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 * Copyright 2020 Cerebras Systems 8 * 9 * Use of this software is governed by the MIT license 10 * 11 * Written by Sven Verdoolaege, K.U.Leuven, Departement 12 * Computerwetenschappen, Celestijnenlaan 200A, B-3001 Leuven, Belgium 13 * and INRIA Saclay - Ile-de-France, Parc Club Orsay Universite, 14 * ZAC des vignes, 4 rue Jacques Monod, 91893 Orsay, France 15 * and Ecole Normale Superieure, 45 rue dUlm, 75230 Paris, France 16 * and Inria Paris - Rocquencourt, Domaine de Voluceau - Rocquencourt, 17 * B.P. 105 - 78153 Le Chesnay, France 18 * and Centre de Recherche Inria de Paris, 2 rue Simone Iff - Voie DQ12, 19 * CS 42112, 75589 Paris Cedex 12, France 20 * and Cerebras Systems, 175 S San Antonio Rd, Los Altos, CA, USA 21 */ 22 23 #include <isl_ctx_private.h> 24 #include "isl_map_private.h" 25 #include <isl_seq.h> 26 #include <isl/options.h> 27 #include "isl_tab.h" 28 #include <isl_mat_private.h> 29 #include <isl_local_space_private.h> 30 #include <isl_val_private.h> 31 #include <isl_vec_private.h> 32 #include <isl_aff_private.h> 33 #include <isl_equalities.h> 34 #include <isl_constraint_private.h> 35 36 #include <set_to_map.c> 37 #include <set_from_map.c> 38 39 #define STATUS_ERROR -1 40 #define STATUS_REDUNDANT 1 41 #define STATUS_VALID 2 42 #define STATUS_SEPARATE 3 43 #define STATUS_CUT 4 44 #define STATUS_ADJ_EQ 5 45 #define STATUS_ADJ_INEQ 6 46 47 static int status_in(isl_int *ineq, struct isl_tab *tab) 48 { 49 enum isl_ineq_type type = isl_tab_ineq_type(tab, ineq); 50 switch (type) { 51 default: 52 case isl_ineq_error: return STATUS_ERROR; 53 case isl_ineq_redundant: return STATUS_VALID; 54 case isl_ineq_separate: return STATUS_SEPARATE; 55 case isl_ineq_cut: return STATUS_CUT; 56 case isl_ineq_adj_eq: return STATUS_ADJ_EQ; 57 case isl_ineq_adj_ineq: return STATUS_ADJ_INEQ; 58 } 59 } 60 61 /* Compute the position of the equalities of basic map "bmap_i" 62 * with respect to the basic map represented by "tab_j". 63 * The resulting array has twice as many entries as the number 64 * of equalities corresponding to the two inequalities to which 65 * each equality corresponds. 66 */ 67 static int *eq_status_in(__isl_keep isl_basic_map *bmap_i, 68 struct isl_tab *tab_j) 69 { 70 int k, l; 71 int *eq; 72 isl_size dim; 73 74 dim = isl_basic_map_dim(bmap_i, isl_dim_all); 75 if (dim < 0) 76 return NULL; 77 78 eq = isl_calloc_array(bmap_i->ctx, int, 2 * bmap_i->n_eq); 79 if (!eq) 80 return NULL; 81 82 for (k = 0; k < bmap_i->n_eq; ++k) { 83 for (l = 0; l < 2; ++l) { 84 isl_seq_neg(bmap_i->eq[k], bmap_i->eq[k], 1+dim); 85 eq[2 * k + l] = status_in(bmap_i->eq[k], tab_j); 86 if (eq[2 * k + l] == STATUS_ERROR) 87 goto error; 88 } 89 } 90 91 return eq; 92 error: 93 free(eq); 94 return NULL; 95 } 96 97 /* Compute the position of the inequalities of basic map "bmap_i" 98 * (also represented by "tab_i", if not NULL) with respect to the basic map 99 * represented by "tab_j". 100 */ 101 static int *ineq_status_in(__isl_keep isl_basic_map *bmap_i, 102 struct isl_tab *tab_i, struct isl_tab *tab_j) 103 { 104 int k; 105 unsigned n_eq = bmap_i->n_eq; 106 int *ineq = isl_calloc_array(bmap_i->ctx, int, bmap_i->n_ineq); 107 108 if (!ineq) 109 return NULL; 110 111 for (k = 0; k < bmap_i->n_ineq; ++k) { 112 if (tab_i && isl_tab_is_redundant(tab_i, n_eq + k)) { 113 ineq[k] = STATUS_REDUNDANT; 114 continue; 115 } 116 ineq[k] = status_in(bmap_i->ineq[k], tab_j); 117 if (ineq[k] == STATUS_ERROR) 118 goto error; 119 if (ineq[k] == STATUS_SEPARATE) 120 break; 121 } 122 123 return ineq; 124 error: 125 free(ineq); 126 return NULL; 127 } 128 129 static int any(int *con, unsigned len, int status) 130 { 131 int i; 132 133 for (i = 0; i < len ; ++i) 134 if (con[i] == status) 135 return 1; 136 return 0; 137 } 138 139 /* Return the first position of "status" in the list "con" of length "len". 140 * Return -1 if there is no such entry. 141 */ 142 static int find(int *con, unsigned len, int status) 143 { 144 int i; 145 146 for (i = 0; i < len ; ++i) 147 if (con[i] == status) 148 return i; 149 return -1; 150 } 151 152 static int count(int *con, unsigned len, int status) 153 { 154 int i; 155 int c = 0; 156 157 for (i = 0; i < len ; ++i) 158 if (con[i] == status) 159 c++; 160 return c; 161 } 162 163 static int all(int *con, unsigned len, int status) 164 { 165 int i; 166 167 for (i = 0; i < len ; ++i) { 168 if (con[i] == STATUS_REDUNDANT) 169 continue; 170 if (con[i] != status) 171 return 0; 172 } 173 return 1; 174 } 175 176 /* Internal information associated to a basic map in a map 177 * that is to be coalesced by isl_map_coalesce. 178 * 179 * "bmap" is the basic map itself (or NULL if "removed" is set) 180 * "tab" is the corresponding tableau (or NULL if "removed" is set) 181 * "hull_hash" identifies the affine space in which "bmap" lives. 182 * "modified" is set if this basic map may not be identical 183 * to any of the basic maps in the input. 184 * "removed" is set if this basic map has been removed from the map 185 * "simplify" is set if this basic map may have some unknown integer 186 * divisions that were not present in the input basic maps. The basic 187 * map should then be simplified such that we may be able to find 188 * a definition among the constraints. 189 * 190 * "eq" and "ineq" are only set if we are currently trying to coalesce 191 * this basic map with another basic map, in which case they represent 192 * the position of the inequalities of this basic map with respect to 193 * the other basic map. The number of elements in the "eq" array 194 * is twice the number of equalities in the "bmap", corresponding 195 * to the two inequalities that make up each equality. 196 */ 197 struct isl_coalesce_info { 198 isl_basic_map *bmap; 199 struct isl_tab *tab; 200 uint32_t hull_hash; 201 int modified; 202 int removed; 203 int simplify; 204 int *eq; 205 int *ineq; 206 }; 207 208 /* Is there any (half of an) equality constraint in the description 209 * of the basic map represented by "info" that 210 * has position "status" with respect to the other basic map? 211 */ 212 static int any_eq(struct isl_coalesce_info *info, int status) 213 { 214 isl_size n_eq; 215 216 n_eq = isl_basic_map_n_equality(info->bmap); 217 return any(info->eq, 2 * n_eq, status); 218 } 219 220 /* Is there any inequality constraint in the description 221 * of the basic map represented by "info" that 222 * has position "status" with respect to the other basic map? 223 */ 224 static int any_ineq(struct isl_coalesce_info *info, int status) 225 { 226 isl_size n_ineq; 227 228 n_ineq = isl_basic_map_n_inequality(info->bmap); 229 return any(info->ineq, n_ineq, status); 230 } 231 232 /* Return the position of the first half on an equality constraint 233 * in the description of the basic map represented by "info" that 234 * has position "status" with respect to the other basic map. 235 * The returned value is twice the position of the equality constraint 236 * plus zero for the negative half and plus one for the positive half. 237 * Return -1 if there is no such entry. 238 */ 239 static int find_eq(struct isl_coalesce_info *info, int status) 240 { 241 isl_size n_eq; 242 243 n_eq = isl_basic_map_n_equality(info->bmap); 244 return find(info->eq, 2 * n_eq, status); 245 } 246 247 /* Return the position of the first inequality constraint in the description 248 * of the basic map represented by "info" that 249 * has position "status" with respect to the other basic map. 250 * Return -1 if there is no such entry. 251 */ 252 static int find_ineq(struct isl_coalesce_info *info, int status) 253 { 254 isl_size n_ineq; 255 256 n_ineq = isl_basic_map_n_inequality(info->bmap); 257 return find(info->ineq, n_ineq, status); 258 } 259 260 /* Return the number of (halves of) equality constraints in the description 261 * of the basic map represented by "info" that 262 * have position "status" with respect to the other basic map. 263 */ 264 static int count_eq(struct isl_coalesce_info *info, int status) 265 { 266 isl_size n_eq; 267 268 n_eq = isl_basic_map_n_equality(info->bmap); 269 return count(info->eq, 2 * n_eq, status); 270 } 271 272 /* Return the number of inequality constraints in the description 273 * of the basic map represented by "info" that 274 * have position "status" with respect to the other basic map. 275 */ 276 static int count_ineq(struct isl_coalesce_info *info, int status) 277 { 278 isl_size n_ineq; 279 280 n_ineq = isl_basic_map_n_inequality(info->bmap); 281 return count(info->ineq, n_ineq, status); 282 } 283 284 /* Are all non-redundant constraints of the basic map represented by "info" 285 * either valid or cut constraints with respect to the other basic map? 286 */ 287 static int all_valid_or_cut(struct isl_coalesce_info *info) 288 { 289 int i; 290 291 for (i = 0; i < 2 * info->bmap->n_eq; ++i) { 292 if (info->eq[i] == STATUS_REDUNDANT) 293 continue; 294 if (info->eq[i] == STATUS_VALID) 295 continue; 296 if (info->eq[i] == STATUS_CUT) 297 continue; 298 return 0; 299 } 300 301 for (i = 0; i < info->bmap->n_ineq; ++i) { 302 if (info->ineq[i] == STATUS_REDUNDANT) 303 continue; 304 if (info->ineq[i] == STATUS_VALID) 305 continue; 306 if (info->ineq[i] == STATUS_CUT) 307 continue; 308 return 0; 309 } 310 311 return 1; 312 } 313 314 /* Compute the hash of the (apparent) affine hull of info->bmap (with 315 * the existentially quantified variables removed) and store it 316 * in info->hash. 317 */ 318 static int coalesce_info_set_hull_hash(struct isl_coalesce_info *info) 319 { 320 isl_basic_map *hull; 321 isl_size n_div; 322 323 hull = isl_basic_map_copy(info->bmap); 324 hull = isl_basic_map_plain_affine_hull(hull); 325 n_div = isl_basic_map_dim(hull, isl_dim_div); 326 if (n_div < 0) 327 hull = isl_basic_map_free(hull); 328 hull = isl_basic_map_drop_constraints_involving_dims(hull, 329 isl_dim_div, 0, n_div); 330 info->hull_hash = isl_basic_map_get_hash(hull); 331 isl_basic_map_free(hull); 332 333 return hull ? 0 : -1; 334 } 335 336 /* Free all the allocated memory in an array 337 * of "n" isl_coalesce_info elements. 338 */ 339 static void clear_coalesce_info(int n, struct isl_coalesce_info *info) 340 { 341 int i; 342 343 if (!info) 344 return; 345 346 for (i = 0; i < n; ++i) { 347 isl_basic_map_free(info[i].bmap); 348 isl_tab_free(info[i].tab); 349 } 350 351 free(info); 352 } 353 354 /* Clear the memory associated to "info". 355 */ 356 static void clear(struct isl_coalesce_info *info) 357 { 358 info->bmap = isl_basic_map_free(info->bmap); 359 isl_tab_free(info->tab); 360 info->tab = NULL; 361 } 362 363 /* Drop the basic map represented by "info". 364 * That is, clear the memory associated to the entry and 365 * mark it as having been removed. 366 */ 367 static void drop(struct isl_coalesce_info *info) 368 { 369 clear(info); 370 info->removed = 1; 371 } 372 373 /* Exchange the information in "info1" with that in "info2". 374 */ 375 static void exchange(struct isl_coalesce_info *info1, 376 struct isl_coalesce_info *info2) 377 { 378 struct isl_coalesce_info info; 379 380 info = *info1; 381 *info1 = *info2; 382 *info2 = info; 383 } 384 385 /* This type represents the kind of change that has been performed 386 * while trying to coalesce two basic maps. 387 * 388 * isl_change_none: nothing was changed 389 * isl_change_drop_first: the first basic map was removed 390 * isl_change_drop_second: the second basic map was removed 391 * isl_change_fuse: the two basic maps were replaced by a new basic map. 392 */ 393 enum isl_change { 394 isl_change_error = -1, 395 isl_change_none = 0, 396 isl_change_drop_first, 397 isl_change_drop_second, 398 isl_change_fuse, 399 }; 400 401 /* Update "change" based on an interchange of the first and the second 402 * basic map. That is, interchange isl_change_drop_first and 403 * isl_change_drop_second. 404 */ 405 static enum isl_change invert_change(enum isl_change change) 406 { 407 switch (change) { 408 case isl_change_error: 409 return isl_change_error; 410 case isl_change_none: 411 return isl_change_none; 412 case isl_change_drop_first: 413 return isl_change_drop_second; 414 case isl_change_drop_second: 415 return isl_change_drop_first; 416 case isl_change_fuse: 417 return isl_change_fuse; 418 } 419 420 return isl_change_error; 421 } 422 423 /* Add the valid constraints of the basic map represented by "info" 424 * to "bmap". "len" is the size of the constraints. 425 * If only one of the pair of inequalities that make up an equality 426 * is valid, then add that inequality. 427 */ 428 static __isl_give isl_basic_map *add_valid_constraints( 429 __isl_take isl_basic_map *bmap, struct isl_coalesce_info *info, 430 unsigned len) 431 { 432 int k, l; 433 434 if (!bmap) 435 return NULL; 436 437 for (k = 0; k < info->bmap->n_eq; ++k) { 438 if (info->eq[2 * k] == STATUS_VALID && 439 info->eq[2 * k + 1] == STATUS_VALID) { 440 l = isl_basic_map_alloc_equality(bmap); 441 if (l < 0) 442 return isl_basic_map_free(bmap); 443 isl_seq_cpy(bmap->eq[l], info->bmap->eq[k], len); 444 } else if (info->eq[2 * k] == STATUS_VALID) { 445 l = isl_basic_map_alloc_inequality(bmap); 446 if (l < 0) 447 return isl_basic_map_free(bmap); 448 isl_seq_neg(bmap->ineq[l], info->bmap->eq[k], len); 449 } else if (info->eq[2 * k + 1] == STATUS_VALID) { 450 l = isl_basic_map_alloc_inequality(bmap); 451 if (l < 0) 452 return isl_basic_map_free(bmap); 453 isl_seq_cpy(bmap->ineq[l], info->bmap->eq[k], len); 454 } 455 } 456 457 for (k = 0; k < info->bmap->n_ineq; ++k) { 458 if (info->ineq[k] != STATUS_VALID) 459 continue; 460 l = isl_basic_map_alloc_inequality(bmap); 461 if (l < 0) 462 return isl_basic_map_free(bmap); 463 isl_seq_cpy(bmap->ineq[l], info->bmap->ineq[k], len); 464 } 465 466 return bmap; 467 } 468 469 /* Is "bmap" defined by a number of (non-redundant) constraints that 470 * is greater than the number of constraints of basic maps i and j combined? 471 * Equalities are counted as two inequalities. 472 */ 473 static int number_of_constraints_increases(int i, int j, 474 struct isl_coalesce_info *info, 475 __isl_keep isl_basic_map *bmap, struct isl_tab *tab) 476 { 477 int k, n_old, n_new; 478 479 n_old = 2 * info[i].bmap->n_eq + info[i].bmap->n_ineq; 480 n_old += 2 * info[j].bmap->n_eq + info[j].bmap->n_ineq; 481 482 n_new = 2 * bmap->n_eq; 483 for (k = 0; k < bmap->n_ineq; ++k) 484 if (!isl_tab_is_redundant(tab, bmap->n_eq + k)) 485 ++n_new; 486 487 return n_new > n_old; 488 } 489 490 /* Replace the pair of basic maps i and j by the basic map bounded 491 * by the valid constraints in both basic maps and the constraints 492 * in extra (if not NULL). 493 * Place the fused basic map in the position that is the smallest of i and j. 494 * 495 * If "detect_equalities" is set, then look for equalities encoded 496 * as pairs of inequalities. 497 * If "check_number" is set, then the original basic maps are only 498 * replaced if the total number of constraints does not increase. 499 * While the number of integer divisions in the two basic maps 500 * is assumed to be the same, the actual definitions may be different. 501 * We only copy the definition from one of the basic maps if it is 502 * the same as that of the other basic map. Otherwise, we mark 503 * the integer division as unknown and simplify the basic map 504 * in an attempt to recover the integer division definition. 505 * If any extra constraints get introduced, then these may 506 * involve integer divisions with a unit coefficient. 507 * Eliminate those that do not appear with any other coefficient 508 * in other constraints, to ensure they get eliminated completely, 509 * improving the chances of further coalescing. 510 */ 511 static enum isl_change fuse(int i, int j, struct isl_coalesce_info *info, 512 __isl_keep isl_mat *extra, int detect_equalities, int check_number) 513 { 514 int k, l; 515 struct isl_basic_map *fused = NULL; 516 struct isl_tab *fused_tab = NULL; 517 isl_size total = isl_basic_map_dim(info[i].bmap, isl_dim_all); 518 unsigned extra_rows = extra ? extra->n_row : 0; 519 unsigned n_eq, n_ineq; 520 int simplify = 0; 521 522 if (total < 0) 523 return isl_change_error; 524 if (j < i) 525 return fuse(j, i, info, extra, detect_equalities, check_number); 526 527 n_eq = info[i].bmap->n_eq + info[j].bmap->n_eq; 528 n_ineq = info[i].bmap->n_ineq + info[j].bmap->n_ineq; 529 fused = isl_basic_map_alloc_space(isl_space_copy(info[i].bmap->dim), 530 info[i].bmap->n_div, n_eq, n_eq + n_ineq + extra_rows); 531 fused = add_valid_constraints(fused, &info[i], 1 + total); 532 fused = add_valid_constraints(fused, &info[j], 1 + total); 533 if (!fused) 534 goto error; 535 if (ISL_F_ISSET(info[i].bmap, ISL_BASIC_MAP_RATIONAL) && 536 ISL_F_ISSET(info[j].bmap, ISL_BASIC_MAP_RATIONAL)) 537 ISL_F_SET(fused, ISL_BASIC_MAP_RATIONAL); 538 539 for (k = 0; k < info[i].bmap->n_div; ++k) { 540 int l = isl_basic_map_alloc_div(fused); 541 if (l < 0) 542 goto error; 543 if (isl_seq_eq(info[i].bmap->div[k], info[j].bmap->div[k], 544 1 + 1 + total)) { 545 isl_seq_cpy(fused->div[l], info[i].bmap->div[k], 546 1 + 1 + total); 547 } else { 548 isl_int_set_si(fused->div[l][0], 0); 549 simplify = 1; 550 } 551 } 552 553 for (k = 0; k < extra_rows; ++k) { 554 l = isl_basic_map_alloc_inequality(fused); 555 if (l < 0) 556 goto error; 557 isl_seq_cpy(fused->ineq[l], extra->row[k], 1 + total); 558 } 559 560 if (detect_equalities) 561 fused = isl_basic_map_detect_inequality_pairs(fused, NULL); 562 fused = isl_basic_map_gauss(fused, NULL); 563 if (simplify || info[j].simplify) { 564 fused = isl_basic_map_simplify(fused); 565 info[i].simplify = 0; 566 } else if (extra_rows > 0) { 567 fused = isl_basic_map_eliminate_pure_unit_divs(fused); 568 } 569 fused = isl_basic_map_finalize(fused); 570 571 fused_tab = isl_tab_from_basic_map(fused, 0); 572 if (isl_tab_detect_redundant(fused_tab) < 0) 573 goto error; 574 575 if (check_number && 576 number_of_constraints_increases(i, j, info, fused, fused_tab)) { 577 isl_tab_free(fused_tab); 578 isl_basic_map_free(fused); 579 return isl_change_none; 580 } 581 582 clear(&info[i]); 583 info[i].bmap = fused; 584 info[i].tab = fused_tab; 585 info[i].modified = 1; 586 drop(&info[j]); 587 588 return isl_change_fuse; 589 error: 590 isl_tab_free(fused_tab); 591 isl_basic_map_free(fused); 592 return isl_change_error; 593 } 594 595 /* Given a pair of basic maps i and j such that all constraints are either 596 * "valid" or "cut", check if the facets corresponding to the "cut" 597 * constraints of i lie entirely within basic map j. 598 * If so, replace the pair by the basic map consisting of the valid 599 * constraints in both basic maps. 600 * Checking whether the facet lies entirely within basic map j 601 * is performed by checking whether the constraints of basic map j 602 * are valid for the facet. These tests are performed on a rational 603 * tableau to avoid the theoretical possibility that a constraint 604 * that was considered to be a cut constraint for the entire basic map i 605 * happens to be considered to be a valid constraint for the facet, 606 * even though it cuts off the same rational points. 607 * 608 * To see that we are not introducing any extra points, call the 609 * two basic maps A and B and the resulting map U and let x 610 * be an element of U \setminus ( A \cup B ). 611 * A line connecting x with an element of A \cup B meets a facet F 612 * of either A or B. Assume it is a facet of B and let c_1 be 613 * the corresponding facet constraint. We have c_1(x) < 0 and 614 * so c_1 is a cut constraint. This implies that there is some 615 * (possibly rational) point x' satisfying the constraints of A 616 * and the opposite of c_1 as otherwise c_1 would have been marked 617 * valid for A. The line connecting x and x' meets a facet of A 618 * in a (possibly rational) point that also violates c_1, but this 619 * is impossible since all cut constraints of B are valid for all 620 * cut facets of A. 621 * In case F is a facet of A rather than B, then we can apply the 622 * above reasoning to find a facet of B separating x from A \cup B first. 623 */ 624 static enum isl_change check_facets(int i, int j, 625 struct isl_coalesce_info *info) 626 { 627 int k, l; 628 struct isl_tab_undo *snap, *snap2; 629 unsigned n_eq = info[i].bmap->n_eq; 630 631 snap = isl_tab_snap(info[i].tab); 632 if (isl_tab_mark_rational(info[i].tab) < 0) 633 return isl_change_error; 634 snap2 = isl_tab_snap(info[i].tab); 635 636 for (k = 0; k < info[i].bmap->n_ineq; ++k) { 637 if (info[i].ineq[k] != STATUS_CUT) 638 continue; 639 if (isl_tab_select_facet(info[i].tab, n_eq + k) < 0) 640 return isl_change_error; 641 for (l = 0; l < info[j].bmap->n_ineq; ++l) { 642 int stat; 643 if (info[j].ineq[l] != STATUS_CUT) 644 continue; 645 stat = status_in(info[j].bmap->ineq[l], info[i].tab); 646 if (stat < 0) 647 return isl_change_error; 648 if (stat != STATUS_VALID) 649 break; 650 } 651 if (isl_tab_rollback(info[i].tab, snap2) < 0) 652 return isl_change_error; 653 if (l < info[j].bmap->n_ineq) 654 break; 655 } 656 657 if (k < info[i].bmap->n_ineq) { 658 if (isl_tab_rollback(info[i].tab, snap) < 0) 659 return isl_change_error; 660 return isl_change_none; 661 } 662 return fuse(i, j, info, NULL, 0, 0); 663 } 664 665 /* Check if info->bmap contains the basic map represented 666 * by the tableau "tab". 667 * For each equality, we check both the constraint itself 668 * (as an inequality) and its negation. Make sure the 669 * equality is returned to its original state before returning. 670 */ 671 static isl_bool contains(struct isl_coalesce_info *info, struct isl_tab *tab) 672 { 673 int k; 674 isl_size dim; 675 isl_basic_map *bmap = info->bmap; 676 677 dim = isl_basic_map_dim(bmap, isl_dim_all); 678 if (dim < 0) 679 return isl_bool_error; 680 for (k = 0; k < bmap->n_eq; ++k) { 681 int stat; 682 isl_seq_neg(bmap->eq[k], bmap->eq[k], 1 + dim); 683 stat = status_in(bmap->eq[k], tab); 684 isl_seq_neg(bmap->eq[k], bmap->eq[k], 1 + dim); 685 if (stat < 0) 686 return isl_bool_error; 687 if (stat != STATUS_VALID) 688 return isl_bool_false; 689 stat = status_in(bmap->eq[k], tab); 690 if (stat < 0) 691 return isl_bool_error; 692 if (stat != STATUS_VALID) 693 return isl_bool_false; 694 } 695 696 for (k = 0; k < bmap->n_ineq; ++k) { 697 int stat; 698 if (info->ineq[k] == STATUS_REDUNDANT) 699 continue; 700 stat = status_in(bmap->ineq[k], tab); 701 if (stat < 0) 702 return isl_bool_error; 703 if (stat != STATUS_VALID) 704 return isl_bool_false; 705 } 706 return isl_bool_true; 707 } 708 709 /* Basic map "i" has an inequality "k" that is adjacent 710 * to some inequality of basic map "j". All the other inequalities 711 * are valid for "j". 712 * If not NULL, then "extra" contains extra wrapping constraints that are valid 713 * for both "i" and "j". 714 * Check if basic map "j" forms an extension of basic map "i", 715 * taking into account the extra constraints, if any. 716 * 717 * Note that this function is only called if some of the equalities or 718 * inequalities of basic map "j" do cut basic map "i". The function is 719 * correct even if there are no such cut constraints, but in that case 720 * the additional checks performed by this function are overkill. 721 * 722 * In particular, we replace constraint k, say f >= 0, by constraint 723 * f <= -1, add the inequalities of "j" that are valid for "i", 724 * as well as the "extra" constraints, if any, 725 * and check if the result is a subset of basic map "j". 726 * To improve the chances of the subset relation being detected, 727 * any variable that only attains a single integer value 728 * in the tableau of "i" is first fixed to that value. 729 * If the result is a subset, then we know that this result is exactly equal 730 * to basic map "j" since all its constraints are valid for basic map "j". 731 * By combining the valid constraints of "i" (all equalities and all 732 * inequalities except "k"), the valid constraints of "j" and 733 * the "extra" constraints, if any, we therefore 734 * obtain a basic map that is equal to their union. 735 * In this case, there is no need to perform a rollback of the tableau 736 * since it is going to be destroyed in fuse(). 737 * 738 * 739 * |\__ |\__ 740 * | \__ | \__ 741 * | \_ => | \__ 742 * |_______| _ |_________\ 743 * 744 * 745 * |\ |\ 746 * | \ | \ 747 * | \ | \ 748 * | | | \ 749 * | ||\ => | \ 750 * | || \ | \ 751 * | || | | | 752 * |__||_/ |_____/ 753 * 754 * 755 * _______ _______ 756 * | | __ | \__ 757 * | ||__| => | __| 758 * |_______| |_______/ 759 */ 760 static enum isl_change is_adj_ineq_extension_with_wraps(int i, int j, int k, 761 struct isl_coalesce_info *info, __isl_keep isl_mat *extra) 762 { 763 struct isl_tab_undo *snap; 764 isl_size n_eq_i, n_ineq_j, n_extra; 765 isl_size total = isl_basic_map_dim(info[i].bmap, isl_dim_all); 766 isl_stat r; 767 isl_bool super; 768 769 if (total < 0) 770 return isl_change_error; 771 772 n_eq_i = isl_basic_map_n_equality(info[i].bmap); 773 n_ineq_j = isl_basic_map_n_inequality(info[j].bmap); 774 n_extra = isl_mat_rows(extra); 775 if (n_eq_i < 0 || n_ineq_j < 0 || n_extra < 0) 776 return isl_change_error; 777 778 if (isl_tab_extend_cons(info[i].tab, 1 + n_ineq_j + n_extra) < 0) 779 return isl_change_error; 780 781 snap = isl_tab_snap(info[i].tab); 782 783 if (isl_tab_unrestrict(info[i].tab, n_eq_i + k) < 0) 784 return isl_change_error; 785 786 isl_seq_neg(info[i].bmap->ineq[k], info[i].bmap->ineq[k], 1 + total); 787 isl_int_sub_ui(info[i].bmap->ineq[k][0], info[i].bmap->ineq[k][0], 1); 788 r = isl_tab_add_ineq(info[i].tab, info[i].bmap->ineq[k]); 789 isl_seq_neg(info[i].bmap->ineq[k], info[i].bmap->ineq[k], 1 + total); 790 isl_int_sub_ui(info[i].bmap->ineq[k][0], info[i].bmap->ineq[k][0], 1); 791 if (r < 0) 792 return isl_change_error; 793 794 for (k = 0; k < n_ineq_j; ++k) { 795 if (info[j].ineq[k] != STATUS_VALID) 796 continue; 797 if (isl_tab_add_ineq(info[i].tab, info[j].bmap->ineq[k]) < 0) 798 return isl_change_error; 799 } 800 for (k = 0; k < n_extra; ++k) { 801 if (isl_tab_add_ineq(info[i].tab, extra->row[k]) < 0) 802 return isl_change_error; 803 } 804 if (isl_tab_detect_constants(info[i].tab) < 0) 805 return isl_change_error; 806 807 super = contains(&info[j], info[i].tab); 808 if (super < 0) 809 return isl_change_error; 810 if (super) 811 return fuse(i, j, info, extra, 0, 0); 812 813 if (isl_tab_rollback(info[i].tab, snap) < 0) 814 return isl_change_error; 815 816 return isl_change_none; 817 } 818 819 /* Given an affine transformation matrix "T", does row "row" represent 820 * anything other than a unit vector (possibly shifted by a constant) 821 * that is not involved in any of the other rows? 822 * 823 * That is, if a constraint involves the variable corresponding to 824 * the row, then could its preimage by "T" have any coefficients 825 * that are different from those in the original constraint? 826 */ 827 static int not_unique_unit_row(__isl_keep isl_mat *T, int row) 828 { 829 int i, j; 830 int len = T->n_col - 1; 831 832 i = isl_seq_first_non_zero(T->row[row] + 1, len); 833 if (i < 0) 834 return 1; 835 if (!isl_int_is_one(T->row[row][1 + i]) && 836 !isl_int_is_negone(T->row[row][1 + i])) 837 return 1; 838 839 j = isl_seq_first_non_zero(T->row[row] + 1 + i + 1, len - (i + 1)); 840 if (j >= 0) 841 return 1; 842 843 for (j = 1; j < T->n_row; ++j) { 844 if (j == row) 845 continue; 846 if (!isl_int_is_zero(T->row[j][1 + i])) 847 return 1; 848 } 849 850 return 0; 851 } 852 853 /* Does inequality constraint "ineq" of "bmap" involve any of 854 * the variables marked in "affected"? 855 * "total" is the total number of variables, i.e., the number 856 * of entries in "affected". 857 */ 858 static isl_bool is_affected(__isl_keep isl_basic_map *bmap, int ineq, 859 int *affected, int total) 860 { 861 int i; 862 863 for (i = 0; i < total; ++i) { 864 if (!affected[i]) 865 continue; 866 if (!isl_int_is_zero(bmap->ineq[ineq][1 + i])) 867 return isl_bool_true; 868 } 869 870 return isl_bool_false; 871 } 872 873 /* Given the compressed version of inequality constraint "ineq" 874 * of info->bmap in "v", check if the constraint can be tightened, 875 * where the compression is based on an equality constraint valid 876 * for info->tab. 877 * If so, add the tightened version of the inequality constraint 878 * to info->tab. "v" may be modified by this function. 879 * 880 * That is, if the compressed constraint is of the form 881 * 882 * m f() + c >= 0 883 * 884 * with 0 < c < m, then it is equivalent to 885 * 886 * f() >= 0 887 * 888 * This means that c can also be subtracted from the original, 889 * uncompressed constraint without affecting the integer points 890 * in info->tab. Add this tightened constraint as an extra row 891 * to info->tab to make this information explicitly available. 892 */ 893 static __isl_give isl_vec *try_tightening(struct isl_coalesce_info *info, 894 int ineq, __isl_take isl_vec *v) 895 { 896 isl_ctx *ctx; 897 isl_stat r; 898 899 if (!v) 900 return NULL; 901 902 ctx = isl_vec_get_ctx(v); 903 isl_seq_gcd(v->el + 1, v->size - 1, &ctx->normalize_gcd); 904 if (isl_int_is_zero(ctx->normalize_gcd) || 905 isl_int_is_one(ctx->normalize_gcd)) { 906 return v; 907 } 908 909 v = isl_vec_cow(v); 910 if (!v) 911 return NULL; 912 913 isl_int_fdiv_r(v->el[0], v->el[0], ctx->normalize_gcd); 914 if (isl_int_is_zero(v->el[0])) 915 return v; 916 917 if (isl_tab_extend_cons(info->tab, 1) < 0) 918 return isl_vec_free(v); 919 920 isl_int_sub(info->bmap->ineq[ineq][0], 921 info->bmap->ineq[ineq][0], v->el[0]); 922 r = isl_tab_add_ineq(info->tab, info->bmap->ineq[ineq]); 923 isl_int_add(info->bmap->ineq[ineq][0], 924 info->bmap->ineq[ineq][0], v->el[0]); 925 926 if (r < 0) 927 return isl_vec_free(v); 928 929 return v; 930 } 931 932 /* Tighten the (non-redundant) constraints on the facet represented 933 * by info->tab. 934 * In particular, on input, info->tab represents the result 935 * of relaxing the "n" inequality constraints of info->bmap in "relaxed" 936 * by one, i.e., replacing f_i >= 0 by f_i + 1 >= 0, and then 937 * replacing the one at index "l" by the corresponding equality, 938 * i.e., f_k + 1 = 0, with k = relaxed[l]. 939 * 940 * Compute a variable compression from the equality constraint f_k + 1 = 0 941 * and use it to tighten the other constraints of info->bmap 942 * (that is, all constraints that have not been relaxed), 943 * updating info->tab (and leaving info->bmap untouched). 944 * The compression handles essentially two cases, one where a variable 945 * is assigned a fixed value and can therefore be eliminated, and one 946 * where one variable is a shifted multiple of some other variable and 947 * can therefore be replaced by that multiple. 948 * Gaussian elimination would also work for the first case, but for 949 * the second case, the effectiveness would depend on the order 950 * of the variables. 951 * After compression, some of the constraints may have coefficients 952 * with a common divisor. If this divisor does not divide the constant 953 * term, then the constraint can be tightened. 954 * The tightening is performed on the tableau info->tab by introducing 955 * extra (temporary) constraints. 956 * 957 * Only constraints that are possibly affected by the compression are 958 * considered. In particular, if the constraint only involves variables 959 * that are directly mapped to a distinct set of other variables, then 960 * no common divisor can be introduced and no tightening can occur. 961 * 962 * It is important to only consider the non-redundant constraints 963 * since the facet constraint has been relaxed prior to the call 964 * to this function, meaning that the constraints that were redundant 965 * prior to the relaxation may no longer be redundant. 966 * These constraints will be ignored in the fused result, so 967 * the fusion detection should not exploit them. 968 */ 969 static isl_stat tighten_on_relaxed_facet(struct isl_coalesce_info *info, 970 int n, int *relaxed, int l) 971 { 972 isl_size total; 973 isl_ctx *ctx; 974 isl_vec *v = NULL; 975 isl_mat *T; 976 int i; 977 int k; 978 int *affected; 979 980 k = relaxed[l]; 981 ctx = isl_basic_map_get_ctx(info->bmap); 982 total = isl_basic_map_dim(info->bmap, isl_dim_all); 983 if (total < 0) 984 return isl_stat_error; 985 isl_int_add_ui(info->bmap->ineq[k][0], info->bmap->ineq[k][0], 1); 986 T = isl_mat_sub_alloc6(ctx, info->bmap->ineq, k, 1, 0, 1 + total); 987 T = isl_mat_variable_compression(T, NULL); 988 isl_int_sub_ui(info->bmap->ineq[k][0], info->bmap->ineq[k][0], 1); 989 if (!T) 990 return isl_stat_error; 991 if (T->n_col == 0) { 992 isl_mat_free(T); 993 return isl_stat_ok; 994 } 995 996 affected = isl_alloc_array(ctx, int, total); 997 if (!affected) 998 goto error; 999 1000 for (i = 0; i < total; ++i) 1001 affected[i] = not_unique_unit_row(T, 1 + i); 1002 1003 for (i = 0; i < info->bmap->n_ineq; ++i) { 1004 isl_bool handle; 1005 if (any(relaxed, n, i)) 1006 continue; 1007 if (info->ineq[i] == STATUS_REDUNDANT) 1008 continue; 1009 handle = is_affected(info->bmap, i, affected, total); 1010 if (handle < 0) 1011 goto error; 1012 if (!handle) 1013 continue; 1014 v = isl_vec_alloc(ctx, 1 + total); 1015 if (!v) 1016 goto error; 1017 isl_seq_cpy(v->el, info->bmap->ineq[i], 1 + total); 1018 v = isl_vec_mat_product(v, isl_mat_copy(T)); 1019 v = try_tightening(info, i, v); 1020 isl_vec_free(v); 1021 if (!v) 1022 goto error; 1023 } 1024 1025 isl_mat_free(T); 1026 free(affected); 1027 return isl_stat_ok; 1028 error: 1029 isl_mat_free(T); 1030 free(affected); 1031 return isl_stat_error; 1032 } 1033 1034 /* Replace the basic maps "i" and "j" by an extension of "i" 1035 * along the "n" inequality constraints in "relax" by one. 1036 * The tableau info[i].tab has already been extended. 1037 * Extend info[i].bmap accordingly by relaxing all constraints in "relax" 1038 * by one. 1039 * Each integer division that does not have exactly the same 1040 * definition in "i" and "j" is marked unknown and the basic map 1041 * is scheduled to be simplified in an attempt to recover 1042 * the integer division definition. 1043 * Place the extension in the position that is the smallest of i and j. 1044 */ 1045 static enum isl_change extend(int i, int j, int n, int *relax, 1046 struct isl_coalesce_info *info) 1047 { 1048 int l; 1049 isl_size total; 1050 1051 info[i].bmap = isl_basic_map_cow(info[i].bmap); 1052 total = isl_basic_map_dim(info[i].bmap, isl_dim_all); 1053 if (total < 0) 1054 return isl_change_error; 1055 for (l = 0; l < info[i].bmap->n_div; ++l) 1056 if (!isl_seq_eq(info[i].bmap->div[l], 1057 info[j].bmap->div[l], 1 + 1 + total)) { 1058 isl_int_set_si(info[i].bmap->div[l][0], 0); 1059 info[i].simplify = 1; 1060 } 1061 for (l = 0; l < n; ++l) 1062 isl_int_add_ui(info[i].bmap->ineq[relax[l]][0], 1063 info[i].bmap->ineq[relax[l]][0], 1); 1064 ISL_F_CLR(info[i].bmap, ISL_BASIC_MAP_NO_REDUNDANT); 1065 ISL_F_SET(info[i].bmap, ISL_BASIC_MAP_FINAL); 1066 drop(&info[j]); 1067 info[i].modified = 1; 1068 if (j < i) 1069 exchange(&info[i], &info[j]); 1070 return isl_change_fuse; 1071 } 1072 1073 /* Basic map "i" has "n" inequality constraints (collected in "relax") 1074 * that are such that they include basic map "j" if they are relaxed 1075 * by one. All the other inequalities are valid for "j". 1076 * Check if basic map "j" forms an extension of basic map "i". 1077 * 1078 * In particular, relax the constraints in "relax", compute the corresponding 1079 * facets one by one and check whether each of these is included 1080 * in the other basic map. 1081 * Before testing for inclusion, the constraints on each facet 1082 * are tightened to increase the chance of an inclusion being detected. 1083 * (Adding the valid constraints of "j" to the tableau of "i", as is done 1084 * in is_adj_ineq_extension, may further increase those chances, but this 1085 * is not currently done.) 1086 * If each facet is included, we know that relaxing the constraints extends 1087 * the basic map with exactly the other basic map (we already know that this 1088 * other basic map is included in the extension, because all other 1089 * inequality constraints are valid of "j") and we can replace the 1090 * two basic maps by this extension. 1091 * 1092 * If any of the relaxed constraints turn out to be redundant, then bail out. 1093 * isl_tab_select_facet refuses to handle such constraints. It may be 1094 * possible to handle them anyway by making a distinction between 1095 * redundant constraints with a corresponding facet that still intersects 1096 * the set (allowing isl_tab_select_facet to handle them) and 1097 * those where the facet does not intersect the set (which can be ignored 1098 * because the empty facet is trivially included in the other disjunct). 1099 * However, relaxed constraints that turn out to be redundant should 1100 * be fairly rare and no such instance has been reported where 1101 * coalescing would be successful. 1102 * ____ _____ 1103 * / || / | 1104 * / || / | 1105 * \ || => \ | 1106 * \ || \ | 1107 * \___|| \____| 1108 * 1109 * 1110 * \ |\ 1111 * |\\ | \ 1112 * | \\ | \ 1113 * | | => | / 1114 * | / | / 1115 * |/ |/ 1116 */ 1117 static enum isl_change is_relaxed_extension(int i, int j, int n, int *relax, 1118 struct isl_coalesce_info *info) 1119 { 1120 int l; 1121 isl_bool super; 1122 struct isl_tab_undo *snap, *snap2; 1123 unsigned n_eq = info[i].bmap->n_eq; 1124 1125 for (l = 0; l < n; ++l) 1126 if (isl_tab_is_equality(info[i].tab, n_eq + relax[l])) 1127 return isl_change_none; 1128 1129 snap = isl_tab_snap(info[i].tab); 1130 for (l = 0; l < n; ++l) 1131 if (isl_tab_relax(info[i].tab, n_eq + relax[l]) < 0) 1132 return isl_change_error; 1133 for (l = 0; l < n; ++l) { 1134 if (!isl_tab_is_redundant(info[i].tab, n_eq + relax[l])) 1135 continue; 1136 if (isl_tab_rollback(info[i].tab, snap) < 0) 1137 return isl_change_error; 1138 return isl_change_none; 1139 } 1140 snap2 = isl_tab_snap(info[i].tab); 1141 for (l = 0; l < n; ++l) { 1142 if (isl_tab_rollback(info[i].tab, snap2) < 0) 1143 return isl_change_error; 1144 if (isl_tab_select_facet(info[i].tab, n_eq + relax[l]) < 0) 1145 return isl_change_error; 1146 if (tighten_on_relaxed_facet(&info[i], n, relax, l) < 0) 1147 return isl_change_error; 1148 super = contains(&info[j], info[i].tab); 1149 if (super < 0) 1150 return isl_change_error; 1151 if (super) 1152 continue; 1153 if (isl_tab_rollback(info[i].tab, snap) < 0) 1154 return isl_change_error; 1155 return isl_change_none; 1156 } 1157 1158 if (isl_tab_rollback(info[i].tab, snap2) < 0) 1159 return isl_change_error; 1160 return extend(i, j, n, relax, info); 1161 } 1162 1163 /* Data structure that keeps track of the wrapping constraints 1164 * and of information to bound the coefficients of those constraints. 1165 * 1166 * "failed" is set if wrapping has failed. 1167 * bound is set if we want to apply a bound on the coefficients 1168 * mat contains the wrapping constraints 1169 * max is the bound on the coefficients (if bound is set) 1170 */ 1171 struct isl_wraps { 1172 int failed; 1173 int bound; 1174 isl_mat *mat; 1175 isl_int max; 1176 }; 1177 1178 /* Update wraps->max to be greater than or equal to the coefficients 1179 * in the equalities and inequalities of info->bmap that can be removed 1180 * if we end up applying wrapping. 1181 */ 1182 static isl_stat wraps_update_max(struct isl_wraps *wraps, 1183 struct isl_coalesce_info *info) 1184 { 1185 int k; 1186 isl_int max_k; 1187 isl_size total = isl_basic_map_dim(info->bmap, isl_dim_all); 1188 1189 if (total < 0) 1190 return isl_stat_error; 1191 isl_int_init(max_k); 1192 1193 for (k = 0; k < info->bmap->n_eq; ++k) { 1194 if (info->eq[2 * k] == STATUS_VALID && 1195 info->eq[2 * k + 1] == STATUS_VALID) 1196 continue; 1197 isl_seq_abs_max(info->bmap->eq[k] + 1, total, &max_k); 1198 if (isl_int_abs_gt(max_k, wraps->max)) 1199 isl_int_set(wraps->max, max_k); 1200 } 1201 1202 for (k = 0; k < info->bmap->n_ineq; ++k) { 1203 if (info->ineq[k] == STATUS_VALID || 1204 info->ineq[k] == STATUS_REDUNDANT) 1205 continue; 1206 isl_seq_abs_max(info->bmap->ineq[k] + 1, total, &max_k); 1207 if (isl_int_abs_gt(max_k, wraps->max)) 1208 isl_int_set(wraps->max, max_k); 1209 } 1210 1211 isl_int_clear(max_k); 1212 1213 return isl_stat_ok; 1214 } 1215 1216 /* Initialize the isl_wraps data structure. 1217 * If we want to bound the coefficients of the wrapping constraints, 1218 * we set wraps->max to the largest coefficient 1219 * in the equalities and inequalities that can be removed if we end up 1220 * applying wrapping. 1221 */ 1222 static isl_stat wraps_init(struct isl_wraps *wraps, __isl_take isl_mat *mat, 1223 struct isl_coalesce_info *info, int i, int j) 1224 { 1225 isl_ctx *ctx; 1226 1227 wraps->failed = 0; 1228 wraps->bound = 0; 1229 wraps->mat = mat; 1230 if (!mat) 1231 return isl_stat_error; 1232 wraps->mat->n_row = 0; 1233 ctx = isl_mat_get_ctx(mat); 1234 wraps->bound = isl_options_get_coalesce_bounded_wrapping(ctx); 1235 if (!wraps->bound) 1236 return isl_stat_ok; 1237 isl_int_init(wraps->max); 1238 isl_int_set_si(wraps->max, 0); 1239 if (wraps_update_max(wraps, &info[i]) < 0) 1240 return isl_stat_error; 1241 if (wraps_update_max(wraps, &info[j]) < 0) 1242 return isl_stat_error; 1243 1244 return isl_stat_ok; 1245 } 1246 1247 /* Free the contents of the isl_wraps data structure. 1248 */ 1249 static void wraps_free(struct isl_wraps *wraps) 1250 { 1251 isl_mat_free(wraps->mat); 1252 if (wraps->bound) 1253 isl_int_clear(wraps->max); 1254 } 1255 1256 /* Mark the wrapping as failed. 1257 */ 1258 static isl_stat wraps_mark_failed(struct isl_wraps *wraps) 1259 { 1260 wraps->failed = 1; 1261 return isl_stat_ok; 1262 } 1263 1264 /* Is the wrapping constraint in row "row" allowed? 1265 * 1266 * If wraps->bound is set, we check that none of the coefficients 1267 * is greater than wraps->max. 1268 */ 1269 static int allow_wrap(struct isl_wraps *wraps, int row) 1270 { 1271 int i; 1272 1273 if (!wraps->bound) 1274 return 1; 1275 1276 for (i = 1; i < wraps->mat->n_col; ++i) 1277 if (isl_int_abs_gt(wraps->mat->row[row][i], wraps->max)) 1278 return 0; 1279 1280 return 1; 1281 } 1282 1283 /* Wrap "ineq" (or its opposite if "negate" is set) around "bound" 1284 * to include "set" and add the result in position "w" of "wraps". 1285 * "len" is the total number of coefficients in "bound" and "ineq". 1286 * Return 1 on success, 0 on failure and -1 on error. 1287 * Wrapping can fail if the result of wrapping is equal to "bound" 1288 * or if we want to bound the sizes of the coefficients and 1289 * the wrapped constraint does not satisfy this bound. 1290 */ 1291 static int add_wrap(struct isl_wraps *wraps, int w, isl_int *bound, 1292 isl_int *ineq, unsigned len, __isl_keep isl_set *set, int negate) 1293 { 1294 isl_seq_cpy(wraps->mat->row[w], bound, len); 1295 if (negate) { 1296 isl_seq_neg(wraps->mat->row[w + 1], ineq, len); 1297 ineq = wraps->mat->row[w + 1]; 1298 } 1299 if (!isl_set_wrap_facet(set, wraps->mat->row[w], ineq)) 1300 return -1; 1301 if (isl_seq_eq(wraps->mat->row[w], bound, len)) 1302 return 0; 1303 if (!allow_wrap(wraps, w)) 1304 return 0; 1305 return 1; 1306 } 1307 1308 /* This function has two modes of operations. 1309 * 1310 * If "add_valid" is set, then all the constraints of info->bmap 1311 * (except the opposite of "bound") are valid for the other basic map. 1312 * In this case, attempts are made to wrap some of these valid constraints 1313 * to more tightly fit around "set". Only successful wrappings are recorded 1314 * and failed wrappings are ignored. 1315 * 1316 * If "add_valid" is not set, then some of the constraints of info->bmap 1317 * are not valid for the other basic map, and only those are considered 1318 * for wrapping. In this case all attempted wrappings need to succeed. 1319 * Otherwise "wraps" is marked as failed. 1320 * Note that the constraints that are valid for the other basic map 1321 * will be added to the combined basic map by default, so there is 1322 * no need to wrap them. 1323 * The caller wrap_in_facets even relies on this function not wrapping 1324 * any constraints that are already valid. 1325 * 1326 * Only consider constraints that are not redundant (as determined 1327 * by info->tab) and that are valid or invalid depending on "add_valid". 1328 * Wrap each constraint around "bound" such that it includes the whole 1329 * set "set" and append the resulting constraint to "wraps". 1330 * "wraps" is assumed to have been pre-allocated to the appropriate size. 1331 * wraps->n_row is the number of actual wrapped constraints that have 1332 * been added. 1333 * If any of the wrapping problems results in a constraint that is 1334 * identical to "bound", then this means that "set" is unbounded in such 1335 * a way that no wrapping is possible. 1336 * Similarly, if we want to bound the coefficients of the wrapping 1337 * constraints and a newly added wrapping constraint does not 1338 * satisfy the bound, then the wrapping is considered to have failed. 1339 * Note though that "wraps" is only marked failed if "add_valid" is not set. 1340 */ 1341 static isl_stat add_selected_wraps(struct isl_wraps *wraps, 1342 struct isl_coalesce_info *info, isl_int *bound, __isl_keep isl_set *set, 1343 int add_valid) 1344 { 1345 int l, m; 1346 int w; 1347 int added; 1348 isl_basic_map *bmap = info->bmap; 1349 isl_size total = isl_basic_map_dim(bmap, isl_dim_all); 1350 unsigned len = 1 + total; 1351 1352 if (total < 0) 1353 return isl_stat_error; 1354 1355 w = wraps->mat->n_row; 1356 1357 for (l = 0; l < bmap->n_ineq; ++l) { 1358 int is_valid = info->ineq[l] == STATUS_VALID; 1359 if ((!add_valid && is_valid) || 1360 info->ineq[l] == STATUS_REDUNDANT) 1361 continue; 1362 if (isl_seq_is_neg(bound, bmap->ineq[l], len)) 1363 continue; 1364 if (isl_seq_eq(bound, bmap->ineq[l], len)) 1365 continue; 1366 if (isl_tab_is_redundant(info->tab, bmap->n_eq + l)) 1367 continue; 1368 1369 added = add_wrap(wraps, w, bound, bmap->ineq[l], len, set, 0); 1370 if (added < 0) 1371 return isl_stat_error; 1372 if (!added && !is_valid) 1373 goto unbounded; 1374 if (added) 1375 ++w; 1376 } 1377 for (l = 0; l < bmap->n_eq; ++l) { 1378 if (isl_seq_is_neg(bound, bmap->eq[l], len)) 1379 continue; 1380 if (isl_seq_eq(bound, bmap->eq[l], len)) 1381 continue; 1382 1383 for (m = 0; m < 2; ++m) { 1384 if (info->eq[2 * l + m] == STATUS_VALID) 1385 continue; 1386 added = add_wrap(wraps, w, bound, bmap->eq[l], len, 1387 set, !m); 1388 if (added < 0) 1389 return isl_stat_error; 1390 if (!added) 1391 goto unbounded; 1392 ++w; 1393 } 1394 } 1395 1396 wraps->mat->n_row = w; 1397 return isl_stat_ok; 1398 unbounded: 1399 return wraps_mark_failed(wraps); 1400 } 1401 1402 /* For each constraint in info->bmap that is not redundant (as determined 1403 * by info->tab) and that is not a valid constraint for the other basic map, 1404 * wrap the constraint around "bound" such that it includes the whole 1405 * set "set" and append the resulting constraint to "wraps". 1406 * Note that the constraints that are valid for the other basic map 1407 * will be added to the combined basic map by default, so there is 1408 * no need to wrap them. 1409 * The caller wrap_in_facets even relies on this function not wrapping 1410 * any constraints that are already valid. 1411 * "wraps" is assumed to have been pre-allocated to the appropriate size. 1412 * wraps->n_row is the number of actual wrapped constraints that have 1413 * been added. 1414 * If any of the wrapping problems results in a constraint that is 1415 * identical to "bound", then this means that "set" is unbounded in such 1416 * a way that no wrapping is possible. If this happens then "wraps" 1417 * is marked as failed. 1418 * Similarly, if we want to bound the coefficients of the wrapping 1419 * constraints and a newly added wrapping constraint does not 1420 * satisfy the bound, then "wraps" is also marked as failed. 1421 */ 1422 static isl_stat add_wraps(struct isl_wraps *wraps, 1423 struct isl_coalesce_info *info, isl_int *bound, __isl_keep isl_set *set) 1424 { 1425 return add_selected_wraps(wraps, info, bound, set, 0); 1426 } 1427 1428 /* Check if the constraints in "wraps" from "first" until the last 1429 * are all valid for the basic set represented by "tab", 1430 * dropping the invalid constraints if "keep" is set and 1431 * marking the wrapping as failed if "keep" is not set and 1432 * any constraint turns out to be invalid. 1433 */ 1434 static isl_stat check_wraps(struct isl_wraps *wraps, int first, 1435 struct isl_tab *tab, int keep) 1436 { 1437 int i; 1438 1439 for (i = wraps->mat->n_row - 1; i >= first; --i) { 1440 enum isl_ineq_type type; 1441 type = isl_tab_ineq_type(tab, wraps->mat->row[i]); 1442 if (type == isl_ineq_error) 1443 return isl_stat_error; 1444 if (type == isl_ineq_redundant) 1445 continue; 1446 if (!keep) 1447 return wraps_mark_failed(wraps); 1448 wraps->mat = isl_mat_drop_rows(wraps->mat, i, 1); 1449 if (!wraps->mat) 1450 return isl_stat_error; 1451 } 1452 1453 return isl_stat_ok; 1454 } 1455 1456 /* Return a set that corresponds to the non-redundant constraints 1457 * (as recorded in tab) of bmap. 1458 * 1459 * It's important to remove the redundant constraints as some 1460 * of the other constraints may have been modified after the 1461 * constraints were marked redundant. 1462 * In particular, a constraint may have been relaxed. 1463 * Redundant constraints are ignored when a constraint is relaxed 1464 * and should therefore continue to be ignored ever after. 1465 * Otherwise, the relaxation might be thwarted by some of 1466 * these constraints. 1467 * 1468 * Update the underlying set to ensure that the dimension doesn't change. 1469 * Otherwise the integer divisions could get dropped if the tab 1470 * turns out to be empty. 1471 */ 1472 static __isl_give isl_set *set_from_updated_bmap(__isl_keep isl_basic_map *bmap, 1473 struct isl_tab *tab) 1474 { 1475 isl_basic_set *bset; 1476 1477 bmap = isl_basic_map_copy(bmap); 1478 bset = isl_basic_map_underlying_set(bmap); 1479 bset = isl_basic_set_cow(bset); 1480 bset = isl_basic_set_update_from_tab(bset, tab); 1481 return isl_set_from_basic_set(bset); 1482 } 1483 1484 /* Does "info" have any cut constraints that are redundant? 1485 */ 1486 static isl_bool has_redundant_cuts(struct isl_coalesce_info *info) 1487 { 1488 int l; 1489 isl_size n_eq, n_ineq; 1490 1491 n_eq = isl_basic_map_n_equality(info->bmap); 1492 n_ineq = isl_basic_map_n_inequality(info->bmap); 1493 if (n_eq < 0 || n_ineq < 0) 1494 return isl_bool_error; 1495 for (l = 0; l < n_ineq; ++l) { 1496 int red; 1497 1498 if (info->ineq[l] != STATUS_CUT) 1499 continue; 1500 red = isl_tab_is_redundant(info->tab, n_eq + l); 1501 if (red < 0) 1502 return isl_bool_error; 1503 if (red) 1504 return isl_bool_true; 1505 } 1506 1507 return isl_bool_false; 1508 } 1509 1510 /* Wrap some constraints of info->bmap that bound the facet defined 1511 * by inequality "k" around (the opposite of) this inequality to 1512 * include "set". "bound" may be used to store the negated inequality. 1513 * 1514 * If "add_valid" is set, then all ridges are already valid and 1515 * the purpose is to wrap "set" more tightly. In this case, 1516 * wrapping doesn't fail, although it is possible that no constraint 1517 * gets wrapped. 1518 * 1519 * If "add_valid" is not set, then some of the ridges are cut constraints 1520 * and only those are wrapped around "set". 1521 * 1522 * Since the wrapped constraints are not guaranteed to contain the whole 1523 * of info->bmap, we check them in check_wraps. 1524 * If any of the wrapped constraints turn out to be invalid, then 1525 * check_wraps will mark "wraps" as failed if "add_valid" is not set. 1526 * If "add_valid" is set, then the offending constraints are 1527 * simply removed. 1528 * 1529 * If the facet turns out to be empty, then no wrapping can be performed. 1530 * This is considered a failure, unless "add_valid" is set. 1531 * 1532 * If any of the cut constraints of info->bmap turn out 1533 * to be redundant with respect to other constraints 1534 * then these will neither be wrapped nor added directly to the result. 1535 * The result may therefore not be correct. 1536 * Skip wrapping and mark "wraps" as failed in this case. 1537 */ 1538 static isl_stat add_selected_wraps_around_facet(struct isl_wraps *wraps, 1539 struct isl_coalesce_info *info, int k, isl_int *bound, 1540 __isl_keep isl_set *set, int add_valid) 1541 { 1542 isl_bool nowrap; 1543 struct isl_tab_undo *snap; 1544 int n; 1545 isl_size total = isl_basic_map_dim(info->bmap, isl_dim_all); 1546 1547 if (total < 0) 1548 return isl_stat_error; 1549 1550 snap = isl_tab_snap(info->tab); 1551 1552 if (isl_tab_select_facet(info->tab, info->bmap->n_eq + k) < 0) 1553 return isl_stat_error; 1554 if (isl_tab_detect_redundant(info->tab) < 0) 1555 return isl_stat_error; 1556 if (info->tab->empty) { 1557 if (isl_tab_rollback(info->tab, snap) < 0) 1558 return isl_stat_error; 1559 if (!add_valid) 1560 return wraps_mark_failed(wraps); 1561 return isl_stat_ok; 1562 } 1563 nowrap = has_redundant_cuts(info); 1564 if (nowrap < 0) 1565 return isl_stat_error; 1566 1567 n = wraps->mat->n_row; 1568 if (!nowrap) { 1569 isl_seq_neg(bound, info->bmap->ineq[k], 1 + total); 1570 1571 if (add_selected_wraps(wraps, info, bound, set, add_valid) < 0) 1572 return isl_stat_error; 1573 } 1574 1575 if (isl_tab_rollback(info->tab, snap) < 0) 1576 return isl_stat_error; 1577 if (nowrap) 1578 return wraps_mark_failed(wraps); 1579 if (check_wraps(wraps, n, info->tab, add_valid) < 0) 1580 return isl_stat_error; 1581 1582 return isl_stat_ok; 1583 } 1584 1585 /* Wrap the constraints of info->bmap that bound the facet defined 1586 * by inequality "k" around (the opposite of) this inequality to 1587 * include "set". "bound" may be used to store the negated inequality. 1588 * If any of the wrapped constraints turn out to be invalid for info->bmap 1589 * itself, then mark "wraps" as failed. 1590 */ 1591 static isl_stat add_wraps_around_facet(struct isl_wraps *wraps, 1592 struct isl_coalesce_info *info, int k, isl_int *bound, 1593 __isl_keep isl_set *set) 1594 { 1595 return add_selected_wraps_around_facet(wraps, info, k, bound, set, 0); 1596 } 1597 1598 /* Wrap the (valid) constraints of info->bmap that bound the facet defined 1599 * by inequality "k" around (the opposite of) this inequality to 1600 * include "set" more tightly. 1601 * "bound" may be used to store the negated inequality. 1602 * Remove any wrapping constraints that turn out to be invalid 1603 * for info->bmap itself. 1604 */ 1605 static isl_stat add_valid_wraps_around_facet(struct isl_wraps *wraps, 1606 struct isl_coalesce_info *info, int k, isl_int *bound, 1607 __isl_keep isl_set *set) 1608 { 1609 return add_selected_wraps_around_facet(wraps, info, k, bound, set, 1); 1610 } 1611 1612 /* Basic map "i" has an inequality (say "k") that is adjacent 1613 * to some inequality of basic map "j". All the other inequalities 1614 * are valid for "j". 1615 * Check if basic map "j" forms an extension of basic map "i". 1616 * 1617 * Note that this function is only called if some of the equalities or 1618 * inequalities of basic map "j" do cut basic map "i". The function is 1619 * correct even if there are no such cut constraints, but in that case 1620 * the additional checks performed by this function are overkill. 1621 * 1622 * First try and wrap the ridges of "k" around "j". 1623 * Note that those ridges are already valid for "j", 1624 * but the wrapped versions may wrap "j" more tightly, 1625 * increasing the chances of "j" being detected as an extension of "i" 1626 */ 1627 static enum isl_change is_adj_ineq_extension(int i, int j, 1628 struct isl_coalesce_info *info) 1629 { 1630 int k; 1631 enum isl_change change; 1632 isl_size total; 1633 isl_size n_eq_i, n_ineq_i; 1634 struct isl_wraps wraps; 1635 isl_ctx *ctx; 1636 isl_mat *mat; 1637 isl_vec *bound; 1638 isl_set *set_j; 1639 isl_stat r; 1640 1641 k = find_ineq(&info[i], STATUS_ADJ_INEQ); 1642 if (k < 0) 1643 isl_die(isl_basic_map_get_ctx(info[i].bmap), isl_error_internal, 1644 "info[i].ineq should have exactly one STATUS_ADJ_INEQ", 1645 return isl_change_error); 1646 1647 total = isl_basic_map_dim(info[i].bmap, isl_dim_all); 1648 n_eq_i = isl_basic_map_n_equality(info[i].bmap); 1649 n_ineq_i = isl_basic_map_n_inequality(info[i].bmap); 1650 if (total < 0 || n_eq_i < 0 || n_ineq_i < 0) 1651 return isl_change_error; 1652 1653 set_j = set_from_updated_bmap(info[j].bmap, info[j].tab); 1654 ctx = isl_basic_map_get_ctx(info[i].bmap); 1655 bound = isl_vec_alloc(ctx, 1 + total); 1656 mat = isl_mat_alloc(ctx, 2 * n_eq_i + n_ineq_i, 1 + total); 1657 if (wraps_init(&wraps, mat, info, i, j) < 0) 1658 goto error; 1659 if (!bound || !set_j) 1660 goto error; 1661 r = add_valid_wraps_around_facet(&wraps, &info[i], k, bound->el, set_j); 1662 if (r < 0) 1663 goto error; 1664 1665 change = is_adj_ineq_extension_with_wraps(i, j, k, info, wraps.mat); 1666 1667 wraps_free(&wraps); 1668 isl_vec_free(bound); 1669 isl_set_free(set_j); 1670 1671 return change; 1672 error: 1673 wraps_free(&wraps); 1674 isl_vec_free(bound); 1675 isl_set_free(set_j); 1676 return isl_change_error; 1677 } 1678 1679 /* Both basic maps have at least one inequality with and adjacent 1680 * (but opposite) inequality in the other basic map. 1681 * Check that there are no cut constraints and that there is only 1682 * a single pair of adjacent inequalities. 1683 * If so, we can replace the pair by a single basic map described 1684 * by all but the pair of adjacent inequalities. 1685 * Any additional points introduced lie strictly between the two 1686 * adjacent hyperplanes and can therefore be integral. 1687 * 1688 * ____ _____ 1689 * / ||\ / \ 1690 * / || \ / \ 1691 * \ || \ => \ \ 1692 * \ || / \ / 1693 * \___||_/ \_____/ 1694 * 1695 * The test for a single pair of adjacent inequalities is important 1696 * for avoiding the combination of two basic maps like the following 1697 * 1698 * /| 1699 * / | 1700 * /__| 1701 * _____ 1702 * | | 1703 * | | 1704 * |___| 1705 * 1706 * If there are some cut constraints on one side, then we may 1707 * still be able to fuse the two basic maps, but we need to perform 1708 * some additional checks in is_adj_ineq_extension. 1709 */ 1710 static enum isl_change check_adj_ineq(int i, int j, 1711 struct isl_coalesce_info *info) 1712 { 1713 int count_i, count_j; 1714 int cut_i, cut_j; 1715 1716 count_i = count_ineq(&info[i], STATUS_ADJ_INEQ); 1717 count_j = count_ineq(&info[j], STATUS_ADJ_INEQ); 1718 1719 if (count_i != 1 && count_j != 1) 1720 return isl_change_none; 1721 1722 cut_i = any_eq(&info[i], STATUS_CUT) || any_ineq(&info[i], STATUS_CUT); 1723 cut_j = any_eq(&info[j], STATUS_CUT) || any_ineq(&info[j], STATUS_CUT); 1724 1725 if (!cut_i && !cut_j && count_i == 1 && count_j == 1) 1726 return fuse(i, j, info, NULL, 0, 0); 1727 1728 if (count_i == 1 && !cut_i) 1729 return is_adj_ineq_extension(i, j, info); 1730 1731 if (count_j == 1 && !cut_j) 1732 return is_adj_ineq_extension(j, i, info); 1733 1734 return isl_change_none; 1735 } 1736 1737 /* Given a basic set i with a constraint k that is adjacent to 1738 * basic set j, check if we can wrap 1739 * both the facet corresponding to k (if "wrap_facet" is set) and basic map j 1740 * (always) around their ridges to include the other set. 1741 * If so, replace the pair of basic sets by their union. 1742 * 1743 * All constraints of i (except k) are assumed to be valid or 1744 * cut constraints for j. 1745 * Wrapping the cut constraints to include basic map j may result 1746 * in constraints that are no longer valid of basic map i 1747 * we have to check that the resulting wrapping constraints are valid for i. 1748 * If "wrap_facet" is not set, then all constraints of i (except k) 1749 * are assumed to be valid for j. 1750 * ____ _____ 1751 * / | / \ 1752 * / || / | 1753 * \ || => \ | 1754 * \ || \ | 1755 * \___|| \____| 1756 * 1757 */ 1758 static enum isl_change can_wrap_in_facet(int i, int j, int k, 1759 struct isl_coalesce_info *info, int wrap_facet) 1760 { 1761 enum isl_change change = isl_change_none; 1762 struct isl_wraps wraps; 1763 isl_ctx *ctx; 1764 isl_mat *mat; 1765 struct isl_set *set_i = NULL; 1766 struct isl_set *set_j = NULL; 1767 struct isl_vec *bound = NULL; 1768 isl_size total = isl_basic_map_dim(info[i].bmap, isl_dim_all); 1769 1770 if (total < 0) 1771 return isl_change_error; 1772 set_i = set_from_updated_bmap(info[i].bmap, info[i].tab); 1773 set_j = set_from_updated_bmap(info[j].bmap, info[j].tab); 1774 ctx = isl_basic_map_get_ctx(info[i].bmap); 1775 mat = isl_mat_alloc(ctx, 2 * (info[i].bmap->n_eq + info[j].bmap->n_eq) + 1776 info[i].bmap->n_ineq + info[j].bmap->n_ineq, 1777 1 + total); 1778 if (wraps_init(&wraps, mat, info, i, j) < 0) 1779 goto error; 1780 bound = isl_vec_alloc(ctx, 1 + total); 1781 if (!set_i || !set_j || !bound) 1782 goto error; 1783 1784 isl_seq_cpy(bound->el, info[i].bmap->ineq[k], 1 + total); 1785 isl_int_add_ui(bound->el[0], bound->el[0], 1); 1786 isl_seq_normalize(ctx, bound->el, 1 + total); 1787 1788 isl_seq_cpy(wraps.mat->row[0], bound->el, 1 + total); 1789 wraps.mat->n_row = 1; 1790 1791 if (add_wraps(&wraps, &info[j], bound->el, set_i) < 0) 1792 goto error; 1793 if (wraps.failed) 1794 goto unbounded; 1795 1796 if (wrap_facet) { 1797 if (add_wraps_around_facet(&wraps, &info[i], k, 1798 bound->el, set_j) < 0) 1799 goto error; 1800 if (wraps.failed) 1801 goto unbounded; 1802 } 1803 1804 change = fuse(i, j, info, wraps.mat, 0, 0); 1805 1806 unbounded: 1807 wraps_free(&wraps); 1808 1809 isl_set_free(set_i); 1810 isl_set_free(set_j); 1811 1812 isl_vec_free(bound); 1813 1814 return change; 1815 error: 1816 wraps_free(&wraps); 1817 isl_vec_free(bound); 1818 isl_set_free(set_i); 1819 isl_set_free(set_j); 1820 return isl_change_error; 1821 } 1822 1823 /* Given a cut constraint t(x) >= 0 of basic map i, stored in row "w" 1824 * of wrap.mat, replace it by its relaxed version t(x) + 1 >= 0, and 1825 * add wrapping constraints to wrap.mat for all constraints 1826 * of basic map j that bound the part of basic map j that sticks out 1827 * of the cut constraint. 1828 * "set_i" is the underlying set of basic map i. 1829 * If any wrapping fails, then wraps->mat.n_row is reset to zero. 1830 * 1831 * In particular, we first intersect basic map j with t(x) + 1 = 0. 1832 * If the result is empty, then t(x) >= 0 was actually a valid constraint 1833 * (with respect to the integer points), so we add t(x) >= 0 instead. 1834 * Otherwise, we wrap the constraints of basic map j that are not 1835 * redundant in this intersection and that are not already valid 1836 * for basic map i over basic map i. 1837 * Note that it is sufficient to wrap the constraints to include 1838 * basic map i, because we will only wrap the constraints that do 1839 * not include basic map i already. The wrapped constraint will 1840 * therefore be more relaxed compared to the original constraint. 1841 * Since the original constraint is valid for basic map j, so is 1842 * the wrapped constraint. 1843 */ 1844 static isl_stat wrap_in_facet(struct isl_wraps *wraps, int w, 1845 struct isl_coalesce_info *info_j, __isl_keep isl_set *set_i, 1846 struct isl_tab_undo *snap) 1847 { 1848 isl_int_add_ui(wraps->mat->row[w][0], wraps->mat->row[w][0], 1); 1849 if (isl_tab_add_eq(info_j->tab, wraps->mat->row[w]) < 0) 1850 return isl_stat_error; 1851 if (isl_tab_detect_redundant(info_j->tab) < 0) 1852 return isl_stat_error; 1853 1854 if (info_j->tab->empty) 1855 isl_int_sub_ui(wraps->mat->row[w][0], wraps->mat->row[w][0], 1); 1856 else if (add_wraps(wraps, info_j, wraps->mat->row[w], set_i) < 0) 1857 return isl_stat_error; 1858 1859 if (isl_tab_rollback(info_j->tab, snap) < 0) 1860 return isl_stat_error; 1861 1862 return isl_stat_ok; 1863 } 1864 1865 /* Given a pair of basic maps i and j such that j sticks out 1866 * of i at n cut constraints, each time by at most one, 1867 * try to compute wrapping constraints and replace the two 1868 * basic maps by a single basic map. 1869 * The other constraints of i are assumed to be valid for j. 1870 * "set_i" is the underlying set of basic map i. 1871 * "wraps" has been initialized to be of the right size. 1872 * 1873 * For each cut constraint t(x) >= 0 of i, we add the relaxed version 1874 * t(x) + 1 >= 0, along with wrapping constraints for all constraints 1875 * of basic map j that bound the part of basic map j that sticks out 1876 * of the cut constraint. 1877 * 1878 * If any wrapping fails, i.e., if we cannot wrap to touch 1879 * the union, then we give up. 1880 * Otherwise, the pair of basic maps is replaced by their union. 1881 */ 1882 static enum isl_change try_wrap_in_facets(int i, int j, 1883 struct isl_coalesce_info *info, struct isl_wraps *wraps, 1884 __isl_keep isl_set *set_i) 1885 { 1886 int k, l, w; 1887 isl_size total; 1888 struct isl_tab_undo *snap; 1889 1890 total = isl_basic_map_dim(info[i].bmap, isl_dim_all); 1891 if (total < 0) 1892 return isl_change_error; 1893 1894 snap = isl_tab_snap(info[j].tab); 1895 1896 for (k = 0; k < info[i].bmap->n_eq; ++k) { 1897 for (l = 0; l < 2; ++l) { 1898 if (info[i].eq[2 * k + l] != STATUS_CUT) 1899 continue; 1900 w = wraps->mat->n_row++; 1901 if (l == 0) 1902 isl_seq_neg(wraps->mat->row[w], 1903 info[i].bmap->eq[k], 1 + total); 1904 else 1905 isl_seq_cpy(wraps->mat->row[w], 1906 info[i].bmap->eq[k], 1 + total); 1907 if (wrap_in_facet(wraps, w, &info[j], set_i, snap) < 0) 1908 return isl_change_error; 1909 1910 if (wraps->failed) 1911 return isl_change_none; 1912 } 1913 } 1914 1915 for (k = 0; k < info[i].bmap->n_ineq; ++k) { 1916 if (info[i].ineq[k] != STATUS_CUT) 1917 continue; 1918 w = wraps->mat->n_row++; 1919 isl_seq_cpy(wraps->mat->row[w], 1920 info[i].bmap->ineq[k], 1 + total); 1921 if (wrap_in_facet(wraps, w, &info[j], set_i, snap) < 0) 1922 return isl_change_error; 1923 1924 if (wraps->failed) 1925 return isl_change_none; 1926 } 1927 1928 return fuse(i, j, info, wraps->mat, 0, 1); 1929 } 1930 1931 /* Given a pair of basic maps i and j such that j sticks out 1932 * of i at n cut constraints, each time by at most one, 1933 * try to compute wrapping constraints and replace the two 1934 * basic maps by a single basic map. 1935 * The other constraints of i are assumed to be valid for j. 1936 * 1937 * The core computation is performed by try_wrap_in_facets. 1938 * This function simply extracts an underlying set representation 1939 * of basic map i and initializes the data structure for keeping 1940 * track of wrapping constraints. 1941 */ 1942 static enum isl_change wrap_in_facets(int i, int j, int n, 1943 struct isl_coalesce_info *info) 1944 { 1945 enum isl_change change = isl_change_none; 1946 struct isl_wraps wraps; 1947 isl_ctx *ctx; 1948 isl_mat *mat; 1949 isl_set *set_i = NULL; 1950 isl_size total = isl_basic_map_dim(info[i].bmap, isl_dim_all); 1951 int max_wrap; 1952 1953 if (total < 0) 1954 return isl_change_error; 1955 if (isl_tab_extend_cons(info[j].tab, 1) < 0) 1956 return isl_change_error; 1957 1958 max_wrap = 1 + 2 * info[j].bmap->n_eq + info[j].bmap->n_ineq; 1959 max_wrap *= n; 1960 1961 set_i = set_from_updated_bmap(info[i].bmap, info[i].tab); 1962 ctx = isl_basic_map_get_ctx(info[i].bmap); 1963 mat = isl_mat_alloc(ctx, max_wrap, 1 + total); 1964 if (wraps_init(&wraps, mat, info, i, j) < 0) 1965 goto error; 1966 if (!set_i) 1967 goto error; 1968 1969 change = try_wrap_in_facets(i, j, info, &wraps, set_i); 1970 1971 wraps_free(&wraps); 1972 isl_set_free(set_i); 1973 1974 return change; 1975 error: 1976 wraps_free(&wraps); 1977 isl_set_free(set_i); 1978 return isl_change_error; 1979 } 1980 1981 /* Return the effect of inequality "ineq" on the tableau "tab", 1982 * after relaxing the constant term of "ineq" by one. 1983 */ 1984 static enum isl_ineq_type type_of_relaxed(struct isl_tab *tab, isl_int *ineq) 1985 { 1986 enum isl_ineq_type type; 1987 1988 isl_int_add_ui(ineq[0], ineq[0], 1); 1989 type = isl_tab_ineq_type(tab, ineq); 1990 isl_int_sub_ui(ineq[0], ineq[0], 1); 1991 1992 return type; 1993 } 1994 1995 /* Given two basic sets i and j, 1996 * check if relaxing all the cut constraints of i by one turns 1997 * them into valid constraint for j and check if we can wrap in 1998 * the bits that are sticking out. 1999 * If so, replace the pair by their union. 2000 * 2001 * We first check if all relaxed cut inequalities of i are valid for j 2002 * and then try to wrap in the intersections of the relaxed cut inequalities 2003 * with j. 2004 * 2005 * During this wrapping, we consider the points of j that lie at a distance 2006 * of exactly 1 from i. In particular, we ignore the points that lie in 2007 * between this lower-dimensional space and the basic map i. 2008 * We can therefore only apply this to integer maps. 2009 * ____ _____ 2010 * / ___|_ / \ 2011 * / | | / | 2012 * \ | | => \ | 2013 * \|____| \ | 2014 * \___| \____/ 2015 * 2016 * _____ ______ 2017 * | ____|_ | \ 2018 * | | | | | 2019 * | | | => | | 2020 * |_| | | | 2021 * |_____| \______| 2022 * 2023 * _______ 2024 * | | 2025 * | |\ | 2026 * | | \ | 2027 * | | \ | 2028 * | | \| 2029 * | | \ 2030 * | |_____\ 2031 * | | 2032 * |_______| 2033 * 2034 * Wrapping can fail if the result of wrapping one of the facets 2035 * around its edges does not produce any new facet constraint. 2036 * In particular, this happens when we try to wrap in unbounded sets. 2037 * 2038 * _______________________________________________________________________ 2039 * | 2040 * | ___ 2041 * | | | 2042 * |_| |_________________________________________________________________ 2043 * |___| 2044 * 2045 * The following is not an acceptable result of coalescing the above two 2046 * sets as it includes extra integer points. 2047 * _______________________________________________________________________ 2048 * | 2049 * | 2050 * | 2051 * | 2052 * \______________________________________________________________________ 2053 */ 2054 static enum isl_change can_wrap_in_set(int i, int j, 2055 struct isl_coalesce_info *info) 2056 { 2057 int k, l; 2058 int n; 2059 isl_size total; 2060 2061 if (ISL_F_ISSET(info[i].bmap, ISL_BASIC_MAP_RATIONAL) || 2062 ISL_F_ISSET(info[j].bmap, ISL_BASIC_MAP_RATIONAL)) 2063 return isl_change_none; 2064 2065 n = count_eq(&info[i], STATUS_CUT) + count_ineq(&info[i], STATUS_CUT); 2066 if (n == 0) 2067 return isl_change_none; 2068 2069 total = isl_basic_map_dim(info[i].bmap, isl_dim_all); 2070 if (total < 0) 2071 return isl_change_error; 2072 for (k = 0; k < info[i].bmap->n_eq; ++k) { 2073 for (l = 0; l < 2; ++l) { 2074 enum isl_ineq_type type; 2075 2076 if (info[i].eq[2 * k + l] != STATUS_CUT) 2077 continue; 2078 2079 if (l == 0) 2080 isl_seq_neg(info[i].bmap->eq[k], 2081 info[i].bmap->eq[k], 1 + total); 2082 type = type_of_relaxed(info[j].tab, 2083 info[i].bmap->eq[k]); 2084 if (l == 0) 2085 isl_seq_neg(info[i].bmap->eq[k], 2086 info[i].bmap->eq[k], 1 + total); 2087 if (type == isl_ineq_error) 2088 return isl_change_error; 2089 if (type != isl_ineq_redundant) 2090 return isl_change_none; 2091 } 2092 } 2093 2094 for (k = 0; k < info[i].bmap->n_ineq; ++k) { 2095 enum isl_ineq_type type; 2096 2097 if (info[i].ineq[k] != STATUS_CUT) 2098 continue; 2099 2100 type = type_of_relaxed(info[j].tab, info[i].bmap->ineq[k]); 2101 if (type == isl_ineq_error) 2102 return isl_change_error; 2103 if (type != isl_ineq_redundant) 2104 return isl_change_none; 2105 } 2106 2107 return wrap_in_facets(i, j, n, info); 2108 } 2109 2110 /* Check if either i or j has only cut constraints that can 2111 * be used to wrap in (a facet of) the other basic set. 2112 * if so, replace the pair by their union. 2113 */ 2114 static enum isl_change check_wrap(int i, int j, struct isl_coalesce_info *info) 2115 { 2116 enum isl_change change = isl_change_none; 2117 2118 change = can_wrap_in_set(i, j, info); 2119 if (change != isl_change_none) 2120 return change; 2121 2122 change = can_wrap_in_set(j, i, info); 2123 return change; 2124 } 2125 2126 /* Check if all inequality constraints of "i" that cut "j" cease 2127 * to be cut constraints if they are relaxed by one. 2128 * If so, collect the cut constraints in "list". 2129 * The caller is responsible for allocating "list". 2130 */ 2131 static isl_bool all_cut_by_one(int i, int j, struct isl_coalesce_info *info, 2132 int *list) 2133 { 2134 int l, n; 2135 2136 n = 0; 2137 for (l = 0; l < info[i].bmap->n_ineq; ++l) { 2138 enum isl_ineq_type type; 2139 2140 if (info[i].ineq[l] != STATUS_CUT) 2141 continue; 2142 type = type_of_relaxed(info[j].tab, info[i].bmap->ineq[l]); 2143 if (type == isl_ineq_error) 2144 return isl_bool_error; 2145 if (type != isl_ineq_redundant) 2146 return isl_bool_false; 2147 list[n++] = l; 2148 } 2149 2150 return isl_bool_true; 2151 } 2152 2153 /* Given two basic maps such that "j" has at least one equality constraint 2154 * that is adjacent to an inequality constraint of "i" and such that "i" has 2155 * exactly one inequality constraint that is adjacent to an equality 2156 * constraint of "j", check whether "i" can be extended to include "j" or 2157 * whether "j" can be wrapped into "i". 2158 * All remaining constraints of "i" and "j" are assumed to be valid 2159 * or cut constraints of the other basic map. 2160 * However, none of the equality constraints of "i" are cut constraints. 2161 * 2162 * If "i" has any "cut" inequality constraints, then check if relaxing 2163 * each of them by one is sufficient for them to become valid. 2164 * If so, check if the inequality constraint adjacent to an equality 2165 * constraint of "j" along with all these cut constraints 2166 * can be relaxed by one to contain exactly "j". 2167 * Otherwise, or if this fails, check if "j" can be wrapped into "i". 2168 */ 2169 static enum isl_change check_single_adj_eq(int i, int j, 2170 struct isl_coalesce_info *info) 2171 { 2172 enum isl_change change = isl_change_none; 2173 int k; 2174 int n_cut; 2175 int *relax; 2176 isl_ctx *ctx; 2177 isl_bool try_relax; 2178 2179 n_cut = count_ineq(&info[i], STATUS_CUT); 2180 2181 k = find_ineq(&info[i], STATUS_ADJ_EQ); 2182 2183 if (n_cut > 0) { 2184 ctx = isl_basic_map_get_ctx(info[i].bmap); 2185 relax = isl_calloc_array(ctx, int, 1 + n_cut); 2186 if (!relax) 2187 return isl_change_error; 2188 relax[0] = k; 2189 try_relax = all_cut_by_one(i, j, info, relax + 1); 2190 if (try_relax < 0) 2191 change = isl_change_error; 2192 } else { 2193 try_relax = isl_bool_true; 2194 relax = &k; 2195 } 2196 if (try_relax && change == isl_change_none) 2197 change = is_relaxed_extension(i, j, 1 + n_cut, relax, info); 2198 if (n_cut > 0) 2199 free(relax); 2200 if (change != isl_change_none) 2201 return change; 2202 2203 change = can_wrap_in_facet(i, j, k, info, n_cut > 0); 2204 2205 return change; 2206 } 2207 2208 /* At least one of the basic maps has an equality that is adjacent 2209 * to an inequality. Make sure that only one of the basic maps has 2210 * such an equality and that the other basic map has exactly one 2211 * inequality adjacent to an equality. 2212 * If the other basic map does not have such an inequality, then 2213 * check if all its constraints are either valid or cut constraints 2214 * and, if so, try wrapping in the first map into the second. 2215 * Otherwise, try to extend one basic map with the other or 2216 * wrap one basic map in the other. 2217 */ 2218 static enum isl_change check_adj_eq(int i, int j, 2219 struct isl_coalesce_info *info) 2220 { 2221 if (any_eq(&info[i], STATUS_ADJ_INEQ) && 2222 any_eq(&info[j], STATUS_ADJ_INEQ)) 2223 /* ADJ EQ TOO MANY */ 2224 return isl_change_none; 2225 2226 if (any_eq(&info[i], STATUS_ADJ_INEQ)) 2227 return check_adj_eq(j, i, info); 2228 2229 /* j has an equality adjacent to an inequality in i */ 2230 2231 if (count_ineq(&info[i], STATUS_ADJ_EQ) != 1) { 2232 if (all_valid_or_cut(&info[i])) 2233 return can_wrap_in_set(i, j, info); 2234 return isl_change_none; 2235 } 2236 if (any_eq(&info[i], STATUS_CUT)) 2237 return isl_change_none; 2238 if (any_ineq(&info[j], STATUS_ADJ_EQ) || 2239 any_ineq(&info[i], STATUS_ADJ_INEQ) || 2240 any_ineq(&info[j], STATUS_ADJ_INEQ)) 2241 /* ADJ EQ TOO MANY */ 2242 return isl_change_none; 2243 2244 return check_single_adj_eq(i, j, info); 2245 } 2246 2247 /* Disjunct "j" lies on a hyperplane that is adjacent to disjunct "i". 2248 * In particular, disjunct "i" has an inequality constraint that is adjacent 2249 * to a (combination of) equality constraint(s) of disjunct "j", 2250 * but disjunct "j" has no explicit equality constraint adjacent 2251 * to an inequality constraint of disjunct "i". 2252 * 2253 * Disjunct "i" is already known not to have any equality constraints 2254 * that are adjacent to an equality or inequality constraint. 2255 * Check that, other than the inequality constraint mentioned above, 2256 * all other constraints of disjunct "i" are valid for disjunct "j". 2257 * If so, try and wrap in disjunct "j". 2258 */ 2259 static enum isl_change check_ineq_adj_eq(int i, int j, 2260 struct isl_coalesce_info *info) 2261 { 2262 int k; 2263 2264 if (any_eq(&info[i], STATUS_CUT)) 2265 return isl_change_none; 2266 if (any_ineq(&info[i], STATUS_CUT)) 2267 return isl_change_none; 2268 if (any_ineq(&info[i], STATUS_ADJ_INEQ)) 2269 return isl_change_none; 2270 if (count_ineq(&info[i], STATUS_ADJ_EQ) != 1) 2271 return isl_change_none; 2272 2273 k = find_ineq(&info[i], STATUS_ADJ_EQ); 2274 2275 return can_wrap_in_facet(i, j, k, info, 0); 2276 } 2277 2278 /* The two basic maps lie on adjacent hyperplanes. In particular, 2279 * basic map "i" has an equality that lies parallel to basic map "j". 2280 * Check if we can wrap the facets around the parallel hyperplanes 2281 * to include the other set. 2282 * 2283 * We perform basically the same operations as can_wrap_in_facet, 2284 * except that we don't need to select a facet of one of the sets. 2285 * _ 2286 * \\ \\ 2287 * \\ => \\ 2288 * \ \| 2289 * 2290 * If there is more than one equality of "i" adjacent to an equality of "j", 2291 * then the result will satisfy one or more equalities that are a linear 2292 * combination of these equalities. These will be encoded as pairs 2293 * of inequalities in the wrapping constraints and need to be made 2294 * explicit. 2295 */ 2296 static enum isl_change check_eq_adj_eq(int i, int j, 2297 struct isl_coalesce_info *info) 2298 { 2299 int k; 2300 enum isl_change change = isl_change_none; 2301 int detect_equalities = 0; 2302 struct isl_wraps wraps; 2303 isl_ctx *ctx; 2304 isl_mat *mat; 2305 struct isl_set *set_i = NULL; 2306 struct isl_set *set_j = NULL; 2307 struct isl_vec *bound = NULL; 2308 isl_size total = isl_basic_map_dim(info[i].bmap, isl_dim_all); 2309 2310 if (total < 0) 2311 return isl_change_error; 2312 if (count_eq(&info[i], STATUS_ADJ_EQ) != 1) 2313 detect_equalities = 1; 2314 2315 k = find_eq(&info[i], STATUS_ADJ_EQ); 2316 2317 set_i = set_from_updated_bmap(info[i].bmap, info[i].tab); 2318 set_j = set_from_updated_bmap(info[j].bmap, info[j].tab); 2319 ctx = isl_basic_map_get_ctx(info[i].bmap); 2320 mat = isl_mat_alloc(ctx, 2 * (info[i].bmap->n_eq + info[j].bmap->n_eq) + 2321 info[i].bmap->n_ineq + info[j].bmap->n_ineq, 2322 1 + total); 2323 if (wraps_init(&wraps, mat, info, i, j) < 0) 2324 goto error; 2325 bound = isl_vec_alloc(ctx, 1 + total); 2326 if (!set_i || !set_j || !bound) 2327 goto error; 2328 2329 if (k % 2 == 0) 2330 isl_seq_neg(bound->el, info[i].bmap->eq[k / 2], 1 + total); 2331 else 2332 isl_seq_cpy(bound->el, info[i].bmap->eq[k / 2], 1 + total); 2333 isl_int_add_ui(bound->el[0], bound->el[0], 1); 2334 2335 isl_seq_cpy(wraps.mat->row[0], bound->el, 1 + total); 2336 wraps.mat->n_row = 1; 2337 2338 if (add_wraps(&wraps, &info[j], bound->el, set_i) < 0) 2339 goto error; 2340 if (wraps.failed) 2341 goto unbounded; 2342 2343 isl_int_sub_ui(bound->el[0], bound->el[0], 1); 2344 isl_seq_neg(bound->el, bound->el, 1 + total); 2345 2346 isl_seq_cpy(wraps.mat->row[wraps.mat->n_row], bound->el, 1 + total); 2347 wraps.mat->n_row++; 2348 2349 if (add_wraps(&wraps, &info[i], bound->el, set_j) < 0) 2350 goto error; 2351 if (wraps.failed) 2352 goto unbounded; 2353 2354 change = fuse(i, j, info, wraps.mat, detect_equalities, 0); 2355 2356 if (0) { 2357 error: change = isl_change_error; 2358 } 2359 unbounded: 2360 2361 wraps_free(&wraps); 2362 isl_set_free(set_i); 2363 isl_set_free(set_j); 2364 isl_vec_free(bound); 2365 2366 return change; 2367 } 2368 2369 /* Initialize the "eq" and "ineq" fields of "info". 2370 */ 2371 static void init_status(struct isl_coalesce_info *info) 2372 { 2373 info->eq = info->ineq = NULL; 2374 } 2375 2376 /* Set info->eq to the positions of the equalities of info->bmap 2377 * with respect to the basic map represented by "tab". 2378 * If info->eq has already been computed, then do not compute it again. 2379 */ 2380 static void set_eq_status_in(struct isl_coalesce_info *info, 2381 struct isl_tab *tab) 2382 { 2383 if (info->eq) 2384 return; 2385 info->eq = eq_status_in(info->bmap, tab); 2386 } 2387 2388 /* Set info->ineq to the positions of the inequalities of info->bmap 2389 * with respect to the basic map represented by "tab". 2390 * If info->ineq has already been computed, then do not compute it again. 2391 */ 2392 static void set_ineq_status_in(struct isl_coalesce_info *info, 2393 struct isl_tab *tab) 2394 { 2395 if (info->ineq) 2396 return; 2397 info->ineq = ineq_status_in(info->bmap, info->tab, tab); 2398 } 2399 2400 /* Free the memory allocated by the "eq" and "ineq" fields of "info". 2401 * This function assumes that init_status has been called on "info" first, 2402 * after which the "eq" and "ineq" fields may or may not have been 2403 * assigned a newly allocated array. 2404 */ 2405 static void clear_status(struct isl_coalesce_info *info) 2406 { 2407 free(info->eq); 2408 free(info->ineq); 2409 } 2410 2411 /* Are all inequality constraints of the basic map represented by "info" 2412 * valid for the other basic map, except for a single constraint 2413 * that is adjacent to an inequality constraint of the other basic map? 2414 */ 2415 static int all_ineq_valid_or_single_adj_ineq(struct isl_coalesce_info *info) 2416 { 2417 int i; 2418 int k = -1; 2419 2420 for (i = 0; i < info->bmap->n_ineq; ++i) { 2421 if (info->ineq[i] == STATUS_REDUNDANT) 2422 continue; 2423 if (info->ineq[i] == STATUS_VALID) 2424 continue; 2425 if (info->ineq[i] != STATUS_ADJ_INEQ) 2426 return 0; 2427 if (k != -1) 2428 return 0; 2429 k = i; 2430 } 2431 2432 return k != -1; 2433 } 2434 2435 /* Basic map "i" has one or more equality constraints that separate it 2436 * from basic map "j". Check if it happens to be an extension 2437 * of basic map "j". 2438 * In particular, check that all constraints of "j" are valid for "i", 2439 * except for one inequality constraint that is adjacent 2440 * to an inequality constraints of "i". 2441 * If so, check for "i" being an extension of "j" by calling 2442 * is_adj_ineq_extension. 2443 * 2444 * Clean up the memory allocated for keeping track of the status 2445 * of the constraints before returning. 2446 */ 2447 static enum isl_change separating_equality(int i, int j, 2448 struct isl_coalesce_info *info) 2449 { 2450 enum isl_change change = isl_change_none; 2451 2452 if (all(info[j].eq, 2 * info[j].bmap->n_eq, STATUS_VALID) && 2453 all_ineq_valid_or_single_adj_ineq(&info[j])) 2454 change = is_adj_ineq_extension(j, i, info); 2455 2456 clear_status(&info[i]); 2457 clear_status(&info[j]); 2458 return change; 2459 } 2460 2461 /* Check if the union of the given pair of basic maps 2462 * can be represented by a single basic map. 2463 * If so, replace the pair by the single basic map and return 2464 * isl_change_drop_first, isl_change_drop_second or isl_change_fuse. 2465 * Otherwise, return isl_change_none. 2466 * The two basic maps are assumed to live in the same local space. 2467 * The "eq" and "ineq" fields of info[i] and info[j] are assumed 2468 * to have been initialized by the caller, either to NULL or 2469 * to valid information. 2470 * 2471 * We first check the effect of each constraint of one basic map 2472 * on the other basic map. 2473 * The constraint may be 2474 * redundant the constraint is redundant in its own 2475 * basic map and should be ignore and removed 2476 * in the end 2477 * valid all (integer) points of the other basic map 2478 * satisfy the constraint 2479 * separate no (integer) point of the other basic map 2480 * satisfies the constraint 2481 * cut some but not all points of the other basic map 2482 * satisfy the constraint 2483 * adj_eq the given constraint is adjacent (on the outside) 2484 * to an equality of the other basic map 2485 * adj_ineq the given constraint is adjacent (on the outside) 2486 * to an inequality of the other basic map 2487 * 2488 * We consider seven cases in which we can replace the pair by a single 2489 * basic map. We ignore all "redundant" constraints. 2490 * 2491 * 1. all constraints of one basic map are valid 2492 * => the other basic map is a subset and can be removed 2493 * 2494 * 2. all constraints of both basic maps are either "valid" or "cut" 2495 * and the facets corresponding to the "cut" constraints 2496 * of one of the basic maps lies entirely inside the other basic map 2497 * => the pair can be replaced by a basic map consisting 2498 * of the valid constraints in both basic maps 2499 * 2500 * 3. there is a single pair of adjacent inequalities 2501 * (all other constraints are "valid") 2502 * => the pair can be replaced by a basic map consisting 2503 * of the valid constraints in both basic maps 2504 * 2505 * 4. one basic map has a single adjacent inequality, while the other 2506 * constraints are "valid". The other basic map has some 2507 * "cut" constraints, but replacing the adjacent inequality by 2508 * its opposite and adding the valid constraints of the other 2509 * basic map results in a subset of the other basic map 2510 * => the pair can be replaced by a basic map consisting 2511 * of the valid constraints in both basic maps 2512 * 2513 * 5. there is a single adjacent pair of an inequality and an equality, 2514 * the other constraints of the basic map containing the inequality are 2515 * "valid". Moreover, if the inequality the basic map is relaxed 2516 * and then turned into an equality, then resulting facet lies 2517 * entirely inside the other basic map 2518 * => the pair can be replaced by the basic map containing 2519 * the inequality, with the inequality relaxed. 2520 * 2521 * 6. there is a single inequality adjacent to an equality, 2522 * the other constraints of the basic map containing the inequality are 2523 * "valid". Moreover, the facets corresponding to both 2524 * the inequality and the equality can be wrapped around their 2525 * ridges to include the other basic map 2526 * => the pair can be replaced by a basic map consisting 2527 * of the valid constraints in both basic maps together 2528 * with all wrapping constraints 2529 * 2530 * 7. one of the basic maps extends beyond the other by at most one. 2531 * Moreover, the facets corresponding to the cut constraints and 2532 * the pieces of the other basic map at offset one from these cut 2533 * constraints can be wrapped around their ridges to include 2534 * the union of the two basic maps 2535 * => the pair can be replaced by a basic map consisting 2536 * of the valid constraints in both basic maps together 2537 * with all wrapping constraints 2538 * 2539 * 8. the two basic maps live in adjacent hyperplanes. In principle 2540 * such sets can always be combined through wrapping, but we impose 2541 * that there is only one such pair, to avoid overeager coalescing. 2542 * 2543 * Throughout the computation, we maintain a collection of tableaus 2544 * corresponding to the basic maps. When the basic maps are dropped 2545 * or combined, the tableaus are modified accordingly. 2546 */ 2547 static enum isl_change coalesce_local_pair_reuse(int i, int j, 2548 struct isl_coalesce_info *info) 2549 { 2550 enum isl_change change = isl_change_none; 2551 2552 set_ineq_status_in(&info[i], info[j].tab); 2553 if (info[i].bmap->n_ineq && !info[i].ineq) 2554 goto error; 2555 if (any_ineq(&info[i], STATUS_ERROR)) 2556 goto error; 2557 if (any_ineq(&info[i], STATUS_SEPARATE)) 2558 goto done; 2559 2560 set_ineq_status_in(&info[j], info[i].tab); 2561 if (info[j].bmap->n_ineq && !info[j].ineq) 2562 goto error; 2563 if (any_ineq(&info[j], STATUS_ERROR)) 2564 goto error; 2565 if (any_ineq(&info[j], STATUS_SEPARATE)) 2566 goto done; 2567 2568 set_eq_status_in(&info[i], info[j].tab); 2569 if (info[i].bmap->n_eq && !info[i].eq) 2570 goto error; 2571 if (any_eq(&info[i], STATUS_ERROR)) 2572 goto error; 2573 2574 set_eq_status_in(&info[j], info[i].tab); 2575 if (info[j].bmap->n_eq && !info[j].eq) 2576 goto error; 2577 if (any_eq(&info[j], STATUS_ERROR)) 2578 goto error; 2579 2580 if (any_eq(&info[i], STATUS_SEPARATE)) 2581 return separating_equality(i, j, info); 2582 if (any_eq(&info[j], STATUS_SEPARATE)) 2583 return separating_equality(j, i, info); 2584 2585 if (all(info[i].eq, 2 * info[i].bmap->n_eq, STATUS_VALID) && 2586 all(info[i].ineq, info[i].bmap->n_ineq, STATUS_VALID)) { 2587 drop(&info[j]); 2588 change = isl_change_drop_second; 2589 } else if (all(info[j].eq, 2 * info[j].bmap->n_eq, STATUS_VALID) && 2590 all(info[j].ineq, info[j].bmap->n_ineq, STATUS_VALID)) { 2591 drop(&info[i]); 2592 change = isl_change_drop_first; 2593 } else if (any_eq(&info[i], STATUS_ADJ_EQ)) { 2594 change = check_eq_adj_eq(i, j, info); 2595 } else if (any_eq(&info[j], STATUS_ADJ_EQ)) { 2596 change = check_eq_adj_eq(j, i, info); 2597 } else if (any_eq(&info[i], STATUS_ADJ_INEQ) || 2598 any_eq(&info[j], STATUS_ADJ_INEQ)) { 2599 change = check_adj_eq(i, j, info); 2600 } else if (any_ineq(&info[i], STATUS_ADJ_EQ)) { 2601 change = check_ineq_adj_eq(i, j, info); 2602 } else if (any_ineq(&info[j], STATUS_ADJ_EQ)) { 2603 change = check_ineq_adj_eq(j, i, info); 2604 } else if (any_ineq(&info[i], STATUS_ADJ_INEQ) || 2605 any_ineq(&info[j], STATUS_ADJ_INEQ)) { 2606 change = check_adj_ineq(i, j, info); 2607 } else { 2608 if (!any_eq(&info[i], STATUS_CUT) && 2609 !any_eq(&info[j], STATUS_CUT)) 2610 change = check_facets(i, j, info); 2611 if (change == isl_change_none) 2612 change = check_wrap(i, j, info); 2613 } 2614 2615 done: 2616 clear_status(&info[i]); 2617 clear_status(&info[j]); 2618 return change; 2619 error: 2620 clear_status(&info[i]); 2621 clear_status(&info[j]); 2622 return isl_change_error; 2623 } 2624 2625 /* Check if the union of the given pair of basic maps 2626 * can be represented by a single basic map. 2627 * If so, replace the pair by the single basic map and return 2628 * isl_change_drop_first, isl_change_drop_second or isl_change_fuse. 2629 * Otherwise, return isl_change_none. 2630 * The two basic maps are assumed to live in the same local space. 2631 */ 2632 static enum isl_change coalesce_local_pair(int i, int j, 2633 struct isl_coalesce_info *info) 2634 { 2635 init_status(&info[i]); 2636 init_status(&info[j]); 2637 return coalesce_local_pair_reuse(i, j, info); 2638 } 2639 2640 /* Shift the integer division at position "div" of the basic map 2641 * represented by "info" by "shift". 2642 * 2643 * That is, if the integer division has the form 2644 * 2645 * floor(f(x)/d) 2646 * 2647 * then replace it by 2648 * 2649 * floor((f(x) + shift * d)/d) - shift 2650 */ 2651 static isl_stat shift_div(struct isl_coalesce_info *info, int div, 2652 isl_int shift) 2653 { 2654 isl_size total, n_div; 2655 2656 info->bmap = isl_basic_map_shift_div(info->bmap, div, 0, shift); 2657 if (!info->bmap) 2658 return isl_stat_error; 2659 2660 total = isl_basic_map_dim(info->bmap, isl_dim_all); 2661 n_div = isl_basic_map_dim(info->bmap, isl_dim_div); 2662 if (total < 0 || n_div < 0) 2663 return isl_stat_error; 2664 total -= n_div; 2665 if (isl_tab_shift_var(info->tab, total + div, shift) < 0) 2666 return isl_stat_error; 2667 2668 return isl_stat_ok; 2669 } 2670 2671 /* If the integer division at position "div" is defined by an equality, 2672 * i.e., a stride constraint, then change the integer division expression 2673 * to have a constant term equal to zero. 2674 * 2675 * Let the equality constraint be 2676 * 2677 * c + f + m a = 0 2678 * 2679 * The integer division expression is then typically of the form 2680 * 2681 * a = floor((-f - c')/m) 2682 * 2683 * The integer division is first shifted by t = floor(c/m), 2684 * turning the equality constraint into 2685 * 2686 * c - m floor(c/m) + f + m a' = 0 2687 * 2688 * i.e., 2689 * 2690 * (c mod m) + f + m a' = 0 2691 * 2692 * That is, 2693 * 2694 * a' = (-f - (c mod m))/m = floor((-f)/m) 2695 * 2696 * because a' is an integer and 0 <= (c mod m) < m. 2697 * The constant term of a' can therefore be zeroed out, 2698 * but only if the integer division expression is of the expected form. 2699 */ 2700 static isl_stat normalize_stride_div(struct isl_coalesce_info *info, int div) 2701 { 2702 isl_bool defined, valid; 2703 isl_stat r; 2704 isl_constraint *c; 2705 isl_int shift, stride; 2706 2707 defined = isl_basic_map_has_defining_equality(info->bmap, isl_dim_div, 2708 div, &c); 2709 if (defined < 0) 2710 return isl_stat_error; 2711 if (!defined) 2712 return isl_stat_ok; 2713 if (!c) 2714 return isl_stat_error; 2715 valid = isl_constraint_is_div_equality(c, div); 2716 isl_int_init(shift); 2717 isl_int_init(stride); 2718 isl_constraint_get_constant(c, &shift); 2719 isl_constraint_get_coefficient(c, isl_dim_div, div, &stride); 2720 isl_int_fdiv_q(shift, shift, stride); 2721 r = shift_div(info, div, shift); 2722 isl_int_clear(stride); 2723 isl_int_clear(shift); 2724 isl_constraint_free(c); 2725 if (r < 0 || valid < 0) 2726 return isl_stat_error; 2727 if (!valid) 2728 return isl_stat_ok; 2729 info->bmap = isl_basic_map_set_div_expr_constant_num_si_inplace( 2730 info->bmap, div, 0); 2731 if (!info->bmap) 2732 return isl_stat_error; 2733 return isl_stat_ok; 2734 } 2735 2736 /* The basic maps represented by "info1" and "info2" are known 2737 * to have the same number of integer divisions. 2738 * Check if pairs of integer divisions are equal to each other 2739 * despite the fact that they differ by a rational constant. 2740 * 2741 * In particular, look for any pair of integer divisions that 2742 * only differ in their constant terms. 2743 * If either of these integer divisions is defined 2744 * by stride constraints, then modify it to have a zero constant term. 2745 * If both are defined by stride constraints then in the end they will have 2746 * the same (zero) constant term. 2747 */ 2748 static isl_stat harmonize_stride_divs(struct isl_coalesce_info *info1, 2749 struct isl_coalesce_info *info2) 2750 { 2751 int i; 2752 isl_size n; 2753 2754 n = isl_basic_map_dim(info1->bmap, isl_dim_div); 2755 if (n < 0) 2756 return isl_stat_error; 2757 for (i = 0; i < n; ++i) { 2758 isl_bool known, harmonize; 2759 2760 known = isl_basic_map_div_is_known(info1->bmap, i); 2761 if (known >= 0 && known) 2762 known = isl_basic_map_div_is_known(info2->bmap, i); 2763 if (known < 0) 2764 return isl_stat_error; 2765 if (!known) 2766 continue; 2767 harmonize = isl_basic_map_equal_div_expr_except_constant( 2768 info1->bmap, i, info2->bmap, i); 2769 if (harmonize < 0) 2770 return isl_stat_error; 2771 if (!harmonize) 2772 continue; 2773 if (normalize_stride_div(info1, i) < 0) 2774 return isl_stat_error; 2775 if (normalize_stride_div(info2, i) < 0) 2776 return isl_stat_error; 2777 } 2778 2779 return isl_stat_ok; 2780 } 2781 2782 /* If "shift" is an integer constant, then shift the integer division 2783 * at position "div" of the basic map represented by "info" by "shift". 2784 * If "shift" is not an integer constant, then do nothing. 2785 * If "shift" is equal to zero, then no shift needs to be performed either. 2786 * 2787 * That is, if the integer division has the form 2788 * 2789 * floor(f(x)/d) 2790 * 2791 * then replace it by 2792 * 2793 * floor((f(x) + shift * d)/d) - shift 2794 */ 2795 static isl_stat shift_if_cst_int(struct isl_coalesce_info *info, int div, 2796 __isl_keep isl_aff *shift) 2797 { 2798 isl_bool cst; 2799 isl_stat r; 2800 isl_int d; 2801 isl_val *c; 2802 2803 cst = isl_aff_is_cst(shift); 2804 if (cst < 0 || !cst) 2805 return cst < 0 ? isl_stat_error : isl_stat_ok; 2806 2807 c = isl_aff_get_constant_val(shift); 2808 cst = isl_val_is_int(c); 2809 if (cst >= 0 && cst) 2810 cst = isl_bool_not(isl_val_is_zero(c)); 2811 if (cst < 0 || !cst) { 2812 isl_val_free(c); 2813 return cst < 0 ? isl_stat_error : isl_stat_ok; 2814 } 2815 2816 isl_int_init(d); 2817 r = isl_val_get_num_isl_int(c, &d); 2818 if (r >= 0) 2819 r = shift_div(info, div, d); 2820 isl_int_clear(d); 2821 2822 isl_val_free(c); 2823 2824 return r; 2825 } 2826 2827 /* Check if some of the divs in the basic map represented by "info1" 2828 * are shifts of the corresponding divs in the basic map represented 2829 * by "info2", taking into account the equality constraints "eq1" of "info1" 2830 * and "eq2" of "info2". If so, align them with those of "info2". 2831 * "info1" and "info2" are assumed to have the same number 2832 * of integer divisions. 2833 * 2834 * An integer division is considered to be a shift of another integer 2835 * division if, after simplification with respect to the equality 2836 * constraints of the other basic map, one is equal to the other 2837 * plus a constant. 2838 * 2839 * In particular, for each pair of integer divisions, if both are known, 2840 * have the same denominator and are not already equal to each other, 2841 * simplify each with respect to the equality constraints 2842 * of the other basic map. If the difference is an integer constant, 2843 * then move this difference outside. 2844 * That is, if, after simplification, one integer division is of the form 2845 * 2846 * floor((f(x) + c_1)/d) 2847 * 2848 * while the other is of the form 2849 * 2850 * floor((f(x) + c_2)/d) 2851 * 2852 * and n = (c_2 - c_1)/d is an integer, then replace the first 2853 * integer division by 2854 * 2855 * floor((f_1(x) + c_1 + n * d)/d) - n, 2856 * 2857 * where floor((f_1(x) + c_1 + n * d)/d) = floor((f2(x) + c_2)/d) 2858 * after simplification with respect to the equality constraints. 2859 */ 2860 static isl_stat harmonize_divs_with_hulls(struct isl_coalesce_info *info1, 2861 struct isl_coalesce_info *info2, __isl_keep isl_basic_set *eq1, 2862 __isl_keep isl_basic_set *eq2) 2863 { 2864 int i; 2865 isl_size total; 2866 isl_local_space *ls1, *ls2; 2867 2868 total = isl_basic_map_dim(info1->bmap, isl_dim_all); 2869 if (total < 0) 2870 return isl_stat_error; 2871 ls1 = isl_local_space_wrap(isl_basic_map_get_local_space(info1->bmap)); 2872 ls2 = isl_local_space_wrap(isl_basic_map_get_local_space(info2->bmap)); 2873 for (i = 0; i < info1->bmap->n_div; ++i) { 2874 isl_stat r; 2875 isl_aff *div1, *div2; 2876 2877 if (!isl_local_space_div_is_known(ls1, i) || 2878 !isl_local_space_div_is_known(ls2, i)) 2879 continue; 2880 if (isl_int_ne(info1->bmap->div[i][0], info2->bmap->div[i][0])) 2881 continue; 2882 if (isl_seq_eq(info1->bmap->div[i] + 1, 2883 info2->bmap->div[i] + 1, 1 + total)) 2884 continue; 2885 div1 = isl_local_space_get_div(ls1, i); 2886 div2 = isl_local_space_get_div(ls2, i); 2887 div1 = isl_aff_substitute_equalities(div1, 2888 isl_basic_set_copy(eq2)); 2889 div2 = isl_aff_substitute_equalities(div2, 2890 isl_basic_set_copy(eq1)); 2891 div2 = isl_aff_sub(div2, div1); 2892 r = shift_if_cst_int(info1, i, div2); 2893 isl_aff_free(div2); 2894 if (r < 0) 2895 break; 2896 } 2897 isl_local_space_free(ls1); 2898 isl_local_space_free(ls2); 2899 2900 if (i < info1->bmap->n_div) 2901 return isl_stat_error; 2902 return isl_stat_ok; 2903 } 2904 2905 /* Check if some of the divs in the basic map represented by "info1" 2906 * are shifts of the corresponding divs in the basic map represented 2907 * by "info2". If so, align them with those of "info2". 2908 * Only do this if "info1" and "info2" have the same number 2909 * of integer divisions. 2910 * 2911 * An integer division is considered to be a shift of another integer 2912 * division if, after simplification with respect to the equality 2913 * constraints of the other basic map, one is equal to the other 2914 * plus a constant. 2915 * 2916 * First check if pairs of integer divisions are equal to each other 2917 * despite the fact that they differ by a rational constant. 2918 * If so, try and arrange for them to have the same constant term. 2919 * 2920 * Then, extract the equality constraints and continue with 2921 * harmonize_divs_with_hulls. 2922 * 2923 * If the equality constraints of both basic maps are the same, 2924 * then there is no need to perform any shifting since 2925 * the coefficients of the integer divisions should have been 2926 * reduced in the same way. 2927 */ 2928 static isl_stat harmonize_divs(struct isl_coalesce_info *info1, 2929 struct isl_coalesce_info *info2) 2930 { 2931 isl_bool equal; 2932 isl_basic_map *bmap1, *bmap2; 2933 isl_basic_set *eq1, *eq2; 2934 isl_stat r; 2935 2936 if (!info1->bmap || !info2->bmap) 2937 return isl_stat_error; 2938 2939 if (info1->bmap->n_div != info2->bmap->n_div) 2940 return isl_stat_ok; 2941 if (info1->bmap->n_div == 0) 2942 return isl_stat_ok; 2943 2944 if (harmonize_stride_divs(info1, info2) < 0) 2945 return isl_stat_error; 2946 2947 bmap1 = isl_basic_map_copy(info1->bmap); 2948 bmap2 = isl_basic_map_copy(info2->bmap); 2949 eq1 = isl_basic_map_wrap(isl_basic_map_plain_affine_hull(bmap1)); 2950 eq2 = isl_basic_map_wrap(isl_basic_map_plain_affine_hull(bmap2)); 2951 equal = isl_basic_set_plain_is_equal(eq1, eq2); 2952 if (equal < 0) 2953 r = isl_stat_error; 2954 else if (equal) 2955 r = isl_stat_ok; 2956 else 2957 r = harmonize_divs_with_hulls(info1, info2, eq1, eq2); 2958 isl_basic_set_free(eq1); 2959 isl_basic_set_free(eq2); 2960 2961 return r; 2962 } 2963 2964 /* Do the two basic maps live in the same local space, i.e., 2965 * do they have the same (known) divs? 2966 * If either basic map has any unknown divs, then we can only assume 2967 * that they do not live in the same local space. 2968 */ 2969 static isl_bool same_divs(__isl_keep isl_basic_map *bmap1, 2970 __isl_keep isl_basic_map *bmap2) 2971 { 2972 int i; 2973 isl_bool known; 2974 isl_size total; 2975 2976 if (!bmap1 || !bmap2) 2977 return isl_bool_error; 2978 if (bmap1->n_div != bmap2->n_div) 2979 return isl_bool_false; 2980 2981 if (bmap1->n_div == 0) 2982 return isl_bool_true; 2983 2984 known = isl_basic_map_divs_known(bmap1); 2985 if (known < 0 || !known) 2986 return known; 2987 known = isl_basic_map_divs_known(bmap2); 2988 if (known < 0 || !known) 2989 return known; 2990 2991 total = isl_basic_map_dim(bmap1, isl_dim_all); 2992 if (total < 0) 2993 return isl_bool_error; 2994 for (i = 0; i < bmap1->n_div; ++i) 2995 if (!isl_seq_eq(bmap1->div[i], bmap2->div[i], 2 + total)) 2996 return isl_bool_false; 2997 2998 return isl_bool_true; 2999 } 3000 3001 /* Assuming that "tab" contains the equality constraints and 3002 * the initial inequality constraints of "bmap", copy the remaining 3003 * inequality constraints of "bmap" to "Tab". 3004 */ 3005 static isl_stat copy_ineq(struct isl_tab *tab, __isl_keep isl_basic_map *bmap) 3006 { 3007 int i, n_ineq; 3008 3009 if (!bmap) 3010 return isl_stat_error; 3011 3012 n_ineq = tab->n_con - tab->n_eq; 3013 for (i = n_ineq; i < bmap->n_ineq; ++i) 3014 if (isl_tab_add_ineq(tab, bmap->ineq[i]) < 0) 3015 return isl_stat_error; 3016 3017 return isl_stat_ok; 3018 } 3019 3020 /* Description of an integer division that is added 3021 * during an expansion. 3022 * "pos" is the position of the corresponding variable. 3023 * "cst" indicates whether this integer division has a fixed value. 3024 * "val" contains the fixed value, if the value is fixed. 3025 */ 3026 struct isl_expanded { 3027 int pos; 3028 isl_bool cst; 3029 isl_int val; 3030 }; 3031 3032 /* For each of the "n" integer division variables "expanded", 3033 * if the variable has a fixed value, then add two inequality 3034 * constraints expressing the fixed value. 3035 * Otherwise, add the corresponding div constraints. 3036 * The caller is responsible for removing the div constraints 3037 * that it added for all these "n" integer divisions. 3038 * 3039 * The div constraints and the pair of inequality constraints 3040 * forcing the fixed value cannot both be added for a given variable 3041 * as the combination may render some of the original constraints redundant. 3042 * These would then be ignored during the coalescing detection, 3043 * while they could remain in the fused result. 3044 * 3045 * The two added inequality constraints are 3046 * 3047 * -a + v >= 0 3048 * a - v >= 0 3049 * 3050 * with "a" the variable and "v" its fixed value. 3051 * The facet corresponding to one of these two constraints is selected 3052 * in the tableau to ensure that the pair of inequality constraints 3053 * is treated as an equality constraint. 3054 * 3055 * The information in info->ineq is thrown away because it was 3056 * computed in terms of div constraints, while some of those 3057 * have now been replaced by these pairs of inequality constraints. 3058 */ 3059 static isl_stat fix_constant_divs(struct isl_coalesce_info *info, 3060 int n, struct isl_expanded *expanded) 3061 { 3062 unsigned o_div; 3063 int i; 3064 isl_vec *ineq; 3065 3066 o_div = isl_basic_map_offset(info->bmap, isl_dim_div) - 1; 3067 ineq = isl_vec_alloc(isl_tab_get_ctx(info->tab), 1 + info->tab->n_var); 3068 if (!ineq) 3069 return isl_stat_error; 3070 isl_seq_clr(ineq->el + 1, info->tab->n_var); 3071 3072 for (i = 0; i < n; ++i) { 3073 if (!expanded[i].cst) { 3074 info->bmap = isl_basic_map_extend_constraints( 3075 info->bmap, 0, 2); 3076 info->bmap = isl_basic_map_add_div_constraints( 3077 info->bmap, expanded[i].pos - o_div); 3078 } else { 3079 isl_int_set_si(ineq->el[1 + expanded[i].pos], -1); 3080 isl_int_set(ineq->el[0], expanded[i].val); 3081 info->bmap = isl_basic_map_add_ineq(info->bmap, 3082 ineq->el); 3083 isl_int_set_si(ineq->el[1 + expanded[i].pos], 1); 3084 isl_int_neg(ineq->el[0], expanded[i].val); 3085 info->bmap = isl_basic_map_add_ineq(info->bmap, 3086 ineq->el); 3087 isl_int_set_si(ineq->el[1 + expanded[i].pos], 0); 3088 } 3089 if (copy_ineq(info->tab, info->bmap) < 0) 3090 break; 3091 if (expanded[i].cst && 3092 isl_tab_select_facet(info->tab, info->tab->n_con - 1) < 0) 3093 break; 3094 } 3095 3096 isl_vec_free(ineq); 3097 3098 clear_status(info); 3099 init_status(info); 3100 3101 return i < n ? isl_stat_error : isl_stat_ok; 3102 } 3103 3104 /* Insert the "n" integer division variables "expanded" 3105 * into info->tab and info->bmap and 3106 * update info->ineq with respect to the redundant constraints 3107 * in the resulting tableau. 3108 * "bmap" contains the result of this insertion in info->bmap, 3109 * while info->bmap is the original version 3110 * of "bmap", i.e., the one that corresponds to the current 3111 * state of info->tab. The number of constraints in info->bmap 3112 * is assumed to be the same as the number of constraints 3113 * in info->tab. This is required to be able to detect 3114 * the extra constraints in "bmap". 3115 * 3116 * In particular, introduce extra variables corresponding 3117 * to the extra integer divisions and add the div constraints 3118 * that were added to "bmap" after info->tab was created 3119 * from info->bmap. 3120 * Furthermore, check if these extra integer divisions happen 3121 * to attain a fixed integer value in info->tab. 3122 * If so, replace the corresponding div constraints by pairs 3123 * of inequality constraints that fix these 3124 * integer divisions to their single integer values. 3125 * Replace info->bmap by "bmap" to match the changes to info->tab. 3126 * info->ineq was computed without a tableau and therefore 3127 * does not take into account the redundant constraints 3128 * in the tableau. Mark them here. 3129 * There is no need to check the newly added div constraints 3130 * since they cannot be redundant. 3131 * The redundancy check is not performed when constants have been discovered 3132 * since info->ineq is completely thrown away in this case. 3133 */ 3134 static isl_stat tab_insert_divs(struct isl_coalesce_info *info, 3135 int n, struct isl_expanded *expanded, __isl_take isl_basic_map *bmap) 3136 { 3137 int i, n_ineq; 3138 unsigned n_eq; 3139 struct isl_tab_undo *snap; 3140 int any; 3141 3142 if (!bmap) 3143 return isl_stat_error; 3144 if (info->bmap->n_eq + info->bmap->n_ineq != info->tab->n_con) 3145 isl_die(isl_basic_map_get_ctx(bmap), isl_error_internal, 3146 "original tableau does not correspond " 3147 "to original basic map", goto error); 3148 3149 if (isl_tab_extend_vars(info->tab, n) < 0) 3150 goto error; 3151 if (isl_tab_extend_cons(info->tab, 2 * n) < 0) 3152 goto error; 3153 3154 for (i = 0; i < n; ++i) { 3155 if (isl_tab_insert_var(info->tab, expanded[i].pos) < 0) 3156 goto error; 3157 } 3158 3159 snap = isl_tab_snap(info->tab); 3160 3161 n_ineq = info->tab->n_con - info->tab->n_eq; 3162 if (copy_ineq(info->tab, bmap) < 0) 3163 goto error; 3164 3165 isl_basic_map_free(info->bmap); 3166 info->bmap = bmap; 3167 3168 any = 0; 3169 for (i = 0; i < n; ++i) { 3170 expanded[i].cst = isl_tab_is_constant(info->tab, 3171 expanded[i].pos, &expanded[i].val); 3172 if (expanded[i].cst < 0) 3173 return isl_stat_error; 3174 if (expanded[i].cst) 3175 any = 1; 3176 } 3177 3178 if (any) { 3179 if (isl_tab_rollback(info->tab, snap) < 0) 3180 return isl_stat_error; 3181 info->bmap = isl_basic_map_cow(info->bmap); 3182 info->bmap = isl_basic_map_free_inequality(info->bmap, 2 * n); 3183 if (!info->bmap) 3184 return isl_stat_error; 3185 3186 return fix_constant_divs(info, n, expanded); 3187 } 3188 3189 n_eq = info->bmap->n_eq; 3190 for (i = 0; i < n_ineq; ++i) { 3191 if (isl_tab_is_redundant(info->tab, n_eq + i)) 3192 info->ineq[i] = STATUS_REDUNDANT; 3193 } 3194 3195 return isl_stat_ok; 3196 error: 3197 isl_basic_map_free(bmap); 3198 return isl_stat_error; 3199 } 3200 3201 /* Expand info->tab and info->bmap in the same way "bmap" was expanded 3202 * in isl_basic_map_expand_divs using the expansion "exp" and 3203 * update info->ineq with respect to the redundant constraints 3204 * in the resulting tableau. info->bmap is the original version 3205 * of "bmap", i.e., the one that corresponds to the current 3206 * state of info->tab. The number of constraints in info->bmap 3207 * is assumed to be the same as the number of constraints 3208 * in info->tab. This is required to be able to detect 3209 * the extra constraints in "bmap". 3210 * 3211 * Extract the positions where extra local variables are introduced 3212 * from "exp" and call tab_insert_divs. 3213 */ 3214 static isl_stat expand_tab(struct isl_coalesce_info *info, int *exp, 3215 __isl_take isl_basic_map *bmap) 3216 { 3217 isl_ctx *ctx; 3218 struct isl_expanded *expanded; 3219 int i, j, k, n; 3220 int extra_var; 3221 isl_size total, n_div; 3222 unsigned pos; 3223 isl_stat r; 3224 3225 total = isl_basic_map_dim(bmap, isl_dim_all); 3226 n_div = isl_basic_map_dim(bmap, isl_dim_div); 3227 if (total < 0 || n_div < 0) 3228 return isl_stat_error; 3229 pos = total - n_div; 3230 extra_var = total - info->tab->n_var; 3231 n = n_div - extra_var; 3232 3233 ctx = isl_basic_map_get_ctx(bmap); 3234 expanded = isl_calloc_array(ctx, struct isl_expanded, extra_var); 3235 if (extra_var && !expanded) 3236 goto error; 3237 3238 i = 0; 3239 k = 0; 3240 for (j = 0; j < n_div; ++j) { 3241 if (i < n && exp[i] == j) { 3242 ++i; 3243 continue; 3244 } 3245 expanded[k++].pos = pos + j; 3246 } 3247 3248 for (k = 0; k < extra_var; ++k) 3249 isl_int_init(expanded[k].val); 3250 3251 r = tab_insert_divs(info, extra_var, expanded, bmap); 3252 3253 for (k = 0; k < extra_var; ++k) 3254 isl_int_clear(expanded[k].val); 3255 free(expanded); 3256 3257 return r; 3258 error: 3259 isl_basic_map_free(bmap); 3260 return isl_stat_error; 3261 } 3262 3263 /* Check if the union of the basic maps represented by info[i] and info[j] 3264 * can be represented by a single basic map, 3265 * after expanding the divs of info[i] to match those of info[j]. 3266 * If so, replace the pair by the single basic map and return 3267 * isl_change_drop_first, isl_change_drop_second or isl_change_fuse. 3268 * Otherwise, return isl_change_none. 3269 * 3270 * The caller has already checked for info[j] being a subset of info[i]. 3271 * If some of the divs of info[j] are unknown, then the expanded info[i] 3272 * will not have the corresponding div constraints. The other patterns 3273 * therefore cannot apply. Skip the computation in this case. 3274 * 3275 * The expansion is performed using the divs "div" and expansion "exp" 3276 * computed by the caller. 3277 * info[i].bmap has already been expanded and the result is passed in 3278 * as "bmap". 3279 * The "eq" and "ineq" fields of info[i] reflect the status of 3280 * the constraints of the expanded "bmap" with respect to info[j].tab. 3281 * However, inequality constraints that are redundant in info[i].tab 3282 * have not yet been marked as such because no tableau was available. 3283 * 3284 * Replace info[i].bmap by "bmap" and expand info[i].tab as well, 3285 * updating info[i].ineq with respect to the redundant constraints. 3286 * Then try and coalesce the expanded info[i] with info[j], 3287 * reusing the information in info[i].eq and info[i].ineq. 3288 * If this does not result in any coalescing or if it results in info[j] 3289 * getting dropped (which should not happen in practice, since the case 3290 * of info[j] being a subset of info[i] has already been checked by 3291 * the caller), then revert info[i] to its original state. 3292 */ 3293 static enum isl_change coalesce_expand_tab_divs(__isl_take isl_basic_map *bmap, 3294 int i, int j, struct isl_coalesce_info *info, __isl_keep isl_mat *div, 3295 int *exp) 3296 { 3297 isl_bool known; 3298 isl_basic_map *bmap_i; 3299 struct isl_tab_undo *snap; 3300 enum isl_change change = isl_change_none; 3301 3302 known = isl_basic_map_divs_known(info[j].bmap); 3303 if (known < 0 || !known) { 3304 clear_status(&info[i]); 3305 isl_basic_map_free(bmap); 3306 return known < 0 ? isl_change_error : isl_change_none; 3307 } 3308 3309 bmap_i = isl_basic_map_copy(info[i].bmap); 3310 snap = isl_tab_snap(info[i].tab); 3311 if (expand_tab(&info[i], exp, bmap) < 0) 3312 change = isl_change_error; 3313 3314 init_status(&info[j]); 3315 if (change == isl_change_none) 3316 change = coalesce_local_pair_reuse(i, j, info); 3317 else 3318 clear_status(&info[i]); 3319 if (change != isl_change_none && change != isl_change_drop_second) { 3320 isl_basic_map_free(bmap_i); 3321 } else { 3322 isl_basic_map_free(info[i].bmap); 3323 info[i].bmap = bmap_i; 3324 3325 if (isl_tab_rollback(info[i].tab, snap) < 0) 3326 change = isl_change_error; 3327 } 3328 3329 return change; 3330 } 3331 3332 /* Check if the union of "bmap" and the basic map represented by info[j] 3333 * can be represented by a single basic map, 3334 * after expanding the divs of "bmap" to match those of info[j]. 3335 * If so, replace the pair by the single basic map and return 3336 * isl_change_drop_first, isl_change_drop_second or isl_change_fuse. 3337 * Otherwise, return isl_change_none. 3338 * 3339 * In particular, check if the expanded "bmap" contains the basic map 3340 * represented by the tableau info[j].tab. 3341 * The expansion is performed using the divs "div" and expansion "exp" 3342 * computed by the caller. 3343 * Then we check if all constraints of the expanded "bmap" are valid for 3344 * info[j].tab. 3345 * 3346 * If "i" is not equal to -1, then "bmap" is equal to info[i].bmap. 3347 * In this case, the positions of the constraints of info[i].bmap 3348 * with respect to the basic map represented by info[j] are stored 3349 * in info[i]. 3350 * 3351 * If the expanded "bmap" does not contain the basic map 3352 * represented by the tableau info[j].tab and if "i" is not -1, 3353 * i.e., if the original "bmap" is info[i].bmap, then expand info[i].tab 3354 * as well and check if that results in coalescing. 3355 */ 3356 static enum isl_change coalesce_with_expanded_divs( 3357 __isl_keep isl_basic_map *bmap, int i, int j, 3358 struct isl_coalesce_info *info, __isl_keep isl_mat *div, int *exp) 3359 { 3360 enum isl_change change = isl_change_none; 3361 struct isl_coalesce_info info_local, *info_i; 3362 3363 info_i = i >= 0 ? &info[i] : &info_local; 3364 init_status(info_i); 3365 bmap = isl_basic_map_copy(bmap); 3366 bmap = isl_basic_map_expand_divs(bmap, isl_mat_copy(div), exp); 3367 bmap = isl_basic_map_mark_final(bmap); 3368 3369 if (!bmap) 3370 goto error; 3371 3372 info_local.bmap = bmap; 3373 info_i->eq = eq_status_in(bmap, info[j].tab); 3374 if (bmap->n_eq && !info_i->eq) 3375 goto error; 3376 if (any_eq(info_i, STATUS_ERROR)) 3377 goto error; 3378 if (any_eq(info_i, STATUS_SEPARATE)) 3379 goto done; 3380 3381 info_i->ineq = ineq_status_in(bmap, NULL, info[j].tab); 3382 if (bmap->n_ineq && !info_i->ineq) 3383 goto error; 3384 if (any_ineq(info_i, STATUS_ERROR)) 3385 goto error; 3386 if (any_ineq(info_i, STATUS_SEPARATE)) 3387 goto done; 3388 3389 if (all(info_i->eq, 2 * bmap->n_eq, STATUS_VALID) && 3390 all(info_i->ineq, bmap->n_ineq, STATUS_VALID)) { 3391 drop(&info[j]); 3392 change = isl_change_drop_second; 3393 } 3394 3395 if (change == isl_change_none && i != -1) 3396 return coalesce_expand_tab_divs(bmap, i, j, info, div, exp); 3397 3398 done: 3399 isl_basic_map_free(bmap); 3400 clear_status(info_i); 3401 return change; 3402 error: 3403 isl_basic_map_free(bmap); 3404 clear_status(info_i); 3405 return isl_change_error; 3406 } 3407 3408 /* Check if the union of "bmap_i" and the basic map represented by info[j] 3409 * can be represented by a single basic map, 3410 * after aligning the divs of "bmap_i" to match those of info[j]. 3411 * If so, replace the pair by the single basic map and return 3412 * isl_change_drop_first, isl_change_drop_second or isl_change_fuse. 3413 * Otherwise, return isl_change_none. 3414 * 3415 * In particular, check if "bmap_i" contains the basic map represented by 3416 * info[j] after aligning the divs of "bmap_i" to those of info[j]. 3417 * Note that this can only succeed if the number of divs of "bmap_i" 3418 * is smaller than (or equal to) the number of divs of info[j]. 3419 * 3420 * We first check if the divs of "bmap_i" are all known and form a subset 3421 * of those of info[j].bmap. If so, we pass control over to 3422 * coalesce_with_expanded_divs. 3423 * 3424 * If "i" is not equal to -1, then "bmap" is equal to info[i].bmap. 3425 */ 3426 static enum isl_change coalesce_after_aligning_divs( 3427 __isl_keep isl_basic_map *bmap_i, int i, int j, 3428 struct isl_coalesce_info *info) 3429 { 3430 isl_bool known; 3431 isl_mat *div_i, *div_j, *div; 3432 int *exp1 = NULL; 3433 int *exp2 = NULL; 3434 isl_ctx *ctx; 3435 enum isl_change change; 3436 3437 known = isl_basic_map_divs_known(bmap_i); 3438 if (known < 0) 3439 return isl_change_error; 3440 if (!known) 3441 return isl_change_none; 3442 3443 ctx = isl_basic_map_get_ctx(bmap_i); 3444 3445 div_i = isl_basic_map_get_divs(bmap_i); 3446 div_j = isl_basic_map_get_divs(info[j].bmap); 3447 3448 if (!div_i || !div_j) 3449 goto error; 3450 3451 exp1 = isl_alloc_array(ctx, int, div_i->n_row); 3452 exp2 = isl_alloc_array(ctx, int, div_j->n_row); 3453 if ((div_i->n_row && !exp1) || (div_j->n_row && !exp2)) 3454 goto error; 3455 3456 div = isl_merge_divs(div_i, div_j, exp1, exp2); 3457 if (!div) 3458 goto error; 3459 3460 if (div->n_row == div_j->n_row) 3461 change = coalesce_with_expanded_divs(bmap_i, 3462 i, j, info, div, exp1); 3463 else 3464 change = isl_change_none; 3465 3466 isl_mat_free(div); 3467 3468 isl_mat_free(div_i); 3469 isl_mat_free(div_j); 3470 3471 free(exp2); 3472 free(exp1); 3473 3474 return change; 3475 error: 3476 isl_mat_free(div_i); 3477 isl_mat_free(div_j); 3478 free(exp1); 3479 free(exp2); 3480 return isl_change_error; 3481 } 3482 3483 /* Check if basic map "j" is a subset of basic map "i" after 3484 * exploiting the extra equalities of "j" to simplify the divs of "i". 3485 * If so, remove basic map "j" and return isl_change_drop_second. 3486 * 3487 * If "j" does not have any equalities or if they are the same 3488 * as those of "i", then we cannot exploit them to simplify the divs. 3489 * Similarly, if there are no divs in "i", then they cannot be simplified. 3490 * If, on the other hand, the affine hulls of "i" and "j" do not intersect, 3491 * then "j" cannot be a subset of "i". 3492 * 3493 * Otherwise, we intersect "i" with the affine hull of "j" and then 3494 * check if "j" is a subset of the result after aligning the divs. 3495 * If so, then "j" is definitely a subset of "i" and can be removed. 3496 * Note that if after intersection with the affine hull of "j". 3497 * "i" still has more divs than "j", then there is no way we can 3498 * align the divs of "i" to those of "j". 3499 */ 3500 static enum isl_change coalesce_subset_with_equalities(int i, int j, 3501 struct isl_coalesce_info *info) 3502 { 3503 isl_basic_map *hull_i, *hull_j, *bmap_i; 3504 int equal, empty; 3505 enum isl_change change; 3506 3507 if (info[j].bmap->n_eq == 0) 3508 return isl_change_none; 3509 if (info[i].bmap->n_div == 0) 3510 return isl_change_none; 3511 3512 hull_i = isl_basic_map_copy(info[i].bmap); 3513 hull_i = isl_basic_map_plain_affine_hull(hull_i); 3514 hull_j = isl_basic_map_copy(info[j].bmap); 3515 hull_j = isl_basic_map_plain_affine_hull(hull_j); 3516 3517 hull_j = isl_basic_map_intersect(hull_j, isl_basic_map_copy(hull_i)); 3518 equal = isl_basic_map_plain_is_equal(hull_i, hull_j); 3519 empty = isl_basic_map_plain_is_empty(hull_j); 3520 isl_basic_map_free(hull_i); 3521 3522 if (equal < 0 || equal || empty < 0 || empty) { 3523 isl_basic_map_free(hull_j); 3524 if (equal < 0 || empty < 0) 3525 return isl_change_error; 3526 return isl_change_none; 3527 } 3528 3529 bmap_i = isl_basic_map_copy(info[i].bmap); 3530 bmap_i = isl_basic_map_intersect(bmap_i, hull_j); 3531 if (!bmap_i) 3532 return isl_change_error; 3533 3534 if (bmap_i->n_div > info[j].bmap->n_div) { 3535 isl_basic_map_free(bmap_i); 3536 return isl_change_none; 3537 } 3538 3539 change = coalesce_after_aligning_divs(bmap_i, -1, j, info); 3540 3541 isl_basic_map_free(bmap_i); 3542 3543 return change; 3544 } 3545 3546 /* Check if the union of the basic maps represented by info[i] and info[j] 3547 * can be represented by a single basic map, by aligning or equating 3548 * their integer divisions. 3549 * If so, replace the pair by the single basic map and return 3550 * isl_change_drop_first, isl_change_drop_second or isl_change_fuse. 3551 * Otherwise, return isl_change_none. 3552 * 3553 * Note that we only perform any test if the number of divs is different 3554 * in the two basic maps. In case the number of divs is the same, 3555 * we have already established that the divs are different 3556 * in the two basic maps. 3557 * In particular, if the number of divs of basic map i is smaller than 3558 * the number of divs of basic map j, then we check if j is a subset of i 3559 * and vice versa. 3560 */ 3561 static enum isl_change coalesce_divs(int i, int j, 3562 struct isl_coalesce_info *info) 3563 { 3564 enum isl_change change = isl_change_none; 3565 3566 if (info[i].bmap->n_div < info[j].bmap->n_div) 3567 change = coalesce_after_aligning_divs(info[i].bmap, i, j, info); 3568 if (change != isl_change_none) 3569 return change; 3570 3571 if (info[j].bmap->n_div < info[i].bmap->n_div) 3572 change = coalesce_after_aligning_divs(info[j].bmap, j, i, info); 3573 if (change != isl_change_none) 3574 return invert_change(change); 3575 3576 change = coalesce_subset_with_equalities(i, j, info); 3577 if (change != isl_change_none) 3578 return change; 3579 3580 change = coalesce_subset_with_equalities(j, i, info); 3581 if (change != isl_change_none) 3582 return invert_change(change); 3583 3584 return isl_change_none; 3585 } 3586 3587 /* Does "bmap" involve any divs that themselves refer to divs? 3588 */ 3589 static isl_bool has_nested_div(__isl_keep isl_basic_map *bmap) 3590 { 3591 int i; 3592 isl_size total; 3593 isl_size n_div; 3594 3595 total = isl_basic_map_dim(bmap, isl_dim_all); 3596 n_div = isl_basic_map_dim(bmap, isl_dim_div); 3597 if (total < 0 || n_div < 0) 3598 return isl_bool_error; 3599 total -= n_div; 3600 3601 for (i = 0; i < n_div; ++i) 3602 if (isl_seq_first_non_zero(bmap->div[i] + 2 + total, 3603 n_div) != -1) 3604 return isl_bool_true; 3605 3606 return isl_bool_false; 3607 } 3608 3609 /* Return a list of affine expressions, one for each integer division 3610 * in "bmap_i". For each integer division that also appears in "bmap_j", 3611 * the affine expression is set to NaN. The number of NaNs in the list 3612 * is equal to the number of integer divisions in "bmap_j". 3613 * For the other integer divisions of "bmap_i", the corresponding 3614 * element in the list is a purely affine expression equal to the integer 3615 * division in "hull". 3616 * If no such list can be constructed, then the number of elements 3617 * in the returned list is smaller than the number of integer divisions 3618 * in "bmap_i". 3619 * The integer division of "bmap_i" and "bmap_j" are assumed to be known and 3620 * not contain any nested divs. 3621 */ 3622 static __isl_give isl_aff_list *set_up_substitutions( 3623 __isl_keep isl_basic_map *bmap_i, __isl_keep isl_basic_map *bmap_j, 3624 __isl_take isl_basic_map *hull) 3625 { 3626 isl_size n_div_i, n_div_j, total; 3627 isl_ctx *ctx; 3628 isl_local_space *ls; 3629 isl_basic_set *wrap_hull; 3630 isl_aff *aff_nan; 3631 isl_aff_list *list; 3632 int i, j; 3633 3634 n_div_i = isl_basic_map_dim(bmap_i, isl_dim_div); 3635 n_div_j = isl_basic_map_dim(bmap_j, isl_dim_div); 3636 total = isl_basic_map_dim(bmap_i, isl_dim_all); 3637 if (!hull || n_div_i < 0 || n_div_j < 0 || total < 0) 3638 return NULL; 3639 3640 ctx = isl_basic_map_get_ctx(hull); 3641 total -= n_div_i; 3642 3643 ls = isl_basic_map_get_local_space(bmap_i); 3644 ls = isl_local_space_wrap(ls); 3645 wrap_hull = isl_basic_map_wrap(hull); 3646 3647 aff_nan = isl_aff_nan_on_domain(isl_local_space_copy(ls)); 3648 list = isl_aff_list_alloc(ctx, n_div_i); 3649 3650 j = 0; 3651 for (i = 0; i < n_div_i; ++i) { 3652 isl_aff *aff; 3653 isl_size n_div; 3654 3655 if (j < n_div_j && 3656 isl_basic_map_equal_div_expr_part(bmap_i, i, bmap_j, j, 3657 0, 2 + total)) { 3658 ++j; 3659 list = isl_aff_list_add(list, isl_aff_copy(aff_nan)); 3660 continue; 3661 } 3662 if (n_div_i - i <= n_div_j - j) 3663 break; 3664 3665 aff = isl_local_space_get_div(ls, i); 3666 aff = isl_aff_substitute_equalities(aff, 3667 isl_basic_set_copy(wrap_hull)); 3668 aff = isl_aff_floor(aff); 3669 n_div = isl_aff_dim(aff, isl_dim_div); 3670 if (n_div < 0) 3671 goto error; 3672 if (n_div != 0) { 3673 isl_aff_free(aff); 3674 break; 3675 } 3676 3677 list = isl_aff_list_add(list, aff); 3678 } 3679 3680 isl_aff_free(aff_nan); 3681 isl_local_space_free(ls); 3682 isl_basic_set_free(wrap_hull); 3683 3684 return list; 3685 error: 3686 isl_aff_free(aff_nan); 3687 isl_local_space_free(ls); 3688 isl_basic_set_free(wrap_hull); 3689 isl_aff_list_free(list); 3690 return NULL; 3691 } 3692 3693 /* Add variables to info->bmap and info->tab corresponding to the elements 3694 * in "list" that are not set to NaN. 3695 * "extra_var" is the number of these elements. 3696 * "dim" is the offset in the variables of "tab" where we should 3697 * start considering the elements in "list". 3698 * When this function returns, the total number of variables in "tab" 3699 * is equal to "dim" plus the number of elements in "list". 3700 * 3701 * The newly added existentially quantified variables are not given 3702 * an explicit representation because the corresponding div constraints 3703 * do not appear in info->bmap. These constraints are not added 3704 * to info->bmap because for internal consistency, they would need to 3705 * be added to info->tab as well, where they could combine with the equality 3706 * that is added later to result in constraints that do not hold 3707 * in the original input. 3708 */ 3709 static isl_stat add_sub_vars(struct isl_coalesce_info *info, 3710 __isl_keep isl_aff_list *list, int dim, int extra_var) 3711 { 3712 int i, j, d; 3713 isl_size n; 3714 3715 info->bmap = isl_basic_map_cow(info->bmap); 3716 info->bmap = isl_basic_map_extend(info->bmap, extra_var, 0, 0); 3717 n = isl_aff_list_n_aff(list); 3718 if (!info->bmap || n < 0) 3719 return isl_stat_error; 3720 for (i = 0; i < n; ++i) { 3721 int is_nan; 3722 isl_aff *aff; 3723 3724 aff = isl_aff_list_get_aff(list, i); 3725 is_nan = isl_aff_is_nan(aff); 3726 isl_aff_free(aff); 3727 if (is_nan < 0) 3728 return isl_stat_error; 3729 if (is_nan) 3730 continue; 3731 3732 if (isl_tab_insert_var(info->tab, dim + i) < 0) 3733 return isl_stat_error; 3734 d = isl_basic_map_alloc_div(info->bmap); 3735 if (d < 0) 3736 return isl_stat_error; 3737 info->bmap = isl_basic_map_mark_div_unknown(info->bmap, d); 3738 for (j = d; j > i; --j) 3739 info->bmap = isl_basic_map_swap_div(info->bmap, 3740 j - 1, j); 3741 if (!info->bmap) 3742 return isl_stat_error; 3743 } 3744 3745 return isl_stat_ok; 3746 } 3747 3748 /* For each element in "list" that is not set to NaN, fix the corresponding 3749 * variable in "tab" to the purely affine expression defined by the element. 3750 * "dim" is the offset in the variables of "tab" where we should 3751 * start considering the elements in "list". 3752 * 3753 * This function assumes that a sufficient number of rows and 3754 * elements in the constraint array are available in the tableau. 3755 */ 3756 static isl_stat add_sub_equalities(struct isl_tab *tab, 3757 __isl_keep isl_aff_list *list, int dim) 3758 { 3759 int i; 3760 isl_size n; 3761 isl_ctx *ctx; 3762 isl_vec *sub; 3763 isl_aff *aff; 3764 3765 n = isl_aff_list_n_aff(list); 3766 if (n < 0) 3767 return isl_stat_error; 3768 3769 ctx = isl_tab_get_ctx(tab); 3770 sub = isl_vec_alloc(ctx, 1 + dim + n); 3771 if (!sub) 3772 return isl_stat_error; 3773 isl_seq_clr(sub->el + 1 + dim, n); 3774 3775 for (i = 0; i < n; ++i) { 3776 aff = isl_aff_list_get_aff(list, i); 3777 if (!aff) 3778 goto error; 3779 if (isl_aff_is_nan(aff)) { 3780 isl_aff_free(aff); 3781 continue; 3782 } 3783 isl_seq_cpy(sub->el, aff->v->el + 1, 1 + dim); 3784 isl_int_neg(sub->el[1 + dim + i], aff->v->el[0]); 3785 if (isl_tab_add_eq(tab, sub->el) < 0) 3786 goto error; 3787 isl_int_set_si(sub->el[1 + dim + i], 0); 3788 isl_aff_free(aff); 3789 } 3790 3791 isl_vec_free(sub); 3792 return isl_stat_ok; 3793 error: 3794 isl_aff_free(aff); 3795 isl_vec_free(sub); 3796 return isl_stat_error; 3797 } 3798 3799 /* Add variables to info->tab and info->bmap corresponding to the elements 3800 * in "list" that are not set to NaN. The value of the added variable 3801 * in info->tab is fixed to the purely affine expression defined by the element. 3802 * "dim" is the offset in the variables of info->tab where we should 3803 * start considering the elements in "list". 3804 * When this function returns, the total number of variables in info->tab 3805 * is equal to "dim" plus the number of elements in "list". 3806 */ 3807 static isl_stat add_subs(struct isl_coalesce_info *info, 3808 __isl_keep isl_aff_list *list, int dim) 3809 { 3810 int extra_var; 3811 isl_size n; 3812 3813 n = isl_aff_list_n_aff(list); 3814 if (n < 0) 3815 return isl_stat_error; 3816 3817 extra_var = n - (info->tab->n_var - dim); 3818 3819 if (isl_tab_extend_vars(info->tab, extra_var) < 0) 3820 return isl_stat_error; 3821 if (isl_tab_extend_cons(info->tab, 2 * extra_var) < 0) 3822 return isl_stat_error; 3823 if (add_sub_vars(info, list, dim, extra_var) < 0) 3824 return isl_stat_error; 3825 3826 return add_sub_equalities(info->tab, list, dim); 3827 } 3828 3829 /* Coalesce basic map "j" into basic map "i" after adding the extra integer 3830 * divisions in "i" but not in "j" to basic map "j", with values 3831 * specified by "list". The total number of elements in "list" 3832 * is equal to the number of integer divisions in "i", while the number 3833 * of NaN elements in the list is equal to the number of integer divisions 3834 * in "j". 3835 * 3836 * If no coalescing can be performed, then we need to revert basic map "j" 3837 * to its original state. We do the same if basic map "i" gets dropped 3838 * during the coalescing, even though this should not happen in practice 3839 * since we have already checked for "j" being a subset of "i" 3840 * before we reach this stage. 3841 */ 3842 static enum isl_change coalesce_with_subs(int i, int j, 3843 struct isl_coalesce_info *info, __isl_keep isl_aff_list *list) 3844 { 3845 isl_basic_map *bmap_j; 3846 struct isl_tab_undo *snap; 3847 isl_size dim, n_div; 3848 enum isl_change change; 3849 3850 bmap_j = isl_basic_map_copy(info[j].bmap); 3851 snap = isl_tab_snap(info[j].tab); 3852 3853 dim = isl_basic_map_dim(bmap_j, isl_dim_all); 3854 n_div = isl_basic_map_dim(bmap_j, isl_dim_div); 3855 if (dim < 0 || n_div < 0) 3856 goto error; 3857 dim -= n_div; 3858 if (add_subs(&info[j], list, dim) < 0) 3859 goto error; 3860 3861 change = coalesce_local_pair(i, j, info); 3862 if (change != isl_change_none && change != isl_change_drop_first) { 3863 isl_basic_map_free(bmap_j); 3864 } else { 3865 isl_basic_map_free(info[j].bmap); 3866 info[j].bmap = bmap_j; 3867 3868 if (isl_tab_rollback(info[j].tab, snap) < 0) 3869 return isl_change_error; 3870 } 3871 3872 return change; 3873 error: 3874 isl_basic_map_free(bmap_j); 3875 return isl_change_error; 3876 } 3877 3878 /* Check if we can coalesce basic map "j" into basic map "i" after copying 3879 * those extra integer divisions in "i" that can be simplified away 3880 * using the extra equalities in "j". 3881 * All divs are assumed to be known and not contain any nested divs. 3882 * 3883 * We first check if there are any extra equalities in "j" that we 3884 * can exploit. Then we check if every integer division in "i" 3885 * either already appears in "j" or can be simplified using the 3886 * extra equalities to a purely affine expression. 3887 * If these tests succeed, then we try to coalesce the two basic maps 3888 * by introducing extra dimensions in "j" corresponding to 3889 * the extra integer divisions "i" fixed to the corresponding 3890 * purely affine expression. 3891 */ 3892 static enum isl_change check_coalesce_into_eq(int i, int j, 3893 struct isl_coalesce_info *info) 3894 { 3895 isl_size n_div_i, n_div_j, n; 3896 isl_basic_map *hull_i, *hull_j; 3897 isl_bool equal, empty; 3898 isl_aff_list *list; 3899 enum isl_change change; 3900 3901 n_div_i = isl_basic_map_dim(info[i].bmap, isl_dim_div); 3902 n_div_j = isl_basic_map_dim(info[j].bmap, isl_dim_div); 3903 if (n_div_i < 0 || n_div_j < 0) 3904 return isl_change_error; 3905 if (n_div_i <= n_div_j) 3906 return isl_change_none; 3907 if (info[j].bmap->n_eq == 0) 3908 return isl_change_none; 3909 3910 hull_i = isl_basic_map_copy(info[i].bmap); 3911 hull_i = isl_basic_map_plain_affine_hull(hull_i); 3912 hull_j = isl_basic_map_copy(info[j].bmap); 3913 hull_j = isl_basic_map_plain_affine_hull(hull_j); 3914 3915 hull_j = isl_basic_map_intersect(hull_j, isl_basic_map_copy(hull_i)); 3916 equal = isl_basic_map_plain_is_equal(hull_i, hull_j); 3917 empty = isl_basic_map_plain_is_empty(hull_j); 3918 isl_basic_map_free(hull_i); 3919 3920 if (equal < 0 || empty < 0) 3921 goto error; 3922 if (equal || empty) { 3923 isl_basic_map_free(hull_j); 3924 return isl_change_none; 3925 } 3926 3927 list = set_up_substitutions(info[i].bmap, info[j].bmap, hull_j); 3928 if (!list) 3929 return isl_change_error; 3930 n = isl_aff_list_n_aff(list); 3931 if (n < 0) 3932 change = isl_change_error; 3933 else if (n < n_div_i) 3934 change = isl_change_none; 3935 else 3936 change = coalesce_with_subs(i, j, info, list); 3937 3938 isl_aff_list_free(list); 3939 3940 return change; 3941 error: 3942 isl_basic_map_free(hull_j); 3943 return isl_change_error; 3944 } 3945 3946 /* Check if we can coalesce basic maps "i" and "j" after copying 3947 * those extra integer divisions in one of the basic maps that can 3948 * be simplified away using the extra equalities in the other basic map. 3949 * We require all divs to be known in both basic maps. 3950 * Furthermore, to simplify the comparison of div expressions, 3951 * we do not allow any nested integer divisions. 3952 */ 3953 static enum isl_change check_coalesce_eq(int i, int j, 3954 struct isl_coalesce_info *info) 3955 { 3956 isl_bool known, nested; 3957 enum isl_change change; 3958 3959 known = isl_basic_map_divs_known(info[i].bmap); 3960 if (known < 0 || !known) 3961 return known < 0 ? isl_change_error : isl_change_none; 3962 known = isl_basic_map_divs_known(info[j].bmap); 3963 if (known < 0 || !known) 3964 return known < 0 ? isl_change_error : isl_change_none; 3965 nested = has_nested_div(info[i].bmap); 3966 if (nested < 0 || nested) 3967 return nested < 0 ? isl_change_error : isl_change_none; 3968 nested = has_nested_div(info[j].bmap); 3969 if (nested < 0 || nested) 3970 return nested < 0 ? isl_change_error : isl_change_none; 3971 3972 change = check_coalesce_into_eq(i, j, info); 3973 if (change != isl_change_none) 3974 return change; 3975 change = check_coalesce_into_eq(j, i, info); 3976 if (change != isl_change_none) 3977 return invert_change(change); 3978 3979 return isl_change_none; 3980 } 3981 3982 /* Check if the union of the given pair of basic maps 3983 * can be represented by a single basic map. 3984 * If so, replace the pair by the single basic map and return 3985 * isl_change_drop_first, isl_change_drop_second or isl_change_fuse. 3986 * Otherwise, return isl_change_none. 3987 * 3988 * We first check if the two basic maps live in the same local space, 3989 * after aligning the divs that differ by only an integer constant. 3990 * If so, we do the complete check. Otherwise, we check if they have 3991 * the same number of integer divisions and can be coalesced, if one is 3992 * an obvious subset of the other or if the extra integer divisions 3993 * of one basic map can be simplified away using the extra equalities 3994 * of the other basic map. 3995 * 3996 * Note that trying to coalesce pairs of disjuncts with the same 3997 * number, but different local variables may drop the explicit 3998 * representation of some of these local variables. 3999 * This operation is therefore not performed when 4000 * the "coalesce_preserve_locals" option is set. 4001 */ 4002 static enum isl_change coalesce_pair(int i, int j, 4003 struct isl_coalesce_info *info) 4004 { 4005 int preserve; 4006 isl_bool same; 4007 enum isl_change change; 4008 isl_ctx *ctx; 4009 4010 if (harmonize_divs(&info[i], &info[j]) < 0) 4011 return isl_change_error; 4012 same = same_divs(info[i].bmap, info[j].bmap); 4013 if (same < 0) 4014 return isl_change_error; 4015 if (same) 4016 return coalesce_local_pair(i, j, info); 4017 4018 ctx = isl_basic_map_get_ctx(info[i].bmap); 4019 preserve = isl_options_get_coalesce_preserve_locals(ctx); 4020 if (!preserve && info[i].bmap->n_div == info[j].bmap->n_div) { 4021 change = coalesce_local_pair(i, j, info); 4022 if (change != isl_change_none) 4023 return change; 4024 } 4025 4026 change = coalesce_divs(i, j, info); 4027 if (change != isl_change_none) 4028 return change; 4029 4030 return check_coalesce_eq(i, j, info); 4031 } 4032 4033 /* Return the maximum of "a" and "b". 4034 */ 4035 static int isl_max(int a, int b) 4036 { 4037 return a > b ? a : b; 4038 } 4039 4040 /* Pairwise coalesce the basic maps in the range [start1, end1[ of "info" 4041 * with those in the range [start2, end2[, skipping basic maps 4042 * that have been removed (either before or within this function). 4043 * 4044 * For each basic map i in the first range, we check if it can be coalesced 4045 * with respect to any previously considered basic map j in the second range. 4046 * If i gets dropped (because it was a subset of some j), then 4047 * we can move on to the next basic map. 4048 * If j gets dropped, we need to continue checking against the other 4049 * previously considered basic maps. 4050 * If the two basic maps got fused, then we recheck the fused basic map 4051 * against the previously considered basic maps, starting at i + 1 4052 * (even if start2 is greater than i + 1). 4053 */ 4054 static int coalesce_range(isl_ctx *ctx, struct isl_coalesce_info *info, 4055 int start1, int end1, int start2, int end2) 4056 { 4057 int i, j; 4058 4059 for (i = end1 - 1; i >= start1; --i) { 4060 if (info[i].removed) 4061 continue; 4062 for (j = isl_max(i + 1, start2); j < end2; ++j) { 4063 enum isl_change changed; 4064 4065 if (info[j].removed) 4066 continue; 4067 if (info[i].removed) 4068 isl_die(ctx, isl_error_internal, 4069 "basic map unexpectedly removed", 4070 return -1); 4071 changed = coalesce_pair(i, j, info); 4072 switch (changed) { 4073 case isl_change_error: 4074 return -1; 4075 case isl_change_none: 4076 case isl_change_drop_second: 4077 continue; 4078 case isl_change_drop_first: 4079 j = end2; 4080 break; 4081 case isl_change_fuse: 4082 j = i; 4083 break; 4084 } 4085 } 4086 } 4087 4088 return 0; 4089 } 4090 4091 /* Pairwise coalesce the basic maps described by the "n" elements of "info". 4092 * 4093 * We consider groups of basic maps that live in the same apparent 4094 * affine hull and we first coalesce within such a group before we 4095 * coalesce the elements in the group with elements of previously 4096 * considered groups. If a fuse happens during the second phase, 4097 * then we also reconsider the elements within the group. 4098 */ 4099 static int coalesce(isl_ctx *ctx, int n, struct isl_coalesce_info *info) 4100 { 4101 int start, end; 4102 4103 for (end = n; end > 0; end = start) { 4104 start = end - 1; 4105 while (start >= 1 && 4106 info[start - 1].hull_hash == info[start].hull_hash) 4107 start--; 4108 if (coalesce_range(ctx, info, start, end, start, end) < 0) 4109 return -1; 4110 if (coalesce_range(ctx, info, start, end, end, n) < 0) 4111 return -1; 4112 } 4113 4114 return 0; 4115 } 4116 4117 /* Update the basic maps in "map" based on the information in "info". 4118 * In particular, remove the basic maps that have been marked removed and 4119 * update the others based on the information in the corresponding tableau. 4120 * Since we detected implicit equalities without calling 4121 * isl_basic_map_gauss, we need to do it now. 4122 * Also call isl_basic_map_simplify if we may have lost the definition 4123 * of one or more integer divisions. 4124 * If a basic map is still equal to the one from which the corresponding "info" 4125 * entry was created, then redundant constraint and 4126 * implicit equality constraint detection have been performed 4127 * on the corresponding tableau and the basic map can be marked as such. 4128 */ 4129 static __isl_give isl_map *update_basic_maps(__isl_take isl_map *map, 4130 int n, struct isl_coalesce_info *info) 4131 { 4132 int i; 4133 4134 if (!map) 4135 return NULL; 4136 4137 for (i = n - 1; i >= 0; --i) { 4138 if (info[i].removed) { 4139 isl_basic_map_free(map->p[i]); 4140 if (i != map->n - 1) 4141 map->p[i] = map->p[map->n - 1]; 4142 map->n--; 4143 continue; 4144 } 4145 4146 info[i].bmap = isl_basic_map_update_from_tab(info[i].bmap, 4147 info[i].tab); 4148 info[i].bmap = isl_basic_map_gauss(info[i].bmap, NULL); 4149 if (info[i].simplify) 4150 info[i].bmap = isl_basic_map_simplify(info[i].bmap); 4151 info[i].bmap = isl_basic_map_finalize(info[i].bmap); 4152 if (!info[i].bmap) 4153 return isl_map_free(map); 4154 if (!info[i].modified) { 4155 ISL_F_SET(info[i].bmap, ISL_BASIC_MAP_NO_IMPLICIT); 4156 ISL_F_SET(info[i].bmap, ISL_BASIC_MAP_NO_REDUNDANT); 4157 } 4158 isl_basic_map_free(map->p[i]); 4159 map->p[i] = info[i].bmap; 4160 info[i].bmap = NULL; 4161 } 4162 4163 return map; 4164 } 4165 4166 /* For each pair of basic maps in the map, check if the union of the two 4167 * can be represented by a single basic map. 4168 * If so, replace the pair by the single basic map and start over. 4169 * 4170 * We factor out any (hidden) common factor from the constraint 4171 * coefficients to improve the detection of adjacent constraints. 4172 * Note that this function does not call isl_basic_map_gauss, 4173 * but it does make sure that only a single copy of the basic map 4174 * is affected. This means that isl_basic_map_gauss may have 4175 * to be called at the end of the computation (in update_basic_maps) 4176 * on this single copy to ensure that 4177 * the basic maps are not left in an unexpected state. 4178 * 4179 * Since we are constructing the tableaus of the basic maps anyway, 4180 * we exploit them to detect implicit equalities and redundant constraints. 4181 * This also helps the coalescing as it can ignore the redundant constraints. 4182 * In order to avoid confusion, we make all implicit equalities explicit 4183 * in the basic maps. If the basic map only has a single reference 4184 * (this happens in particular if it was modified by 4185 * isl_basic_map_reduce_coefficients), then isl_basic_map_gauss 4186 * does not get called on the result. The call to 4187 * isl_basic_map_gauss in update_basic_maps resolves this as well. 4188 * For each basic map, we also compute the hash of the apparent affine hull 4189 * for use in coalesce. 4190 */ 4191 __isl_give isl_map *isl_map_coalesce(__isl_take isl_map *map) 4192 { 4193 int i; 4194 unsigned n; 4195 isl_ctx *ctx; 4196 struct isl_coalesce_info *info = NULL; 4197 4198 map = isl_map_remove_empty_parts(map); 4199 if (!map) 4200 return NULL; 4201 4202 if (map->n <= 1) 4203 return map; 4204 4205 ctx = isl_map_get_ctx(map); 4206 map = isl_map_sort_divs(map); 4207 map = isl_map_cow(map); 4208 4209 if (!map) 4210 return NULL; 4211 4212 n = map->n; 4213 4214 info = isl_calloc_array(map->ctx, struct isl_coalesce_info, n); 4215 if (!info) 4216 goto error; 4217 4218 for (i = 0; i < map->n; ++i) { 4219 map->p[i] = isl_basic_map_reduce_coefficients(map->p[i]); 4220 if (!map->p[i]) 4221 goto error; 4222 info[i].bmap = isl_basic_map_copy(map->p[i]); 4223 info[i].tab = isl_tab_from_basic_map(info[i].bmap, 0); 4224 if (!info[i].tab) 4225 goto error; 4226 if (!ISL_F_ISSET(info[i].bmap, ISL_BASIC_MAP_NO_IMPLICIT)) 4227 if (isl_tab_detect_implicit_equalities(info[i].tab) < 0) 4228 goto error; 4229 info[i].bmap = isl_tab_make_equalities_explicit(info[i].tab, 4230 info[i].bmap); 4231 if (!info[i].bmap) 4232 goto error; 4233 if (!ISL_F_ISSET(info[i].bmap, ISL_BASIC_MAP_NO_REDUNDANT)) 4234 if (isl_tab_detect_redundant(info[i].tab) < 0) 4235 goto error; 4236 if (coalesce_info_set_hull_hash(&info[i]) < 0) 4237 goto error; 4238 } 4239 for (i = map->n - 1; i >= 0; --i) 4240 if (info[i].tab->empty) 4241 drop(&info[i]); 4242 4243 if (coalesce(ctx, n, info) < 0) 4244 goto error; 4245 4246 map = update_basic_maps(map, n, info); 4247 4248 clear_coalesce_info(n, info); 4249 4250 return map; 4251 error: 4252 clear_coalesce_info(n, info); 4253 isl_map_free(map); 4254 return NULL; 4255 } 4256 4257 /* For each pair of basic sets in the set, check if the union of the two 4258 * can be represented by a single basic set. 4259 * If so, replace the pair by the single basic set and start over. 4260 */ 4261 __isl_give isl_set *isl_set_coalesce(__isl_take isl_set *set) 4262 { 4263 return set_from_map(isl_map_coalesce(set_to_map(set))); 4264 } 4265