isl_map_coalesce: handle divs that have been simplified away in one basic map
[isl.git] / isl_coalesce.c
blobce396a95ebe22c5eab81c65dac81ada451e65516
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
7 * Use of this software is governed by the MIT license
9 * Written by Sven Verdoolaege, K.U.Leuven, Departement
10 * Computerwetenschappen, Celestijnenlaan 200A, B-3001 Leuven, Belgium
11 * and INRIA Saclay - Ile-de-France, Parc Club Orsay Universite,
12 * ZAC des vignes, 4 rue Jacques Monod, 91893 Orsay, France
13 * and Ecole Normale Superieure, 45 rue d’Ulm, 75230 Paris, France
14 * and Inria Paris - Rocquencourt, Domaine de Voluceau - Rocquencourt,
15 * B.P. 105 - 78153 Le Chesnay, France
18 #include "isl_map_private.h"
19 #include <isl_seq.h>
20 #include <isl/options.h>
21 #include "isl_tab.h"
22 #include <isl_mat_private.h>
23 #include <isl_local_space_private.h>
24 #include <isl_vec_private.h>
25 #include <isl_aff_private.h>
27 #define STATUS_ERROR -1
28 #define STATUS_REDUNDANT 1
29 #define STATUS_VALID 2
30 #define STATUS_SEPARATE 3
31 #define STATUS_CUT 4
32 #define STATUS_ADJ_EQ 5
33 #define STATUS_ADJ_INEQ 6
35 static int status_in(isl_int *ineq, struct isl_tab *tab)
37 enum isl_ineq_type type = isl_tab_ineq_type(tab, ineq);
38 switch (type) {
39 default:
40 case isl_ineq_error: return STATUS_ERROR;
41 case isl_ineq_redundant: return STATUS_VALID;
42 case isl_ineq_separate: return STATUS_SEPARATE;
43 case isl_ineq_cut: return STATUS_CUT;
44 case isl_ineq_adj_eq: return STATUS_ADJ_EQ;
45 case isl_ineq_adj_ineq: return STATUS_ADJ_INEQ;
49 /* Compute the position of the equalities of basic map "bmap_i"
50 * with respect to the basic map represented by "tab_j".
51 * The resulting array has twice as many entries as the number
52 * of equalities corresponding to the two inequalties to which
53 * each equality corresponds.
55 static int *eq_status_in(__isl_keep isl_basic_map *bmap_i,
56 struct isl_tab *tab_j)
58 int k, l;
59 int *eq = isl_calloc_array(bmap_i->ctx, int, 2 * bmap_i->n_eq);
60 unsigned dim;
62 if (!eq)
63 return NULL;
65 dim = isl_basic_map_total_dim(bmap_i);
66 for (k = 0; k < bmap_i->n_eq; ++k) {
67 for (l = 0; l < 2; ++l) {
68 isl_seq_neg(bmap_i->eq[k], bmap_i->eq[k], 1+dim);
69 eq[2 * k + l] = status_in(bmap_i->eq[k], tab_j);
70 if (eq[2 * k + l] == STATUS_ERROR)
71 goto error;
73 if (eq[2 * k] == STATUS_SEPARATE ||
74 eq[2 * k + 1] == STATUS_SEPARATE)
75 break;
78 return eq;
79 error:
80 free(eq);
81 return NULL;
84 /* Compute the position of the inequalities of basic map "bmap_i"
85 * (also represented by "tab_i", if not NULL) with respect to the basic map
86 * represented by "tab_j".
88 static int *ineq_status_in(__isl_keep isl_basic_map *bmap_i,
89 struct isl_tab *tab_i, struct isl_tab *tab_j)
91 int k;
92 unsigned n_eq = bmap_i->n_eq;
93 int *ineq = isl_calloc_array(bmap_i->ctx, int, bmap_i->n_ineq);
95 if (!ineq)
96 return NULL;
98 for (k = 0; k < bmap_i->n_ineq; ++k) {
99 if (tab_i && isl_tab_is_redundant(tab_i, n_eq + k)) {
100 ineq[k] = STATUS_REDUNDANT;
101 continue;
103 ineq[k] = status_in(bmap_i->ineq[k], tab_j);
104 if (ineq[k] == STATUS_ERROR)
105 goto error;
106 if (ineq[k] == STATUS_SEPARATE)
107 break;
110 return ineq;
111 error:
112 free(ineq);
113 return NULL;
116 static int any(int *con, unsigned len, int status)
118 int i;
120 for (i = 0; i < len ; ++i)
121 if (con[i] == status)
122 return 1;
123 return 0;
126 static int count(int *con, unsigned len, int status)
128 int i;
129 int c = 0;
131 for (i = 0; i < len ; ++i)
132 if (con[i] == status)
133 c++;
134 return c;
137 static int all(int *con, unsigned len, int status)
139 int i;
141 for (i = 0; i < len ; ++i) {
142 if (con[i] == STATUS_REDUNDANT)
143 continue;
144 if (con[i] != status)
145 return 0;
147 return 1;
150 /* Internal information associated to a basic map in a map
151 * that is to be coalesced by isl_map_coalesce.
153 * "bmap" is the basic map itself (or NULL if "removed" is set)
154 * "tab" is the corresponding tableau (or NULL if "removed" is set)
155 * "removed" is set if this basic map has been removed from the map
157 * "eq" and "ineq" are only set if we are currently trying to coalesce
158 * this basic map with another basic map, in which case they represent
159 * the position of the inequalities of this basic map with respect to
160 * the other basic map. The number of elements in the "eq" array
161 * is twice the number of equalities in the "bmap", corresponding
162 * to the two inequalities that make up each equality.
164 struct isl_coalesce_info {
165 isl_basic_map *bmap;
166 struct isl_tab *tab;
167 int removed;
168 int *eq;
169 int *ineq;
172 /* Free all the allocated memory in an array
173 * of "n" isl_coalesce_info elements.
175 static void clear_coalesce_info(int n, struct isl_coalesce_info *info)
177 int i;
179 if (!info)
180 return;
182 for (i = 0; i < n; ++i) {
183 isl_basic_map_free(info[i].bmap);
184 isl_tab_free(info[i].tab);
187 free(info);
190 /* Drop the basic map represented by "info".
191 * That is, clear the memory associated to the entry and
192 * mark it as having been removed.
194 static void drop(struct isl_coalesce_info *info)
196 info->bmap = isl_basic_map_free(info->bmap);
197 isl_tab_free(info->tab);
198 info->tab = NULL;
199 info->removed = 1;
202 /* Exchange the information in "info1" with that in "info2".
204 static void exchange(struct isl_coalesce_info *info1,
205 struct isl_coalesce_info *info2)
207 struct isl_coalesce_info info;
209 info = *info1;
210 *info1 = *info2;
211 *info2 = info;
214 /* This type represents the kind of change that has been performed
215 * while trying to coalesce two basic maps.
217 * isl_change_none: nothing was changed
218 * isl_change_drop_first: the first basic map was removed
219 * isl_change_drop_second: the second basic map was removed
220 * isl_change_fuse: the two basic maps were replaced by a new basic map.
222 enum isl_change {
223 isl_change_error = -1,
224 isl_change_none = 0,
225 isl_change_drop_first,
226 isl_change_drop_second,
227 isl_change_fuse,
230 /* Update "change" based on an interchange of the first and the second
231 * basic map. That is, interchange isl_change_drop_first and
232 * isl_change_drop_second.
234 static enum isl_change invert_change(enum isl_change change)
236 switch (change) {
237 case isl_change_error:
238 return isl_change_error;
239 case isl_change_none:
240 return isl_change_none;
241 case isl_change_drop_first:
242 return isl_change_drop_second;
243 case isl_change_drop_second:
244 return isl_change_drop_first;
245 case isl_change_fuse:
246 return isl_change_fuse;
250 /* Add the valid constraints of the basic map represented by "info"
251 * to "bmap". "len" is the size of the constraints.
252 * If only one of the pair of inequalities that make up an equality
253 * is valid, then add that inequality.
255 static __isl_give isl_basic_map *add_valid_constraints(
256 __isl_take isl_basic_map *bmap, struct isl_coalesce_info *info,
257 unsigned len)
259 int k, l;
261 if (!bmap)
262 return NULL;
264 for (k = 0; k < info->bmap->n_eq; ++k) {
265 if (info->eq[2 * k] == STATUS_VALID &&
266 info->eq[2 * k + 1] == STATUS_VALID) {
267 l = isl_basic_map_alloc_equality(bmap);
268 if (l < 0)
269 return isl_basic_map_free(bmap);
270 isl_seq_cpy(bmap->eq[l], info->bmap->eq[k], len);
271 } else if (info->eq[2 * k] == STATUS_VALID) {
272 l = isl_basic_map_alloc_inequality(bmap);
273 if (l < 0)
274 return isl_basic_map_free(bmap);
275 isl_seq_neg(bmap->ineq[l], info->bmap->eq[k], len);
276 } else if (info->eq[2 * k + 1] == STATUS_VALID) {
277 l = isl_basic_map_alloc_inequality(bmap);
278 if (l < 0)
279 return isl_basic_map_free(bmap);
280 isl_seq_cpy(bmap->ineq[l], info->bmap->eq[k], len);
284 for (k = 0; k < info->bmap->n_ineq; ++k) {
285 if (info->ineq[k] != STATUS_VALID)
286 continue;
287 l = isl_basic_map_alloc_inequality(bmap);
288 if (l < 0)
289 return isl_basic_map_free(bmap);
290 isl_seq_cpy(bmap->ineq[l], info->bmap->ineq[k], len);
293 return bmap;
296 /* Is "bmap" defined by a number of (non-redundant) constraints that
297 * is greater than the number of constraints of basic maps i and j combined?
298 * Equalities are counted as two inequalities.
300 static int number_of_constraints_increases(int i, int j,
301 struct isl_coalesce_info *info,
302 __isl_keep isl_basic_map *bmap, struct isl_tab *tab)
304 int k, n_old, n_new;
306 n_old = 2 * info[i].bmap->n_eq + info[i].bmap->n_ineq;
307 n_old += 2 * info[j].bmap->n_eq + info[j].bmap->n_ineq;
309 n_new = 2 * bmap->n_eq;
310 for (k = 0; k < bmap->n_ineq; ++k)
311 if (!isl_tab_is_redundant(tab, bmap->n_eq + k))
312 ++n_new;
314 return n_new > n_old;
317 /* Replace the pair of basic maps i and j by the basic map bounded
318 * by the valid constraints in both basic maps and the constraints
319 * in extra (if not NULL).
320 * Place the fused basic map in the position that is the smallest of i and j.
322 * If "detect_equalities" is set, then look for equalities encoded
323 * as pairs of inequalities.
324 * If "check_number" is set, then the original basic maps are only
325 * replaced if the total number of constraints does not increase.
327 static enum isl_change fuse(int i, int j, struct isl_coalesce_info *info,
328 __isl_keep isl_mat *extra, int detect_equalities, int check_number)
330 int k, l;
331 struct isl_basic_map *fused = NULL;
332 struct isl_tab *fused_tab = NULL;
333 unsigned total = isl_basic_map_total_dim(info[i].bmap);
334 unsigned extra_rows = extra ? extra->n_row : 0;
335 unsigned n_eq, n_ineq;
337 if (j < i)
338 return fuse(j, i, info, extra, detect_equalities, check_number);
340 n_eq = info[i].bmap->n_eq + info[j].bmap->n_eq;
341 n_ineq = info[i].bmap->n_ineq + info[j].bmap->n_ineq;
342 fused = isl_basic_map_alloc_space(isl_space_copy(info[i].bmap->dim),
343 info[i].bmap->n_div, n_eq, n_eq + n_ineq + extra_rows);
344 fused = add_valid_constraints(fused, &info[i], 1 + total);
345 fused = add_valid_constraints(fused, &info[j], 1 + total);
346 if (!fused)
347 goto error;
349 for (k = 0; k < info[i].bmap->n_div; ++k) {
350 int l = isl_basic_map_alloc_div(fused);
351 if (l < 0)
352 goto error;
353 isl_seq_cpy(fused->div[l], info[i].bmap->div[k], 1 + 1 + total);
356 for (k = 0; k < extra_rows; ++k) {
357 l = isl_basic_map_alloc_inequality(fused);
358 if (l < 0)
359 goto error;
360 isl_seq_cpy(fused->ineq[l], extra->row[k], 1 + total);
363 if (detect_equalities)
364 fused = isl_basic_map_detect_inequality_pairs(fused, NULL);
365 fused = isl_basic_map_gauss(fused, NULL);
366 ISL_F_SET(fused, ISL_BASIC_MAP_FINAL);
367 if (ISL_F_ISSET(info[i].bmap, ISL_BASIC_MAP_RATIONAL) &&
368 ISL_F_ISSET(info[j].bmap, ISL_BASIC_MAP_RATIONAL))
369 ISL_F_SET(fused, ISL_BASIC_MAP_RATIONAL);
371 fused_tab = isl_tab_from_basic_map(fused, 0);
372 if (isl_tab_detect_redundant(fused_tab) < 0)
373 goto error;
375 if (check_number &&
376 number_of_constraints_increases(i, j, info, fused, fused_tab)) {
377 isl_tab_free(fused_tab);
378 isl_basic_map_free(fused);
379 return isl_change_none;
382 isl_basic_map_free(info[i].bmap);
383 info[i].bmap = fused;
384 isl_tab_free(info[i].tab);
385 info[i].tab = fused_tab;
386 drop(&info[j]);
388 return isl_change_fuse;
389 error:
390 isl_tab_free(fused_tab);
391 isl_basic_map_free(fused);
392 return isl_change_error;
395 /* Given a pair of basic maps i and j such that all constraints are either
396 * "valid" or "cut", check if the facets corresponding to the "cut"
397 * constraints of i lie entirely within basic map j.
398 * If so, replace the pair by the basic map consisting of the valid
399 * constraints in both basic maps.
400 * Checking whether the facet lies entirely within basic map j
401 * is performed by checking whether the constraints of basic map j
402 * are valid for the facet. These tests are performed on a rational
403 * tableau to avoid the theoretical possibility that a constraint
404 * that was considered to be a cut constraint for the entire basic map i
405 * happens to be considered to be a valid constraint for the facet,
406 * even though it cuts off the same rational points.
408 * To see that we are not introducing any extra points, call the
409 * two basic maps A and B and the resulting map U and let x
410 * be an element of U \setminus ( A \cup B ).
411 * A line connecting x with an element of A \cup B meets a facet F
412 * of either A or B. Assume it is a facet of B and let c_1 be
413 * the corresponding facet constraint. We have c_1(x) < 0 and
414 * so c_1 is a cut constraint. This implies that there is some
415 * (possibly rational) point x' satisfying the constraints of A
416 * and the opposite of c_1 as otherwise c_1 would have been marked
417 * valid for A. The line connecting x and x' meets a facet of A
418 * in a (possibly rational) point that also violates c_1, but this
419 * is impossible since all cut constraints of B are valid for all
420 * cut facets of A.
421 * In case F is a facet of A rather than B, then we can apply the
422 * above reasoning to find a facet of B separating x from A \cup B first.
424 static enum isl_change check_facets(int i, int j,
425 struct isl_coalesce_info *info)
427 int k, l;
428 struct isl_tab_undo *snap, *snap2;
429 unsigned n_eq = info[i].bmap->n_eq;
431 snap = isl_tab_snap(info[i].tab);
432 if (isl_tab_mark_rational(info[i].tab) < 0)
433 return isl_change_error;
434 snap2 = isl_tab_snap(info[i].tab);
436 for (k = 0; k < info[i].bmap->n_ineq; ++k) {
437 if (info[i].ineq[k] != STATUS_CUT)
438 continue;
439 if (isl_tab_select_facet(info[i].tab, n_eq + k) < 0)
440 return isl_change_error;
441 for (l = 0; l < info[j].bmap->n_ineq; ++l) {
442 int stat;
443 if (info[j].ineq[l] != STATUS_CUT)
444 continue;
445 stat = status_in(info[j].bmap->ineq[l], info[i].tab);
446 if (stat != STATUS_VALID)
447 break;
449 if (isl_tab_rollback(info[i].tab, snap2) < 0)
450 return isl_change_error;
451 if (l < info[j].bmap->n_ineq)
452 break;
455 if (k < info[i].bmap->n_ineq) {
456 if (isl_tab_rollback(info[i].tab, snap) < 0)
457 return isl_change_error;
458 return isl_change_none;
460 return fuse(i, j, info, NULL, 0, 0);
463 /* Check if info->bmap contains the basic map represented
464 * by the tableau "tab".
465 * For each equality, we check both the constraint itself
466 * (as an inequality) and its negation. Make sure the
467 * equality is returned to its original state before returning.
469 static int contains(struct isl_coalesce_info *info, struct isl_tab *tab)
471 int k;
472 unsigned dim;
473 isl_basic_map *bmap = info->bmap;
475 dim = isl_basic_map_total_dim(bmap);
476 for (k = 0; k < bmap->n_eq; ++k) {
477 int stat;
478 isl_seq_neg(bmap->eq[k], bmap->eq[k], 1 + dim);
479 stat = status_in(bmap->eq[k], tab);
480 isl_seq_neg(bmap->eq[k], bmap->eq[k], 1 + dim);
481 if (stat != STATUS_VALID)
482 return 0;
483 stat = status_in(bmap->eq[k], tab);
484 if (stat != STATUS_VALID)
485 return 0;
488 for (k = 0; k < bmap->n_ineq; ++k) {
489 int stat;
490 if (info->ineq[k] == STATUS_REDUNDANT)
491 continue;
492 stat = status_in(bmap->ineq[k], tab);
493 if (stat != STATUS_VALID)
494 return 0;
496 return 1;
499 /* Basic map "i" has an inequality (say "k") that is adjacent
500 * to some inequality of basic map "j". All the other inequalities
501 * are valid for "j".
502 * Check if basic map "j" forms an extension of basic map "i".
504 * Note that this function is only called if some of the equalities or
505 * inequalities of basic map "j" do cut basic map "i". The function is
506 * correct even if there are no such cut constraints, but in that case
507 * the additional checks performed by this function are overkill.
509 * In particular, we replace constraint k, say f >= 0, by constraint
510 * f <= -1, add the inequalities of "j" that are valid for "i"
511 * and check if the result is a subset of basic map "j".
512 * If so, then we know that this result is exactly equal to basic map "j"
513 * since all its constraints are valid for basic map "j".
514 * By combining the valid constraints of "i" (all equalities and all
515 * inequalities except "k") and the valid constraints of "j" we therefore
516 * obtain a basic map that is equal to their union.
517 * In this case, there is no need to perform a rollback of the tableau
518 * since it is going to be destroyed in fuse().
521 * |\__ |\__
522 * | \__ | \__
523 * | \_ => | \__
524 * |_______| _ |_________\
527 * |\ |\
528 * | \ | \
529 * | \ | \
530 * | | | \
531 * | ||\ => | \
532 * | || \ | \
533 * | || | | |
534 * |__||_/ |_____/
536 static enum isl_change is_adj_ineq_extension(int i, int j,
537 struct isl_coalesce_info *info)
539 int k;
540 struct isl_tab_undo *snap;
541 unsigned n_eq = info[i].bmap->n_eq;
542 unsigned total = isl_basic_map_total_dim(info[i].bmap);
543 int r;
545 if (isl_tab_extend_cons(info[i].tab, 1 + info[j].bmap->n_ineq) < 0)
546 return isl_change_error;
548 for (k = 0; k < info[i].bmap->n_ineq; ++k)
549 if (info[i].ineq[k] == STATUS_ADJ_INEQ)
550 break;
551 if (k >= info[i].bmap->n_ineq)
552 isl_die(isl_basic_map_get_ctx(info[i].bmap), isl_error_internal,
553 "info[i].ineq should have exactly one STATUS_ADJ_INEQ",
554 return isl_change_error);
556 snap = isl_tab_snap(info[i].tab);
558 if (isl_tab_unrestrict(info[i].tab, n_eq + k) < 0)
559 return isl_change_error;
561 isl_seq_neg(info[i].bmap->ineq[k], info[i].bmap->ineq[k], 1 + total);
562 isl_int_sub_ui(info[i].bmap->ineq[k][0], info[i].bmap->ineq[k][0], 1);
563 r = isl_tab_add_ineq(info[i].tab, info[i].bmap->ineq[k]);
564 isl_seq_neg(info[i].bmap->ineq[k], info[i].bmap->ineq[k], 1 + total);
565 isl_int_sub_ui(info[i].bmap->ineq[k][0], info[i].bmap->ineq[k][0], 1);
566 if (r < 0)
567 return isl_change_error;
569 for (k = 0; k < info[j].bmap->n_ineq; ++k) {
570 if (info[j].ineq[k] != STATUS_VALID)
571 continue;
572 if (isl_tab_add_ineq(info[i].tab, info[j].bmap->ineq[k]) < 0)
573 return isl_change_error;
576 if (contains(&info[j], info[i].tab))
577 return fuse(i, j, info, NULL, 0, 0);
579 if (isl_tab_rollback(info[i].tab, snap) < 0)
580 return isl_change_error;
582 return isl_change_none;
586 /* Both basic maps have at least one inequality with and adjacent
587 * (but opposite) inequality in the other basic map.
588 * Check that there are no cut constraints and that there is only
589 * a single pair of adjacent inequalities.
590 * If so, we can replace the pair by a single basic map described
591 * by all but the pair of adjacent inequalities.
592 * Any additional points introduced lie strictly between the two
593 * adjacent hyperplanes and can therefore be integral.
595 * ____ _____
596 * / ||\ / \
597 * / || \ / \
598 * \ || \ => \ \
599 * \ || / \ /
600 * \___||_/ \_____/
602 * The test for a single pair of adjancent inequalities is important
603 * for avoiding the combination of two basic maps like the following
605 * /|
606 * / |
607 * /__|
608 * _____
609 * | |
610 * | |
611 * |___|
613 * If there are some cut constraints on one side, then we may
614 * still be able to fuse the two basic maps, but we need to perform
615 * some additional checks in is_adj_ineq_extension.
617 static enum isl_change check_adj_ineq(int i, int j,
618 struct isl_coalesce_info *info)
620 int count_i, count_j;
621 int cut_i, cut_j;
623 count_i = count(info[i].ineq, info[i].bmap->n_ineq, STATUS_ADJ_INEQ);
624 count_j = count(info[j].ineq, info[j].bmap->n_ineq, STATUS_ADJ_INEQ);
626 if (count_i != 1 && count_j != 1)
627 return isl_change_none;
629 cut_i = any(info[i].eq, 2 * info[i].bmap->n_eq, STATUS_CUT) ||
630 any(info[i].ineq, info[i].bmap->n_ineq, STATUS_CUT);
631 cut_j = any(info[j].eq, 2 * info[j].bmap->n_eq, STATUS_CUT) ||
632 any(info[j].ineq, info[j].bmap->n_ineq, STATUS_CUT);
634 if (!cut_i && !cut_j && count_i == 1 && count_j == 1)
635 return fuse(i, j, info, NULL, 0, 0);
637 if (count_i == 1 && !cut_i)
638 return is_adj_ineq_extension(i, j, info);
640 if (count_j == 1 && !cut_j)
641 return is_adj_ineq_extension(j, i, info);
643 return isl_change_none;
646 /* Basic map "i" has an inequality "k" that is adjacent to some equality
647 * of basic map "j". All the other inequalities are valid for "j".
648 * Check if basic map "j" forms an extension of basic map "i".
650 * In particular, we relax constraint "k", compute the corresponding
651 * facet and check whether it is included in the other basic map.
652 * If so, we know that relaxing the constraint extends the basic
653 * map with exactly the other basic map (we already know that this
654 * other basic map is included in the extension, because there
655 * were no "cut" inequalities in "i") and we can replace the
656 * two basic maps by this extension.
657 * Place this extension in the position that is the smallest of i and j.
658 * ____ _____
659 * / || / |
660 * / || / |
661 * \ || => \ |
662 * \ || \ |
663 * \___|| \____|
665 static enum isl_change is_adj_eq_extension(int i, int j, int k,
666 struct isl_coalesce_info *info)
668 int change = isl_change_none;
669 int super;
670 struct isl_tab_undo *snap, *snap2;
671 unsigned n_eq = info[i].bmap->n_eq;
673 if (isl_tab_is_equality(info[i].tab, n_eq + k))
674 return isl_change_none;
676 snap = isl_tab_snap(info[i].tab);
677 if (isl_tab_relax(info[i].tab, n_eq + k) < 0)
678 return isl_change_error;
679 snap2 = isl_tab_snap(info[i].tab);
680 if (isl_tab_select_facet(info[i].tab, n_eq + k) < 0)
681 return isl_change_error;
682 super = contains(&info[j], info[i].tab);
683 if (super) {
684 if (isl_tab_rollback(info[i].tab, snap2) < 0)
685 return isl_change_error;
686 info[i].bmap = isl_basic_map_cow(info[i].bmap);
687 if (!info[i].bmap)
688 return isl_change_error;
689 isl_int_add_ui(info[i].bmap->ineq[k][0],
690 info[i].bmap->ineq[k][0], 1);
691 ISL_F_SET(info[i].bmap, ISL_BASIC_MAP_FINAL);
692 drop(&info[j]);
693 if (j < i)
694 exchange(&info[i], &info[j]);
695 change = isl_change_fuse;
696 } else
697 if (isl_tab_rollback(info[i].tab, snap) < 0)
698 return isl_change_error;
700 return change;
703 /* Data structure that keeps track of the wrapping constraints
704 * and of information to bound the coefficients of those constraints.
706 * bound is set if we want to apply a bound on the coefficients
707 * mat contains the wrapping constraints
708 * max is the bound on the coefficients (if bound is set)
710 struct isl_wraps {
711 int bound;
712 isl_mat *mat;
713 isl_int max;
716 /* Update wraps->max to be greater than or equal to the coefficients
717 * in the equalities and inequalities of info->bmap that can be removed
718 * if we end up applying wrapping.
720 static void wraps_update_max(struct isl_wraps *wraps,
721 struct isl_coalesce_info *info)
723 int k;
724 isl_int max_k;
725 unsigned total = isl_basic_map_total_dim(info->bmap);
727 isl_int_init(max_k);
729 for (k = 0; k < info->bmap->n_eq; ++k) {
730 if (info->eq[2 * k] == STATUS_VALID &&
731 info->eq[2 * k + 1] == STATUS_VALID)
732 continue;
733 isl_seq_abs_max(info->bmap->eq[k] + 1, total, &max_k);
734 if (isl_int_abs_gt(max_k, wraps->max))
735 isl_int_set(wraps->max, max_k);
738 for (k = 0; k < info->bmap->n_ineq; ++k) {
739 if (info->ineq[k] == STATUS_VALID ||
740 info->ineq[k] == STATUS_REDUNDANT)
741 continue;
742 isl_seq_abs_max(info->bmap->ineq[k] + 1, total, &max_k);
743 if (isl_int_abs_gt(max_k, wraps->max))
744 isl_int_set(wraps->max, max_k);
747 isl_int_clear(max_k);
750 /* Initialize the isl_wraps data structure.
751 * If we want to bound the coefficients of the wrapping constraints,
752 * we set wraps->max to the largest coefficient
753 * in the equalities and inequalities that can be removed if we end up
754 * applying wrapping.
756 static void wraps_init(struct isl_wraps *wraps, __isl_take isl_mat *mat,
757 struct isl_coalesce_info *info, int i, int j)
759 isl_ctx *ctx;
761 wraps->bound = 0;
762 wraps->mat = mat;
763 if (!mat)
764 return;
765 ctx = isl_mat_get_ctx(mat);
766 wraps->bound = isl_options_get_coalesce_bounded_wrapping(ctx);
767 if (!wraps->bound)
768 return;
769 isl_int_init(wraps->max);
770 isl_int_set_si(wraps->max, 0);
771 wraps_update_max(wraps, &info[i]);
772 wraps_update_max(wraps, &info[j]);
775 /* Free the contents of the isl_wraps data structure.
777 static void wraps_free(struct isl_wraps *wraps)
779 isl_mat_free(wraps->mat);
780 if (wraps->bound)
781 isl_int_clear(wraps->max);
784 /* Is the wrapping constraint in row "row" allowed?
786 * If wraps->bound is set, we check that none of the coefficients
787 * is greater than wraps->max.
789 static int allow_wrap(struct isl_wraps *wraps, int row)
791 int i;
793 if (!wraps->bound)
794 return 1;
796 for (i = 1; i < wraps->mat->n_col; ++i)
797 if (isl_int_abs_gt(wraps->mat->row[row][i], wraps->max))
798 return 0;
800 return 1;
803 /* Wrap "ineq" (or its opposite if "negate" is set) around "bound"
804 * to include "set" and add the result in position "w" of "wraps".
805 * "len" is the total number of coefficients in "bound" and "ineq".
806 * Return 1 on success, 0 on failure and -1 on error.
807 * Wrapping can fail if the result of wrapping is equal to "bound"
808 * or if we want to bound the sizes of the coefficients and
809 * the wrapped constraint does not satisfy this bound.
811 static int add_wrap(struct isl_wraps *wraps, int w, isl_int *bound,
812 isl_int *ineq, unsigned len, __isl_keep isl_set *set, int negate)
814 isl_seq_cpy(wraps->mat->row[w], bound, len);
815 if (negate) {
816 isl_seq_neg(wraps->mat->row[w + 1], ineq, len);
817 ineq = wraps->mat->row[w + 1];
819 if (!isl_set_wrap_facet(set, wraps->mat->row[w], ineq))
820 return -1;
821 if (isl_seq_eq(wraps->mat->row[w], bound, len))
822 return 0;
823 if (!allow_wrap(wraps, w))
824 return 0;
825 return 1;
828 /* For each constraint in info->bmap that is not redundant (as determined
829 * by info->tab) and that is not a valid constraint for the other basic map,
830 * wrap the constraint around "bound" such that it includes the whole
831 * set "set" and append the resulting constraint to "wraps".
832 * Note that the constraints that are valid for the other basic map
833 * will be added to the combined basic map by default, so there is
834 * no need to wrap them.
835 * The caller wrap_in_facets even relies on this function not wrapping
836 * any constraints that are already valid.
837 * "wraps" is assumed to have been pre-allocated to the appropriate size.
838 * wraps->n_row is the number of actual wrapped constraints that have
839 * been added.
840 * If any of the wrapping problems results in a constraint that is
841 * identical to "bound", then this means that "set" is unbounded in such
842 * way that no wrapping is possible. If this happens then wraps->n_row
843 * is reset to zero.
844 * Similarly, if we want to bound the coefficients of the wrapping
845 * constraints and a newly added wrapping constraint does not
846 * satisfy the bound, then wraps->n_row is also reset to zero.
848 static int add_wraps(struct isl_wraps *wraps, struct isl_coalesce_info *info,
849 isl_int *bound, __isl_keep isl_set *set)
851 int l, m;
852 int w;
853 int added;
854 isl_basic_map *bmap = info->bmap;
855 unsigned len = 1 + isl_basic_map_total_dim(bmap);
857 w = wraps->mat->n_row;
859 for (l = 0; l < bmap->n_ineq; ++l) {
860 if (info->ineq[l] == STATUS_VALID ||
861 info->ineq[l] == STATUS_REDUNDANT)
862 continue;
863 if (isl_seq_is_neg(bound, bmap->ineq[l], len))
864 continue;
865 if (isl_seq_eq(bound, bmap->ineq[l], len))
866 continue;
867 if (isl_tab_is_redundant(info->tab, bmap->n_eq + l))
868 continue;
870 added = add_wrap(wraps, w, bound, bmap->ineq[l], len, set, 0);
871 if (added < 0)
872 return -1;
873 if (!added)
874 goto unbounded;
875 ++w;
877 for (l = 0; l < bmap->n_eq; ++l) {
878 if (isl_seq_is_neg(bound, bmap->eq[l], len))
879 continue;
880 if (isl_seq_eq(bound, bmap->eq[l], len))
881 continue;
883 for (m = 0; m < 2; ++m) {
884 if (info->eq[2 * l + m] == STATUS_VALID)
885 continue;
886 added = add_wrap(wraps, w, bound, bmap->eq[l], len,
887 set, !m);
888 if (added < 0)
889 return -1;
890 if (!added)
891 goto unbounded;
892 ++w;
896 wraps->mat->n_row = w;
897 return 0;
898 unbounded:
899 wraps->mat->n_row = 0;
900 return 0;
903 /* Check if the constraints in "wraps" from "first" until the last
904 * are all valid for the basic set represented by "tab".
905 * If not, wraps->n_row is set to zero.
907 static int check_wraps(__isl_keep isl_mat *wraps, int first,
908 struct isl_tab *tab)
910 int i;
912 for (i = first; i < wraps->n_row; ++i) {
913 enum isl_ineq_type type;
914 type = isl_tab_ineq_type(tab, wraps->row[i]);
915 if (type == isl_ineq_error)
916 return -1;
917 if (type == isl_ineq_redundant)
918 continue;
919 wraps->n_row = 0;
920 return 0;
923 return 0;
926 /* Return a set that corresponds to the non-redundant constraints
927 * (as recorded in tab) of bmap.
929 * It's important to remove the redundant constraints as some
930 * of the other constraints may have been modified after the
931 * constraints were marked redundant.
932 * In particular, a constraint may have been relaxed.
933 * Redundant constraints are ignored when a constraint is relaxed
934 * and should therefore continue to be ignored ever after.
935 * Otherwise, the relaxation might be thwarted by some of
936 * these constraints.
938 * Update the underlying set to ensure that the dimension doesn't change.
939 * Otherwise the integer divisions could get dropped if the tab
940 * turns out to be empty.
942 static __isl_give isl_set *set_from_updated_bmap(__isl_keep isl_basic_map *bmap,
943 struct isl_tab *tab)
945 isl_basic_set *bset;
947 bmap = isl_basic_map_copy(bmap);
948 bset = isl_basic_map_underlying_set(bmap);
949 bset = isl_basic_set_cow(bset);
950 bset = isl_basic_set_update_from_tab(bset, tab);
951 return isl_set_from_basic_set(bset);
954 /* Wrap the constraints of info->bmap that bound the facet defined
955 * by inequality "k" around (the opposite of) this inequality to
956 * include "set". "bound" may be used to store the negated inequality.
957 * Since the wrapped constraints are not guaranteed to contain the whole
958 * of info->bmap, we check them in check_wraps.
959 * If any of the wrapped constraints turn out to be invalid, then
960 * check_wraps will reset wrap->n_row to zero.
962 static int add_wraps_around_facet(struct isl_wraps *wraps,
963 struct isl_coalesce_info *info, int k, isl_int *bound,
964 __isl_keep isl_set *set)
966 struct isl_tab_undo *snap;
967 int n;
968 unsigned total = isl_basic_map_total_dim(info->bmap);
970 snap = isl_tab_snap(info->tab);
972 if (isl_tab_select_facet(info->tab, info->bmap->n_eq + k) < 0)
973 return -1;
974 if (isl_tab_detect_redundant(info->tab) < 0)
975 return -1;
977 isl_seq_neg(bound, info->bmap->ineq[k], 1 + total);
979 n = wraps->mat->n_row;
980 if (add_wraps(wraps, info, bound, set) < 0)
981 return -1;
983 if (isl_tab_rollback(info->tab, snap) < 0)
984 return -1;
985 if (check_wraps(wraps->mat, n, info->tab) < 0)
986 return -1;
988 return 0;
991 /* Given a basic set i with a constraint k that is adjacent to
992 * basic set j, check if we can wrap
993 * both the facet corresponding to k (if "wrap_facet" is set) and basic map j
994 * (always) around their ridges to include the other set.
995 * If so, replace the pair of basic sets by their union.
997 * All constraints of i (except k) are assumed to be valid or
998 * cut constraints for j.
999 * Wrapping the cut constraints to include basic map j may result
1000 * in constraints that are no longer valid of basic map i
1001 * we have to check that the resulting wrapping constraints are valid for i.
1002 * If "wrap_facet" is not set, then all constraints of i (except k)
1003 * are assumed to be valid for j.
1004 * ____ _____
1005 * / | / \
1006 * / || / |
1007 * \ || => \ |
1008 * \ || \ |
1009 * \___|| \____|
1012 static enum isl_change can_wrap_in_facet(int i, int j, int k,
1013 struct isl_coalesce_info *info, int wrap_facet)
1015 enum isl_change change = isl_change_none;
1016 struct isl_wraps wraps;
1017 isl_ctx *ctx;
1018 isl_mat *mat;
1019 struct isl_set *set_i = NULL;
1020 struct isl_set *set_j = NULL;
1021 struct isl_vec *bound = NULL;
1022 unsigned total = isl_basic_map_total_dim(info[i].bmap);
1024 set_i = set_from_updated_bmap(info[i].bmap, info[i].tab);
1025 set_j = set_from_updated_bmap(info[j].bmap, info[j].tab);
1026 ctx = isl_basic_map_get_ctx(info[i].bmap);
1027 mat = isl_mat_alloc(ctx, 2 * (info[i].bmap->n_eq + info[j].bmap->n_eq) +
1028 info[i].bmap->n_ineq + info[j].bmap->n_ineq,
1029 1 + total);
1030 wraps_init(&wraps, mat, info, i, j);
1031 bound = isl_vec_alloc(ctx, 1 + total);
1032 if (!set_i || !set_j || !wraps.mat || !bound)
1033 goto error;
1035 isl_seq_cpy(bound->el, info[i].bmap->ineq[k], 1 + total);
1036 isl_int_add_ui(bound->el[0], bound->el[0], 1);
1038 isl_seq_cpy(wraps.mat->row[0], bound->el, 1 + total);
1039 wraps.mat->n_row = 1;
1041 if (add_wraps(&wraps, &info[j], bound->el, set_i) < 0)
1042 goto error;
1043 if (!wraps.mat->n_row)
1044 goto unbounded;
1046 if (wrap_facet) {
1047 if (add_wraps_around_facet(&wraps, &info[i], k,
1048 bound->el, set_j) < 0)
1049 goto error;
1050 if (!wraps.mat->n_row)
1051 goto unbounded;
1054 change = fuse(i, j, info, wraps.mat, 0, 0);
1056 unbounded:
1057 wraps_free(&wraps);
1059 isl_set_free(set_i);
1060 isl_set_free(set_j);
1062 isl_vec_free(bound);
1064 return change;
1065 error:
1066 wraps_free(&wraps);
1067 isl_vec_free(bound);
1068 isl_set_free(set_i);
1069 isl_set_free(set_j);
1070 return isl_change_error;
1073 /* Given a pair of basic maps i and j such that j sticks out
1074 * of i at n cut constraints, each time by at most one,
1075 * try to compute wrapping constraints and replace the two
1076 * basic maps by a single basic map.
1077 * The other constraints of i are assumed to be valid for j.
1079 * For each cut constraint t(x) >= 0 of i, we add the relaxed version
1080 * t(x) + 1 >= 0, along with wrapping constraints for all constraints
1081 * of basic map j that bound the part of basic map j that sticks out
1082 * of the cut constraint.
1083 * In particular, we first intersect basic map j with t(x) + 1 = 0.
1084 * If the result is empty, then t(x) >= 0 was actually a valid constraint
1085 * (with respect to the integer points), so we add t(x) >= 0 instead.
1086 * Otherwise, we wrap the constraints of basic map j that are not
1087 * redundant in this intersection and that are not already valid
1088 * for basic map i over basic map i.
1089 * Note that it is sufficient to wrap the constraints to include
1090 * basic map i, because we will only wrap the constraints that do
1091 * not include basic map i already. The wrapped constraint will
1092 * therefore be more relaxed compared to the original constraint.
1093 * Since the original constraint is valid for basic map j, so is
1094 * the wrapped constraint.
1096 * If any wrapping fails, i.e., if we cannot wrap to touch
1097 * the union, then we give up.
1098 * Otherwise, the pair of basic maps is replaced by their union.
1100 static enum isl_change wrap_in_facets(int i, int j, int *cuts, int n,
1101 struct isl_coalesce_info *info)
1103 enum isl_change change = isl_change_none;
1104 struct isl_wraps wraps;
1105 isl_ctx *ctx;
1106 isl_mat *mat;
1107 isl_set *set_i = NULL;
1108 unsigned total = isl_basic_map_total_dim(info[i].bmap);
1109 int max_wrap;
1110 int k, w;
1111 struct isl_tab_undo *snap;
1113 if (isl_tab_extend_cons(info[j].tab, 1) < 0)
1114 goto error;
1116 max_wrap = 1 + 2 * info[j].bmap->n_eq + info[j].bmap->n_ineq;
1117 max_wrap *= n;
1119 set_i = set_from_updated_bmap(info[i].bmap, info[i].tab);
1120 ctx = isl_basic_map_get_ctx(info[i].bmap);
1121 mat = isl_mat_alloc(ctx, max_wrap, 1 + total);
1122 wraps_init(&wraps, mat, info, i, j);
1123 if (!set_i || !wraps.mat)
1124 goto error;
1126 snap = isl_tab_snap(info[j].tab);
1128 wraps.mat->n_row = 0;
1130 for (k = 0; k < n; ++k) {
1131 w = wraps.mat->n_row++;
1132 isl_seq_cpy(wraps.mat->row[w],
1133 info[i].bmap->ineq[cuts[k]], 1 + total);
1134 isl_int_add_ui(wraps.mat->row[w][0], wraps.mat->row[w][0], 1);
1135 if (isl_tab_add_eq(info[j].tab, wraps.mat->row[w]) < 0)
1136 goto error;
1137 if (isl_tab_detect_redundant(info[j].tab) < 0)
1138 goto error;
1140 if (info[j].tab->empty)
1141 isl_int_sub_ui(wraps.mat->row[w][0],
1142 wraps.mat->row[w][0], 1);
1143 else if (add_wraps(&wraps, &info[j],
1144 wraps.mat->row[w], set_i) < 0)
1145 goto error;
1147 if (isl_tab_rollback(info[j].tab, snap) < 0)
1148 goto error;
1150 if (!wraps.mat->n_row)
1151 break;
1154 if (k == n)
1155 change = fuse(i, j, info, wraps.mat, 0, 1);
1157 wraps_free(&wraps);
1158 isl_set_free(set_i);
1160 return change;
1161 error:
1162 wraps_free(&wraps);
1163 isl_set_free(set_i);
1164 return isl_change_error;
1167 /* Given two basic sets i and j such that i has no cut equalities,
1168 * check if relaxing all the cut inequalities of i by one turns
1169 * them into valid constraint for j and check if we can wrap in
1170 * the bits that are sticking out.
1171 * If so, replace the pair by their union.
1173 * We first check if all relaxed cut inequalities of i are valid for j
1174 * and then try to wrap in the intersections of the relaxed cut inequalities
1175 * with j.
1177 * During this wrapping, we consider the points of j that lie at a distance
1178 * of exactly 1 from i. In particular, we ignore the points that lie in
1179 * between this lower-dimensional space and the basic map i.
1180 * We can therefore only apply this to integer maps.
1181 * ____ _____
1182 * / ___|_ / \
1183 * / | | / |
1184 * \ | | => \ |
1185 * \|____| \ |
1186 * \___| \____/
1188 * _____ ______
1189 * | ____|_ | \
1190 * | | | | |
1191 * | | | => | |
1192 * |_| | | |
1193 * |_____| \______|
1195 * _______
1196 * | |
1197 * | |\ |
1198 * | | \ |
1199 * | | \ |
1200 * | | \|
1201 * | | \
1202 * | |_____\
1203 * | |
1204 * |_______|
1206 * Wrapping can fail if the result of wrapping one of the facets
1207 * around its edges does not produce any new facet constraint.
1208 * In particular, this happens when we try to wrap in unbounded sets.
1210 * _______________________________________________________________________
1212 * | ___
1213 * | | |
1214 * |_| |_________________________________________________________________
1215 * |___|
1217 * The following is not an acceptable result of coalescing the above two
1218 * sets as it includes extra integer points.
1219 * _______________________________________________________________________
1221 * |
1222 * |
1224 * \______________________________________________________________________
1226 static enum isl_change can_wrap_in_set(int i, int j,
1227 struct isl_coalesce_info *info)
1229 enum isl_change change = isl_change_none;
1230 int k, m;
1231 int n;
1232 int *cuts = NULL;
1233 isl_ctx *ctx;
1235 if (ISL_F_ISSET(info[i].bmap, ISL_BASIC_MAP_RATIONAL) ||
1236 ISL_F_ISSET(info[j].bmap, ISL_BASIC_MAP_RATIONAL))
1237 return isl_change_none;
1239 n = count(info[i].ineq, info[i].bmap->n_ineq, STATUS_CUT);
1240 if (n == 0)
1241 return isl_change_none;
1243 ctx = isl_basic_map_get_ctx(info[i].bmap);
1244 cuts = isl_alloc_array(ctx, int, n);
1245 if (!cuts)
1246 return isl_change_error;
1248 for (k = 0, m = 0; m < n; ++k) {
1249 enum isl_ineq_type type;
1251 if (info[i].ineq[k] != STATUS_CUT)
1252 continue;
1254 isl_int_add_ui(info[i].bmap->ineq[k][0],
1255 info[i].bmap->ineq[k][0], 1);
1256 type = isl_tab_ineq_type(info[j].tab, info[i].bmap->ineq[k]);
1257 isl_int_sub_ui(info[i].bmap->ineq[k][0],
1258 info[i].bmap->ineq[k][0], 1);
1259 if (type == isl_ineq_error)
1260 goto error;
1261 if (type != isl_ineq_redundant)
1262 break;
1263 cuts[m] = k;
1264 ++m;
1267 if (m == n)
1268 change = wrap_in_facets(i, j, cuts, n, info);
1270 free(cuts);
1272 return change;
1273 error:
1274 free(cuts);
1275 return isl_change_error;
1278 /* Check if either i or j has only cut inequalities that can
1279 * be used to wrap in (a facet of) the other basic set.
1280 * if so, replace the pair by their union.
1282 static enum isl_change check_wrap(int i, int j, struct isl_coalesce_info *info)
1284 enum isl_change change = isl_change_none;
1286 if (!any(info[i].eq, 2 * info[i].bmap->n_eq, STATUS_CUT))
1287 change = can_wrap_in_set(i, j, info);
1288 if (change != isl_change_none)
1289 return change;
1291 if (!any(info[j].eq, 2 * info[j].bmap->n_eq, STATUS_CUT))
1292 change = can_wrap_in_set(j, i, info);
1293 return change;
1296 /* At least one of the basic maps has an equality that is adjacent
1297 * to inequality. Make sure that only one of the basic maps has
1298 * such an equality and that the other basic map has exactly one
1299 * inequality adjacent to an equality.
1300 * We call the basic map that has the inequality "i" and the basic
1301 * map that has the equality "j".
1302 * If "i" has any "cut" (in)equality, then relaxing the inequality
1303 * by one would not result in a basic map that contains the other
1304 * basic map. However, it may still be possible to wrap in the other
1305 * basic map.
1307 static enum isl_change check_adj_eq(int i, int j,
1308 struct isl_coalesce_info *info)
1310 enum isl_change change = isl_change_none;
1311 int k;
1312 int any_cut;
1314 if (any(info[i].eq, 2 * info[i].bmap->n_eq, STATUS_ADJ_INEQ) &&
1315 any(info[j].eq, 2 * info[j].bmap->n_eq, STATUS_ADJ_INEQ))
1316 /* ADJ EQ TOO MANY */
1317 return isl_change_none;
1319 if (any(info[i].eq, 2 * info[i].bmap->n_eq, STATUS_ADJ_INEQ))
1320 return check_adj_eq(j, i, info);
1322 /* j has an equality adjacent to an inequality in i */
1324 if (any(info[i].eq, 2 * info[i].bmap->n_eq, STATUS_CUT))
1325 return isl_change_none;
1326 any_cut = any(info[i].ineq, info[i].bmap->n_ineq, STATUS_CUT);
1327 if (count(info[i].ineq, info[i].bmap->n_ineq, STATUS_ADJ_EQ) != 1 ||
1328 any(info[j].ineq, info[j].bmap->n_ineq, STATUS_ADJ_EQ) ||
1329 any(info[i].ineq, info[i].bmap->n_ineq, STATUS_ADJ_INEQ) ||
1330 any(info[j].ineq, info[j].bmap->n_ineq, STATUS_ADJ_INEQ))
1331 /* ADJ EQ TOO MANY */
1332 return isl_change_none;
1334 for (k = 0; k < info[i].bmap->n_ineq; ++k)
1335 if (info[i].ineq[k] == STATUS_ADJ_EQ)
1336 break;
1338 if (!any_cut) {
1339 change = is_adj_eq_extension(i, j, k, info);
1340 if (change != isl_change_none)
1341 return change;
1344 change = can_wrap_in_facet(i, j, k, info, any_cut);
1346 return change;
1349 /* The two basic maps lie on adjacent hyperplanes. In particular,
1350 * basic map "i" has an equality that lies parallel to basic map "j".
1351 * Check if we can wrap the facets around the parallel hyperplanes
1352 * to include the other set.
1354 * We perform basically the same operations as can_wrap_in_facet,
1355 * except that we don't need to select a facet of one of the sets.
1357 * \\ \\
1358 * \\ => \\
1359 * \ \|
1361 * If there is more than one equality of "i" adjacent to an equality of "j",
1362 * then the result will satisfy one or more equalities that are a linear
1363 * combination of these equalities. These will be encoded as pairs
1364 * of inequalities in the wrapping constraints and need to be made
1365 * explicit.
1367 static enum isl_change check_eq_adj_eq(int i, int j,
1368 struct isl_coalesce_info *info)
1370 int k;
1371 enum isl_change change = isl_change_none;
1372 int detect_equalities = 0;
1373 struct isl_wraps wraps;
1374 isl_ctx *ctx;
1375 isl_mat *mat;
1376 struct isl_set *set_i = NULL;
1377 struct isl_set *set_j = NULL;
1378 struct isl_vec *bound = NULL;
1379 unsigned total = isl_basic_map_total_dim(info[i].bmap);
1381 if (count(info[i].eq, 2 * info[i].bmap->n_eq, STATUS_ADJ_EQ) != 1)
1382 detect_equalities = 1;
1384 for (k = 0; k < 2 * info[i].bmap->n_eq ; ++k)
1385 if (info[i].eq[k] == STATUS_ADJ_EQ)
1386 break;
1388 set_i = set_from_updated_bmap(info[i].bmap, info[i].tab);
1389 set_j = set_from_updated_bmap(info[j].bmap, info[j].tab);
1390 ctx = isl_basic_map_get_ctx(info[i].bmap);
1391 mat = isl_mat_alloc(ctx, 2 * (info[i].bmap->n_eq + info[j].bmap->n_eq) +
1392 info[i].bmap->n_ineq + info[j].bmap->n_ineq,
1393 1 + total);
1394 wraps_init(&wraps, mat, info, i, j);
1395 bound = isl_vec_alloc(ctx, 1 + total);
1396 if (!set_i || !set_j || !wraps.mat || !bound)
1397 goto error;
1399 if (k % 2 == 0)
1400 isl_seq_neg(bound->el, info[i].bmap->eq[k / 2], 1 + total);
1401 else
1402 isl_seq_cpy(bound->el, info[i].bmap->eq[k / 2], 1 + total);
1403 isl_int_add_ui(bound->el[0], bound->el[0], 1);
1405 isl_seq_cpy(wraps.mat->row[0], bound->el, 1 + total);
1406 wraps.mat->n_row = 1;
1408 if (add_wraps(&wraps, &info[j], bound->el, set_i) < 0)
1409 goto error;
1410 if (!wraps.mat->n_row)
1411 goto unbounded;
1413 isl_int_sub_ui(bound->el[0], bound->el[0], 1);
1414 isl_seq_neg(bound->el, bound->el, 1 + total);
1416 isl_seq_cpy(wraps.mat->row[wraps.mat->n_row], bound->el, 1 + total);
1417 wraps.mat->n_row++;
1419 if (add_wraps(&wraps, &info[i], bound->el, set_j) < 0)
1420 goto error;
1421 if (!wraps.mat->n_row)
1422 goto unbounded;
1424 change = fuse(i, j, info, wraps.mat, detect_equalities, 0);
1426 if (0) {
1427 error: change = isl_change_error;
1429 unbounded:
1431 wraps_free(&wraps);
1432 isl_set_free(set_i);
1433 isl_set_free(set_j);
1434 isl_vec_free(bound);
1436 return change;
1439 /* Check if the union of the given pair of basic maps
1440 * can be represented by a single basic map.
1441 * If so, replace the pair by the single basic map and return
1442 * isl_change_drop_first, isl_change_drop_second or isl_change_fuse.
1443 * Otherwise, return isl_change_none.
1444 * The two basic maps are assumed to live in the same local space.
1446 * We first check the effect of each constraint of one basic map
1447 * on the other basic map.
1448 * The constraint may be
1449 * redundant the constraint is redundant in its own
1450 * basic map and should be ignore and removed
1451 * in the end
1452 * valid all (integer) points of the other basic map
1453 * satisfy the constraint
1454 * separate no (integer) point of the other basic map
1455 * satisfies the constraint
1456 * cut some but not all points of the other basic map
1457 * satisfy the constraint
1458 * adj_eq the given constraint is adjacent (on the outside)
1459 * to an equality of the other basic map
1460 * adj_ineq the given constraint is adjacent (on the outside)
1461 * to an inequality of the other basic map
1463 * We consider seven cases in which we can replace the pair by a single
1464 * basic map. We ignore all "redundant" constraints.
1466 * 1. all constraints of one basic map are valid
1467 * => the other basic map is a subset and can be removed
1469 * 2. all constraints of both basic maps are either "valid" or "cut"
1470 * and the facets corresponding to the "cut" constraints
1471 * of one of the basic maps lies entirely inside the other basic map
1472 * => the pair can be replaced by a basic map consisting
1473 * of the valid constraints in both basic maps
1475 * 3. there is a single pair of adjacent inequalities
1476 * (all other constraints are "valid")
1477 * => the pair can be replaced by a basic map consisting
1478 * of the valid constraints in both basic maps
1480 * 4. one basic map has a single adjacent inequality, while the other
1481 * constraints are "valid". The other basic map has some
1482 * "cut" constraints, but replacing the adjacent inequality by
1483 * its opposite and adding the valid constraints of the other
1484 * basic map results in a subset of the other basic map
1485 * => the pair can be replaced by a basic map consisting
1486 * of the valid constraints in both basic maps
1488 * 5. there is a single adjacent pair of an inequality and an equality,
1489 * the other constraints of the basic map containing the inequality are
1490 * "valid". Moreover, if the inequality the basic map is relaxed
1491 * and then turned into an equality, then resulting facet lies
1492 * entirely inside the other basic map
1493 * => the pair can be replaced by the basic map containing
1494 * the inequality, with the inequality relaxed.
1496 * 6. there is a single adjacent pair of an inequality and an equality,
1497 * the other constraints of the basic map containing the inequality are
1498 * "valid". Moreover, the facets corresponding to both
1499 * the inequality and the equality can be wrapped around their
1500 * ridges to include the other basic map
1501 * => the pair can be replaced by a basic map consisting
1502 * of the valid constraints in both basic maps together
1503 * with all wrapping constraints
1505 * 7. one of the basic maps extends beyond the other by at most one.
1506 * Moreover, the facets corresponding to the cut constraints and
1507 * the pieces of the other basic map at offset one from these cut
1508 * constraints can be wrapped around their ridges to include
1509 * the union of the two basic maps
1510 * => the pair can be replaced by a basic map consisting
1511 * of the valid constraints in both basic maps together
1512 * with all wrapping constraints
1514 * 8. the two basic maps live in adjacent hyperplanes. In principle
1515 * such sets can always be combined through wrapping, but we impose
1516 * that there is only one such pair, to avoid overeager coalescing.
1518 * Throughout the computation, we maintain a collection of tableaus
1519 * corresponding to the basic maps. When the basic maps are dropped
1520 * or combined, the tableaus are modified accordingly.
1522 static enum isl_change coalesce_local_pair(int i, int j,
1523 struct isl_coalesce_info *info)
1525 enum isl_change change = isl_change_none;
1527 info[i].eq = info[i].ineq = NULL;
1528 info[j].eq = info[j].ineq = NULL;
1530 info[i].eq = eq_status_in(info[i].bmap, info[j].tab);
1531 if (info[i].bmap->n_eq && !info[i].eq)
1532 goto error;
1533 if (any(info[i].eq, 2 * info[i].bmap->n_eq, STATUS_ERROR))
1534 goto error;
1535 if (any(info[i].eq, 2 * info[i].bmap->n_eq, STATUS_SEPARATE))
1536 goto done;
1538 info[j].eq = eq_status_in(info[j].bmap, info[i].tab);
1539 if (info[j].bmap->n_eq && !info[j].eq)
1540 goto error;
1541 if (any(info[j].eq, 2 * info[j].bmap->n_eq, STATUS_ERROR))
1542 goto error;
1543 if (any(info[j].eq, 2 * info[j].bmap->n_eq, STATUS_SEPARATE))
1544 goto done;
1546 info[i].ineq = ineq_status_in(info[i].bmap, info[i].tab, info[j].tab);
1547 if (info[i].bmap->n_ineq && !info[i].ineq)
1548 goto error;
1549 if (any(info[i].ineq, info[i].bmap->n_ineq, STATUS_ERROR))
1550 goto error;
1551 if (any(info[i].ineq, info[i].bmap->n_ineq, STATUS_SEPARATE))
1552 goto done;
1554 info[j].ineq = ineq_status_in(info[j].bmap, info[j].tab, info[i].tab);
1555 if (info[j].bmap->n_ineq && !info[j].ineq)
1556 goto error;
1557 if (any(info[j].ineq, info[j].bmap->n_ineq, STATUS_ERROR))
1558 goto error;
1559 if (any(info[j].ineq, info[j].bmap->n_ineq, STATUS_SEPARATE))
1560 goto done;
1562 if (all(info[i].eq, 2 * info[i].bmap->n_eq, STATUS_VALID) &&
1563 all(info[i].ineq, info[i].bmap->n_ineq, STATUS_VALID)) {
1564 drop(&info[j]);
1565 change = isl_change_drop_second;
1566 } else if (all(info[j].eq, 2 * info[j].bmap->n_eq, STATUS_VALID) &&
1567 all(info[j].ineq, info[j].bmap->n_ineq, STATUS_VALID)) {
1568 drop(&info[i]);
1569 change = isl_change_drop_first;
1570 } else if (any(info[i].eq, 2 * info[i].bmap->n_eq, STATUS_ADJ_EQ)) {
1571 change = check_eq_adj_eq(i, j, info);
1572 } else if (any(info[j].eq, 2 * info[j].bmap->n_eq, STATUS_ADJ_EQ)) {
1573 change = check_eq_adj_eq(j, i, info);
1574 } else if (any(info[i].eq, 2 * info[i].bmap->n_eq, STATUS_ADJ_INEQ) ||
1575 any(info[j].eq, 2 * info[j].bmap->n_eq, STATUS_ADJ_INEQ)) {
1576 change = check_adj_eq(i, j, info);
1577 } else if (any(info[i].ineq, info[i].bmap->n_ineq, STATUS_ADJ_EQ) ||
1578 any(info[j].ineq, info[j].bmap->n_ineq, STATUS_ADJ_EQ)) {
1579 /* Can't happen */
1580 /* BAD ADJ INEQ */
1581 } else if (any(info[i].ineq, info[i].bmap->n_ineq, STATUS_ADJ_INEQ) ||
1582 any(info[j].ineq, info[j].bmap->n_ineq, STATUS_ADJ_INEQ)) {
1583 change = check_adj_ineq(i, j, info);
1584 } else {
1585 if (!any(info[i].eq, 2 * info[i].bmap->n_eq, STATUS_CUT) &&
1586 !any(info[j].eq, 2 * info[j].bmap->n_eq, STATUS_CUT))
1587 change = check_facets(i, j, info);
1588 if (change == isl_change_none)
1589 change = check_wrap(i, j, info);
1592 done:
1593 free(info[i].eq);
1594 free(info[j].eq);
1595 free(info[i].ineq);
1596 free(info[j].ineq);
1597 return change;
1598 error:
1599 free(info[i].eq);
1600 free(info[j].eq);
1601 free(info[i].ineq);
1602 free(info[j].ineq);
1603 return isl_change_error;
1606 /* Do the two basic maps live in the same local space, i.e.,
1607 * do they have the same (known) divs?
1608 * If either basic map has any unknown divs, then we can only assume
1609 * that they do not live in the same local space.
1611 static int same_divs(__isl_keep isl_basic_map *bmap1,
1612 __isl_keep isl_basic_map *bmap2)
1614 int i;
1615 int known;
1616 int total;
1618 if (!bmap1 || !bmap2)
1619 return -1;
1620 if (bmap1->n_div != bmap2->n_div)
1621 return 0;
1623 if (bmap1->n_div == 0)
1624 return 1;
1626 known = isl_basic_map_divs_known(bmap1);
1627 if (known < 0 || !known)
1628 return known;
1629 known = isl_basic_map_divs_known(bmap2);
1630 if (known < 0 || !known)
1631 return known;
1633 total = isl_basic_map_total_dim(bmap1);
1634 for (i = 0; i < bmap1->n_div; ++i)
1635 if (!isl_seq_eq(bmap1->div[i], bmap2->div[i], 2 + total))
1636 return 0;
1638 return 1;
1641 /* Does "bmap" contain the basic map represented by the tableau "tab"
1642 * after expanding the divs of "bmap" to match those of "tab"?
1643 * The expansion is performed using the divs "div" and expansion "exp"
1644 * computed by the caller.
1645 * Then we check if all constraints of the expanded "bmap" are valid for "tab".
1647 static int contains_with_expanded_divs(__isl_keep isl_basic_map *bmap,
1648 struct isl_tab *tab, __isl_keep isl_mat *div, int *exp)
1650 int superset = 0;
1651 int *eq_i = NULL;
1652 int *ineq_i = NULL;
1654 bmap = isl_basic_map_copy(bmap);
1655 bmap = isl_basic_set_expand_divs(bmap, isl_mat_copy(div), exp);
1657 if (!bmap)
1658 goto error;
1660 eq_i = eq_status_in(bmap, tab);
1661 if (bmap->n_eq && !eq_i)
1662 goto error;
1663 if (any(eq_i, 2 * bmap->n_eq, STATUS_ERROR))
1664 goto error;
1665 if (any(eq_i, 2 * bmap->n_eq, STATUS_SEPARATE))
1666 goto done;
1668 ineq_i = ineq_status_in(bmap, NULL, tab);
1669 if (bmap->n_ineq && !ineq_i)
1670 goto error;
1671 if (any(ineq_i, bmap->n_ineq, STATUS_ERROR))
1672 goto error;
1673 if (any(ineq_i, bmap->n_ineq, STATUS_SEPARATE))
1674 goto done;
1676 if (all(eq_i, 2 * bmap->n_eq, STATUS_VALID) &&
1677 all(ineq_i, bmap->n_ineq, STATUS_VALID))
1678 superset = 1;
1680 done:
1681 isl_basic_map_free(bmap);
1682 free(eq_i);
1683 free(ineq_i);
1684 return superset;
1685 error:
1686 isl_basic_map_free(bmap);
1687 free(eq_i);
1688 free(ineq_i);
1689 return -1;
1692 /* Does "bmap_i" contain the basic map represented by "info_j"
1693 * after aligning the divs of "bmap_i" to those of "info_j".
1694 * Note that this can only succeed if the number of divs of "bmap_i"
1695 * is smaller than (or equal to) the number of divs of "info_j".
1697 * We first check if the divs of "bmap_i" are all known and form a subset
1698 * of those of "bmap_j". If so, we pass control over to
1699 * contains_with_expanded_divs.
1701 static int contains_after_aligning_divs(__isl_keep isl_basic_map *bmap_i,
1702 struct isl_coalesce_info *info_j)
1704 int known;
1705 isl_mat *div_i, *div_j, *div;
1706 int *exp1 = NULL;
1707 int *exp2 = NULL;
1708 isl_ctx *ctx;
1709 int subset;
1711 known = isl_basic_map_divs_known(bmap_i);
1712 if (known < 0 || !known)
1713 return known;
1715 ctx = isl_basic_map_get_ctx(bmap_i);
1717 div_i = isl_basic_map_get_divs(bmap_i);
1718 div_j = isl_basic_map_get_divs(info_j->bmap);
1720 if (!div_i || !div_j)
1721 goto error;
1723 exp1 = isl_alloc_array(ctx, int, div_i->n_row);
1724 exp2 = isl_alloc_array(ctx, int, div_j->n_row);
1725 if ((div_i->n_row && !exp1) || (div_j->n_row && !exp2))
1726 goto error;
1728 div = isl_merge_divs(div_i, div_j, exp1, exp2);
1729 if (!div)
1730 goto error;
1732 if (div->n_row == div_j->n_row)
1733 subset = contains_with_expanded_divs(bmap_i,
1734 info_j->tab, div, exp1);
1735 else
1736 subset = 0;
1738 isl_mat_free(div);
1740 isl_mat_free(div_i);
1741 isl_mat_free(div_j);
1743 free(exp2);
1744 free(exp1);
1746 return subset;
1747 error:
1748 isl_mat_free(div_i);
1749 isl_mat_free(div_j);
1750 free(exp1);
1751 free(exp2);
1752 return -1;
1755 /* Check if the basic map "j" is a subset of basic map "i",
1756 * if "i" has fewer divs that "j".
1757 * If so, remove basic map "j".
1759 * If the two basic maps have the same number of divs, then
1760 * they must necessarily be different. Otherwise, we would have
1761 * called coalesce_local_pair. We therefore don't try anything
1762 * in this case.
1764 static int coalesced_subset(int i, int j, struct isl_coalesce_info *info)
1766 int superset;
1768 if (info[i].bmap->n_div >= info[j].bmap->n_div)
1769 return 0;
1771 superset = contains_after_aligning_divs(info[i].bmap, &info[j]);
1772 if (superset < 0)
1773 return -1;
1774 if (superset)
1775 drop(&info[j]);
1777 return superset;
1780 /* Check if basic map "j" is a subset of basic map "i" after
1781 * exploiting the extra equalities of "j" to simplify the divs of "i".
1782 * If so, remove basic map "j".
1784 * If "j" does not have any equalities or if they are the same
1785 * as those of "i", then we cannot exploit them to simplify the divs.
1786 * Similarly, if there are no divs in "i", then they cannot be simplified.
1787 * If, on the other hand, the affine hulls of "i" and "j" do not intersect,
1788 * then "j" cannot be a subset of "i".
1790 * Otherwise, we intersect "i" with the affine hull of "j" and then
1791 * check if "j" is a subset of the result after aligning the divs.
1792 * If so, then "j" is definitely a subset of "i" and can be removed.
1793 * Note that if after intersection with the affine hull of "j".
1794 * "i" still has more divs than "j", then there is no way we can
1795 * align the divs of "i" to those of "j".
1797 static int coalesced_subset_with_equalities(int i, int j,
1798 struct isl_coalesce_info *info)
1800 isl_basic_map *hull_i, *hull_j, *bmap_i;
1801 int equal, empty, subset;
1803 if (info[j].bmap->n_eq == 0)
1804 return 0;
1805 if (info[i].bmap->n_div == 0)
1806 return 0;
1808 hull_i = isl_basic_map_copy(info[i].bmap);
1809 hull_i = isl_basic_map_plain_affine_hull(hull_i);
1810 hull_j = isl_basic_map_copy(info[j].bmap);
1811 hull_j = isl_basic_map_plain_affine_hull(hull_j);
1813 hull_j = isl_basic_map_intersect(hull_j, isl_basic_map_copy(hull_i));
1814 equal = isl_basic_map_plain_is_equal(hull_i, hull_j);
1815 empty = isl_basic_map_plain_is_empty(hull_j);
1816 isl_basic_map_free(hull_i);
1818 if (equal < 0 || equal || empty < 0 || empty) {
1819 isl_basic_map_free(hull_j);
1820 return equal < 0 || empty < 0 ? -1 : 0;
1823 bmap_i = isl_basic_map_copy(info[i].bmap);
1824 bmap_i = isl_basic_map_intersect(bmap_i, hull_j);
1825 if (!bmap_i)
1826 return -1;
1828 if (bmap_i->n_div > info[j].bmap->n_div) {
1829 isl_basic_map_free(bmap_i);
1830 return 0;
1833 subset = contains_after_aligning_divs(bmap_i, &info[j]);
1835 isl_basic_map_free(bmap_i);
1837 if (subset < 0)
1838 return -1;
1839 if (subset)
1840 drop(&info[j]);
1842 return subset;
1845 /* Check if one of the basic maps is a subset of the other and, if so,
1846 * drop the subset.
1847 * Note that we only perform any test if the number of divs is different
1848 * in the two basic maps. In case the number of divs is the same,
1849 * we have already established that the divs are different
1850 * in the two basic maps.
1851 * In particular, if the number of divs of basic map i is smaller than
1852 * the number of divs of basic map j, then we check if j is a subset of i
1853 * and vice versa.
1855 static enum isl_change check_coalesce_subset(int i, int j,
1856 struct isl_coalesce_info *info)
1858 int changed;
1860 changed = coalesced_subset(i, j, info);
1861 if (changed < 0 || changed)
1862 return changed < 0 ? isl_change_error : isl_change_drop_second;
1864 changed = coalesced_subset(j, i, info);
1865 if (changed < 0 || changed)
1866 return changed < 0 ? isl_change_error : isl_change_drop_first;
1868 changed = coalesced_subset_with_equalities(i, j, info);
1869 if (changed < 0 || changed)
1870 return changed < 0 ? isl_change_error : isl_change_drop_second;
1872 changed = coalesced_subset_with_equalities(j, i, info);
1873 if (changed < 0 || changed)
1874 return changed < 0 ? isl_change_error : isl_change_drop_first;
1876 return isl_change_none;
1879 /* Does "bmap" involve any divs that themselves refer to divs?
1881 static int has_nested_div(__isl_keep isl_basic_map *bmap)
1883 int i;
1884 unsigned total;
1885 unsigned n_div;
1887 total = isl_basic_map_dim(bmap, isl_dim_all);
1888 n_div = isl_basic_map_dim(bmap, isl_dim_div);
1889 total -= n_div;
1891 for (i = 0; i < n_div; ++i)
1892 if (isl_seq_first_non_zero(bmap->div[i] + 2 + total,
1893 n_div) != -1)
1894 return 1;
1896 return 0;
1899 /* Return a list of affine expressions, one for each integer division
1900 * in "bmap_i". For each integer division that also appears in "bmap_j",
1901 * the affine expression is set to NaN. The number of NaNs in the list
1902 * is equal to the number of integer divisions in "bmap_j".
1903 * For the other integer divisions of "bmap_i", the corresponding
1904 * element in the list is a purely affine expression equal to the integer
1905 * division in "hull".
1906 * If no such list can be constructed, then the number of elements
1907 * in the returned list is smaller than the number of integer divisions
1908 * in "bmap_i".
1910 static __isl_give isl_aff_list *set_up_substitutions(
1911 __isl_keep isl_basic_map *bmap_i, __isl_keep isl_basic_map *bmap_j,
1912 __isl_take isl_basic_map *hull)
1914 unsigned n_div_i, n_div_j, total;
1915 isl_ctx *ctx;
1916 isl_local_space *ls;
1917 isl_basic_set *wrap_hull;
1918 isl_aff *aff_nan;
1919 isl_aff_list *list;
1920 int i, j;
1922 if (!hull)
1923 return NULL;
1925 ctx = isl_basic_map_get_ctx(hull);
1927 n_div_i = isl_basic_map_dim(bmap_i, isl_dim_div);
1928 n_div_j = isl_basic_map_dim(bmap_j, isl_dim_div);
1929 total = isl_basic_map_total_dim(bmap_i) - n_div_i;
1931 ls = isl_basic_map_get_local_space(bmap_i);
1932 ls = isl_local_space_wrap(ls);
1933 wrap_hull = isl_basic_map_wrap(hull);
1935 aff_nan = isl_aff_nan_on_domain(isl_local_space_copy(ls));
1936 list = isl_aff_list_alloc(ctx, n_div_i);
1938 j = 0;
1939 for (i = 0; i < n_div_i; ++i) {
1940 isl_aff *aff;
1942 if (j < n_div_j &&
1943 isl_seq_eq(bmap_i->div[i], bmap_j->div[j], 2 + total)) {
1944 ++j;
1945 list = isl_aff_list_add(list, isl_aff_copy(aff_nan));
1946 continue;
1948 if (n_div_i - i <= n_div_j - j)
1949 break;
1951 aff = isl_local_space_get_div(ls, i);
1952 aff = isl_aff_substitute_equalities(aff,
1953 isl_basic_set_copy(wrap_hull));
1954 aff = isl_aff_floor(aff);
1955 if (!aff)
1956 goto error;
1957 if (isl_aff_dim(aff, isl_dim_div) != 0) {
1958 isl_aff_free(aff);
1959 break;
1962 list = isl_aff_list_add(list, aff);
1965 isl_aff_free(aff_nan);
1966 isl_local_space_free(ls);
1967 isl_basic_set_free(wrap_hull);
1969 return list;
1970 error:
1971 isl_local_space_free(ls);
1972 isl_basic_set_free(wrap_hull);
1973 isl_aff_list_free(list);
1974 return NULL;
1977 /* Add variables to "tab" corresponding to the elements in "list"
1978 * that are not set to NaN. The value of the added variable
1979 * is fixed to the purely affine expression defined by the element.
1980 * "dim" is the offset in the variables of "tab" where we should
1981 * start considering the elements in "list".
1982 * When this function returns, the total number of variables in "tab"
1983 * is equal to "dim" plus the number of elements in "list".
1985 static int add_subs(struct isl_tab *tab, __isl_keep isl_aff_list *list, int dim)
1987 int i, extra;
1988 isl_ctx *ctx;
1989 isl_vec *sub;
1990 isl_aff *aff;
1991 int n;
1993 if (!list)
1994 return -1;
1996 n = isl_aff_list_n_aff(list);
1997 extra = n - (tab->n_var - dim);
1999 if (isl_tab_extend_vars(tab, extra) < 0)
2000 return -1;
2001 if (isl_tab_extend_cons(tab, 2 * extra) < 0)
2002 return -1;
2004 ctx = isl_tab_get_ctx(tab);
2005 sub = isl_vec_alloc(ctx, 1 + dim + n);
2006 if (!sub)
2007 return -1;
2008 isl_seq_clr(sub->el + 1 + dim, n);
2010 for (i = 0; i < n; ++i) {
2011 aff = isl_aff_list_get_aff(list, i);
2012 if (!aff)
2013 goto error;
2014 if (isl_aff_is_nan(aff)) {
2015 isl_aff_free(aff);
2016 continue;
2018 if (isl_tab_insert_var(tab, dim + i) < 0)
2019 goto error;
2020 isl_seq_cpy(sub->el, aff->v->el + 1, 1 + dim);
2021 isl_int_neg(sub->el[1 + dim + i], aff->v->el[0]);
2022 if (isl_tab_add_eq(tab, sub->el) < 0)
2023 goto error;
2024 isl_int_set_si(sub->el[1 + dim + i], 0);
2025 isl_aff_free(aff);
2028 isl_vec_free(sub);
2029 return 0;
2030 error:
2031 isl_aff_free(aff);
2032 isl_vec_free(sub);
2033 return -1;
2036 /* Coalesce basic map "j" into basic map "i" after adding the extra integer
2037 * divisions in "i" but not in "j" to basic map "j", with values
2038 * specified by "list". The total number of elements in "list"
2039 * is equal to the number of integer divisions in "i", while the number
2040 * of NaN elements in the list is equal to the number of integer divisions
2041 * in "j".
2042 * If no coalescing can be performed, then we need to revert basic map "j"
2043 * to its original state. We do the same if basic map "i" gets dropped
2044 * during the coalescing, even though this should not happen in practice
2045 * since we have already checked for "j" being a subset of "i"
2046 * before we reach this stage.
2048 static enum isl_change coalesce_with_subs(int i, int j,
2049 struct isl_coalesce_info *info, __isl_keep isl_aff_list *list)
2051 isl_basic_map *bmap_j;
2052 struct isl_tab_undo *snap;
2053 unsigned dim;
2054 enum isl_change change;
2056 bmap_j = isl_basic_map_copy(info[j].bmap);
2057 info[j].bmap = isl_basic_map_align_divs(info[j].bmap, info[i].bmap);
2058 if (!info[j].bmap)
2059 goto error;
2061 snap = isl_tab_snap(info[j].tab);
2063 dim = isl_basic_map_dim(bmap_j, isl_dim_all);
2064 dim -= isl_basic_map_dim(bmap_j, isl_dim_div);
2065 if (add_subs(info[j].tab, list, dim) < 0)
2066 goto error;
2068 change = coalesce_local_pair(i, j, info);
2069 if (change != isl_change_none && change != isl_change_drop_first) {
2070 isl_basic_map_free(bmap_j);
2071 } else {
2072 isl_basic_map_free(info[j].bmap);
2073 info[j].bmap = bmap_j;
2075 if (isl_tab_rollback(info[j].tab, snap) < 0)
2076 return isl_change_error;
2079 return change;
2080 error:
2081 isl_basic_map_free(bmap_j);
2082 return isl_change_error;
2085 /* Check if we can coalesce basic map "j" into basic map "i" after copying
2086 * those extra integer divisions in "i" that can be simplified away
2087 * using the extra equalities in "j".
2088 * All divs are assumed to be known and not contain any nested divs.
2090 * We first check if there are any extra equalities in "j" that we
2091 * can exploit. Then we check if every integer division in "i"
2092 * either already appears in "j" or can be simplified using the
2093 * extra equalities to a purely affine expression.
2094 * If these tests succeed, then we try to coalesce the two basic maps
2095 * by introducing extra dimensions in "j" corresponding to
2096 * the extra integer divsisions "i" fixed to the corresponding
2097 * purely affine expression.
2099 static enum isl_change check_coalesce_into_eq(int i, int j,
2100 struct isl_coalesce_info *info)
2102 unsigned n_div_i, n_div_j;
2103 isl_basic_map *hull_i, *hull_j;
2104 int equal, empty;
2105 isl_aff_list *list;
2106 enum isl_change change;
2108 n_div_i = isl_basic_map_dim(info[i].bmap, isl_dim_div);
2109 n_div_j = isl_basic_map_dim(info[j].bmap, isl_dim_div);
2110 if (n_div_i <= n_div_j)
2111 return isl_change_none;
2112 if (info[j].bmap->n_eq == 0)
2113 return isl_change_none;
2115 hull_i = isl_basic_map_copy(info[i].bmap);
2116 hull_i = isl_basic_map_plain_affine_hull(hull_i);
2117 hull_j = isl_basic_map_copy(info[j].bmap);
2118 hull_j = isl_basic_map_plain_affine_hull(hull_j);
2120 hull_j = isl_basic_map_intersect(hull_j, isl_basic_map_copy(hull_i));
2121 equal = isl_basic_map_plain_is_equal(hull_i, hull_j);
2122 empty = isl_basic_map_plain_is_empty(hull_j);
2123 isl_basic_map_free(hull_i);
2125 if (equal < 0 || empty < 0)
2126 goto error;
2127 if (equal || empty) {
2128 isl_basic_map_free(hull_j);
2129 return isl_change_none;
2132 list = set_up_substitutions(info[i].bmap, info[j].bmap, hull_j);
2133 if (!list)
2134 goto error;
2135 if (isl_aff_list_n_aff(list) < n_div_i)
2136 change = isl_change_none;
2137 else
2138 change = coalesce_with_subs(i, j, info, list);
2140 isl_aff_list_free(list);
2142 return change;
2143 error:
2144 isl_basic_map_free(hull_j);
2145 return isl_change_error;
2148 /* Check if we can coalesce basic maps "i" and "j" after copying
2149 * those extra integer divisions in one of the basic maps that can
2150 * be simplified away using the extra equalities in the other basic map.
2151 * We require all divs to be known in both basic maps.
2152 * Furthermore, to simplify the comparison of div expressions,
2153 * we do not allow any nested integer divisions.
2155 static enum isl_change check_coalesce_eq(int i, int j,
2156 struct isl_coalesce_info *info)
2158 int known, nested;
2159 enum isl_change change;
2161 known = isl_basic_map_divs_known(info[i].bmap);
2162 if (known < 0 || !known)
2163 return known < 0 ? isl_change_error : isl_change_none;
2164 known = isl_basic_map_divs_known(info[j].bmap);
2165 if (known < 0 || !known)
2166 return known < 0 ? isl_change_error : isl_change_none;
2167 nested = has_nested_div(info[i].bmap);
2168 if (nested < 0 || nested)
2169 return nested < 0 ? isl_change_error : isl_change_none;
2170 nested = has_nested_div(info[j].bmap);
2171 if (nested < 0 || nested)
2172 return nested < 0 ? isl_change_error : isl_change_none;
2174 change = check_coalesce_into_eq(i, j, info);
2175 if (change != isl_change_none)
2176 return change;
2177 change = check_coalesce_into_eq(j, i, info);
2178 if (change != isl_change_none)
2179 return invert_change(change);
2181 return isl_change_none;
2184 /* Check if the union of the given pair of basic maps
2185 * can be represented by a single basic map.
2186 * If so, replace the pair by the single basic map and return
2187 * isl_change_drop_first, isl_change_drop_second or isl_change_fuse.
2188 * Otherwise, return isl_change_none.
2190 * We first check if the two basic maps live in the same local space.
2191 * If so, we do the complete check. Otherwise, we check if one is
2192 * an obvious subset of the other or if the extra integer divisions
2193 * of one basic map can be simplified away using the extra equalities
2194 * of the other basic map.
2196 static enum isl_change coalesce_pair(int i, int j,
2197 struct isl_coalesce_info *info)
2199 int same;
2200 enum isl_change change;
2202 same = same_divs(info[i].bmap, info[j].bmap);
2203 if (same < 0)
2204 return isl_change_error;
2205 if (same)
2206 return coalesce_local_pair(i, j, info);
2208 change = check_coalesce_subset(i, j, info);
2209 if (change != isl_change_none)
2210 return change;
2212 return check_coalesce_eq(i, j, info);
2215 /* Pairwise coalesce the basic maps described by the "n" elements of "info",
2216 * skipping basic maps that have been removed (either before or within
2217 * this function).
2219 * For each basic map i, we check if it can be coalesced with respect
2220 * to any previously considered basic map j.
2221 * If i gets dropped (because it was a subset of some j), then
2222 * we can move on to the next basic map.
2223 * If j gets dropped, we need to continue checking against the other
2224 * previously considered basic maps.
2225 * If the two basic maps got fused, then we recheck the fused basic map
2226 * against the previously considered basic maps.
2228 static int coalesce(isl_ctx *ctx, int n, struct isl_coalesce_info *info)
2230 int i, j;
2232 for (i = n - 2; i >= 0; --i) {
2233 if (info[i].removed)
2234 continue;
2235 for (j = i + 1; j < n; ++j) {
2236 enum isl_change changed;
2238 if (info[j].removed)
2239 continue;
2240 if (info[i].removed)
2241 isl_die(ctx, isl_error_internal,
2242 "basic map unexpectedly removed",
2243 return -1);
2244 changed = coalesce_pair(i, j, info);
2245 switch (changed) {
2246 case isl_change_error:
2247 return -1;
2248 case isl_change_none:
2249 case isl_change_drop_second:
2250 continue;
2251 case isl_change_drop_first:
2252 j = n;
2253 break;
2254 case isl_change_fuse:
2255 j = i;
2256 break;
2261 return 0;
2264 /* Update the basic maps in "map" based on the information in "info".
2265 * In particular, remove the basic maps that have been marked removed and
2266 * update the others based on the information in the corresponding tableau.
2267 * Since we detected implicit equalities without calling
2268 * isl_basic_map_gauss, we need to do it now.
2270 static __isl_give isl_map *update_basic_maps(__isl_take isl_map *map,
2271 int n, struct isl_coalesce_info *info)
2273 int i;
2275 if (!map)
2276 return NULL;
2278 for (i = n - 1; i >= 0; --i) {
2279 if (info[i].removed) {
2280 isl_basic_map_free(map->p[i]);
2281 if (i != map->n - 1)
2282 map->p[i] = map->p[map->n - 1];
2283 map->n--;
2284 continue;
2287 info[i].bmap = isl_basic_map_update_from_tab(info[i].bmap,
2288 info[i].tab);
2289 info[i].bmap = isl_basic_map_gauss(info[i].bmap, NULL);
2290 info[i].bmap = isl_basic_map_finalize(info[i].bmap);
2291 if (!info[i].bmap)
2292 return isl_map_free(map);
2293 ISL_F_SET(info[i].bmap, ISL_BASIC_MAP_NO_IMPLICIT);
2294 ISL_F_SET(info[i].bmap, ISL_BASIC_MAP_NO_REDUNDANT);
2295 isl_basic_map_free(map->p[i]);
2296 map->p[i] = info[i].bmap;
2297 info[i].bmap = NULL;
2300 return map;
2303 /* For each pair of basic maps in the map, check if the union of the two
2304 * can be represented by a single basic map.
2305 * If so, replace the pair by the single basic map and start over.
2307 * Since we are constructing the tableaus of the basic maps anyway,
2308 * we exploit them to detect implicit equalities and redundant constraints.
2309 * This also helps the coalescing as it can ignore the redundant constraints.
2310 * In order to avoid confusion, we make all implicit equalities explicit
2311 * in the basic maps. We don't call isl_basic_map_gauss, though,
2312 * as that may affect the number of constraints.
2313 * This means that we have to call isl_basic_map_gauss at the end
2314 * of the computation (in update_basic_maps) to ensure that
2315 * the basic maps are not left in an unexpected state.
2317 struct isl_map *isl_map_coalesce(struct isl_map *map)
2319 int i;
2320 unsigned n;
2321 isl_ctx *ctx;
2322 struct isl_coalesce_info *info = NULL;
2324 map = isl_map_remove_empty_parts(map);
2325 if (!map)
2326 return NULL;
2328 if (map->n <= 1)
2329 return map;
2331 ctx = isl_map_get_ctx(map);
2332 map = isl_map_sort_divs(map);
2333 map = isl_map_cow(map);
2335 if (!map)
2336 return NULL;
2338 n = map->n;
2340 info = isl_calloc_array(map->ctx, struct isl_coalesce_info, n);
2341 if (!info)
2342 goto error;
2344 for (i = 0; i < map->n; ++i) {
2345 info[i].bmap = isl_basic_map_copy(map->p[i]);
2346 info[i].tab = isl_tab_from_basic_map(info[i].bmap, 0);
2347 if (!info[i].tab)
2348 goto error;
2349 if (!ISL_F_ISSET(info[i].bmap, ISL_BASIC_MAP_NO_IMPLICIT))
2350 if (isl_tab_detect_implicit_equalities(info[i].tab) < 0)
2351 goto error;
2352 info[i].bmap = isl_tab_make_equalities_explicit(info[i].tab,
2353 info[i].bmap);
2354 if (!info[i].bmap)
2355 goto error;
2356 if (!ISL_F_ISSET(info[i].bmap, ISL_BASIC_MAP_NO_REDUNDANT))
2357 if (isl_tab_detect_redundant(info[i].tab) < 0)
2358 goto error;
2360 for (i = map->n - 1; i >= 0; --i)
2361 if (info[i].tab->empty)
2362 drop(&info[i]);
2364 if (coalesce(ctx, n, info) < 0)
2365 goto error;
2367 map = update_basic_maps(map, n, info);
2369 clear_coalesce_info(n, info);
2371 return map;
2372 error:
2373 clear_coalesce_info(n, info);
2374 isl_map_free(map);
2375 return NULL;
2378 /* For each pair of basic sets in the set, check if the union of the two
2379 * can be represented by a single basic set.
2380 * If so, replace the pair by the single basic set and start over.
2382 struct isl_set *isl_set_coalesce(struct isl_set *set)
2384 return (struct isl_set *)isl_map_coalesce((struct isl_map *)set);