero.
1323 static isl_stat add_wraps(struct isl_wraps *wraps,
1324 struct isl_coalesce_info *info, isl_int *bound, __isl_keep isl_set *set)
1326 int l, m;
1327 int w;
1328 int added;
1329 isl_basic_map *bmap = info->bmap;
1330 unsigned len = 1 + isl_basic_map_total_dim(bmap);
1332 w = wraps->mat->n_row;
1334 for (l = 0; l < bmap->n_ineq; ++l) {
1335 if (info->ineq[l] == STATUS_VALID ||
1336 info->ineq[l] == STATUS_REDUNDANT)
1337 continue;
1338 if (isl_seq_is_neg(bound, bmap->ineq[l], len))
1339 continue;
1340 if (isl_seq_eq(bound, bmap->ineq[l], len))
1341 continue;
1342 if (isl_tab_is_redundant(info->tab, bmap->n_eq + l))
1343 continue;
1345 added = add_wrap(wraps, w, bound, bmap->ineq[l], len, set, 0);
1346 if (added < 0)
1347 return isl_stat_error;
1348 if (!added)
1349 goto unbounded;
1350 ++w;
1352 for (l = 0; l < bmap->n_eq; ++l) {
1353 if (isl_seq_is_neg(bound, bmap->eq[l], len))
1354 continue;
1355 if (isl_seq_eq(bound, bmap->eq[l], len))
1356 continue;
1358 for (m = 0; m < 2; ++m) {
1359 if (info->eq[2 * l + m] == STATUS_VALID)
1360 continue;
1361 added = add_wrap(wraps, w, bound, bmap->eq[l], len,
1362 set, !m);
1363 if (added < 0)
1364 return isl_stat_error;
1365 if (!added)
1366 goto unbounded;
1367 ++w;
1371 wraps->mat->n_row = w;
1372 return isl_stat_ok;
1373 unbounded:
1374 wraps->mat->n_row = 0;
1375 return isl_stat_ok;
1378 /* Check if the constraints in "wraps" from "first" until the last
1379 * are all valid for the basic set represented by "tab".
1380 * If not, wraps->n_row is set to zero.
1382 static int check_wraps(__isl_keep isl_mat *wraps, int first,
1383 struct isl_tab *tab)
1385 int i;
1387 for (i = first; i < wraps->n_row; ++i) {
1388 enum isl_ineq_type type;
1389 type = isl_tab_ineq_type(tab, wraps->row[i]);
1390 if (type == isl_ineq_error)
1391 return -1;
1392 if (type == isl_ineq_redundant)
1393 continue;
1394 wraps->n_row = 0;
1395 return 0;
1398 return 0;
1401 /* Return a set that corresponds to the non-redundant constraints
1402 * (as recorded in tab) of bmap.
1404 * It's important to remove the redundant constraints as some
1405 * of the other constraints may have been modified after the
1406 * constraints were marked redundant.
1407 * In particular, a constraint may have been relaxed.
1408 * Redundant constraints are ignored when a constraint is relaxed
1409 * and should therefore continue to be ignored ever after.
1410 * Otherwise, the relaxation might be thwarted by some of
1411 * these constraints.
1413 * Update the underlying set to ensure that the dimension doesn't change.
1414 * Otherwise the integer divisions could get dropped if the tab
1415 * turns out to be empty.
1417 static __isl_give isl_set *set_from_updated_bmap(__isl_keep isl_basic_map *bmap,
1418 struct isl_tab *tab)
1420 isl_basic_set *bset;
1422 bmap = isl_basic_map_copy(bmap);
1423 bset = isl_basic_map_underlying_set(bmap);
1424 bset = isl_basic_set_cow(bset);
1425 bset = isl_basic_set_update_from_tab(bset, tab);
1426 return isl_set_from_basic_set(bset);
1429 /* Wrap the constraints of info->bmap that bound the facet defined
1430 * by inequality "k" around (the opposite of) this inequality to
1431 * include "set". "bound" may be used to store the negated inequality.
1432 * Since the wrapped constraints are not guaranteed to contain the whole
1433 * of info->bmap, we check them in check_wraps.
1434 * If any of the wrapped constraints turn out to be invalid, then
1435 * check_wraps will reset wrap->n_row to zero.
1437 static isl_stat add_wraps_around_facet(struct isl_wraps *wraps,
1438 struct isl_coalesce_info *info, int k, isl_int *bound,
1439 __isl_keep isl_set *set)
1441 struct isl_tab_undo *snap;
1442 int n;
1443 unsigned total = isl_basic_map_total_dim(info->bmap);
1445 snap = isl_tab_snap(info->tab);
1447 if (isl_tab_select_facet(info->tab, info->bmap->n_eq + k) < 0)
1448 return isl_stat_error;
1449 if (isl_tab_detect_redundant(info->tab) < 0)
1450 return isl_stat_error;
1452 isl_seq_neg(bound, info->bmap->ineq[k], 1 + total);
1454 n = wraps->mat->n_row;
1455 if (add_wraps(wraps, info, bound, set) < 0)
1456 return isl_stat_error;
1458 if (isl_tab_rollback(info->tab, snap) < 0)
1459 return isl_stat_error;
1460 if (check_wraps(wraps->mat, n, info->tab) < 0)
1461 return isl_stat_error;
1463 return isl_stat_ok;
1466 /* Given a basic set i with a constraint k that is adjacent to
1467 * basic set j, check if we can wrap
1468 * both the facet corresponding to k (if "wrap_facet" is set) and basic map j
1469 * (always) around their ridges to include the other set.
1470 * If so, replace the pair of basic sets by their union.
1472 * All constraints of i (except k) are assumed to be valid or
1473 * cut constraints for j.
1474 * Wrapping the cut constraints to include basic map j may result
1475 * in constraints that are no longer valid of basic map i
1476 * we have to check that the resulting wrapping constraints are valid for i.
1477 * If "wrap_facet" is not set, then all constraints of i (except k)
1478 * are assumed to be valid for j.
1479 * ____ _____
1480 * / | / \
1481 * / || / |
1482 * \ || => \ |
1483 * \ || \ |
1484 * \___|| \____|
1487 static enum isl_change can_wrap_in_facet(int i, int j, int k,
1488 struct isl_coalesce_info *info, int wrap_facet)
1490 enum isl_change change = isl_change_none;
1491 struct isl_wraps wraps;
1492 isl_ctx *ctx;
1493 isl_mat *mat;
1494 struct isl_set *set_i = NULL;
1495 struct isl_set *set_j = NULL;
1496 struct isl_vec *bound = NULL;
1497 unsigned total = isl_basic_map_total_dim(info[i].bmap);
1499 set_i = set_from_updated_bmap(info[i].bmap, info[i].tab);
1500 set_j = set_from_updated_bmap(info[j].bmap, info[j].tab);
1501 ctx = isl_basic_map_get_ctx(info[i].bmap);
1502 mat = isl_mat_alloc(ctx, 2 * (info[i].bmap->n_eq + info[j].bmap->n_eq) +
1503 info[i].bmap->n_ineq + info[j].bmap->n_ineq,
1504 1 + total);
1505 if (wraps_init(&wraps, mat, info, i, j) < 0)
1506 goto error;
1507 bound = isl_vec_alloc(ctx, 1 + total);
1508 if (!set_i || !set_j || !bound)
1509 goto error;
1511 isl_seq_cpy(bound->el, info[i].bmap->ineq[k], 1 + total);
1512 isl_int_add_ui(bound->el[0], bound->el[0], 1);
1514 isl_seq_cpy(wraps.mat->row[0], bound->el, 1 + total);
1515 wraps.mat->n_row = 1;
1517 if (add_wraps(&wraps, &info[j], bound->el, set_i) < 0)
1518 goto error;
1519 if (!wraps.mat->n_row)
1520 goto unbounded;
1522 if (wrap_facet) {
1523 if (add_wraps_around_facet(&wraps, &info[i], k,
1524 bound->el, set_j) < 0)
1525 goto error;
1526 if (!wraps.mat->n_row)
1527 goto unbounded;
1530 change = fuse(i, j, info, wraps.mat, 0, 0);
1532 unbounded:
1533 wraps_free(&wraps);
1535 isl_set_free(set_i);
1536 isl_set_free(set_j);
1538 isl_vec_free(bound);
1540 return change;
1541 error:
1542 wraps_free(&wraps);
1543 isl_vec_free(bound);
1544 isl_set_free(set_i);
1545 isl_set_free(set_j);
1546 return isl_change_error;
1549 /* Given a cut constraint t(x) >= 0 of basic map i, stored in row "w"
1550 * of wrap.mat, replace it by its relaxed version t(x) + 1 >= 0, and
1551 * add wrapping constraints to wrap.mat for all constraints
1552 * of basic map j that bound the part of basic map j that sticks out
1553 * of the cut constraint.
1554 * "set_i" is the underlying set of basic map i.
1555 * If any wrapping fails, then wraps->mat.n_row is reset to zero.
1557 * In particular, we first intersect basic map j with t(x) + 1 = 0.
1558 * If the result is empty, then t(x) >= 0 was actually a valid constraint
1559 * (with respect to the integer points), so we add t(x) >= 0 instead.
1560 * Otherwise, we wrap the constraints of basic map j that are not
1561 * redundant in this intersection and that are not already valid
1562 * for basic map i over basic map i.
1563 * Note that it is sufficient to wrap the constraints to include
1564 * basic map i, because we will only wrap the constraints that do
1565 * not include basic map i already. The wrapped constraint will
1566 * therefore be more relaxed compared to the original constraint.
1567 * Since the original constraint is valid for basic map j, so is
1568 * the wrapped constraint.
1570 static isl_stat wrap_in_facet(struct isl_wraps *wraps, int w,
1571 struct isl_coalesce_info *info_j, __isl_keep isl_set *set_i,
1572 struct isl_tab_undo *snap)
1574 isl_int_add_ui(wraps->mat->row[w][0], wraps->mat->row[w][0], 1);
1575 if (isl_tab_add_eq(info_j->tab, wraps->mat->row[w]) < 0)
1576 return isl_stat_error;
1577 if (isl_tab_detect_redundant(info_j->tab) < 0)
1578 return isl_stat_error;
1580 if (info_j->tab->empty)
1581 isl_int_sub_ui(wraps->mat->row[w][0], wraps->mat->row[w][0], 1);
1582 else if (add_wraps(wraps, info_j, wraps->mat->row[w], set_i) < 0)
1583 return isl_stat_error;
1585 if (isl_tab_rollback(info_j->tab, snap) < 0)
1586 return isl_stat_error;
1588 return isl_stat_ok;
1591 /* Given a pair of basic maps i and j such that j sticks out
1592 * of i at n cut constraints, each time by at most one,
1593 * try to compute wrapping constraints and replace the two
1594 * basic maps by a single basic map.
1595 * The other constraints of i are assumed to be valid for j.
1596 * "set_i" is the underlying set of basic map i.
1597 * "wraps" has been initialized to be of the right size.
1599 * For each cut constraint t(x) >= 0 of i, we add the relaxed version
1600 * t(x) + 1 >= 0, along with wrapping constraints for all constraints
1601 * of basic map j that bound the part of basic map j that sticks out
1602 * of the cut constraint.
1604 * If any wrapping fails, i.e., if we cannot wrap to touch
1605 * the union, then we give up.
1606 * Otherwise, the pair of basic maps is replaced by their union.
1608 static enum isl_change try_wrap_in_facets(int i, int j,
1609 struct isl_coalesce_info *info, struct isl_wraps *wraps,
1610 __isl_keep isl_set *set_i)
1612 int k, l, w;
1613 unsigned total;
1614 struct isl_tab_undo *snap;
1616 total = isl_basic_map_total_dim(info[i].bmap);
1618 snap = isl_tab_snap(info[j].tab);
1620 wraps->mat->n_row = 0;
1622 for (k = 0; k < info[i].bmap->n_eq; ++k) {
1623 for (l = 0; l < 2; ++l) {
1624 if (info[i].eq[2 * k + l] != STATUS_CUT)
1625 continue;
1626 w = wraps->mat->n_row++;
1627 if (l == 0)
1628 isl_seq_neg(wraps->mat->row[w],
1629 info[i].bmap->eq[k], 1 + total);
1630 else
1631 isl_seq_cpy(wraps->mat->row[w],
1632 info[i].bmap->eq[k], 1 + total);
1633 if (wrap_in_facet(wraps, w, &info[j], set_i, snap) < 0)
1634 return isl_change_error;
1636 if (!wraps->mat->n_row)
1637 return isl_change_none;
1641 for (k = 0; k < info[i].bmap->n_ineq; ++k) {
1642 if (info[i].ineq[k] != STATUS_CUT)
1643 continue;
1644 w = wraps->mat->n_row++;
1645 isl_seq_cpy(wraps->mat->row[w],
1646 info[i].bmap->ineq[k], 1 + total);
1647 if (wrap_in_facet(wraps, w, &info[j], set_i, snap) < 0)
1648 return isl_change_error;
1650 if (!wraps->mat->n_row)
1651 return isl_change_none;
1654 return fuse(i, j, info, wraps->mat, 0, 1);
1657 /* Given a pair of basic maps i and j such that j sticks out
1658 * of i at n cut constraints, each time by at most one,
1659 * try to compute wrapping constraints and replace the two
1660 * basic maps by a single basic map.
1661 * The other constraints of i are assumed to be valid for j.
1663 * The core computation is performed by try_wrap_in_facets.
1664 * This function simply extracts an underlying set representation
1665 * of basic map i and initializes the data structure for keeping
1666 * track of wrapping constraints.
1668 static enum isl_change wrap_in_facets(int i, int j, int n,
1669 struct isl_coalesce_info *info)
1671 enum isl_change change = isl_change_none;
1672 struct isl_wraps wraps;
1673 isl_ctx *ctx;
1674 isl_mat *mat;
1675 isl_set *set_i = NULL;
1676 unsigned total = isl_basic_map_total_dim(info[i].bmap);
1677 int max_wrap;
1679 if (isl_tab_extend_cons(info[j].tab, 1) < 0)
1680 return isl_change_error;
1682 max_wrap = 1 + 2 * info[j].bmap->n_eq + info[j].bmap->n_ineq;
1683 max_wrap *= n;
1685 set_i = set_from_updated_bmap(info[i].bmap, info[i].tab);
1686 ctx = isl_basic_map_get_ctx(info[i].bmap);
1687 mat = isl_mat_alloc(ctx, max_wrap, 1 + total);
1688 if (wraps_init(&wraps, mat, info, i, j) < 0)
1689 goto error;
1690 if (!set_i)
1691 goto error;
1693 change = try_wrap_in_facets(i, j, info, &wraps, set_i);
1695 wraps_free(&wraps);
1696 isl_set_free(set_i);
1698 return change;
1699 error:
1700 wraps_free(&wraps);
1701 isl_set_free(set_i);
1702 return isl_change_error;
1705 /* Return the effect of inequality "ineq" on the tableau "tab",
1706 * after relaxing the constant term of "ineq" by one.
1708 static enum isl_ineq_type type_of_relaxed(struct isl_tab *tab, isl_int *ineq)
1710 enum isl_ineq_type type;
1712 isl_int_add_ui(ineq[0], ineq[0], 1);
1713 type = isl_tab_ineq_type(tab, ineq);
1714 isl_int_sub_ui(ineq[0], ineq[0], 1);
1716 return type;
1719 /* Given two basic sets i and j,
1720 * check if relaxing all the cut constraints of i by one turns
1721 * them into valid constraint for j and check if we can wrap in
1722 * the bits that are sticking out.
1723 * If so, replace the pair by their union.
1725 * We first check if all relaxed cut inequalities of i are valid for j
1726 * and then try to wrap in the intersections of the relaxed cut inequalities
1727 * with j.
1729 * During this wrapping, we consider the points of j that lie at a distance
1730 * of exactly 1 from i. In particular, we ignore the points that lie in
1731 * between this lower-dimensional space and the basic map i.
1732 * We can therefore only apply this to integer maps.
1733 * ____ _____
1734 * / ___|_ / \
1735 * / | | / |
1736 * \ | | => \ |
1737 * \|____| \ |
1738 * \___| \____/
1740 * _____ ______
1741 * | ____|_ | \
1742 * | | | | |
1743 * | | | => | |
1744 * |_| | | |
1745 * |_____| \______|
1747 * _______
1748 * | |
1749 * | |\ |
1750 * | | \ |
1751 * | | \ |
1752 * | | \|
1753 * | | \
1754 * | |_____\
1755 * | |
1756 * |_______|
1758 * Wrapping can fail if the result of wrapping one of the facets
1759 * around its edges does not produce any new facet constraint.
1760 * In particular, this happens when we try to wrap in unbounded sets.
1762 * _______________________________________________________________________
1764 * | ___
1765 * | | |
1766 * |_| |_________________________________________________________________
1767 * |___|
1769 * The following is not an acceptable result of coalescing the above two
1770 * sets as it includes extra integer points.
1771 * _______________________________________________________________________
1773 * |
1774 * |
1776 * \______________________________________________________________________
1778 static enum isl_change can_wrap_in_set(int i, int j,
1779 struct isl_coalesce_info *info)
1781 int k, l;
1782 int n;
1783 unsigned total;
1785 if (ISL_F_ISSET(info[i].bmap, ISL_BASIC_MAP_RATIONAL) ||
1786 ISL_F_ISSET(info[j].bmap, ISL_BASIC_MAP_RATIONAL))
1787 return isl_change_none;
1789 n = count_eq(&info[i], STATUS_CUT) + count_ineq(&info[i], STATUS_CUT);
1790 if (n == 0)
1791 return isl_change_none;
1793 total = isl_basic_map_total_dim(info[i].bmap);
1794 for (k = 0; k < info[i].bmap->n_eq; ++k) {
1795 for (l = 0; l < 2; ++l) {
1796 enum isl_ineq_type type;
1798 if (info[i].eq[2 * k + l] != STATUS_CUT)
1799 continue;
1801 if (l == 0)
1802 isl_seq_neg(info[i].bmap->eq[k],
1803 info[i].bmap->eq[k], 1 + total);
1804 type = type_of_relaxed(info[j].tab,
1805 info[i].bmap->eq[k]);
1806 if (l == 0)
1807 isl_seq_neg(info[i].bmap->eq[k],
1808 info[i].bmap->eq[k], 1 + total);
1809 if (type == isl_ineq_error)
1810 return isl_change_error;
1811 if (type != isl_ineq_redundant)
1812 return isl_change_none;
1816 for (k = 0; k < info[i].bmap->n_ineq; ++k) {
1817 enum isl_ineq_type type;
1819 if (info[i].ineq[k] != STATUS_CUT)
1820 continue;
1822 type = type_of_relaxed(info[j].tab, info[i].bmap->ineq[k]);
1823 if (type == isl_ineq_error)
1824 return isl_change_error;
1825 if (type != isl_ineq_redundant)
1826 return isl_change_none;
1829 return wrap_in_facets(i, j, n, info);
1832 /* Check if either i or j has only cut constraints that can
1833 * be used to wrap in (a facet of) the other basic set.
1834 * if so, replace the pair by their union.
1836 static enum isl_change check_wrap(int i, int j, struct isl_coalesce_info *info)
1838 enum isl_change change = isl_change_none;
1840 change = can_wrap_in_set(i, j, info);
1841 if (change != isl_change_none)
1842 return change;
1844 change = can_wrap_in_set(j, i, info);
1845 return change;
1848 /* Check if all inequality constraints of "i" that cut "j" cease
1849 * to be cut constraints if they are relaxed by one.
1850 * If so, collect the cut constraints in "list".
1851 * The caller is responsible for allocating "list".
1853 static isl_bool all_cut_by_one(int i, int j, struct isl_coalesce_info *info,
1854 int *list)
1856 int l, n;
1858 n = 0;
1859 for (l = 0; l < info[i].bmap->n_ineq; ++l) {
1860 enum isl_ineq_type type;
1862 if (info[i].ineq[l] != STATUS_CUT)
1863 continue;
1864 type = type_of_relaxed(info[j].tab, info[i].bmap->ineq[l]);
1865 if (type == isl_ineq_error)
1866 return isl_bool_error;
1867 if (type != isl_ineq_redundant)
1868 return isl_bool_false;
1869 list[n++] = l;
1872 return isl_bool_true;
1875 /* Given two basic maps such that "j" has at least one equality constraint
1876 * that is adjacent to an inequality constraint of "i" and such that "i" has
1877 * exactly one inequality constraint that is adjacent to an equality
1878 * constraint of "j", check whether "i" can be extended to include "j" or
1879 * whether "j" can be wrapped into "i".
1880 * All remaining constraints of "i" and "j" are assumed to be valid
1881 * or cut constraints of the other basic map.
1882 * However, none of the equality constraints of "i" are cut constraints.
1884 * If "i" has any "cut" inequality constraints, then check if relaxing
1885 * each of them by one is sufficient for them to become valid.
1886 * If so, check if the inequality constraint adjacent to an equality
1887 * constraint of "j" along with all these cut constraints
1888 * can be relaxed by one to contain exactly "j".
1889 * Otherwise, or if this fails, check if "j" can be wrapped into "i".
1891 static enum isl_change check_single_adj_eq(int i, int j,
1892 struct isl_coalesce_info *info)
1894 enum isl_change change = isl_change_none;
1895 int k;
1896 int n_cut;
1897 int *relax;
1898 isl_ctx *ctx;
1899 isl_bool try_relax;
1901 n_cut = count_ineq(&info[i], STATUS_CUT);
1903 k = find_ineq(&info[i], STATUS_ADJ_EQ);
1905 if (n_cut > 0) {
1906 ctx = isl_basic_map_get_ctx(info[i].bmap);
1907 relax = isl_calloc_array(ctx, int, 1 + n_cut);
1908 if (!relax)
1909 return isl_change_error;
1910 relax[0] = k;
1911 try_relax = all_cut_by_one(i, j, info, relax + 1);
1912 if (try_relax < 0)
1913 change = isl_change_error;
1914 } else {
1915 try_relax = isl_bool_true;
1916 relax = &k;
1918 if (try_relax && change == isl_change_none)
1919 change = is_relaxed_extension(i, j, 1 + n_cut, relax, info);
1920 if (n_cut > 0)
1921 free(relax);
1922 if (change != isl_change_none)
1923 return change;
1925 change = can_wrap_in_facet(i, j, k, info, n_cut > 0);
1927 return change;
1930 /* At least one of the basic maps has an equality that is adjacent
1931 * to an inequality. Make sure that only one of the basic maps has
1932 * such an equality and that the other basic map has exactly one
1933 * inequality adjacent to an equality.
1934 * If the other basic map does not have such an inequality, then
1935 * check if all its constraints are either valid or cut constraints
1936 * and, if so, try wrapping in the first map into the second.
1937 * Otherwise, try to extend one basic map with the other or
1938 * wrap one basic map in the other.
1940 static enum isl_change check_adj_eq(int i, int j,
1941 struct isl_coalesce_info *info)
1943 if (any_eq(&info[i], STATUS_ADJ_INEQ) &&
1944 any_eq(&info[j], STATUS_ADJ_INEQ))
1945 /* ADJ EQ TOO MANY */
1946 return isl_change_none;
1948 if (any_eq(&info[i], STATUS_ADJ_INEQ))
1949 return check_adj_eq(j, i, info);
1951 /* j has an equality adjacent to an inequality in i */
1953 if (count_ineq(&info[i], STATUS_ADJ_EQ) != 1) {
1954 if (all_valid_or_cut(&info[i]))
1955 return can_wrap_in_set(i, j, info);
1956 return isl_change_none;
1958 if (any_eq(&info[i], STATUS_CUT))
1959 return isl_change_none;
1960 if (any_ineq(&info[j], STATUS_ADJ_EQ) ||
1961 any_ineq(&info[i], STATUS_ADJ_INEQ) ||
1962 any_ineq(&info[j], STATUS_ADJ_INEQ))
1963 /* ADJ EQ TOO MANY */
1964 return isl_change_none;
1966 return check_single_adj_eq(i, j, info);
1969 /* Disjunct "j" lies on a hyperplane that is adjacent to disjunct "i".
1970 * In particular, disjunct "i" has an inequality constraint that is adjacent
1971 * to a (combination of) equality constraint(s) of disjunct "j",
1972 * but disjunct "j" has no explicit equality constraint adjacent
1973 * to an inequality constraint of disjunct "i".
1975 * Disjunct "i" is already known not to have any equality constraints
1976 * that are adjacent to an equality or inequality constraint.
1977 * Check that, other than the inequality constraint mentioned above,
1978 * all other constraints of disjunct "i" are valid for disjunct "j".
1979 * If so, try and wrap in disjunct "j".
1981 static enum isl_change check_ineq_adj_eq(int i, int j,
1982 struct isl_coalesce_info *info)
1984 int k;
1986 if (any_eq(&info[i], STATUS_CUT))
1987 return isl_change_none;
1988 if (any_ineq(&info[i], STATUS_CUT))
1989 return isl_change_none;
1990 if (any_ineq(&info[i], STATUS_ADJ_INEQ))
1991 return isl_change_none;
1992 if (count_ineq(&info[i], STATUS_ADJ_EQ) != 1)
1993 return isl_change_none;
1995 k = find_ineq(&info[i], STATUS_ADJ_EQ);
1997 return can_wrap_in_facet(i, j, k, info, 0);
2000 /* The two basic maps lie on adjacent hyperplanes. In particular,
2001 * basic map "i" has an equality that lies parallel to basic map "j".
2002 * Check if we can wrap the facets around the parallel hyperplanes
2003 * to include the other set.
2005 * We perform basically the same operations as can_wrap_in_facet,
2006 * except that we don't need to select a facet of one of the sets.
2008 * \\ \\
2009 * \\ => \\
2010 * \ \|
2012 * If there is more than one equality of "i" adjacent to an equality of "j",
2013 * then the result will satisfy one or more equalities that are a linear
2014 * combination of these equalities. These will be encoded as pairs
2015 * of inequalities in the wrapping constraints and need to be made
2016 * explicit.
2018 static enum isl_change check_eq_adj_eq(int i, int j,
2019 struct isl_coalesce_info *info)
2021 int k;
2022 enum isl_change change = isl_change_none;
2023 int detect_equalities = 0;
2024 struct isl_wraps wraps;
2025 isl_ctx *ctx;
2026 isl_mat *mat;
2027 struct isl_set *set_i = NULL;
2028 struct isl_set *set_j = NULL;
2029 struct isl_vec *bound = NULL;
2030 unsigned total = isl_basic_map_total_dim(info[i].bmap);
2032 if (count_eq(&info[i], STATUS_ADJ_EQ) != 1)
2033 detect_equalities = 1;
2035 k = find_eq(&info[i], STATUS_ADJ_EQ);
2037 set_i = set_from_updated_bmap(info[i].bmap, info[i].tab);
2038 set_j = set_from_updated_bmap(info[j].bmap, info[j].tab);
2039 ctx = isl_basic_map_get_ctx(info[i].bmap);
2040 mat = isl_mat_alloc(ctx, 2 * (info[i].bmap->n_eq + info[j].bmap->n_eq) +
2041 info[i].bmap->n_ineq + info[j].bmap->n_ineq,
2042 1 + total);
2043 if (wraps_init(&wraps, mat, info, i, j) < 0)
2044 goto error;
2045 bound = isl_vec_alloc(ctx, 1 + total);
2046 if (!set_i || !set_j || !bound)
2047 goto error;
2049 if (k % 2 == 0)
2050 isl_seq_neg(bound->el, info[i].bmap->eq[k / 2], 1 + total);
2051 else
2052 isl_seq_cpy(bound->el, info[i].bmap->eq[k / 2], 1 + total);
2053 isl_int_add_ui(bound->el[0], bound->el[0], 1);
2055 isl_seq_cpy(wraps.mat->row[0], bound->el, 1 + total);
2056 wraps.mat->n_row = 1;
2058 if (add_wraps(&wraps, &info[j], bound->el, set_i) < 0)
2059 goto error;
2060 if (!wraps.mat->n_row)
2061 goto unbounded;
2063 isl_int_sub_ui(bound->el[0], bound->el[0], 1);
2064 isl_seq_neg(bound->el, bound->el, 1 + total);
2066 isl_seq_cpy(wraps.mat->row[wraps.mat->n_row], bound->el, 1 + total);
2067 wraps.mat->n_row++;
2069 if (add_wraps(&wraps, &info[i], bound->el, set_j) < 0)
2070 goto error;
2071 if (!wraps.mat->n_row)
2072 goto unbounded;
2074 change = fuse(i, j, info, wraps.mat, detect_equalities, 0);
2076 if (0) {
2077 error: change = isl_change_error;
2079 unbounded:
2081 wraps_free(&wraps);
2082 isl_set_free(set_i);
2083 isl_set_free(set_j);
2084 isl_vec_free(bound);
2086 return change;
2089 /* Initialize the "eq" and "ineq" fields of "info".
2091 static void init_status(struct isl_coalesce_info *info)
2093 info->eq = info->ineq = NULL;
2096 /* Set info->eq to the positions of the equalities of info->bmap
2097 * with respect to the basic map represented by "tab".
2098 * If info->eq has already been computed, then do not compute it again.
2100 static void set_eq_status_in(struct isl_coalesce_info *info,
2101 struct isl_tab *tab)
2103 if (info->eq)
2104 return;
2105 info->eq = eq_status_in(info->bmap, tab);
2108 /* Set info->ineq to the positions of the inequalities of info->bmap
2109 * with respect to the basic map represented by "tab".
2110 * If info->ineq has already been computed, then do not compute it again.
2112 static void set_ineq_status_in(struct isl_coalesce_info *info,
2113 struct isl_tab *tab)
2115 if (info->ineq)
2116 return;
2117 info->ineq = ineq_status_in(info->bmap, info->tab, tab);
2120 /* Free the memory allocated by the "eq" and "ineq" fields of "info".
2121 * This function assumes that init_status has been called on "info" first,
2122 * after which the "eq" and "ineq" fields may or may not have been
2123 * assigned a newly allocated array.
2125 static void clear_status(struct isl_coalesce_info *info)
2127 free(info->eq);
2128 free(info->ineq);
2131 /* Are all inequality constraints of the basic map represented by "info"
2132 * valid for the other basic map, except for a single constraint
2133 * that is adjacent to an inequality constraint of the other basic map?
2135 static int all_ineq_valid_or_single_adj_ineq(struct isl_coalesce_info *info)
2137 int i;
2138 int k = -1;
2140 for (i = 0; i < info->bmap->n_ineq; ++i) {
2141 if (info->ineq[i] == STATUS_REDUNDANT)
2142 continue;
2143 if (info->ineq[i] == STATUS_VALID)
2144 continue;
2145 if (info->ineq[i] != STATUS_ADJ_INEQ)
2146 return 0;
2147 if (k != -1)
2148 return 0;
2149 k = i;
2152 return k != -1;
2155 /* Basic map "i" has one or more equality constraints that separate it
2156 * from basic map "j". Check if it happens to be an extension
2157 * of basic map "j".
2158 * In particular, check that all constraints of "j" are valid for "i",
2159 * except for one inequality constraint that is adjacent
2160 * to an inequality constraints of "i".
2161 * If so, check for "i" being an extension of "j" by calling
2162 * is_adj_ineq_extension.
2164 * Clean up the memory allocated for keeping track of the status
2165 * of the constraints before returning.
2167 static enum isl_change separating_equality(int i, int j,
2168 struct isl_coalesce_info *info)
2170 enum isl_change change = isl_change_none;
2172 if (all(info[j].eq, 2 * info[j].bmap->n_eq, STATUS_VALID) &&
2173 all_ineq_valid_or_single_adj_ineq(&info[j]))
2174 change = is_adj_ineq_extension(j, i, info);
2176 clear_status(&info[i]);
2177 clear_status(&info[j]);
2178 return change;
2181 /* Check if the union of the given pair of basic maps
2182 * can be represented by a single basic map.
2183 * If so, replace the pair by the single basic map and return
2184 * isl_change_drop_first, isl_change_drop_second or isl_change_fuse.
2185 * Otherwise, return isl_change_none.
2186 * The two basic maps are assumed to live in the same local space.
2187 * The "eq" and "ineq" fields of info[i] and info[j] are assumed
2188 * to have been initialized by the caller, either to NULL or
2189 * to valid information.
2191 * We first check the effect of each constraint of one basic map
2192 * on the other basic map.
2193 * The constraint may be
2194 * redundant the constraint is redundant in its own
2195 * basic map and should be ignore and removed
2196 * in the end
2197 * valid all (integer) points of the other basic map
2198 * satisfy the constraint
2199 * separate no (integer) point of the other basic map
2200 * satisfies the constraint
2201 * cut some but not all points of the other basic map
2202 * satisfy the constraint
2203 * adj_eq the given constraint is adjacent (on the outside)
2204 * to an equality of the other basic map
2205 * adj_ineq the given constraint is adjacent (on the outside)
2206 * to an inequality of the other basic map
2208 * We consider seven cases in which we can replace the pair by a single
2209 * basic map. We ignore all "redundant" constraints.
2211 * 1. all constraints of one basic map are valid
2212 * => the other basic map is a subset and can be removed
2214 * 2. all constraints of both basic maps are either "valid" or "cut"
2215 * and the facets corresponding to the "cut" constraints
2216 * of one of the basic maps lies entirely inside the other basic map
2217 * => the pair can be replaced by a basic map consisting
2218 * of the valid constraints in both basic maps
2220 * 3. there is a single pair of adjacent inequalities
2221 * (all other constraints are "valid")
2222 * => the pair can be replaced by a basic map consisting
2223 * of the valid constraints in both basic maps
2225 * 4. one basic map has a single adjacent inequality, while the other
2226 * constraints are "valid". The other basic map has some
2227 * "cut" constraints, but replacing the adjacent inequality by
2228 * its opposite and adding the valid constraints of the other
2229 * basic map results in a subset of the other basic map
2230 * => the pair can be replaced by a basic map consisting
2231 * of the valid constraints in both basic maps
2233 * 5. there is a single adjacent pair of an inequality and an equality,
2234 * the other constraints of the basic map containing the inequality are
2235 * "valid". Moreover, if the inequality the basic map is relaxed
2236 * and then turned into an equality, then resulting facet lies
2237 * entirely inside the other basic map
2238 * => the pair can be replaced by the basic map containing
2239 * the inequality, with the inequality relaxed.
2241 * 6. there is a single inequality adjacent to an equality,
2242 * the other constraints of the basic map containing the inequality are
2243 * "valid". Moreover, the facets corresponding to both
2244 * the inequality and the equality can be wrapped around their
2245 * ridges to include the other basic map
2246 * => the pair can be replaced by a basic map consisting
2247 * of the valid constraints in both basic maps together
2248 * with all wrapping constraints
2250 * 7. one of the basic maps extends beyond the other by at most one.
2251 * Moreover, the facets corresponding to the cut constraints and
2252 * the pieces of the other basic map at offset one from these cut
2253 * constraints can be wrapped around their ridges to include
2254 * the union of the two basic maps
2255 * => the pair can be replaced by a basic map consisting
2256 * of the valid constraints in both basic maps together
2257 * with all wrapping constraints
2259 * 8. the two basic maps live in adjacent hyperplanes. In principle
2260 * such sets can always be combined through wrapping, but we impose
2261 * that there is only one such pair, to avoid overeager coalescing.
2263 * Throughout the computation, we maintain a collection of tableaus
2264 * corresponding to the basic maps. When the basic maps are dropped
2265 * or combined, the tableaus are modified accordingly.
2267 static enum isl_change coalesce_local_pair_reuse(int i, int j,
2268 struct isl_coalesce_info *info)
2270 enum isl_change change = isl_change_none;
2272 set_ineq_status_in(&info[i], info[j].tab);
2273 if (info[i].bmap->n_ineq && !info[i].ineq)
2274 goto error;
2275 if (any_ineq(&info[i], STATUS_ERROR))
2276 goto error;
2277 if (any_ineq(&info[i], STATUS_SEPARATE))
2278 goto done;
2280 set_ineq_status_in(&info[j], info[i].tab);
2281 if (info[j].bmap->n_ineq && !info[j].ineq)
2282 goto error;
2283 if (any_ineq(&info[j], STATUS_ERROR))
2284 goto error;
2285 if (any_ineq(&info[j], STATUS_SEPARATE))
2286 goto done;
2288 set_eq_status_in(&info[i], info[j].tab);
2289 if (info[i].bmap->n_eq && !info[i].eq)
2290 goto error;
2291 if (any_eq(&info[i], STATUS_ERROR))
2292 goto error;
2294 set_eq_status_in(&info[j], info[i].tab);
2295 if (info[j].bmap->n_eq && !info[j].eq)
2296 goto error;
2297 if (any_eq(&info[j], STATUS_ERROR))
2298 goto error;
2300 if (any_eq(&info[i], STATUS_SEPARATE))
2301 return separating_equality(i, j, info);
2302 if (any_eq(&info[j], STATUS_SEPARATE))
2303 return separating_equality(j, i, info);
2305 if (all(info[i].eq, 2 * info[i].bmap->n_eq, STATUS_VALID) &&
2306 all(info[i].ineq, info[i].bmap->n_ineq, STATUS_VALID)) {
2307 drop(&info[j]);
2308 change = isl_change_drop_second;
2309 } else if (all(info[j].eq, 2 * info[j].bmap->n_eq, STATUS_VALID) &&
2310 all(info[j].ineq, info[j].bmap->n_ineq, STATUS_VALID)) {
2311 drop(&info[i]);
2312 change = isl_change_drop_first;
2313 } else if (any_eq(&info[i], STATUS_ADJ_EQ)) {
2314 change = check_eq_adj_eq(i, j, info);
2315 } else if (any_eq(&info[j], STATUS_ADJ_EQ)) {
2316 change = check_eq_adj_eq(j, i, info);
2317 } else if (any_eq(&info[i], STATUS_ADJ_INEQ) ||
2318 any_eq(&info[j], STATUS_ADJ_INEQ)) {
2319 change = check_adj_eq(i, j, info);
2320 } else if (any_ineq(&info[i], STATUS_ADJ_EQ)) {
2321 change = check_ineq_adj_eq(i, j, info);
2322 } else if (any_ineq(&info[j], STATUS_ADJ_EQ)) {
2323 change = check_ineq_adj_eq(j, i, info);
2324 } else if (any_ineq(&info[i], STATUS_ADJ_INEQ) ||
2325 any_ineq(&info[j], STATUS_ADJ_INEQ)) {
2326 change = check_adj_ineq(i, j, info);
2327 } else {
2328 if (!any_eq(&info[i], STATUS_CUT) &&
2329 !any_eq(&info[j], STATUS_CUT))
2330 change = check_facets(i, j, info);
2331 if (change == isl_change_none)
2332 change = check_wrap(i, j, info);
2335 done:
2336 clear_status(&info[i]);
2337 clear_status(&info[j]);
2338 return change;
2339 error:
2340 clear_status(&info[i]);
2341 clear_status(&info[j]);
2342 return isl_change_error;
2345 /* Check if the union of the given pair of basic maps
2346 * can be represented by a single basic map.
2347 * If so, replace the pair by the single basic map and return
2348 * isl_change_drop_first, isl_change_drop_second or isl_change_fuse.
2349 * Otherwise, return isl_change_none.
2350 * The two basic maps are assumed to live in the same local space.
2352 static enum isl_change coalesce_local_pair(int i, int j,
2353 struct isl_coalesce_info *info)
2355 init_status(&info[i]);
2356 init_status(&info[j]);
2357 return coalesce_local_pair_reuse(i, j, info);
2360 /* Shift the integer division at position "div" of the basic map
2361 * represented by "info" by "shift".
2363 * That is, if the integer division has the form
2365 * floor(f(x)/d)
2367 * then replace it by
2369 * floor((f(x) + shift * d)/d) - shift
2371 static isl_stat shift_div(struct isl_coalesce_info *info, int div,
2372 isl_int shift)
2374 unsigned total;
2376 info->bmap = isl_basic_map_shift_div(info->bmap, div, 0, shift);
2377 if (!info->bmap)
2378 return isl_stat_error;
2380 total = isl_basic_map_dim(info->bmap, isl_dim_all);
2381 total -= isl_basic_map_dim(info->bmap, isl_dim_div);
2382 if (isl_tab_shift_var(info->tab, total + div, shift) < 0)
2383 return isl_stat_error;
2385 return isl_stat_ok;
2388 /* If the integer division at position "div" is defined by an equality,
2389 * i.e., a stride constraint, then change the integer division expression
2390 * to have a constant term equal to zero.
2392 * Let the equality constraint be
2394 * c + f + m a = 0
2396 * The integer division expression is then of the form
2398 * a = floor((-f - c')/m)
2400 * The integer division is first shifted by t = floor(c/m),
2401 * turning the equality constraint into
2403 * c - m floor(c/m) + f + m a' = 0
2405 * i.e.,
2407 * (c mod m) + f + m a' = 0
2409 * That is,
2411 * a' = (-f - (c mod m))/m = floor((-f)/m)
2413 * because a' is an integer and 0 <= (c mod m) < m.
2414 * The constant term of a' can therefore be zeroed out.
2416 static isl_stat normalize_stride_div(struct isl_coalesce_info *info, int div)
2418 isl_bool defined;
2419 isl_stat r;
2420 isl_constraint *c;
2421 isl_int shift, stride;
2423 defined = isl_basic_map_has_defining_equality(info->bmap, isl_dim_div,
2424 div, &c);
2425 if (defined < 0)
2426 return isl_stat_error;
2427 if (!defined)
2428 return isl_stat_ok;
2429 if (!c)
2430 return isl_stat_error;
2431 isl_int_init(shift);
2432 isl_int_init(stride);
2433 isl_constraint_get_constant(c, &shift);
2434 isl_constraint_get_coefficient(c, isl_dim_div, div, &stride);
2435 isl_int_fdiv_q(shift, shift, stride);
2436 r = shift_div(info, div, shift);
2437 isl_int_clear(stride);
2438 isl_int_clear(shift);
2439 isl_constraint_free(c);
2440 if (r < 0)
2441 return isl_stat_error;
2442 info->bmap = isl_basic_map_set_div_expr_constant_num_si_inplace(
2443 info->bmap, div, 0);
2444 if (!info->bmap)
2445 return isl_stat_error;
2446 return isl_stat_ok;
2449 /* The basic maps represented by "info1" and "info2" are known
2450 * to have the same number of integer divisions.
2451 * Check if pairs of integer divisions are equal to each other
2452 * despite the fact that they differ by a rational constant.
2454 * In particular, look for any pair of integer divisions that
2455 * only differ in their constant terms.
2456 * If either of these integer divisions is defined
2457 * by stride constraints, then modify it to have a zero constant term.
2458 * If both are defined by stride constraints then in the end they will have
2459 * the same (zero) constant term.
2461 static isl_stat harmonize_stride_divs(struct isl_coalesce_info *info1,
2462 struct isl_coalesce_info *info2)
2464 int i, n;
2466 n = isl_basic_map_dim(info1->bmap, isl_dim_div);
2467 for (i = 0; i < n; ++i) {
2468 isl_bool known, harmonize;
2470 known = isl_basic_map_div_is_known(info1->bmap, i);
2471 if (known >= 0 && known)
2472 known = isl_basic_map_div_is_known(info2->bmap, i);
2473 if (known < 0)
2474 return isl_stat_error;
2475 if (!known)
2476 continue;
2477 harmonize = isl_basic_map_equal_div_expr_except_constant(
2478 info1->bmap, i, info2->bmap, i);
2479 if (harmonize < 0)
2480 return isl_stat_error;
2481 if (!harmonize)
2482 continue;
2483 if (normalize_stride_div(info1, i) < 0)
2484 return isl_stat_error;
2485 if (normalize_stride_div(info2, i) < 0)
2486 return isl_stat_error;
2489 return isl_stat_ok;
2492 /* If "shift" is an integer constant, then shift the integer division
2493 * at position "div" of the basic map represented by "info" by "shift".
2494 * If "shift" is not an integer constant, then do nothing.
2495 * If "shift" is equal to zero, then no shift needs to be performed either.
2497 * That is, if the integer division has the form
2499 * floor(f(x)/d)
2501 * then replace it by
2503 * floor((f(x) + shift * d)/d) - shift
2505 static isl_stat shift_if_cst_int(struct isl_coalesce_info *info, int div,
2506 __isl_keep isl_aff *shift)
2508 isl_bool cst;
2509 isl_stat r;
2510 isl_int d;
2511 isl_val *c;
2513 cst = isl_aff_is_cst(shift);
2514 if (cst < 0 || !cst)
2515 return cst < 0 ? isl_stat_error : isl_stat_ok;
2517 c = isl_aff_get_constant_val(shift);
2518 cst = isl_val_is_int(c);
2519 if (cst >= 0 && cst)
2520 cst = isl_bool_not(isl_val_is_zero(c));
2521 if (cst < 0 || !cst) {
2522 isl_val_free(c);
2523 return cst < 0 ? isl_stat_error : isl_stat_ok;
2526 isl_int_init(d);
2527 r = isl_val_get_num_isl_int(c, &d);
2528 if (r >= 0)
2529 r = shift_div(info, div, d);
2530 isl_int_clear(d);
2532 isl_val_free(c);
2534 return r;
2537 /* Check if some of the divs in the basic map represented by "info1"
2538 * are shifts of the corresponding divs in the basic map represented
2539 * by "info2", taking into account the equality constraints "eq1" of "info1"
2540 * and "eq2" of "info2". If so, align them with those of "info2".
2541 * "info1" and "info2" are assumed to have the same number
2542 * of integer divisions.
2544 * An integer division is considered to be a shift of another integer
2545 * division if, after simplification with respect to the equality
2546 * constraints of the other basic map, one is equal to the other
2547 * plus a constant.
2549 * In particular, for each pair of integer divisions, if both are known,
2550 * have the same denominator and are not already equal to each other,
2551 * simplify each with respect to the equality constraints
2552 * of the other basic map. If the difference is an integer constant,
2553 * then move this difference outside.
2554 * That is, if, after simplification, one integer division is of the form
2556 * floor((f(x) + c_1)/d)
2558 * while the other is of the form
2560 * floor((f(x) + c_2)/d)
2562 * and n = (c_2 - c_1)/d is an integer, then replace the first
2563 * integer division by
2565 * floor((f_1(x) + c_1 + n * d)/d) - n,
2567 * where floor((f_1(x) + c_1 + n * d)/d) = floor((f2(x) + c_2)/d)
2568 * after simplification with respect to the equality constraints.
2570 static isl_stat harmonize_divs_with_hulls(struct isl_coalesce_info *info1,
2571 struct isl_coalesce_info *info2, __isl_keep isl_basic_set *eq1,
2572 __isl_keep isl_basic_set *eq2)
2574 int i;
2575 int total;
2576 isl_local_space *ls1, *ls2;
2578 total = isl_basic_map_total_dim(info1->bmap);
2579 ls1 = isl_local_space_wrap(isl_basic_map_get_local_space(info1->bmap));
2580 ls2 = isl_local_space_wrap(isl_basic_map_get_local_space(info2->bmap));
2581 for (i = 0; i < info1->bmap->n_div; ++i) {
2582 isl_stat r;
2583 isl_aff *div1, *div2;
2585 if (!isl_local_space_div_is_known(ls1, i) ||
2586 !isl_local_space_div_is_known(ls2, i))
2587 continue;
2588 if (isl_int_ne(info1->bmap->div[i][0], info2->bmap->div[i][0]))
2589 continue;
2590 if (isl_seq_eq(info1->bmap->div[i] + 1,
2591 info2->bmap->div[i] + 1, 1 + total))
2592 continue;
2593 div1 = isl_local_space_get_div(ls1, i);
2594 div2 = isl_local_space_get_div(ls2, i);
2595 div1 = isl_aff_substitute_equalities(div1,
2596 isl_basic_set_copy(eq2));
2597 div2 = isl_aff_substitute_equalities(div2,
2598 isl_basic_set_copy(eq1));
2599 div2 = isl_aff_sub(div2, div1);
2600 r = shift_if_cst_int(info1, i, div2);
2601 isl_aff_free(div2);
2602 if (r < 0)
2603 break;
2605 isl_local_space_free(ls1);
2606 isl_local_space_free(ls2);
2608 if (i < info1->bmap->n_div)
2609 return isl_stat_error;
2610 return isl_stat_ok;
2613 /* Check if some of the divs in the basic map represented by "info1"
2614 * are shifts of the corresponding divs in the basic map represented
2615 * by "info2". If so, align them with those of "info2".
2616 * Only do this if "info1" and "info2" have the same number
2617 * of integer divisions.
2619 * An integer division is considered to be a shift of another integer
2620 * division if, after simplification with respect to the equality
2621 * constraints of the other basic map, one is equal to the other
2622 * plus a constant.
2624 * First check if pairs of integer divisions are equal to each other
2625 * despite the fact that they differ by a rational constant.
2626 * If so, try and arrange for them to have the same constant term.
2628 * Then, extract the equality constraints and continue with
2629 * harmonize_divs_with_hulls.
2631 * If the equality constraints of both basic maps are the same,
2632 * then there is no need to perform any shifting since
2633 * the coefficients of the integer divisions should have been
2634 * reduced in the same way.
2636 static isl_stat harmonize_divs(struct isl_coalesce_info *info1,
2637 struct isl_coalesce_info *info2)
2639 isl_bool equal;
2640 isl_basic_map *bmap1, *bmap2;
2641 isl_basic_set *eq1, *eq2;
2642 isl_stat r;
2644 if (!info1->bmap || !info2->bmap)
2645 return isl_stat_error;
2647 if (info1->bmap->n_div != info2->bmap->n_div)
2648 return isl_stat_ok;
2649 if (info1->bmap->n_div == 0)
2650 return isl_stat_ok;
2652 if (harmonize_stride_divs(info1, info2) < 0)
2653 return isl_stat_error;
2655 bmap1 = isl_basic_map_copy(info1->bmap);
2656 bmap2 = isl_basic_map_copy(info2->bmap);
2657 eq1 = isl_basic_map_wrap(isl_basic_map_plain_affine_hull(bmap1));
2658 eq2 = isl_basic_map_wrap(isl_basic_map_plain_affine_hull(bmap2));
2659 equal = isl_basic_set_plain_is_equal(eq1, eq2);
2660 if (equal < 0)
2661 r = isl_stat_error;
2662 else if (equal)
2663 r = isl_stat_ok;
2664 else
2665 r = harmonize_divs_with_hulls(info1, info2, eq1, eq2);
2666 isl_basic_set_free(eq1);
2667 isl_basic_set_free(eq2);
2669 return r;
2672 /* Do the two basic maps live in the same local space, i.e.,
2673 * do they have the same (known) divs?
2674 * If either basic map has any unknown divs, then we can only assume
2675 * that they do not live in the same local space.
2677 static isl_bool same_divs(__isl_keep isl_basic_map *bmap1,
2678 __isl_keep isl_basic_map *bmap2)
2680 int i;
2681 isl_bool known;
2682 int total;
2684 if (!bmap1 || !bmap2)
2685 return isl_bool_error;
2686 if (bmap1->n_div != bmap2->n_div)
2687 return isl_bool_false;
2689 if (bmap1->n_div == 0)
2690 return isl_bool_true;
2692 known = isl_basic_map_divs_known(bmap1);
2693 if (known < 0 || !known)
2694 return known;
2695 known = isl_basic_map_divs_known(bmap2);
2696 if (known < 0 || !known)
2697 return known;
2699 total = isl_basic_map_total_dim(bmap1);
2700 for (i = 0; i < bmap1->n_div; ++i)
2701 if (!isl_seq_eq(bmap1->div[i], bmap2->div[i], 2 + total))
2702 return 0;
2704 return 1;
2707 /* Assuming that "tab" contains the equality constraints and
2708 * the initial inequality constraints of "bmap", copy the remaining
2709 * inequality constraints of "bmap" to "Tab".
2711 static isl_stat copy_ineq(struct isl_tab *tab, __isl_keep isl_basic_map *bmap)
2713 int i, n_ineq;
2715 if (!bmap)
2716 return isl_stat_error;
2718 n_ineq = tab->n_con - tab->n_eq;
2719 for (i = n_ineq; i < bmap->n_ineq; ++i)
2720 if (isl_tab_add_ineq(tab, bmap->ineq[i]) < 0)
2721 return isl_stat_error;
2723 return isl_stat_ok;
2726 /* Description of an integer division that is added
2727 * during an expansion.
2728 * "pos" is the position of the corresponding variable.
2729 * "cst" indicates whether this integer division has a fixed value.
2730 * "val" contains the fixed value, if the value is fixed.
2732 struct isl_expanded {
2733 int pos;
2734 isl_bool cst;
2735 isl_int val;
2738 /* For each of the "n" integer division variables "expanded",
2739 * if the variable has a fixed value, then add two inequality
2740 * constraints expressing the fixed value.
2741 * Otherwise, add the corresponding div constraints.
2742 * The caller is responsible for removing the div constraints
2743 * that it added for all these "n" integer divisions.
2745 * The div constraints and the pair of inequality constraints
2746 * forcing the fixed value cannot both be added for a given variable
2747 * as the combination may render some of the original constraints redundant.
2748 * These would then be ignored during the coalescing detection,
2749 * while they could remain in the fused result.
2751 * The two added inequality constraints are
2753 * -a + v >= 0
2754 * a - v >= 0
2756 * with "a" the variable and "v" its fixed value.
2757 * The facet corresponding to one of these two constraints is selected
2758 * in the tableau to ensure that the pair of inequality constraints
2759 * is treated as an equality constraint.
2761 * The information in info->ineq is thrown away because it was
2762 * computed in terms of div constraints, while some of those
2763 * have now been replaced by these pairs of inequality constraints.
2765 static isl_stat fix_constant_divs(struct isl_coalesce_info *info,
2766 int n, struct isl_expanded *expanded)
2768 unsigned o_div;
2769 int i;
2770 isl_vec *ineq;
2772 o_div = isl_basic_map_offset(info->bmap, isl_dim_div) - 1;
2773 ineq = isl_vec_alloc(isl_tab_get_ctx(info->tab), 1 + info->tab->n_var);
2774 if (!ineq)
2775 return isl_stat_error;
2776 isl_seq_clr(ineq->el + 1, info->tab->n_var);
2778 for (i = 0; i < n; ++i) {
2779 if (!expanded[i].cst) {
2780 info->bmap = isl_basic_map_extend_constraints(
2781 info->bmap, 0, 2);
2782 if (isl_basic_map_add_div_constraints(info->bmap,
2783 expanded[i].pos - o_div) < 0)
2784 break;
2785 } else {
2786 isl_int_set_si(ineq->el[1 + expanded[i].pos], -1);
2787 isl_int_set(ineq->el[0], expanded[i].val);
2788 info->bmap = isl_basic_map_add_ineq(info->bmap,
2789 ineq->el);
2790 isl_int_set_si(ineq->el[1 + expanded[i].pos], 1);
2791 isl_int_neg(ineq->el[0], expanded[i].val);
2792 info->bmap = isl_basic_map_add_ineq(info->bmap,
2793 ineq->el);
2794 isl_int_set_si(ineq->el[1 + expanded[i].pos], 0);
2796 if (copy_ineq(info->tab, info->bmap) < 0)
2797 break;
2798 if (expanded[i].cst &&
2799 isl_tab_select_facet(info->tab, info->tab->n_con - 1) < 0)
2800 break;
2803 isl_vec_free(ineq);
2805 clear_status(info);
2806 init_status(info);
2808 return i < n ? isl_stat_error : isl_stat_ok;
2811 /* Insert the "n" integer division variables "expanded"
2812 * into info->tab and info->bmap and
2813 * update info->ineq with respect to the redundant constraints
2814 * in the resulting tableau.
2815 * "bmap" contains the result of this insertion in info->bmap,
2816 * while info->bmap is the original version
2817 * of "bmap", i.e., the one that corresponds to the current
2818 * state of info->tab. The number of constraints in info->bmap
2819 * is assumed to be the same as the number of constraints
2820 * in info->tab. This is required to be able to detect
2821 * the extra constraints in "bmap".
2823 * In particular, introduce extra variables corresponding
2824 * to the extra integer divisions and add the div constraints
2825 * that were added to "bmap" after info->tab was created
2826 * from info->bmap.
2827 * Furthermore, check if these extra integer divisions happen
2828 * to attain a fixed integer value in info->tab.
2829 * If so, replace the corresponding div constraints by pairs
2830 * of inequality constraints that fix these
2831 * integer divisions to their single integer values.
2832 * Replace info->bmap by "bmap" to match the changes to info->tab.
2833 * info->ineq was computed without a tableau and therefore
2834 * does not take into account the redundant constraints
2835 * in the tableau. Mark them here.
2836 * There is no need to check the newly added div constraints
2837 * since they cannot be redundant.
2838 * The redundancy check is not performed when constants have been discovered
2839 * since info->ineq is completely thrown away in this case.
2841 static isl_stat tab_insert_divs(struct isl_coalesce_info *info,
2842 int n, struct isl_expanded *expanded, __isl_take isl_basic_map *bmap)
2844 int i, n_ineq;
2845 unsigned n_eq;
2846 struct isl_tab_undo *snap;
2847 int any;
2849 if (!bmap)
2850 return isl_stat_error;
2851 if (info->bmap->n_eq + info->bmap->n_ineq != info->tab->n_con)
2852 isl_die(isl_basic_map_get_ctx(bmap), isl_error_internal,
2853 "original tableau does not correspond "
2854 "to original basic map", goto error);
2856 if (isl_tab_extend_vars(info->tab, n) < 0)
2857 goto error;
2858 if (isl_tab_extend_cons(info->tab, 2 * n) < 0)
2859 goto error;
2861 for (i = 0; i < n; ++i) {
2862 if (isl_tab_insert_var(info->tab, expanded[i].pos) < 0)
2863 goto error;
2866 snap = isl_tab_snap(info->tab);
2868 n_ineq = info->tab->n_con - info->tab->n_eq;
2869 if (copy_ineq(info->tab, bmap) < 0)
2870 goto error;
2872 isl_basic_map_free(info->bmap);
2873 info->bmap = bmap;
2875 any = 0;
2876 for (i = 0; i < n; ++i) {
2877 expanded[i].cst = isl_tab_is_constant(info->tab,
2878 expanded[i].pos, &expanded[i].val);
2879 if (expanded[i].cst < 0)
2880 return isl_stat_error;
2881 if (expanded[i].cst)
2882 any = 1;
2885 if (any) {
2886 if (isl_tab_rollback(info->tab, snap) < 0)
2887 return isl_stat_error;
2888 info->bmap = isl_basic_map_cow(info->bmap);
2889 if (isl_basic_map_free_inequality(info->bmap, 2 * n) < 0)
2890 return isl_stat_error;
2892 return fix_constant_divs(info, n, expanded);
2895 n_eq = info->bmap->n_eq;
2896 for (i = 0; i < n_ineq; ++i) {
2897 if (isl_tab_is_redundant(info->tab, n_eq + i))
2898 info->ineq[i] = STATUS_REDUNDANT;
2901 return isl_stat_ok;
2902 error:
2903 isl_basic_map_free(bmap);
2904 return isl_stat_error;
2907 /* Expand info->tab and info->bmap in the same way "bmap" was expanded
2908 * in isl_basic_map_expand_divs using the expansion "exp" and
2909 * update info->ineq with respect to the redundant constraints
2910 * in the resulting tableau. info->bmap is the original version
2911 * of "bmap", i.e., the one that corresponds to the current
2912 * state of info->tab. The number of constraints in info->bmap
2913 * is assumed to be the same as the number of constraints
2914 * in info->tab. This is required to be able to detect
2915 * the extra constraints in "bmap".
2917 * Extract the positions where extra local variables are introduced
2918 * from "exp" and call tab_insert_divs.
2920 static isl_stat expand_tab(struct isl_coalesce_info *info, int *exp,
2921 __isl_take isl_basic_map *bmap)
2923 isl_ctx *ctx;
2924 struct isl_expanded *expanded;
2925 int i, j, k, n;
2926 int extra_var;
2927 unsigned total, pos, n_div;
2928 isl_stat r;
2930 total = isl_basic_map_dim(bmap, isl_dim_all);
2931 n_div = isl_basic_map_dim(bmap, isl_dim_div);
2932 pos = total - n_div;
2933 extra_var = total - info->tab->n_var;
2934 n = n_div - extra_var;
2936 ctx = isl_basic_map_get_ctx(bmap);
2937 expanded = isl_calloc_array(ctx, struct isl_expanded, extra_var);
2938 if (extra_var && !expanded)
2939 goto error;
2941 i = 0;
2942 k = 0;
2943 for (j = 0; j < n_div; ++j) {
2944 if (i < n && exp[i] == j) {
2945 ++i;
2946 continue;
2948 expanded[k++].pos = pos + j;
2951 for (k = 0; k < extra_var; ++k)
2952 isl_int_init(expanded[k].val);
2954 r = tab_insert_divs(info, extra_var, expanded, bmap);
2956 for (k = 0; k < extra_var; ++k)
2957 isl_int_clear(expanded[k].val);
2958 free(expanded);
2960 return r;
2961 error:
2962 isl_basic_map_free(bmap);
2963 return isl_stat_error;
2966 /* Check if the union of the basic maps represented by info[i] and info[j]
2967 * can be represented by a single basic map,
2968 * after expanding the divs of info[i] to match those of info[j].
2969 * If so, replace the pair by the single basic map and return
2970 * isl_change_drop_first, isl_change_drop_second or isl_change_fuse.
2971 * Otherwise, return isl_change_none.
2973 * The caller has already checked for info[j] being a subset of info[i].
2974 * If some of the divs of info[j] are unknown, then the expanded info[i]
2975 * will not have the corresponding div constraints. The other patterns
2976 * therefore cannot apply. Skip the computation in this case.
2978 * The expansion is performed using the divs "div" and expansion "exp"
2979 * computed by the caller.
2980 * info[i].bmap has already been expanded and the result is passed in
2981 * as "bmap".
2982 * The "eq" and "ineq" fields of info[i] reflect the status of
2983 * the constraints of the expanded "bmap" with respect to info[j].tab.
2984 * However, inequality constraints that are redundant in info[i].tab
2985 * have not yet been marked as such because no tableau was available.
2987 * Replace info[i].bmap by "bmap" and expand info[i].tab as well,
2988 * updating info[i].ineq with respect to the redundant constraints.
2989 * Then try and coalesce the expanded info[i] with info[j],
2990 * reusing the information in info[i].eq and info[i].ineq.
2991 * If this does not result in any coalescing or if it results in info[j]
2992 * getting dropped (which should not happen in practice, since the case
2993 * of info[j] being a subset of info[i] has already been checked by
2994 * the caller), then revert info[i] to its original state.
2996 static enum isl_change coalesce_expand_tab_divs(__isl_take isl_basic_map *bmap,
2997 int i, int j, struct isl_coalesce_info *info, __isl_keep isl_mat *div,
2998 int *exp)
3000 isl_bool known;
3001 isl_basic_map *bmap_i;
3002 struct isl_tab_undo *snap;
3003 enum isl_change change = isl_change_none;
3005 known = isl_basic_map_divs_known(info[j].bmap);
3006 if (known < 0 || !known) {
3007 clear_status(&info[i]);
3008 isl_basic_map_free(bmap);
3009 return known < 0 ? isl_change_error : isl_change_none;
3012 bmap_i = isl_basic_map_copy(info[i].bmap);
3013 snap = isl_tab_snap(info[i].tab);
3014 if (expand_tab(&info[i], exp, bmap) < 0)
3015 change = isl_change_error;
3017 init_status(&info[j]);
3018 if (change == isl_change_none)
3019 change = coalesce_local_pair_reuse(i, j, info);
3020 else
3021 clear_status(&info[i]);
3022 if (change != isl_change_none && change != isl_change_drop_second) {
3023 isl_basic_map_free(bmap_i);
3024 } else {
3025 isl_basic_map_free(info[i].bmap);
3026 info[i].bmap = bmap_i;
3028 if (isl_tab_rollback(info[i].tab, snap) < 0)
3029 change = isl_change_error;
3032 return change;
3035 /* Check if the union of "bmap" and the basic map represented by info[j]
3036 * can be represented by a single basic map,
3037 * after expanding the divs of "bmap" to match those of info[j].
3038 * If so, replace the pair by the single basic map and return
3039 * isl_change_drop_first, isl_change_drop_second or isl_change_fuse.
3040 * Otherwise, return isl_change_none.
3042 * In particular, check if the expanded "bmap" contains the basic map
3043 * represented by the tableau info[j].tab.
3044 * The expansion is performed using the divs "div" and expansion "exp"
3045 * computed by the caller.
3046 * Then we check if all constraints of the expanded "bmap" are valid for
3047 * info[j].tab.
3049 * If "i" is not equal to -1, then "bmap" is equal to info[i].bmap.
3050 * In this case, the positions of the constraints of info[i].bmap
3051 * with respect to the basic map represented by info[j] are stored
3052 * in info[i].
3054 * If the expanded "bmap" does not contain the basic map
3055 * represented by the tableau info[j].tab and if "i" is not -1,
3056 * i.e., if the original "bmap" is info[i].bmap, then expand info[i].tab
3057 * as well and check if that results in coalescing.
3059 static enum isl_change coalesce_with_expanded_divs(
3060 __isl_keep isl_basic_map *bmap, int i, int j,
3061 struct isl_coalesce_info *info, __isl_keep isl_mat *div, int *exp)
3063 enum isl_change change = isl_change_none;
3064 struct isl_coalesce_info info_local, *info_i;
3066 info_i = i >= 0 ? &info[i] : &info_local;
3067 init_status(info_i);
3068 bmap = isl_basic_map_copy(bmap);
3069 bmap = isl_basic_map_expand_divs(bmap, isl_mat_copy(div), exp);
3070 bmap = isl_basic_map_mark_final(bmap);
3072 if (!bmap)
3073 goto error;
3075 info_local.bmap = bmap;
3076 info_i->eq = eq_status_in(bmap, info[j].tab);
3077 if (bmap->n_eq && !info_i->eq)
3078 goto error;
3079 if (any_eq(info_i, STATUS_ERROR))
3080 goto error;
3081 if (any_eq(info_i, STATUS_SEPARATE))
3082 goto done;
3084 info_i->ineq = ineq_status_in(bmap, NULL, info[j].tab);
3085 if (bmap->n_ineq && !info_i->ineq)
3086 goto error;
3087 if (any_ineq(info_i, STATUS_ERROR))
3088 goto error;
3089 if (any_ineq(info_i, STATUS_SEPARATE))
3090 goto done;
3092 if (all(info_i->eq, 2 * bmap->n_eq, STATUS_VALID) &&
3093 all(info_i->ineq, bmap->n_ineq, STATUS_VALID)) {
3094 drop(&info[j]);
3095 change = isl_change_drop_second;
3098 if (change == isl_change_none && i != -1)
3099 return coalesce_expand_tab_divs(bmap, i, j, info, div, exp);
3101 done:
3102 isl_basic_map_free(bmap);
3103 clear_status(info_i);
3104 return change;
3105 error:
3106 isl_basic_map_free(bmap);
3107 clear_status(info_i);
3108 return isl_change_error;
3111 /* Check if the union of "bmap_i" and the basic map represented by info[j]
3112 * can be represented by a single basic map,
3113 * after aligning the divs of "bmap_i" to match those of info[j].
3114 * If so, replace the pair by the single basic map and return
3115 * isl_change_drop_first, isl_change_drop_second or isl_change_fuse.
3116 * Otherwise, return isl_change_none.
3118 * In particular, check if "bmap_i" contains the basic map represented by
3119 * info[j] after aligning the divs of "bmap_i" to those of info[j].
3120 * Note that this can only succeed if the number of divs of "bmap_i"
3121 * is smaller than (or equal to) the number of divs of info[j].
3123 * We first check if the divs of "bmap_i" are all known and form a subset
3124 * of those of info[j].bmap. If so, we pass control over to
3125 * coalesce_with_expanded_divs.
3127 * If "i" is not equal to -1, then "bmap" is equal to info[i].bmap.
3129 static enum isl_change coalesce_after_aligning_divs(
3130 __isl_keep isl_basic_map *bmap_i, int i, int j,
3131 struct isl_coalesce_info *info)
3133 int known;
3134 isl_mat *div_i, *div_j, *div;
3135 int *exp1 = NULL;
3136 int *exp2 = NULL;
3137 isl_ctx *ctx;
3138 enum isl_change change;
3140 known = isl_basic_map_divs_known(bmap_i);
3141 if (known < 0 || !known)
3142 return known;
3144 ctx = isl_basic_map_get_ctx(bmap_i);
3146 div_i = isl_basic_map_get_divs(bmap_i);
3147 div_j = isl_basic_map_get_divs(info[j].bmap);
3149 if (!div_i || !div_j)
3150 goto error;
3152 exp1 = isl_alloc_array(ctx, int, div_i->n_row);
3153 exp2 = isl_alloc_array(ctx, int, div_j->n_row);
3154 if ((div_i->n_row && !exp1) || (div_j->n_row && !exp2))
3155 goto error;
3157 div = isl_merge_divs(div_i, div_j, exp1, exp2);
3158 if (!div)
3159 goto error;
3161 if (div->n_row == div_j->n_row)
3162 change = coalesce_with_expanded_divs(bmap_i,
3163 i, j, info, div, exp1);
3164 else
3165 change = isl_change_none;
3167 isl_mat_free(div);
3169 isl_mat_free(div_i);
3170 isl_mat_free(div_j);
3172 free(exp2);
3173 free(exp1);
3175 return change;
3176 error:
3177 isl_mat_free(div_i);
3178 isl_mat_free(div_j);
3179 free(exp1);
3180 free(exp2);
3181 return isl_change_error;
3184 /* Check if basic map "j" is a subset of basic map "i" after
3185 * exploiting the extra equalities of "j" to simplify the divs of "i".
3186 * If so, remove basic map "j" and return isl_change_drop_second.
3188 * If "j" does not have any equalities or if they are the same
3189 * as those of "i", then we cannot exploit them to simplify the divs.
3190 * Similarly, if there are no divs in "i", then they cannot be simplified.
3191 * If, on the other hand, the affine hulls of "i" and "j" do not intersect,
3192 * then "j" cannot be a subset of "i".
3194 * Otherwise, we intersect "i" with the affine hull of "j" and then
3195 * check if "j" is a subset of the result after aligning the divs.
3196 * If so, then "j" is definitely a subset of "i" and can be removed.
3197 * Note that if after intersection with the affine hull of "j".
3198 * "i" still has more divs than "j", then there is no way we can
3199 * align the divs of "i" to those of "j".
3201 static enum isl_change coalesce_subset_with_equalities(int i, int j,
3202 struct isl_coalesce_info *info)
3204 isl_basic_map *hull_i, *hull_j, *bmap_i;
3205 int equal, empty;
3206 enum isl_change change;
3208 if (info[j].bmap->n_eq == 0)
3209 return isl_change_none;
3210 if (info[i].bmap->n_div == 0)
3211 return isl_change_none;
3213 hull_i = isl_basic_map_copy(info[i].bmap);
3214 hull_i = isl_basic_map_plain_affine_hull(hull_i);
3215 hull_j = isl_basic_map_copy(info[j].bmap);
3216 hull_j = isl_basic_map_plain_affine_hull(hull_j);
3218 hull_j = isl_basic_map_intersect(hull_j, isl_basic_map_copy(hull_i));
3219 equal = isl_basic_map_plain_is_equal(hull_i, hull_j);
3220 empty = isl_basic_map_plain_is_empty(hull_j);
3221 isl_basic_map_free(hull_i);
3223 if (equal < 0 || equal || empty < 0 || empty) {
3224 isl_basic_map_free(hull_j);
3225 if (equal < 0 || empty < 0)
3226 return isl_change_error;
3227 return isl_change_none;
3230 bmap_i = isl_basic_map_copy(info[i].bmap);
3231 bmap_i = isl_basic_map_intersect(bmap_i, hull_j);
3232 if (!bmap_i)
3233 return isl_change_error;
3235 if (bmap_i->n_div > info[j].bmap->n_div) {
3236 isl_basic_map_free(bmap_i);
3237 return isl_change_none;
3240 change = coalesce_after_aligning_divs(bmap_i, -1, j, info);
3242 isl_basic_map_free(bmap_i);
3244 return change;
3247 /* Check if the union of and the basic maps represented by info[i] and info[j]
3248 * can be represented by a single basic map, by aligning or equating
3249 * their integer divisions.
3250 * If so, replace the pair by the single basic map and return
3251 * isl_change_drop_first, isl_change_drop_second or isl_change_fuse.
3252 * Otherwise, return isl_change_none.
3254 * Note that we only perform any test if the number of divs is different
3255 * in the two basic maps. In case the number of divs is the same,
3256 * we have already established that the divs are different
3257 * in the two basic maps.
3258 * In particular, if the number of divs of basic map i is smaller than
3259 * the number of divs of basic map j, then we check if j is a subset of i
3260 * and vice versa.
3262 static enum isl_change coalesce_divs(int i, int j,
3263 struct isl_coalesce_info *info)
3265 enum isl_change change = isl_change_none;
3267 if (info[i].bmap->n_div < info[j].bmap->n_div)
3268 change = coalesce_after_aligning_divs(info[i].bmap, i, j, info);
3269 if (change != isl_change_none)
3270 return change;
3272 if (info[j].bmap->n_div < info[i].bmap->n_div)
3273 change = coalesce_after_aligning_divs(info[j].bmap, j, i, info);
3274 if (change != isl_change_none)
3275 return invert_change(change);
3277 change = coalesce_subset_with_equalities(i, j, info);
3278 if (change != isl_change_none)
3279 return change;
3281 change = coalesce_subset_with_equalities(j, i, info);
3282 if (change != isl_change_none)
3283 return invert_change(change);
3285 return isl_change_none;
3288 /* Does "bmap" involve any divs that themselves refer to divs?
3290 static isl_bool has_nested_div(__isl_keep isl_basic_map *bmap)
3292 int i;
3293 unsigned total;
3294 unsigned n_div;
3296 total = isl_basic_map_dim(bmap, isl_dim_all);
3297 n_div = isl_basic_map_dim(bmap, isl_dim_div);
3298 total -= n_div;
3300 for (i = 0; i < n_div; ++i)
3301 if (isl_seq_first_non_zero(bmap->div[i] + 2 + total,
3302 n_div) != -1)
3303 return isl_bool_true;
3305 return isl_bool_false;
3308 /* Return a list of affine expressions, one for each integer division
3309 * in "bmap_i". For each integer division that also appears in "bmap_j",
3310 * the affine expression is set to NaN. The number of NaNs in the list
3311 * is equal to the number of integer divisions in "bmap_j".
3312 * For the other integer divisions of "bmap_i", the corresponding
3313 * element in the list is a purely affine expression equal to the integer
3314 * division in "hull".
3315 * If no such list can be constructed, then the number of elements
3316 * in the returned list is smaller than the number of integer divisions
3317 * in "bmap_i".
3319 static __isl_give isl_aff_list *set_up_substitutions(
3320 __isl_keep isl_basic_map *bmap_i, __isl_keep isl_basic_map *bmap_j,
3321 __isl_take isl_basic_map *hull)
3323 unsigned n_div_i, n_div_j, total;
3324 isl_ctx *ctx;
3325 isl_local_space *ls;
3326 isl_basic_set *wrap_hull;
3327 isl_aff *aff_nan;
3328 isl_aff_list *list;
3329 int i, j;
3331 if (!hull)
3332 return NULL;
3334 ctx = isl_basic_map_get_ctx(hull);
3336 n_div_i = isl_basic_map_dim(bmap_i, isl_dim_div);
3337 n_div_j = isl_basic_map_dim(bmap_j, isl_dim_div);
3338 total = isl_basic_map_total_dim(bmap_i) - n_div_i;
3340 ls = isl_basic_map_get_local_space(bmap_i);
3341 ls = isl_local_space_wrap(ls);
3342 wrap_hull = isl_basic_map_wrap(hull);
3344 aff_nan = isl_aff_nan_on_domain(isl_local_space_copy(ls));
3345 list = isl_aff_list_alloc(ctx, n_div_i);
3347 j = 0;
3348 for (i = 0; i < n_div_i; ++i) {
3349 isl_aff *aff;
3351 if (j < n_div_j &&
3352 isl_basic_map_equal_div_expr_part(bmap_i, i, bmap_j, j,
3353 0, 2 + total)) {
3354 ++j;
3355 list = isl_aff_list_add(list, isl_aff_copy(aff_nan));
3356 continue;
3358 if (n_div_i - i <= n_div_j - j)
3359 break;
3361 aff = isl_local_space_get_div(ls, i);
3362 aff = isl_aff_substitute_equalities(aff,
3363 isl_basic_set_copy(wrap_hull));
3364 aff = isl_aff_floor(aff);
3365 if (!aff)
3366 goto error;
3367 if (isl_aff_dim(aff, isl_dim_div) != 0) {
3368 isl_aff_free(aff);
3369 break;
3372 list = isl_aff_list_add(list, aff);
3375 isl_aff_free(aff_nan);
3376 isl_local_space_free(ls);
3377 isl_basic_set_free(wrap_hull);
3379 return list;
3380 error:
3381 isl_aff_free(aff_nan);
3382 isl_local_space_free(ls);
3383 isl_basic_set_free(wrap_hull);
3384 isl_aff_list_free(list);
3385 return NULL;
3388 /* Add variables to info->bmap and info->tab corresponding to the elements
3389 * in "list" that are not set to NaN.
3390 * "extra_var" is the number of these elements.
3391 * "dim" is the offset in the variables of "tab" where we should
3392 * start considering the elements in "list".
3393 * When this function returns, the total number of variables in "tab"
3394 * is equal to "dim" plus the number of elements in "list".
3396 * The newly added existentially quantified variables are not given
3397 * an explicit representation because the corresponding div constraints
3398 * do not appear in info->bmap. These constraints are not added
3399 * to info->bmap because for internal consistency, they would need to
3400 * be added to info->tab as well, where they could combine with the equality
3401 * that is added later to result in constraints that do not hold
3402 * in the original input.
3404 static isl_stat add_sub_vars(struct isl_coalesce_info *info,
3405 __isl_keep isl_aff_list *list, int dim, int extra_var)
3407 int i, j, n, d;
3408 isl_space *space;
3410 space = isl_basic_map_get_space(info->bmap);
3411 info->bmap = isl_basic_map_cow(info->bmap);
3412 info->bmap = isl_basic_map_extend_space(info->bmap, space,
3413 extra_var, 0, 0);
3414 if (!info->bmap)
3415 return isl_stat_error;
3416 n = isl_aff_list_n_aff(list);
3417 for (i = 0; i < n; ++i) {
3418 int is_nan;
3419 isl_aff *aff;
3421 aff = isl_aff_list_get_aff(list, i);
3422 is_nan = isl_aff_is_nan(aff);
3423 isl_aff_free(aff);
3424 if (is_nan < 0)
3425 return isl_stat_error;
3426 if (is_nan)
3427 continue;
3429 if (isl_tab_insert_var(info->tab, dim + i) < 0)
3430 return isl_stat_error;
3431 d = isl_basic_map_alloc_div(info->bmap);
3432 if (d < 0)
3433 return isl_stat_error;
3434 info->bmap = isl_basic_map_mark_div_unknown(info->bmap, d);
3435 if (!info->bmap)
3436 return isl_stat_error;
3437 for (j = d; j > i; --j)
3438 isl_basic_map_swap_div(info->bmap, j - 1, j);
3441 return isl_stat_ok;
3444 /* For each element in "list" that is not set to NaN, fix the corresponding
3445 * variable in "tab" to the purely affine expression defined by the element.
3446 * "dim" is the offset in the variables of "tab" where we should
3447 * start considering the elements in "list".
3449 * This function assumes that a sufficient number of rows and
3450 * elements in the constraint array are available in the tableau.
3452 static int add_sub_equalities(struct isl_tab *tab,
3453 __isl_keep isl_aff_list *list, int dim)
3455 int i, n;
3456 isl_ctx *ctx;
3457 isl_vec *sub;
3458 isl_aff *aff;
3460 n = isl_aff_list_n_aff(list);
3462 ctx = isl_tab_get_ctx(tab);
3463 sub = isl_vec_alloc(ctx, 1 + dim + n);
3464 if (!sub)
3465 return -1;
3466 isl_seq_clr(sub->el + 1 + dim, n);
3468 for (i = 0; i < n; ++i) {
3469 aff = isl_aff_list_get_aff(list, i);
3470 if (!aff)
3471 goto error;
3472 if (isl_aff_is_nan(aff)) {
3473 isl_aff_free(aff);
3474 continue;
3476 isl_seq_cpy(sub->el, aff->v->el + 1, 1 + dim);
3477 isl_int_neg(sub->el[1 + dim + i], aff->v->el[0]);
3478 if (isl_tab_add_eq(tab, sub->el) < 0)
3479 goto error;
3480 isl_int_set_si(sub->el[1 + dim + i], 0);
3481 isl_aff_free(aff);
3484 isl_vec_free(sub);
3485 return 0;
3486 error:
3487 isl_aff_free(aff);
3488 isl_vec_free(sub);
3489 return -1;
3492 /* Add variables to info->tab and info->bmap corresponding to the elements
3493 * in "list" that are not set to NaN. The value of the added variable
3494 * in info->tab is fixed to the purely affine expression defined by the element.
3495 * "dim" is the offset in the variables of info->tab where we should
3496 * start considering the elements in "list".
3497 * When this function returns, the total number of variables in info->tab
3498 * is equal to "dim" plus the number of elements in "list".
3500 static int add_subs(struct isl_coalesce_info *info,
3501 __isl_keep isl_aff_list *list, int dim)
3503 int extra_var;
3504 int n;
3506 if (!list)
3507 return -1;
3509 n = isl_aff_list_n_aff(list);
3510 extra_var = n - (info->tab->n_var - dim);
3512 if (isl_tab_extend_vars(info->tab, extra_var) < 0)
3513 return -1;
3514 if (isl_tab_extend_cons(info->tab, 2 * extra_var) < 0)
3515 return -1;
3516 if (add_sub_vars(info, list, dim, extra_var) < 0)
3517 return -1;
3519 return add_sub_equalities(info->tab, list, dim);
3522 /* Coalesce basic map "j" into basic map "i" after adding the extra integer
3523 * divisions in "i" but not in "j" to basic map "j", with values
3524 * specified by "list". The total number of elements in "list"
3525 * is equal to the number of integer divisions in "i", while the number
3526 * of NaN elements in the list is equal to the number of integer divisions
3527 * in "j".
3529 * If no coalescing can be performed, then we need to revert basic map "j"
3530 * to its original state. We do the same if basic map "i" gets dropped
3531 * during the coalescing, even though this should not happen in practice
3532 * since we have already checked for "j" being a subset of "i"
3533 * before we reach this stage.
3535 static enum isl_change coalesce_with_subs(int i, int j,
3536 struct isl_coalesce_info *info, __isl_keep isl_aff_list *list)
3538 isl_basic_map *bmap_j;
3539 struct isl_tab_undo *snap;
3540 unsigned dim;
3541 enum isl_change change;
3543 bmap_j = isl_basic_map_copy(info[j].bmap);
3544 snap = isl_tab_snap(info[j].tab);
3546 dim = isl_basic_map_dim(bmap_j, isl_dim_all);
3547 dim -= isl_basic_map_dim(bmap_j, isl_dim_div);
3548 if (add_subs(&info[j], list, dim) < 0)
3549 goto error;
3551 change = coalesce_local_pair(i, j, info);
3552 if (change != isl_change_none && change != isl_change_drop_first) {
3553 isl_basic_map_free(bmap_j);
3554 } else {
3555 isl_basic_map_free(info[j].bmap);
3556 info[j].bmap = bmap_j;
3558 if (isl_tab_rollback(info[j].tab, snap) < 0)
3559 return isl_change_error;
3562 return change;
3563 error:
3564 isl_basic_map_free(bmap_j);
3565 return isl_change_error;
3568 /* Check if we can coalesce basic map "j" into basic map "i" after copying
3569 * those extra integer divisions in "i" that can be simplified away
3570 * using the extra equalities in "j".
3571 * All divs are assumed to be known and not contain any nested divs.
3573 * We first check if there are any extra equalities in "j" that we
3574 * can exploit. Then we check if every integer division in "i"
3575 * either already appears in "j" or can be simplified using the
3576 * extra equalities to a purely affine expression.
3577 * If these tests succeed, then we try to coalesce the two basic maps
3578 * by introducing extra dimensions in "j" corresponding to
3579 * the extra integer divsisions "i" fixed to the corresponding
3580 * purely affine expression.
3582 static enum isl_change check_coalesce_into_eq(int i, int j,
3583 struct isl_coalesce_info *info)
3585 unsigned n_div_i, n_div_j;
3586 isl_basic_map *hull_i, *hull_j;
3587 int equal, empty;
3588 isl_aff_list *list;
3589 enum isl_change change;
3591 n_div_i = isl_basic_map_dim(info[i].bmap, isl_dim_div);
3592 n_div_j = isl_basic_map_dim(info[j].bmap, isl_dim_div);
3593 if (n_div_i <= n_div_j)
3594 return isl_change_none;
3595 if (info[j].bmap->n_eq == 0)
3596 return isl_change_none;
3598 hull_i = isl_basic_map_copy(info[i].bmap);
3599 hull_i = isl_basic_map_plain_affine_hull(hull_i);
3600 hull_j = isl_basic_map_copy(info[j].bmap);
3601 hull_j = isl_basic_map_plain_affine_hull(hull_j);
3603 hull_j = isl_basic_map_intersect(hull_j, isl_basic_map_copy(hull_i));
3604 equal = isl_basic_map_plain_is_equal(hull_i, hull_j);
3605 empty = isl_basic_map_plain_is_empty(hull_j);
3606 isl_basic_map_free(hull_i);
3608 if (equal < 0 || empty < 0)
3609 goto error;
3610 if (equal || empty) {
3611 isl_basic_map_free(hull_j);
3612 return isl_change_none;
3615 list = set_up_substitutions(info[i].bmap, info[j].bmap, hull_j);
3616 if (!list)
3617 return isl_change_error;
3618 if (isl_aff_list_n_aff(list) < n_div_i)
3619 change = isl_change_none;
3620 else
3621 change = coalesce_with_subs(i, j, info, list);
3623 isl_aff_list_free(list);
3625 return change;
3626 error:
3627 isl_basic_map_free(hull_j);
3628 return isl_change_error;
3631 /* Check if we can coalesce basic maps "i" and "j" after copying
3632 * those extra integer divisions in one of the basic maps that can
3633 * be simplified away using the extra equalities in the other basic map.
3634 * We require all divs to be known in both basic maps.
3635 * Furthermore, to simplify the comparison of div expressions,
3636 * we do not allow any nested integer divisions.
3638 static enum isl_change check_coalesce_eq(int i, int j,
3639 struct isl_coalesce_info *info)
3641 isl_bool known, nested;
3642 enum isl_change change;
3644 known = isl_basic_map_divs_known(info[i].bmap);
3645 if (known < 0 || !known)
3646 return known < 0 ? isl_change_error : isl_change_none;
3647 known = isl_basic_map_divs_known(info[j].bmap);
3648 if (known < 0 || !known)
3649 return known < 0 ? isl_change_error : isl_change_none;
3650 nested = has_nested_div(info[i].bmap);
3651 if (nested < 0 || nested)
3652 return nested < 0 ? isl_change_error : isl_change_none;
3653 nested = has_nested_div(info[j].bmap);
3654 if (nested < 0 || nested)
3655 return nested < 0 ? isl_change_error : isl_change_none;
3657 change = check_coalesce_into_eq(i, j, info);
3658 if (change != isl_change_none)
3659 return change;
3660 change = check_coalesce_into_eq(j, i, info);
3661 if (change != isl_change_none)
3662 return invert_change(change);
3664 return isl_change_none;
3667 /* Check if the union of the given pair of basic maps
3668 * can be represented by a single basic map.
3669 * If so, replace the pair by the single basic map and return
3670 * isl_change_drop_first, isl_change_drop_second or isl_change_fuse.
3671 * Otherwise, return isl_change_none.
3673 * We first check if the two basic maps live in the same local space,
3674 * after aligning the divs that differ by only an integer constant.
3675 * If so, we do the complete check. Otherwise, we check if they have
3676 * the same number of integer divisions and can be coalesced, if one is
3677 * an obvious subset of the other or if the extra integer divisions
3678 * of one basic map can be simplified away using the extra equalities
3679 * of the other basic map.
3681 static enum isl_change coalesce_pair(int i, int j,
3682 struct isl_coalesce_info *info)
3684 isl_bool same;
3685 enum isl_change change;
3687 if (harmonize_divs(&info[i], &info[j]) < 0)
3688 return isl_change_error;
3689 same = same_divs(info[i].bmap, info[j].bmap);
3690 if (same < 0)
3691 return isl_change_error;
3692 if (same)
3693 return coalesce_local_pair(i, j, info);
3695 if (info[i].bmap->n_div == info[j].bmap->n_div) {
3696 change = coalesce_local_pair(i, j, info);
3697 if (change != isl_change_none)
3698 return change;
3701 change = coalesce_divs(i, j, info);
3702 if (change != isl_change_none)
3703 return change;
3705 return check_coalesce_eq(i, j, info);
3708 /* Return the maximum of "a" and "b".
3710 static int isl_max(int a, int b)
3712 return a > b ? a : b;
3715 /* Pairwise coalesce the basic maps in the range [start1, end1[ of "info"
3716 * with those in the range [start2, end2[, skipping basic maps
3717 * that have been removed (either before or within this function).
3719 * For each basic map i in the first range, we check if it can be coalesced
3720 * with respect to any previously considered basic map j in the second range.
3721 * If i gets dropped (because it was a subset of some j), then
3722 * we can move on to the next basic map.
3723 * If j gets dropped, we need to continue checking against the other
3724 * previously considered basic maps.
3725 * If the two basic maps got fused, then we recheck the fused basic map
3726 * against the previously considered basic maps, starting at i + 1
3727 * (even if start2 is greater than i + 1).
3729 static int coalesce_range(isl_ctx *ctx, struct isl_coalesce_info *info,
3730 int start1, int end1, int start2, int end2)
3732 int i, j;
3734 for (i = end1 - 1; i >= start1; --i) {
3735 if (info[i].removed)
3736 continue;
3737 for (j = isl_max(i + 1, start2); j < end2; ++j) {
3738 enum isl_change changed;
3740 if (info[j].removed)
3741 continue;
3742 if (info[i].removed)
3743 isl_die(ctx, isl_error_internal,
3744 "basic map unexpectedly removed",
3745 return -1);
3746 changed = coalesce_pair(i, j, info);
3747 switch (changed) {
3748 case isl_change_error:
3749 return -1;
3750 case isl_change_none:
3751 case isl_change_drop_second:
3752 continue;
3753 case isl_change_drop_first:
3754 j = end2;
3755 break;
3756 case isl_change_fuse:
3757 j = i;
3758 break;
3763 return 0;
3766 /* Pairwise coalesce the basic maps described by the "n" elements of "info".
3768 * We consider groups of basic maps that live in the same apparent
3769 * affine hull and we first coalesce within such a group before we
3770 * coalesce the elements in the group with elements of previously
3771 * considered groups. If a fuse happens during the second phase,
3772 * then we also reconsider the elements within the group.
3774 static int coalesce(isl_ctx *ctx, int n, struct isl_coalesce_info *info)
3776 int start, end;
3778 for (end = n; end > 0; end = start) {
3779 start = end - 1;
3780 while (start >= 1 &&
3781 info[start - 1].hull_hash == info[start].hull_hash)
3782 start--;
3783 if (coalesce_range(ctx, info, start, end, start, end) < 0)
3784 return -1;
3785 if (coalesce_range(ctx, info, start, end, end, n) < 0)
3786 return -1;
3789 return 0;
3792 /* Update the basic maps in "map" based on the information in "info".
3793 * In particular, remove the basic maps that have been marked removed and
3794 * update the others based on the information in the corresponding tableau.
3795 * Since we detected implicit equalities without calling
3796 * isl_basic_map_gauss, we need to do it now.
3797 * Also call isl_basic_map_simplify if we may have lost the definition
3798 * of one or more integer divisions.
3800 static __isl_give isl_map *update_basic_maps(__isl_take isl_map *map,
3801 int n, struct isl_coalesce_info *info)
3803 int i;
3805 if (!map)
3806 return NULL;
3808 for (i = n - 1; i >= 0; --i) {
3809 if (info[i].removed) {
3810 isl_basic_map_free(map->p[i]);
3811 if (i != map->n - 1)
3812 map->p[i] = map->p[map->n - 1];
3813 map->n--;
3814 continue;
3817 info[i].bmap = isl_basic_map_update_from_tab(info[i].bmap,
3818 info[i].tab);
3819 info[i].bmap = isl_basic_map_gauss(info[i].bmap, NULL);
3820 if (info[i].simplify)
3821 info[i].bmap = isl_basic_map_simplify(info[i].bmap);
3822 info[i].bmap = isl_basic_map_finalize(info[i].bmap);
3823 if (!info[i].bmap)
3824 return isl_map_free(map);
3825 ISL_F_SET(info[i].bmap, ISL_BASIC_MAP_NO_IMPLICIT);
3826 ISL_F_SET(info[i].bmap, ISL_BASIC_MAP_NO_REDUNDANT);
3827 isl_basic_map_free(map->p[i]);
3828 map->p[i] = info[i].bmap;
3829 info[i].bmap = NULL;
3832 return map;
3835 /* For each pair of basic maps in the map, check if the union of the two
3836 * can be represented by a single basic map.
3837 * If so, replace the pair by the single basic map and start over.
3839 * We factor out any (hidden) common factor from the constraint
3840 * coefficients to improve the detection of adjacent constraints.
3842 * Since we are constructing the tableaus of the basic maps anyway,
3843 * we exploit them to detect implicit equalities and redundant constraints.
3844 * This also helps the coalescing as it can ignore the redundant constraints.
3845 * In order to avoid confusion, we make all implicit equalities explicit
3846 * in the basic maps. We don't call isl_basic_map_gauss, though,
3847 * as that may affect the number of constraints.
3848 * This means that we have to call isl_basic_map_gauss at the end
3849 * of the computation (in update_basic_maps) to ensure that
3850 * the basic maps are not left in an unexpected state.
3851 * For each basic map, we also compute the hash of the apparent affine hull
3852 * for use in coalesce.
3854 __isl_give isl_map *isl_map_coalesce(__isl_take isl_map *map)
3856 int i;
3857 unsigned n;
3858 isl_ctx *ctx;
3859 struct isl_coalesce_info *info = NULL;
3861 map = isl_map_remove_empty_parts(map);
3862 if (!map)
3863 return NULL;
3865 if (map->n <= 1)
3866 return map;
3868 ctx = isl_map_get_ctx(map);
3869 map = isl_map_sort_divs(map);
3870 map = isl_map_cow(map);
3872 if (!map)
3873 return NULL;
3875 n = map->n;
3877 info = isl_calloc_array(map->ctx, struct isl_coalesce_info, n);
3878 if (!info)
3879 goto error;
3881 for (i = 0; i < map->n; ++i) {
3882 map->p[i] = isl_basic_map_reduce_coefficients(map->p[i]);
3883 if (!map->p[i])
3884 goto error;
3885 info[i].bmap = isl_basic_map_copy(map->p[i]);
3886 info[i].tab = isl_tab_from_basic_map(info[i].bmap, 0);
3887 if (!info[i].tab)
3888 goto error;
3889 if (!ISL_F_ISSET(info[i].bmap, ISL_BASIC_MAP_NO_IMPLICIT))
3890 if (isl_tab_detect_implicit_equalities(info[i].tab) < 0)
3891 goto error;
3892 info[i].bmap = isl_tab_make_equalities_explicit(info[i].tab,
3893 info[i].bmap);
3894 if (!info[i].bmap)
3895 goto error;
3896 if (!ISL_F_ISSET(info[i].bmap, ISL_BASIC_MAP_NO_REDUNDANT))
3897 if (isl_tab_detect_redundant(info[i].tab) < 0)
3898 goto error;
3899 if (coalesce_info_set_hull_hash(&info[i]) < 0)
3900 goto error;
3902 for (i = map->n - 1; i >= 0; --i)
3903 if (info[i].tab->empty)
3904 drop(&info[i]);
3906 if (coalesce(ctx, n, info) < 0)
3907 goto error;
3909 map = update_basic_maps(map, n, info);
3911 clear_coalesce_info(n, info);
3913 return map;
3914 error:
3915 clear_coalesce_info(n, info);
3916 isl_map_free(map);
3917 return NULL;
3920 /* For each pair of basic sets in the set, check if the union of the two
3921 * can be represented by a single basic set.
3922 * If so, replace the pair by the single basic set and start over.
3924 struct isl_set *isl_set_coalesce(struct isl_set *set)
3926 return set_from_map(isl_map_coalesce(set_to_map(set)));