isl_qpolynomial_as_polynomial_on_domain: return isl_stat
[isl.git] / isl_coalesce.c
blob57e437647962acf2cdefde5abcbb2c6407f70891
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 inequalties 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 /* Are all non-redundant constraints of the basic map represented by "info"
200 * either valid or cut constraints with respect to the other basic map?
202 static int all_valid_or_cut(struct isl_coalesce_info *info)
204 int i;
206 for (i = 0; i < 2 * info->bmap->n_eq; ++i) {
207 if (info->eq[i] == STATUS_REDUNDANT)
208 continue;
209 if (info->eq[i] == STATUS_VALID)
210 continue;
211 if (info->eq[i] == STATUS_CUT)
212 continue;
213 return 0;
216 for (i = 0; i < info->bmap->n_ineq; ++i) {
217 if (info->ineq[i] == STATUS_REDUNDANT)
218 continue;
219 if (info->ineq[i] == STATUS_VALID)
220 continue;
221 if (info->ineq[i] == STATUS_CUT)
222 continue;
223 return 0;
226 return 1;
229 /* Compute the hash of the (apparent) affine hull of info->bmap (with
230 * the existentially quantified variables removed) and store it
231 * in info->hash.
233 static int coalesce_info_set_hull_hash(struct isl_coalesce_info *info)
235 isl_basic_map *hull;
236 unsigned n_div;
238 hull = isl_basic_map_copy(info->bmap);
239 hull = isl_basic_map_plain_affine_hull(hull);
240 n_div = isl_basic_map_dim(hull, isl_dim_div);
241 hull = isl_basic_map_drop_constraints_involving_dims(hull,
242 isl_dim_div, 0, n_div);
243 info->hull_hash = isl_basic_map_get_hash(hull);
244 isl_basic_map_free(hull);
246 return hull ? 0 : -1;
249 /* Free all the allocated memory in an array
250 * of "n" isl_coalesce_info elements.
252 static void clear_coalesce_info(int n, struct isl_coalesce_info *info)
254 int i;
256 if (!info)
257 return;
259 for (i = 0; i < n; ++i) {
260 isl_basic_map_free(info[i].bmap);
261 isl_tab_free(info[i].tab);
264 free(info);
267 /* Drop the basic map represented by "info".
268 * That is, clear the memory associated to the entry and
269 * mark it as having been removed.
271 static void drop(struct isl_coalesce_info *info)
273 info->bmap = isl_basic_map_free(info->bmap);
274 isl_tab_free(info->tab);
275 info->tab = NULL;
276 info->removed = 1;
279 /* Exchange the information in "info1" with that in "info2".
281 static void exchange(struct isl_coalesce_info *info1,
282 struct isl_coalesce_info *info2)
284 struct isl_coalesce_info info;
286 info = *info1;
287 *info1 = *info2;
288 *info2 = info;
291 /* This type represents the kind of change that has been performed
292 * while trying to coalesce two basic maps.
294 * isl_change_none: nothing was changed
295 * isl_change_drop_first: the first basic map was removed
296 * isl_change_drop_second: the second basic map was removed
297 * isl_change_fuse: the two basic maps were replaced by a new basic map.
299 enum isl_change {
300 isl_change_error = -1,
301 isl_change_none = 0,
302 isl_change_drop_first,
303 isl_change_drop_second,
304 isl_change_fuse,
307 /* Update "change" based on an interchange of the first and the second
308 * basic map. That is, interchange isl_change_drop_first and
309 * isl_change_drop_second.
311 static enum isl_change invert_change(enum isl_change change)
313 switch (change) {
314 case isl_change_error:
315 return isl_change_error;
316 case isl_change_none:
317 return isl_change_none;
318 case isl_change_drop_first:
319 return isl_change_drop_second;
320 case isl_change_drop_second:
321 return isl_change_drop_first;
322 case isl_change_fuse:
323 return isl_change_fuse;
326 return isl_change_error;
329 /* Add the valid constraints of the basic map represented by "info"
330 * to "bmap". "len" is the size of the constraints.
331 * If only one of the pair of inequalities that make up an equality
332 * is valid, then add that inequality.
334 static __isl_give isl_basic_map *add_valid_constraints(
335 __isl_take isl_basic_map *bmap, struct isl_coalesce_info *info,
336 unsigned len)
338 int k, l;
340 if (!bmap)
341 return NULL;
343 for (k = 0; k < info->bmap->n_eq; ++k) {
344 if (info->eq[2 * k] == STATUS_VALID &&
345 info->eq[2 * k + 1] == STATUS_VALID) {
346 l = isl_basic_map_alloc_equality(bmap);
347 if (l < 0)
348 return isl_basic_map_free(bmap);
349 isl_seq_cpy(bmap->eq[l], info->bmap->eq[k], len);
350 } else if (info->eq[2 * k] == STATUS_VALID) {
351 l = isl_basic_map_alloc_inequality(bmap);
352 if (l < 0)
353 return isl_basic_map_free(bmap);
354 isl_seq_neg(bmap->ineq[l], info->bmap->eq[k], len);
355 } else if (info->eq[2 * k + 1] == STATUS_VALID) {
356 l = isl_basic_map_alloc_inequality(bmap);
357 if (l < 0)
358 return isl_basic_map_free(bmap);
359 isl_seq_cpy(bmap->ineq[l], info->bmap->eq[k], len);
363 for (k = 0; k < info->bmap->n_ineq; ++k) {
364 if (info->ineq[k] != STATUS_VALID)
365 continue;
366 l = isl_basic_map_alloc_inequality(bmap);
367 if (l < 0)
368 return isl_basic_map_free(bmap);
369 isl_seq_cpy(bmap->ineq[l], info->bmap->ineq[k], len);
372 return bmap;
375 /* Is "bmap" defined by a number of (non-redundant) constraints that
376 * is greater than the number of constraints of basic maps i and j combined?
377 * Equalities are counted as two inequalities.
379 static int number_of_constraints_increases(int i, int j,
380 struct isl_coalesce_info *info,
381 __isl_keep isl_basic_map *bmap, struct isl_tab *tab)
383 int k, n_old, n_new;
385 n_old = 2 * info[i].bmap->n_eq + info[i].bmap->n_ineq;
386 n_old += 2 * info[j].bmap->n_eq + info[j].bmap->n_ineq;
388 n_new = 2 * bmap->n_eq;
389 for (k = 0; k < bmap->n_ineq; ++k)
390 if (!isl_tab_is_redundant(tab, bmap->n_eq + k))
391 ++n_new;
393 return n_new > n_old;
396 /* Replace the pair of basic maps i and j by the basic map bounded
397 * by the valid constraints in both basic maps and the constraints
398 * in extra (if not NULL).
399 * Place the fused basic map in the position that is the smallest of i and j.
401 * If "detect_equalities" is set, then look for equalities encoded
402 * as pairs of inequalities.
403 * If "check_number" is set, then the original basic maps are only
404 * replaced if the total number of constraints does not increase.
405 * While the number of integer divisions in the two basic maps
406 * is assumed to be the same, the actual definitions may be different.
407 * We only copy the definition from one of the basic map if it is
408 * the same as that of the other basic map. Otherwise, we mark
409 * the integer division as unknown and simplify the basic map
410 * in an attempt to recover the integer division definition.
412 static enum isl_change fuse(int i, int j, struct isl_coalesce_info *info,
413 __isl_keep isl_mat *extra, int detect_equalities, int check_number)
415 int k, l;
416 struct isl_basic_map *fused = NULL;
417 struct isl_tab *fused_tab = NULL;
418 unsigned total = isl_basic_map_total_dim(info[i].bmap);
419 unsigned extra_rows = extra ? extra->n_row : 0;
420 unsigned n_eq, n_ineq;
421 int simplify = 0;
423 if (j < i)
424 return fuse(j, i, info, extra, detect_equalities, check_number);
426 n_eq = info[i].bmap->n_eq + info[j].bmap->n_eq;
427 n_ineq = info[i].bmap->n_ineq + info[j].bmap->n_ineq;
428 fused = isl_basic_map_alloc_space(isl_space_copy(info[i].bmap->dim),
429 info[i].bmap->n_div, n_eq, n_eq + n_ineq + extra_rows);
430 fused = add_valid_constraints(fused, &info[i], 1 + total);
431 fused = add_valid_constraints(fused, &info[j], 1 + total);
432 if (!fused)
433 goto error;
434 if (ISL_F_ISSET(info[i].bmap, ISL_BASIC_MAP_RATIONAL) &&
435 ISL_F_ISSET(info[j].bmap, ISL_BASIC_MAP_RATIONAL))
436 ISL_F_SET(fused, ISL_BASIC_MAP_RATIONAL);
438 for (k = 0; k < info[i].bmap->n_div; ++k) {
439 int l = isl_basic_map_alloc_div(fused);
440 if (l < 0)
441 goto error;
442 if (isl_seq_eq(info[i].bmap->div[k], info[j].bmap->div[k],
443 1 + 1 + total)) {
444 isl_seq_cpy(fused->div[l], info[i].bmap->div[k],
445 1 + 1 + total);
446 } else {
447 isl_int_set_si(fused->div[l][0], 0);
448 simplify = 1;
452 for (k = 0; k < extra_rows; ++k) {
453 l = isl_basic_map_alloc_inequality(fused);
454 if (l < 0)
455 goto error;
456 isl_seq_cpy(fused->ineq[l], extra->row[k], 1 + total);
459 if (detect_equalities)
460 fused = isl_basic_map_detect_inequality_pairs(fused, NULL);
461 fused = isl_basic_map_gauss(fused, NULL);
462 if (simplify || info[j].simplify) {
463 fused = isl_basic_map_simplify(fused);
464 info[i].simplify = 0;
466 fused = isl_basic_map_finalize(fused);
468 fused_tab = isl_tab_from_basic_map(fused, 0);
469 if (isl_tab_detect_redundant(fused_tab) < 0)
470 goto error;
472 if (check_number &&
473 number_of_constraints_increases(i, j, info, fused, fused_tab)) {
474 isl_tab_free(fused_tab);
475 isl_basic_map_free(fused);
476 return isl_change_none;
479 isl_basic_map_free(info[i].bmap);
480 info[i].bmap = fused;
481 isl_tab_free(info[i].tab);
482 info[i].tab = fused_tab;
483 drop(&info[j]);
485 return isl_change_fuse;
486 error:
487 isl_tab_free(fused_tab);
488 isl_basic_map_free(fused);
489 return isl_change_error;
492 /* Given a pair of basic maps i and j such that all constraints are either
493 * "valid" or "cut", check if the facets corresponding to the "cut"
494 * constraints of i lie entirely within basic map j.
495 * If so, replace the pair by the basic map consisting of the valid
496 * constraints in both basic maps.
497 * Checking whether the facet lies entirely within basic map j
498 * is performed by checking whether the constraints of basic map j
499 * are valid for the facet. These tests are performed on a rational
500 * tableau to avoid the theoretical possibility that a constraint
501 * that was considered to be a cut constraint for the entire basic map i
502 * happens to be considered to be a valid constraint for the facet,
503 * even though it cuts off the same rational points.
505 * To see that we are not introducing any extra points, call the
506 * two basic maps A and B and the resulting map U and let x
507 * be an element of U \setminus ( A \cup B ).
508 * A line connecting x with an element of A \cup B meets a facet F
509 * of either A or B. Assume it is a facet of B and let c_1 be
510 * the corresponding facet constraint. We have c_1(x) < 0 and
511 * so c_1 is a cut constraint. This implies that there is some
512 * (possibly rational) point x' satisfying the constraints of A
513 * and the opposite of c_1 as otherwise c_1 would have been marked
514 * valid for A. The line connecting x and x' meets a facet of A
515 * in a (possibly rational) point that also violates c_1, but this
516 * is impossible since all cut constraints of B are valid for all
517 * cut facets of A.
518 * In case F is a facet of A rather than B, then we can apply the
519 * above reasoning to find a facet of B separating x from A \cup B first.
521 static enum isl_change check_facets(int i, int j,
522 struct isl_coalesce_info *info)
524 int k, l;
525 struct isl_tab_undo *snap, *snap2;
526 unsigned n_eq = info[i].bmap->n_eq;
528 snap = isl_tab_snap(info[i].tab);
529 if (isl_tab_mark_rational(info[i].tab) < 0)
530 return isl_change_error;
531 snap2 = isl_tab_snap(info[i].tab);
533 for (k = 0; k < info[i].bmap->n_ineq; ++k) {
534 if (info[i].ineq[k] != STATUS_CUT)
535 continue;
536 if (isl_tab_select_facet(info[i].tab, n_eq + k) < 0)
537 return isl_change_error;
538 for (l = 0; l < info[j].bmap->n_ineq; ++l) {
539 int stat;
540 if (info[j].ineq[l] != STATUS_CUT)
541 continue;
542 stat = status_in(info[j].bmap->ineq[l], info[i].tab);
543 if (stat < 0)
544 return isl_change_error;
545 if (stat != STATUS_VALID)
546 break;
548 if (isl_tab_rollback(info[i].tab, snap2) < 0)
549 return isl_change_error;
550 if (l < info[j].bmap->n_ineq)
551 break;
554 if (k < info[i].bmap->n_ineq) {
555 if (isl_tab_rollback(info[i].tab, snap) < 0)
556 return isl_change_error;
557 return isl_change_none;
559 return fuse(i, j, info, NULL, 0, 0);
562 /* Check if info->bmap contains the basic map represented
563 * by the tableau "tab".
564 * For each equality, we check both the constraint itself
565 * (as an inequality) and its negation. Make sure the
566 * equality is returned to its original state before returning.
568 static isl_bool contains(struct isl_coalesce_info *info, struct isl_tab *tab)
570 int k;
571 unsigned dim;
572 isl_basic_map *bmap = info->bmap;
574 dim = isl_basic_map_total_dim(bmap);
575 for (k = 0; k < bmap->n_eq; ++k) {
576 int stat;
577 isl_seq_neg(bmap->eq[k], bmap->eq[k], 1 + dim);
578 stat = status_in(bmap->eq[k], tab);
579 isl_seq_neg(bmap->eq[k], bmap->eq[k], 1 + dim);
580 if (stat < 0)
581 return isl_bool_error;
582 if (stat != STATUS_VALID)
583 return isl_bool_false;
584 stat = status_in(bmap->eq[k], tab);
585 if (stat < 0)
586 return isl_bool_error;
587 if (stat != STATUS_VALID)
588 return isl_bool_false;
591 for (k = 0; k < bmap->n_ineq; ++k) {
592 int stat;
593 if (info->ineq[k] == STATUS_REDUNDANT)
594 continue;
595 stat = status_in(bmap->ineq[k], tab);
596 if (stat < 0)
597 return isl_bool_error;
598 if (stat != STATUS_VALID)
599 return isl_bool_false;
601 return isl_bool_true;
604 /* Basic map "i" has an inequality (say "k") that is adjacent
605 * to some inequality of basic map "j". All the other inequalities
606 * are valid for "j".
607 * Check if basic map "j" forms an extension of basic map "i".
609 * Note that this function is only called if some of the equalities or
610 * inequalities of basic map "j" do cut basic map "i". The function is
611 * correct even if there are no such cut constraints, but in that case
612 * the additional checks performed by this function are overkill.
614 * In particular, we replace constraint k, say f >= 0, by constraint
615 * f <= -1, add the inequalities of "j" that are valid for "i"
616 * and check if the result is a subset of basic map "j".
617 * To improve the chances of the subset relation being detected,
618 * any variable that only attains a single integer value
619 * in the tableau of "i" is first fixed to that value.
620 * If the result is a subset, then we know that this result is exactly equal
621 * to basic map "j" since all its constraints are valid for basic map "j".
622 * By combining the valid constraints of "i" (all equalities and all
623 * inequalities except "k") and the valid constraints of "j" we therefore
624 * obtain a basic map that is equal to their union.
625 * In this case, there is no need to perform a rollback of the tableau
626 * since it is going to be destroyed in fuse().
629 * |\__ |\__
630 * | \__ | \__
631 * | \_ => | \__
632 * |_______| _ |_________\
635 * |\ |\
636 * | \ | \
637 * | \ | \
638 * | | | \
639 * | ||\ => | \
640 * | || \ | \
641 * | || | | |
642 * |__||_/ |_____/
644 static enum isl_change is_adj_ineq_extension(int i, int j,
645 struct isl_coalesce_info *info)
647 int k;
648 struct isl_tab_undo *snap;
649 unsigned n_eq = info[i].bmap->n_eq;
650 unsigned total = isl_basic_map_total_dim(info[i].bmap);
651 isl_stat r;
652 isl_bool super;
654 if (isl_tab_extend_cons(info[i].tab, 1 + info[j].bmap->n_ineq) < 0)
655 return isl_change_error;
657 k = find(info[i].ineq, info[i].bmap->n_ineq, STATUS_ADJ_INEQ);
658 if (k < 0)
659 isl_die(isl_basic_map_get_ctx(info[i].bmap), isl_error_internal,
660 "info[i].ineq should have exactly one STATUS_ADJ_INEQ",
661 return isl_change_error);
663 snap = isl_tab_snap(info[i].tab);
665 if (isl_tab_unrestrict(info[i].tab, n_eq + k) < 0)
666 return isl_change_error;
668 isl_seq_neg(info[i].bmap->ineq[k], info[i].bmap->ineq[k], 1 + total);
669 isl_int_sub_ui(info[i].bmap->ineq[k][0], info[i].bmap->ineq[k][0], 1);
670 r = isl_tab_add_ineq(info[i].tab, info[i].bmap->ineq[k]);
671 isl_seq_neg(info[i].bmap->ineq[k], info[i].bmap->ineq[k], 1 + total);
672 isl_int_sub_ui(info[i].bmap->ineq[k][0], info[i].bmap->ineq[k][0], 1);
673 if (r < 0)
674 return isl_change_error;
676 for (k = 0; k < info[j].bmap->n_ineq; ++k) {
677 if (info[j].ineq[k] != STATUS_VALID)
678 continue;
679 if (isl_tab_add_ineq(info[i].tab, info[j].bmap->ineq[k]) < 0)
680 return isl_change_error;
682 if (isl_tab_detect_constants(info[i].tab) < 0)
683 return isl_change_error;
685 super = contains(&info[j], info[i].tab);
686 if (super < 0)
687 return isl_change_error;
688 if (super)
689 return fuse(i, j, info, NULL, 0, 0);
691 if (isl_tab_rollback(info[i].tab, snap) < 0)
692 return isl_change_error;
694 return isl_change_none;
698 /* Both basic maps have at least one inequality with and adjacent
699 * (but opposite) inequality in the other basic map.
700 * Check that there are no cut constraints and that there is only
701 * a single pair of adjacent inequalities.
702 * If so, we can replace the pair by a single basic map described
703 * by all but the pair of adjacent inequalities.
704 * Any additional points introduced lie strictly between the two
705 * adjacent hyperplanes and can therefore be integral.
707 * ____ _____
708 * / ||\ / \
709 * / || \ / \
710 * \ || \ => \ \
711 * \ || / \ /
712 * \___||_/ \_____/
714 * The test for a single pair of adjancent inequalities is important
715 * for avoiding the combination of two basic maps like the following
717 * /|
718 * / |
719 * /__|
720 * _____
721 * | |
722 * | |
723 * |___|
725 * If there are some cut constraints on one side, then we may
726 * still be able to fuse the two basic maps, but we need to perform
727 * some additional checks in is_adj_ineq_extension.
729 static enum isl_change check_adj_ineq(int i, int j,
730 struct isl_coalesce_info *info)
732 int count_i, count_j;
733 int cut_i, cut_j;
735 count_i = count(info[i].ineq, info[i].bmap->n_ineq, STATUS_ADJ_INEQ);
736 count_j = count(info[j].ineq, info[j].bmap->n_ineq, STATUS_ADJ_INEQ);
738 if (count_i != 1 && count_j != 1)
739 return isl_change_none;
741 cut_i = any(info[i].eq, 2 * info[i].bmap->n_eq, STATUS_CUT) ||
742 any(info[i].ineq, info[i].bmap->n_ineq, STATUS_CUT);
743 cut_j = any(info[j].eq, 2 * info[j].bmap->n_eq, STATUS_CUT) ||
744 any(info[j].ineq, info[j].bmap->n_ineq, STATUS_CUT);
746 if (!cut_i && !cut_j && count_i == 1 && count_j == 1)
747 return fuse(i, j, info, NULL, 0, 0);
749 if (count_i == 1 && !cut_i)
750 return is_adj_ineq_extension(i, j, info);
752 if (count_j == 1 && !cut_j)
753 return is_adj_ineq_extension(j, i, info);
755 return isl_change_none;
758 /* Given an affine transformation matrix "T", does row "row" represent
759 * anything other than a unit vector (possibly shifted by a constant)
760 * that is not involved in any of the other rows?
762 * That is, if a constraint involves the variable corresponding to
763 * the row, then could its preimage by "T" have any coefficients
764 * that are different from those in the original constraint?
766 static int not_unique_unit_row(__isl_keep isl_mat *T, int row)
768 int i, j;
769 int len = T->n_col - 1;
771 i = isl_seq_first_non_zero(T->row[row] + 1, len);
772 if (i < 0)
773 return 1;
774 if (!isl_int_is_one(T->row[row][1 + i]) &&
775 !isl_int_is_negone(T->row[row][1 + i]))
776 return 1;
778 j = isl_seq_first_non_zero(T->row[row] + 1 + i + 1, len - (i + 1));
779 if (j >= 0)
780 return 1;
782 for (j = 1; j < T->n_row; ++j) {
783 if (j == row)
784 continue;
785 if (!isl_int_is_zero(T->row[j][1 + i]))
786 return 1;
789 return 0;
792 /* Does inequality constraint "ineq" of "bmap" involve any of
793 * the variables marked in "affected"?
794 * "total" is the total number of variables, i.e., the number
795 * of entries in "affected".
797 static isl_bool is_affected(__isl_keep isl_basic_map *bmap, int ineq,
798 int *affected, int total)
800 int i;
802 for (i = 0; i < total; ++i) {
803 if (!affected[i])
804 continue;
805 if (!isl_int_is_zero(bmap->ineq[ineq][1 + i]))
806 return isl_bool_true;
809 return isl_bool_false;
812 /* Given the compressed version of inequality constraint "ineq"
813 * of info->bmap in "v", check if the constraint can be tightened,
814 * where the compression is based on an equality constraint valid
815 * for info->tab.
816 * If so, add the tightened version of the inequality constraint
817 * to info->tab. "v" may be modified by this function.
819 * That is, if the compressed constraint is of the form
821 * m f() + c >= 0
823 * with 0 < c < m, then it is equivalent to
825 * f() >= 0
827 * This means that c can also be subtracted from the original,
828 * uncompressed constraint without affecting the integer points
829 * in info->tab. Add this tightened constraint as an extra row
830 * to info->tab to make this information explicitly available.
832 static __isl_give isl_vec *try_tightening(struct isl_coalesce_info *info,
833 int ineq, __isl_take isl_vec *v)
835 isl_ctx *ctx;
836 isl_stat r;
838 if (!v)
839 return NULL;
841 ctx = isl_vec_get_ctx(v);
842 isl_seq_gcd(v->el + 1, v->size - 1, &ctx->normalize_gcd);
843 if (isl_int_is_zero(ctx->normalize_gcd) ||
844 isl_int_is_one(ctx->normalize_gcd)) {
845 return v;
848 v = isl_vec_cow(v);
849 if (!v)
850 return NULL;
852 isl_int_fdiv_r(v->el[0], v->el[0], ctx->normalize_gcd);
853 if (isl_int_is_zero(v->el[0]))
854 return v;
856 if (isl_tab_extend_cons(info->tab, 1) < 0)
857 return isl_vec_free(v);
859 isl_int_sub(info->bmap->ineq[ineq][0],
860 info->bmap->ineq[ineq][0], v->el[0]);
861 r = isl_tab_add_ineq(info->tab, info->bmap->ineq[ineq]);
862 isl_int_add(info->bmap->ineq[ineq][0],
863 info->bmap->ineq[ineq][0], v->el[0]);
865 if (r < 0)
866 return isl_vec_free(v);
868 return v;
871 /* Tighten the (non-redundant) constraints on the facet represented
872 * by info->tab.
873 * In particular, on input, info->tab represents the result
874 * of relaxing the "n" inequality constraints of info->bmap in "relaxed"
875 * by one, i.e., replacing f_i >= 0 by f_i + 1 >= 0, and then
876 * replacing the one at index "l" by the corresponding equality,
877 * i.e., f_k + 1 = 0, with k = relaxed[l].
879 * Compute a variable compression from the equality constraint f_k + 1 = 0
880 * and use it to tighten the other constraints of info->bmap
881 * (that is, all constraints that have not been relaxed),
882 * updating info->tab (and leaving info->bmap untouched).
883 * The compression handles essentially two cases, one where a variable
884 * is assigned a fixed value and can therefore be eliminated, and one
885 * where one variable is a shifted multiple of some other variable and
886 * can therefore be replaced by that multiple.
887 * Gaussian elimination would also work for the first case, but for
888 * the second case, the effectiveness would depend on the order
889 * of the variables.
890 * After compression, some of the constraints may have coefficients
891 * with a common divisor. If this divisor does not divide the constant
892 * term, then the constraint can be tightened.
893 * The tightening is performed on the tableau info->tab by introducing
894 * extra (temporary) constraints.
896 * Only constraints that are possibly affected by the compression are
897 * considered. In particular, if the constraint only involves variables
898 * that are directly mapped to a distinct set of other variables, then
899 * no common divisor can be introduced and no tightening can occur.
901 * It is important to only consider the non-redundant constraints
902 * since the facet constraint has been relaxed prior to the call
903 * to this function, meaning that the constraints that were redundant
904 * prior to the relaxation may no longer be redundant.
905 * These constraints will be ignored in the fused result, so
906 * the fusion detection should not exploit them.
908 static isl_stat tighten_on_relaxed_facet(struct isl_coalesce_info *info,
909 int n, int *relaxed, int l)
911 unsigned total;
912 isl_ctx *ctx;
913 isl_vec *v = NULL;
914 isl_mat *T;
915 int i;
916 int k;
917 int *affected;
919 k = relaxed[l];
920 ctx = isl_basic_map_get_ctx(info->bmap);
921 total = isl_basic_map_total_dim(info->bmap);
922 isl_int_add_ui(info->bmap->ineq[k][0], info->bmap->ineq[k][0], 1);
923 T = isl_mat_sub_alloc6(ctx, info->bmap->ineq, k, 1, 0, 1 + total);
924 T = isl_mat_variable_compression(T, NULL);
925 isl_int_sub_ui(info->bmap->ineq[k][0], info->bmap->ineq[k][0], 1);
926 if (!T)
927 return isl_stat_error;
928 if (T->n_col == 0) {
929 isl_mat_free(T);
930 return isl_stat_ok;
933 affected = isl_alloc_array(ctx, int, total);
934 if (!affected)
935 goto error;
937 for (i = 0; i < total; ++i)
938 affected[i] = not_unique_unit_row(T, 1 + i);
940 for (i = 0; i < info->bmap->n_ineq; ++i) {
941 isl_bool handle;
942 if (any(relaxed, n, i))
943 continue;
944 if (info->ineq[i] == STATUS_REDUNDANT)
945 continue;
946 handle = is_affected(info->bmap, i, affected, total);
947 if (handle < 0)
948 goto error;
949 if (!handle)
950 continue;
951 v = isl_vec_alloc(ctx, 1 + total);
952 if (!v)
953 goto error;
954 isl_seq_cpy(v->el, info->bmap->ineq[i], 1 + total);
955 v = isl_vec_mat_product(v, isl_mat_copy(T));
956 v = try_tightening(info, i, v);
957 isl_vec_free(v);
958 if (!v)
959 goto error;
962 isl_mat_free(T);
963 free(affected);
964 return isl_stat_ok;
965 error:
966 isl_mat_free(T);
967 free(affected);
968 return isl_stat_error;
971 /* Replace the basic maps "i" and "j" by an extension of "i"
972 * along the "n" inequality constraints in "relax" by one.
973 * The tableau info[i].tab has already been extended.
974 * Extend info[i].bmap accordingly by relaxing all constraints in "relax"
975 * by one.
976 * Each integer division that does not have exactly the same
977 * definition in "i" and "j" is marked unknown and the basic map
978 * is scheduled to be simplified in an attempt to recover
979 * the integer division definition.
980 * Place the extension in the position that is the smallest of i and j.
982 static enum isl_change extend(int i, int j, int n, int *relax,
983 struct isl_coalesce_info *info)
985 int l;
986 unsigned total;
988 info[i].bmap = isl_basic_map_cow(info[i].bmap);
989 if (!info[i].bmap)
990 return isl_change_error;
991 total = isl_basic_map_total_dim(info[i].bmap);
992 for (l = 0; l < info[i].bmap->n_div; ++l)
993 if (!isl_seq_eq(info[i].bmap->div[l],
994 info[j].bmap->div[l], 1 + 1 + total)) {
995 isl_int_set_si(info[i].bmap->div[l][0], 0);
996 info[i].simplify = 1;
998 for (l = 0; l < n; ++l)
999 isl_int_add_ui(info[i].bmap->ineq[relax[l]][0],
1000 info[i].bmap->ineq[relax[l]][0], 1);
1001 ISL_F_SET(info[i].bmap, ISL_BASIC_MAP_FINAL);
1002 drop(&info[j]);
1003 if (j < i)
1004 exchange(&info[i], &info[j]);
1005 return isl_change_fuse;
1008 /* Basic map "i" has "n" inequality constraints (collected in "relax")
1009 * that are such that they include basic map "j" if they are relaxed
1010 * by one. All the other inequalities are valid for "j".
1011 * Check if basic map "j" forms an extension of basic map "i".
1013 * In particular, relax the constraints in "relax", compute the corresponding
1014 * facets one by one and check whether each of these is included
1015 * in the other basic map.
1016 * Before testing for inclusion, the constraints on each facet
1017 * are tightened to increase the chance of an inclusion being detected.
1018 * (Adding the valid constraints of "j" to the tableau of "i", as is done
1019 * in is_adj_ineq_extension, may further increase those chances, but this
1020 * is not currently done.)
1021 * If each facet is included, we know that relaxing the constraints extends
1022 * the basic map with exactly the other basic map (we already know that this
1023 * other basic map is included in the extension, because all other
1024 * inequality constraints are valid of "j") and we can replace the
1025 * two basic maps by this extension.
1026 * ____ _____
1027 * / || / |
1028 * / || / |
1029 * \ || => \ |
1030 * \ || \ |
1031 * \___|| \____|
1034 * \ |\
1035 * |\\ | \
1036 * | \\ | \
1037 * | | => | /
1038 * | / | /
1039 * |/ |/
1041 static enum isl_change is_relaxed_extension(int i, int j, int n, int *relax,
1042 struct isl_coalesce_info *info)
1044 int l;
1045 isl_bool super;
1046 struct isl_tab_undo *snap, *snap2;
1047 unsigned n_eq = info[i].bmap->n_eq;
1049 for (l = 0; l < n; ++l)
1050 if (isl_tab_is_equality(info[i].tab, n_eq + relax[l]))
1051 return isl_change_none;
1053 snap = isl_tab_snap(info[i].tab);
1054 for (l = 0; l < n; ++l)
1055 if (isl_tab_relax(info[i].tab, n_eq + relax[l]) < 0)
1056 return isl_change_error;
1057 snap2 = isl_tab_snap(info[i].tab);
1058 for (l = 0; l < n; ++l) {
1059 if (isl_tab_rollback(info[i].tab, snap2) < 0)
1060 return isl_change_error;
1061 if (isl_tab_select_facet(info[i].tab, n_eq + relax[l]) < 0)
1062 return isl_change_error;
1063 if (tighten_on_relaxed_facet(&info[i], n, relax, l) < 0)
1064 return isl_change_error;
1065 super = contains(&info[j], info[i].tab);
1066 if (super < 0)
1067 return isl_change_error;
1068 if (super)
1069 continue;
1070 if (isl_tab_rollback(info[i].tab, snap) < 0)
1071 return isl_change_error;
1072 return isl_change_none;
1075 if (isl_tab_rollback(info[i].tab, snap2) < 0)
1076 return isl_change_error;
1077 return extend(i, j, n, relax, info);
1080 /* Data structure that keeps track of the wrapping constraints
1081 * and of information to bound the coefficients of those constraints.
1083 * bound is set if we want to apply a bound on the coefficients
1084 * mat contains the wrapping constraints
1085 * max is the bound on the coefficients (if bound is set)
1087 struct isl_wraps {
1088 int bound;
1089 isl_mat *mat;
1090 isl_int max;
1093 /* Update wraps->max to be greater than or equal to the coefficients
1094 * in the equalities and inequalities of info->bmap that can be removed
1095 * if we end up applying wrapping.
1097 static void wraps_update_max(struct isl_wraps *wraps,
1098 struct isl_coalesce_info *info)
1100 int k;
1101 isl_int max_k;
1102 unsigned total = isl_basic_map_total_dim(info->bmap);
1104 isl_int_init(max_k);
1106 for (k = 0; k < info->bmap->n_eq; ++k) {
1107 if (info->eq[2 * k] == STATUS_VALID &&
1108 info->eq[2 * k + 1] == STATUS_VALID)
1109 continue;
1110 isl_seq_abs_max(info->bmap->eq[k] + 1, total, &max_k);
1111 if (isl_int_abs_gt(max_k, wraps->max))
1112 isl_int_set(wraps->max, max_k);
1115 for (k = 0; k < info->bmap->n_ineq; ++k) {
1116 if (info->ineq[k] == STATUS_VALID ||
1117 info->ineq[k] == STATUS_REDUNDANT)
1118 continue;
1119 isl_seq_abs_max(info->bmap->ineq[k] + 1, total, &max_k);
1120 if (isl_int_abs_gt(max_k, wraps->max))
1121 isl_int_set(wraps->max, max_k);
1124 isl_int_clear(max_k);
1127 /* Initialize the isl_wraps data structure.
1128 * If we want to bound the coefficients of the wrapping constraints,
1129 * we set wraps->max to the largest coefficient
1130 * in the equalities and inequalities that can be removed if we end up
1131 * applying wrapping.
1133 static void wraps_init(struct isl_wraps *wraps, __isl_take isl_mat *mat,
1134 struct isl_coalesce_info *info, int i, int j)
1136 isl_ctx *ctx;
1138 wraps->bound = 0;
1139 wraps->mat = mat;
1140 if (!mat)
1141 return;
1142 ctx = isl_mat_get_ctx(mat);
1143 wraps->bound = isl_options_get_coalesce_bounded_wrapping(ctx);
1144 if (!wraps->bound)
1145 return;
1146 isl_int_init(wraps->max);
1147 isl_int_set_si(wraps->max, 0);
1148 wraps_update_max(wraps, &info[i]);
1149 wraps_update_max(wraps, &info[j]);
1152 /* Free the contents of the isl_wraps data structure.
1154 static void wraps_free(struct isl_wraps *wraps)
1156 isl_mat_free(wraps->mat);
1157 if (wraps->bound)
1158 isl_int_clear(wraps->max);
1161 /* Is the wrapping constraint in row "row" allowed?
1163 * If wraps->bound is set, we check that none of the coefficients
1164 * is greater than wraps->max.
1166 static int allow_wrap(struct isl_wraps *wraps, int row)
1168 int i;
1170 if (!wraps->bound)
1171 return 1;
1173 for (i = 1; i < wraps->mat->n_col; ++i)
1174 if (isl_int_abs_gt(wraps->mat->row[row][i], wraps->max))
1175 return 0;
1177 return 1;
1180 /* Wrap "ineq" (or its opposite if "negate" is set) around "bound"
1181 * to include "set" and add the result in position "w" of "wraps".
1182 * "len" is the total number of coefficients in "bound" and "ineq".
1183 * Return 1 on success, 0 on failure and -1 on error.
1184 * Wrapping can fail if the result of wrapping is equal to "bound"
1185 * or if we want to bound the sizes of the coefficients and
1186 * the wrapped constraint does not satisfy this bound.
1188 static int add_wrap(struct isl_wraps *wraps, int w, isl_int *bound,
1189 isl_int *ineq, unsigned len, __isl_keep isl_set *set, int negate)
1191 isl_seq_cpy(wraps->mat->row[w], bound, len);
1192 if (negate) {
1193 isl_seq_neg(wraps->mat->row[w + 1], ineq, len);
1194 ineq = wraps->mat->row[w + 1];
1196 if (!isl_set_wrap_facet(set, wraps->mat->row[w], ineq))
1197 return -1;
1198 if (isl_seq_eq(wraps->mat->row[w], bound, len))
1199 return 0;
1200 if (!allow_wrap(wraps, w))
1201 return 0;
1202 return 1;
1205 /* For each constraint in info->bmap that is not redundant (as determined
1206 * by info->tab) and that is not a valid constraint for the other basic map,
1207 * wrap the constraint around "bound" such that it includes the whole
1208 * set "set" and append the resulting constraint to "wraps".
1209 * Note that the constraints that are valid for the other basic map
1210 * will be added to the combined basic map by default, so there is
1211 * no need to wrap them.
1212 * The caller wrap_in_facets even relies on this function not wrapping
1213 * any constraints that are already valid.
1214 * "wraps" is assumed to have been pre-allocated to the appropriate size.
1215 * wraps->n_row is the number of actual wrapped constraints that have
1216 * been added.
1217 * If any of the wrapping problems results in a constraint that is
1218 * identical to "bound", then this means that "set" is unbounded in such
1219 * way that no wrapping is possible. If this happens then wraps->n_row
1220 * is reset to zero.
1221 * Similarly, if we want to bound the coefficients of the wrapping
1222 * constraints and a newly added wrapping constraint does not
1223 * satisfy the bound, then wraps->n_row is also reset to zero.
1225 static isl_stat add_wraps(struct isl_wraps *wraps,
1226 struct isl_coalesce_info *info, isl_int *bound, __isl_keep isl_set *set)
1228 int l, m;
1229 int w;
1230 int added;
1231 isl_basic_map *bmap = info->bmap;
1232 unsigned len = 1 + isl_basic_map_total_dim(bmap);
1234 w = wraps->mat->n_row;
1236 for (l = 0; l < bmap->n_ineq; ++l) {
1237 if (info->ineq[l] == STATUS_VALID ||
1238 info->ineq[l] == STATUS_REDUNDANT)
1239 continue;
1240 if (isl_seq_is_neg(bound, bmap->ineq[l], len))
1241 continue;
1242 if (isl_seq_eq(bound, bmap->ineq[l], len))
1243 continue;
1244 if (isl_tab_is_redundant(info->tab, bmap->n_eq + l))
1245 continue;
1247 added = add_wrap(wraps, w, bound, bmap->ineq[l], len, set, 0);
1248 if (added < 0)
1249 return isl_stat_error;
1250 if (!added)
1251 goto unbounded;
1252 ++w;
1254 for (l = 0; l < bmap->n_eq; ++l) {
1255 if (isl_seq_is_neg(bound, bmap->eq[l], len))
1256 continue;
1257 if (isl_seq_eq(bound, bmap->eq[l], len))
1258 continue;
1260 for (m = 0; m < 2; ++m) {
1261 if (info->eq[2 * l + m] == STATUS_VALID)
1262 continue;
1263 added = add_wrap(wraps, w, bound, bmap->eq[l], len,
1264 set, !m);
1265 if (added < 0)
1266 return isl_stat_error;
1267 if (!added)
1268 goto unbounded;
1269 ++w;
1273 wraps->mat->n_row = w;
1274 return isl_stat_ok;
1275 unbounded:
1276 wraps->mat->n_row = 0;
1277 return isl_stat_ok;
1280 /* Check if the constraints in "wraps" from "first" until the last
1281 * are all valid for the basic set represented by "tab".
1282 * If not, wraps->n_row is set to zero.
1284 static int check_wraps(__isl_keep isl_mat *wraps, int first,
1285 struct isl_tab *tab)
1287 int i;
1289 for (i = first; i < wraps->n_row; ++i) {
1290 enum isl_ineq_type type;
1291 type = isl_tab_ineq_type(tab, wraps->row[i]);
1292 if (type == isl_ineq_error)
1293 return -1;
1294 if (type == isl_ineq_redundant)
1295 continue;
1296 wraps->n_row = 0;
1297 return 0;
1300 return 0;
1303 /* Return a set that corresponds to the non-redundant constraints
1304 * (as recorded in tab) of bmap.
1306 * It's important to remove the redundant constraints as some
1307 * of the other constraints may have been modified after the
1308 * constraints were marked redundant.
1309 * In particular, a constraint may have been relaxed.
1310 * Redundant constraints are ignored when a constraint is relaxed
1311 * and should therefore continue to be ignored ever after.
1312 * Otherwise, the relaxation might be thwarted by some of
1313 * these constraints.
1315 * Update the underlying set to ensure that the dimension doesn't change.
1316 * Otherwise the integer divisions could get dropped if the tab
1317 * turns out to be empty.
1319 static __isl_give isl_set *set_from_updated_bmap(__isl_keep isl_basic_map *bmap,
1320 struct isl_tab *tab)
1322 isl_basic_set *bset;
1324 bmap = isl_basic_map_copy(bmap);
1325 bset = isl_basic_map_underlying_set(bmap);
1326 bset = isl_basic_set_cow(bset);
1327 bset = isl_basic_set_update_from_tab(bset, tab);
1328 return isl_set_from_basic_set(bset);
1331 /* Wrap the constraints of info->bmap that bound the facet defined
1332 * by inequality "k" around (the opposite of) this inequality to
1333 * include "set". "bound" may be used to store the negated inequality.
1334 * Since the wrapped constraints are not guaranteed to contain the whole
1335 * of info->bmap, we check them in check_wraps.
1336 * If any of the wrapped constraints turn out to be invalid, then
1337 * check_wraps will reset wrap->n_row to zero.
1339 static isl_stat add_wraps_around_facet(struct isl_wraps *wraps,
1340 struct isl_coalesce_info *info, int k, isl_int *bound,
1341 __isl_keep isl_set *set)
1343 struct isl_tab_undo *snap;
1344 int n;
1345 unsigned total = isl_basic_map_total_dim(info->bmap);
1347 snap = isl_tab_snap(info->tab);
1349 if (isl_tab_select_facet(info->tab, info->bmap->n_eq + k) < 0)
1350 return isl_stat_error;
1351 if (isl_tab_detect_redundant(info->tab) < 0)
1352 return isl_stat_error;
1354 isl_seq_neg(bound, info->bmap->ineq[k], 1 + total);
1356 n = wraps->mat->n_row;
1357 if (add_wraps(wraps, info, bound, set) < 0)
1358 return isl_stat_error;
1360 if (isl_tab_rollback(info->tab, snap) < 0)
1361 return isl_stat_error;
1362 if (check_wraps(wraps->mat, n, info->tab) < 0)
1363 return isl_stat_error;
1365 return isl_stat_ok;
1368 /* Given a basic set i with a constraint k that is adjacent to
1369 * basic set j, check if we can wrap
1370 * both the facet corresponding to k (if "wrap_facet" is set) and basic map j
1371 * (always) around their ridges to include the other set.
1372 * If so, replace the pair of basic sets by their union.
1374 * All constraints of i (except k) are assumed to be valid or
1375 * cut constraints for j.
1376 * Wrapping the cut constraints to include basic map j may result
1377 * in constraints that are no longer valid of basic map i
1378 * we have to check that the resulting wrapping constraints are valid for i.
1379 * If "wrap_facet" is not set, then all constraints of i (except k)
1380 * are assumed to be valid for j.
1381 * ____ _____
1382 * / | / \
1383 * / || / |
1384 * \ || => \ |
1385 * \ || \ |
1386 * \___|| \____|
1389 static enum isl_change can_wrap_in_facet(int i, int j, int k,
1390 struct isl_coalesce_info *info, int wrap_facet)
1392 enum isl_change change = isl_change_none;
1393 struct isl_wraps wraps;
1394 isl_ctx *ctx;
1395 isl_mat *mat;
1396 struct isl_set *set_i = NULL;
1397 struct isl_set *set_j = NULL;
1398 struct isl_vec *bound = NULL;
1399 unsigned total = isl_basic_map_total_dim(info[i].bmap);
1401 set_i = set_from_updated_bmap(info[i].bmap, info[i].tab);
1402 set_j = set_from_updated_bmap(info[j].bmap, info[j].tab);
1403 ctx = isl_basic_map_get_ctx(info[i].bmap);
1404 mat = isl_mat_alloc(ctx, 2 * (info[i].bmap->n_eq + info[j].bmap->n_eq) +
1405 info[i].bmap->n_ineq + info[j].bmap->n_ineq,
1406 1 + total);
1407 wraps_init(&wraps, mat, info, i, j);
1408 bound = isl_vec_alloc(ctx, 1 + total);
1409 if (!set_i || !set_j || !wraps.mat || !bound)
1410 goto error;
1412 isl_seq_cpy(bound->el, info[i].bmap->ineq[k], 1 + total);
1413 isl_int_add_ui(bound->el[0], bound->el[0], 1);
1415 isl_seq_cpy(wraps.mat->row[0], bound->el, 1 + total);
1416 wraps.mat->n_row = 1;
1418 if (add_wraps(&wraps, &info[j], bound->el, set_i) < 0)
1419 goto error;
1420 if (!wraps.mat->n_row)
1421 goto unbounded;
1423 if (wrap_facet) {
1424 if (add_wraps_around_facet(&wraps, &info[i], k,
1425 bound->el, set_j) < 0)
1426 goto error;
1427 if (!wraps.mat->n_row)
1428 goto unbounded;
1431 change = fuse(i, j, info, wraps.mat, 0, 0);
1433 unbounded:
1434 wraps_free(&wraps);
1436 isl_set_free(set_i);
1437 isl_set_free(set_j);
1439 isl_vec_free(bound);
1441 return change;
1442 error:
1443 wraps_free(&wraps);
1444 isl_vec_free(bound);
1445 isl_set_free(set_i);
1446 isl_set_free(set_j);
1447 return isl_change_error;
1450 /* Given a cut constraint t(x) >= 0 of basic map i, stored in row "w"
1451 * of wrap.mat, replace it by its relaxed version t(x) + 1 >= 0, and
1452 * add wrapping constraints to wrap.mat for all constraints
1453 * of basic map j that bound the part of basic map j that sticks out
1454 * of the cut constraint.
1455 * "set_i" is the underlying set of basic map i.
1456 * If any wrapping fails, then wraps->mat.n_row is reset to zero.
1458 * In particular, we first intersect basic map j with t(x) + 1 = 0.
1459 * If the result is empty, then t(x) >= 0 was actually a valid constraint
1460 * (with respect to the integer points), so we add t(x) >= 0 instead.
1461 * Otherwise, we wrap the constraints of basic map j that are not
1462 * redundant in this intersection and that are not already valid
1463 * for basic map i over basic map i.
1464 * Note that it is sufficient to wrap the constraints to include
1465 * basic map i, because we will only wrap the constraints that do
1466 * not include basic map i already. The wrapped constraint will
1467 * therefore be more relaxed compared to the original constraint.
1468 * Since the original constraint is valid for basic map j, so is
1469 * the wrapped constraint.
1471 static isl_stat wrap_in_facet(struct isl_wraps *wraps, int w,
1472 struct isl_coalesce_info *info_j, __isl_keep isl_set *set_i,
1473 struct isl_tab_undo *snap)
1475 isl_int_add_ui(wraps->mat->row[w][0], wraps->mat->row[w][0], 1);
1476 if (isl_tab_add_eq(info_j->tab, wraps->mat->row[w]) < 0)
1477 return isl_stat_error;
1478 if (isl_tab_detect_redundant(info_j->tab) < 0)
1479 return isl_stat_error;
1481 if (info_j->tab->empty)
1482 isl_int_sub_ui(wraps->mat->row[w][0], wraps->mat->row[w][0], 1);
1483 else if (add_wraps(wraps, info_j, wraps->mat->row[w], set_i) < 0)
1484 return isl_stat_error;
1486 if (isl_tab_rollback(info_j->tab, snap) < 0)
1487 return isl_stat_error;
1489 return isl_stat_ok;
1492 /* Given a pair of basic maps i and j such that j sticks out
1493 * of i at n cut constraints, each time by at most one,
1494 * try to compute wrapping constraints and replace the two
1495 * basic maps by a single basic map.
1496 * The other constraints of i are assumed to be valid for j.
1497 * "set_i" is the underlying set of basic map i.
1498 * "wraps" has been initialized to be of the right size.
1500 * For each cut constraint t(x) >= 0 of i, we add the relaxed version
1501 * t(x) + 1 >= 0, along with wrapping constraints for all constraints
1502 * of basic map j that bound the part of basic map j that sticks out
1503 * of the cut constraint.
1505 * If any wrapping fails, i.e., if we cannot wrap to touch
1506 * the union, then we give up.
1507 * Otherwise, the pair of basic maps is replaced by their union.
1509 static enum isl_change try_wrap_in_facets(int i, int j,
1510 struct isl_coalesce_info *info, struct isl_wraps *wraps,
1511 __isl_keep isl_set *set_i)
1513 int k, l, w;
1514 unsigned total;
1515 struct isl_tab_undo *snap;
1517 total = isl_basic_map_total_dim(info[i].bmap);
1519 snap = isl_tab_snap(info[j].tab);
1521 wraps->mat->n_row = 0;
1523 for (k = 0; k < info[i].bmap->n_eq; ++k) {
1524 for (l = 0; l < 2; ++l) {
1525 if (info[i].eq[2 * k + l] != STATUS_CUT)
1526 continue;
1527 w = wraps->mat->n_row++;
1528 if (l == 0)
1529 isl_seq_neg(wraps->mat->row[w],
1530 info[i].bmap->eq[k], 1 + total);
1531 else
1532 isl_seq_cpy(wraps->mat->row[w],
1533 info[i].bmap->eq[k], 1 + total);
1534 if (wrap_in_facet(wraps, w, &info[j], set_i, snap) < 0)
1535 return isl_change_error;
1537 if (!wraps->mat->n_row)
1538 return isl_change_none;
1542 for (k = 0; k < info[i].bmap->n_ineq; ++k) {
1543 if (info[i].ineq[k] != STATUS_CUT)
1544 continue;
1545 w = wraps->mat->n_row++;
1546 isl_seq_cpy(wraps->mat->row[w],
1547 info[i].bmap->ineq[k], 1 + total);
1548 if (wrap_in_facet(wraps, w, &info[j], set_i, snap) < 0)
1549 return isl_change_error;
1551 if (!wraps->mat->n_row)
1552 return isl_change_none;
1555 return fuse(i, j, info, wraps->mat, 0, 1);
1558 /* Given a pair of basic maps i and j such that j sticks out
1559 * of i at n cut constraints, each time by at most one,
1560 * try to compute wrapping constraints and replace the two
1561 * basic maps by a single basic map.
1562 * The other constraints of i are assumed to be valid for j.
1564 * The core computation is performed by try_wrap_in_facets.
1565 * This function simply extracts an underlying set representation
1566 * of basic map i and initializes the data structure for keeping
1567 * track of wrapping constraints.
1569 static enum isl_change wrap_in_facets(int i, int j, int n,
1570 struct isl_coalesce_info *info)
1572 enum isl_change change = isl_change_none;
1573 struct isl_wraps wraps;
1574 isl_ctx *ctx;
1575 isl_mat *mat;
1576 isl_set *set_i = NULL;
1577 unsigned total = isl_basic_map_total_dim(info[i].bmap);
1578 int max_wrap;
1580 if (isl_tab_extend_cons(info[j].tab, 1) < 0)
1581 return isl_change_error;
1583 max_wrap = 1 + 2 * info[j].bmap->n_eq + info[j].bmap->n_ineq;
1584 max_wrap *= n;
1586 set_i = set_from_updated_bmap(info[i].bmap, info[i].tab);
1587 ctx = isl_basic_map_get_ctx(info[i].bmap);
1588 mat = isl_mat_alloc(ctx, max_wrap, 1 + total);
1589 wraps_init(&wraps, mat, info, i, j);
1590 if (!set_i || !wraps.mat)
1591 goto error;
1593 change = try_wrap_in_facets(i, j, info, &wraps, set_i);
1595 wraps_free(&wraps);
1596 isl_set_free(set_i);
1598 return change;
1599 error:
1600 wraps_free(&wraps);
1601 isl_set_free(set_i);
1602 return isl_change_error;
1605 /* Return the effect of inequality "ineq" on the tableau "tab",
1606 * after relaxing the constant term of "ineq" by one.
1608 static enum isl_ineq_type type_of_relaxed(struct isl_tab *tab, isl_int *ineq)
1610 enum isl_ineq_type type;
1612 isl_int_add_ui(ineq[0], ineq[0], 1);
1613 type = isl_tab_ineq_type(tab, ineq);
1614 isl_int_sub_ui(ineq[0], ineq[0], 1);
1616 return type;
1619 /* Given two basic sets i and j,
1620 * check if relaxing all the cut constraints of i by one turns
1621 * them into valid constraint for j and check if we can wrap in
1622 * the bits that are sticking out.
1623 * If so, replace the pair by their union.
1625 * We first check if all relaxed cut inequalities of i are valid for j
1626 * and then try to wrap in the intersections of the relaxed cut inequalities
1627 * with j.
1629 * During this wrapping, we consider the points of j that lie at a distance
1630 * of exactly 1 from i. In particular, we ignore the points that lie in
1631 * between this lower-dimensional space and the basic map i.
1632 * We can therefore only apply this to integer maps.
1633 * ____ _____
1634 * / ___|_ / \
1635 * / | | / |
1636 * \ | | => \ |
1637 * \|____| \ |
1638 * \___| \____/
1640 * _____ ______
1641 * | ____|_ | \
1642 * | | | | |
1643 * | | | => | |
1644 * |_| | | |
1645 * |_____| \______|
1647 * _______
1648 * | |
1649 * | |\ |
1650 * | | \ |
1651 * | | \ |
1652 * | | \|
1653 * | | \
1654 * | |_____\
1655 * | |
1656 * |_______|
1658 * Wrapping can fail if the result of wrapping one of the facets
1659 * around its edges does not produce any new facet constraint.
1660 * In particular, this happens when we try to wrap in unbounded sets.
1662 * _______________________________________________________________________
1664 * | ___
1665 * | | |
1666 * |_| |_________________________________________________________________
1667 * |___|
1669 * The following is not an acceptable result of coalescing the above two
1670 * sets as it includes extra integer points.
1671 * _______________________________________________________________________
1673 * |
1674 * |
1676 * \______________________________________________________________________
1678 static enum isl_change can_wrap_in_set(int i, int j,
1679 struct isl_coalesce_info *info)
1681 int k, l;
1682 int n;
1683 unsigned total;
1685 if (ISL_F_ISSET(info[i].bmap, ISL_BASIC_MAP_RATIONAL) ||
1686 ISL_F_ISSET(info[j].bmap, ISL_BASIC_MAP_RATIONAL))
1687 return isl_change_none;
1689 n = count(info[i].eq, 2 * info[i].bmap->n_eq, STATUS_CUT);
1690 n += count(info[i].ineq, info[i].bmap->n_ineq, STATUS_CUT);
1691 if (n == 0)
1692 return isl_change_none;
1694 total = isl_basic_map_total_dim(info[i].bmap);
1695 for (k = 0; k < info[i].bmap->n_eq; ++k) {
1696 for (l = 0; l < 2; ++l) {
1697 enum isl_ineq_type type;
1699 if (info[i].eq[2 * k + l] != STATUS_CUT)
1700 continue;
1702 if (l == 0)
1703 isl_seq_neg(info[i].bmap->eq[k],
1704 info[i].bmap->eq[k], 1 + total);
1705 type = type_of_relaxed(info[j].tab,
1706 info[i].bmap->eq[k]);
1707 if (l == 0)
1708 isl_seq_neg(info[i].bmap->eq[k],
1709 info[i].bmap->eq[k], 1 + total);
1710 if (type == isl_ineq_error)
1711 return isl_change_error;
1712 if (type != isl_ineq_redundant)
1713 return isl_change_none;
1717 for (k = 0; k < info[i].bmap->n_ineq; ++k) {
1718 enum isl_ineq_type type;
1720 if (info[i].ineq[k] != STATUS_CUT)
1721 continue;
1723 type = type_of_relaxed(info[j].tab, info[i].bmap->ineq[k]);
1724 if (type == isl_ineq_error)
1725 return isl_change_error;
1726 if (type != isl_ineq_redundant)
1727 return isl_change_none;
1730 return wrap_in_facets(i, j, n, info);
1733 /* Check if either i or j has only cut constraints that can
1734 * be used to wrap in (a facet of) the other basic set.
1735 * if so, replace the pair by their union.
1737 static enum isl_change check_wrap(int i, int j, struct isl_coalesce_info *info)
1739 enum isl_change change = isl_change_none;
1741 change = can_wrap_in_set(i, j, info);
1742 if (change != isl_change_none)
1743 return change;
1745 change = can_wrap_in_set(j, i, info);
1746 return change;
1749 /* Check if all inequality constraints of "i" that cut "j" cease
1750 * to be cut constraints if they are relaxed by one.
1751 * If so, collect the cut constraints in "list".
1752 * The caller is responsible for allocating "list".
1754 static isl_bool all_cut_by_one(int i, int j, struct isl_coalesce_info *info,
1755 int *list)
1757 int l, n;
1759 n = 0;
1760 for (l = 0; l < info[i].bmap->n_ineq; ++l) {
1761 enum isl_ineq_type type;
1763 if (info[i].ineq[l] != STATUS_CUT)
1764 continue;
1765 type = type_of_relaxed(info[j].tab, info[i].bmap->ineq[l]);
1766 if (type == isl_ineq_error)
1767 return isl_bool_error;
1768 if (type != isl_ineq_redundant)
1769 return isl_bool_false;
1770 list[n++] = l;
1773 return isl_bool_true;
1776 /* Given two basic maps such that "j" has at least one equality constraint
1777 * that is adjacent to an inequality constraint of "i" and such that "i" has
1778 * exactly one inequality constraint that is adjacent to an equality
1779 * constraint of "j", check whether "i" can be extended to include "j" or
1780 * whether "j" can be wrapped into "i".
1781 * All remaining constraints of "i" and "j" are assumed to be valid
1782 * or cut constraints of the other basic map.
1783 * However, none of the equality constraints of "i" are cut constraints.
1785 * If "i" has any "cut" inequality constraints, then check if relaxing
1786 * each of them by one is sufficient for them to become valid.
1787 * If so, check if the inequality constraint adjacent to an equality
1788 * constraint of "j" along with all these cut constraints
1789 * can be relaxed by one to contain exactly "j".
1790 * Otherwise, or if this fails, check if "j" can be wrapped into "i".
1792 static enum isl_change check_single_adj_eq(int i, int j,
1793 struct isl_coalesce_info *info)
1795 enum isl_change change = isl_change_none;
1796 int k;
1797 int n_cut;
1798 int *relax;
1799 isl_ctx *ctx;
1800 isl_bool try_relax;
1802 n_cut = count(info[i].ineq, info[i].bmap->n_ineq, STATUS_CUT);
1804 k = find(info[i].ineq, info[i].bmap->n_ineq, STATUS_ADJ_EQ);
1806 if (n_cut > 0) {
1807 ctx = isl_basic_map_get_ctx(info[i].bmap);
1808 relax = isl_calloc_array(ctx, int, 1 + n_cut);
1809 if (!relax)
1810 return isl_change_error;
1811 relax[0] = k;
1812 try_relax = all_cut_by_one(i, j, info, relax + 1);
1813 if (try_relax < 0)
1814 change = isl_change_error;
1815 } else {
1816 try_relax = isl_bool_true;
1817 relax = &k;
1819 if (try_relax && change == isl_change_none)
1820 change = is_relaxed_extension(i, j, 1 + n_cut, relax, info);
1821 if (n_cut > 0)
1822 free(relax);
1823 if (change != isl_change_none)
1824 return change;
1826 change = can_wrap_in_facet(i, j, k, info, n_cut > 0);
1828 return change;
1831 /* At least one of the basic maps has an equality that is adjacent
1832 * to inequality. Make sure that only one of the basic maps has
1833 * such an equality and that the other basic map has exactly one
1834 * inequality adjacent to an equality.
1835 * If the other basic map does not have such an inequality, then
1836 * check if all its constraints are either valid or cut constraints
1837 * and, if so, try wrapping in the first map into the second.
1838 * Otherwise, try to extend one basic map with the other or
1839 * wrap one basic map in the other.
1841 static enum isl_change check_adj_eq(int i, int j,
1842 struct isl_coalesce_info *info)
1844 if (any(info[i].eq, 2 * info[i].bmap->n_eq, STATUS_ADJ_INEQ) &&
1845 any(info[j].eq, 2 * info[j].bmap->n_eq, STATUS_ADJ_INEQ))
1846 /* ADJ EQ TOO MANY */
1847 return isl_change_none;
1849 if (any(info[i].eq, 2 * info[i].bmap->n_eq, STATUS_ADJ_INEQ))
1850 return check_adj_eq(j, i, info);
1852 /* j has an equality adjacent to an inequality in i */
1854 if (count(info[i].ineq, info[i].bmap->n_ineq, STATUS_ADJ_EQ) != 1) {
1855 if (all_valid_or_cut(&info[i]))
1856 return can_wrap_in_set(i, j, info);
1857 return isl_change_none;
1859 if (any(info[i].eq, 2 * info[i].bmap->n_eq, STATUS_CUT))
1860 return isl_change_none;
1861 if (any(info[j].ineq, info[j].bmap->n_ineq, STATUS_ADJ_EQ) ||
1862 any(info[i].ineq, info[i].bmap->n_ineq, STATUS_ADJ_INEQ) ||
1863 any(info[j].ineq, info[j].bmap->n_ineq, STATUS_ADJ_INEQ))
1864 /* ADJ EQ TOO MANY */
1865 return isl_change_none;
1867 return check_single_adj_eq(i, j, info);
1870 /* The two basic maps lie on adjacent hyperplanes. In particular,
1871 * basic map "i" has an equality that lies parallel to basic map "j".
1872 * Check if we can wrap the facets around the parallel hyperplanes
1873 * to include the other set.
1875 * We perform basically the same operations as can_wrap_in_facet,
1876 * except that we don't need to select a facet of one of the sets.
1878 * \\ \\
1879 * \\ => \\
1880 * \ \|
1882 * If there is more than one equality of "i" adjacent to an equality of "j",
1883 * then the result will satisfy one or more equalities that are a linear
1884 * combination of these equalities. These will be encoded as pairs
1885 * of inequalities in the wrapping constraints and need to be made
1886 * explicit.
1888 static enum isl_change check_eq_adj_eq(int i, int j,
1889 struct isl_coalesce_info *info)
1891 int k;
1892 enum isl_change change = isl_change_none;
1893 int detect_equalities = 0;
1894 struct isl_wraps wraps;
1895 isl_ctx *ctx;
1896 isl_mat *mat;
1897 struct isl_set *set_i = NULL;
1898 struct isl_set *set_j = NULL;
1899 struct isl_vec *bound = NULL;
1900 unsigned total = isl_basic_map_total_dim(info[i].bmap);
1902 if (count(info[i].eq, 2 * info[i].bmap->n_eq, STATUS_ADJ_EQ) != 1)
1903 detect_equalities = 1;
1905 k = find(info[i].eq, 2 * info[i].bmap->n_eq, STATUS_ADJ_EQ);
1907 set_i = set_from_updated_bmap(info[i].bmap, info[i].tab);
1908 set_j = set_from_updated_bmap(info[j].bmap, info[j].tab);
1909 ctx = isl_basic_map_get_ctx(info[i].bmap);
1910 mat = isl_mat_alloc(ctx, 2 * (info[i].bmap->n_eq + info[j].bmap->n_eq) +
1911 info[i].bmap->n_ineq + info[j].bmap->n_ineq,
1912 1 + total);
1913 wraps_init(&wraps, mat, info, i, j);
1914 bound = isl_vec_alloc(ctx, 1 + total);
1915 if (!set_i || !set_j || !wraps.mat || !bound)
1916 goto error;
1918 if (k % 2 == 0)
1919 isl_seq_neg(bound->el, info[i].bmap->eq[k / 2], 1 + total);
1920 else
1921 isl_seq_cpy(bound->el, info[i].bmap->eq[k / 2], 1 + total);
1922 isl_int_add_ui(bound->el[0], bound->el[0], 1);
1924 isl_seq_cpy(wraps.mat->row[0], bound->el, 1 + total);
1925 wraps.mat->n_row = 1;
1927 if (add_wraps(&wraps, &info[j], bound->el, set_i) < 0)
1928 goto error;
1929 if (!wraps.mat->n_row)
1930 goto unbounded;
1932 isl_int_sub_ui(bound->el[0], bound->el[0], 1);
1933 isl_seq_neg(bound->el, bound->el, 1 + total);
1935 isl_seq_cpy(wraps.mat->row[wraps.mat->n_row], bound->el, 1 + total);
1936 wraps.mat->n_row++;
1938 if (add_wraps(&wraps, &info[i], bound->el, set_j) < 0)
1939 goto error;
1940 if (!wraps.mat->n_row)
1941 goto unbounded;
1943 change = fuse(i, j, info, wraps.mat, detect_equalities, 0);
1945 if (0) {
1946 error: change = isl_change_error;
1948 unbounded:
1950 wraps_free(&wraps);
1951 isl_set_free(set_i);
1952 isl_set_free(set_j);
1953 isl_vec_free(bound);
1955 return change;
1958 /* Initialize the "eq" and "ineq" fields of "info".
1960 static void init_status(struct isl_coalesce_info *info)
1962 info->eq = info->ineq = NULL;
1965 /* Set info->eq to the positions of the equalities of info->bmap
1966 * with respect to the basic map represented by "tab".
1967 * If info->eq has already been computed, then do not compute it again.
1969 static void set_eq_status_in(struct isl_coalesce_info *info,
1970 struct isl_tab *tab)
1972 if (info->eq)
1973 return;
1974 info->eq = eq_status_in(info->bmap, tab);
1977 /* Set info->ineq to the positions of the inequalities of info->bmap
1978 * with respect to the basic map represented by "tab".
1979 * If info->ineq has already been computed, then do not compute it again.
1981 static void set_ineq_status_in(struct isl_coalesce_info *info,
1982 struct isl_tab *tab)
1984 if (info->ineq)
1985 return;
1986 info->ineq = ineq_status_in(info->bmap, info->tab, tab);
1989 /* Free the memory allocated by the "eq" and "ineq" fields of "info".
1990 * This function assumes that init_status has been called on "info" first,
1991 * after which the "eq" and "ineq" fields may or may not have been
1992 * assigned a newly allocated array.
1994 static void clear_status(struct isl_coalesce_info *info)
1996 free(info->eq);
1997 free(info->ineq);
2000 /* Are all inequality constraints of the basic map represented by "info"
2001 * valid for the other basic map, except for a single constraint
2002 * that is adjacent to an inequality constraint of the other basic map?
2004 static int all_ineq_valid_or_single_adj_ineq(struct isl_coalesce_info *info)
2006 int i;
2007 int k = -1;
2009 for (i = 0; i < info->bmap->n_ineq; ++i) {
2010 if (info->ineq[i] == STATUS_REDUNDANT)
2011 continue;
2012 if (info->ineq[i] == STATUS_VALID)
2013 continue;
2014 if (info->ineq[i] != STATUS_ADJ_INEQ)
2015 return 0;
2016 if (k != -1)
2017 return 0;
2018 k = i;
2021 return k != -1;
2024 /* Basic map "i" has one or more equality constraints that separate it
2025 * from basic map "j". Check if it happens to be an extension
2026 * of basic map "j".
2027 * In particular, check that all constraints of "j" are valid for "i",
2028 * except for one inequality constraint that is adjacent
2029 * to an inequality constraints of "i".
2030 * If so, check for "i" being an extension of "j" by calling
2031 * is_adj_ineq_extension.
2033 * Clean up the memory allocated for keeping track of the status
2034 * of the constraints before returning.
2036 static enum isl_change separating_equality(int i, int j,
2037 struct isl_coalesce_info *info)
2039 enum isl_change change = isl_change_none;
2041 if (all(info[j].eq, 2 * info[j].bmap->n_eq, STATUS_VALID) &&
2042 all_ineq_valid_or_single_adj_ineq(&info[j]))
2043 change = is_adj_ineq_extension(j, i, info);
2045 clear_status(&info[i]);
2046 clear_status(&info[j]);
2047 return change;
2050 /* Check if the union of the given pair of basic maps
2051 * can be represented by a single basic map.
2052 * If so, replace the pair by the single basic map and return
2053 * isl_change_drop_first, isl_change_drop_second or isl_change_fuse.
2054 * Otherwise, return isl_change_none.
2055 * The two basic maps are assumed to live in the same local space.
2056 * The "eq" and "ineq" fields of info[i] and info[j] are assumed
2057 * to have been initialized by the caller, either to NULL or
2058 * to valid information.
2060 * We first check the effect of each constraint of one basic map
2061 * on the other basic map.
2062 * The constraint may be
2063 * redundant the constraint is redundant in its own
2064 * basic map and should be ignore and removed
2065 * in the end
2066 * valid all (integer) points of the other basic map
2067 * satisfy the constraint
2068 * separate no (integer) point of the other basic map
2069 * satisfies the constraint
2070 * cut some but not all points of the other basic map
2071 * satisfy the constraint
2072 * adj_eq the given constraint is adjacent (on the outside)
2073 * to an equality of the other basic map
2074 * adj_ineq the given constraint is adjacent (on the outside)
2075 * to an inequality of the other basic map
2077 * We consider seven cases in which we can replace the pair by a single
2078 * basic map. We ignore all "redundant" constraints.
2080 * 1. all constraints of one basic map are valid
2081 * => the other basic map is a subset and can be removed
2083 * 2. all constraints of both basic maps are either "valid" or "cut"
2084 * and the facets corresponding to the "cut" constraints
2085 * of one of the basic maps lies entirely inside the other basic map
2086 * => the pair can be replaced by a basic map consisting
2087 * of the valid constraints in both basic maps
2089 * 3. there is a single pair of adjacent inequalities
2090 * (all other constraints are "valid")
2091 * => the pair can be replaced by a basic map consisting
2092 * of the valid constraints in both basic maps
2094 * 4. one basic map has a single adjacent inequality, while the other
2095 * constraints are "valid". The other basic map has some
2096 * "cut" constraints, but replacing the adjacent inequality by
2097 * its opposite and adding the valid constraints of the other
2098 * basic map results in a subset of the other basic map
2099 * => the pair can be replaced by a basic map consisting
2100 * of the valid constraints in both basic maps
2102 * 5. there is a single adjacent pair of an inequality and an equality,
2103 * the other constraints of the basic map containing the inequality are
2104 * "valid". Moreover, if the inequality the basic map is relaxed
2105 * and then turned into an equality, then resulting facet lies
2106 * entirely inside the other basic map
2107 * => the pair can be replaced by the basic map containing
2108 * the inequality, with the inequality relaxed.
2110 * 6. there is a single adjacent pair of an inequality and an equality,
2111 * the other constraints of the basic map containing the inequality are
2112 * "valid". Moreover, the facets corresponding to both
2113 * the inequality and the equality can be wrapped around their
2114 * ridges to include the other basic map
2115 * => the pair can be replaced by a basic map consisting
2116 * of the valid constraints in both basic maps together
2117 * with all wrapping constraints
2119 * 7. one of the basic maps extends beyond the other by at most one.
2120 * Moreover, the facets corresponding to the cut constraints and
2121 * the pieces of the other basic map at offset one from these cut
2122 * constraints can be wrapped around their ridges to include
2123 * the union of the two basic maps
2124 * => the pair can be replaced by a basic map consisting
2125 * of the valid constraints in both basic maps together
2126 * with all wrapping constraints
2128 * 8. the two basic maps live in adjacent hyperplanes. In principle
2129 * such sets can always be combined through wrapping, but we impose
2130 * that there is only one such pair, to avoid overeager coalescing.
2132 * Throughout the computation, we maintain a collection of tableaus
2133 * corresponding to the basic maps. When the basic maps are dropped
2134 * or combined, the tableaus are modified accordingly.
2136 static enum isl_change coalesce_local_pair_reuse(int i, int j,
2137 struct isl_coalesce_info *info)
2139 enum isl_change change = isl_change_none;
2141 set_ineq_status_in(&info[i], info[j].tab);
2142 if (info[i].bmap->n_ineq && !info[i].ineq)
2143 goto error;
2144 if (any(info[i].ineq, info[i].bmap->n_ineq, STATUS_ERROR))
2145 goto error;
2146 if (any(info[i].ineq, info[i].bmap->n_ineq, STATUS_SEPARATE))
2147 goto done;
2149 set_ineq_status_in(&info[j], info[i].tab);
2150 if (info[j].bmap->n_ineq && !info[j].ineq)
2151 goto error;
2152 if (any(info[j].ineq, info[j].bmap->n_ineq, STATUS_ERROR))
2153 goto error;
2154 if (any(info[j].ineq, info[j].bmap->n_ineq, STATUS_SEPARATE))
2155 goto done;
2157 set_eq_status_in(&info[i], info[j].tab);
2158 if (info[i].bmap->n_eq && !info[i].eq)
2159 goto error;
2160 if (any(info[i].eq, 2 * info[i].bmap->n_eq, STATUS_ERROR))
2161 goto error;
2163 set_eq_status_in(&info[j], info[i].tab);
2164 if (info[j].bmap->n_eq && !info[j].eq)
2165 goto error;
2166 if (any(info[j].eq, 2 * info[j].bmap->n_eq, STATUS_ERROR))
2167 goto error;
2169 if (any(info[i].eq, 2 * info[i].bmap->n_eq, STATUS_SEPARATE))
2170 return separating_equality(i, j, info);
2171 if (any(info[j].eq, 2 * info[j].bmap->n_eq, STATUS_SEPARATE))
2172 return separating_equality(j, i, info);
2174 if (all(info[i].eq, 2 * info[i].bmap->n_eq, STATUS_VALID) &&
2175 all(info[i].ineq, info[i].bmap->n_ineq, STATUS_VALID)) {
2176 drop(&info[j]);
2177 change = isl_change_drop_second;
2178 } else if (all(info[j].eq, 2 * info[j].bmap->n_eq, STATUS_VALID) &&
2179 all(info[j].ineq, info[j].bmap->n_ineq, STATUS_VALID)) {
2180 drop(&info[i]);
2181 change = isl_change_drop_first;
2182 } else if (any(info[i].eq, 2 * info[i].bmap->n_eq, STATUS_ADJ_EQ)) {
2183 change = check_eq_adj_eq(i, j, info);
2184 } else if (any(info[j].eq, 2 * info[j].bmap->n_eq, STATUS_ADJ_EQ)) {
2185 change = check_eq_adj_eq(j, i, info);
2186 } else if (any(info[i].eq, 2 * info[i].bmap->n_eq, STATUS_ADJ_INEQ) ||
2187 any(info[j].eq, 2 * info[j].bmap->n_eq, STATUS_ADJ_INEQ)) {
2188 change = check_adj_eq(i, j, info);
2189 } else if (any(info[i].ineq, info[i].bmap->n_ineq, STATUS_ADJ_EQ) ||
2190 any(info[j].ineq, info[j].bmap->n_ineq, STATUS_ADJ_EQ)) {
2191 /* Can't happen */
2192 /* BAD ADJ INEQ */
2193 } else if (any(info[i].ineq, info[i].bmap->n_ineq, STATUS_ADJ_INEQ) ||
2194 any(info[j].ineq, info[j].bmap->n_ineq, STATUS_ADJ_INEQ)) {
2195 change = check_adj_ineq(i, j, info);
2196 } else {
2197 if (!any(info[i].eq, 2 * info[i].bmap->n_eq, STATUS_CUT) &&
2198 !any(info[j].eq, 2 * info[j].bmap->n_eq, STATUS_CUT))
2199 change = check_facets(i, j, info);
2200 if (change == isl_change_none)
2201 change = check_wrap(i, j, info);
2204 done:
2205 clear_status(&info[i]);
2206 clear_status(&info[j]);
2207 return change;
2208 error:
2209 clear_status(&info[i]);
2210 clear_status(&info[j]);
2211 return isl_change_error;
2214 /* Check if the union of the given pair of basic maps
2215 * can be represented by a single basic map.
2216 * If so, replace the pair by the single basic map and return
2217 * isl_change_drop_first, isl_change_drop_second or isl_change_fuse.
2218 * Otherwise, return isl_change_none.
2219 * The two basic maps are assumed to live in the same local space.
2221 static enum isl_change coalesce_local_pair(int i, int j,
2222 struct isl_coalesce_info *info)
2224 init_status(&info[i]);
2225 init_status(&info[j]);
2226 return coalesce_local_pair_reuse(i, j, info);
2229 /* Shift the integer division at position "div" of the basic map
2230 * represented by "info" by "shift".
2232 * That is, if the integer division has the form
2234 * floor(f(x)/d)
2236 * then replace it by
2238 * floor((f(x) + shift * d)/d) - shift
2240 static isl_stat shift_div(struct isl_coalesce_info *info, int div,
2241 isl_int shift)
2243 unsigned total;
2245 info->bmap = isl_basic_map_shift_div(info->bmap, div, 0, shift);
2246 if (!info->bmap)
2247 return isl_stat_error;
2249 total = isl_basic_map_dim(info->bmap, isl_dim_all);
2250 total -= isl_basic_map_dim(info->bmap, isl_dim_div);
2251 if (isl_tab_shift_var(info->tab, total + div, shift) < 0)
2252 return isl_stat_error;
2254 return isl_stat_ok;
2257 /* If the integer division at position "div" is defined by an equality,
2258 * i.e., a stride constraint, then change the integer division expression
2259 * to have a constant term equal to zero.
2261 * Let the equality constraint be
2263 * c + f + m a = 0
2265 * The integer division expression is then of the form
2267 * a = floor((-f - c')/m)
2269 * The integer division is first shifted by t = floor(c/m),
2270 * turning the equality constraint into
2272 * c - m floor(c/m) + f + m a' = 0
2274 * i.e.,
2276 * (c mod m) + f + m a' = 0
2278 * That is,
2280 * a' = (-f - (c mod m))/m = floor((-f)/m)
2282 * because a' is an integer and 0 <= (c mod m) < m.
2283 * The constant term of a' can therefore be zeroed out.
2285 static isl_stat normalize_stride_div(struct isl_coalesce_info *info, int div)
2287 isl_bool defined;
2288 isl_stat r;
2289 isl_constraint *c;
2290 isl_int shift, stride;
2292 defined = isl_basic_map_has_defining_equality(info->bmap, isl_dim_div,
2293 div, &c);
2294 if (defined < 0)
2295 return isl_stat_error;
2296 if (!defined)
2297 return isl_stat_ok;
2298 if (!c)
2299 return isl_stat_error;
2300 isl_int_init(shift);
2301 isl_int_init(stride);
2302 isl_constraint_get_constant(c, &shift);
2303 isl_constraint_get_coefficient(c, isl_dim_div, div, &stride);
2304 isl_int_fdiv_q(shift, shift, stride);
2305 r = shift_div(info, div, shift);
2306 isl_int_clear(stride);
2307 isl_int_clear(shift);
2308 isl_constraint_free(c);
2309 if (r < 0)
2310 return isl_stat_error;
2311 info->bmap = isl_basic_map_set_div_expr_constant_num_si_inplace(
2312 info->bmap, div, 0);
2313 if (!info->bmap)
2314 return isl_stat_error;
2315 return isl_stat_ok;
2318 /* The basic maps represented by "info1" and "info2" are known
2319 * to have the same number of integer divisions.
2320 * Check if pairs of integer divisions are equal to each other
2321 * despite the fact that they differ by a rational constant.
2323 * In particular, look for any pair of integer divisions that
2324 * only differ in their constant terms.
2325 * If either of these integer divisions is defined
2326 * by stride constraints, then modify it to have a zero constant term.
2327 * If both are defined by stride constraints then in the end they will have
2328 * the same (zero) constant term.
2330 static isl_stat harmonize_stride_divs(struct isl_coalesce_info *info1,
2331 struct isl_coalesce_info *info2)
2333 int i, n;
2334 int total;
2336 total = isl_basic_map_total_dim(info1->bmap);
2337 n = isl_basic_map_dim(info1->bmap, isl_dim_div);
2338 for (i = 0; i < n; ++i) {
2339 isl_bool known, harmonize;
2341 known = isl_basic_map_div_is_known(info1->bmap, i);
2342 if (known >= 0 && known)
2343 known = isl_basic_map_div_is_known(info2->bmap, i);
2344 if (known < 0)
2345 return isl_stat_error;
2346 if (!known)
2347 continue;
2348 harmonize = isl_basic_map_equal_div_expr_except_constant(
2349 info1->bmap, i, info2->bmap, i);
2350 if (harmonize < 0)
2351 return isl_stat_error;
2352 if (!harmonize)
2353 continue;
2354 if (normalize_stride_div(info1, i) < 0)
2355 return isl_stat_error;
2356 if (normalize_stride_div(info2, i) < 0)
2357 return isl_stat_error;
2360 return isl_stat_ok;
2363 /* If "shift" is an integer constant, then shift the integer division
2364 * at position "div" of the basic map represented by "info" by "shift".
2365 * If "shift" is not an integer constant, then do nothing.
2366 * If "shift" is equal to zero, then no shift needs to be performed either.
2368 * That is, if the integer division has the form
2370 * floor(f(x)/d)
2372 * then replace it by
2374 * floor((f(x) + shift * d)/d) - shift
2376 static isl_stat shift_if_cst_int(struct isl_coalesce_info *info, int div,
2377 __isl_keep isl_aff *shift)
2379 isl_bool cst;
2380 isl_stat r;
2381 isl_int d;
2382 isl_val *c;
2384 cst = isl_aff_is_cst(shift);
2385 if (cst < 0 || !cst)
2386 return cst < 0 ? isl_stat_error : isl_stat_ok;
2388 c = isl_aff_get_constant_val(shift);
2389 cst = isl_val_is_int(c);
2390 if (cst >= 0 && cst)
2391 cst = isl_bool_not(isl_val_is_zero(c));
2392 if (cst < 0 || !cst) {
2393 isl_val_free(c);
2394 return cst < 0 ? isl_stat_error : isl_stat_ok;
2397 isl_int_init(d);
2398 r = isl_val_get_num_isl_int(c, &d);
2399 if (r >= 0)
2400 r = shift_div(info, div, d);
2401 isl_int_clear(d);
2403 isl_val_free(c);
2405 return r;
2408 /* Check if some of the divs in the basic map represented by "info1"
2409 * are shifts of the corresponding divs in the basic map represented
2410 * by "info2", taking into account the equality constraints "eq1" of "info1"
2411 * and "eq2" of "info2". If so, align them with those of "info2".
2412 * "info1" and "info2" are assumed to have the same number
2413 * of integer divisions.
2415 * An integer division is considered to be a shift of another integer
2416 * division if, after simplification with respect to the equality
2417 * constraints of the other basic map, one is equal to the other
2418 * plus a constant.
2420 * In particular, for each pair of integer divisions, if both are known,
2421 * have the same denominator and are not already equal to each other,
2422 * simplify each with respect to the equality constraints
2423 * of the other basic map. If the difference is an integer constant,
2424 * then move this difference outside.
2425 * That is, if, after simplification, one integer division is of the form
2427 * floor((f(x) + c_1)/d)
2429 * while the other is of the form
2431 * floor((f(x) + c_2)/d)
2433 * and n = (c_2 - c_1)/d is an integer, then replace the first
2434 * integer division by
2436 * floor((f_1(x) + c_1 + n * d)/d) - n,
2438 * where floor((f_1(x) + c_1 + n * d)/d) = floor((f2(x) + c_2)/d)
2439 * after simplification with respect to the equality constraints.
2441 static isl_stat harmonize_divs_with_hulls(struct isl_coalesce_info *info1,
2442 struct isl_coalesce_info *info2, __isl_keep isl_basic_set *eq1,
2443 __isl_keep isl_basic_set *eq2)
2445 int i;
2446 int total;
2447 isl_local_space *ls1, *ls2;
2449 total = isl_basic_map_total_dim(info1->bmap);
2450 ls1 = isl_local_space_wrap(isl_basic_map_get_local_space(info1->bmap));
2451 ls2 = isl_local_space_wrap(isl_basic_map_get_local_space(info2->bmap));
2452 for (i = 0; i < info1->bmap->n_div; ++i) {
2453 isl_stat r;
2454 isl_aff *div1, *div2;
2456 if (!isl_local_space_div_is_known(ls1, i) ||
2457 !isl_local_space_div_is_known(ls2, i))
2458 continue;
2459 if (isl_int_ne(info1->bmap->div[i][0], info2->bmap->div[i][0]))
2460 continue;
2461 if (isl_seq_eq(info1->bmap->div[i] + 1,
2462 info2->bmap->div[i] + 1, 1 + total))
2463 continue;
2464 div1 = isl_local_space_get_div(ls1, i);
2465 div2 = isl_local_space_get_div(ls2, i);
2466 div1 = isl_aff_substitute_equalities(div1,
2467 isl_basic_set_copy(eq2));
2468 div2 = isl_aff_substitute_equalities(div2,
2469 isl_basic_set_copy(eq1));
2470 div2 = isl_aff_sub(div2, div1);
2471 r = shift_if_cst_int(info1, i, div2);
2472 isl_aff_free(div2);
2473 if (r < 0)
2474 break;
2476 isl_local_space_free(ls1);
2477 isl_local_space_free(ls2);
2479 if (i < info1->bmap->n_div)
2480 return isl_stat_error;
2481 return isl_stat_ok;
2484 /* Check if some of the divs in the basic map represented by "info1"
2485 * are shifts of the corresponding divs in the basic map represented
2486 * by "info2". If so, align them with those of "info2".
2487 * Only do this if "info1" and "info2" have the same number
2488 * of integer divisions.
2490 * An integer division is considered to be a shift of another integer
2491 * division if, after simplification with respect to the equality
2492 * constraints of the other basic map, one is equal to the other
2493 * plus a constant.
2495 * First check if pairs of integer divisions are equal to each other
2496 * despite the fact that they differ by a rational constant.
2497 * If so, try and arrange for them to have the same constant term.
2499 * Then, extract the equality constraints and continue with
2500 * harmonize_divs_with_hulls.
2502 * If the equality constraints of both basic maps are the same,
2503 * then there is no need to perform any shifting since
2504 * the coefficients of the integer divisions should have been
2505 * reduced in the same way.
2507 static isl_stat harmonize_divs(struct isl_coalesce_info *info1,
2508 struct isl_coalesce_info *info2)
2510 isl_bool equal;
2511 isl_basic_map *bmap1, *bmap2;
2512 isl_basic_set *eq1, *eq2;
2513 isl_stat r;
2515 if (!info1->bmap || !info2->bmap)
2516 return isl_stat_error;
2518 if (info1->bmap->n_div != info2->bmap->n_div)
2519 return isl_stat_ok;
2520 if (info1->bmap->n_div == 0)
2521 return isl_stat_ok;
2523 if (harmonize_stride_divs(info1, info2) < 0)
2524 return isl_stat_error;
2526 bmap1 = isl_basic_map_copy(info1->bmap);
2527 bmap2 = isl_basic_map_copy(info2->bmap);
2528 eq1 = isl_basic_map_wrap(isl_basic_map_plain_affine_hull(bmap1));
2529 eq2 = isl_basic_map_wrap(isl_basic_map_plain_affine_hull(bmap2));
2530 equal = isl_basic_set_plain_is_equal(eq1, eq2);
2531 if (equal < 0)
2532 r = isl_stat_error;
2533 else if (equal)
2534 r = isl_stat_ok;
2535 else
2536 r = harmonize_divs_with_hulls(info1, info2, eq1, eq2);
2537 isl_basic_set_free(eq1);
2538 isl_basic_set_free(eq2);
2540 return r;
2543 /* Do the two basic maps live in the same local space, i.e.,
2544 * do they have the same (known) divs?
2545 * If either basic map has any unknown divs, then we can only assume
2546 * that they do not live in the same local space.
2548 static isl_bool same_divs(__isl_keep isl_basic_map *bmap1,
2549 __isl_keep isl_basic_map *bmap2)
2551 int i;
2552 isl_bool known;
2553 int total;
2555 if (!bmap1 || !bmap2)
2556 return isl_bool_error;
2557 if (bmap1->n_div != bmap2->n_div)
2558 return isl_bool_false;
2560 if (bmap1->n_div == 0)
2561 return isl_bool_true;
2563 known = isl_basic_map_divs_known(bmap1);
2564 if (known < 0 || !known)
2565 return known;
2566 known = isl_basic_map_divs_known(bmap2);
2567 if (known < 0 || !known)
2568 return known;
2570 total = isl_basic_map_total_dim(bmap1);
2571 for (i = 0; i < bmap1->n_div; ++i)
2572 if (!isl_seq_eq(bmap1->div[i], bmap2->div[i], 2 + total))
2573 return 0;
2575 return 1;
2578 /* Assuming that "tab" contains the equality constraints and
2579 * the initial inequality constraints of "bmap", copy the remaining
2580 * inequality constraints of "bmap" to "Tab".
2582 static isl_stat copy_ineq(struct isl_tab *tab, __isl_keep isl_basic_map *bmap)
2584 int i, n_ineq;
2586 if (!bmap)
2587 return isl_stat_error;
2589 n_ineq = tab->n_con - tab->n_eq;
2590 for (i = n_ineq; i < bmap->n_ineq; ++i)
2591 if (isl_tab_add_ineq(tab, bmap->ineq[i]) < 0)
2592 return isl_stat_error;
2594 return isl_stat_ok;
2597 /* Description of an integer division that is added
2598 * during an expansion.
2599 * "pos" is the position of the corresponding variable.
2600 * "cst" indicates whether this integer division has a fixed value.
2601 * "val" contains the fixed value, if the value is fixed.
2603 struct isl_expanded {
2604 int pos;
2605 isl_bool cst;
2606 isl_int val;
2609 /* For each of the "n" integer division variables "expanded",
2610 * if the variable has a fixed value, then add two inequality
2611 * constraints expressing the fixed value.
2612 * Otherwise, add the corresponding div constraints.
2613 * The caller is responsible for removing the div constraints
2614 * that it added for all these "n" integer divisions.
2616 * The div constraints and the pair of inequality constraints
2617 * forcing the fixed value cannot both be added for a given variable
2618 * as the combination may render some of the original constraints redundant.
2619 * These would then be ignored during the coalescing detection,
2620 * while they could remain in the fused result.
2622 * The two added inequality constraints are
2624 * -a + v >= 0
2625 * a - v >= 0
2627 * with "a" the variable and "v" its fixed value.
2628 * The facet corresponding to one of these two constraints is selected
2629 * in the tableau to ensure that the pair of inequality constraints
2630 * is treated as an equality constraint.
2632 * The information in info->ineq is thrown away because it was
2633 * computed in terms of div constraints, while some of those
2634 * have now been replaced by these pairs of inequality constraints.
2636 static isl_stat fix_constant_divs(struct isl_coalesce_info *info,
2637 int n, struct isl_expanded *expanded)
2639 unsigned o_div;
2640 int i;
2641 isl_vec *ineq;
2643 o_div = isl_basic_map_offset(info->bmap, isl_dim_div) - 1;
2644 ineq = isl_vec_alloc(isl_tab_get_ctx(info->tab), 1 + info->tab->n_var);
2645 if (!ineq)
2646 return isl_stat_error;
2647 isl_seq_clr(ineq->el + 1, info->tab->n_var);
2649 for (i = 0; i < n; ++i) {
2650 if (!expanded[i].cst) {
2651 info->bmap = isl_basic_map_extend_constraints(
2652 info->bmap, 0, 2);
2653 if (isl_basic_map_add_div_constraints(info->bmap,
2654 expanded[i].pos - o_div) < 0)
2655 break;
2656 } else {
2657 isl_int_set_si(ineq->el[1 + expanded[i].pos], -1);
2658 isl_int_set(ineq->el[0], expanded[i].val);
2659 info->bmap = isl_basic_map_add_ineq(info->bmap,
2660 ineq->el);
2661 isl_int_set_si(ineq->el[1 + expanded[i].pos], 1);
2662 isl_int_neg(ineq->el[0], expanded[i].val);
2663 info->bmap = isl_basic_map_add_ineq(info->bmap,
2664 ineq->el);
2665 isl_int_set_si(ineq->el[1 + expanded[i].pos], 0);
2667 if (copy_ineq(info->tab, info->bmap) < 0)
2668 break;
2669 if (expanded[i].cst &&
2670 isl_tab_select_facet(info->tab, info->tab->n_con - 1) < 0)
2671 break;
2674 isl_vec_free(ineq);
2676 clear_status(info);
2677 init_status(info);
2679 return i < n ? isl_stat_error : isl_stat_ok;
2682 /* Insert the "n" integer division variables "expanded"
2683 * into info->tab and info->bmap and
2684 * update info->ineq with respect to the redundant constraints
2685 * in the resulting tableau.
2686 * "bmap" contains the result of this insertion in info->bmap,
2687 * while info->bmap is the original version
2688 * of "bmap", i.e., the one that corresponds to the current
2689 * state of info->tab. The number of constraints in info->bmap
2690 * is assumed to be the same as the number of constraints
2691 * in info->tab. This is required to be able to detect
2692 * the extra constraints in "bmap".
2694 * In particular, introduce extra variables corresponding
2695 * to the extra integer divisions and add the div constraints
2696 * that were added to "bmap" after info->tab was created
2697 * from info->bmap.
2698 * Furthermore, check if these extra integer divisions happen
2699 * to attain a fixed integer value in info->tab.
2700 * If so, replace the corresponding div constraints by pairs
2701 * of inequality constraints that fix these
2702 * integer divisions to their single integer values.
2703 * Replace info->bmap by "bmap" to match the changes to info->tab.
2704 * info->ineq was computed without a tableau and therefore
2705 * does not take into account the redundant constraints
2706 * in the tableau. Mark them here.
2707 * There is no need to check the newly added div constraints
2708 * since they cannot be redundant.
2709 * The redundancy check is not performed when constants have been discovered
2710 * since info->ineq is completely thrown away in this case.
2712 static isl_stat tab_insert_divs(struct isl_coalesce_info *info,
2713 int n, struct isl_expanded *expanded, __isl_take isl_basic_map *bmap)
2715 int i, n_ineq;
2716 unsigned n_eq;
2717 struct isl_tab_undo *snap;
2718 int any;
2720 if (!bmap)
2721 return isl_stat_error;
2722 if (info->bmap->n_eq + info->bmap->n_ineq != info->tab->n_con)
2723 isl_die(isl_basic_map_get_ctx(bmap), isl_error_internal,
2724 "original tableau does not correspond "
2725 "to original basic map", goto error);
2727 if (isl_tab_extend_vars(info->tab, n) < 0)
2728 goto error;
2729 if (isl_tab_extend_cons(info->tab, 2 * n) < 0)
2730 goto error;
2732 for (i = 0; i < n; ++i) {
2733 if (isl_tab_insert_var(info->tab, expanded[i].pos) < 0)
2734 goto error;
2737 snap = isl_tab_snap(info->tab);
2739 n_ineq = info->tab->n_con - info->tab->n_eq;
2740 if (copy_ineq(info->tab, bmap) < 0)
2741 goto error;
2743 isl_basic_map_free(info->bmap);
2744 info->bmap = bmap;
2746 any = 0;
2747 for (i = 0; i < n; ++i) {
2748 expanded[i].cst = isl_tab_is_constant(info->tab,
2749 expanded[i].pos, &expanded[i].val);
2750 if (expanded[i].cst < 0)
2751 return isl_stat_error;
2752 if (expanded[i].cst)
2753 any = 1;
2756 if (any) {
2757 if (isl_tab_rollback(info->tab, snap) < 0)
2758 return isl_stat_error;
2759 info->bmap = isl_basic_map_cow(info->bmap);
2760 if (isl_basic_map_free_inequality(info->bmap, 2 * n) < 0)
2761 return isl_stat_error;
2763 return fix_constant_divs(info, n, expanded);
2766 n_eq = info->bmap->n_eq;
2767 for (i = 0; i < n_ineq; ++i) {
2768 if (isl_tab_is_redundant(info->tab, n_eq + i))
2769 info->ineq[i] = STATUS_REDUNDANT;
2772 return isl_stat_ok;
2773 error:
2774 isl_basic_map_free(bmap);
2775 return isl_stat_error;
2778 /* Expand info->tab and info->bmap in the same way "bmap" was expanded
2779 * in isl_basic_map_expand_divs using the expansion "exp" and
2780 * update info->ineq with respect to the redundant constraints
2781 * in the resulting tableau. info->bmap is the original version
2782 * of "bmap", i.e., the one that corresponds to the current
2783 * state of info->tab. The number of constraints in info->bmap
2784 * is assumed to be the same as the number of constraints
2785 * in info->tab. This is required to be able to detect
2786 * the extra constraints in "bmap".
2788 * Extract the positions where extra local variables are introduced
2789 * from "exp" and call tab_insert_divs.
2791 static isl_stat expand_tab(struct isl_coalesce_info *info, int *exp,
2792 __isl_take isl_basic_map *bmap)
2794 isl_ctx *ctx;
2795 struct isl_expanded *expanded;
2796 int i, j, k, n;
2797 int extra_var;
2798 unsigned total, pos, n_div;
2799 isl_stat r;
2801 total = isl_basic_map_dim(bmap, isl_dim_all);
2802 n_div = isl_basic_map_dim(bmap, isl_dim_div);
2803 pos = total - n_div;
2804 extra_var = total - info->tab->n_var;
2805 n = n_div - extra_var;
2807 ctx = isl_basic_map_get_ctx(bmap);
2808 expanded = isl_calloc_array(ctx, struct isl_expanded, extra_var);
2809 if (extra_var && !expanded)
2810 goto error;
2812 i = 0;
2813 k = 0;
2814 for (j = 0; j < n_div; ++j) {
2815 if (i < n && exp[i] == j) {
2816 ++i;
2817 continue;
2819 expanded[k++].pos = pos + j;
2822 for (k = 0; k < extra_var; ++k)
2823 isl_int_init(expanded[k].val);
2825 r = tab_insert_divs(info, extra_var, expanded, bmap);
2827 for (k = 0; k < extra_var; ++k)
2828 isl_int_clear(expanded[k].val);
2829 free(expanded);
2831 return r;
2832 error:
2833 isl_basic_map_free(bmap);
2834 return isl_stat_error;
2837 /* Check if the union of the basic maps represented by info[i] and info[j]
2838 * can be represented by a single basic map,
2839 * after expanding the divs of info[i] to match those of info[j].
2840 * If so, replace the pair by the single basic map and return
2841 * isl_change_drop_first, isl_change_drop_second or isl_change_fuse.
2842 * Otherwise, return isl_change_none.
2844 * The caller has already checked for info[j] being a subset of info[i].
2845 * If some of the divs of info[j] are unknown, then the expanded info[i]
2846 * will not have the corresponding div constraints. The other patterns
2847 * therefore cannot apply. Skip the computation in this case.
2849 * The expansion is performed using the divs "div" and expansion "exp"
2850 * computed by the caller.
2851 * info[i].bmap has already been expanded and the result is passed in
2852 * as "bmap".
2853 * The "eq" and "ineq" fields of info[i] reflect the status of
2854 * the constraints of the expanded "bmap" with respect to info[j].tab.
2855 * However, inequality constraints that are redundant in info[i].tab
2856 * have not yet been marked as such because no tableau was available.
2858 * Replace info[i].bmap by "bmap" and expand info[i].tab as well,
2859 * updating info[i].ineq with respect to the redundant constraints.
2860 * Then try and coalesce the expanded info[i] with info[j],
2861 * reusing the information in info[i].eq and info[i].ineq.
2862 * If this does not result in any coalescing or if it results in info[j]
2863 * getting dropped (which should not happen in practice, since the case
2864 * of info[j] being a subset of info[i] has already been checked by
2865 * the caller), then revert info[i] to its original state.
2867 static enum isl_change coalesce_expand_tab_divs(__isl_take isl_basic_map *bmap,
2868 int i, int j, struct isl_coalesce_info *info, __isl_keep isl_mat *div,
2869 int *exp)
2871 isl_bool known;
2872 isl_basic_map *bmap_i;
2873 struct isl_tab_undo *snap;
2874 enum isl_change change = isl_change_none;
2876 known = isl_basic_map_divs_known(info[j].bmap);
2877 if (known < 0 || !known) {
2878 clear_status(&info[i]);
2879 isl_basic_map_free(bmap);
2880 return known < 0 ? isl_change_error : isl_change_none;
2883 bmap_i = isl_basic_map_copy(info[i].bmap);
2884 snap = isl_tab_snap(info[i].tab);
2885 if (expand_tab(&info[i], exp, bmap) < 0)
2886 change = isl_change_error;
2888 init_status(&info[j]);
2889 if (change == isl_change_none)
2890 change = coalesce_local_pair_reuse(i, j, info);
2891 else
2892 clear_status(&info[i]);
2893 if (change != isl_change_none && change != isl_change_drop_second) {
2894 isl_basic_map_free(bmap_i);
2895 } else {
2896 isl_basic_map_free(info[i].bmap);
2897 info[i].bmap = bmap_i;
2899 if (isl_tab_rollback(info[i].tab, snap) < 0)
2900 change = isl_change_error;
2903 return change;
2906 /* Check if the union of "bmap" and the basic map represented by info[j]
2907 * can be represented by a single basic map,
2908 * after expanding the divs of "bmap" to match those of info[j].
2909 * If so, replace the pair by the single basic map and return
2910 * isl_change_drop_first, isl_change_drop_second or isl_change_fuse.
2911 * Otherwise, return isl_change_none.
2913 * In particular, check if the expanded "bmap" contains the basic map
2914 * represented by the tableau info[j].tab.
2915 * The expansion is performed using the divs "div" and expansion "exp"
2916 * computed by the caller.
2917 * Then we check if all constraints of the expanded "bmap" are valid for
2918 * info[j].tab.
2920 * If "i" is not equal to -1, then "bmap" is equal to info[i].bmap.
2921 * In this case, the positions of the constraints of info[i].bmap
2922 * with respect to the basic map represented by info[j] are stored
2923 * in info[i].
2925 * If the expanded "bmap" does not contain the basic map
2926 * represented by the tableau info[j].tab and if "i" is not -1,
2927 * i.e., if the original "bmap" is info[i].bmap, then expand info[i].tab
2928 * as well and check if that results in coalescing.
2930 static enum isl_change coalesce_with_expanded_divs(
2931 __isl_keep isl_basic_map *bmap, int i, int j,
2932 struct isl_coalesce_info *info, __isl_keep isl_mat *div, int *exp)
2934 enum isl_change change = isl_change_none;
2935 struct isl_coalesce_info info_local, *info_i;
2937 info_i = i >= 0 ? &info[i] : &info_local;
2938 init_status(info_i);
2939 bmap = isl_basic_map_copy(bmap);
2940 bmap = isl_basic_map_expand_divs(bmap, isl_mat_copy(div), exp);
2941 bmap = isl_basic_map_mark_final(bmap);
2943 if (!bmap)
2944 goto error;
2946 info_i->eq = eq_status_in(bmap, info[j].tab);
2947 if (bmap->n_eq && !info_i->eq)
2948 goto error;
2949 if (any(info_i->eq, 2 * bmap->n_eq, STATUS_ERROR))
2950 goto error;
2951 if (any(info_i->eq, 2 * bmap->n_eq, STATUS_SEPARATE))
2952 goto done;
2954 info_i->ineq = ineq_status_in(bmap, NULL, info[j].tab);
2955 if (bmap->n_ineq && !info_i->ineq)
2956 goto error;
2957 if (any(info_i->ineq, bmap->n_ineq, STATUS_ERROR))
2958 goto error;
2959 if (any(info_i->ineq, bmap->n_ineq, STATUS_SEPARATE))
2960 goto done;
2962 if (all(info_i->eq, 2 * bmap->n_eq, STATUS_VALID) &&
2963 all(info_i->ineq, bmap->n_ineq, STATUS_VALID)) {
2964 drop(&info[j]);
2965 change = isl_change_drop_second;
2968 if (change == isl_change_none && i != -1)
2969 return coalesce_expand_tab_divs(bmap, i, j, info, div, exp);
2971 done:
2972 isl_basic_map_free(bmap);
2973 clear_status(info_i);
2974 return change;
2975 error:
2976 isl_basic_map_free(bmap);
2977 clear_status(info_i);
2978 return isl_change_error;
2981 /* Check if the union of "bmap_i" and the basic map represented by info[j]
2982 * can be represented by a single basic map,
2983 * after aligning the divs of "bmap_i" to match those of info[j].
2984 * If so, replace the pair by the single basic map and return
2985 * isl_change_drop_first, isl_change_drop_second or isl_change_fuse.
2986 * Otherwise, return isl_change_none.
2988 * In particular, check if "bmap_i" contains the basic map represented by
2989 * info[j] after aligning the divs of "bmap_i" to those of info[j].
2990 * Note that this can only succeed if the number of divs of "bmap_i"
2991 * is smaller than (or equal to) the number of divs of info[j].
2993 * We first check if the divs of "bmap_i" are all known and form a subset
2994 * of those of info[j].bmap. If so, we pass control over to
2995 * coalesce_with_expanded_divs.
2997 * If "i" is not equal to -1, then "bmap" is equal to info[i].bmap.
2999 static enum isl_change coalesce_after_aligning_divs(
3000 __isl_keep isl_basic_map *bmap_i, int i, int j,
3001 struct isl_coalesce_info *info)
3003 int known;
3004 isl_mat *div_i, *div_j, *div;
3005 int *exp1 = NULL;
3006 int *exp2 = NULL;
3007 isl_ctx *ctx;
3008 enum isl_change change;
3010 known = isl_basic_map_divs_known(bmap_i);
3011 if (known < 0 || !known)
3012 return known;
3014 ctx = isl_basic_map_get_ctx(bmap_i);
3016 div_i = isl_basic_map_get_divs(bmap_i);
3017 div_j = isl_basic_map_get_divs(info[j].bmap);
3019 if (!div_i || !div_j)
3020 goto error;
3022 exp1 = isl_alloc_array(ctx, int, div_i->n_row);
3023 exp2 = isl_alloc_array(ctx, int, div_j->n_row);
3024 if ((div_i->n_row && !exp1) || (div_j->n_row && !exp2))
3025 goto error;
3027 div = isl_merge_divs(div_i, div_j, exp1, exp2);
3028 if (!div)
3029 goto error;
3031 if (div->n_row == div_j->n_row)
3032 change = coalesce_with_expanded_divs(bmap_i,
3033 i, j, info, div, exp1);
3034 else
3035 change = isl_change_none;
3037 isl_mat_free(div);
3039 isl_mat_free(div_i);
3040 isl_mat_free(div_j);
3042 free(exp2);
3043 free(exp1);
3045 return change;
3046 error:
3047 isl_mat_free(div_i);
3048 isl_mat_free(div_j);
3049 free(exp1);
3050 free(exp2);
3051 return isl_change_error;
3054 /* Check if basic map "j" is a subset of basic map "i" after
3055 * exploiting the extra equalities of "j" to simplify the divs of "i".
3056 * If so, remove basic map "j" and return isl_change_drop_second.
3058 * If "j" does not have any equalities or if they are the same
3059 * as those of "i", then we cannot exploit them to simplify the divs.
3060 * Similarly, if there are no divs in "i", then they cannot be simplified.
3061 * If, on the other hand, the affine hulls of "i" and "j" do not intersect,
3062 * then "j" cannot be a subset of "i".
3064 * Otherwise, we intersect "i" with the affine hull of "j" and then
3065 * check if "j" is a subset of the result after aligning the divs.
3066 * If so, then "j" is definitely a subset of "i" and can be removed.
3067 * Note that if after intersection with the affine hull of "j".
3068 * "i" still has more divs than "j", then there is no way we can
3069 * align the divs of "i" to those of "j".
3071 static enum isl_change coalesce_subset_with_equalities(int i, int j,
3072 struct isl_coalesce_info *info)
3074 isl_basic_map *hull_i, *hull_j, *bmap_i;
3075 int equal, empty;
3076 enum isl_change change;
3078 if (info[j].bmap->n_eq == 0)
3079 return isl_change_none;
3080 if (info[i].bmap->n_div == 0)
3081 return isl_change_none;
3083 hull_i = isl_basic_map_copy(info[i].bmap);
3084 hull_i = isl_basic_map_plain_affine_hull(hull_i);
3085 hull_j = isl_basic_map_copy(info[j].bmap);
3086 hull_j = isl_basic_map_plain_affine_hull(hull_j);
3088 hull_j = isl_basic_map_intersect(hull_j, isl_basic_map_copy(hull_i));
3089 equal = isl_basic_map_plain_is_equal(hull_i, hull_j);
3090 empty = isl_basic_map_plain_is_empty(hull_j);
3091 isl_basic_map_free(hull_i);
3093 if (equal < 0 || equal || empty < 0 || empty) {
3094 isl_basic_map_free(hull_j);
3095 if (equal < 0 || empty < 0)
3096 return isl_change_error;
3097 return isl_change_none;
3100 bmap_i = isl_basic_map_copy(info[i].bmap);
3101 bmap_i = isl_basic_map_intersect(bmap_i, hull_j);
3102 if (!bmap_i)
3103 return isl_change_error;
3105 if (bmap_i->n_div > info[j].bmap->n_div) {
3106 isl_basic_map_free(bmap_i);
3107 return isl_change_none;
3110 change = coalesce_after_aligning_divs(bmap_i, -1, j, info);
3112 isl_basic_map_free(bmap_i);
3114 return change;
3117 /* Check if the union of and the basic maps represented by info[i] and info[j]
3118 * can be represented by a single basic map, by aligning or equating
3119 * their integer divisions.
3120 * If so, replace the pair by the single basic map and return
3121 * isl_change_drop_first, isl_change_drop_second or isl_change_fuse.
3122 * Otherwise, return isl_change_none.
3124 * Note that we only perform any test if the number of divs is different
3125 * in the two basic maps. In case the number of divs is the same,
3126 * we have already established that the divs are different
3127 * in the two basic maps.
3128 * In particular, if the number of divs of basic map i is smaller than
3129 * the number of divs of basic map j, then we check if j is a subset of i
3130 * and vice versa.
3132 static enum isl_change coalesce_divs(int i, int j,
3133 struct isl_coalesce_info *info)
3135 enum isl_change change = isl_change_none;
3137 if (info[i].bmap->n_div < info[j].bmap->n_div)
3138 change = coalesce_after_aligning_divs(info[i].bmap, i, j, info);
3139 if (change != isl_change_none)
3140 return change;
3142 if (info[j].bmap->n_div < info[i].bmap->n_div)
3143 change = coalesce_after_aligning_divs(info[j].bmap, j, i, info);
3144 if (change != isl_change_none)
3145 return invert_change(change);
3147 change = coalesce_subset_with_equalities(i, j, info);
3148 if (change != isl_change_none)
3149 return change;
3151 change = coalesce_subset_with_equalities(j, i, info);
3152 if (change != isl_change_none)
3153 return invert_change(change);
3155 return isl_change_none;
3158 /* Does "bmap" involve any divs that themselves refer to divs?
3160 static isl_bool has_nested_div(__isl_keep isl_basic_map *bmap)
3162 int i;
3163 unsigned total;
3164 unsigned n_div;
3166 total = isl_basic_map_dim(bmap, isl_dim_all);
3167 n_div = isl_basic_map_dim(bmap, isl_dim_div);
3168 total -= n_div;
3170 for (i = 0; i < n_div; ++i)
3171 if (isl_seq_first_non_zero(bmap->div[i] + 2 + total,
3172 n_div) != -1)
3173 return isl_bool_true;
3175 return isl_bool_false;
3178 /* Return a list of affine expressions, one for each integer division
3179 * in "bmap_i". For each integer division that also appears in "bmap_j",
3180 * the affine expression is set to NaN. The number of NaNs in the list
3181 * is equal to the number of integer divisions in "bmap_j".
3182 * For the other integer divisions of "bmap_i", the corresponding
3183 * element in the list is a purely affine expression equal to the integer
3184 * division in "hull".
3185 * If no such list can be constructed, then the number of elements
3186 * in the returned list is smaller than the number of integer divisions
3187 * in "bmap_i".
3189 static __isl_give isl_aff_list *set_up_substitutions(
3190 __isl_keep isl_basic_map *bmap_i, __isl_keep isl_basic_map *bmap_j,
3191 __isl_take isl_basic_map *hull)
3193 unsigned n_div_i, n_div_j, total;
3194 isl_ctx *ctx;
3195 isl_local_space *ls;
3196 isl_basic_set *wrap_hull;
3197 isl_aff *aff_nan;
3198 isl_aff_list *list;
3199 int i, j;
3201 if (!hull)
3202 return NULL;
3204 ctx = isl_basic_map_get_ctx(hull);
3206 n_div_i = isl_basic_map_dim(bmap_i, isl_dim_div);
3207 n_div_j = isl_basic_map_dim(bmap_j, isl_dim_div);
3208 total = isl_basic_map_total_dim(bmap_i) - n_div_i;
3210 ls = isl_basic_map_get_local_space(bmap_i);
3211 ls = isl_local_space_wrap(ls);
3212 wrap_hull = isl_basic_map_wrap(hull);
3214 aff_nan = isl_aff_nan_on_domain(isl_local_space_copy(ls));
3215 list = isl_aff_list_alloc(ctx, n_div_i);
3217 j = 0;
3218 for (i = 0; i < n_div_i; ++i) {
3219 isl_aff *aff;
3221 if (j < n_div_j &&
3222 isl_basic_map_equal_div_expr_part(bmap_i, i, bmap_j, j,
3223 0, 2 + total)) {
3224 ++j;
3225 list = isl_aff_list_add(list, isl_aff_copy(aff_nan));
3226 continue;
3228 if (n_div_i - i <= n_div_j - j)
3229 break;
3231 aff = isl_local_space_get_div(ls, i);
3232 aff = isl_aff_substitute_equalities(aff,
3233 isl_basic_set_copy(wrap_hull));
3234 aff = isl_aff_floor(aff);
3235 if (!aff)
3236 goto error;
3237 if (isl_aff_dim(aff, isl_dim_div) != 0) {
3238 isl_aff_free(aff);
3239 break;
3242 list = isl_aff_list_add(list, aff);
3245 isl_aff_free(aff_nan);
3246 isl_local_space_free(ls);
3247 isl_basic_set_free(wrap_hull);
3249 return list;
3250 error:
3251 isl_aff_free(aff_nan);
3252 isl_local_space_free(ls);
3253 isl_basic_set_free(wrap_hull);
3254 isl_aff_list_free(list);
3255 return NULL;
3258 /* Add variables to info->bmap and info->tab corresponding to the elements
3259 * in "list" that are not set to NaN.
3260 * "extra_var" is the number of these elements.
3261 * "dim" is the offset in the variables of "tab" where we should
3262 * start considering the elements in "list".
3263 * When this function returns, the total number of variables in "tab"
3264 * is equal to "dim" plus the number of elements in "list".
3266 * The newly added existentially quantified variables are not given
3267 * an explicit representation because the corresponding div constraints
3268 * do not appear in info->bmap. These constraints are not added
3269 * to info->bmap because for internal consistency, they would need to
3270 * be added to info->tab as well, where they could combine with the equality
3271 * that is added later to result in constraints that do not hold
3272 * in the original input.
3274 static isl_stat add_sub_vars(struct isl_coalesce_info *info,
3275 __isl_keep isl_aff_list *list, int dim, int extra_var)
3277 int i, j, n, d;
3278 isl_space *space;
3280 space = isl_basic_map_get_space(info->bmap);
3281 info->bmap = isl_basic_map_cow(info->bmap);
3282 info->bmap = isl_basic_map_extend_space(info->bmap, space,
3283 extra_var, 0, 0);
3284 if (!info->bmap)
3285 return isl_stat_error;
3286 n = isl_aff_list_n_aff(list);
3287 for (i = 0; i < n; ++i) {
3288 int is_nan;
3289 isl_aff *aff;
3291 aff = isl_aff_list_get_aff(list, i);
3292 is_nan = isl_aff_is_nan(aff);
3293 isl_aff_free(aff);
3294 if (is_nan < 0)
3295 return isl_stat_error;
3296 if (is_nan)
3297 continue;
3299 if (isl_tab_insert_var(info->tab, dim + i) < 0)
3300 return isl_stat_error;
3301 d = isl_basic_map_alloc_div(info->bmap);
3302 if (d < 0)
3303 return isl_stat_error;
3304 info->bmap = isl_basic_map_mark_div_unknown(info->bmap, d);
3305 if (!info->bmap)
3306 return isl_stat_error;
3307 for (j = d; j > i; --j)
3308 isl_basic_map_swap_div(info->bmap, j - 1, j);
3311 return isl_stat_ok;
3314 /* For each element in "list" that is not set to NaN, fix the corresponding
3315 * variable in "tab" to the purely affine expression defined by the element.
3316 * "dim" is the offset in the variables of "tab" where we should
3317 * start considering the elements in "list".
3319 * This function assumes that a sufficient number of rows and
3320 * elements in the constraint array are available in the tableau.
3322 static int add_sub_equalities(struct isl_tab *tab,
3323 __isl_keep isl_aff_list *list, int dim)
3325 int i, n;
3326 isl_ctx *ctx;
3327 isl_vec *sub;
3328 isl_aff *aff;
3330 n = isl_aff_list_n_aff(list);
3332 ctx = isl_tab_get_ctx(tab);
3333 sub = isl_vec_alloc(ctx, 1 + dim + n);
3334 if (!sub)
3335 return -1;
3336 isl_seq_clr(sub->el + 1 + dim, n);
3338 for (i = 0; i < n; ++i) {
3339 aff = isl_aff_list_get_aff(list, i);
3340 if (!aff)
3341 goto error;
3342 if (isl_aff_is_nan(aff)) {
3343 isl_aff_free(aff);
3344 continue;
3346 isl_seq_cpy(sub->el, aff->v->el + 1, 1 + dim);
3347 isl_int_neg(sub->el[1 + dim + i], aff->v->el[0]);
3348 if (isl_tab_add_eq(tab, sub->el) < 0)
3349 goto error;
3350 isl_int_set_si(sub->el[1 + dim + i], 0);
3351 isl_aff_free(aff);
3354 isl_vec_free(sub);
3355 return 0;
3356 error:
3357 isl_aff_free(aff);
3358 isl_vec_free(sub);
3359 return -1;
3362 /* Add variables to info->tab and info->bmap corresponding to the elements
3363 * in "list" that are not set to NaN. The value of the added variable
3364 * in info->tab is fixed to the purely affine expression defined by the element.
3365 * "dim" is the offset in the variables of info->tab where we should
3366 * start considering the elements in "list".
3367 * When this function returns, the total number of variables in info->tab
3368 * is equal to "dim" plus the number of elements in "list".
3370 static int add_subs(struct isl_coalesce_info *info,
3371 __isl_keep isl_aff_list *list, int dim)
3373 int extra_var;
3374 int n;
3376 if (!list)
3377 return -1;
3379 n = isl_aff_list_n_aff(list);
3380 extra_var = n - (info->tab->n_var - dim);
3382 if (isl_tab_extend_vars(info->tab, extra_var) < 0)
3383 return -1;
3384 if (isl_tab_extend_cons(info->tab, 2 * extra_var) < 0)
3385 return -1;
3386 if (add_sub_vars(info, list, dim, extra_var) < 0)
3387 return -1;
3389 return add_sub_equalities(info->tab, list, dim);
3392 /* Coalesce basic map "j" into basic map "i" after adding the extra integer
3393 * divisions in "i" but not in "j" to basic map "j", with values
3394 * specified by "list". The total number of elements in "list"
3395 * is equal to the number of integer divisions in "i", while the number
3396 * of NaN elements in the list is equal to the number of integer divisions
3397 * in "j".
3399 * If no coalescing can be performed, then we need to revert basic map "j"
3400 * to its original state. We do the same if basic map "i" gets dropped
3401 * during the coalescing, even though this should not happen in practice
3402 * since we have already checked for "j" being a subset of "i"
3403 * before we reach this stage.
3405 static enum isl_change coalesce_with_subs(int i, int j,
3406 struct isl_coalesce_info *info, __isl_keep isl_aff_list *list)
3408 isl_basic_map *bmap_j;
3409 struct isl_tab_undo *snap;
3410 unsigned dim;
3411 enum isl_change change;
3413 bmap_j = isl_basic_map_copy(info[j].bmap);
3414 snap = isl_tab_snap(info[j].tab);
3416 dim = isl_basic_map_dim(bmap_j, isl_dim_all);
3417 dim -= isl_basic_map_dim(bmap_j, isl_dim_div);
3418 if (add_subs(&info[j], list, dim) < 0)
3419 goto error;
3421 change = coalesce_local_pair(i, j, info);
3422 if (change != isl_change_none && change != isl_change_drop_first) {
3423 isl_basic_map_free(bmap_j);
3424 } else {
3425 isl_basic_map_free(info[j].bmap);
3426 info[j].bmap = bmap_j;
3428 if (isl_tab_rollback(info[j].tab, snap) < 0)
3429 return isl_change_error;
3432 return change;
3433 error:
3434 isl_basic_map_free(bmap_j);
3435 return isl_change_error;
3438 /* Check if we can coalesce basic map "j" into basic map "i" after copying
3439 * those extra integer divisions in "i" that can be simplified away
3440 * using the extra equalities in "j".
3441 * All divs are assumed to be known and not contain any nested divs.
3443 * We first check if there are any extra equalities in "j" that we
3444 * can exploit. Then we check if every integer division in "i"
3445 * either already appears in "j" or can be simplified using the
3446 * extra equalities to a purely affine expression.
3447 * If these tests succeed, then we try to coalesce the two basic maps
3448 * by introducing extra dimensions in "j" corresponding to
3449 * the extra integer divsisions "i" fixed to the corresponding
3450 * purely affine expression.
3452 static enum isl_change check_coalesce_into_eq(int i, int j,
3453 struct isl_coalesce_info *info)
3455 unsigned n_div_i, n_div_j;
3456 isl_basic_map *hull_i, *hull_j;
3457 int equal, empty;
3458 isl_aff_list *list;
3459 enum isl_change change;
3461 n_div_i = isl_basic_map_dim(info[i].bmap, isl_dim_div);
3462 n_div_j = isl_basic_map_dim(info[j].bmap, isl_dim_div);
3463 if (n_div_i <= n_div_j)
3464 return isl_change_none;
3465 if (info[j].bmap->n_eq == 0)
3466 return isl_change_none;
3468 hull_i = isl_basic_map_copy(info[i].bmap);
3469 hull_i = isl_basic_map_plain_affine_hull(hull_i);
3470 hull_j = isl_basic_map_copy(info[j].bmap);
3471 hull_j = isl_basic_map_plain_affine_hull(hull_j);
3473 hull_j = isl_basic_map_intersect(hull_j, isl_basic_map_copy(hull_i));
3474 equal = isl_basic_map_plain_is_equal(hull_i, hull_j);
3475 empty = isl_basic_map_plain_is_empty(hull_j);
3476 isl_basic_map_free(hull_i);
3478 if (equal < 0 || empty < 0)
3479 goto error;
3480 if (equal || empty) {
3481 isl_basic_map_free(hull_j);
3482 return isl_change_none;
3485 list = set_up_substitutions(info[i].bmap, info[j].bmap, hull_j);
3486 if (!list)
3487 return isl_change_error;
3488 if (isl_aff_list_n_aff(list) < n_div_i)
3489 change = isl_change_none;
3490 else
3491 change = coalesce_with_subs(i, j, info, list);
3493 isl_aff_list_free(list);
3495 return change;
3496 error:
3497 isl_basic_map_free(hull_j);
3498 return isl_change_error;
3501 /* Check if we can coalesce basic maps "i" and "j" after copying
3502 * those extra integer divisions in one of the basic maps that can
3503 * be simplified away using the extra equalities in the other basic map.
3504 * We require all divs to be known in both basic maps.
3505 * Furthermore, to simplify the comparison of div expressions,
3506 * we do not allow any nested integer divisions.
3508 static enum isl_change check_coalesce_eq(int i, int j,
3509 struct isl_coalesce_info *info)
3511 isl_bool known, nested;
3512 enum isl_change change;
3514 known = isl_basic_map_divs_known(info[i].bmap);
3515 if (known < 0 || !known)
3516 return known < 0 ? isl_change_error : isl_change_none;
3517 known = isl_basic_map_divs_known(info[j].bmap);
3518 if (known < 0 || !known)
3519 return known < 0 ? isl_change_error : isl_change_none;
3520 nested = has_nested_div(info[i].bmap);
3521 if (nested < 0 || nested)
3522 return nested < 0 ? isl_change_error : isl_change_none;
3523 nested = has_nested_div(info[j].bmap);
3524 if (nested < 0 || nested)
3525 return nested < 0 ? isl_change_error : isl_change_none;
3527 change = check_coalesce_into_eq(i, j, info);
3528 if (change != isl_change_none)
3529 return change;
3530 change = check_coalesce_into_eq(j, i, info);
3531 if (change != isl_change_none)
3532 return invert_change(change);
3534 return isl_change_none;
3537 /* Check if the union of the given pair of basic maps
3538 * can be represented by a single basic map.
3539 * If so, replace the pair by the single basic map and return
3540 * isl_change_drop_first, isl_change_drop_second or isl_change_fuse.
3541 * Otherwise, return isl_change_none.
3543 * We first check if the two basic maps live in the same local space,
3544 * after aligning the divs that differ by only an integer constant.
3545 * If so, we do the complete check. Otherwise, we check if they have
3546 * the same number of integer divisions and can be coalesced, if one is
3547 * an obvious subset of the other or if the extra integer divisions
3548 * of one basic map can be simplified away using the extra equalities
3549 * of the other basic map.
3551 static enum isl_change coalesce_pair(int i, int j,
3552 struct isl_coalesce_info *info)
3554 isl_bool same;
3555 enum isl_change change;
3557 if (harmonize_divs(&info[i], &info[j]) < 0)
3558 return isl_change_error;
3559 same = same_divs(info[i].bmap, info[j].bmap);
3560 if (same < 0)
3561 return isl_change_error;
3562 if (same)
3563 return coalesce_local_pair(i, j, info);
3565 if (info[i].bmap->n_div == info[j].bmap->n_div) {
3566 change = coalesce_local_pair(i, j, info);
3567 if (change != isl_change_none)
3568 return change;
3571 change = coalesce_divs(i, j, info);
3572 if (change != isl_change_none)
3573 return change;
3575 return check_coalesce_eq(i, j, info);
3578 /* Return the maximum of "a" and "b".
3580 static int isl_max(int a, int b)
3582 return a > b ? a : b;
3585 /* Pairwise coalesce the basic maps in the range [start1, end1[ of "info"
3586 * with those in the range [start2, end2[, skipping basic maps
3587 * that have been removed (either before or within this function).
3589 * For each basic map i in the first range, we check if it can be coalesced
3590 * with respect to any previously considered basic map j in the second range.
3591 * If i gets dropped (because it was a subset of some j), then
3592 * we can move on to the next basic map.
3593 * If j gets dropped, we need to continue checking against the other
3594 * previously considered basic maps.
3595 * If the two basic maps got fused, then we recheck the fused basic map
3596 * against the previously considered basic maps, starting at i + 1
3597 * (even if start2 is greater than i + 1).
3599 static int coalesce_range(isl_ctx *ctx, struct isl_coalesce_info *info,
3600 int start1, int end1, int start2, int end2)
3602 int i, j;
3604 for (i = end1 - 1; i >= start1; --i) {
3605 if (info[i].removed)
3606 continue;
3607 for (j = isl_max(i + 1, start2); j < end2; ++j) {
3608 enum isl_change changed;
3610 if (info[j].removed)
3611 continue;
3612 if (info[i].removed)
3613 isl_die(ctx, isl_error_internal,
3614 "basic map unexpectedly removed",
3615 return -1);
3616 changed = coalesce_pair(i, j, info);
3617 switch (changed) {
3618 case isl_change_error:
3619 return -1;
3620 case isl_change_none:
3621 case isl_change_drop_second:
3622 continue;
3623 case isl_change_drop_first:
3624 j = end2;
3625 break;
3626 case isl_change_fuse:
3627 j = i;
3628 break;
3633 return 0;
3636 /* Pairwise coalesce the basic maps described by the "n" elements of "info".
3638 * We consider groups of basic maps that live in the same apparent
3639 * affine hull and we first coalesce within such a group before we
3640 * coalesce the elements in the group with elements of previously
3641 * considered groups. If a fuse happens during the second phase,
3642 * then we also reconsider the elements within the group.
3644 static int coalesce(isl_ctx *ctx, int n, struct isl_coalesce_info *info)
3646 int start, end;
3648 for (end = n; end > 0; end = start) {
3649 start = end - 1;
3650 while (start >= 1 &&
3651 info[start - 1].hull_hash == info[start].hull_hash)
3652 start--;
3653 if (coalesce_range(ctx, info, start, end, start, end) < 0)
3654 return -1;
3655 if (coalesce_range(ctx, info, start, end, end, n) < 0)
3656 return -1;
3659 return 0;
3662 /* Update the basic maps in "map" based on the information in "info".
3663 * In particular, remove the basic maps that have been marked removed and
3664 * update the others based on the information in the corresponding tableau.
3665 * Since we detected implicit equalities without calling
3666 * isl_basic_map_gauss, we need to do it now.
3667 * Also call isl_basic_map_simplify if we may have lost the definition
3668 * of one or more integer divisions.
3670 static __isl_give isl_map *update_basic_maps(__isl_take isl_map *map,
3671 int n, struct isl_coalesce_info *info)
3673 int i;
3675 if (!map)
3676 return NULL;
3678 for (i = n - 1; i >= 0; --i) {
3679 if (info[i].removed) {
3680 isl_basic_map_free(map->p[i]);
3681 if (i != map->n - 1)
3682 map->p[i] = map->p[map->n - 1];
3683 map->n--;
3684 continue;
3687 info[i].bmap = isl_basic_map_update_from_tab(info[i].bmap,
3688 info[i].tab);
3689 info[i].bmap = isl_basic_map_gauss(info[i].bmap, NULL);
3690 if (info[i].simplify)
3691 info[i].bmap = isl_basic_map_simplify(info[i].bmap);
3692 info[i].bmap = isl_basic_map_finalize(info[i].bmap);
3693 if (!info[i].bmap)
3694 return isl_map_free(map);
3695 ISL_F_SET(info[i].bmap, ISL_BASIC_MAP_NO_IMPLICIT);
3696 ISL_F_SET(info[i].bmap, ISL_BASIC_MAP_NO_REDUNDANT);
3697 isl_basic_map_free(map->p[i]);
3698 map->p[i] = info[i].bmap;
3699 info[i].bmap = NULL;
3702 return map;
3705 /* For each pair of basic maps in the map, check if the union of the two
3706 * can be represented by a single basic map.
3707 * If so, replace the pair by the single basic map and start over.
3709 * We factor out any (hidden) common factor from the constraint
3710 * coefficients to improve the detection of adjacent constraints.
3712 * Since we are constructing the tableaus of the basic maps anyway,
3713 * we exploit them to detect implicit equalities and redundant constraints.
3714 * This also helps the coalescing as it can ignore the redundant constraints.
3715 * In order to avoid confusion, we make all implicit equalities explicit
3716 * in the basic maps. We don't call isl_basic_map_gauss, though,
3717 * as that may affect the number of constraints.
3718 * This means that we have to call isl_basic_map_gauss at the end
3719 * of the computation (in update_basic_maps) to ensure that
3720 * the basic maps are not left in an unexpected state.
3721 * For each basic map, we also compute the hash of the apparent affine hull
3722 * for use in coalesce.
3724 struct isl_map *isl_map_coalesce(struct isl_map *map)
3726 int i;
3727 unsigned n;
3728 isl_ctx *ctx;
3729 struct isl_coalesce_info *info = NULL;
3731 map = isl_map_remove_empty_parts(map);
3732 if (!map)
3733 return NULL;
3735 if (map->n <= 1)
3736 return map;
3738 ctx = isl_map_get_ctx(map);
3739 map = isl_map_sort_divs(map);
3740 map = isl_map_cow(map);
3742 if (!map)
3743 return NULL;
3745 n = map->n;
3747 info = isl_calloc_array(map->ctx, struct isl_coalesce_info, n);
3748 if (!info)
3749 goto error;
3751 for (i = 0; i < map->n; ++i) {
3752 map->p[i] = isl_basic_map_reduce_coefficients(map->p[i]);
3753 if (!map->p[i])
3754 goto error;
3755 info[i].bmap = isl_basic_map_copy(map->p[i]);
3756 info[i].tab = isl_tab_from_basic_map(info[i].bmap, 0);
3757 if (!info[i].tab)
3758 goto error;
3759 if (!ISL_F_ISSET(info[i].bmap, ISL_BASIC_MAP_NO_IMPLICIT))
3760 if (isl_tab_detect_implicit_equalities(info[i].tab) < 0)
3761 goto error;
3762 info[i].bmap = isl_tab_make_equalities_explicit(info[i].tab,
3763 info[i].bmap);
3764 if (!info[i].bmap)
3765 goto error;
3766 if (!ISL_F_ISSET(info[i].bmap, ISL_BASIC_MAP_NO_REDUNDANT))
3767 if (isl_tab_detect_redundant(info[i].tab) < 0)
3768 goto error;
3769 if (coalesce_info_set_hull_hash(&info[i]) < 0)
3770 goto error;
3772 for (i = map->n - 1; i >= 0; --i)
3773 if (info[i].tab->empty)
3774 drop(&info[i]);
3776 if (coalesce(ctx, n, info) < 0)
3777 goto error;
3779 map = update_basic_maps(map, n, info);
3781 clear_coalesce_info(n, info);
3783 return map;
3784 error:
3785 clear_coalesce_info(n, info);
3786 isl_map_free(map);
3787 return NULL;
3790 /* For each pair of basic sets in the set, check if the union of the two
3791 * can be represented by a single basic set.
3792 * If so, replace the pair by the single basic set and start over.
3794 struct isl_set *isl_set_coalesce(struct isl_set *set)
3796 return set_from_map(isl_map_coalesce(set_to_map(set)));