File: | tools/polly/lib/External/isl/isl_coalesce.c |
Warning: | line 2284, column 2 Value stored to 'total' is never read |
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 |
7 | * |
8 | * Use of this software is governed by the MIT license |
9 | * |
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 |
19 | */ |
20 | |
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> |
33 | |
34 | #include <set_to_map.c> |
35 | #include <set_from_map.c> |
36 | |
37 | #define STATUS_ERROR-1 -1 |
38 | #define STATUS_REDUNDANT1 1 |
39 | #define STATUS_VALID2 2 |
40 | #define STATUS_SEPARATE3 3 |
41 | #define STATUS_CUT4 4 |
42 | #define STATUS_ADJ_EQ5 5 |
43 | #define STATUS_ADJ_INEQ6 6 |
44 | |
45 | static int status_in(isl_int *ineq, struct isl_tab *tab) |
46 | { |
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-1; |
51 | case isl_ineq_redundant: return STATUS_VALID2; |
52 | case isl_ineq_separate: return STATUS_SEPARATE3; |
53 | case isl_ineq_cut: return STATUS_CUT4; |
54 | case isl_ineq_adj_eq: return STATUS_ADJ_EQ5; |
55 | case isl_ineq_adj_ineq: return STATUS_ADJ_INEQ6; |
56 | } |
57 | } |
58 | |
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. |
64 | */ |
65 | static int *eq_status_in(__isl_keep isl_basic_map *bmap_i, |
66 | struct isl_tab *tab_j) |
67 | { |
68 | int k, l; |
69 | int *eq = isl_calloc_array(bmap_i->ctx, int, 2 * bmap_i->n_eq)((int *)isl_calloc_or_die(bmap_i->ctx, 2 * bmap_i->n_eq , sizeof(int))); |
70 | unsigned dim; |
71 | |
72 | if (!eq) |
73 | return NULL((void*)0); |
74 | |
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-1) |
81 | goto error; |
82 | } |
83 | if (eq[2 * k] == STATUS_SEPARATE3 || |
84 | eq[2 * k + 1] == STATUS_SEPARATE3) |
85 | break; |
86 | } |
87 | |
88 | return eq; |
89 | error: |
90 | free(eq); |
91 | return NULL((void*)0); |
92 | } |
93 | |
94 | /* Compute the position of the inequalities of basic map "bmap_i" |
95 | * (also represented by "tab_i", if not NULL) with respect to the basic map |
96 | * represented by "tab_j". |
97 | */ |
98 | static int *ineq_status_in(__isl_keep isl_basic_map *bmap_i, |
99 | struct isl_tab *tab_i, struct isl_tab *tab_j) |
100 | { |
101 | int k; |
102 | unsigned n_eq = bmap_i->n_eq; |
103 | int *ineq = isl_calloc_array(bmap_i->ctx, int, bmap_i->n_ineq)((int *)isl_calloc_or_die(bmap_i->ctx, bmap_i->n_ineq, sizeof (int))); |
104 | |
105 | if (!ineq) |
106 | return NULL((void*)0); |
107 | |
108 | for (k = 0; k < bmap_i->n_ineq; ++k) { |
109 | if (tab_i && isl_tab_is_redundant(tab_i, n_eq + k)) { |
110 | ineq[k] = STATUS_REDUNDANT1; |
111 | continue; |
112 | } |
113 | ineq[k] = status_in(bmap_i->ineq[k], tab_j); |
114 | if (ineq[k] == STATUS_ERROR-1) |
115 | goto error; |
116 | if (ineq[k] == STATUS_SEPARATE3) |
117 | break; |
118 | } |
119 | |
120 | return ineq; |
121 | error: |
122 | free(ineq); |
123 | return NULL((void*)0); |
124 | } |
125 | |
126 | static int any(int *con, unsigned len, int status) |
127 | { |
128 | int i; |
129 | |
130 | for (i = 0; i < len ; ++i) |
131 | if (con[i] == status) |
132 | return 1; |
133 | return 0; |
134 | } |
135 | |
136 | /* Return the first position of "status" in the list "con" of length "len". |
137 | * Return -1 if there is no such entry. |
138 | */ |
139 | static int find(int *con, unsigned len, int status) |
140 | { |
141 | int i; |
142 | |
143 | for (i = 0; i < len ; ++i) |
144 | if (con[i] == status) |
145 | return i; |
146 | return -1; |
147 | } |
148 | |
149 | static int count(int *con, unsigned len, int status) |
150 | { |
151 | int i; |
152 | int c = 0; |
153 | |
154 | for (i = 0; i < len ; ++i) |
155 | if (con[i] == status) |
156 | c++; |
157 | return c; |
158 | } |
159 | |
160 | static int all(int *con, unsigned len, int status) |
161 | { |
162 | int i; |
163 | |
164 | for (i = 0; i < len ; ++i) { |
165 | if (con[i] == STATUS_REDUNDANT1) |
166 | continue; |
167 | if (con[i] != status) |
168 | return 0; |
169 | } |
170 | return 1; |
171 | } |
172 | |
173 | /* Internal information associated to a basic map in a map |
174 | * that is to be coalesced by isl_map_coalesce. |
175 | * |
176 | * "bmap" is the basic map itself (or NULL if "removed" is set) |
177 | * "tab" is the corresponding tableau (or NULL if "removed" is set) |
178 | * "hull_hash" identifies the affine space in which "bmap" lives. |
179 | * "removed" is set if this basic map has been removed from the map |
180 | * "simplify" is set if this basic map may have some unknown integer |
181 | * divisions that were not present in the input basic maps. The basic |
182 | * map should then be simplified such that we may be able to find |
183 | * a definition among the constraints. |
184 | * |
185 | * "eq" and "ineq" are only set if we are currently trying to coalesce |
186 | * this basic map with another basic map, in which case they represent |
187 | * the position of the inequalities of this basic map with respect to |
188 | * the other basic map. The number of elements in the "eq" array |
189 | * is twice the number of equalities in the "bmap", corresponding |
190 | * to the two inequalities that make up each equality. |
191 | */ |
192 | struct isl_coalesce_info { |
193 | isl_basic_map *bmap; |
194 | struct isl_tab *tab; |
195 | uint32_t hull_hash; |
196 | int removed; |
197 | int simplify; |
198 | int *eq; |
199 | int *ineq; |
200 | }; |
201 | |
202 | /* Are all non-redundant constraints of the basic map represented by "info" |
203 | * either valid or cut constraints with respect to the other basic map? |
204 | */ |
205 | static int all_valid_or_cut(struct isl_coalesce_info *info) |
206 | { |
207 | int i; |
208 | |
209 | for (i = 0; i < 2 * info->bmap->n_eq; ++i) { |
210 | if (info->eq[i] == STATUS_REDUNDANT1) |
211 | continue; |
212 | if (info->eq[i] == STATUS_VALID2) |
213 | continue; |
214 | if (info->eq[i] == STATUS_CUT4) |
215 | continue; |
216 | return 0; |
217 | } |
218 | |
219 | for (i = 0; i < info->bmap->n_ineq; ++i) { |
220 | if (info->ineq[i] == STATUS_REDUNDANT1) |
221 | continue; |
222 | if (info->ineq[i] == STATUS_VALID2) |
223 | continue; |
224 | if (info->ineq[i] == STATUS_CUT4) |
225 | continue; |
226 | return 0; |
227 | } |
228 | |
229 | return 1; |
230 | } |
231 | |
232 | /* Compute the hash of the (apparent) affine hull of info->bmap (with |
233 | * the existentially quantified variables removed) and store it |
234 | * in info->hash. |
235 | */ |
236 | static int coalesce_info_set_hull_hash(struct isl_coalesce_info *info) |
237 | { |
238 | isl_basic_map *hull; |
239 | unsigned n_div; |
240 | |
241 | hull = isl_basic_map_copy(info->bmap); |
242 | hull = isl_basic_map_plain_affine_hull(hull); |
243 | n_div = isl_basic_map_dim(hull, isl_dim_div); |
244 | hull = isl_basic_map_drop_constraints_involving_dims(hull, |
245 | isl_dim_div, 0, n_div); |
246 | info->hull_hash = isl_basic_map_get_hash(hull); |
247 | isl_basic_map_free(hull); |
248 | |
249 | return hull ? 0 : -1; |
250 | } |
251 | |
252 | /* Free all the allocated memory in an array |
253 | * of "n" isl_coalesce_info elements. |
254 | */ |
255 | static void clear_coalesce_info(int n, struct isl_coalesce_info *info) |
256 | { |
257 | int i; |
258 | |
259 | if (!info) |
260 | return; |
261 | |
262 | for (i = 0; i < n; ++i) { |
263 | isl_basic_map_free(info[i].bmap); |
264 | isl_tab_free(info[i].tab); |
265 | } |
266 | |
267 | free(info); |
268 | } |
269 | |
270 | /* Drop the basic map represented by "info". |
271 | * That is, clear the memory associated to the entry and |
272 | * mark it as having been removed. |
273 | */ |
274 | static void drop(struct isl_coalesce_info *info) |
275 | { |
276 | info->bmap = isl_basic_map_free(info->bmap); |
277 | isl_tab_free(info->tab); |
278 | info->tab = NULL((void*)0); |
279 | info->removed = 1; |
280 | } |
281 | |
282 | /* Exchange the information in "info1" with that in "info2". |
283 | */ |
284 | static void exchange(struct isl_coalesce_info *info1, |
285 | struct isl_coalesce_info *info2) |
286 | { |
287 | struct isl_coalesce_info info; |
288 | |
289 | info = *info1; |
290 | *info1 = *info2; |
291 | *info2 = info; |
292 | } |
293 | |
294 | /* This type represents the kind of change that has been performed |
295 | * while trying to coalesce two basic maps. |
296 | * |
297 | * isl_change_none: nothing was changed |
298 | * isl_change_drop_first: the first basic map was removed |
299 | * isl_change_drop_second: the second basic map was removed |
300 | * isl_change_fuse: the two basic maps were replaced by a new basic map. |
301 | */ |
302 | enum isl_change { |
303 | isl_change_error = -1, |
304 | isl_change_none = 0, |
305 | isl_change_drop_first, |
306 | isl_change_drop_second, |
307 | isl_change_fuse, |
308 | }; |
309 | |
310 | /* Update "change" based on an interchange of the first and the second |
311 | * basic map. That is, interchange isl_change_drop_first and |
312 | * isl_change_drop_second. |
313 | */ |
314 | static enum isl_change invert_change(enum isl_change change) |
315 | { |
316 | switch (change) { |
317 | case isl_change_error: |
318 | return isl_change_error; |
319 | case isl_change_none: |
320 | return isl_change_none; |
321 | case isl_change_drop_first: |
322 | return isl_change_drop_second; |
323 | case isl_change_drop_second: |
324 | return isl_change_drop_first; |
325 | case isl_change_fuse: |
326 | return isl_change_fuse; |
327 | } |
328 | |
329 | return isl_change_error; |
330 | } |
331 | |
332 | /* Add the valid constraints of the basic map represented by "info" |
333 | * to "bmap". "len" is the size of the constraints. |
334 | * If only one of the pair of inequalities that make up an equality |
335 | * is valid, then add that inequality. |
336 | */ |
337 | static __isl_give isl_basic_map *add_valid_constraints( |
338 | __isl_take isl_basic_map *bmap, struct isl_coalesce_info *info, |
339 | unsigned len) |
340 | { |
341 | int k, l; |
342 | |
343 | if (!bmap) |
344 | return NULL((void*)0); |
345 | |
346 | for (k = 0; k < info->bmap->n_eq; ++k) { |
347 | if (info->eq[2 * k] == STATUS_VALID2 && |
348 | info->eq[2 * k + 1] == STATUS_VALID2) { |
349 | l = isl_basic_map_alloc_equality(bmap); |
350 | if (l < 0) |
351 | return isl_basic_map_free(bmap); |
352 | isl_seq_cpy(bmap->eq[l], info->bmap->eq[k], len); |
353 | } else if (info->eq[2 * k] == STATUS_VALID2) { |
354 | l = isl_basic_map_alloc_inequality(bmap); |
355 | if (l < 0) |
356 | return isl_basic_map_free(bmap); |
357 | isl_seq_neg(bmap->ineq[l], info->bmap->eq[k], len); |
358 | } else if (info->eq[2 * k + 1] == STATUS_VALID2) { |
359 | l = isl_basic_map_alloc_inequality(bmap); |
360 | if (l < 0) |
361 | return isl_basic_map_free(bmap); |
362 | isl_seq_cpy(bmap->ineq[l], info->bmap->eq[k], len); |
363 | } |
364 | } |
365 | |
366 | for (k = 0; k < info->bmap->n_ineq; ++k) { |
367 | if (info->ineq[k] != STATUS_VALID2) |
368 | continue; |
369 | l = isl_basic_map_alloc_inequality(bmap); |
370 | if (l < 0) |
371 | return isl_basic_map_free(bmap); |
372 | isl_seq_cpy(bmap->ineq[l], info->bmap->ineq[k], len); |
373 | } |
374 | |
375 | return bmap; |
376 | } |
377 | |
378 | /* Is "bmap" defined by a number of (non-redundant) constraints that |
379 | * is greater than the number of constraints of basic maps i and j combined? |
380 | * Equalities are counted as two inequalities. |
381 | */ |
382 | static int number_of_constraints_increases(int i, int j, |
383 | struct isl_coalesce_info *info, |
384 | __isl_keep isl_basic_map *bmap, struct isl_tab *tab) |
385 | { |
386 | int k, n_old, n_new; |
387 | |
388 | n_old = 2 * info[i].bmap->n_eq + info[i].bmap->n_ineq; |
389 | n_old += 2 * info[j].bmap->n_eq + info[j].bmap->n_ineq; |
390 | |
391 | n_new = 2 * bmap->n_eq; |
392 | for (k = 0; k < bmap->n_ineq; ++k) |
393 | if (!isl_tab_is_redundant(tab, bmap->n_eq + k)) |
394 | ++n_new; |
395 | |
396 | return n_new > n_old; |
397 | } |
398 | |
399 | /* Replace the pair of basic maps i and j by the basic map bounded |
400 | * by the valid constraints in both basic maps and the constraints |
401 | * in extra (if not NULL). |
402 | * Place the fused basic map in the position that is the smallest of i and j. |
403 | * |
404 | * If "detect_equalities" is set, then look for equalities encoded |
405 | * as pairs of inequalities. |
406 | * If "check_number" is set, then the original basic maps are only |
407 | * replaced if the total number of constraints does not increase. |
408 | * While the number of integer divisions in the two basic maps |
409 | * is assumed to be the same, the actual definitions may be different. |
410 | * We only copy the definition from one of the basic map if it is |
411 | * the same as that of the other basic map. Otherwise, we mark |
412 | * the integer division as unknown and simplify the basic map |
413 | * in an attempt to recover the integer division definition. |
414 | */ |
415 | static enum isl_change fuse(int i, int j, struct isl_coalesce_info *info, |
416 | __isl_keep isl_mat *extra, int detect_equalities, int check_number) |
417 | { |
418 | int k, l; |
419 | struct isl_basic_map *fused = NULL((void*)0); |
420 | struct isl_tab *fused_tab = NULL((void*)0); |
421 | unsigned total = isl_basic_map_total_dim(info[i].bmap); |
422 | unsigned extra_rows = extra ? extra->n_row : 0; |
423 | unsigned n_eq, n_ineq; |
424 | int simplify = 0; |
425 | |
426 | if (j < i) |
427 | return fuse(j, i, info, extra, detect_equalities, check_number); |
428 | |
429 | n_eq = info[i].bmap->n_eq + info[j].bmap->n_eq; |
430 | n_ineq = info[i].bmap->n_ineq + info[j].bmap->n_ineq; |
431 | fused = isl_basic_map_alloc_space(isl_space_copy(info[i].bmap->dim), |
432 | info[i].bmap->n_div, n_eq, n_eq + n_ineq + extra_rows); |
433 | fused = add_valid_constraints(fused, &info[i], 1 + total); |
434 | fused = add_valid_constraints(fused, &info[j], 1 + total); |
435 | if (!fused) |
436 | goto error; |
437 | if (ISL_F_ISSET(info[i].bmap, ISL_BASIC_MAP_RATIONAL)(!!(((info[i].bmap)->flags) & ((1 << 4)))) && |
438 | ISL_F_ISSET(info[j].bmap, ISL_BASIC_MAP_RATIONAL)(!!(((info[j].bmap)->flags) & ((1 << 4))))) |
439 | ISL_F_SET(fused, ISL_BASIC_MAP_RATIONAL)(((fused)->flags) |= ((1 << 4))); |
440 | |
441 | for (k = 0; k < info[i].bmap->n_div; ++k) { |
442 | int l = isl_basic_map_alloc_div(fused); |
443 | if (l < 0) |
444 | goto error; |
445 | if (isl_seq_eq(info[i].bmap->div[k], info[j].bmap->div[k], |
446 | 1 + 1 + total)) { |
447 | isl_seq_cpy(fused->div[l], info[i].bmap->div[k], |
448 | 1 + 1 + total); |
449 | } else { |
450 | isl_int_set_si(fused->div[l][0], 0)isl_sioimath_set_si((fused->div[l][0]), 0); |
451 | simplify = 1; |
452 | } |
453 | } |
454 | |
455 | for (k = 0; k < extra_rows; ++k) { |
456 | l = isl_basic_map_alloc_inequality(fused); |
457 | if (l < 0) |
458 | goto error; |
459 | isl_seq_cpy(fused->ineq[l], extra->row[k], 1 + total); |
460 | } |
461 | |
462 | if (detect_equalities) |
463 | fused = isl_basic_map_detect_inequality_pairs(fused, NULL((void*)0)); |
464 | fused = isl_basic_map_gauss(fused, NULL((void*)0)); |
465 | if (simplify || info[j].simplify) { |
466 | fused = isl_basic_map_simplify(fused); |
467 | info[i].simplify = 0; |
468 | } |
469 | fused = isl_basic_map_finalize(fused); |
470 | |
471 | fused_tab = isl_tab_from_basic_map(fused, 0); |
472 | if (isl_tab_detect_redundant(fused_tab) < 0) |
473 | goto error; |
474 | |
475 | if (check_number && |
476 | number_of_constraints_increases(i, j, info, fused, fused_tab)) { |
477 | isl_tab_free(fused_tab); |
478 | isl_basic_map_free(fused); |
479 | return isl_change_none; |
480 | } |
481 | |
482 | isl_basic_map_free(info[i].bmap); |
483 | info[i].bmap = fused; |
484 | isl_tab_free(info[i].tab); |
485 | info[i].tab = fused_tab; |
486 | drop(&info[j]); |
487 | |
488 | return isl_change_fuse; |
489 | error: |
490 | isl_tab_free(fused_tab); |
491 | isl_basic_map_free(fused); |
492 | return isl_change_error; |
493 | } |
494 | |
495 | /* Given a pair of basic maps i and j such that all constraints are either |
496 | * "valid" or "cut", check if the facets corresponding to the "cut" |
497 | * constraints of i lie entirely within basic map j. |
498 | * If so, replace the pair by the basic map consisting of the valid |
499 | * constraints in both basic maps. |
500 | * Checking whether the facet lies entirely within basic map j |
501 | * is performed by checking whether the constraints of basic map j |
502 | * are valid for the facet. These tests are performed on a rational |
503 | * tableau to avoid the theoretical possibility that a constraint |
504 | * that was considered to be a cut constraint for the entire basic map i |
505 | * happens to be considered to be a valid constraint for the facet, |
506 | * even though it cuts off the same rational points. |
507 | * |
508 | * To see that we are not introducing any extra points, call the |
509 | * two basic maps A and B and the resulting map U and let x |
510 | * be an element of U \setminus ( A \cup B ). |
511 | * A line connecting x with an element of A \cup B meets a facet F |
512 | * of either A or B. Assume it is a facet of B and let c_1 be |
513 | * the corresponding facet constraint. We have c_1(x) < 0 and |
514 | * so c_1 is a cut constraint. This implies that there is some |
515 | * (possibly rational) point x' satisfying the constraints of A |
516 | * and the opposite of c_1 as otherwise c_1 would have been marked |
517 | * valid for A. The line connecting x and x' meets a facet of A |
518 | * in a (possibly rational) point that also violates c_1, but this |
519 | * is impossible since all cut constraints of B are valid for all |
520 | * cut facets of A. |
521 | * In case F is a facet of A rather than B, then we can apply the |
522 | * above reasoning to find a facet of B separating x from A \cup B first. |
523 | */ |
524 | static enum isl_change check_facets(int i, int j, |
525 | struct isl_coalesce_info *info) |
526 | { |
527 | int k, l; |
528 | struct isl_tab_undo *snap, *snap2; |
529 | unsigned n_eq = info[i].bmap->n_eq; |
530 | |
531 | snap = isl_tab_snap(info[i].tab); |
532 | if (isl_tab_mark_rational(info[i].tab) < 0) |
533 | return isl_change_error; |
534 | snap2 = isl_tab_snap(info[i].tab); |
535 | |
536 | for (k = 0; k < info[i].bmap->n_ineq; ++k) { |
537 | if (info[i].ineq[k] != STATUS_CUT4) |
538 | continue; |
539 | if (isl_tab_select_facet(info[i].tab, n_eq + k) < 0) |
540 | return isl_change_error; |
541 | for (l = 0; l < info[j].bmap->n_ineq; ++l) { |
542 | int stat; |
543 | if (info[j].ineq[l] != STATUS_CUT4) |
544 | continue; |
545 | stat = status_in(info[j].bmap->ineq[l], info[i].tab); |
546 | if (stat < 0) |
547 | return isl_change_error; |
548 | if (stat != STATUS_VALID2) |
549 | break; |
550 | } |
551 | if (isl_tab_rollback(info[i].tab, snap2) < 0) |
552 | return isl_change_error; |
553 | if (l < info[j].bmap->n_ineq) |
554 | break; |
555 | } |
556 | |
557 | if (k < info[i].bmap->n_ineq) { |
558 | if (isl_tab_rollback(info[i].tab, snap) < 0) |
559 | return isl_change_error; |
560 | return isl_change_none; |
561 | } |
562 | return fuse(i, j, info, NULL((void*)0), 0, 0); |
563 | } |
564 | |
565 | /* Check if info->bmap contains the basic map represented |
566 | * by the tableau "tab". |
567 | * For each equality, we check both the constraint itself |
568 | * (as an inequality) and its negation. Make sure the |
569 | * equality is returned to its original state before returning. |
570 | */ |
571 | static int contains(struct isl_coalesce_info *info, struct isl_tab *tab) |
572 | { |
573 | int k; |
574 | unsigned dim; |
575 | isl_basic_map *bmap = info->bmap; |
576 | |
577 | dim = isl_basic_map_total_dim(bmap); |
578 | for (k = 0; k < bmap->n_eq; ++k) { |
579 | int stat; |
580 | isl_seq_neg(bmap->eq[k], bmap->eq[k], 1 + dim); |
581 | stat = status_in(bmap->eq[k], tab); |
582 | isl_seq_neg(bmap->eq[k], bmap->eq[k], 1 + dim); |
583 | if (stat < 0) |
584 | return -1; |
585 | if (stat != STATUS_VALID2) |
586 | return 0; |
587 | stat = status_in(bmap->eq[k], tab); |
588 | if (stat < 0) |
589 | return -1; |
590 | if (stat != STATUS_VALID2) |
591 | return 0; |
592 | } |
593 | |
594 | for (k = 0; k < bmap->n_ineq; ++k) { |
595 | int stat; |
596 | if (info->ineq[k] == STATUS_REDUNDANT1) |
597 | continue; |
598 | stat = status_in(bmap->ineq[k], tab); |
599 | if (stat < 0) |
600 | return -1; |
601 | if (stat != STATUS_VALID2) |
602 | return 0; |
603 | } |
604 | return 1; |
605 | } |
606 | |
607 | /* Basic map "i" has an inequality (say "k") that is adjacent |
608 | * to some inequality of basic map "j". All the other inequalities |
609 | * are valid for "j". |
610 | * Check if basic map "j" forms an extension of basic map "i". |
611 | * |
612 | * Note that this function is only called if some of the equalities or |
613 | * inequalities of basic map "j" do cut basic map "i". The function is |
614 | * correct even if there are no such cut constraints, but in that case |
615 | * the additional checks performed by this function are overkill. |
616 | * |
617 | * In particular, we replace constraint k, say f >= 0, by constraint |
618 | * f <= -1, add the inequalities of "j" that are valid for "i" |
619 | * and check if the result is a subset of basic map "j". |
620 | * To improve the chances of the subset relation being detected, |
621 | * any variable that only attains a single integer value |
622 | * in the tableau of "i" is first fixed to that value. |
623 | * If the result is a subset, then we know that this result is exactly equal |
624 | * to basic map "j" since all its constraints are valid for basic map "j". |
625 | * By combining the valid constraints of "i" (all equalities and all |
626 | * inequalities except "k") and the valid constraints of "j" we therefore |
627 | * obtain a basic map that is equal to their union. |
628 | * In this case, there is no need to perform a rollback of the tableau |
629 | * since it is going to be destroyed in fuse(). |
630 | * |
631 | * |
632 | * |\__ |\__ |
633 | * | \__ | \__ |
634 | * | \_ => | \__ |
635 | * |_______| _ |_________\ |
636 | * |
637 | * |
638 | * |\ |\ |
639 | * | \ | \ |
640 | * | \ | \ |
641 | * | | | \ |
642 | * | ||\ => | \ |
643 | * | || \ | \ |
644 | * | || | | | |
645 | * |__||_/ |_____/ |
646 | */ |
647 | static enum isl_change is_adj_ineq_extension(int i, int j, |
648 | struct isl_coalesce_info *info) |
649 | { |
650 | int k; |
651 | struct isl_tab_undo *snap; |
652 | unsigned n_eq = info[i].bmap->n_eq; |
653 | unsigned total = isl_basic_map_total_dim(info[i].bmap); |
654 | int r; |
655 | int super; |
656 | |
657 | if (isl_tab_extend_cons(info[i].tab, 1 + info[j].bmap->n_ineq) < 0) |
658 | return isl_change_error; |
659 | |
660 | k = find(info[i].ineq, info[i].bmap->n_ineq, STATUS_ADJ_INEQ6); |
661 | if (k < 0) |
662 | isl_die(isl_basic_map_get_ctx(info[i].bmap), isl_error_internal,do { isl_handle_error(isl_basic_map_get_ctx(info[i].bmap), isl_error_internal , "info[i].ineq should have exactly one STATUS_ADJ_INEQ", "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn295388/tools/polly/lib/External/isl/isl_coalesce.c" , 664); return isl_change_error; } while (0) |
663 | "info[i].ineq should have exactly one STATUS_ADJ_INEQ",do { isl_handle_error(isl_basic_map_get_ctx(info[i].bmap), isl_error_internal , "info[i].ineq should have exactly one STATUS_ADJ_INEQ", "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn295388/tools/polly/lib/External/isl/isl_coalesce.c" , 664); return isl_change_error; } while (0) |
664 | return isl_change_error)do { isl_handle_error(isl_basic_map_get_ctx(info[i].bmap), isl_error_internal , "info[i].ineq should have exactly one STATUS_ADJ_INEQ", "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn295388/tools/polly/lib/External/isl/isl_coalesce.c" , 664); return isl_change_error; } while (0); |
665 | |
666 | snap = isl_tab_snap(info[i].tab); |
667 | |
668 | if (isl_tab_unrestrict(info[i].tab, n_eq + k) < 0) |
669 | return isl_change_error; |
670 | |
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)isl_sioimath_sub_ui((info[i].bmap->ineq[k][0]), *(info[i]. bmap->ineq[k][0]), 1); |
673 | r = isl_tab_add_ineq(info[i].tab, info[i].bmap->ineq[k]); |
674 | isl_seq_neg(info[i].bmap->ineq[k], info[i].bmap->ineq[k], 1 + total); |
675 | isl_int_sub_ui(info[i].bmap->ineq[k][0], info[i].bmap->ineq[k][0], 1)isl_sioimath_sub_ui((info[i].bmap->ineq[k][0]), *(info[i]. bmap->ineq[k][0]), 1); |
676 | if (r < 0) |
677 | return isl_change_error; |
678 | |
679 | for (k = 0; k < info[j].bmap->n_ineq; ++k) { |
680 | if (info[j].ineq[k] != STATUS_VALID2) |
681 | continue; |
682 | if (isl_tab_add_ineq(info[i].tab, info[j].bmap->ineq[k]) < 0) |
683 | return isl_change_error; |
684 | } |
685 | if (isl_tab_detect_constants(info[i].tab) < 0) |
686 | return isl_change_error; |
687 | |
688 | super = contains(&info[j], info[i].tab); |
689 | if (super < 0) |
690 | return isl_change_error; |
691 | if (super) |
692 | return fuse(i, j, info, NULL((void*)0), 0, 0); |
693 | |
694 | if (isl_tab_rollback(info[i].tab, snap) < 0) |
695 | return isl_change_error; |
696 | |
697 | return isl_change_none; |
698 | } |
699 | |
700 | |
701 | /* Both basic maps have at least one inequality with and adjacent |
702 | * (but opposite) inequality in the other basic map. |
703 | * Check that there are no cut constraints and that there is only |
704 | * a single pair of adjacent inequalities. |
705 | * If so, we can replace the pair by a single basic map described |
706 | * by all but the pair of adjacent inequalities. |
707 | * Any additional points introduced lie strictly between the two |
708 | * adjacent hyperplanes and can therefore be integral. |
709 | * |
710 | * ____ _____ |
711 | * / ||\ / \ |
712 | * / || \ / \ |
713 | * \ || \ => \ \ |
714 | * \ || / \ / |
715 | * \___||_/ \_____/ |
716 | * |
717 | * The test for a single pair of adjancent inequalities is important |
718 | * for avoiding the combination of two basic maps like the following |
719 | * |
720 | * /| |
721 | * / | |
722 | * /__| |
723 | * _____ |
724 | * | | |
725 | * | | |
726 | * |___| |
727 | * |
728 | * If there are some cut constraints on one side, then we may |
729 | * still be able to fuse the two basic maps, but we need to perform |
730 | * some additional checks in is_adj_ineq_extension. |
731 | */ |
732 | static enum isl_change check_adj_ineq(int i, int j, |
733 | struct isl_coalesce_info *info) |
734 | { |
735 | int count_i, count_j; |
736 | int cut_i, cut_j; |
737 | |
738 | count_i = count(info[i].ineq, info[i].bmap->n_ineq, STATUS_ADJ_INEQ6); |
739 | count_j = count(info[j].ineq, info[j].bmap->n_ineq, STATUS_ADJ_INEQ6); |
740 | |
741 | if (count_i != 1 && count_j != 1) |
742 | return isl_change_none; |
743 | |
744 | cut_i = any(info[i].eq, 2 * info[i].bmap->n_eq, STATUS_CUT4) || |
745 | any(info[i].ineq, info[i].bmap->n_ineq, STATUS_CUT4); |
746 | cut_j = any(info[j].eq, 2 * info[j].bmap->n_eq, STATUS_CUT4) || |
747 | any(info[j].ineq, info[j].bmap->n_ineq, STATUS_CUT4); |
748 | |
749 | if (!cut_i && !cut_j && count_i == 1 && count_j == 1) |
750 | return fuse(i, j, info, NULL((void*)0), 0, 0); |
751 | |
752 | if (count_i == 1 && !cut_i) |
753 | return is_adj_ineq_extension(i, j, info); |
754 | |
755 | if (count_j == 1 && !cut_j) |
756 | return is_adj_ineq_extension(j, i, info); |
757 | |
758 | return isl_change_none; |
759 | } |
760 | |
761 | /* Given an affine transformation matrix "T", does row "row" represent |
762 | * anything other than a unit vector (possibly shifted by a constant) |
763 | * that is not involved in any of the other rows? |
764 | * |
765 | * That is, if a constraint involves the variable corresponding to |
766 | * the row, then could its preimage by "T" have any coefficients |
767 | * that are different from those in the original constraint? |
768 | */ |
769 | static int not_unique_unit_row(__isl_keep isl_mat *T, int row) |
770 | { |
771 | int i, j; |
772 | int len = T->n_col - 1; |
773 | |
774 | i = isl_seq_first_non_zero(T->row[row] + 1, len); |
775 | if (i < 0) |
776 | return 1; |
777 | if (!isl_int_is_one(T->row[row][1 + i])(isl_sioimath_cmp_si(*(T->row[row][1 + i]), 1) == 0) && |
778 | !isl_int_is_negone(T->row[row][1 + i])(isl_sioimath_cmp_si(*(T->row[row][1 + i]), -1) == 0)) |
779 | return 1; |
780 | |
781 | j = isl_seq_first_non_zero(T->row[row] + 1 + i + 1, len - (i + 1)); |
782 | if (j >= 0) |
783 | return 1; |
784 | |
785 | for (j = 1; j < T->n_row; ++j) { |
786 | if (j == row) |
787 | continue; |
788 | if (!isl_int_is_zero(T->row[j][1 + i])(isl_sioimath_sgn(*(T->row[j][1 + i])) == 0)) |
789 | return 1; |
790 | } |
791 | |
792 | return 0; |
793 | } |
794 | |
795 | /* Does inequality constraint "ineq" of "bmap" involve any of |
796 | * the variables marked in "affected"? |
797 | * "total" is the total number of variables, i.e., the number |
798 | * of entries in "affected". |
799 | */ |
800 | static int is_affected(__isl_keep isl_basic_map *bmap, int ineq, int *affected, |
801 | int total) |
802 | { |
803 | int i; |
804 | |
805 | for (i = 0; i < total; ++i) { |
806 | if (!affected[i]) |
807 | continue; |
808 | if (!isl_int_is_zero(bmap->ineq[ineq][1 + i])(isl_sioimath_sgn(*(bmap->ineq[ineq][1 + i])) == 0)) |
809 | return 1; |
810 | } |
811 | |
812 | return 0; |
813 | } |
814 | |
815 | /* Given the compressed version of inequality constraint "ineq" |
816 | * of info->bmap in "v", check if the constraint can be tightened, |
817 | * where the compression is based on an equality constraint valid |
818 | * for info->tab. |
819 | * If so, add the tightened version of the inequality constraint |
820 | * to info->tab. "v" may be modified by this function. |
821 | * |
822 | * That is, if the compressed constraint is of the form |
823 | * |
824 | * m f() + c >= 0 |
825 | * |
826 | * with 0 < c < m, then it is equivalent to |
827 | * |
828 | * f() >= 0 |
829 | * |
830 | * This means that c can also be subtracted from the original, |
831 | * uncompressed constraint without affecting the integer points |
832 | * in info->tab. Add this tightened constraint as an extra row |
833 | * to info->tab to make this information explicitly available. |
834 | */ |
835 | static __isl_give isl_vec *try_tightening(struct isl_coalesce_info *info, |
836 | int ineq, __isl_take isl_vec *v) |
837 | { |
838 | isl_ctx *ctx; |
839 | int r; |
840 | |
841 | if (!v) |
842 | return NULL((void*)0); |
843 | |
844 | ctx = isl_vec_get_ctx(v); |
845 | isl_seq_gcd(v->el + 1, v->size - 1, &ctx->normalize_gcd); |
846 | if (isl_int_is_zero(ctx->normalize_gcd)(isl_sioimath_sgn(*(ctx->normalize_gcd)) == 0) || |
847 | isl_int_is_one(ctx->normalize_gcd)(isl_sioimath_cmp_si(*(ctx->normalize_gcd), 1) == 0)) { |
848 | return v; |
849 | } |
850 | |
851 | v = isl_vec_cow(v); |
852 | if (!v) |
853 | return NULL((void*)0); |
854 | |
855 | isl_int_fdiv_r(v->el[0], v->el[0], ctx->normalize_gcd)isl_sioimath_fdiv_r((v->el[0]), *(v->el[0]), *(ctx-> normalize_gcd)); |
856 | if (isl_int_is_zero(v->el[0])(isl_sioimath_sgn(*(v->el[0])) == 0)) |
857 | return v; |
858 | |
859 | if (isl_tab_extend_cons(info->tab, 1) < 0) |
860 | return isl_vec_free(v); |
861 | |
862 | isl_int_sub(info->bmap->ineq[ineq][0],isl_sioimath_sub((info->bmap->ineq[ineq][0]), *(info-> bmap->ineq[ineq][0]), *(v->el[0])) |
863 | info->bmap->ineq[ineq][0], v->el[0])isl_sioimath_sub((info->bmap->ineq[ineq][0]), *(info-> bmap->ineq[ineq][0]), *(v->el[0])); |
864 | r = isl_tab_add_ineq(info->tab, info->bmap->ineq[ineq]); |
865 | isl_int_add(info->bmap->ineq[ineq][0],isl_sioimath_add((info->bmap->ineq[ineq][0]), *(info-> bmap->ineq[ineq][0]), *(v->el[0])) |
866 | info->bmap->ineq[ineq][0], v->el[0])isl_sioimath_add((info->bmap->ineq[ineq][0]), *(info-> bmap->ineq[ineq][0]), *(v->el[0])); |
867 | |
868 | if (r < 0) |
869 | return isl_vec_free(v); |
870 | |
871 | return v; |
872 | } |
873 | |
874 | /* Tighten the (non-redundant) constraints on the facet represented |
875 | * by info->tab. |
876 | * In particular, on input, info->tab represents the result |
877 | * of relaxing the "n" inequality constraints of info->bmap in "relaxed" |
878 | * by one, i.e., replacing f_i >= 0 by f_i + 1 >= 0, and then |
879 | * replacing the one at index "l" by the corresponding equality, |
880 | * i.e., f_k + 1 = 0, with k = relaxed[l]. |
881 | * |
882 | * Compute a variable compression from the equality constraint f_k + 1 = 0 |
883 | * and use it to tighten the other constraints of info->bmap |
884 | * (that is, all constraints that have not been relaxed), |
885 | * updating info->tab (and leaving info->bmap untouched). |
886 | * The compression handles essentially two cases, one where a variable |
887 | * is assigned a fixed value and can therefore be eliminated, and one |
888 | * where one variable is a shifted multiple of some other variable and |
889 | * can therefore be replaced by that multiple. |
890 | * Gaussian elimination would also work for the first case, but for |
891 | * the second case, the effectiveness would depend on the order |
892 | * of the variables. |
893 | * After compression, some of the constraints may have coefficients |
894 | * with a common divisor. If this divisor does not divide the constant |
895 | * term, then the constraint can be tightened. |
896 | * The tightening is performed on the tableau info->tab by introducing |
897 | * extra (temporary) constraints. |
898 | * |
899 | * Only constraints that are possibly affected by the compression are |
900 | * considered. In particular, if the constraint only involves variables |
901 | * that are directly mapped to a distinct set of other variables, then |
902 | * no common divisor can be introduced and no tightening can occur. |
903 | * |
904 | * It is important to only consider the non-redundant constraints |
905 | * since the facet constraint has been relaxed prior to the call |
906 | * to this function, meaning that the constraints that were redundant |
907 | * prior to the relaxation may no longer be redundant. |
908 | * These constraints will be ignored in the fused result, so |
909 | * the fusion detection should not exploit them. |
910 | */ |
911 | static isl_stat tighten_on_relaxed_facet(struct isl_coalesce_info *info, |
912 | int n, int *relaxed, int l) |
913 | { |
914 | unsigned total; |
915 | isl_ctx *ctx; |
916 | isl_vec *v = NULL((void*)0); |
917 | isl_mat *T; |
918 | int i; |
919 | int k; |
920 | int *affected; |
921 | |
922 | k = relaxed[l]; |
923 | ctx = isl_basic_map_get_ctx(info->bmap); |
924 | total = isl_basic_map_total_dim(info->bmap); |
925 | isl_int_add_ui(info->bmap->ineq[k][0], info->bmap->ineq[k][0], 1)isl_sioimath_add_ui((info->bmap->ineq[k][0]), *(info-> bmap->ineq[k][0]), 1); |
926 | T = isl_mat_sub_alloc6(ctx, info->bmap->ineq, k, 1, 0, 1 + total); |
927 | T = isl_mat_variable_compression(T, NULL((void*)0)); |
928 | isl_int_sub_ui(info->bmap->ineq[k][0], info->bmap->ineq[k][0], 1)isl_sioimath_sub_ui((info->bmap->ineq[k][0]), *(info-> bmap->ineq[k][0]), 1); |
929 | if (!T) |
930 | return isl_stat_error; |
931 | if (T->n_col == 0) { |
932 | isl_mat_free(T); |
933 | return isl_stat_ok; |
934 | } |
935 | |
936 | affected = isl_alloc_array(ctx, int, total)((int *)isl_malloc_or_die(ctx, (total)*sizeof(int))); |
937 | if (!affected) |
938 | goto error; |
939 | |
940 | for (i = 0; i < total; ++i) |
941 | affected[i] = not_unique_unit_row(T, 1 + i); |
942 | |
943 | for (i = 0; i < info->bmap->n_ineq; ++i) { |
944 | if (any(relaxed, n, i)) |
945 | continue; |
946 | if (info->ineq[i] == STATUS_REDUNDANT1) |
947 | continue; |
948 | if (!is_affected(info->bmap, i, affected, total)) |
949 | continue; |
950 | v = isl_vec_alloc(ctx, 1 + total); |
951 | if (!v) |
952 | goto error; |
953 | isl_seq_cpy(v->el, info->bmap->ineq[i], 1 + total); |
954 | v = isl_vec_mat_product(v, isl_mat_copy(T)); |
955 | v = try_tightening(info, i, v); |
956 | isl_vec_free(v); |
957 | if (!v) |
958 | goto error; |
959 | } |
960 | |
961 | isl_mat_free(T); |
962 | free(affected); |
963 | return isl_stat_ok; |
964 | error: |
965 | isl_mat_free(T); |
966 | free(affected); |
967 | return isl_stat_error; |
968 | } |
969 | |
970 | /* Replace the basic maps "i" and "j" by an extension of "i" |
971 | * along the "n" inequality constraints in "relax" by one. |
972 | * The tableau info[i].tab has already been extended. |
973 | * Extend info[i].bmap accordingly by relaxing all constraints in "relax" |
974 | * by one. |
975 | * Each integer division that does not have exactly the same |
976 | * definition in "i" and "j" is marked unknown and the basic map |
977 | * is scheduled to be simplified in an attempt to recover |
978 | * the integer division definition. |
979 | * Place the extension in the position that is the smallest of i and j. |
980 | */ |
981 | static enum isl_change extend(int i, int j, int n, int *relax, |
982 | struct isl_coalesce_info *info) |
983 | { |
984 | int l; |
985 | unsigned total; |
986 | |
987 | info[i].bmap = isl_basic_map_cow(info[i].bmap); |
988 | if (!info[i].bmap) |
989 | return isl_change_error; |
990 | total = isl_basic_map_total_dim(info[i].bmap); |
991 | for (l = 0; l < info[i].bmap->n_div; ++l) |
992 | if (!isl_seq_eq(info[i].bmap->div[l], |
993 | info[j].bmap->div[l], 1 + 1 + total)) { |
994 | isl_int_set_si(info[i].bmap->div[l][0], 0)isl_sioimath_set_si((info[i].bmap->div[l][0]), 0); |
995 | info[i].simplify = 1; |
996 | } |
997 | for (l = 0; l < n; ++l) |
998 | isl_int_add_ui(info[i].bmap->ineq[relax[l]][0],isl_sioimath_add_ui((info[i].bmap->ineq[relax[l]][0]), *(info [i].bmap->ineq[relax[l]][0]), 1) |
999 | info[i].bmap->ineq[relax[l]][0], 1)isl_sioimath_add_ui((info[i].bmap->ineq[relax[l]][0]), *(info [i].bmap->ineq[relax[l]][0]), 1); |
1000 | ISL_F_SET(info[i].bmap, ISL_BASIC_MAP_FINAL)(((info[i].bmap)->flags) |= ((1 << 0))); |
1001 | drop(&info[j]); |
1002 | if (j < i) |
1003 | exchange(&info[i], &info[j]); |
1004 | return isl_change_fuse; |
1005 | } |
1006 | |
1007 | /* Basic map "i" has "n" inequality constraints (collected in "relax") |
1008 | * that are such that they include basic map "j" if they are relaxed |
1009 | * by one. All the other inequalities are valid for "j". |
1010 | * Check if basic map "j" forms an extension of basic map "i". |
1011 | * |
1012 | * In particular, relax the constraints in "relax", compute the corresponding |
1013 | * facets one by one and check whether each of these is included |
1014 | * in the other basic map. |
1015 | * Before testing for inclusion, the constraints on each facet |
1016 | * are tightened to increase the chance of an inclusion being detected. |
1017 | * (Adding the valid constraints of "j" to the tableau of "i", as is done |
1018 | * in is_adj_ineq_extension, may further increase those chances, but this |
1019 | * is not currently done.) |
1020 | * If each facet is included, we know that relaxing the constraints extends |
1021 | * the basic map with exactly the other basic map (we already know that this |
1022 | * other basic map is included in the extension, because all other |
1023 | * inequality constraints are valid of "j") and we can replace the |
1024 | * two basic maps by this extension. |
1025 | * ____ _____ |
1026 | * / || / | |
1027 | * / || / | |
1028 | * \ || => \ | |
1029 | * \ || \ | |
1030 | * \___|| \____| |
1031 | * |
1032 | * |
1033 | * \ |\ |
1034 | * |\\ | \ |
1035 | * | \\ | \ |
1036 | * | | => | / |
1037 | * | / | / |
1038 | * |/ |/ |
1039 | */ |
1040 | static enum isl_change is_relaxed_extension(int i, int j, int n, int *relax, |
1041 | struct isl_coalesce_info *info) |
1042 | { |
1043 | int l; |
1044 | int super; |
1045 | struct isl_tab_undo *snap, *snap2; |
1046 | unsigned n_eq = info[i].bmap->n_eq; |
1047 | |
1048 | for (l = 0; l < n; ++l) |
1049 | if (isl_tab_is_equality(info[i].tab, n_eq + relax[l])) |
1050 | return isl_change_none; |
1051 | |
1052 | snap = isl_tab_snap(info[i].tab); |
1053 | for (l = 0; l < n; ++l) |
1054 | if (isl_tab_relax(info[i].tab, n_eq + relax[l]) < 0) |
1055 | return isl_change_error; |
1056 | snap2 = isl_tab_snap(info[i].tab); |
1057 | for (l = 0; l < n; ++l) { |
1058 | if (isl_tab_rollback(info[i].tab, snap2) < 0) |
1059 | return isl_change_error; |
1060 | if (isl_tab_select_facet(info[i].tab, n_eq + relax[l]) < 0) |
1061 | return isl_change_error; |
1062 | if (tighten_on_relaxed_facet(&info[i], n, relax, l) < 0) |
1063 | return isl_change_error; |
1064 | super = contains(&info[j], info[i].tab); |
1065 | if (super < 0) |
1066 | return isl_change_error; |
1067 | if (super) |
1068 | continue; |
1069 | if (isl_tab_rollback(info[i].tab, snap) < 0) |
1070 | return isl_change_error; |
1071 | return isl_change_none; |
1072 | } |
1073 | |
1074 | if (isl_tab_rollback(info[i].tab, snap2) < 0) |
1075 | return isl_change_error; |
1076 | return extend(i, j, n, relax, info); |
1077 | } |
1078 | |
1079 | /* Data structure that keeps track of the wrapping constraints |
1080 | * and of information to bound the coefficients of those constraints. |
1081 | * |
1082 | * bound is set if we want to apply a bound on the coefficients |
1083 | * mat contains the wrapping constraints |
1084 | * max is the bound on the coefficients (if bound is set) |
1085 | */ |
1086 | struct isl_wraps { |
1087 | int bound; |
1088 | isl_mat *mat; |
1089 | isl_int max; |
1090 | }; |
1091 | |
1092 | /* Update wraps->max to be greater than or equal to the coefficients |
1093 | * in the equalities and inequalities of info->bmap that can be removed |
1094 | * if we end up applying wrapping. |
1095 | */ |
1096 | static void wraps_update_max(struct isl_wraps *wraps, |
1097 | struct isl_coalesce_info *info) |
1098 | { |
1099 | int k; |
1100 | isl_int max_k; |
1101 | unsigned total = isl_basic_map_total_dim(info->bmap); |
1102 | |
1103 | isl_int_init(max_k)isl_sioimath_init((max_k)); |
1104 | |
1105 | for (k = 0; k < info->bmap->n_eq; ++k) { |
1106 | if (info->eq[2 * k] == STATUS_VALID2 && |
1107 | info->eq[2 * k + 1] == STATUS_VALID2) |
1108 | continue; |
1109 | isl_seq_abs_max(info->bmap->eq[k] + 1, total, &max_k); |
1110 | if (isl_int_abs_gt(max_k, wraps->max)(isl_sioimath_abs_cmp(*(max_k), *(wraps->max)) > 0)) |
1111 | isl_int_set(wraps->max, max_k)isl_sioimath_set((wraps->max), *(max_k)); |
1112 | } |
1113 | |
1114 | for (k = 0; k < info->bmap->n_ineq; ++k) { |
1115 | if (info->ineq[k] == STATUS_VALID2 || |
1116 | info->ineq[k] == STATUS_REDUNDANT1) |
1117 | continue; |
1118 | isl_seq_abs_max(info->bmap->ineq[k] + 1, total, &max_k); |
1119 | if (isl_int_abs_gt(max_k, wraps->max)(isl_sioimath_abs_cmp(*(max_k), *(wraps->max)) > 0)) |
1120 | isl_int_set(wraps->max, max_k)isl_sioimath_set((wraps->max), *(max_k)); |
1121 | } |
1122 | |
1123 | isl_int_clear(max_k)isl_sioimath_clear((max_k)); |
1124 | } |
1125 | |
1126 | /* Initialize the isl_wraps data structure. |
1127 | * If we want to bound the coefficients of the wrapping constraints, |
1128 | * we set wraps->max to the largest coefficient |
1129 | * in the equalities and inequalities that can be removed if we end up |
1130 | * applying wrapping. |
1131 | */ |
1132 | static void wraps_init(struct isl_wraps *wraps, __isl_take isl_mat *mat, |
1133 | struct isl_coalesce_info *info, int i, int j) |
1134 | { |
1135 | isl_ctx *ctx; |
1136 | |
1137 | wraps->bound = 0; |
1138 | wraps->mat = mat; |
1139 | if (!mat) |
1140 | return; |
1141 | ctx = isl_mat_get_ctx(mat); |
1142 | wraps->bound = isl_options_get_coalesce_bounded_wrapping(ctx); |
1143 | if (!wraps->bound) |
1144 | return; |
1145 | isl_int_init(wraps->max)isl_sioimath_init((wraps->max)); |
1146 | isl_int_set_si(wraps->max, 0)isl_sioimath_set_si((wraps->max), 0); |
1147 | wraps_update_max(wraps, &info[i]); |
1148 | wraps_update_max(wraps, &info[j]); |
1149 | } |
1150 | |
1151 | /* Free the contents of the isl_wraps data structure. |
1152 | */ |
1153 | static void wraps_free(struct isl_wraps *wraps) |
1154 | { |
1155 | isl_mat_free(wraps->mat); |
1156 | if (wraps->bound) |
1157 | isl_int_clear(wraps->max)isl_sioimath_clear((wraps->max)); |
1158 | } |
1159 | |
1160 | /* Is the wrapping constraint in row "row" allowed? |
1161 | * |
1162 | * If wraps->bound is set, we check that none of the coefficients |
1163 | * is greater than wraps->max. |
1164 | */ |
1165 | static int allow_wrap(struct isl_wraps *wraps, int row) |
1166 | { |
1167 | int i; |
1168 | |
1169 | if (!wraps->bound) |
1170 | return 1; |
1171 | |
1172 | for (i = 1; i < wraps->mat->n_col; ++i) |
1173 | if (isl_int_abs_gt(wraps->mat->row[row][i], wraps->max)(isl_sioimath_abs_cmp(*(wraps->mat->row[row][i]), *(wraps ->max)) > 0)) |
1174 | return 0; |
1175 | |
1176 | return 1; |
1177 | } |
1178 | |
1179 | /* Wrap "ineq" (or its opposite if "negate" is set) around "bound" |
1180 | * to include "set" and add the result in position "w" of "wraps". |
1181 | * "len" is the total number of coefficients in "bound" and "ineq". |
1182 | * Return 1 on success, 0 on failure and -1 on error. |
1183 | * Wrapping can fail if the result of wrapping is equal to "bound" |
1184 | * or if we want to bound the sizes of the coefficients and |
1185 | * the wrapped constraint does not satisfy this bound. |
1186 | */ |
1187 | static int add_wrap(struct isl_wraps *wraps, int w, isl_int *bound, |
1188 | isl_int *ineq, unsigned len, __isl_keep isl_setisl_map *set, int negate) |
1189 | { |
1190 | isl_seq_cpy(wraps->mat->row[w], bound, len); |
1191 | if (negate) { |
1192 | isl_seq_neg(wraps->mat->row[w + 1], ineq, len); |
1193 | ineq = wraps->mat->row[w + 1]; |
1194 | } |
1195 | if (!isl_set_wrap_facet(set, wraps->mat->row[w], ineq)) |
1196 | return -1; |
1197 | if (isl_seq_eq(wraps->mat->row[w], bound, len)) |
1198 | return 0; |
1199 | if (!allow_wrap(wraps, w)) |
1200 | return 0; |
1201 | return 1; |
1202 | } |
1203 | |
1204 | /* For each constraint in info->bmap that is not redundant (as determined |
1205 | * by info->tab) and that is not a valid constraint for the other basic map, |
1206 | * wrap the constraint around "bound" such that it includes the whole |
1207 | * set "set" and append the resulting constraint to "wraps". |
1208 | * Note that the constraints that are valid for the other basic map |
1209 | * will be added to the combined basic map by default, so there is |
1210 | * no need to wrap them. |
1211 | * The caller wrap_in_facets even relies on this function not wrapping |
1212 | * any constraints that are already valid. |
1213 | * "wraps" is assumed to have been pre-allocated to the appropriate size. |
1214 | * wraps->n_row is the number of actual wrapped constraints that have |
1215 | * been added. |
1216 | * If any of the wrapping problems results in a constraint that is |
1217 | * identical to "bound", then this means that "set" is unbounded in such |
1218 | * way that no wrapping is possible. If this happens then wraps->n_row |
1219 | * is reset to zero. |
1220 | * Similarly, if we want to bound the coefficients of the wrapping |
1221 | * constraints and a newly added wrapping constraint does not |
1222 | * satisfy the bound, then wraps->n_row is also reset to zero. |
1223 | */ |
1224 | static int add_wraps(struct isl_wraps *wraps, struct isl_coalesce_info *info, |
1225 | isl_int *bound, __isl_keep isl_setisl_map *set) |
1226 | { |
1227 | int l, m; |
1228 | int w; |
1229 | int added; |
1230 | isl_basic_map *bmap = info->bmap; |
1231 | unsigned len = 1 + isl_basic_map_total_dim(bmap); |
1232 | |
1233 | w = wraps->mat->n_row; |
1234 | |
1235 | for (l = 0; l < bmap->n_ineq; ++l) { |
1236 | if (info->ineq[l] == STATUS_VALID2 || |
1237 | info->ineq[l] == STATUS_REDUNDANT1) |
1238 | continue; |
1239 | if (isl_seq_is_neg(bound, bmap->ineq[l], len)) |
1240 | continue; |
1241 | if (isl_seq_eq(bound, bmap->ineq[l], len)) |
1242 | continue; |
1243 | if (isl_tab_is_redundant(info->tab, bmap->n_eq + l)) |
1244 | continue; |
1245 | |
1246 | added = add_wrap(wraps, w, bound, bmap->ineq[l], len, set, 0); |
1247 | if (added < 0) |
1248 | return -1; |
1249 | if (!added) |
1250 | goto unbounded; |
1251 | ++w; |
1252 | } |
1253 | for (l = 0; l < bmap->n_eq; ++l) { |
1254 | if (isl_seq_is_neg(bound, bmap->eq[l], len)) |
1255 | continue; |
1256 | if (isl_seq_eq(bound, bmap->eq[l], len)) |
1257 | continue; |
1258 | |
1259 | for (m = 0; m < 2; ++m) { |
1260 | if (info->eq[2 * l + m] == STATUS_VALID2) |
1261 | continue; |
1262 | added = add_wrap(wraps, w, bound, bmap->eq[l], len, |
1263 | set, !m); |
1264 | if (added < 0) |
1265 | return -1; |
1266 | if (!added) |
1267 | goto unbounded; |
1268 | ++w; |
1269 | } |
1270 | } |
1271 | |
1272 | wraps->mat->n_row = w; |
1273 | return 0; |
1274 | unbounded: |
1275 | wraps->mat->n_row = 0; |
1276 | return 0; |
1277 | } |
1278 | |
1279 | /* Check if the constraints in "wraps" from "first" until the last |
1280 | * are all valid for the basic set represented by "tab". |
1281 | * If not, wraps->n_row is set to zero. |
1282 | */ |
1283 | static int check_wraps(__isl_keep isl_mat *wraps, int first, |
1284 | struct isl_tab *tab) |
1285 | { |
1286 | int i; |
1287 | |
1288 | for (i = first; i < wraps->n_row; ++i) { |
1289 | enum isl_ineq_type type; |
1290 | type = isl_tab_ineq_type(tab, wraps->row[i]); |
1291 | if (type == isl_ineq_error) |
1292 | return -1; |
1293 | if (type == isl_ineq_redundant) |
1294 | continue; |
1295 | wraps->n_row = 0; |
1296 | return 0; |
1297 | } |
1298 | |
1299 | return 0; |
1300 | } |
1301 | |
1302 | /* Return a set that corresponds to the non-redundant constraints |
1303 | * (as recorded in tab) of bmap. |
1304 | * |
1305 | * It's important to remove the redundant constraints as some |
1306 | * of the other constraints may have been modified after the |
1307 | * constraints were marked redundant. |
1308 | * In particular, a constraint may have been relaxed. |
1309 | * Redundant constraints are ignored when a constraint is relaxed |
1310 | * and should therefore continue to be ignored ever after. |
1311 | * Otherwise, the relaxation might be thwarted by some of |
1312 | * these constraints. |
1313 | * |
1314 | * Update the underlying set to ensure that the dimension doesn't change. |
1315 | * Otherwise the integer divisions could get dropped if the tab |
1316 | * turns out to be empty. |
1317 | */ |
1318 | static __isl_give isl_setisl_map *set_from_updated_bmap(__isl_keep isl_basic_map *bmap, |
1319 | struct isl_tab *tab) |
1320 | { |
1321 | isl_basic_setisl_basic_map *bset; |
1322 | |
1323 | bmap = isl_basic_map_copy(bmap); |
1324 | bset = isl_basic_map_underlying_set(bmap); |
1325 | bset = isl_basic_set_cow(bset); |
1326 | bset = isl_basic_set_update_from_tab(bset, tab); |
1327 | return isl_set_from_basic_set(bset); |
1328 | } |
1329 | |
1330 | /* Wrap the constraints of info->bmap that bound the facet defined |
1331 | * by inequality "k" around (the opposite of) this inequality to |
1332 | * include "set". "bound" may be used to store the negated inequality. |
1333 | * Since the wrapped constraints are not guaranteed to contain the whole |
1334 | * of info->bmap, we check them in check_wraps. |
1335 | * If any of the wrapped constraints turn out to be invalid, then |
1336 | * check_wraps will reset wrap->n_row to zero. |
1337 | */ |
1338 | static int add_wraps_around_facet(struct isl_wraps *wraps, |
1339 | struct isl_coalesce_info *info, int k, isl_int *bound, |
1340 | __isl_keep isl_setisl_map *set) |
1341 | { |
1342 | struct isl_tab_undo *snap; |
1343 | int n; |
1344 | unsigned total = isl_basic_map_total_dim(info->bmap); |
1345 | |
1346 | snap = isl_tab_snap(info->tab); |
1347 | |
1348 | if (isl_tab_select_facet(info->tab, info->bmap->n_eq + k) < 0) |
1349 | return -1; |
1350 | if (isl_tab_detect_redundant(info->tab) < 0) |
1351 | return -1; |
1352 | |
1353 | isl_seq_neg(bound, info->bmap->ineq[k], 1 + total); |
1354 | |
1355 | n = wraps->mat->n_row; |
1356 | if (add_wraps(wraps, info, bound, set) < 0) |
1357 | return -1; |
1358 | |
1359 | if (isl_tab_rollback(info->tab, snap) < 0) |
1360 | return -1; |
1361 | if (check_wraps(wraps->mat, n, info->tab) < 0) |
1362 | return -1; |
1363 | |
1364 | return 0; |
1365 | } |
1366 | |
1367 | /* Given a basic set i with a constraint k that is adjacent to |
1368 | * basic set j, check if we can wrap |
1369 | * both the facet corresponding to k (if "wrap_facet" is set) and basic map j |
1370 | * (always) around their ridges to include the other set. |
1371 | * If so, replace the pair of basic sets by their union. |
1372 | * |
1373 | * All constraints of i (except k) are assumed to be valid or |
1374 | * cut constraints for j. |
1375 | * Wrapping the cut constraints to include basic map j may result |
1376 | * in constraints that are no longer valid of basic map i |
1377 | * we have to check that the resulting wrapping constraints are valid for i. |
1378 | * If "wrap_facet" is not set, then all constraints of i (except k) |
1379 | * are assumed to be valid for j. |
1380 | * ____ _____ |
1381 | * / | / \ |
1382 | * / || / | |
1383 | * \ || => \ | |
1384 | * \ || \ | |
1385 | * \___|| \____| |
1386 | * |
1387 | */ |
1388 | static enum isl_change can_wrap_in_facet(int i, int j, int k, |
1389 | struct isl_coalesce_info *info, int wrap_facet) |
1390 | { |
1391 | enum isl_change change = isl_change_none; |
1392 | struct isl_wraps wraps; |
1393 | isl_ctx *ctx; |
1394 | isl_mat *mat; |
1395 | struct isl_setisl_map *set_i = NULL((void*)0); |
1396 | struct isl_setisl_map *set_j = NULL((void*)0); |
1397 | struct isl_vec *bound = NULL((void*)0); |
1398 | unsigned total = isl_basic_map_total_dim(info[i].bmap); |
1399 | |
1400 | set_i = set_from_updated_bmap(info[i].bmap, info[i].tab); |
1401 | set_j = set_from_updated_bmap(info[j].bmap, info[j].tab); |
1402 | ctx = isl_basic_map_get_ctx(info[i].bmap); |
1403 | mat = isl_mat_alloc(ctx, 2 * (info[i].bmap->n_eq + info[j].bmap->n_eq) + |
1404 | info[i].bmap->n_ineq + info[j].bmap->n_ineq, |
1405 | 1 + total); |
1406 | wraps_init(&wraps, mat, info, i, j); |
1407 | bound = isl_vec_alloc(ctx, 1 + total); |
1408 | if (!set_i || !set_j || !wraps.mat || !bound) |
1409 | goto error; |
1410 | |
1411 | isl_seq_cpy(bound->el, info[i].bmap->ineq[k], 1 + total); |
1412 | isl_int_add_ui(bound->el[0], bound->el[0], 1)isl_sioimath_add_ui((bound->el[0]), *(bound->el[0]), 1); |
1413 | |
1414 | isl_seq_cpy(wraps.mat->row[0], bound->el, 1 + total); |
1415 | wraps.mat->n_row = 1; |
1416 | |
1417 | if (add_wraps(&wraps, &info[j], bound->el, set_i) < 0) |
1418 | goto error; |
1419 | if (!wraps.mat->n_row) |
1420 | goto unbounded; |
1421 | |
1422 | if (wrap_facet) { |
1423 | if (add_wraps_around_facet(&wraps, &info[i], k, |
1424 | bound->el, set_j) < 0) |
1425 | goto error; |
1426 | if (!wraps.mat->n_row) |
1427 | goto unbounded; |
1428 | } |
1429 | |
1430 | change = fuse(i, j, info, wraps.mat, 0, 0); |
1431 | |
1432 | unbounded: |
1433 | wraps_free(&wraps); |
1434 | |
1435 | isl_set_free(set_i); |
1436 | isl_set_free(set_j); |
1437 | |
1438 | isl_vec_free(bound); |
1439 | |
1440 | return change; |
1441 | error: |
1442 | wraps_free(&wraps); |
1443 | isl_vec_free(bound); |
1444 | isl_set_free(set_i); |
1445 | isl_set_free(set_j); |
1446 | return isl_change_error; |
1447 | } |
1448 | |
1449 | /* Given a cut constraint t(x) >= 0 of basic map i, stored in row "w" |
1450 | * of wrap.mat, replace it by its relaxed version t(x) + 1 >= 0, and |
1451 | * add wrapping constraints to wrap.mat for all constraints |
1452 | * of basic map j that bound the part of basic map j that sticks out |
1453 | * of the cut constraint. |
1454 | * "set_i" is the underlying set of basic map i. |
1455 | * If any wrapping fails, then wraps->mat.n_row is reset to zero. |
1456 | * |
1457 | * In particular, we first intersect basic map j with t(x) + 1 = 0. |
1458 | * If the result is empty, then t(x) >= 0 was actually a valid constraint |
1459 | * (with respect to the integer points), so we add t(x) >= 0 instead. |
1460 | * Otherwise, we wrap the constraints of basic map j that are not |
1461 | * redundant in this intersection and that are not already valid |
1462 | * for basic map i over basic map i. |
1463 | * Note that it is sufficient to wrap the constraints to include |
1464 | * basic map i, because we will only wrap the constraints that do |
1465 | * not include basic map i already. The wrapped constraint will |
1466 | * therefore be more relaxed compared to the original constraint. |
1467 | * Since the original constraint is valid for basic map j, so is |
1468 | * the wrapped constraint. |
1469 | */ |
1470 | static isl_stat wrap_in_facet(struct isl_wraps *wraps, int w, |
1471 | struct isl_coalesce_info *info_j, __isl_keep isl_setisl_map *set_i, |
1472 | struct isl_tab_undo *snap) |
1473 | { |
1474 | isl_int_add_ui(wraps->mat->row[w][0], wraps->mat->row[w][0], 1)isl_sioimath_add_ui((wraps->mat->row[w][0]), *(wraps-> mat->row[w][0]), 1); |
1475 | if (isl_tab_add_eq(info_j->tab, wraps->mat->row[w]) < 0) |
1476 | return isl_stat_error; |
1477 | if (isl_tab_detect_redundant(info_j->tab) < 0) |
1478 | return isl_stat_error; |
1479 | |
1480 | if (info_j->tab->empty) |
1481 | isl_int_sub_ui(wraps->mat->row[w][0], wraps->mat->row[w][0], 1)isl_sioimath_sub_ui((wraps->mat->row[w][0]), *(wraps-> mat->row[w][0]), 1); |
1482 | else if (add_wraps(wraps, info_j, wraps->mat->row[w], set_i) < 0) |
1483 | return isl_stat_error; |
1484 | |
1485 | if (isl_tab_rollback(info_j->tab, snap) < 0) |
1486 | return isl_stat_error; |
1487 | |
1488 | return isl_stat_ok; |
1489 | } |
1490 | |
1491 | /* Given a pair of basic maps i and j such that j sticks out |
1492 | * of i at n cut constraints, each time by at most one, |
1493 | * try to compute wrapping constraints and replace the two |
1494 | * basic maps by a single basic map. |
1495 | * The other constraints of i are assumed to be valid for j. |
1496 | * "set_i" is the underlying set of basic map i. |
1497 | * "wraps" has been initialized to be of the right size. |
1498 | * |
1499 | * For each cut constraint t(x) >= 0 of i, we add the relaxed version |
1500 | * t(x) + 1 >= 0, along with wrapping constraints for all constraints |
1501 | * of basic map j that bound the part of basic map j that sticks out |
1502 | * of the cut constraint. |
1503 | * |
1504 | * If any wrapping fails, i.e., if we cannot wrap to touch |
1505 | * the union, then we give up. |
1506 | * Otherwise, the pair of basic maps is replaced by their union. |
1507 | */ |
1508 | static enum isl_change try_wrap_in_facets(int i, int j, |
1509 | struct isl_coalesce_info *info, struct isl_wraps *wraps, |
1510 | __isl_keep isl_setisl_map *set_i) |
1511 | { |
1512 | int k, l, w; |
1513 | unsigned total; |
1514 | struct isl_tab_undo *snap; |
1515 | |
1516 | total = isl_basic_map_total_dim(info[i].bmap); |
1517 | |
1518 | snap = isl_tab_snap(info[j].tab); |
1519 | |
1520 | wraps->mat->n_row = 0; |
1521 | |
1522 | for (k = 0; k < info[i].bmap->n_eq; ++k) { |
1523 | for (l = 0; l < 2; ++l) { |
1524 | if (info[i].eq[2 * k + l] != STATUS_CUT4) |
1525 | continue; |
1526 | w = wraps->mat->n_row++; |
1527 | if (l == 0) |
1528 | isl_seq_neg(wraps->mat->row[w], |
1529 | info[i].bmap->eq[k], 1 + total); |
1530 | else |
1531 | isl_seq_cpy(wraps->mat->row[w], |
1532 | info[i].bmap->eq[k], 1 + total); |
1533 | if (wrap_in_facet(wraps, w, &info[j], set_i, snap) < 0) |
1534 | return isl_change_error; |
1535 | |
1536 | if (!wraps->mat->n_row) |
1537 | return isl_change_none; |
1538 | } |
1539 | } |
1540 | |
1541 | for (k = 0; k < info[i].bmap->n_ineq; ++k) { |
1542 | if (info[i].ineq[k] != STATUS_CUT4) |
1543 | continue; |
1544 | w = wraps->mat->n_row++; |
1545 | isl_seq_cpy(wraps->mat->row[w], |
1546 | info[i].bmap->ineq[k], 1 + total); |
1547 | if (wrap_in_facet(wraps, w, &info[j], set_i, snap) < 0) |
1548 | return isl_change_error; |
1549 | |
1550 | if (!wraps->mat->n_row) |
1551 | return isl_change_none; |
1552 | } |
1553 | |
1554 | return fuse(i, j, info, wraps->mat, 0, 1); |
1555 | } |
1556 | |
1557 | /* Given a pair of basic maps i and j such that j sticks out |
1558 | * of i at n cut constraints, each time by at most one, |
1559 | * try to compute wrapping constraints and replace the two |
1560 | * basic maps by a single basic map. |
1561 | * The other constraints of i are assumed to be valid for j. |
1562 | * |
1563 | * The core computation is performed by try_wrap_in_facets. |
1564 | * This function simply extracts an underlying set representation |
1565 | * of basic map i and initializes the data structure for keeping |
1566 | * track of wrapping constraints. |
1567 | */ |
1568 | static enum isl_change wrap_in_facets(int i, int j, int n, |
1569 | struct isl_coalesce_info *info) |
1570 | { |
1571 | enum isl_change change = isl_change_none; |
1572 | struct isl_wraps wraps; |
1573 | isl_ctx *ctx; |
1574 | isl_mat *mat; |
1575 | isl_setisl_map *set_i = NULL((void*)0); |
1576 | unsigned total = isl_basic_map_total_dim(info[i].bmap); |
1577 | int max_wrap; |
1578 | |
1579 | if (isl_tab_extend_cons(info[j].tab, 1) < 0) |
1580 | return isl_change_error; |
1581 | |
1582 | max_wrap = 1 + 2 * info[j].bmap->n_eq + info[j].bmap->n_ineq; |
1583 | max_wrap *= n; |
1584 | |
1585 | set_i = set_from_updated_bmap(info[i].bmap, info[i].tab); |
1586 | ctx = isl_basic_map_get_ctx(info[i].bmap); |
1587 | mat = isl_mat_alloc(ctx, max_wrap, 1 + total); |
1588 | wraps_init(&wraps, mat, info, i, j); |
1589 | if (!set_i || !wraps.mat) |
1590 | goto error; |
1591 | |
1592 | change = try_wrap_in_facets(i, j, info, &wraps, set_i); |
1593 | |
1594 | wraps_free(&wraps); |
1595 | isl_set_free(set_i); |
1596 | |
1597 | return change; |
1598 | error: |
1599 | wraps_free(&wraps); |
1600 | isl_set_free(set_i); |
1601 | return isl_change_error; |
1602 | } |
1603 | |
1604 | /* Return the effect of inequality "ineq" on the tableau "tab", |
1605 | * after relaxing the constant term of "ineq" by one. |
1606 | */ |
1607 | static enum isl_ineq_type type_of_relaxed(struct isl_tab *tab, isl_int *ineq) |
1608 | { |
1609 | enum isl_ineq_type type; |
1610 | |
1611 | isl_int_add_ui(ineq[0], ineq[0], 1)isl_sioimath_add_ui((ineq[0]), *(ineq[0]), 1); |
1612 | type = isl_tab_ineq_type(tab, ineq); |
1613 | isl_int_sub_ui(ineq[0], ineq[0], 1)isl_sioimath_sub_ui((ineq[0]), *(ineq[0]), 1); |
1614 | |
1615 | return type; |
1616 | } |
1617 | |
1618 | /* Given two basic sets i and j, |
1619 | * check if relaxing all the cut constraints of i by one turns |
1620 | * them into valid constraint for j and check if we can wrap in |
1621 | * the bits that are sticking out. |
1622 | * If so, replace the pair by their union. |
1623 | * |
1624 | * We first check if all relaxed cut inequalities of i are valid for j |
1625 | * and then try to wrap in the intersections of the relaxed cut inequalities |
1626 | * with j. |
1627 | * |
1628 | * During this wrapping, we consider the points of j that lie at a distance |
1629 | * of exactly 1 from i. In particular, we ignore the points that lie in |
1630 | * between this lower-dimensional space and the basic map i. |
1631 | * We can therefore only apply this to integer maps. |
1632 | * ____ _____ |
1633 | * / ___|_ / \ |
1634 | * / | | / | |
1635 | * \ | | => \ | |
1636 | * \|____| \ | |
1637 | * \___| \____/ |
1638 | * |
1639 | * _____ ______ |
1640 | * | ____|_ | \ |
1641 | * | | | | | |
1642 | * | | | => | | |
1643 | * |_| | | | |
1644 | * |_____| \______| |
1645 | * |
1646 | * _______ |
1647 | * | | |
1648 | * | |\ | |
1649 | * | | \ | |
1650 | * | | \ | |
1651 | * | | \| |
1652 | * | | \ |
1653 | * | |_____\ |
1654 | * | | |
1655 | * |_______| |
1656 | * |
1657 | * Wrapping can fail if the result of wrapping one of the facets |
1658 | * around its edges does not produce any new facet constraint. |
1659 | * In particular, this happens when we try to wrap in unbounded sets. |
1660 | * |
1661 | * _______________________________________________________________________ |
1662 | * | |
1663 | * | ___ |
1664 | * | | | |
1665 | * |_| |_________________________________________________________________ |
1666 | * |___| |
1667 | * |
1668 | * The following is not an acceptable result of coalescing the above two |
1669 | * sets as it includes extra integer points. |
1670 | * _______________________________________________________________________ |
1671 | * | |
1672 | * | |
1673 | * | |
1674 | * | |
1675 | * \______________________________________________________________________ |
1676 | */ |
1677 | static enum isl_change can_wrap_in_set(int i, int j, |
1678 | struct isl_coalesce_info *info) |
1679 | { |
1680 | int k, l; |
1681 | int n; |
1682 | unsigned total; |
1683 | |
1684 | if (ISL_F_ISSET(info[i].bmap, ISL_BASIC_MAP_RATIONAL)(!!(((info[i].bmap)->flags) & ((1 << 4)))) || |
1685 | ISL_F_ISSET(info[j].bmap, ISL_BASIC_MAP_RATIONAL)(!!(((info[j].bmap)->flags) & ((1 << 4))))) |
1686 | return isl_change_none; |
1687 | |
1688 | n = count(info[i].eq, 2 * info[i].bmap->n_eq, STATUS_CUT4); |
1689 | n += count(info[i].ineq, info[i].bmap->n_ineq, STATUS_CUT4); |
1690 | if (n == 0) |
1691 | return isl_change_none; |
1692 | |
1693 | total = isl_basic_map_total_dim(info[i].bmap); |
1694 | for (k = 0; k < info[i].bmap->n_eq; ++k) { |
1695 | for (l = 0; l < 2; ++l) { |
1696 | enum isl_ineq_type type; |
1697 | |
1698 | if (info[i].eq[2 * k + l] != STATUS_CUT4) |
1699 | continue; |
1700 | |
1701 | if (l == 0) |
1702 | isl_seq_neg(info[i].bmap->eq[k], |
1703 | info[i].bmap->eq[k], 1 + total); |
1704 | type = type_of_relaxed(info[j].tab, |
1705 | info[i].bmap->eq[k]); |
1706 | if (l == 0) |
1707 | isl_seq_neg(info[i].bmap->eq[k], |
1708 | info[i].bmap->eq[k], 1 + total); |
1709 | if (type == isl_ineq_error) |
1710 | return isl_change_error; |
1711 | if (type != isl_ineq_redundant) |
1712 | return isl_change_none; |
1713 | } |
1714 | } |
1715 | |
1716 | for (k = 0; k < info[i].bmap->n_ineq; ++k) { |
1717 | enum isl_ineq_type type; |
1718 | |
1719 | if (info[i].ineq[k] != STATUS_CUT4) |
1720 | continue; |
1721 | |
1722 | type = type_of_relaxed(info[j].tab, info[i].bmap->ineq[k]); |
1723 | if (type == isl_ineq_error) |
1724 | return isl_change_error; |
1725 | if (type != isl_ineq_redundant) |
1726 | return isl_change_none; |
1727 | } |
1728 | |
1729 | return wrap_in_facets(i, j, n, info); |
1730 | } |
1731 | |
1732 | /* Check if either i or j has only cut constraints that can |
1733 | * be used to wrap in (a facet of) the other basic set. |
1734 | * if so, replace the pair by their union. |
1735 | */ |
1736 | static enum isl_change check_wrap(int i, int j, struct isl_coalesce_info *info) |
1737 | { |
1738 | enum isl_change change = isl_change_none; |
1739 | |
1740 | change = can_wrap_in_set(i, j, info); |
1741 | if (change != isl_change_none) |
1742 | return change; |
1743 | |
1744 | change = can_wrap_in_set(j, i, info); |
1745 | return change; |
1746 | } |
1747 | |
1748 | /* Check if all inequality constraints of "i" that cut "j" cease |
1749 | * to be cut constraints if they are relaxed by one. |
1750 | * If so, collect the cut constraints in "list". |
1751 | * The caller is responsible for allocating "list". |
1752 | */ |
1753 | static isl_bool all_cut_by_one(int i, int j, struct isl_coalesce_info *info, |
1754 | int *list) |
1755 | { |
1756 | int l, n; |
1757 | |
1758 | n = 0; |
1759 | for (l = 0; l < info[i].bmap->n_ineq; ++l) { |
1760 | enum isl_ineq_type type; |
1761 | |
1762 | if (info[i].ineq[l] != STATUS_CUT4) |
1763 | continue; |
1764 | type = type_of_relaxed(info[j].tab, info[i].bmap->ineq[l]); |
1765 | if (type == isl_ineq_error) |
1766 | return isl_bool_error; |
1767 | if (type != isl_ineq_redundant) |
1768 | return isl_bool_false; |
1769 | list[n++] = l; |
1770 | } |
1771 | |
1772 | return isl_bool_true; |
1773 | } |
1774 | |
1775 | /* Given two basic maps such that "j" has at least one equality constraint |
1776 | * that is adjacent to an inequality constraint of "i" and such that "i" has |
1777 | * exactly one inequality constraint that is adjacent to an equality |
1778 | * constraint of "j", check whether "i" can be extended to include "j" or |
1779 | * whether "j" can be wrapped into "i". |
1780 | * All remaining constraints of "i" and "j" are assumed to be valid |
1781 | * or cut constraints of the other basic map. |
1782 | * However, none of the equality constraints of "i" are cut constraints. |
1783 | * |
1784 | * If "i" has any "cut" inequality constraints, then check if relaxing |
1785 | * each of them by one is sufficient for them to become valid. |
1786 | * If so, check if the inequality constraint adjacent to an equality |
1787 | * constraint of "j" along with all these cut constraints |
1788 | * can be relaxed by one to contain exactly "j". |
1789 | * Otherwise, or if this fails, check if "j" can be wrapped into "i". |
1790 | */ |
1791 | static enum isl_change check_single_adj_eq(int i, int j, |
1792 | struct isl_coalesce_info *info) |
1793 | { |
1794 | enum isl_change change = isl_change_none; |
1795 | int k; |
1796 | int n_cut; |
1797 | int *relax; |
1798 | isl_ctx *ctx; |
1799 | isl_bool try_relax; |
1800 | |
1801 | n_cut = count(info[i].ineq, info[i].bmap->n_ineq, STATUS_CUT4); |
1802 | |
1803 | k = find(info[i].ineq, info[i].bmap->n_ineq, STATUS_ADJ_EQ5); |
1804 | |
1805 | if (n_cut > 0) { |
1806 | ctx = isl_basic_map_get_ctx(info[i].bmap); |
1807 | relax = isl_calloc_array(ctx, int, 1 + n_cut)((int *)isl_calloc_or_die(ctx, 1 + n_cut, sizeof(int))); |
1808 | if (!relax) |
1809 | return isl_change_error; |
1810 | relax[0] = k; |
1811 | try_relax = all_cut_by_one(i, j, info, relax + 1); |
1812 | if (try_relax < 0) |
1813 | change = isl_change_error; |
1814 | } else { |
1815 | try_relax = isl_bool_true; |
1816 | relax = &k; |
1817 | } |
1818 | if (try_relax && change == isl_change_none) |
1819 | change = is_relaxed_extension(i, j, 1 + n_cut, relax, info); |
1820 | if (n_cut > 0) |
1821 | free(relax); |
1822 | if (change != isl_change_none) |
1823 | return change; |
1824 | |
1825 | change = can_wrap_in_facet(i, j, k, info, n_cut > 0); |
1826 | |
1827 | return change; |
1828 | } |
1829 | |
1830 | /* At least one of the basic maps has an equality that is adjacent |
1831 | * to inequality. Make sure that only one of the basic maps has |
1832 | * such an equality and that the other basic map has exactly one |
1833 | * inequality adjacent to an equality. |
1834 | * If the other basic map does not have such an inequality, then |
1835 | * check if all its constraints are either valid or cut constraints |
1836 | * and, if so, try wrapping in the first map into the second. |
1837 | * Otherwise, try to extend one basic map with the other or |
1838 | * wrap one basic map in the other. |
1839 | */ |
1840 | static enum isl_change check_adj_eq(int i, int j, |
1841 | struct isl_coalesce_info *info) |
1842 | { |
1843 | if (any(info[i].eq, 2 * info[i].bmap->n_eq, STATUS_ADJ_INEQ6) && |
1844 | any(info[j].eq, 2 * info[j].bmap->n_eq, STATUS_ADJ_INEQ6)) |
1845 | /* ADJ EQ TOO MANY */ |
1846 | return isl_change_none; |
1847 | |
1848 | if (any(info[i].eq, 2 * info[i].bmap->n_eq, STATUS_ADJ_INEQ6)) |
1849 | return check_adj_eq(j, i, info); |
1850 | |
1851 | /* j has an equality adjacent to an inequality in i */ |
1852 | |
1853 | if (count(info[i].ineq, info[i].bmap->n_ineq, STATUS_ADJ_EQ5) != 1) { |
1854 | if (all_valid_or_cut(&info[i])) |
1855 | return can_wrap_in_set(i, j, info); |
1856 | return isl_change_none; |
1857 | } |
1858 | if (any(info[i].eq, 2 * info[i].bmap->n_eq, STATUS_CUT4)) |
1859 | return isl_change_none; |
1860 | if (any(info[j].ineq, info[j].bmap->n_ineq, STATUS_ADJ_EQ5) || |
1861 | any(info[i].ineq, info[i].bmap->n_ineq, STATUS_ADJ_INEQ6) || |
1862 | any(info[j].ineq, info[j].bmap->n_ineq, STATUS_ADJ_INEQ6)) |
1863 | /* ADJ EQ TOO MANY */ |
1864 | return isl_change_none; |
1865 | |
1866 | return check_single_adj_eq(i, j, info); |
1867 | } |
1868 | |
1869 | /* The two basic maps lie on adjacent hyperplanes. In particular, |
1870 | * basic map "i" has an equality that lies parallel to basic map "j". |
1871 | * Check if we can wrap the facets around the parallel hyperplanes |
1872 | * to include the other set. |
1873 | * |
1874 | * We perform basically the same operations as can_wrap_in_facet, |
1875 | * except that we don't need to select a facet of one of the sets. |
1876 | * _ |
1877 | * \\ \\ |
1878 | * \\ => \\ |
1879 | * \ \| |
1880 | * |
1881 | * If there is more than one equality of "i" adjacent to an equality of "j", |
1882 | * then the result will satisfy one or more equalities that are a linear |
1883 | * combination of these equalities. These will be encoded as pairs |
1884 | * of inequalities in the wrapping constraints and need to be made |
1885 | * explicit. |
1886 | */ |
1887 | static enum isl_change check_eq_adj_eq(int i, int j, |
1888 | struct isl_coalesce_info *info) |
1889 | { |
1890 | int k; |
1891 | enum isl_change change = isl_change_none; |
1892 | int detect_equalities = 0; |
1893 | struct isl_wraps wraps; |
1894 | isl_ctx *ctx; |
1895 | isl_mat *mat; |
1896 | struct isl_setisl_map *set_i = NULL((void*)0); |
1897 | struct isl_setisl_map *set_j = NULL((void*)0); |
1898 | struct isl_vec *bound = NULL((void*)0); |
1899 | unsigned total = isl_basic_map_total_dim(info[i].bmap); |
1900 | |
1901 | if (count(info[i].eq, 2 * info[i].bmap->n_eq, STATUS_ADJ_EQ5) != 1) |
1902 | detect_equalities = 1; |
1903 | |
1904 | k = find(info[i].eq, 2 * info[i].bmap->n_eq, STATUS_ADJ_EQ5); |
1905 | |
1906 | set_i = set_from_updated_bmap(info[i].bmap, info[i].tab); |
1907 | set_j = set_from_updated_bmap(info[j].bmap, info[j].tab); |
1908 | ctx = isl_basic_map_get_ctx(info[i].bmap); |
1909 | mat = isl_mat_alloc(ctx, 2 * (info[i].bmap->n_eq + info[j].bmap->n_eq) + |
1910 | info[i].bmap->n_ineq + info[j].bmap->n_ineq, |
1911 | 1 + total); |
1912 | wraps_init(&wraps, mat, info, i, j); |
1913 | bound = isl_vec_alloc(ctx, 1 + total); |
1914 | if (!set_i || !set_j || !wraps.mat || !bound) |
1915 | goto error; |
1916 | |
1917 | if (k % 2 == 0) |
1918 | isl_seq_neg(bound->el, info[i].bmap->eq[k / 2], 1 + total); |
1919 | else |
1920 | isl_seq_cpy(bound->el, info[i].bmap->eq[k / 2], 1 + total); |
1921 | isl_int_add_ui(bound->el[0], bound->el[0], 1)isl_sioimath_add_ui((bound->el[0]), *(bound->el[0]), 1); |
1922 | |
1923 | isl_seq_cpy(wraps.mat->row[0], bound->el, 1 + total); |
1924 | wraps.mat->n_row = 1; |
1925 | |
1926 | if (add_wraps(&wraps, &info[j], bound->el, set_i) < 0) |
1927 | goto error; |
1928 | if (!wraps.mat->n_row) |
1929 | goto unbounded; |
1930 | |
1931 | isl_int_sub_ui(bound->el[0], bound->el[0], 1)isl_sioimath_sub_ui((bound->el[0]), *(bound->el[0]), 1); |
1932 | isl_seq_neg(bound->el, bound->el, 1 + total); |
1933 | |
1934 | isl_seq_cpy(wraps.mat->row[wraps.mat->n_row], bound->el, 1 + total); |
1935 | wraps.mat->n_row++; |
1936 | |
1937 | if (add_wraps(&wraps, &info[i], bound->el, set_j) < 0) |
1938 | goto error; |
1939 | if (!wraps.mat->n_row) |
1940 | goto unbounded; |
1941 | |
1942 | change = fuse(i, j, info, wraps.mat, detect_equalities, 0); |
1943 | |
1944 | if (0) { |
1945 | error: change = isl_change_error; |
1946 | } |
1947 | unbounded: |
1948 | |
1949 | wraps_free(&wraps); |
1950 | isl_set_free(set_i); |
1951 | isl_set_free(set_j); |
1952 | isl_vec_free(bound); |
1953 | |
1954 | return change; |
1955 | } |
1956 | |
1957 | /* Initialize the "eq" and "ineq" fields of "info". |
1958 | */ |
1959 | static void init_status(struct isl_coalesce_info *info) |
1960 | { |
1961 | info->eq = info->ineq = NULL((void*)0); |
1962 | } |
1963 | |
1964 | /* Set info->eq to the positions of the equalities of info->bmap |
1965 | * with respect to the basic map represented by "tab". |
1966 | * If info->eq has already been computed, then do not compute it again. |
1967 | */ |
1968 | static void set_eq_status_in(struct isl_coalesce_info *info, |
1969 | struct isl_tab *tab) |
1970 | { |
1971 | if (info->eq) |
1972 | return; |
1973 | info->eq = eq_status_in(info->bmap, tab); |
1974 | } |
1975 | |
1976 | /* Set info->ineq to the positions of the inequalities of info->bmap |
1977 | * with respect to the basic map represented by "tab". |
1978 | * If info->ineq has already been computed, then do not compute it again. |
1979 | */ |
1980 | static void set_ineq_status_in(struct isl_coalesce_info *info, |
1981 | struct isl_tab *tab) |
1982 | { |
1983 | if (info->ineq) |
1984 | return; |
1985 | info->ineq = ineq_status_in(info->bmap, info->tab, tab); |
1986 | } |
1987 | |
1988 | /* Free the memory allocated by the "eq" and "ineq" fields of "info". |
1989 | * This function assumes that init_status has been called on "info" first, |
1990 | * after which the "eq" and "ineq" fields may or may not have been |
1991 | * assigned a newly allocated array. |
1992 | */ |
1993 | static void clear_status(struct isl_coalesce_info *info) |
1994 | { |
1995 | free(info->eq); |
1996 | free(info->ineq); |
1997 | } |
1998 | |
1999 | /* Check if the union of the given pair of basic maps |
2000 | * can be represented by a single basic map. |
2001 | * If so, replace the pair by the single basic map and return |
2002 | * isl_change_drop_first, isl_change_drop_second or isl_change_fuse. |
2003 | * Otherwise, return isl_change_none. |
2004 | * The two basic maps are assumed to live in the same local space. |
2005 | * The "eq" and "ineq" fields of info[i] and info[j] are assumed |
2006 | * to have been initialized by the caller, either to NULL or |
2007 | * to valid information. |
2008 | * |
2009 | * We first check the effect of each constraint of one basic map |
2010 | * on the other basic map. |
2011 | * The constraint may be |
2012 | * redundant the constraint is redundant in its own |
2013 | * basic map and should be ignore and removed |
2014 | * in the end |
2015 | * valid all (integer) points of the other basic map |
2016 | * satisfy the constraint |
2017 | * separate no (integer) point of the other basic map |
2018 | * satisfies the constraint |
2019 | * cut some but not all points of the other basic map |
2020 | * satisfy the constraint |
2021 | * adj_eq the given constraint is adjacent (on the outside) |
2022 | * to an equality of the other basic map |
2023 | * adj_ineq the given constraint is adjacent (on the outside) |
2024 | * to an inequality of the other basic map |
2025 | * |
2026 | * We consider seven cases in which we can replace the pair by a single |
2027 | * basic map. We ignore all "redundant" constraints. |
2028 | * |
2029 | * 1. all constraints of one basic map are valid |
2030 | * => the other basic map is a subset and can be removed |
2031 | * |
2032 | * 2. all constraints of both basic maps are either "valid" or "cut" |
2033 | * and the facets corresponding to the "cut" constraints |
2034 | * of one of the basic maps lies entirely inside the other basic map |
2035 | * => the pair can be replaced by a basic map consisting |
2036 | * of the valid constraints in both basic maps |
2037 | * |
2038 | * 3. there is a single pair of adjacent inequalities |
2039 | * (all other constraints are "valid") |
2040 | * => the pair can be replaced by a basic map consisting |
2041 | * of the valid constraints in both basic maps |
2042 | * |
2043 | * 4. one basic map has a single adjacent inequality, while the other |
2044 | * constraints are "valid". The other basic map has some |
2045 | * "cut" constraints, but replacing the adjacent inequality by |
2046 | * its opposite and adding the valid constraints of the other |
2047 | * basic map results in a subset of the other basic map |
2048 | * => the pair can be replaced by a basic map consisting |
2049 | * of the valid constraints in both basic maps |
2050 | * |
2051 | * 5. there is a single adjacent pair of an inequality and an equality, |
2052 | * the other constraints of the basic map containing the inequality are |
2053 | * "valid". Moreover, if the inequality the basic map is relaxed |
2054 | * and then turned into an equality, then resulting facet lies |
2055 | * entirely inside the other basic map |
2056 | * => the pair can be replaced by the basic map containing |
2057 | * the inequality, with the inequality relaxed. |
2058 | * |
2059 | * 6. there is a single adjacent pair of an inequality and an equality, |
2060 | * the other constraints of the basic map containing the inequality are |
2061 | * "valid". Moreover, the facets corresponding to both |
2062 | * the inequality and the equality can be wrapped around their |
2063 | * ridges to include the other basic map |
2064 | * => the pair can be replaced by a basic map consisting |
2065 | * of the valid constraints in both basic maps together |
2066 | * with all wrapping constraints |
2067 | * |
2068 | * 7. one of the basic maps extends beyond the other by at most one. |
2069 | * Moreover, the facets corresponding to the cut constraints and |
2070 | * the pieces of the other basic map at offset one from these cut |
2071 | * constraints can be wrapped around their ridges to include |
2072 | * the union of the two basic maps |
2073 | * => the pair can be replaced by a basic map consisting |
2074 | * of the valid constraints in both basic maps together |
2075 | * with all wrapping constraints |
2076 | * |
2077 | * 8. the two basic maps live in adjacent hyperplanes. In principle |
2078 | * such sets can always be combined through wrapping, but we impose |
2079 | * that there is only one such pair, to avoid overeager coalescing. |
2080 | * |
2081 | * Throughout the computation, we maintain a collection of tableaus |
2082 | * corresponding to the basic maps. When the basic maps are dropped |
2083 | * or combined, the tableaus are modified accordingly. |
2084 | */ |
2085 | static enum isl_change coalesce_local_pair_reuse(int i, int j, |
2086 | struct isl_coalesce_info *info) |
2087 | { |
2088 | enum isl_change change = isl_change_none; |
2089 | |
2090 | set_eq_status_in(&info[i], info[j].tab); |
2091 | if (info[i].bmap->n_eq && !info[i].eq) |
2092 | goto error; |
2093 | if (any(info[i].eq, 2 * info[i].bmap->n_eq, STATUS_ERROR-1)) |
2094 | goto error; |
2095 | if (any(info[i].eq, 2 * info[i].bmap->n_eq, STATUS_SEPARATE3)) |
2096 | goto done; |
2097 | |
2098 | set_eq_status_in(&info[j], info[i].tab); |
2099 | if (info[j].bmap->n_eq && !info[j].eq) |
2100 | goto error; |
2101 | if (any(info[j].eq, 2 * info[j].bmap->n_eq, STATUS_ERROR-1)) |
2102 | goto error; |
2103 | if (any(info[j].eq, 2 * info[j].bmap->n_eq, STATUS_SEPARATE3)) |
2104 | goto done; |
2105 | |
2106 | set_ineq_status_in(&info[i], info[j].tab); |
2107 | if (info[i].bmap->n_ineq && !info[i].ineq) |
2108 | goto error; |
2109 | if (any(info[i].ineq, info[i].bmap->n_ineq, STATUS_ERROR-1)) |
2110 | goto error; |
2111 | if (any(info[i].ineq, info[i].bmap->n_ineq, STATUS_SEPARATE3)) |
2112 | goto done; |
2113 | |
2114 | set_ineq_status_in(&info[j], info[i].tab); |
2115 | if (info[j].bmap->n_ineq && !info[j].ineq) |
2116 | goto error; |
2117 | if (any(info[j].ineq, info[j].bmap->n_ineq, STATUS_ERROR-1)) |
2118 | goto error; |
2119 | if (any(info[j].ineq, info[j].bmap->n_ineq, STATUS_SEPARATE3)) |
2120 | goto done; |
2121 | |
2122 | if (all(info[i].eq, 2 * info[i].bmap->n_eq, STATUS_VALID2) && |
2123 | all(info[i].ineq, info[i].bmap->n_ineq, STATUS_VALID2)) { |
2124 | drop(&info[j]); |
2125 | change = isl_change_drop_second; |
2126 | } else if (all(info[j].eq, 2 * info[j].bmap->n_eq, STATUS_VALID2) && |
2127 | all(info[j].ineq, info[j].bmap->n_ineq, STATUS_VALID2)) { |
2128 | drop(&info[i]); |
2129 | change = isl_change_drop_first; |
2130 | } else if (any(info[i].eq, 2 * info[i].bmap->n_eq, STATUS_ADJ_EQ5)) { |
2131 | change = check_eq_adj_eq(i, j, info); |
2132 | } else if (any(info[j].eq, 2 * info[j].bmap->n_eq, STATUS_ADJ_EQ5)) { |
2133 | change = check_eq_adj_eq(j, i, info); |
2134 | } else if (any(info[i].eq, 2 * info[i].bmap->n_eq, STATUS_ADJ_INEQ6) || |
2135 | any(info[j].eq, 2 * info[j].bmap->n_eq, STATUS_ADJ_INEQ6)) { |
2136 | change = check_adj_eq(i, j, info); |
2137 | } else if (any(info[i].ineq, info[i].bmap->n_ineq, STATUS_ADJ_EQ5) || |
2138 | any(info[j].ineq, info[j].bmap->n_ineq, STATUS_ADJ_EQ5)) { |
2139 | /* Can't happen */ |
2140 | /* BAD ADJ INEQ */ |
2141 | } else if (any(info[i].ineq, info[i].bmap->n_ineq, STATUS_ADJ_INEQ6) || |
2142 | any(info[j].ineq, info[j].bmap->n_ineq, STATUS_ADJ_INEQ6)) { |
2143 | change = check_adj_ineq(i, j, info); |
2144 | } else { |
2145 | if (!any(info[i].eq, 2 * info[i].bmap->n_eq, STATUS_CUT4) && |
2146 | !any(info[j].eq, 2 * info[j].bmap->n_eq, STATUS_CUT4)) |
2147 | change = check_facets(i, j, info); |
2148 | if (change == isl_change_none) |
2149 | change = check_wrap(i, j, info); |
2150 | } |
2151 | |
2152 | done: |
2153 | clear_status(&info[i]); |
2154 | clear_status(&info[j]); |
2155 | return change; |
2156 | error: |
2157 | clear_status(&info[i]); |
2158 | clear_status(&info[j]); |
2159 | return isl_change_error; |
2160 | } |
2161 | |
2162 | /* Check if the union of the given pair of basic maps |
2163 | * can be represented by a single basic map. |
2164 | * If so, replace the pair by the single basic map and return |
2165 | * isl_change_drop_first, isl_change_drop_second or isl_change_fuse. |
2166 | * Otherwise, return isl_change_none. |
2167 | * The two basic maps are assumed to live in the same local space. |
2168 | */ |
2169 | static enum isl_change coalesce_local_pair(int i, int j, |
2170 | struct isl_coalesce_info *info) |
2171 | { |
2172 | init_status(&info[i]); |
2173 | init_status(&info[j]); |
2174 | return coalesce_local_pair_reuse(i, j, info); |
2175 | } |
2176 | |
2177 | /* Shift the integer division at position "div" of the basic map |
2178 | * represented by "info" by "shift". |
2179 | * |
2180 | * That is, if the integer division has the form |
2181 | * |
2182 | * floor(f(x)/d) |
2183 | * |
2184 | * then replace it by |
2185 | * |
2186 | * floor((f(x) + shift * d)/d) - shift |
2187 | */ |
2188 | static isl_stat shift_div(struct isl_coalesce_info *info, int div, |
2189 | isl_int shift) |
2190 | { |
2191 | unsigned total; |
2192 | |
2193 | info->bmap = isl_basic_map_shift_div(info->bmap, div, 0, shift); |
2194 | if (!info->bmap) |
2195 | return isl_stat_error; |
2196 | |
2197 | total = isl_basic_map_dim(info->bmap, isl_dim_all); |
2198 | total -= isl_basic_map_dim(info->bmap, isl_dim_div); |
2199 | if (isl_tab_shift_var(info->tab, total + div, shift) < 0) |
2200 | return isl_stat_error; |
2201 | |
2202 | return isl_stat_ok; |
2203 | } |
2204 | |
2205 | /* If the integer division at position "div" is defined by an equality, |
2206 | * i.e., a stride constraint, then change the integer division expression |
2207 | * to have a constant term equal to zero. |
2208 | * |
2209 | * Let the equality constraint be |
2210 | * |
2211 | * c + f + m a = 0 |
2212 | * |
2213 | * The integer division expression is then of the form |
2214 | * |
2215 | * a = floor((-f - c')/m) |
2216 | * |
2217 | * The integer division is first shifted by t = floor(c/m), |
2218 | * turning the equality constraint into |
2219 | * |
2220 | * c - m floor(c/m) + f + m a' = 0 |
2221 | * |
2222 | * i.e., |
2223 | * |
2224 | * (c mod m) + f + m a' = 0 |
2225 | * |
2226 | * That is, |
2227 | * |
2228 | * a' = (-f - (c mod m))/m = floor((-f)/m) |
2229 | * |
2230 | * because a' is an integer and 0 <= (c mod m) < m. |
2231 | * The constant term of a' can therefore be zeroed out. |
2232 | */ |
2233 | static isl_stat normalize_stride_div(struct isl_coalesce_info *info, int div) |
2234 | { |
2235 | isl_bool defined; |
2236 | isl_stat r; |
2237 | isl_constraint *c; |
2238 | isl_int shift, stride; |
2239 | |
2240 | defined = isl_basic_map_has_defining_equality(info->bmap, isl_dim_div, |
2241 | div, &c); |
2242 | if (defined < 0) |
2243 | return isl_stat_error; |
2244 | if (!defined) |
2245 | return isl_stat_ok; |
2246 | if (!c) |
2247 | return isl_stat_error; |
2248 | isl_int_init(shift)isl_sioimath_init((shift)); |
2249 | isl_int_init(stride)isl_sioimath_init((stride)); |
2250 | isl_constraint_get_constant(c, &shift); |
2251 | isl_constraint_get_coefficient(c, isl_dim_div, div, &stride); |
2252 | isl_int_fdiv_q(shift, shift, stride)isl_sioimath_fdiv_q((shift), *(shift), *(stride)); |
2253 | r = shift_div(info, div, shift); |
2254 | isl_int_clear(stride)isl_sioimath_clear((stride)); |
2255 | isl_int_clear(shift)isl_sioimath_clear((shift)); |
2256 | isl_constraint_free(c); |
2257 | if (r < 0) |
2258 | return isl_stat_error; |
2259 | info->bmap = isl_basic_map_set_div_expr_constant_num_si_inplace( |
2260 | info->bmap, div, 0); |
2261 | if (!info->bmap) |
2262 | return isl_stat_error; |
2263 | return isl_stat_ok; |
2264 | } |
2265 | |
2266 | /* The basic maps represented by "info1" and "info2" are known |
2267 | * to have the same number of integer divisions. |
2268 | * Check if pairs of integer divisions are equal to each other |
2269 | * despite the fact that they differ by a rational constant. |
2270 | * |
2271 | * In particular, look for any pair of integer divisions that |
2272 | * only differ in their constant terms. |
2273 | * If either of these integer divisions is defined |
2274 | * by stride constraints, then modify it to have a zero constant term. |
2275 | * If both are defined by stride constraints then in the end they will have |
2276 | * the same (zero) constant term. |
2277 | */ |
2278 | static isl_stat harmonize_stride_divs(struct isl_coalesce_info *info1, |
2279 | struct isl_coalesce_info *info2) |
2280 | { |
2281 | int i, n; |
2282 | int total; |
2283 | |
2284 | total = isl_basic_map_total_dim(info1->bmap); |
Value stored to 'total' is never read | |
2285 | n = isl_basic_map_dim(info1->bmap, isl_dim_div); |
2286 | for (i = 0; i < n; ++i) { |
2287 | isl_bool known, harmonize; |
2288 | |
2289 | known = isl_basic_map_div_is_known(info1->bmap, i); |
2290 | if (known >= 0 && known) |
2291 | known = isl_basic_map_div_is_known(info2->bmap, i); |
2292 | if (known < 0) |
2293 | return isl_stat_error; |
2294 | if (!known) |
2295 | continue; |
2296 | harmonize = isl_basic_map_equal_div_expr_except_constant( |
2297 | info1->bmap, i, info2->bmap, i); |
2298 | if (harmonize < 0) |
2299 | return isl_stat_error; |
2300 | if (!harmonize) |
2301 | continue; |
2302 | if (normalize_stride_div(info1, i) < 0) |
2303 | return isl_stat_error; |
2304 | if (normalize_stride_div(info2, i) < 0) |
2305 | return isl_stat_error; |
2306 | } |
2307 | |
2308 | return isl_stat_ok; |
2309 | } |
2310 | |
2311 | /* If "shift" is an integer constant, then shift the integer division |
2312 | * at position "div" of the basic map represented by "info" by "shift". |
2313 | * If "shift" is not an integer constant, then do nothing. |
2314 | * If "shift" is equal to zero, then no shift needs to be performed either. |
2315 | * |
2316 | * That is, if the integer division has the form |
2317 | * |
2318 | * floor(f(x)/d) |
2319 | * |
2320 | * then replace it by |
2321 | * |
2322 | * floor((f(x) + shift * d)/d) - shift |
2323 | */ |
2324 | static isl_stat shift_if_cst_int(struct isl_coalesce_info *info, int div, |
2325 | __isl_keep isl_aff *shift) |
2326 | { |
2327 | isl_bool cst; |
2328 | isl_stat r; |
2329 | isl_int d; |
2330 | isl_val *c; |
2331 | |
2332 | cst = isl_aff_is_cst(shift); |
2333 | if (cst < 0 || !cst) |
2334 | return cst < 0 ? isl_stat_error : isl_stat_ok; |
2335 | |
2336 | c = isl_aff_get_constant_val(shift); |
2337 | cst = isl_val_is_int(c); |
2338 | if (cst >= 0 && cst) |
2339 | cst = isl_bool_not(isl_val_is_zero(c)); |
2340 | if (cst < 0 || !cst) { |
2341 | isl_val_free(c); |
2342 | return cst < 0 ? isl_stat_error : isl_stat_ok; |
2343 | } |
2344 | |
2345 | isl_int_init(d)isl_sioimath_init((d)); |
2346 | r = isl_val_get_num_isl_int(c, &d); |
2347 | if (r >= 0) |
2348 | r = shift_div(info, div, d); |
2349 | isl_int_clear(d)isl_sioimath_clear((d)); |
2350 | |
2351 | isl_val_free(c); |
2352 | |
2353 | return r; |
2354 | } |
2355 | |
2356 | /* Check if some of the divs in the basic map represented by "info1" |
2357 | * are shifts of the corresponding divs in the basic map represented |
2358 | * by "info2", taking into account the equality constraints "eq1" of "info1" |
2359 | * and "eq2" of "info2". If so, align them with those of "info2". |
2360 | * "info1" and "info2" are assumed to have the same number |
2361 | * of integer divisions. |
2362 | * |
2363 | * An integer division is considered to be a shift of another integer |
2364 | * division if, after simplification with respect to the equality |
2365 | * constraints of the other basic map, one is equal to the other |
2366 | * plus a constant. |
2367 | * |
2368 | * In particular, for each pair of integer divisions, if both are known, |
2369 | * have the same denominator and are not already equal to each other, |
2370 | * simplify each with respect to the equality constraints |
2371 | * of the other basic map. If the difference is an integer constant, |
2372 | * then move this difference outside. |
2373 | * That is, if, after simplification, one integer division is of the form |
2374 | * |
2375 | * floor((f(x) + c_1)/d) |
2376 | * |
2377 | * while the other is of the form |
2378 | * |
2379 | * floor((f(x) + c_2)/d) |
2380 | * |
2381 | * and n = (c_2 - c_1)/d is an integer, then replace the first |
2382 | * integer division by |
2383 | * |
2384 | * floor((f_1(x) + c_1 + n * d)/d) - n, |
2385 | * |
2386 | * where floor((f_1(x) + c_1 + n * d)/d) = floor((f2(x) + c_2)/d) |
2387 | * after simplification with respect to the equality constraints. |
2388 | */ |
2389 | static isl_stat harmonize_divs_with_hulls(struct isl_coalesce_info *info1, |
2390 | struct isl_coalesce_info *info2, __isl_keep isl_basic_setisl_basic_map *eq1, |
2391 | __isl_keep isl_basic_setisl_basic_map *eq2) |
2392 | { |
2393 | int i; |
2394 | int total; |
2395 | isl_local_space *ls1, *ls2; |
2396 | |
2397 | total = isl_basic_map_total_dim(info1->bmap); |
2398 | ls1 = isl_local_space_wrap(isl_basic_map_get_local_space(info1->bmap)); |
2399 | ls2 = isl_local_space_wrap(isl_basic_map_get_local_space(info2->bmap)); |
2400 | for (i = 0; i < info1->bmap->n_div; ++i) { |
2401 | isl_stat r; |
2402 | isl_aff *div1, *div2; |
2403 | |
2404 | if (!isl_local_space_div_is_known(ls1, i) || |
2405 | !isl_local_space_div_is_known(ls2, i)) |
2406 | continue; |
2407 | if (isl_int_ne(info1->bmap->div[i][0], info2->bmap->div[i][0])(isl_sioimath_cmp(*(info1->bmap->div[i][0]), *(info2-> bmap->div[i][0])) != 0)) |
2408 | continue; |
2409 | if (isl_seq_eq(info1->bmap->div[i] + 1, |
2410 | info2->bmap->div[i] + 1, 1 + total)) |
2411 | continue; |
2412 | div1 = isl_local_space_get_div(ls1, i); |
2413 | div2 = isl_local_space_get_div(ls2, i); |
2414 | div1 = isl_aff_substitute_equalities(div1, |
2415 | isl_basic_set_copy(eq2)); |
2416 | div2 = isl_aff_substitute_equalities(div2, |
2417 | isl_basic_set_copy(eq1)); |
2418 | div2 = isl_aff_sub(div2, div1); |
2419 | r = shift_if_cst_int(info1, i, div2); |
2420 | isl_aff_free(div2); |
2421 | if (r < 0) |
2422 | break; |
2423 | } |
2424 | isl_local_space_free(ls1); |
2425 | isl_local_space_free(ls2); |
2426 | |
2427 | if (i < info1->bmap->n_div) |
2428 | return isl_stat_error; |
2429 | return isl_stat_ok; |
2430 | } |
2431 | |
2432 | /* Check if some of the divs in the basic map represented by "info1" |
2433 | * are shifts of the corresponding divs in the basic map represented |
2434 | * by "info2". If so, align them with those of "info2". |
2435 | * Only do this if "info1" and "info2" have the same number |
2436 | * of integer divisions. |
2437 | * |
2438 | * An integer division is considered to be a shift of another integer |
2439 | * division if, after simplification with respect to the equality |
2440 | * constraints of the other basic map, one is equal to the other |
2441 | * plus a constant. |
2442 | * |
2443 | * First check if pairs of integer divisions are equal to each other |
2444 | * despite the fact that they differ by a rational constant. |
2445 | * If so, try and arrange for them to have the same constant term. |
2446 | * |
2447 | * Then, extract the equality constraints and continue with |
2448 | * harmonize_divs_with_hulls. |
2449 | */ |
2450 | static isl_stat harmonize_divs(struct isl_coalesce_info *info1, |
2451 | struct isl_coalesce_info *info2) |
2452 | { |
2453 | isl_basic_map *bmap1, *bmap2; |
2454 | isl_basic_setisl_basic_map *eq1, *eq2; |
2455 | isl_stat r; |
2456 | |
2457 | if (!info1->bmap || !info2->bmap) |
2458 | return isl_stat_error; |
2459 | |
2460 | if (info1->bmap->n_div != info2->bmap->n_div) |
2461 | return isl_stat_ok; |
2462 | if (info1->bmap->n_div == 0) |
2463 | return isl_stat_ok; |
2464 | |
2465 | if (harmonize_stride_divs(info1, info2) < 0) |
2466 | return isl_stat_error; |
2467 | |
2468 | bmap1 = isl_basic_map_copy(info1->bmap); |
2469 | bmap2 = isl_basic_map_copy(info2->bmap); |
2470 | eq1 = isl_basic_map_wrap(isl_basic_map_plain_affine_hull(bmap1)); |
2471 | eq2 = isl_basic_map_wrap(isl_basic_map_plain_affine_hull(bmap2)); |
2472 | r = harmonize_divs_with_hulls(info1, info2, eq1, eq2); |
2473 | isl_basic_set_free(eq1); |
2474 | isl_basic_set_free(eq2); |
2475 | |
2476 | return r; |
2477 | } |
2478 | |
2479 | /* Do the two basic maps live in the same local space, i.e., |
2480 | * do they have the same (known) divs? |
2481 | * If either basic map has any unknown divs, then we can only assume |
2482 | * that they do not live in the same local space. |
2483 | */ |
2484 | static int same_divs(__isl_keep isl_basic_map *bmap1, |
2485 | __isl_keep isl_basic_map *bmap2) |
2486 | { |
2487 | int i; |
2488 | int known; |
2489 | int total; |
2490 | |
2491 | if (!bmap1 || !bmap2) |
2492 | return -1; |
2493 | if (bmap1->n_div != bmap2->n_div) |
2494 | return 0; |
2495 | |
2496 | if (bmap1->n_div == 0) |
2497 | return 1; |
2498 | |
2499 | known = isl_basic_map_divs_known(bmap1); |
2500 | if (known < 0 || !known) |
2501 | return known; |
2502 | known = isl_basic_map_divs_known(bmap2); |
2503 | if (known < 0 || !known) |
2504 | return known; |
2505 | |
2506 | total = isl_basic_map_total_dim(bmap1); |
2507 | for (i = 0; i < bmap1->n_div; ++i) |
2508 | if (!isl_seq_eq(bmap1->div[i], bmap2->div[i], 2 + total)) |
2509 | return 0; |
2510 | |
2511 | return 1; |
2512 | } |
2513 | |
2514 | /* Assuming that "tab" contains the equality constraints and |
2515 | * the initial inequality constraints of "bmap", copy the remaining |
2516 | * inequality constraints of "bmap" to "Tab". |
2517 | */ |
2518 | static isl_stat copy_ineq(struct isl_tab *tab, __isl_keep isl_basic_map *bmap) |
2519 | { |
2520 | int i, n_ineq; |
2521 | |
2522 | if (!bmap) |
2523 | return isl_stat_error; |
2524 | |
2525 | n_ineq = tab->n_con - tab->n_eq; |
2526 | for (i = n_ineq; i < bmap->n_ineq; ++i) |
2527 | if (isl_tab_add_ineq(tab, bmap->ineq[i]) < 0) |
2528 | return isl_stat_error; |
2529 | |
2530 | return isl_stat_ok; |
2531 | } |
2532 | |
2533 | /* Description of an integer division that is added |
2534 | * during an expansion. |
2535 | * "pos" is the position of the corresponding variable. |
2536 | * "cst" indicates whether this integer division has a fixed value. |
2537 | * "val" contains the fixed value, if the value is fixed. |
2538 | */ |
2539 | struct isl_expanded { |
2540 | int pos; |
2541 | isl_bool cst; |
2542 | isl_int val; |
2543 | }; |
2544 | |
2545 | /* For each of the "n" integer division variables "expanded", |
2546 | * if the variable has a fixed value, then add two inequality |
2547 | * constraints expressing the fixed value. |
2548 | * Otherwise, add the corresponding div constraints. |
2549 | * The caller is responsible for removing the div constraints |
2550 | * that it added for all these "n" integer divisions. |
2551 | * |
2552 | * The div constraints and the pair of inequality constraints |
2553 | * forcing the fixed value cannot both be added for a given variable |
2554 | * as the combination may render some of the original constraints redundant. |
2555 | * These would then be ignored during the coalescing detection, |
2556 | * while they could remain in the fused result. |
2557 | * |
2558 | * The two added inequality constraints are |
2559 | * |
2560 | * -a + v >= 0 |
2561 | * a - v >= 0 |
2562 | * |
2563 | * with "a" the variable and "v" its fixed value. |
2564 | * The facet corresponding to one of these two constraints is selected |
2565 | * in the tableau to ensure that the pair of inequality constraints |
2566 | * is treated as an equality constraint. |
2567 | * |
2568 | * The information in info->ineq is thrown away because it was |
2569 | * computed in terms of div constraints, while some of those |
2570 | * have now been replaced by these pairs of inequality constraints. |
2571 | */ |
2572 | static isl_stat fix_constant_divs(struct isl_coalesce_info *info, |
2573 | int n, struct isl_expanded *expanded) |
2574 | { |
2575 | unsigned o_div; |
2576 | int i; |
2577 | isl_vec *ineq; |
2578 | |
2579 | o_div = isl_basic_map_offset(info->bmap, isl_dim_div) - 1; |
2580 | ineq = isl_vec_alloc(isl_tab_get_ctx(info->tab), 1 + info->tab->n_var); |
2581 | if (!ineq) |
2582 | return isl_stat_error; |
2583 | isl_seq_clr(ineq->el + 1, info->tab->n_var); |
2584 | |
2585 | for (i = 0; i < n; ++i) { |
2586 | if (!expanded[i].cst) { |
2587 | info->bmap = isl_basic_map_extend_constraints( |
2588 | info->bmap, 0, 2); |
2589 | if (isl_basic_map_add_div_constraints(info->bmap, |
2590 | expanded[i].pos - o_div) < 0) |
2591 | break; |
2592 | } else { |
2593 | isl_int_set_si(ineq->el[1 + expanded[i].pos], -1)isl_sioimath_set_si((ineq->el[1 + expanded[i].pos]), -1); |
2594 | isl_int_set(ineq->el[0], expanded[i].val)isl_sioimath_set((ineq->el[0]), *(expanded[i].val)); |
2595 | info->bmap = isl_basic_map_add_ineq(info->bmap, |
2596 | ineq->el); |
2597 | isl_int_set_si(ineq->el[1 + expanded[i].pos], 1)isl_sioimath_set_si((ineq->el[1 + expanded[i].pos]), 1); |
2598 | isl_int_neg(ineq->el[0], expanded[i].val)isl_sioimath_neg((ineq->el[0]), *(expanded[i].val)); |
2599 | info->bmap = isl_basic_map_add_ineq(info->bmap, |
2600 | ineq->el); |
2601 | isl_int_set_si(ineq->el[1 + expanded[i].pos], 0)isl_sioimath_set_si((ineq->el[1 + expanded[i].pos]), 0); |
2602 | } |
2603 | if (copy_ineq(info->tab, info->bmap) < 0) |
2604 | break; |
2605 | if (expanded[i].cst && |
2606 | isl_tab_select_facet(info->tab, info->tab->n_con - 1) < 0) |
2607 | break; |
2608 | } |
2609 | |
2610 | isl_vec_free(ineq); |
2611 | |
2612 | clear_status(info); |
2613 | init_status(info); |
2614 | |
2615 | return i < n ? isl_stat_error : isl_stat_ok; |
2616 | } |
2617 | |
2618 | /* Insert the "n" integer division variables "expanded" |
2619 | * into info->tab and info->bmap and |
2620 | * update info->ineq with respect to the redundant constraints |
2621 | * in the resulting tableau. |
2622 | * "bmap" contains the result of this insertion in info->bmap, |
2623 | * while info->bmap is the original version |
2624 | * of "bmap", i.e., the one that corresponds to the current |
2625 | * state of info->tab. The number of constraints in info->bmap |
2626 | * is assumed to be the same as the number of constraints |
2627 | * in info->tab. This is required to be able to detect |
2628 | * the extra constraints in "bmap". |
2629 | * |
2630 | * In particular, introduce extra variables corresponding |
2631 | * to the extra integer divisions and add the div constraints |
2632 | * that were added to "bmap" after info->tab was created |
2633 | * from info->bmap. |
2634 | * Furthermore, check if these extra integer divisions happen |
2635 | * to attain a fixed integer value in info->tab. |
2636 | * If so, replace the corresponding div constraints by pairs |
2637 | * of inequality constraints that fix these |
2638 | * integer divisions to their single integer values. |
2639 | * Replace info->bmap by "bmap" to match the changes to info->tab. |
2640 | * info->ineq was computed without a tableau and therefore |
2641 | * does not take into account the redundant constraints |
2642 | * in the tableau. Mark them here. |
2643 | * There is no need to check the newly added div constraints |
2644 | * since they cannot be redundant. |
2645 | * The redundancy check is not performed when constants have been discovered |
2646 | * since info->ineq is completely thrown away in this case. |
2647 | */ |
2648 | static isl_stat tab_insert_divs(struct isl_coalesce_info *info, |
2649 | int n, struct isl_expanded *expanded, __isl_take isl_basic_map *bmap) |
2650 | { |
2651 | int i, n_ineq; |
2652 | unsigned n_eq; |
2653 | struct isl_tab_undo *snap; |
2654 | int any; |
2655 | |
2656 | if (!bmap) |
2657 | return isl_stat_error; |
2658 | if (info->bmap->n_eq + info->bmap->n_ineq != info->tab->n_con) |
2659 | isl_die(isl_basic_map_get_ctx(bmap), isl_error_internal,do { isl_handle_error(isl_basic_map_get_ctx(bmap), isl_error_internal , "original tableau does not correspond " "to original basic map" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn295388/tools/polly/lib/External/isl/isl_coalesce.c" , 2661); goto error; } while (0) |
2660 | "original tableau does not correspond "do { isl_handle_error(isl_basic_map_get_ctx(bmap), isl_error_internal , "original tableau does not correspond " "to original basic map" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn295388/tools/polly/lib/External/isl/isl_coalesce.c" , 2661); goto error; } while (0) |
2661 | "to original basic map", goto error)do { isl_handle_error(isl_basic_map_get_ctx(bmap), isl_error_internal , "original tableau does not correspond " "to original basic map" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn295388/tools/polly/lib/External/isl/isl_coalesce.c" , 2661); goto error; } while (0); |
2662 | |
2663 | if (isl_tab_extend_vars(info->tab, n) < 0) |
2664 | goto error; |
2665 | if (isl_tab_extend_cons(info->tab, 2 * n) < 0) |
2666 | goto error; |
2667 | |
2668 | for (i = 0; i < n; ++i) { |
2669 | if (isl_tab_insert_var(info->tab, expanded[i].pos) < 0) |
2670 | goto error; |
2671 | } |
2672 | |
2673 | snap = isl_tab_snap(info->tab); |
2674 | |
2675 | n_ineq = info->tab->n_con - info->tab->n_eq; |
2676 | if (copy_ineq(info->tab, bmap) < 0) |
2677 | goto error; |
2678 | |
2679 | isl_basic_map_free(info->bmap); |
2680 | info->bmap = bmap; |
2681 | |
2682 | any = 0; |
2683 | for (i = 0; i < n; ++i) { |
2684 | expanded[i].cst = isl_tab_is_constant(info->tab, |
2685 | expanded[i].pos, &expanded[i].val); |
2686 | if (expanded[i].cst < 0) |
2687 | return isl_stat_error; |
2688 | if (expanded[i].cst) |
2689 | any = 1; |
2690 | } |
2691 | |
2692 | if (any) { |
2693 | if (isl_tab_rollback(info->tab, snap) < 0) |
2694 | return isl_stat_error; |
2695 | info->bmap = isl_basic_map_cow(info->bmap); |
2696 | if (isl_basic_map_free_inequality(info->bmap, 2 * n) < 0) |
2697 | return isl_stat_error; |
2698 | |
2699 | return fix_constant_divs(info, n, expanded); |
2700 | } |
2701 | |
2702 | n_eq = info->bmap->n_eq; |
2703 | for (i = 0; i < n_ineq; ++i) { |
2704 | if (isl_tab_is_redundant(info->tab, n_eq + i)) |
2705 | info->ineq[i] = STATUS_REDUNDANT1; |
2706 | } |
2707 | |
2708 | return isl_stat_ok; |
2709 | error: |
2710 | isl_basic_map_free(bmap); |
2711 | return isl_stat_error; |
2712 | } |
2713 | |
2714 | /* Expand info->tab and info->bmap in the same way "bmap" was expanded |
2715 | * in isl_basic_map_expand_divs using the expansion "exp" and |
2716 | * update info->ineq with respect to the redundant constraints |
2717 | * in the resulting tableau. info->bmap is the original version |
2718 | * of "bmap", i.e., the one that corresponds to the current |
2719 | * state of info->tab. The number of constraints in info->bmap |
2720 | * is assumed to be the same as the number of constraints |
2721 | * in info->tab. This is required to be able to detect |
2722 | * the extra constraints in "bmap". |
2723 | * |
2724 | * Extract the positions where extra local variables are introduced |
2725 | * from "exp" and call tab_insert_divs. |
2726 | */ |
2727 | static isl_stat expand_tab(struct isl_coalesce_info *info, int *exp, |
2728 | __isl_take isl_basic_map *bmap) |
2729 | { |
2730 | isl_ctx *ctx; |
2731 | struct isl_expanded *expanded; |
2732 | int i, j, k, n; |
2733 | int extra_var; |
2734 | unsigned total, pos, n_div; |
2735 | isl_stat r; |
2736 | |
2737 | total = isl_basic_map_dim(bmap, isl_dim_all); |
2738 | n_div = isl_basic_map_dim(bmap, isl_dim_div); |
2739 | pos = total - n_div; |
2740 | extra_var = total - info->tab->n_var; |
2741 | n = n_div - extra_var; |
2742 | |
2743 | ctx = isl_basic_map_get_ctx(bmap); |
2744 | expanded = isl_calloc_array(ctx, struct isl_expanded, extra_var)((struct isl_expanded *)isl_calloc_or_die(ctx, extra_var, sizeof (struct isl_expanded))); |
2745 | if (extra_var && !expanded) |
2746 | goto error; |
2747 | |
2748 | i = 0; |
2749 | k = 0; |
2750 | for (j = 0; j < n_div; ++j) { |
2751 | if (i < n && exp[i] == j) { |
2752 | ++i; |
2753 | continue; |
2754 | } |
2755 | expanded[k++].pos = pos + j; |
2756 | } |
2757 | |
2758 | for (k = 0; k < extra_var; ++k) |
2759 | isl_int_init(expanded[k].val)isl_sioimath_init((expanded[k].val)); |
2760 | |
2761 | r = tab_insert_divs(info, extra_var, expanded, bmap); |
2762 | |
2763 | for (k = 0; k < extra_var; ++k) |
2764 | isl_int_clear(expanded[k].val)isl_sioimath_clear((expanded[k].val)); |
2765 | free(expanded); |
2766 | |
2767 | return r; |
2768 | error: |
2769 | isl_basic_map_free(bmap); |
2770 | return isl_stat_error; |
2771 | } |
2772 | |
2773 | /* Check if the union of the basic maps represented by info[i] and info[j] |
2774 | * can be represented by a single basic map, |
2775 | * after expanding the divs of info[i] to match those of info[j]. |
2776 | * If so, replace the pair by the single basic map and return |
2777 | * isl_change_drop_first, isl_change_drop_second or isl_change_fuse. |
2778 | * Otherwise, return isl_change_none. |
2779 | * |
2780 | * The caller has already checked for info[j] being a subset of info[i]. |
2781 | * If some of the divs of info[j] are unknown, then the expanded info[i] |
2782 | * will not have the corresponding div constraints. The other patterns |
2783 | * therefore cannot apply. Skip the computation in this case. |
2784 | * |
2785 | * The expansion is performed using the divs "div" and expansion "exp" |
2786 | * computed by the caller. |
2787 | * info[i].bmap has already been expanded and the result is passed in |
2788 | * as "bmap". |
2789 | * The "eq" and "ineq" fields of info[i] reflect the status of |
2790 | * the constraints of the expanded "bmap" with respect to info[j].tab. |
2791 | * However, inequality constraints that are redundant in info[i].tab |
2792 | * have not yet been marked as such because no tableau was available. |
2793 | * |
2794 | * Replace info[i].bmap by "bmap" and expand info[i].tab as well, |
2795 | * updating info[i].ineq with respect to the redundant constraints. |
2796 | * Then try and coalesce the expanded info[i] with info[j], |
2797 | * reusing the information in info[i].eq and info[i].ineq. |
2798 | * If this does not result in any coalescing or if it results in info[j] |
2799 | * getting dropped (which should not happen in practice, since the case |
2800 | * of info[j] being a subset of info[i] has already been checked by |
2801 | * the caller), then revert info[i] to its original state. |
2802 | */ |
2803 | static enum isl_change coalesce_expand_tab_divs(__isl_take isl_basic_map *bmap, |
2804 | int i, int j, struct isl_coalesce_info *info, __isl_keep isl_mat *div, |
2805 | int *exp) |
2806 | { |
2807 | isl_bool known; |
2808 | isl_basic_map *bmap_i; |
2809 | struct isl_tab_undo *snap; |
2810 | enum isl_change change = isl_change_none; |
2811 | |
2812 | known = isl_basic_map_divs_known(info[j].bmap); |
2813 | if (known < 0 || !known) { |
2814 | clear_status(&info[i]); |
2815 | isl_basic_map_free(bmap); |
2816 | return known < 0 ? isl_change_error : isl_change_none; |
2817 | } |
2818 | |
2819 | bmap_i = isl_basic_map_copy(info[i].bmap); |
2820 | snap = isl_tab_snap(info[i].tab); |
2821 | if (expand_tab(&info[i], exp, bmap) < 0) |
2822 | change = isl_change_error; |
2823 | |
2824 | init_status(&info[j]); |
2825 | if (change == isl_change_none) |
2826 | change = coalesce_local_pair_reuse(i, j, info); |
2827 | else |
2828 | clear_status(&info[i]); |
2829 | if (change != isl_change_none && change != isl_change_drop_second) { |
2830 | isl_basic_map_free(bmap_i); |
2831 | } else { |
2832 | isl_basic_map_free(info[i].bmap); |
2833 | info[i].bmap = bmap_i; |
2834 | |
2835 | if (isl_tab_rollback(info[i].tab, snap) < 0) |
2836 | change = isl_change_error; |
2837 | } |
2838 | |
2839 | return change; |
2840 | } |
2841 | |
2842 | /* Check if the union of "bmap" and the basic map represented by info[j] |
2843 | * can be represented by a single basic map, |
2844 | * after expanding the divs of "bmap" to match those of info[j]. |
2845 | * If so, replace the pair by the single basic map and return |
2846 | * isl_change_drop_first, isl_change_drop_second or isl_change_fuse. |
2847 | * Otherwise, return isl_change_none. |
2848 | * |
2849 | * In particular, check if the expanded "bmap" contains the basic map |
2850 | * represented by the tableau info[j].tab. |
2851 | * The expansion is performed using the divs "div" and expansion "exp" |
2852 | * computed by the caller. |
2853 | * Then we check if all constraints of the expanded "bmap" are valid for |
2854 | * info[j].tab. |
2855 | * |
2856 | * If "i" is not equal to -1, then "bmap" is equal to info[i].bmap. |
2857 | * In this case, the positions of the constraints of info[i].bmap |
2858 | * with respect to the basic map represented by info[j] are stored |
2859 | * in info[i]. |
2860 | * |
2861 | * If the expanded "bmap" does not contain the basic map |
2862 | * represented by the tableau info[j].tab and if "i" is not -1, |
2863 | * i.e., if the original "bmap" is info[i].bmap, then expand info[i].tab |
2864 | * as well and check if that results in coalescing. |
2865 | */ |
2866 | static enum isl_change coalesce_with_expanded_divs( |
2867 | __isl_keep isl_basic_map *bmap, int i, int j, |
2868 | struct isl_coalesce_info *info, __isl_keep isl_mat *div, int *exp) |
2869 | { |
2870 | enum isl_change change = isl_change_none; |
2871 | struct isl_coalesce_info info_local, *info_i; |
2872 | |
2873 | info_i = i >= 0 ? &info[i] : &info_local; |
2874 | init_status(info_i); |
2875 | bmap = isl_basic_map_copy(bmap); |
2876 | bmap = isl_basic_map_expand_divs(bmap, isl_mat_copy(div), exp); |
2877 | bmap = isl_basic_map_mark_final(bmap); |
2878 | |
2879 | if (!bmap) |
2880 | goto error; |
2881 | |
2882 | info_i->eq = eq_status_in(bmap, info[j].tab); |
2883 | if (bmap->n_eq && !info_i->eq) |
2884 | goto error; |
2885 | if (any(info_i->eq, 2 * bmap->n_eq, STATUS_ERROR-1)) |
2886 | goto error; |
2887 | if (any(info_i->eq, 2 * bmap->n_eq, STATUS_SEPARATE3)) |
2888 | goto done; |
2889 | |
2890 | info_i->ineq = ineq_status_in(bmap, NULL((void*)0), info[j].tab); |
2891 | if (bmap->n_ineq && !info_i->ineq) |
2892 | goto error; |
2893 | if (any(info_i->ineq, bmap->n_ineq, STATUS_ERROR-1)) |
2894 | goto error; |
2895 | if (any(info_i->ineq, bmap->n_ineq, STATUS_SEPARATE3)) |
2896 | goto done; |
2897 | |
2898 | if (all(info_i->eq, 2 * bmap->n_eq, STATUS_VALID2) && |
2899 | all(info_i->ineq, bmap->n_ineq, STATUS_VALID2)) { |
2900 | drop(&info[j]); |
2901 | change = isl_change_drop_second; |
2902 | } |
2903 | |
2904 | if (change == isl_change_none && i != -1) |
2905 | return coalesce_expand_tab_divs(bmap, i, j, info, div, exp); |
2906 | |
2907 | done: |
2908 | isl_basic_map_free(bmap); |
2909 | clear_status(info_i); |
2910 | return change; |
2911 | error: |
2912 | isl_basic_map_free(bmap); |
2913 | clear_status(info_i); |
2914 | return isl_change_error; |
2915 | } |
2916 | |
2917 | /* Check if the union of "bmap_i" and the basic map represented by info[j] |
2918 | * can be represented by a single basic map, |
2919 | * after aligning the divs of "bmap_i" to match those of info[j]. |
2920 | * If so, replace the pair by the single basic map and return |
2921 | * isl_change_drop_first, isl_change_drop_second or isl_change_fuse. |
2922 | * Otherwise, return isl_change_none. |
2923 | * |
2924 | * In particular, check if "bmap_i" contains the basic map represented by |
2925 | * info[j] after aligning the divs of "bmap_i" to those of info[j]. |
2926 | * Note that this can only succeed if the number of divs of "bmap_i" |
2927 | * is smaller than (or equal to) the number of divs of info[j]. |
2928 | * |
2929 | * We first check if the divs of "bmap_i" are all known and form a subset |
2930 | * of those of info[j].bmap. If so, we pass control over to |
2931 | * coalesce_with_expanded_divs. |
2932 | * |
2933 | * If "i" is not equal to -1, then "bmap" is equal to info[i].bmap. |
2934 | */ |
2935 | static enum isl_change coalesce_after_aligning_divs( |
2936 | __isl_keep isl_basic_map *bmap_i, int i, int j, |
2937 | struct isl_coalesce_info *info) |
2938 | { |
2939 | int known; |
2940 | isl_mat *div_i, *div_j, *div; |
2941 | int *exp1 = NULL((void*)0); |
2942 | int *exp2 = NULL((void*)0); |
2943 | isl_ctx *ctx; |
2944 | enum isl_change change; |
2945 | |
2946 | known = isl_basic_map_divs_known(bmap_i); |
2947 | if (known < 0 || !known) |
2948 | return known; |
2949 | |
2950 | ctx = isl_basic_map_get_ctx(bmap_i); |
2951 | |
2952 | div_i = isl_basic_map_get_divs(bmap_i); |
2953 | div_j = isl_basic_map_get_divs(info[j].bmap); |
2954 | |
2955 | if (!div_i || !div_j) |
2956 | goto error; |
2957 | |
2958 | exp1 = isl_alloc_array(ctx, int, div_i->n_row)((int *)isl_malloc_or_die(ctx, (div_i->n_row)*sizeof(int)) ); |
2959 | exp2 = isl_alloc_array(ctx, int, div_j->n_row)((int *)isl_malloc_or_die(ctx, (div_j->n_row)*sizeof(int)) ); |
2960 | if ((div_i->n_row && !exp1) || (div_j->n_row && !exp2)) |
2961 | goto error; |
2962 | |
2963 | div = isl_merge_divs(div_i, div_j, exp1, exp2); |
2964 | if (!div) |
2965 | goto error; |
2966 | |
2967 | if (div->n_row == div_j->n_row) |
2968 | change = coalesce_with_expanded_divs(bmap_i, |
2969 | i, j, info, div, exp1); |
2970 | else |
2971 | change = isl_change_none; |
2972 | |
2973 | isl_mat_free(div); |
2974 | |
2975 | isl_mat_free(div_i); |
2976 | isl_mat_free(div_j); |
2977 | |
2978 | free(exp2); |
2979 | free(exp1); |
2980 | |
2981 | return change; |
2982 | error: |
2983 | isl_mat_free(div_i); |
2984 | isl_mat_free(div_j); |
2985 | free(exp1); |
2986 | free(exp2); |
2987 | return isl_change_error; |
2988 | } |
2989 | |
2990 | /* Check if basic map "j" is a subset of basic map "i" after |
2991 | * exploiting the extra equalities of "j" to simplify the divs of "i". |
2992 | * If so, remove basic map "j" and return isl_change_drop_second. |
2993 | * |
2994 | * If "j" does not have any equalities or if they are the same |
2995 | * as those of "i", then we cannot exploit them to simplify the divs. |
2996 | * Similarly, if there are no divs in "i", then they cannot be simplified. |
2997 | * If, on the other hand, the affine hulls of "i" and "j" do not intersect, |
2998 | * then "j" cannot be a subset of "i". |
2999 | * |
3000 | * Otherwise, we intersect "i" with the affine hull of "j" and then |
3001 | * check if "j" is a subset of the result after aligning the divs. |
3002 | * If so, then "j" is definitely a subset of "i" and can be removed. |
3003 | * Note that if after intersection with the affine hull of "j". |
3004 | * "i" still has more divs than "j", then there is no way we can |
3005 | * align the divs of "i" to those of "j". |
3006 | */ |
3007 | static enum isl_change coalesce_subset_with_equalities(int i, int j, |
3008 | struct isl_coalesce_info *info) |
3009 | { |
3010 | isl_basic_map *hull_i, *hull_j, *bmap_i; |
3011 | int equal, empty; |
3012 | enum isl_change change; |
3013 | |
3014 | if (info[j].bmap->n_eq == 0) |
3015 | return isl_change_none; |
3016 | if (info[i].bmap->n_div == 0) |
3017 | return isl_change_none; |
3018 | |
3019 | hull_i = isl_basic_map_copy(info[i].bmap); |
3020 | hull_i = isl_basic_map_plain_affine_hull(hull_i); |
3021 | hull_j = isl_basic_map_copy(info[j].bmap); |
3022 | hull_j = isl_basic_map_plain_affine_hull(hull_j); |
3023 | |
3024 | hull_j = isl_basic_map_intersect(hull_j, isl_basic_map_copy(hull_i)); |
3025 | equal = isl_basic_map_plain_is_equal(hull_i, hull_j); |
3026 | empty = isl_basic_map_plain_is_empty(hull_j); |
3027 | isl_basic_map_free(hull_i); |
3028 | |
3029 | if (equal < 0 || equal || empty < 0 || empty) { |
3030 | isl_basic_map_free(hull_j); |
3031 | if (equal < 0 || empty < 0) |
3032 | return isl_change_error; |
3033 | return isl_change_none; |
3034 | } |
3035 | |
3036 | bmap_i = isl_basic_map_copy(info[i].bmap); |
3037 | bmap_i = isl_basic_map_intersect(bmap_i, hull_j); |
3038 | if (!bmap_i) |
3039 | return isl_change_error; |
3040 | |
3041 | if (bmap_i->n_div > info[j].bmap->n_div) { |
3042 | isl_basic_map_free(bmap_i); |
3043 | return isl_change_none; |
3044 | } |
3045 | |
3046 | change = coalesce_after_aligning_divs(bmap_i, -1, j, info); |
3047 | |
3048 | isl_basic_map_free(bmap_i); |
3049 | |
3050 | return change; |
3051 | } |
3052 | |
3053 | /* Check if the union of and the basic maps represented by info[i] and info[j] |
3054 | * can be represented by a single basic map, by aligning or equating |
3055 | * their integer divisions. |
3056 | * If so, replace the pair by the single basic map and return |
3057 | * isl_change_drop_first, isl_change_drop_second or isl_change_fuse. |
3058 | * Otherwise, return isl_change_none. |
3059 | * |
3060 | * Note that we only perform any test if the number of divs is different |
3061 | * in the two basic maps. In case the number of divs is the same, |
3062 | * we have already established that the divs are different |
3063 | * in the two basic maps. |
3064 | * In particular, if the number of divs of basic map i is smaller than |
3065 | * the number of divs of basic map j, then we check if j is a subset of i |
3066 | * and vice versa. |
3067 | */ |
3068 | static enum isl_change coalesce_divs(int i, int j, |
3069 | struct isl_coalesce_info *info) |
3070 | { |
3071 | enum isl_change change = isl_change_none; |
3072 | |
3073 | if (info[i].bmap->n_div < info[j].bmap->n_div) |
3074 | change = coalesce_after_aligning_divs(info[i].bmap, i, j, info); |
3075 | if (change != isl_change_none) |
3076 | return change; |
3077 | |
3078 | if (info[j].bmap->n_div < info[i].bmap->n_div) |
3079 | change = coalesce_after_aligning_divs(info[j].bmap, j, i, info); |
3080 | if (change != isl_change_none) |
3081 | return invert_change(change); |
3082 | |
3083 | change = coalesce_subset_with_equalities(i, j, info); |
3084 | if (change != isl_change_none) |
3085 | return change; |
3086 | |
3087 | change = coalesce_subset_with_equalities(j, i, info); |
3088 | if (change != isl_change_none) |
3089 | return invert_change(change); |
3090 | |
3091 | return isl_change_none; |
3092 | } |
3093 | |
3094 | /* Does "bmap" involve any divs that themselves refer to divs? |
3095 | */ |
3096 | static int has_nested_div(__isl_keep isl_basic_map *bmap) |
3097 | { |
3098 | int i; |
3099 | unsigned total; |
3100 | unsigned n_div; |
3101 | |
3102 | total = isl_basic_map_dim(bmap, isl_dim_all); |
3103 | n_div = isl_basic_map_dim(bmap, isl_dim_div); |
3104 | total -= n_div; |
3105 | |
3106 | for (i = 0; i < n_div; ++i) |
3107 | if (isl_seq_first_non_zero(bmap->div[i] + 2 + total, |
3108 | n_div) != -1) |
3109 | return 1; |
3110 | |
3111 | return 0; |
3112 | } |
3113 | |
3114 | /* Return a list of affine expressions, one for each integer division |
3115 | * in "bmap_i". For each integer division that also appears in "bmap_j", |
3116 | * the affine expression is set to NaN. The number of NaNs in the list |
3117 | * is equal to the number of integer divisions in "bmap_j". |
3118 | * For the other integer divisions of "bmap_i", the corresponding |
3119 | * element in the list is a purely affine expression equal to the integer |
3120 | * division in "hull". |
3121 | * If no such list can be constructed, then the number of elements |
3122 | * in the returned list is smaller than the number of integer divisions |
3123 | * in "bmap_i". |
3124 | */ |
3125 | static __isl_give isl_aff_list *set_up_substitutions( |
3126 | __isl_keep isl_basic_map *bmap_i, __isl_keep isl_basic_map *bmap_j, |
3127 | __isl_take isl_basic_map *hull) |
3128 | { |
3129 | unsigned n_div_i, n_div_j, total; |
3130 | isl_ctx *ctx; |
3131 | isl_local_space *ls; |
3132 | isl_basic_setisl_basic_map *wrap_hull; |
3133 | isl_aff *aff_nan; |
3134 | isl_aff_list *list; |
3135 | int i, j; |
3136 | |
3137 | if (!hull) |
3138 | return NULL((void*)0); |
3139 | |
3140 | ctx = isl_basic_map_get_ctx(hull); |
3141 | |
3142 | n_div_i = isl_basic_map_dim(bmap_i, isl_dim_div); |
3143 | n_div_j = isl_basic_map_dim(bmap_j, isl_dim_div); |
3144 | total = isl_basic_map_total_dim(bmap_i) - n_div_i; |
3145 | |
3146 | ls = isl_basic_map_get_local_space(bmap_i); |
3147 | ls = isl_local_space_wrap(ls); |
3148 | wrap_hull = isl_basic_map_wrap(hull); |
3149 | |
3150 | aff_nan = isl_aff_nan_on_domain(isl_local_space_copy(ls)); |
3151 | list = isl_aff_list_alloc(ctx, n_div_i); |
3152 | |
3153 | j = 0; |
3154 | for (i = 0; i < n_div_i; ++i) { |
3155 | isl_aff *aff; |
3156 | |
3157 | if (j < n_div_j && |
3158 | isl_basic_map_equal_div_expr_part(bmap_i, i, bmap_j, j, |
3159 | 0, 2 + total)) { |
3160 | ++j; |
3161 | list = isl_aff_list_add(list, isl_aff_copy(aff_nan)); |
3162 | continue; |
3163 | } |
3164 | if (n_div_i - i <= n_div_j - j) |
3165 | break; |
3166 | |
3167 | aff = isl_local_space_get_div(ls, i); |
3168 | aff = isl_aff_substitute_equalities(aff, |
3169 | isl_basic_set_copy(wrap_hull)); |
3170 | aff = isl_aff_floor(aff); |
3171 | if (!aff) |
3172 | goto error; |
3173 | if (isl_aff_dim(aff, isl_dim_div) != 0) { |
3174 | isl_aff_free(aff); |
3175 | break; |
3176 | } |
3177 | |
3178 | list = isl_aff_list_add(list, aff); |
3179 | } |
3180 | |
3181 | isl_aff_free(aff_nan); |
3182 | isl_local_space_free(ls); |
3183 | isl_basic_set_free(wrap_hull); |
3184 | |
3185 | return list; |
3186 | error: |
3187 | isl_aff_free(aff_nan); |
3188 | isl_local_space_free(ls); |
3189 | isl_basic_set_free(wrap_hull); |
3190 | isl_aff_list_free(list); |
3191 | return NULL((void*)0); |
3192 | } |
3193 | |
3194 | /* Add variables to info->bmap and info->tab corresponding to the elements |
3195 | * in "list" that are not set to NaN. |
3196 | * "extra_var" is the number of these elements. |
3197 | * "dim" is the offset in the variables of "tab" where we should |
3198 | * start considering the elements in "list". |
3199 | * When this function returns, the total number of variables in "tab" |
3200 | * is equal to "dim" plus the number of elements in "list". |
3201 | * |
3202 | * The newly added existentially quantified variables are not given |
3203 | * an explicit representation because the corresponding div constraints |
3204 | * do not appear in info->bmap. These constraints are not added |
3205 | * to info->bmap because for internal consistency, they would need to |
3206 | * be added to info->tab as well, where they could combine with the equality |
3207 | * that is added later to result in constraints that do not hold |
3208 | * in the original input. |
3209 | */ |
3210 | static int add_sub_vars(struct isl_coalesce_info *info, |
3211 | __isl_keep isl_aff_list *list, int dim, int extra_var) |
3212 | { |
3213 | int i, j, n, d; |
3214 | isl_space *space; |
3215 | |
3216 | space = isl_basic_map_get_space(info->bmap); |
3217 | info->bmap = isl_basic_map_cow(info->bmap); |
3218 | info->bmap = isl_basic_map_extend_space(info->bmap, space, |
3219 | extra_var, 0, 0); |
3220 | if (!info->bmap) |
3221 | return -1; |
3222 | n = isl_aff_list_n_aff(list); |
3223 | for (i = 0; i < n; ++i) { |
3224 | int is_nan; |
3225 | isl_aff *aff; |
3226 | |
3227 | aff = isl_aff_list_get_aff(list, i); |
3228 | is_nan = isl_aff_is_nan(aff); |
3229 | isl_aff_free(aff); |
3230 | if (is_nan < 0) |
3231 | return -1; |
3232 | if (is_nan) |
3233 | continue; |
3234 | |
3235 | if (isl_tab_insert_var(info->tab, dim + i) < 0) |
3236 | return -1; |
3237 | d = isl_basic_map_alloc_div(info->bmap); |
3238 | if (d < 0) |
3239 | return -1; |
3240 | info->bmap = isl_basic_map_mark_div_unknown(info->bmap, d); |
3241 | if (!info->bmap) |
3242 | return -1; |
3243 | for (j = d; j > i; --j) |
3244 | isl_basic_map_swap_div(info->bmap, j - 1, j); |
3245 | } |
3246 | |
3247 | return 0; |
3248 | } |
3249 | |
3250 | /* For each element in "list" that is not set to NaN, fix the corresponding |
3251 | * variable in "tab" to the purely affine expression defined by the element. |
3252 | * "dim" is the offset in the variables of "tab" where we should |
3253 | * start considering the elements in "list". |
3254 | * |
3255 | * This function assumes that a sufficient number of rows and |
3256 | * elements in the constraint array are available in the tableau. |
3257 | */ |
3258 | static int add_sub_equalities(struct isl_tab *tab, |
3259 | __isl_keep isl_aff_list *list, int dim) |
3260 | { |
3261 | int i, n; |
3262 | isl_ctx *ctx; |
3263 | isl_vec *sub; |
3264 | isl_aff *aff; |
3265 | |
3266 | n = isl_aff_list_n_aff(list); |
3267 | |
3268 | ctx = isl_tab_get_ctx(tab); |
3269 | sub = isl_vec_alloc(ctx, 1 + dim + n); |
3270 | if (!sub) |
3271 | return -1; |
3272 | isl_seq_clr(sub->el + 1 + dim, n); |
3273 | |
3274 | for (i = 0; i < n; ++i) { |
3275 | aff = isl_aff_list_get_aff(list, i); |
3276 | if (!aff) |
3277 | goto error; |
3278 | if (isl_aff_is_nan(aff)) { |
3279 | isl_aff_free(aff); |
3280 | continue; |
3281 | } |
3282 | isl_seq_cpy(sub->el, aff->v->el + 1, 1 + dim); |
3283 | isl_int_neg(sub->el[1 + dim + i], aff->v->el[0])isl_sioimath_neg((sub->el[1 + dim + i]), *(aff->v->el [0])); |
3284 | if (isl_tab_add_eq(tab, sub->el) < 0) |
3285 | goto error; |
3286 | isl_int_set_si(sub->el[1 + dim + i], 0)isl_sioimath_set_si((sub->el[1 + dim + i]), 0); |
3287 | isl_aff_free(aff); |
3288 | } |
3289 | |
3290 | isl_vec_free(sub); |
3291 | return 0; |
3292 | error: |
3293 | isl_aff_free(aff); |
3294 | isl_vec_free(sub); |
3295 | return -1; |
3296 | } |
3297 | |
3298 | /* Add variables to info->tab and info->bmap corresponding to the elements |
3299 | * in "list" that are not set to NaN. The value of the added variable |
3300 | * in info->tab is fixed to the purely affine expression defined by the element. |
3301 | * "dim" is the offset in the variables of info->tab where we should |
3302 | * start considering the elements in "list". |
3303 | * When this function returns, the total number of variables in info->tab |
3304 | * is equal to "dim" plus the number of elements in "list". |
3305 | */ |
3306 | static int add_subs(struct isl_coalesce_info *info, |
3307 | __isl_keep isl_aff_list *list, int dim) |
3308 | { |
3309 | int extra_var; |
3310 | int n; |
3311 | |
3312 | if (!list) |
3313 | return -1; |
3314 | |
3315 | n = isl_aff_list_n_aff(list); |
3316 | extra_var = n - (info->tab->n_var - dim); |
3317 | |
3318 | if (isl_tab_extend_vars(info->tab, extra_var) < 0) |
3319 | return -1; |
3320 | if (isl_tab_extend_cons(info->tab, 2 * extra_var) < 0) |
3321 | return -1; |
3322 | if (add_sub_vars(info, list, dim, extra_var) < 0) |
3323 | return -1; |
3324 | |
3325 | return add_sub_equalities(info->tab, list, dim); |
3326 | } |
3327 | |
3328 | /* Coalesce basic map "j" into basic map "i" after adding the extra integer |
3329 | * divisions in "i" but not in "j" to basic map "j", with values |
3330 | * specified by "list". The total number of elements in "list" |
3331 | * is equal to the number of integer divisions in "i", while the number |
3332 | * of NaN elements in the list is equal to the number of integer divisions |
3333 | * in "j". |
3334 | * |
3335 | * If no coalescing can be performed, then we need to revert basic map "j" |
3336 | * to its original state. We do the same if basic map "i" gets dropped |
3337 | * during the coalescing, even though this should not happen in practice |
3338 | * since we have already checked for "j" being a subset of "i" |
3339 | * before we reach this stage. |
3340 | */ |
3341 | static enum isl_change coalesce_with_subs(int i, int j, |
3342 | struct isl_coalesce_info *info, __isl_keep isl_aff_list *list) |
3343 | { |
3344 | isl_basic_map *bmap_j; |
3345 | struct isl_tab_undo *snap; |
3346 | unsigned dim; |
3347 | enum isl_change change; |
3348 | |
3349 | bmap_j = isl_basic_map_copy(info[j].bmap); |
3350 | snap = isl_tab_snap(info[j].tab); |
3351 | |
3352 | dim = isl_basic_map_dim(bmap_j, isl_dim_all); |
3353 | dim -= isl_basic_map_dim(bmap_j, isl_dim_div); |
3354 | if (add_subs(&info[j], list, dim) < 0) |
3355 | goto error; |
3356 | |
3357 | change = coalesce_local_pair(i, j, info); |
3358 | if (change != isl_change_none && change != isl_change_drop_first) { |
3359 | isl_basic_map_free(bmap_j); |
3360 | } else { |
3361 | isl_basic_map_free(info[j].bmap); |
3362 | info[j].bmap = bmap_j; |
3363 | |
3364 | if (isl_tab_rollback(info[j].tab, snap) < 0) |
3365 | return isl_change_error; |
3366 | } |
3367 | |
3368 | return change; |
3369 | error: |
3370 | isl_basic_map_free(bmap_j); |
3371 | return isl_change_error; |
3372 | } |
3373 | |
3374 | /* Check if we can coalesce basic map "j" into basic map "i" after copying |
3375 | * those extra integer divisions in "i" that can be simplified away |
3376 | * using the extra equalities in "j". |
3377 | * All divs are assumed to be known and not contain any nested divs. |
3378 | * |
3379 | * We first check if there are any extra equalities in "j" that we |
3380 | * can exploit. Then we check if every integer division in "i" |
3381 | * either already appears in "j" or can be simplified using the |
3382 | * extra equalities to a purely affine expression. |
3383 | * If these tests succeed, then we try to coalesce the two basic maps |
3384 | * by introducing extra dimensions in "j" corresponding to |
3385 | * the extra integer divsisions "i" fixed to the corresponding |
3386 | * purely affine expression. |
3387 | */ |
3388 | static enum isl_change check_coalesce_into_eq(int i, int j, |
3389 | struct isl_coalesce_info *info) |
3390 | { |
3391 | unsigned n_div_i, n_div_j; |
3392 | isl_basic_map *hull_i, *hull_j; |
3393 | int equal, empty; |
3394 | isl_aff_list *list; |
3395 | enum isl_change change; |
3396 | |
3397 | n_div_i = isl_basic_map_dim(info[i].bmap, isl_dim_div); |
3398 | n_div_j = isl_basic_map_dim(info[j].bmap, isl_dim_div); |
3399 | if (n_div_i <= n_div_j) |
3400 | return isl_change_none; |
3401 | if (info[j].bmap->n_eq == 0) |
3402 | return isl_change_none; |
3403 | |
3404 | hull_i = isl_basic_map_copy(info[i].bmap); |
3405 | hull_i = isl_basic_map_plain_affine_hull(hull_i); |
3406 | hull_j = isl_basic_map_copy(info[j].bmap); |
3407 | hull_j = isl_basic_map_plain_affine_hull(hull_j); |
3408 | |
3409 | hull_j = isl_basic_map_intersect(hull_j, isl_basic_map_copy(hull_i)); |
3410 | equal = isl_basic_map_plain_is_equal(hull_i, hull_j); |
3411 | empty = isl_basic_map_plain_is_empty(hull_j); |
3412 | isl_basic_map_free(hull_i); |
3413 | |
3414 | if (equal < 0 || empty < 0) |
3415 | goto error; |
3416 | if (equal || empty) { |
3417 | isl_basic_map_free(hull_j); |
3418 | return isl_change_none; |
3419 | } |
3420 | |
3421 | list = set_up_substitutions(info[i].bmap, info[j].bmap, hull_j); |
3422 | if (!list) |
3423 | return isl_change_error; |
3424 | if (isl_aff_list_n_aff(list) < n_div_i) |
3425 | change = isl_change_none; |
3426 | else |
3427 | change = coalesce_with_subs(i, j, info, list); |
3428 | |
3429 | isl_aff_list_free(list); |
3430 | |
3431 | return change; |
3432 | error: |
3433 | isl_basic_map_free(hull_j); |
3434 | return isl_change_error; |
3435 | } |
3436 | |
3437 | /* Check if we can coalesce basic maps "i" and "j" after copying |
3438 | * those extra integer divisions in one of the basic maps that can |
3439 | * be simplified away using the extra equalities in the other basic map. |
3440 | * We require all divs to be known in both basic maps. |
3441 | * Furthermore, to simplify the comparison of div expressions, |
3442 | * we do not allow any nested integer divisions. |
3443 | */ |
3444 | static enum isl_change check_coalesce_eq(int i, int j, |
3445 | struct isl_coalesce_info *info) |
3446 | { |
3447 | int known, nested; |
3448 | enum isl_change change; |
3449 | |
3450 | known = isl_basic_map_divs_known(info[i].bmap); |
3451 | if (known < 0 || !known) |
3452 | return known < 0 ? isl_change_error : isl_change_none; |
3453 | known = isl_basic_map_divs_known(info[j].bmap); |
3454 | if (known < 0 || !known) |
3455 | return known < 0 ? isl_change_error : isl_change_none; |
3456 | nested = has_nested_div(info[i].bmap); |
3457 | if (nested < 0 || nested) |
3458 | return nested < 0 ? isl_change_error : isl_change_none; |
3459 | nested = has_nested_div(info[j].bmap); |
3460 | if (nested < 0 || nested) |
3461 | return nested < 0 ? isl_change_error : isl_change_none; |
3462 | |
3463 | change = check_coalesce_into_eq(i, j, info); |
3464 | if (change != isl_change_none) |
3465 | return change; |
3466 | change = check_coalesce_into_eq(j, i, info); |
3467 | if (change != isl_change_none) |
3468 | return invert_change(change); |
3469 | |
3470 | return isl_change_none; |
3471 | } |
3472 | |
3473 | /* Check if the union of the given pair of basic maps |
3474 | * can be represented by a single basic map. |
3475 | * If so, replace the pair by the single basic map and return |
3476 | * isl_change_drop_first, isl_change_drop_second or isl_change_fuse. |
3477 | * Otherwise, return isl_change_none. |
3478 | * |
3479 | * We first check if the two basic maps live in the same local space, |
3480 | * after aligning the divs that differ by only an integer constant. |
3481 | * If so, we do the complete check. Otherwise, we check if they have |
3482 | * the same number of integer divisions and can be coalesced, if one is |
3483 | * an obvious subset of the other or if the extra integer divisions |
3484 | * of one basic map can be simplified away using the extra equalities |
3485 | * of the other basic map. |
3486 | */ |
3487 | static enum isl_change coalesce_pair(int i, int j, |
3488 | struct isl_coalesce_info *info) |
3489 | { |
3490 | int same; |
3491 | enum isl_change change; |
3492 | |
3493 | if (harmonize_divs(&info[i], &info[j]) < 0) |
3494 | return isl_change_error; |
3495 | same = same_divs(info[i].bmap, info[j].bmap); |
3496 | if (same < 0) |
3497 | return isl_change_error; |
3498 | if (same) |
3499 | return coalesce_local_pair(i, j, info); |
3500 | |
3501 | if (info[i].bmap->n_div == info[j].bmap->n_div) { |
3502 | change = coalesce_local_pair(i, j, info); |
3503 | if (change != isl_change_none) |
3504 | return change; |
3505 | } |
3506 | |
3507 | change = coalesce_divs(i, j, info); |
3508 | if (change != isl_change_none) |
3509 | return change; |
3510 | |
3511 | return check_coalesce_eq(i, j, info); |
3512 | } |
3513 | |
3514 | /* Return the maximum of "a" and "b". |
3515 | */ |
3516 | static int isl_max(int a, int b) |
3517 | { |
3518 | return a > b ? a : b; |
3519 | } |
3520 | |
3521 | /* Pairwise coalesce the basic maps in the range [start1, end1[ of "info" |
3522 | * with those in the range [start2, end2[, skipping basic maps |
3523 | * that have been removed (either before or within this function). |
3524 | * |
3525 | * For each basic map i in the first range, we check if it can be coalesced |
3526 | * with respect to any previously considered basic map j in the second range. |
3527 | * If i gets dropped (because it was a subset of some j), then |
3528 | * we can move on to the next basic map. |
3529 | * If j gets dropped, we need to continue checking against the other |
3530 | * previously considered basic maps. |
3531 | * If the two basic maps got fused, then we recheck the fused basic map |
3532 | * against the previously considered basic maps, starting at i + 1 |
3533 | * (even if start2 is greater than i + 1). |
3534 | */ |
3535 | static int coalesce_range(isl_ctx *ctx, struct isl_coalesce_info *info, |
3536 | int start1, int end1, int start2, int end2) |
3537 | { |
3538 | int i, j; |
3539 | |
3540 | for (i = end1 - 1; i >= start1; --i) { |
3541 | if (info[i].removed) |
3542 | continue; |
3543 | for (j = isl_max(i + 1, start2); j < end2; ++j) { |
3544 | enum isl_change changed; |
3545 | |
3546 | if (info[j].removed) |
3547 | continue; |
3548 | if (info[i].removed) |
3549 | isl_die(ctx, isl_error_internal,do { isl_handle_error(ctx, isl_error_internal, "basic map unexpectedly removed" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn295388/tools/polly/lib/External/isl/isl_coalesce.c" , 3551); return -1; } while (0) |
3550 | "basic map unexpectedly removed",do { isl_handle_error(ctx, isl_error_internal, "basic map unexpectedly removed" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn295388/tools/polly/lib/External/isl/isl_coalesce.c" , 3551); return -1; } while (0) |
3551 | return -1)do { isl_handle_error(ctx, isl_error_internal, "basic map unexpectedly removed" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn295388/tools/polly/lib/External/isl/isl_coalesce.c" , 3551); return -1; } while (0); |
3552 | changed = coalesce_pair(i, j, info); |
3553 | switch (changed) { |
3554 | case isl_change_error: |
3555 | return -1; |
3556 | case isl_change_none: |
3557 | case isl_change_drop_second: |
3558 | continue; |
3559 | case isl_change_drop_first: |
3560 | j = end2; |
3561 | break; |
3562 | case isl_change_fuse: |
3563 | j = i; |
3564 | break; |
3565 | } |
3566 | } |
3567 | } |
3568 | |
3569 | return 0; |
3570 | } |
3571 | |
3572 | /* Pairwise coalesce the basic maps described by the "n" elements of "info". |
3573 | * |
3574 | * We consider groups of basic maps that live in the same apparent |
3575 | * affine hull and we first coalesce within such a group before we |
3576 | * coalesce the elements in the group with elements of previously |
3577 | * considered groups. If a fuse happens during the second phase, |
3578 | * then we also reconsider the elements within the group. |
3579 | */ |
3580 | static int coalesce(isl_ctx *ctx, int n, struct isl_coalesce_info *info) |
3581 | { |
3582 | int start, end; |
3583 | |
3584 | for (end = n; end > 0; end = start) { |
3585 | start = end - 1; |
3586 | while (start >= 1 && |
3587 | info[start - 1].hull_hash == info[start].hull_hash) |
3588 | start--; |
3589 | if (coalesce_range(ctx, info, start, end, start, end) < 0) |
3590 | return -1; |
3591 | if (coalesce_range(ctx, info, start, end, end, n) < 0) |
3592 | return -1; |
3593 | } |
3594 | |
3595 | return 0; |
3596 | } |
3597 | |
3598 | /* Update the basic maps in "map" based on the information in "info". |
3599 | * In particular, remove the basic maps that have been marked removed and |
3600 | * update the others based on the information in the corresponding tableau. |
3601 | * Since we detected implicit equalities without calling |
3602 | * isl_basic_map_gauss, we need to do it now. |
3603 | * Also call isl_basic_map_simplify if we may have lost the definition |
3604 | * of one or more integer divisions. |
3605 | */ |
3606 | static __isl_give isl_map *update_basic_maps(__isl_take isl_map *map, |
3607 | int n, struct isl_coalesce_info *info) |
3608 | { |
3609 | int i; |
3610 | |
3611 | if (!map) |
3612 | return NULL((void*)0); |
3613 | |
3614 | for (i = n - 1; i >= 0; --i) { |
3615 | if (info[i].removed) { |
3616 | isl_basic_map_free(map->p[i]); |
3617 | if (i != map->n - 1) |
3618 | map->p[i] = map->p[map->n - 1]; |
3619 | map->n--; |
3620 | continue; |
3621 | } |
3622 | |
3623 | info[i].bmap = isl_basic_map_update_from_tab(info[i].bmap, |
3624 | info[i].tab); |
3625 | info[i].bmap = isl_basic_map_gauss(info[i].bmap, NULL((void*)0)); |
3626 | if (info[i].simplify) |
3627 | info[i].bmap = isl_basic_map_simplify(info[i].bmap); |
3628 | info[i].bmap = isl_basic_map_finalize(info[i].bmap); |
3629 | if (!info[i].bmap) |
3630 | return isl_map_free(map); |
3631 | ISL_F_SET(info[i].bmap, ISL_BASIC_MAP_NO_IMPLICIT)(((info[i].bmap)->flags) |= ((1 << 2))); |
3632 | ISL_F_SET(info[i].bmap, ISL_BASIC_MAP_NO_REDUNDANT)(((info[i].bmap)->flags) |= ((1 << 3))); |
3633 | isl_basic_map_free(map->p[i]); |
3634 | map->p[i] = info[i].bmap; |
3635 | info[i].bmap = NULL((void*)0); |
3636 | } |
3637 | |
3638 | return map; |
3639 | } |
3640 | |
3641 | /* For each pair of basic maps in the map, check if the union of the two |
3642 | * can be represented by a single basic map. |
3643 | * If so, replace the pair by the single basic map and start over. |
3644 | * |
3645 | * We factor out any (hidden) common factor from the constraint |
3646 | * coefficients to improve the detection of adjacent constraints. |
3647 | * |
3648 | * Since we are constructing the tableaus of the basic maps anyway, |
3649 | * we exploit them to detect implicit equalities and redundant constraints. |
3650 | * This also helps the coalescing as it can ignore the redundant constraints. |
3651 | * In order to avoid confusion, we make all implicit equalities explicit |
3652 | * in the basic maps. We don't call isl_basic_map_gauss, though, |
3653 | * as that may affect the number of constraints. |
3654 | * This means that we have to call isl_basic_map_gauss at the end |
3655 | * of the computation (in update_basic_maps) to ensure that |
3656 | * the basic maps are not left in an unexpected state. |
3657 | * For each basic map, we also compute the hash of the apparent affine hull |
3658 | * for use in coalesce. |
3659 | */ |
3660 | struct isl_map *isl_map_coalesce(struct isl_map *map) |
3661 | { |
3662 | int i; |
3663 | unsigned n; |
3664 | isl_ctx *ctx; |
3665 | struct isl_coalesce_info *info = NULL((void*)0); |
3666 | |
3667 | map = isl_map_remove_empty_parts(map); |
3668 | if (!map) |
3669 | return NULL((void*)0); |
3670 | |
3671 | if (map->n <= 1) |
3672 | return map; |
3673 | |
3674 | ctx = isl_map_get_ctx(map); |
3675 | map = isl_map_sort_divs(map); |
3676 | map = isl_map_cow(map); |
3677 | |
3678 | if (!map) |
3679 | return NULL((void*)0); |
3680 | |
3681 | n = map->n; |
3682 | |
3683 | info = isl_calloc_array(map->ctx, struct isl_coalesce_info, n)((struct isl_coalesce_info *)isl_calloc_or_die(map->ctx, n , sizeof(struct isl_coalesce_info))); |
3684 | if (!info) |
3685 | goto error; |
3686 | |
3687 | for (i = 0; i < map->n; ++i) { |
3688 | map->p[i] = isl_basic_map_reduce_coefficients(map->p[i]); |
3689 | if (!map->p[i]) |
3690 | goto error; |
3691 | info[i].bmap = isl_basic_map_copy(map->p[i]); |
3692 | info[i].tab = isl_tab_from_basic_map(info[i].bmap, 0); |
3693 | if (!info[i].tab) |
3694 | goto error; |
3695 | if (!ISL_F_ISSET(info[i].bmap, ISL_BASIC_MAP_NO_IMPLICIT)(!!(((info[i].bmap)->flags) & ((1 << 2))))) |
3696 | if (isl_tab_detect_implicit_equalities(info[i].tab) < 0) |
3697 | goto error; |
3698 | info[i].bmap = isl_tab_make_equalities_explicit(info[i].tab, |
3699 | info[i].bmap); |
3700 | if (!info[i].bmap) |
3701 | goto error; |
3702 | if (!ISL_F_ISSET(info[i].bmap, ISL_BASIC_MAP_NO_REDUNDANT)(!!(((info[i].bmap)->flags) & ((1 << 3))))) |
3703 | if (isl_tab_detect_redundant(info[i].tab) < 0) |
3704 | goto error; |
3705 | if (coalesce_info_set_hull_hash(&info[i]) < 0) |
3706 | goto error; |
3707 | } |
3708 | for (i = map->n - 1; i >= 0; --i) |
3709 | if (info[i].tab->empty) |
3710 | drop(&info[i]); |
3711 | |
3712 | if (coalesce(ctx, n, info) < 0) |
3713 | goto error; |
3714 | |
3715 | map = update_basic_maps(map, n, info); |
3716 | |
3717 | clear_coalesce_info(n, info); |
3718 | |
3719 | return map; |
3720 | error: |
3721 | clear_coalesce_info(n, info); |
3722 | isl_map_free(map); |
3723 | return NULL((void*)0); |
3724 | } |
3725 | |
3726 | /* For each pair of basic sets in the set, check if the union of the two |
3727 | * can be represented by a single basic set. |
3728 | * If so, replace the pair by the single basic set and start over. |
3729 | */ |
3730 | struct isl_setisl_map *isl_set_coalesce(struct isl_setisl_map *set) |
3731 | { |
3732 | return set_from_map(isl_map_coalesce(set_to_map(set))); |
3733 | } |