1 1.1 mrg /* 2 1.1 mrg * Copyright 2008-2009 Katholieke Universiteit Leuven 3 1.1 mrg * Copyright 2010 INRIA Saclay 4 1.1 mrg * Copyright 2012-2013 Ecole Normale Superieure 5 1.1 mrg * Copyright 2014 INRIA Rocquencourt 6 1.1 mrg * Copyright 2016 INRIA Paris 7 1.1 mrg * Copyright 2020 Cerebras Systems 8 1.1 mrg * 9 1.1 mrg * Use of this software is governed by the MIT license 10 1.1 mrg * 11 1.1 mrg * Written by Sven Verdoolaege, K.U.Leuven, Departement 12 1.1 mrg * Computerwetenschappen, Celestijnenlaan 200A, B-3001 Leuven, Belgium 13 1.1 mrg * and INRIA Saclay - Ile-de-France, Parc Club Orsay Universite, 14 1.1 mrg * ZAC des vignes, 4 rue Jacques Monod, 91893 Orsay, France 15 1.1 mrg * and Ecole Normale Superieure, 45 rue dUlm, 75230 Paris, France 16 1.1 mrg * and Inria Paris - Rocquencourt, Domaine de Voluceau - Rocquencourt, 17 1.1 mrg * B.P. 105 - 78153 Le Chesnay, France 18 1.1 mrg * and Centre de Recherche Inria de Paris, 2 rue Simone Iff - Voie DQ12, 19 1.1 mrg * CS 42112, 75589 Paris Cedex 12, France 20 1.1 mrg * and Cerebras Systems, 175 S San Antonio Rd, Los Altos, CA, USA 21 1.1 mrg */ 22 1.1 mrg 23 1.1 mrg #include <isl_ctx_private.h> 24 1.1 mrg #include "isl_map_private.h" 25 1.1 mrg #include <isl_seq.h> 26 1.1 mrg #include <isl/options.h> 27 1.1 mrg #include "isl_tab.h" 28 1.1 mrg #include <isl_mat_private.h> 29 1.1 mrg #include <isl_local_space_private.h> 30 1.1 mrg #include <isl_val_private.h> 31 1.1 mrg #include <isl_vec_private.h> 32 1.1 mrg #include <isl_aff_private.h> 33 1.1 mrg #include <isl_equalities.h> 34 1.1 mrg #include <isl_constraint_private.h> 35 1.1 mrg 36 1.1 mrg #include <set_to_map.c> 37 1.1 mrg #include <set_from_map.c> 38 1.1 mrg 39 1.1 mrg #define STATUS_ERROR -1 40 1.1 mrg #define STATUS_REDUNDANT 1 41 1.1 mrg #define STATUS_VALID 2 42 1.1 mrg #define STATUS_SEPARATE 3 43 1.1 mrg #define STATUS_CUT 4 44 1.1 mrg #define STATUS_ADJ_EQ 5 45 1.1 mrg #define STATUS_ADJ_INEQ 6 46 1.1 mrg 47 1.1 mrg static int status_in(isl_int *ineq, struct isl_tab *tab) 48 1.1 mrg { 49 1.1 mrg enum isl_ineq_type type = isl_tab_ineq_type(tab, ineq); 50 1.1 mrg switch (type) { 51 1.1 mrg default: 52 1.1 mrg case isl_ineq_error: return STATUS_ERROR; 53 1.1 mrg case isl_ineq_redundant: return STATUS_VALID; 54 1.1 mrg case isl_ineq_separate: return STATUS_SEPARATE; 55 1.1 mrg case isl_ineq_cut: return STATUS_CUT; 56 1.1 mrg case isl_ineq_adj_eq: return STATUS_ADJ_EQ; 57 1.1 mrg case isl_ineq_adj_ineq: return STATUS_ADJ_INEQ; 58 1.1 mrg } 59 1.1 mrg } 60 1.1 mrg 61 1.1 mrg /* Compute the position of the equalities of basic map "bmap_i" 62 1.1 mrg * with respect to the basic map represented by "tab_j". 63 1.1 mrg * The resulting array has twice as many entries as the number 64 1.1 mrg * of equalities corresponding to the two inequalities to which 65 1.1 mrg * each equality corresponds. 66 1.1 mrg */ 67 1.1 mrg static int *eq_status_in(__isl_keep isl_basic_map *bmap_i, 68 1.1 mrg struct isl_tab *tab_j) 69 1.1 mrg { 70 1.1 mrg int k, l; 71 1.1 mrg int *eq; 72 1.1 mrg isl_size dim; 73 1.1 mrg 74 1.1 mrg dim = isl_basic_map_dim(bmap_i, isl_dim_all); 75 1.1 mrg if (dim < 0) 76 1.1 mrg return NULL; 77 1.1 mrg 78 1.1 mrg eq = isl_calloc_array(bmap_i->ctx, int, 2 * bmap_i->n_eq); 79 1.1 mrg if (!eq) 80 1.1 mrg return NULL; 81 1.1 mrg 82 1.1 mrg for (k = 0; k < bmap_i->n_eq; ++k) { 83 1.1 mrg for (l = 0; l < 2; ++l) { 84 1.1 mrg isl_seq_neg(bmap_i->eq[k], bmap_i->eq[k], 1+dim); 85 1.1 mrg eq[2 * k + l] = status_in(bmap_i->eq[k], tab_j); 86 1.1 mrg if (eq[2 * k + l] == STATUS_ERROR) 87 1.1 mrg goto error; 88 1.1 mrg } 89 1.1 mrg } 90 1.1 mrg 91 1.1 mrg return eq; 92 1.1 mrg error: 93 1.1 mrg free(eq); 94 1.1 mrg return NULL; 95 1.1 mrg } 96 1.1 mrg 97 1.1 mrg /* Compute the position of the inequalities of basic map "bmap_i" 98 1.1 mrg * (also represented by "tab_i", if not NULL) with respect to the basic map 99 1.1 mrg * represented by "tab_j". 100 1.1 mrg */ 101 1.1 mrg static int *ineq_status_in(__isl_keep isl_basic_map *bmap_i, 102 1.1 mrg struct isl_tab *tab_i, struct isl_tab *tab_j) 103 1.1 mrg { 104 1.1 mrg int k; 105 1.1 mrg unsigned n_eq = bmap_i->n_eq; 106 1.1 mrg int *ineq = isl_calloc_array(bmap_i->ctx, int, bmap_i->n_ineq); 107 1.1 mrg 108 1.1 mrg if (!ineq) 109 1.1 mrg return NULL; 110 1.1 mrg 111 1.1 mrg for (k = 0; k < bmap_i->n_ineq; ++k) { 112 1.1 mrg if (tab_i && isl_tab_is_redundant(tab_i, n_eq + k)) { 113 1.1 mrg ineq[k] = STATUS_REDUNDANT; 114 1.1 mrg continue; 115 1.1 mrg } 116 1.1 mrg ineq[k] = status_in(bmap_i->ineq[k], tab_j); 117 1.1 mrg if (ineq[k] == STATUS_ERROR) 118 1.1 mrg goto error; 119 1.1 mrg if (ineq[k] == STATUS_SEPARATE) 120 1.1 mrg break; 121 1.1 mrg } 122 1.1 mrg 123 1.1 mrg return ineq; 124 1.1 mrg error: 125 1.1 mrg free(ineq); 126 1.1 mrg return NULL; 127 1.1 mrg } 128 1.1 mrg 129 1.1 mrg static int any(int *con, unsigned len, int status) 130 1.1 mrg { 131 1.1 mrg int i; 132 1.1 mrg 133 1.1 mrg for (i = 0; i < len ; ++i) 134 1.1 mrg if (con[i] == status) 135 1.1 mrg return 1; 136 1.1 mrg return 0; 137 1.1 mrg } 138 1.1 mrg 139 1.1 mrg /* Return the first position of "status" in the list "con" of length "len". 140 1.1 mrg * Return -1 if there is no such entry. 141 1.1 mrg */ 142 1.1 mrg static int find(int *con, unsigned len, int status) 143 1.1 mrg { 144 1.1 mrg int i; 145 1.1 mrg 146 1.1 mrg for (i = 0; i < len ; ++i) 147 1.1 mrg if (con[i] == status) 148 1.1 mrg return i; 149 1.1 mrg return -1; 150 1.1 mrg } 151 1.1 mrg 152 1.1 mrg static int count(int *con, unsigned len, int status) 153 1.1 mrg { 154 1.1 mrg int i; 155 1.1 mrg int c = 0; 156 1.1 mrg 157 1.1 mrg for (i = 0; i < len ; ++i) 158 1.1 mrg if (con[i] == status) 159 1.1 mrg c++; 160 1.1 mrg return c; 161 1.1 mrg } 162 1.1 mrg 163 1.1 mrg static int all(int *con, unsigned len, int status) 164 1.1 mrg { 165 1.1 mrg int i; 166 1.1 mrg 167 1.1 mrg for (i = 0; i < len ; ++i) { 168 1.1 mrg if (con[i] == STATUS_REDUNDANT) 169 1.1 mrg continue; 170 1.1 mrg if (con[i] != status) 171 1.1 mrg return 0; 172 1.1 mrg } 173 1.1 mrg return 1; 174 1.1 mrg } 175 1.1 mrg 176 1.1 mrg /* Internal information associated to a basic map in a map 177 1.1 mrg * that is to be coalesced by isl_map_coalesce. 178 1.1 mrg * 179 1.1 mrg * "bmap" is the basic map itself (or NULL if "removed" is set) 180 1.1 mrg * "tab" is the corresponding tableau (or NULL if "removed" is set) 181 1.1 mrg * "hull_hash" identifies the affine space in which "bmap" lives. 182 1.1 mrg * "modified" is set if this basic map may not be identical 183 1.1 mrg * to any of the basic maps in the input. 184 1.1 mrg * "removed" is set if this basic map has been removed from the map 185 1.1 mrg * "simplify" is set if this basic map may have some unknown integer 186 1.1 mrg * divisions that were not present in the input basic maps. The basic 187 1.1 mrg * map should then be simplified such that we may be able to find 188 1.1 mrg * a definition among the constraints. 189 1.1 mrg * 190 1.1 mrg * "eq" and "ineq" are only set if we are currently trying to coalesce 191 1.1 mrg * this basic map with another basic map, in which case they represent 192 1.1 mrg * the position of the inequalities of this basic map with respect to 193 1.1 mrg * the other basic map. The number of elements in the "eq" array 194 1.1 mrg * is twice the number of equalities in the "bmap", corresponding 195 1.1 mrg * to the two inequalities that make up each equality. 196 1.1 mrg */ 197 1.1 mrg struct isl_coalesce_info { 198 1.1 mrg isl_basic_map *bmap; 199 1.1 mrg struct isl_tab *tab; 200 1.1 mrg uint32_t hull_hash; 201 1.1 mrg int modified; 202 1.1 mrg int removed; 203 1.1 mrg int simplify; 204 1.1 mrg int *eq; 205 1.1 mrg int *ineq; 206 1.1 mrg }; 207 1.1 mrg 208 1.1 mrg /* Is there any (half of an) equality constraint in the description 209 1.1 mrg * of the basic map represented by "info" that 210 1.1 mrg * has position "status" with respect to the other basic map? 211 1.1 mrg */ 212 1.1 mrg static int any_eq(struct isl_coalesce_info *info, int status) 213 1.1 mrg { 214 1.1 mrg isl_size n_eq; 215 1.1 mrg 216 1.1 mrg n_eq = isl_basic_map_n_equality(info->bmap); 217 1.1 mrg return any(info->eq, 2 * n_eq, status); 218 1.1 mrg } 219 1.1 mrg 220 1.1 mrg /* Is there any inequality constraint in the description 221 1.1 mrg * of the basic map represented by "info" that 222 1.1 mrg * has position "status" with respect to the other basic map? 223 1.1 mrg */ 224 1.1 mrg static int any_ineq(struct isl_coalesce_info *info, int status) 225 1.1 mrg { 226 1.1 mrg isl_size n_ineq; 227 1.1 mrg 228 1.1 mrg n_ineq = isl_basic_map_n_inequality(info->bmap); 229 1.1 mrg return any(info->ineq, n_ineq, status); 230 1.1 mrg } 231 1.1 mrg 232 1.1 mrg /* Return the position of the first half on an equality constraint 233 1.1 mrg * in the description of the basic map represented by "info" that 234 1.1 mrg * has position "status" with respect to the other basic map. 235 1.1 mrg * The returned value is twice the position of the equality constraint 236 1.1 mrg * plus zero for the negative half and plus one for the positive half. 237 1.1 mrg * Return -1 if there is no such entry. 238 1.1 mrg */ 239 1.1 mrg static int find_eq(struct isl_coalesce_info *info, int status) 240 1.1 mrg { 241 1.1 mrg isl_size n_eq; 242 1.1 mrg 243 1.1 mrg n_eq = isl_basic_map_n_equality(info->bmap); 244 1.1 mrg return find(info->eq, 2 * n_eq, status); 245 1.1 mrg } 246 1.1 mrg 247 1.1 mrg /* Return the position of the first inequality constraint in the description 248 1.1 mrg * of the basic map represented by "info" that 249 1.1 mrg * has position "status" with respect to the other basic map. 250 1.1 mrg * Return -1 if there is no such entry. 251 1.1 mrg */ 252 1.1 mrg static int find_ineq(struct isl_coalesce_info *info, int status) 253 1.1 mrg { 254 1.1 mrg isl_size n_ineq; 255 1.1 mrg 256 1.1 mrg n_ineq = isl_basic_map_n_inequality(info->bmap); 257 1.1 mrg return find(info->ineq, n_ineq, status); 258 1.1 mrg } 259 1.1 mrg 260 1.1 mrg /* Return the number of (halves of) equality constraints in the description 261 1.1 mrg * of the basic map represented by "info" that 262 1.1 mrg * have position "status" with respect to the other basic map. 263 1.1 mrg */ 264 1.1 mrg static int count_eq(struct isl_coalesce_info *info, int status) 265 1.1 mrg { 266 1.1 mrg isl_size n_eq; 267 1.1 mrg 268 1.1 mrg n_eq = isl_basic_map_n_equality(info->bmap); 269 1.1 mrg return count(info->eq, 2 * n_eq, status); 270 1.1 mrg } 271 1.1 mrg 272 1.1 mrg /* Return the number of inequality constraints in the description 273 1.1 mrg * of the basic map represented by "info" that 274 1.1 mrg * have position "status" with respect to the other basic map. 275 1.1 mrg */ 276 1.1 mrg static int count_ineq(struct isl_coalesce_info *info, int status) 277 1.1 mrg { 278 1.1 mrg isl_size n_ineq; 279 1.1 mrg 280 1.1 mrg n_ineq = isl_basic_map_n_inequality(info->bmap); 281 1.1 mrg return count(info->ineq, n_ineq, status); 282 1.1 mrg } 283 1.1 mrg 284 1.1 mrg /* Are all non-redundant constraints of the basic map represented by "info" 285 1.1 mrg * either valid or cut constraints with respect to the other basic map? 286 1.1 mrg */ 287 1.1 mrg static int all_valid_or_cut(struct isl_coalesce_info *info) 288 1.1 mrg { 289 1.1 mrg int i; 290 1.1 mrg 291 1.1 mrg for (i = 0; i < 2 * info->bmap->n_eq; ++i) { 292 1.1 mrg if (info->eq[i] == STATUS_REDUNDANT) 293 1.1 mrg continue; 294 1.1 mrg if (info->eq[i] == STATUS_VALID) 295 1.1 mrg continue; 296 1.1 mrg if (info->eq[i] == STATUS_CUT) 297 1.1 mrg continue; 298 1.1 mrg return 0; 299 1.1 mrg } 300 1.1 mrg 301 1.1 mrg for (i = 0; i < info->bmap->n_ineq; ++i) { 302 1.1 mrg if (info->ineq[i] == STATUS_REDUNDANT) 303 1.1 mrg continue; 304 1.1 mrg if (info->ineq[i] == STATUS_VALID) 305 1.1 mrg continue; 306 1.1 mrg if (info->ineq[i] == STATUS_CUT) 307 1.1 mrg continue; 308 1.1 mrg return 0; 309 1.1 mrg } 310 1.1 mrg 311 1.1 mrg return 1; 312 1.1 mrg } 313 1.1 mrg 314 1.1 mrg /* Compute the hash of the (apparent) affine hull of info->bmap (with 315 1.1 mrg * the existentially quantified variables removed) and store it 316 1.1 mrg * in info->hash. 317 1.1 mrg */ 318 1.1 mrg static int coalesce_info_set_hull_hash(struct isl_coalesce_info *info) 319 1.1 mrg { 320 1.1 mrg isl_basic_map *hull; 321 1.1 mrg isl_size n_div; 322 1.1 mrg 323 1.1 mrg hull = isl_basic_map_copy(info->bmap); 324 1.1 mrg hull = isl_basic_map_plain_affine_hull(hull); 325 1.1 mrg n_div = isl_basic_map_dim(hull, isl_dim_div); 326 1.1 mrg if (n_div < 0) 327 1.1 mrg hull = isl_basic_map_free(hull); 328 1.1 mrg hull = isl_basic_map_drop_constraints_involving_dims(hull, 329 1.1 mrg isl_dim_div, 0, n_div); 330 1.1 mrg info->hull_hash = isl_basic_map_get_hash(hull); 331 1.1 mrg isl_basic_map_free(hull); 332 1.1 mrg 333 1.1 mrg return hull ? 0 : -1; 334 1.1 mrg } 335 1.1 mrg 336 1.1 mrg /* Free all the allocated memory in an array 337 1.1 mrg * of "n" isl_coalesce_info elements. 338 1.1 mrg */ 339 1.1 mrg static void clear_coalesce_info(int n, struct isl_coalesce_info *info) 340 1.1 mrg { 341 1.1 mrg int i; 342 1.1 mrg 343 1.1 mrg if (!info) 344 1.1 mrg return; 345 1.1 mrg 346 1.1 mrg for (i = 0; i < n; ++i) { 347 1.1 mrg isl_basic_map_free(info[i].bmap); 348 1.1 mrg isl_tab_free(info[i].tab); 349 1.1 mrg } 350 1.1 mrg 351 1.1 mrg free(info); 352 1.1 mrg } 353 1.1 mrg 354 1.1 mrg /* Clear the memory associated to "info". 355 1.1 mrg */ 356 1.1 mrg static void clear(struct isl_coalesce_info *info) 357 1.1 mrg { 358 1.1 mrg info->bmap = isl_basic_map_free(info->bmap); 359 1.1 mrg isl_tab_free(info->tab); 360 1.1 mrg info->tab = NULL; 361 1.1 mrg } 362 1.1 mrg 363 1.1 mrg /* Drop the basic map represented by "info". 364 1.1 mrg * That is, clear the memory associated to the entry and 365 1.1 mrg * mark it as having been removed. 366 1.1 mrg */ 367 1.1 mrg static void drop(struct isl_coalesce_info *info) 368 1.1 mrg { 369 1.1 mrg clear(info); 370 1.1 mrg info->removed = 1; 371 1.1 mrg } 372 1.1 mrg 373 1.1 mrg /* Exchange the information in "info1" with that in "info2". 374 1.1 mrg */ 375 1.1 mrg static void exchange(struct isl_coalesce_info *info1, 376 1.1 mrg struct isl_coalesce_info *info2) 377 1.1 mrg { 378 1.1 mrg struct isl_coalesce_info info; 379 1.1 mrg 380 1.1 mrg info = *info1; 381 1.1 mrg *info1 = *info2; 382 1.1 mrg *info2 = info; 383 1.1 mrg } 384 1.1 mrg 385 1.1 mrg /* This type represents the kind of change that has been performed 386 1.1 mrg * while trying to coalesce two basic maps. 387 1.1 mrg * 388 1.1 mrg * isl_change_none: nothing was changed 389 1.1 mrg * isl_change_drop_first: the first basic map was removed 390 1.1 mrg * isl_change_drop_second: the second basic map was removed 391 1.1 mrg * isl_change_fuse: the two basic maps were replaced by a new basic map. 392 1.1 mrg */ 393 1.1 mrg enum isl_change { 394 1.1 mrg isl_change_error = -1, 395 1.1 mrg isl_change_none = 0, 396 1.1 mrg isl_change_drop_first, 397 1.1 mrg isl_change_drop_second, 398 1.1 mrg isl_change_fuse, 399 1.1 mrg }; 400 1.1 mrg 401 1.1 mrg /* Update "change" based on an interchange of the first and the second 402 1.1 mrg * basic map. That is, interchange isl_change_drop_first and 403 1.1 mrg * isl_change_drop_second. 404 1.1 mrg */ 405 1.1 mrg static enum isl_change invert_change(enum isl_change change) 406 1.1 mrg { 407 1.1 mrg switch (change) { 408 1.1 mrg case isl_change_error: 409 1.1 mrg return isl_change_error; 410 1.1 mrg case isl_change_none: 411 1.1 mrg return isl_change_none; 412 1.1 mrg case isl_change_drop_first: 413 1.1 mrg return isl_change_drop_second; 414 1.1 mrg case isl_change_drop_second: 415 1.1 mrg return isl_change_drop_first; 416 1.1 mrg case isl_change_fuse: 417 1.1 mrg return isl_change_fuse; 418 1.1 mrg } 419 1.1 mrg 420 1.1 mrg return isl_change_error; 421 1.1 mrg } 422 1.1 mrg 423 1.1 mrg /* Add the valid constraints of the basic map represented by "info" 424 1.1 mrg * to "bmap". "len" is the size of the constraints. 425 1.1 mrg * If only one of the pair of inequalities that make up an equality 426 1.1 mrg * is valid, then add that inequality. 427 1.1 mrg */ 428 1.1 mrg static __isl_give isl_basic_map *add_valid_constraints( 429 1.1 mrg __isl_take isl_basic_map *bmap, struct isl_coalesce_info *info, 430 1.1 mrg unsigned len) 431 1.1 mrg { 432 1.1 mrg int k, l; 433 1.1 mrg 434 1.1 mrg if (!bmap) 435 1.1 mrg return NULL; 436 1.1 mrg 437 1.1 mrg for (k = 0; k < info->bmap->n_eq; ++k) { 438 1.1 mrg if (info->eq[2 * k] == STATUS_VALID && 439 1.1 mrg info->eq[2 * k + 1] == STATUS_VALID) { 440 1.1 mrg l = isl_basic_map_alloc_equality(bmap); 441 1.1 mrg if (l < 0) 442 1.1 mrg return isl_basic_map_free(bmap); 443 1.1 mrg isl_seq_cpy(bmap->eq[l], info->bmap->eq[k], len); 444 1.1 mrg } else if (info->eq[2 * k] == STATUS_VALID) { 445 1.1 mrg l = isl_basic_map_alloc_inequality(bmap); 446 1.1 mrg if (l < 0) 447 1.1 mrg return isl_basic_map_free(bmap); 448 1.1 mrg isl_seq_neg(bmap->ineq[l], info->bmap->eq[k], len); 449 1.1 mrg } else if (info->eq[2 * k + 1] == STATUS_VALID) { 450 1.1 mrg l = isl_basic_map_alloc_inequality(bmap); 451 1.1 mrg if (l < 0) 452 1.1 mrg return isl_basic_map_free(bmap); 453 1.1 mrg isl_seq_cpy(bmap->ineq[l], info->bmap->eq[k], len); 454 1.1 mrg } 455 1.1 mrg } 456 1.1 mrg 457 1.1 mrg for (k = 0; k < info->bmap->n_ineq; ++k) { 458 1.1 mrg if (info->ineq[k] != STATUS_VALID) 459 1.1 mrg continue; 460 1.1 mrg l = isl_basic_map_alloc_inequality(bmap); 461 1.1 mrg if (l < 0) 462 1.1 mrg return isl_basic_map_free(bmap); 463 1.1 mrg isl_seq_cpy(bmap->ineq[l], info->bmap->ineq[k], len); 464 1.1 mrg } 465 1.1 mrg 466 1.1 mrg return bmap; 467 1.1 mrg } 468 1.1 mrg 469 1.1 mrg /* Is "bmap" defined by a number of (non-redundant) constraints that 470 1.1 mrg * is greater than the number of constraints of basic maps i and j combined? 471 1.1 mrg * Equalities are counted as two inequalities. 472 1.1 mrg */ 473 1.1 mrg static int number_of_constraints_increases(int i, int j, 474 1.1 mrg struct isl_coalesce_info *info, 475 1.1 mrg __isl_keep isl_basic_map *bmap, struct isl_tab *tab) 476 1.1 mrg { 477 1.1 mrg int k, n_old, n_new; 478 1.1 mrg 479 1.1 mrg n_old = 2 * info[i].bmap->n_eq + info[i].bmap->n_ineq; 480 1.1 mrg n_old += 2 * info[j].bmap->n_eq + info[j].bmap->n_ineq; 481 1.1 mrg 482 1.1 mrg n_new = 2 * bmap->n_eq; 483 1.1 mrg for (k = 0; k < bmap->n_ineq; ++k) 484 1.1 mrg if (!isl_tab_is_redundant(tab, bmap->n_eq + k)) 485 1.1 mrg ++n_new; 486 1.1 mrg 487 1.1 mrg return n_new > n_old; 488 1.1 mrg } 489 1.1 mrg 490 1.1 mrg /* Replace the pair of basic maps i and j by the basic map bounded 491 1.1 mrg * by the valid constraints in both basic maps and the constraints 492 1.1 mrg * in extra (if not NULL). 493 1.1 mrg * Place the fused basic map in the position that is the smallest of i and j. 494 1.1 mrg * 495 1.1 mrg * If "detect_equalities" is set, then look for equalities encoded 496 1.1 mrg * as pairs of inequalities. 497 1.1 mrg * If "check_number" is set, then the original basic maps are only 498 1.1 mrg * replaced if the total number of constraints does not increase. 499 1.1 mrg * While the number of integer divisions in the two basic maps 500 1.1 mrg * is assumed to be the same, the actual definitions may be different. 501 1.1 mrg * We only copy the definition from one of the basic maps if it is 502 1.1 mrg * the same as that of the other basic map. Otherwise, we mark 503 1.1 mrg * the integer division as unknown and simplify the basic map 504 1.1 mrg * in an attempt to recover the integer division definition. 505 1.1 mrg * If any extra constraints get introduced, then these may 506 1.1 mrg * involve integer divisions with a unit coefficient. 507 1.1 mrg * Eliminate those that do not appear with any other coefficient 508 1.1 mrg * in other constraints, to ensure they get eliminated completely, 509 1.1 mrg * improving the chances of further coalescing. 510 1.1 mrg */ 511 1.1 mrg static enum isl_change fuse(int i, int j, struct isl_coalesce_info *info, 512 1.1 mrg __isl_keep isl_mat *extra, int detect_equalities, int check_number) 513 1.1 mrg { 514 1.1 mrg int k, l; 515 1.1 mrg struct isl_basic_map *fused = NULL; 516 1.1 mrg struct isl_tab *fused_tab = NULL; 517 1.1 mrg isl_size total = isl_basic_map_dim(info[i].bmap, isl_dim_all); 518 1.1 mrg unsigned extra_rows = extra ? extra->n_row : 0; 519 1.1 mrg unsigned n_eq, n_ineq; 520 1.1 mrg int simplify = 0; 521 1.1 mrg 522 1.1 mrg if (total < 0) 523 1.1 mrg return isl_change_error; 524 1.1 mrg if (j < i) 525 1.1 mrg return fuse(j, i, info, extra, detect_equalities, check_number); 526 1.1 mrg 527 1.1 mrg n_eq = info[i].bmap->n_eq + info[j].bmap->n_eq; 528 1.1 mrg n_ineq = info[i].bmap->n_ineq + info[j].bmap->n_ineq; 529 1.1 mrg fused = isl_basic_map_alloc_space(isl_space_copy(info[i].bmap->dim), 530 1.1 mrg info[i].bmap->n_div, n_eq, n_eq + n_ineq + extra_rows); 531 1.1 mrg fused = add_valid_constraints(fused, &info[i], 1 + total); 532 1.1 mrg fused = add_valid_constraints(fused, &info[j], 1 + total); 533 1.1 mrg if (!fused) 534 1.1 mrg goto error; 535 1.1 mrg if (ISL_F_ISSET(info[i].bmap, ISL_BASIC_MAP_RATIONAL) && 536 1.1 mrg ISL_F_ISSET(info[j].bmap, ISL_BASIC_MAP_RATIONAL)) 537 1.1 mrg ISL_F_SET(fused, ISL_BASIC_MAP_RATIONAL); 538 1.1 mrg 539 1.1 mrg for (k = 0; k < info[i].bmap->n_div; ++k) { 540 1.1 mrg int l = isl_basic_map_alloc_div(fused); 541 1.1 mrg if (l < 0) 542 1.1 mrg goto error; 543 1.1 mrg if (isl_seq_eq(info[i].bmap->div[k], info[j].bmap->div[k], 544 1.1 mrg 1 + 1 + total)) { 545 1.1 mrg isl_seq_cpy(fused->div[l], info[i].bmap->div[k], 546 1.1 mrg 1 + 1 + total); 547 1.1 mrg } else { 548 1.1 mrg isl_int_set_si(fused->div[l][0], 0); 549 1.1 mrg simplify = 1; 550 1.1 mrg } 551 1.1 mrg } 552 1.1 mrg 553 1.1 mrg for (k = 0; k < extra_rows; ++k) { 554 1.1 mrg l = isl_basic_map_alloc_inequality(fused); 555 1.1 mrg if (l < 0) 556 1.1 mrg goto error; 557 1.1 mrg isl_seq_cpy(fused->ineq[l], extra->row[k], 1 + total); 558 1.1 mrg } 559 1.1 mrg 560 1.1 mrg if (detect_equalities) 561 1.1 mrg fused = isl_basic_map_detect_inequality_pairs(fused, NULL); 562 1.1 mrg fused = isl_basic_map_gauss(fused, NULL); 563 1.1 mrg if (simplify || info[j].simplify) { 564 1.1 mrg fused = isl_basic_map_simplify(fused); 565 1.1 mrg info[i].simplify = 0; 566 1.1 mrg } else if (extra_rows > 0) { 567 1.1 mrg fused = isl_basic_map_eliminate_pure_unit_divs(fused); 568 1.1 mrg } 569 1.1 mrg fused = isl_basic_map_finalize(fused); 570 1.1 mrg 571 1.1 mrg fused_tab = isl_tab_from_basic_map(fused, 0); 572 1.1 mrg if (isl_tab_detect_redundant(fused_tab) < 0) 573 1.1 mrg goto error; 574 1.1 mrg 575 1.1 mrg if (check_number && 576 1.1 mrg number_of_constraints_increases(i, j, info, fused, fused_tab)) { 577 1.1 mrg isl_tab_free(fused_tab); 578 1.1 mrg isl_basic_map_free(fused); 579 1.1 mrg return isl_change_none; 580 1.1 mrg } 581 1.1 mrg 582 1.1 mrg clear(&info[i]); 583 1.1 mrg info[i].bmap = fused; 584 1.1 mrg info[i].tab = fused_tab; 585 1.1 mrg info[i].modified = 1; 586 1.1 mrg drop(&info[j]); 587 1.1 mrg 588 1.1 mrg return isl_change_fuse; 589 1.1 mrg error: 590 1.1 mrg isl_tab_free(fused_tab); 591 1.1 mrg isl_basic_map_free(fused); 592 1.1 mrg return isl_change_error; 593 1.1 mrg } 594 1.1 mrg 595 1.1 mrg /* Given a pair of basic maps i and j such that all constraints are either 596 1.1 mrg * "valid" or "cut", check if the facets corresponding to the "cut" 597 1.1 mrg * constraints of i lie entirely within basic map j. 598 1.1 mrg * If so, replace the pair by the basic map consisting of the valid 599 1.1 mrg * constraints in both basic maps. 600 1.1 mrg * Checking whether the facet lies entirely within basic map j 601 1.1 mrg * is performed by checking whether the constraints of basic map j 602 1.1 mrg * are valid for the facet. These tests are performed on a rational 603 1.1 mrg * tableau to avoid the theoretical possibility that a constraint 604 1.1 mrg * that was considered to be a cut constraint for the entire basic map i 605 1.1 mrg * happens to be considered to be a valid constraint for the facet, 606 1.1 mrg * even though it cuts off the same rational points. 607 1.1 mrg * 608 1.1 mrg * To see that we are not introducing any extra points, call the 609 1.1 mrg * two basic maps A and B and the resulting map U and let x 610 1.1 mrg * be an element of U \setminus ( A \cup B ). 611 1.1 mrg * A line connecting x with an element of A \cup B meets a facet F 612 1.1 mrg * of either A or B. Assume it is a facet of B and let c_1 be 613 1.1 mrg * the corresponding facet constraint. We have c_1(x) < 0 and 614 1.1 mrg * so c_1 is a cut constraint. This implies that there is some 615 1.1 mrg * (possibly rational) point x' satisfying the constraints of A 616 1.1 mrg * and the opposite of c_1 as otherwise c_1 would have been marked 617 1.1 mrg * valid for A. The line connecting x and x' meets a facet of A 618 1.1 mrg * in a (possibly rational) point that also violates c_1, but this 619 1.1 mrg * is impossible since all cut constraints of B are valid for all 620 1.1 mrg * cut facets of A. 621 1.1 mrg * In case F is a facet of A rather than B, then we can apply the 622 1.1 mrg * above reasoning to find a facet of B separating x from A \cup B first. 623 1.1 mrg */ 624 1.1 mrg static enum isl_change check_facets(int i, int j, 625 1.1 mrg struct isl_coalesce_info *info) 626 1.1 mrg { 627 1.1 mrg int k, l; 628 1.1 mrg struct isl_tab_undo *snap, *snap2; 629 1.1 mrg unsigned n_eq = info[i].bmap->n_eq; 630 1.1 mrg 631 1.1 mrg snap = isl_tab_snap(info[i].tab); 632 1.1 mrg if (isl_tab_mark_rational(info[i].tab) < 0) 633 1.1 mrg return isl_change_error; 634 1.1 mrg snap2 = isl_tab_snap(info[i].tab); 635 1.1 mrg 636 1.1 mrg for (k = 0; k < info[i].bmap->n_ineq; ++k) { 637 1.1 mrg if (info[i].ineq[k] != STATUS_CUT) 638 1.1 mrg continue; 639 1.1 mrg if (isl_tab_select_facet(info[i].tab, n_eq + k) < 0) 640 1.1 mrg return isl_change_error; 641 1.1 mrg for (l = 0; l < info[j].bmap->n_ineq; ++l) { 642 1.1 mrg int stat; 643 1.1 mrg if (info[j].ineq[l] != STATUS_CUT) 644 1.1 mrg continue; 645 1.1 mrg stat = status_in(info[j].bmap->ineq[l], info[i].tab); 646 1.1 mrg if (stat < 0) 647 1.1 mrg return isl_change_error; 648 1.1 mrg if (stat != STATUS_VALID) 649 1.1 mrg break; 650 1.1 mrg } 651 1.1 mrg if (isl_tab_rollback(info[i].tab, snap2) < 0) 652 1.1 mrg return isl_change_error; 653 1.1 mrg if (l < info[j].bmap->n_ineq) 654 1.1 mrg break; 655 1.1 mrg } 656 1.1 mrg 657 1.1 mrg if (k < info[i].bmap->n_ineq) { 658 1.1 mrg if (isl_tab_rollback(info[i].tab, snap) < 0) 659 1.1 mrg return isl_change_error; 660 1.1 mrg return isl_change_none; 661 1.1 mrg } 662 1.1 mrg return fuse(i, j, info, NULL, 0, 0); 663 1.1 mrg } 664 1.1 mrg 665 1.1 mrg /* Check if info->bmap contains the basic map represented 666 1.1 mrg * by the tableau "tab". 667 1.1 mrg * For each equality, we check both the constraint itself 668 1.1 mrg * (as an inequality) and its negation. Make sure the 669 1.1 mrg * equality is returned to its original state before returning. 670 1.1 mrg */ 671 1.1 mrg static isl_bool contains(struct isl_coalesce_info *info, struct isl_tab *tab) 672 1.1 mrg { 673 1.1 mrg int k; 674 1.1 mrg isl_size dim; 675 1.1 mrg isl_basic_map *bmap = info->bmap; 676 1.1 mrg 677 1.1 mrg dim = isl_basic_map_dim(bmap, isl_dim_all); 678 1.1 mrg if (dim < 0) 679 1.1 mrg return isl_bool_error; 680 1.1 mrg for (k = 0; k < bmap->n_eq; ++k) { 681 1.1 mrg int stat; 682 1.1 mrg isl_seq_neg(bmap->eq[k], bmap->eq[k], 1 + dim); 683 1.1 mrg stat = status_in(bmap->eq[k], tab); 684 1.1 mrg isl_seq_neg(bmap->eq[k], bmap->eq[k], 1 + dim); 685 1.1 mrg if (stat < 0) 686 1.1 mrg return isl_bool_error; 687 1.1 mrg if (stat != STATUS_VALID) 688 1.1 mrg return isl_bool_false; 689 1.1 mrg stat = status_in(bmap->eq[k], tab); 690 1.1 mrg if (stat < 0) 691 1.1 mrg return isl_bool_error; 692 1.1 mrg if (stat != STATUS_VALID) 693 1.1 mrg return isl_bool_false; 694 1.1 mrg } 695 1.1 mrg 696 1.1 mrg for (k = 0; k < bmap->n_ineq; ++k) { 697 1.1 mrg int stat; 698 1.1 mrg if (info->ineq[k] == STATUS_REDUNDANT) 699 1.1 mrg continue; 700 1.1 mrg stat = status_in(bmap->ineq[k], tab); 701 1.1 mrg if (stat < 0) 702 1.1 mrg return isl_bool_error; 703 1.1 mrg if (stat != STATUS_VALID) 704 1.1 mrg return isl_bool_false; 705 1.1 mrg } 706 1.1 mrg return isl_bool_true; 707 1.1 mrg } 708 1.1 mrg 709 1.1 mrg /* Basic map "i" has an inequality "k" that is adjacent 710 1.1 mrg * to some inequality of basic map "j". All the other inequalities 711 1.1 mrg * are valid for "j". 712 1.1 mrg * If not NULL, then "extra" contains extra wrapping constraints that are valid 713 1.1 mrg * for both "i" and "j". 714 1.1 mrg * Check if basic map "j" forms an extension of basic map "i", 715 1.1 mrg * taking into account the extra constraints, if any. 716 1.1 mrg * 717 1.1 mrg * Note that this function is only called if some of the equalities or 718 1.1 mrg * inequalities of basic map "j" do cut basic map "i". The function is 719 1.1 mrg * correct even if there are no such cut constraints, but in that case 720 1.1 mrg * the additional checks performed by this function are overkill. 721 1.1 mrg * 722 1.1 mrg * In particular, we replace constraint k, say f >= 0, by constraint 723 1.1 mrg * f <= -1, add the inequalities of "j" that are valid for "i", 724 1.1 mrg * as well as the "extra" constraints, if any, 725 1.1 mrg * and check if the result is a subset of basic map "j". 726 1.1 mrg * To improve the chances of the subset relation being detected, 727 1.1 mrg * any variable that only attains a single integer value 728 1.1 mrg * in the tableau of "i" is first fixed to that value. 729 1.1 mrg * If the result is a subset, then we know that this result is exactly equal 730 1.1 mrg * to basic map "j" since all its constraints are valid for basic map "j". 731 1.1 mrg * By combining the valid constraints of "i" (all equalities and all 732 1.1 mrg * inequalities except "k"), the valid constraints of "j" and 733 1.1 mrg * the "extra" constraints, if any, we therefore 734 1.1 mrg * obtain a basic map that is equal to their union. 735 1.1 mrg * In this case, there is no need to perform a rollback of the tableau 736 1.1 mrg * since it is going to be destroyed in fuse(). 737 1.1 mrg * 738 1.1 mrg * 739 1.1 mrg * |\__ |\__ 740 1.1 mrg * | \__ | \__ 741 1.1 mrg * | \_ => | \__ 742 1.1 mrg * |_______| _ |_________\ 743 1.1 mrg * 744 1.1 mrg * 745 1.1 mrg * |\ |\ 746 1.1 mrg * | \ | \ 747 1.1 mrg * | \ | \ 748 1.1 mrg * | | | \ 749 1.1 mrg * | ||\ => | \ 750 1.1 mrg * | || \ | \ 751 1.1 mrg * | || | | | 752 1.1 mrg * |__||_/ |_____/ 753 1.1 mrg * 754 1.1 mrg * 755 1.1 mrg * _______ _______ 756 1.1 mrg * | | __ | \__ 757 1.1 mrg * | ||__| => | __| 758 1.1 mrg * |_______| |_______/ 759 1.1 mrg */ 760 1.1 mrg static enum isl_change is_adj_ineq_extension_with_wraps(int i, int j, int k, 761 1.1 mrg struct isl_coalesce_info *info, __isl_keep isl_mat *extra) 762 1.1 mrg { 763 1.1 mrg struct isl_tab_undo *snap; 764 1.1 mrg isl_size n_eq_i, n_ineq_j, n_extra; 765 1.1 mrg isl_size total = isl_basic_map_dim(info[i].bmap, isl_dim_all); 766 1.1 mrg isl_stat r; 767 1.1 mrg isl_bool super; 768 1.1 mrg 769 1.1 mrg if (total < 0) 770 1.1 mrg return isl_change_error; 771 1.1 mrg 772 1.1 mrg n_eq_i = isl_basic_map_n_equality(info[i].bmap); 773 1.1 mrg n_ineq_j = isl_basic_map_n_inequality(info[j].bmap); 774 1.1 mrg n_extra = isl_mat_rows(extra); 775 1.1 mrg if (n_eq_i < 0 || n_ineq_j < 0 || n_extra < 0) 776 1.1 mrg return isl_change_error; 777 1.1 mrg 778 1.1 mrg if (isl_tab_extend_cons(info[i].tab, 1 + n_ineq_j + n_extra) < 0) 779 1.1 mrg return isl_change_error; 780 1.1 mrg 781 1.1 mrg snap = isl_tab_snap(info[i].tab); 782 1.1 mrg 783 1.1 mrg if (isl_tab_unrestrict(info[i].tab, n_eq_i + k) < 0) 784 1.1 mrg return isl_change_error; 785 1.1 mrg 786 1.1 mrg isl_seq_neg(info[i].bmap->ineq[k], info[i].bmap->ineq[k], 1 + total); 787 1.1 mrg isl_int_sub_ui(info[i].bmap->ineq[k][0], info[i].bmap->ineq[k][0], 1); 788 1.1 mrg r = isl_tab_add_ineq(info[i].tab, info[i].bmap->ineq[k]); 789 1.1 mrg isl_seq_neg(info[i].bmap->ineq[k], info[i].bmap->ineq[k], 1 + total); 790 1.1 mrg isl_int_sub_ui(info[i].bmap->ineq[k][0], info[i].bmap->ineq[k][0], 1); 791 1.1 mrg if (r < 0) 792 1.1 mrg return isl_change_error; 793 1.1 mrg 794 1.1 mrg for (k = 0; k < n_ineq_j; ++k) { 795 1.1 mrg if (info[j].ineq[k] != STATUS_VALID) 796 1.1 mrg continue; 797 1.1 mrg if (isl_tab_add_ineq(info[i].tab, info[j].bmap->ineq[k]) < 0) 798 1.1 mrg return isl_change_error; 799 1.1 mrg } 800 1.1 mrg for (k = 0; k < n_extra; ++k) { 801 1.1 mrg if (isl_tab_add_ineq(info[i].tab, extra->row[k]) < 0) 802 1.1 mrg return isl_change_error; 803 1.1 mrg } 804 1.1 mrg if (isl_tab_detect_constants(info[i].tab) < 0) 805 1.1 mrg return isl_change_error; 806 1.1 mrg 807 1.1 mrg super = contains(&info[j], info[i].tab); 808 1.1 mrg if (super < 0) 809 1.1 mrg return isl_change_error; 810 1.1 mrg if (super) 811 1.1 mrg return fuse(i, j, info, extra, 0, 0); 812 1.1 mrg 813 1.1 mrg if (isl_tab_rollback(info[i].tab, snap) < 0) 814 1.1 mrg return isl_change_error; 815 1.1 mrg 816 1.1 mrg return isl_change_none; 817 1.1 mrg } 818 1.1 mrg 819 1.1 mrg /* Given an affine transformation matrix "T", does row "row" represent 820 1.1 mrg * anything other than a unit vector (possibly shifted by a constant) 821 1.1 mrg * that is not involved in any of the other rows? 822 1.1 mrg * 823 1.1 mrg * That is, if a constraint involves the variable corresponding to 824 1.1 mrg * the row, then could its preimage by "T" have any coefficients 825 1.1 mrg * that are different from those in the original constraint? 826 1.1 mrg */ 827 1.1 mrg static int not_unique_unit_row(__isl_keep isl_mat *T, int row) 828 1.1 mrg { 829 1.1 mrg int i, j; 830 1.1 mrg int len = T->n_col - 1; 831 1.1 mrg 832 1.1 mrg i = isl_seq_first_non_zero(T->row[row] + 1, len); 833 1.1 mrg if (i < 0) 834 1.1 mrg return 1; 835 1.1 mrg if (!isl_int_is_one(T->row[row][1 + i]) && 836 1.1 mrg !isl_int_is_negone(T->row[row][1 + i])) 837 1.1 mrg return 1; 838 1.1 mrg 839 1.1 mrg j = isl_seq_first_non_zero(T->row[row] + 1 + i + 1, len - (i + 1)); 840 1.1 mrg if (j >= 0) 841 1.1 mrg return 1; 842 1.1 mrg 843 1.1 mrg for (j = 1; j < T->n_row; ++j) { 844 1.1 mrg if (j == row) 845 1.1 mrg continue; 846 1.1 mrg if (!isl_int_is_zero(T->row[j][1 + i])) 847 1.1 mrg return 1; 848 1.1 mrg } 849 1.1 mrg 850 1.1 mrg return 0; 851 1.1 mrg } 852 1.1 mrg 853 1.1 mrg /* Does inequality constraint "ineq" of "bmap" involve any of 854 1.1 mrg * the variables marked in "affected"? 855 1.1 mrg * "total" is the total number of variables, i.e., the number 856 1.1 mrg * of entries in "affected". 857 1.1 mrg */ 858 1.1 mrg static isl_bool is_affected(__isl_keep isl_basic_map *bmap, int ineq, 859 1.1 mrg int *affected, int total) 860 1.1 mrg { 861 1.1 mrg int i; 862 1.1 mrg 863 1.1 mrg for (i = 0; i < total; ++i) { 864 1.1 mrg if (!affected[i]) 865 1.1 mrg continue; 866 1.1 mrg if (!isl_int_is_zero(bmap->ineq[ineq][1 + i])) 867 1.1 mrg return isl_bool_true; 868 1.1 mrg } 869 1.1 mrg 870 1.1 mrg return isl_bool_false; 871 1.1 mrg } 872 1.1 mrg 873 1.1 mrg /* Given the compressed version of inequality constraint "ineq" 874 1.1 mrg * of info->bmap in "v", check if the constraint can be tightened, 875 1.1 mrg * where the compression is based on an equality constraint valid 876 1.1 mrg * for info->tab. 877 1.1 mrg * If so, add the tightened version of the inequality constraint 878 1.1 mrg * to info->tab. "v" may be modified by this function. 879 1.1 mrg * 880 1.1 mrg * That is, if the compressed constraint is of the form 881 1.1 mrg * 882 1.1 mrg * m f() + c >= 0 883 1.1 mrg * 884 1.1 mrg * with 0 < c < m, then it is equivalent to 885 1.1 mrg * 886 1.1 mrg * f() >= 0 887 1.1 mrg * 888 1.1 mrg * This means that c can also be subtracted from the original, 889 1.1 mrg * uncompressed constraint without affecting the integer points 890 1.1 mrg * in info->tab. Add this tightened constraint as an extra row 891 1.1 mrg * to info->tab to make this information explicitly available. 892 1.1 mrg */ 893 1.1 mrg static __isl_give isl_vec *try_tightening(struct isl_coalesce_info *info, 894 1.1 mrg int ineq, __isl_take isl_vec *v) 895 1.1 mrg { 896 1.1 mrg isl_ctx *ctx; 897 1.1 mrg isl_stat r; 898 1.1 mrg 899 1.1 mrg if (!v) 900 1.1 mrg return NULL; 901 1.1 mrg 902 1.1 mrg ctx = isl_vec_get_ctx(v); 903 1.1 mrg isl_seq_gcd(v->el + 1, v->size - 1, &ctx->normalize_gcd); 904 1.1 mrg if (isl_int_is_zero(ctx->normalize_gcd) || 905 1.1 mrg isl_int_is_one(ctx->normalize_gcd)) { 906 1.1 mrg return v; 907 1.1 mrg } 908 1.1 mrg 909 1.1 mrg v = isl_vec_cow(v); 910 1.1 mrg if (!v) 911 1.1 mrg return NULL; 912 1.1 mrg 913 1.1 mrg isl_int_fdiv_r(v->el[0], v->el[0], ctx->normalize_gcd); 914 1.1 mrg if (isl_int_is_zero(v->el[0])) 915 1.1 mrg return v; 916 1.1 mrg 917 1.1 mrg if (isl_tab_extend_cons(info->tab, 1) < 0) 918 1.1 mrg return isl_vec_free(v); 919 1.1 mrg 920 1.1 mrg isl_int_sub(info->bmap->ineq[ineq][0], 921 1.1 mrg info->bmap->ineq[ineq][0], v->el[0]); 922 1.1 mrg r = isl_tab_add_ineq(info->tab, info->bmap->ineq[ineq]); 923 1.1 mrg isl_int_add(info->bmap->ineq[ineq][0], 924 1.1 mrg info->bmap->ineq[ineq][0], v->el[0]); 925 1.1 mrg 926 1.1 mrg if (r < 0) 927 1.1 mrg return isl_vec_free(v); 928 1.1 mrg 929 1.1 mrg return v; 930 1.1 mrg } 931 1.1 mrg 932 1.1 mrg /* Tighten the (non-redundant) constraints on the facet represented 933 1.1 mrg * by info->tab. 934 1.1 mrg * In particular, on input, info->tab represents the result 935 1.1 mrg * of relaxing the "n" inequality constraints of info->bmap in "relaxed" 936 1.1 mrg * by one, i.e., replacing f_i >= 0 by f_i + 1 >= 0, and then 937 1.1 mrg * replacing the one at index "l" by the corresponding equality, 938 1.1 mrg * i.e., f_k + 1 = 0, with k = relaxed[l]. 939 1.1 mrg * 940 1.1 mrg * Compute a variable compression from the equality constraint f_k + 1 = 0 941 1.1 mrg * and use it to tighten the other constraints of info->bmap 942 1.1 mrg * (that is, all constraints that have not been relaxed), 943 1.1 mrg * updating info->tab (and leaving info->bmap untouched). 944 1.1 mrg * The compression handles essentially two cases, one where a variable 945 1.1 mrg * is assigned a fixed value and can therefore be eliminated, and one 946 1.1 mrg * where one variable is a shifted multiple of some other variable and 947 1.1 mrg * can therefore be replaced by that multiple. 948 1.1 mrg * Gaussian elimination would also work for the first case, but for 949 1.1 mrg * the second case, the effectiveness would depend on the order 950 1.1 mrg * of the variables. 951 1.1 mrg * After compression, some of the constraints may have coefficients 952 1.1 mrg * with a common divisor. If this divisor does not divide the constant 953 1.1 mrg * term, then the constraint can be tightened. 954 1.1 mrg * The tightening is performed on the tableau info->tab by introducing 955 1.1 mrg * extra (temporary) constraints. 956 1.1 mrg * 957 1.1 mrg * Only constraints that are possibly affected by the compression are 958 1.1 mrg * considered. In particular, if the constraint only involves variables 959 1.1 mrg * that are directly mapped to a distinct set of other variables, then 960 1.1 mrg * no common divisor can be introduced and no tightening can occur. 961 1.1 mrg * 962 1.1 mrg * It is important to only consider the non-redundant constraints 963 1.1 mrg * since the facet constraint has been relaxed prior to the call 964 1.1 mrg * to this function, meaning that the constraints that were redundant 965 1.1 mrg * prior to the relaxation may no longer be redundant. 966 1.1 mrg * These constraints will be ignored in the fused result, so 967 1.1 mrg * the fusion detection should not exploit them. 968 1.1 mrg */ 969 1.1 mrg static isl_stat tighten_on_relaxed_facet(struct isl_coalesce_info *info, 970 1.1 mrg int n, int *relaxed, int l) 971 1.1 mrg { 972 1.1 mrg isl_size total; 973 1.1 mrg isl_ctx *ctx; 974 1.1 mrg isl_vec *v = NULL; 975 1.1 mrg isl_mat *T; 976 1.1 mrg int i; 977 1.1 mrg int k; 978 1.1 mrg int *affected; 979 1.1 mrg 980 1.1 mrg k = relaxed[l]; 981 1.1 mrg ctx = isl_basic_map_get_ctx(info->bmap); 982 1.1 mrg total = isl_basic_map_dim(info->bmap, isl_dim_all); 983 1.1 mrg if (total < 0) 984 1.1 mrg return isl_stat_error; 985 1.1 mrg isl_int_add_ui(info->bmap->ineq[k][0], info->bmap->ineq[k][0], 1); 986 1.1 mrg T = isl_mat_sub_alloc6(ctx, info->bmap->ineq, k, 1, 0, 1 + total); 987 1.1 mrg T = isl_mat_variable_compression(T, NULL); 988 1.1 mrg isl_int_sub_ui(info->bmap->ineq[k][0], info->bmap->ineq[k][0], 1); 989 1.1 mrg if (!T) 990 1.1 mrg return isl_stat_error; 991 1.1 mrg if (T->n_col == 0) { 992 1.1 mrg isl_mat_free(T); 993 1.1 mrg return isl_stat_ok; 994 1.1 mrg } 995 1.1 mrg 996 1.1 mrg affected = isl_alloc_array(ctx, int, total); 997 1.1 mrg if (!affected) 998 1.1 mrg goto error; 999 1.1 mrg 1000 1.1 mrg for (i = 0; i < total; ++i) 1001 1.1 mrg affected[i] = not_unique_unit_row(T, 1 + i); 1002 1.1 mrg 1003 1.1 mrg for (i = 0; i < info->bmap->n_ineq; ++i) { 1004 1.1 mrg isl_bool handle; 1005 1.1 mrg if (any(relaxed, n, i)) 1006 1.1 mrg continue; 1007 1.1 mrg if (info->ineq[i] == STATUS_REDUNDANT) 1008 1.1 mrg continue; 1009 1.1 mrg handle = is_affected(info->bmap, i, affected, total); 1010 1.1 mrg if (handle < 0) 1011 1.1 mrg goto error; 1012 1.1 mrg if (!handle) 1013 1.1 mrg continue; 1014 1.1 mrg v = isl_vec_alloc(ctx, 1 + total); 1015 1.1 mrg if (!v) 1016 1.1 mrg goto error; 1017 1.1 mrg isl_seq_cpy(v->el, info->bmap->ineq[i], 1 + total); 1018 1.1 mrg v = isl_vec_mat_product(v, isl_mat_copy(T)); 1019 1.1 mrg v = try_tightening(info, i, v); 1020 1.1 mrg isl_vec_free(v); 1021 1.1 mrg if (!v) 1022 1.1 mrg goto error; 1023 1.1 mrg } 1024 1.1 mrg 1025 1.1 mrg isl_mat_free(T); 1026 1.1 mrg free(affected); 1027 1.1 mrg return isl_stat_ok; 1028 1.1 mrg error: 1029 1.1 mrg isl_mat_free(T); 1030 1.1 mrg free(affected); 1031 1.1 mrg return isl_stat_error; 1032 1.1 mrg } 1033 1.1 mrg 1034 1.1 mrg /* Replace the basic maps "i" and "j" by an extension of "i" 1035 1.1 mrg * along the "n" inequality constraints in "relax" by one. 1036 1.1 mrg * The tableau info[i].tab has already been extended. 1037 1.1 mrg * Extend info[i].bmap accordingly by relaxing all constraints in "relax" 1038 1.1 mrg * by one. 1039 1.1 mrg * Each integer division that does not have exactly the same 1040 1.1 mrg * definition in "i" and "j" is marked unknown and the basic map 1041 1.1 mrg * is scheduled to be simplified in an attempt to recover 1042 1.1 mrg * the integer division definition. 1043 1.1 mrg * Place the extension in the position that is the smallest of i and j. 1044 1.1 mrg */ 1045 1.1 mrg static enum isl_change extend(int i, int j, int n, int *relax, 1046 1.1 mrg struct isl_coalesce_info *info) 1047 1.1 mrg { 1048 1.1 mrg int l; 1049 1.1 mrg isl_size total; 1050 1.1 mrg 1051 1.1 mrg info[i].bmap = isl_basic_map_cow(info[i].bmap); 1052 1.1 mrg total = isl_basic_map_dim(info[i].bmap, isl_dim_all); 1053 1.1 mrg if (total < 0) 1054 1.1 mrg return isl_change_error; 1055 1.1 mrg for (l = 0; l < info[i].bmap->n_div; ++l) 1056 1.1 mrg if (!isl_seq_eq(info[i].bmap->div[l], 1057 1.1 mrg info[j].bmap->div[l], 1 + 1 + total)) { 1058 1.1 mrg isl_int_set_si(info[i].bmap->div[l][0], 0); 1059 1.1 mrg info[i].simplify = 1; 1060 1.1 mrg } 1061 1.1 mrg for (l = 0; l < n; ++l) 1062 1.1 mrg isl_int_add_ui(info[i].bmap->ineq[relax[l]][0], 1063 1.1 mrg info[i].bmap->ineq[relax[l]][0], 1); 1064 1.1 mrg ISL_F_CLR(info[i].bmap, ISL_BASIC_MAP_NO_REDUNDANT); 1065 1.1 mrg ISL_F_SET(info[i].bmap, ISL_BASIC_MAP_FINAL); 1066 1.1 mrg drop(&info[j]); 1067 1.1 mrg info[i].modified = 1; 1068 1.1 mrg if (j < i) 1069 1.1 mrg exchange(&info[i], &info[j]); 1070 1.1 mrg return isl_change_fuse; 1071 1.1 mrg } 1072 1.1 mrg 1073 1.1 mrg /* Basic map "i" has "n" inequality constraints (collected in "relax") 1074 1.1 mrg * that are such that they include basic map "j" if they are relaxed 1075 1.1 mrg * by one. All the other inequalities are valid for "j". 1076 1.1 mrg * Check if basic map "j" forms an extension of basic map "i". 1077 1.1 mrg * 1078 1.1 mrg * In particular, relax the constraints in "relax", compute the corresponding 1079 1.1 mrg * facets one by one and check whether each of these is included 1080 1.1 mrg * in the other basic map. 1081 1.1 mrg * Before testing for inclusion, the constraints on each facet 1082 1.1 mrg * are tightened to increase the chance of an inclusion being detected. 1083 1.1 mrg * (Adding the valid constraints of "j" to the tableau of "i", as is done 1084 1.1 mrg * in is_adj_ineq_extension, may further increase those chances, but this 1085 1.1 mrg * is not currently done.) 1086 1.1 mrg * If each facet is included, we know that relaxing the constraints extends 1087 1.1 mrg * the basic map with exactly the other basic map (we already know that this 1088 1.1 mrg * other basic map is included in the extension, because all other 1089 1.1 mrg * inequality constraints are valid of "j") and we can replace the 1090 1.1 mrg * two basic maps by this extension. 1091 1.1 mrg * 1092 1.1 mrg * If any of the relaxed constraints turn out to be redundant, then bail out. 1093 1.1 mrg * isl_tab_select_facet refuses to handle such constraints. It may be 1094 1.1 mrg * possible to handle them anyway by making a distinction between 1095 1.1 mrg * redundant constraints with a corresponding facet that still intersects 1096 1.1 mrg * the set (allowing isl_tab_select_facet to handle them) and 1097 1.1 mrg * those where the facet does not intersect the set (which can be ignored 1098 1.1 mrg * because the empty facet is trivially included in the other disjunct). 1099 1.1 mrg * However, relaxed constraints that turn out to be redundant should 1100 1.1 mrg * be fairly rare and no such instance has been reported where 1101 1.1 mrg * coalescing would be successful. 1102 1.1 mrg * ____ _____ 1103 1.1 mrg * / || / | 1104 1.1 mrg * / || / | 1105 1.1 mrg * \ || => \ | 1106 1.1 mrg * \ || \ | 1107 1.1 mrg * \___|| \____| 1108 1.1 mrg * 1109 1.1 mrg * 1110 1.1 mrg * \ |\ 1111 1.1 mrg * |\\ | \ 1112 1.1 mrg * | \\ | \ 1113 1.1 mrg * | | => | / 1114 1.1 mrg * | / | / 1115 1.1 mrg * |/ |/ 1116 1.1 mrg */ 1117 1.1 mrg static enum isl_change is_relaxed_extension(int i, int j, int n, int *relax, 1118 1.1 mrg struct isl_coalesce_info *info) 1119 1.1 mrg { 1120 1.1 mrg int l; 1121 1.1 mrg isl_bool super; 1122 1.1 mrg struct isl_tab_undo *snap, *snap2; 1123 1.1 mrg unsigned n_eq = info[i].bmap->n_eq; 1124 1.1 mrg 1125 1.1 mrg for (l = 0; l < n; ++l) 1126 1.1 mrg if (isl_tab_is_equality(info[i].tab, n_eq + relax[l])) 1127 1.1 mrg return isl_change_none; 1128 1.1 mrg 1129 1.1 mrg snap = isl_tab_snap(info[i].tab); 1130 1.1 mrg for (l = 0; l < n; ++l) 1131 1.1 mrg if (isl_tab_relax(info[i].tab, n_eq + relax[l]) < 0) 1132 1.1 mrg return isl_change_error; 1133 1.1 mrg for (l = 0; l < n; ++l) { 1134 1.1 mrg if (!isl_tab_is_redundant(info[i].tab, n_eq + relax[l])) 1135 1.1 mrg continue; 1136 1.1 mrg if (isl_tab_rollback(info[i].tab, snap) < 0) 1137 1.1 mrg return isl_change_error; 1138 1.1 mrg return isl_change_none; 1139 1.1 mrg } 1140 1.1 mrg snap2 = isl_tab_snap(info[i].tab); 1141 1.1 mrg for (l = 0; l < n; ++l) { 1142 1.1 mrg if (isl_tab_rollback(info[i].tab, snap2) < 0) 1143 1.1 mrg return isl_change_error; 1144 1.1 mrg if (isl_tab_select_facet(info[i].tab, n_eq + relax[l]) < 0) 1145 1.1 mrg return isl_change_error; 1146 1.1 mrg if (tighten_on_relaxed_facet(&info[i], n, relax, l) < 0) 1147 1.1 mrg return isl_change_error; 1148 1.1 mrg super = contains(&info[j], info[i].tab); 1149 1.1 mrg if (super < 0) 1150 1.1 mrg return isl_change_error; 1151 1.1 mrg if (super) 1152 1.1 mrg continue; 1153 1.1 mrg if (isl_tab_rollback(info[i].tab, snap) < 0) 1154 1.1 mrg return isl_change_error; 1155 1.1 mrg return isl_change_none; 1156 1.1 mrg } 1157 1.1 mrg 1158 1.1 mrg if (isl_tab_rollback(info[i].tab, snap2) < 0) 1159 1.1 mrg return isl_change_error; 1160 1.1 mrg return extend(i, j, n, relax, info); 1161 1.1 mrg } 1162 1.1 mrg 1163 1.1 mrg /* Data structure that keeps track of the wrapping constraints 1164 1.1 mrg * and of information to bound the coefficients of those constraints. 1165 1.1 mrg * 1166 1.1 mrg * "failed" is set if wrapping has failed. 1167 1.1 mrg * bound is set if we want to apply a bound on the coefficients 1168 1.1 mrg * mat contains the wrapping constraints 1169 1.1 mrg * max is the bound on the coefficients (if bound is set) 1170 1.1 mrg */ 1171 1.1 mrg struct isl_wraps { 1172 1.1 mrg int failed; 1173 1.1 mrg int bound; 1174 1.1 mrg isl_mat *mat; 1175 1.1 mrg isl_int max; 1176 1.1 mrg }; 1177 1.1 mrg 1178 1.1 mrg /* Update wraps->max to be greater than or equal to the coefficients 1179 1.1 mrg * in the equalities and inequalities of info->bmap that can be removed 1180 1.1 mrg * if we end up applying wrapping. 1181 1.1 mrg */ 1182 1.1 mrg static isl_stat wraps_update_max(struct isl_wraps *wraps, 1183 1.1 mrg struct isl_coalesce_info *info) 1184 1.1 mrg { 1185 1.1 mrg int k; 1186 1.1 mrg isl_int max_k; 1187 1.1 mrg isl_size total = isl_basic_map_dim(info->bmap, isl_dim_all); 1188 1.1 mrg 1189 1.1 mrg if (total < 0) 1190 1.1 mrg return isl_stat_error; 1191 1.1 mrg isl_int_init(max_k); 1192 1.1 mrg 1193 1.1 mrg for (k = 0; k < info->bmap->n_eq; ++k) { 1194 1.1 mrg if (info->eq[2 * k] == STATUS_VALID && 1195 1.1 mrg info->eq[2 * k + 1] == STATUS_VALID) 1196 1.1 mrg continue; 1197 1.1 mrg isl_seq_abs_max(info->bmap->eq[k] + 1, total, &max_k); 1198 1.1 mrg if (isl_int_abs_gt(max_k, wraps->max)) 1199 1.1 mrg isl_int_set(wraps->max, max_k); 1200 1.1 mrg } 1201 1.1 mrg 1202 1.1 mrg for (k = 0; k < info->bmap->n_ineq; ++k) { 1203 1.1 mrg if (info->ineq[k] == STATUS_VALID || 1204 1.1 mrg info->ineq[k] == STATUS_REDUNDANT) 1205 1.1 mrg continue; 1206 1.1 mrg isl_seq_abs_max(info->bmap->ineq[k] + 1, total, &max_k); 1207 1.1 mrg if (isl_int_abs_gt(max_k, wraps->max)) 1208 1.1 mrg isl_int_set(wraps->max, max_k); 1209 1.1 mrg } 1210 1.1 mrg 1211 1.1 mrg isl_int_clear(max_k); 1212 1.1 mrg 1213 1.1 mrg return isl_stat_ok; 1214 1.1 mrg } 1215 1.1 mrg 1216 1.1 mrg /* Initialize the isl_wraps data structure. 1217 1.1 mrg * If we want to bound the coefficients of the wrapping constraints, 1218 1.1 mrg * we set wraps->max to the largest coefficient 1219 1.1 mrg * in the equalities and inequalities that can be removed if we end up 1220 1.1 mrg * applying wrapping. 1221 1.1 mrg */ 1222 1.1 mrg static isl_stat wraps_init(struct isl_wraps *wraps, __isl_take isl_mat *mat, 1223 1.1 mrg struct isl_coalesce_info *info, int i, int j) 1224 1.1 mrg { 1225 1.1 mrg isl_ctx *ctx; 1226 1.1 mrg 1227 1.1 mrg wraps->failed = 0; 1228 1.1 mrg wraps->bound = 0; 1229 1.1 mrg wraps->mat = mat; 1230 1.1 mrg if (!mat) 1231 1.1 mrg return isl_stat_error; 1232 1.1 mrg wraps->mat->n_row = 0; 1233 1.1 mrg ctx = isl_mat_get_ctx(mat); 1234 1.1 mrg wraps->bound = isl_options_get_coalesce_bounded_wrapping(ctx); 1235 1.1 mrg if (!wraps->bound) 1236 1.1 mrg return isl_stat_ok; 1237 1.1 mrg isl_int_init(wraps->max); 1238 1.1 mrg isl_int_set_si(wraps->max, 0); 1239 1.1 mrg if (wraps_update_max(wraps, &info[i]) < 0) 1240 1.1 mrg return isl_stat_error; 1241 1.1 mrg if (wraps_update_max(wraps, &info[j]) < 0) 1242 1.1 mrg return isl_stat_error; 1243 1.1 mrg 1244 1.1 mrg return isl_stat_ok; 1245 1.1 mrg } 1246 1.1 mrg 1247 1.1 mrg /* Free the contents of the isl_wraps data structure. 1248 1.1 mrg */ 1249 1.1 mrg static void wraps_free(struct isl_wraps *wraps) 1250 1.1 mrg { 1251 1.1 mrg isl_mat_free(wraps->mat); 1252 1.1 mrg if (wraps->bound) 1253 1.1 mrg isl_int_clear(wraps->max); 1254 1.1 mrg } 1255 1.1 mrg 1256 1.1 mrg /* Mark the wrapping as failed. 1257 1.1 mrg */ 1258 1.1 mrg static isl_stat wraps_mark_failed(struct isl_wraps *wraps) 1259 1.1 mrg { 1260 1.1 mrg wraps->failed = 1; 1261 1.1 mrg return isl_stat_ok; 1262 1.1 mrg } 1263 1.1 mrg 1264 1.1 mrg /* Is the wrapping constraint in row "row" allowed? 1265 1.1 mrg * 1266 1.1 mrg * If wraps->bound is set, we check that none of the coefficients 1267 1.1 mrg * is greater than wraps->max. 1268 1.1 mrg */ 1269 1.1 mrg static int allow_wrap(struct isl_wraps *wraps, int row) 1270 1.1 mrg { 1271 1.1 mrg int i; 1272 1.1 mrg 1273 1.1 mrg if (!wraps->bound) 1274 1.1 mrg return 1; 1275 1.1 mrg 1276 1.1 mrg for (i = 1; i < wraps->mat->n_col; ++i) 1277 1.1 mrg if (isl_int_abs_gt(wraps->mat->row[row][i], wraps->max)) 1278 1.1 mrg return 0; 1279 1.1 mrg 1280 1.1 mrg return 1; 1281 1.1 mrg } 1282 1.1 mrg 1283 1.1 mrg /* Wrap "ineq" (or its opposite if "negate" is set) around "bound" 1284 1.1 mrg * to include "set" and add the result in position "w" of "wraps". 1285 1.1 mrg * "len" is the total number of coefficients in "bound" and "ineq". 1286 1.1 mrg * Return 1 on success, 0 on failure and -1 on error. 1287 1.1 mrg * Wrapping can fail if the result of wrapping is equal to "bound" 1288 1.1 mrg * or if we want to bound the sizes of the coefficients and 1289 1.1 mrg * the wrapped constraint does not satisfy this bound. 1290 1.1 mrg */ 1291 1.1 mrg static int add_wrap(struct isl_wraps *wraps, int w, isl_int *bound, 1292 1.1 mrg isl_int *ineq, unsigned len, __isl_keep isl_set *set, int negate) 1293 1.1 mrg { 1294 1.1 mrg isl_seq_cpy(wraps->mat->row[w], bound, len); 1295 1.1 mrg if (negate) { 1296 1.1 mrg isl_seq_neg(wraps->mat->row[w + 1], ineq, len); 1297 1.1 mrg ineq = wraps->mat->row[w + 1]; 1298 1.1 mrg } 1299 1.1 mrg if (!isl_set_wrap_facet(set, wraps->mat->row[w], ineq)) 1300 1.1 mrg return -1; 1301 1.1 mrg if (isl_seq_eq(wraps->mat->row[w], bound, len)) 1302 1.1 mrg return 0; 1303 1.1 mrg if (!allow_wrap(wraps, w)) 1304 1.1 mrg return 0; 1305 1.1 mrg return 1; 1306 1.1 mrg } 1307 1.1 mrg 1308 1.1 mrg /* This function has two modes of operations. 1309 1.1 mrg * 1310 1.1 mrg * If "add_valid" is set, then all the constraints of info->bmap 1311 1.1 mrg * (except the opposite of "bound") are valid for the other basic map. 1312 1.1 mrg * In this case, attempts are made to wrap some of these valid constraints 1313 1.1 mrg * to more tightly fit around "set". Only successful wrappings are recorded 1314 1.1 mrg * and failed wrappings are ignored. 1315 1.1 mrg * 1316 1.1 mrg * If "add_valid" is not set, then some of the constraints of info->bmap 1317 1.1 mrg * are not valid for the other basic map, and only those are considered 1318 1.1 mrg * for wrapping. In this case all attempted wrappings need to succeed. 1319 1.1 mrg * Otherwise "wraps" is marked as failed. 1320 1.1 mrg * Note that the constraints that are valid for the other basic map 1321 1.1 mrg * will be added to the combined basic map by default, so there is 1322 1.1 mrg * no need to wrap them. 1323 1.1 mrg * The caller wrap_in_facets even relies on this function not wrapping 1324 1.1 mrg * any constraints that are already valid. 1325 1.1 mrg * 1326 1.1 mrg * Only consider constraints that are not redundant (as determined 1327 1.1 mrg * by info->tab) and that are valid or invalid depending on "add_valid". 1328 1.1 mrg * Wrap each constraint around "bound" such that it includes the whole 1329 1.1 mrg * set "set" and append the resulting constraint to "wraps". 1330 1.1 mrg * "wraps" is assumed to have been pre-allocated to the appropriate size. 1331 1.1 mrg * wraps->n_row is the number of actual wrapped constraints that have 1332 1.1 mrg * been added. 1333 1.1 mrg * If any of the wrapping problems results in a constraint that is 1334 1.1 mrg * identical to "bound", then this means that "set" is unbounded in such 1335 1.1 mrg * a way that no wrapping is possible. 1336 1.1 mrg * Similarly, if we want to bound the coefficients of the wrapping 1337 1.1 mrg * constraints and a newly added wrapping constraint does not 1338 1.1 mrg * satisfy the bound, then the wrapping is considered to have failed. 1339 1.1 mrg * Note though that "wraps" is only marked failed if "add_valid" is not set. 1340 1.1 mrg */ 1341 1.1 mrg static isl_stat add_selected_wraps(struct isl_wraps *wraps, 1342 1.1 mrg struct isl_coalesce_info *info, isl_int *bound, __isl_keep isl_set *set, 1343 1.1 mrg int add_valid) 1344 1.1 mrg { 1345 1.1 mrg int l, m; 1346 1.1 mrg int w; 1347 1.1 mrg int added; 1348 1.1 mrg isl_basic_map *bmap = info->bmap; 1349 1.1 mrg isl_size total = isl_basic_map_dim(bmap, isl_dim_all); 1350 1.1 mrg unsigned len = 1 + total; 1351 1.1 mrg 1352 1.1 mrg if (total < 0) 1353 1.1 mrg return isl_stat_error; 1354 1.1 mrg 1355 1.1 mrg w = wraps->mat->n_row; 1356 1.1 mrg 1357 1.1 mrg for (l = 0; l < bmap->n_ineq; ++l) { 1358 1.1 mrg int is_valid = info->ineq[l] == STATUS_VALID; 1359 1.1 mrg if ((!add_valid && is_valid) || 1360 1.1 mrg info->ineq[l] == STATUS_REDUNDANT) 1361 1.1 mrg continue; 1362 1.1 mrg if (isl_seq_is_neg(bound, bmap->ineq[l], len)) 1363 1.1 mrg continue; 1364 1.1 mrg if (isl_seq_eq(bound, bmap->ineq[l], len)) 1365 1.1 mrg continue; 1366 1.1 mrg if (isl_tab_is_redundant(info->tab, bmap->n_eq + l)) 1367 1.1 mrg continue; 1368 1.1 mrg 1369 1.1 mrg added = add_wrap(wraps, w, bound, bmap->ineq[l], len, set, 0); 1370 1.1 mrg if (added < 0) 1371 1.1 mrg return isl_stat_error; 1372 1.1 mrg if (!added && !is_valid) 1373 1.1 mrg goto unbounded; 1374 1.1 mrg if (added) 1375 1.1 mrg ++w; 1376 1.1 mrg } 1377 1.1 mrg for (l = 0; l < bmap->n_eq; ++l) { 1378 1.1 mrg if (isl_seq_is_neg(bound, bmap->eq[l], len)) 1379 1.1 mrg continue; 1380 1.1 mrg if (isl_seq_eq(bound, bmap->eq[l], len)) 1381 1.1 mrg continue; 1382 1.1 mrg 1383 1.1 mrg for (m = 0; m < 2; ++m) { 1384 1.1 mrg if (info->eq[2 * l + m] == STATUS_VALID) 1385 1.1 mrg continue; 1386 1.1 mrg added = add_wrap(wraps, w, bound, bmap->eq[l], len, 1387 1.1 mrg set, !m); 1388 1.1 mrg if (added < 0) 1389 1.1 mrg return isl_stat_error; 1390 1.1 mrg if (!added) 1391 1.1 mrg goto unbounded; 1392 1.1 mrg ++w; 1393 1.1 mrg } 1394 1.1 mrg } 1395 1.1 mrg 1396 1.1 mrg wraps->mat->n_row = w; 1397 1.1 mrg return isl_stat_ok; 1398 1.1 mrg unbounded: 1399 1.1 mrg return wraps_mark_failed(wraps); 1400 1.1 mrg } 1401 1.1 mrg 1402 1.1 mrg /* For each constraint in info->bmap that is not redundant (as determined 1403 1.1 mrg * by info->tab) and that is not a valid constraint for the other basic map, 1404 1.1 mrg * wrap the constraint around "bound" such that it includes the whole 1405 1.1 mrg * set "set" and append the resulting constraint to "wraps". 1406 1.1 mrg * Note that the constraints that are valid for the other basic map 1407 1.1 mrg * will be added to the combined basic map by default, so there is 1408 1.1 mrg * no need to wrap them. 1409 1.1 mrg * The caller wrap_in_facets even relies on this function not wrapping 1410 1.1 mrg * any constraints that are already valid. 1411 1.1 mrg * "wraps" is assumed to have been pre-allocated to the appropriate size. 1412 1.1 mrg * wraps->n_row is the number of actual wrapped constraints that have 1413 1.1 mrg * been added. 1414 1.1 mrg * If any of the wrapping problems results in a constraint that is 1415 1.1 mrg * identical to "bound", then this means that "set" is unbounded in such 1416 1.1 mrg * a way that no wrapping is possible. If this happens then "wraps" 1417 1.1 mrg * is marked as failed. 1418 1.1 mrg * Similarly, if we want to bound the coefficients of the wrapping 1419 1.1 mrg * constraints and a newly added wrapping constraint does not 1420 1.1 mrg * satisfy the bound, then "wraps" is also marked as failed. 1421 1.1 mrg */ 1422 1.1 mrg static isl_stat add_wraps(struct isl_wraps *wraps, 1423 1.1 mrg struct isl_coalesce_info *info, isl_int *bound, __isl_keep isl_set *set) 1424 1.1 mrg { 1425 1.1 mrg return add_selected_wraps(wraps, info, bound, set, 0); 1426 1.1 mrg } 1427 1.1 mrg 1428 1.1 mrg /* Check if the constraints in "wraps" from "first" until the last 1429 1.1 mrg * are all valid for the basic set represented by "tab", 1430 1.1 mrg * dropping the invalid constraints if "keep" is set and 1431 1.1 mrg * marking the wrapping as failed if "keep" is not set and 1432 1.1 mrg * any constraint turns out to be invalid. 1433 1.1 mrg */ 1434 1.1 mrg static isl_stat check_wraps(struct isl_wraps *wraps, int first, 1435 1.1 mrg struct isl_tab *tab, int keep) 1436 1.1 mrg { 1437 1.1 mrg int i; 1438 1.1 mrg 1439 1.1 mrg for (i = wraps->mat->n_row - 1; i >= first; --i) { 1440 1.1 mrg enum isl_ineq_type type; 1441 1.1 mrg type = isl_tab_ineq_type(tab, wraps->mat->row[i]); 1442 1.1 mrg if (type == isl_ineq_error) 1443 1.1 mrg return isl_stat_error; 1444 1.1 mrg if (type == isl_ineq_redundant) 1445 1.1 mrg continue; 1446 1.1 mrg if (!keep) 1447 1.1 mrg return wraps_mark_failed(wraps); 1448 1.1 mrg wraps->mat = isl_mat_drop_rows(wraps->mat, i, 1); 1449 1.1 mrg if (!wraps->mat) 1450 1.1 mrg return isl_stat_error; 1451 1.1 mrg } 1452 1.1 mrg 1453 1.1 mrg return isl_stat_ok; 1454 1.1 mrg } 1455 1.1 mrg 1456 1.1 mrg /* Return a set that corresponds to the non-redundant constraints 1457 1.1 mrg * (as recorded in tab) of bmap. 1458 1.1 mrg * 1459 1.1 mrg * It's important to remove the redundant constraints as some 1460 1.1 mrg * of the other constraints may have been modified after the 1461 1.1 mrg * constraints were marked redundant. 1462 1.1 mrg * In particular, a constraint may have been relaxed. 1463 1.1 mrg * Redundant constraints are ignored when a constraint is relaxed 1464 1.1 mrg * and should therefore continue to be ignored ever after. 1465 1.1 mrg * Otherwise, the relaxation might be thwarted by some of 1466 1.1 mrg * these constraints. 1467 1.1 mrg * 1468 1.1 mrg * Update the underlying set to ensure that the dimension doesn't change. 1469 1.1 mrg * Otherwise the integer divisions could get dropped if the tab 1470 1.1 mrg * turns out to be empty. 1471 1.1 mrg */ 1472 1.1 mrg static __isl_give isl_set *set_from_updated_bmap(__isl_keep isl_basic_map *bmap, 1473 1.1 mrg struct isl_tab *tab) 1474 1.1 mrg { 1475 1.1 mrg isl_basic_set *bset; 1476 1.1 mrg 1477 1.1 mrg bmap = isl_basic_map_copy(bmap); 1478 1.1 mrg bset = isl_basic_map_underlying_set(bmap); 1479 1.1 mrg bset = isl_basic_set_cow(bset); 1480 1.1 mrg bset = isl_basic_set_update_from_tab(bset, tab); 1481 1.1 mrg return isl_set_from_basic_set(bset); 1482 1.1 mrg } 1483 1.1 mrg 1484 1.1 mrg /* Does "info" have any cut constraints that are redundant? 1485 1.1 mrg */ 1486 1.1 mrg static isl_bool has_redundant_cuts(struct isl_coalesce_info *info) 1487 1.1 mrg { 1488 1.1 mrg int l; 1489 1.1 mrg isl_size n_eq, n_ineq; 1490 1.1 mrg 1491 1.1 mrg n_eq = isl_basic_map_n_equality(info->bmap); 1492 1.1 mrg n_ineq = isl_basic_map_n_inequality(info->bmap); 1493 1.1 mrg if (n_eq < 0 || n_ineq < 0) 1494 1.1 mrg return isl_bool_error; 1495 1.1 mrg for (l = 0; l < n_ineq; ++l) { 1496 1.1 mrg int red; 1497 1.1 mrg 1498 1.1 mrg if (info->ineq[l] != STATUS_CUT) 1499 1.1 mrg continue; 1500 1.1 mrg red = isl_tab_is_redundant(info->tab, n_eq + l); 1501 1.1 mrg if (red < 0) 1502 1.1 mrg return isl_bool_error; 1503 1.1 mrg if (red) 1504 1.1 mrg return isl_bool_true; 1505 1.1 mrg } 1506 1.1 mrg 1507 1.1 mrg return isl_bool_false; 1508 1.1 mrg } 1509 1.1 mrg 1510 1.1 mrg /* Wrap some constraints of info->bmap that bound the facet defined 1511 1.1 mrg * by inequality "k" around (the opposite of) this inequality to 1512 1.1 mrg * include "set". "bound" may be used to store the negated inequality. 1513 1.1 mrg * 1514 1.1 mrg * If "add_valid" is set, then all ridges are already valid and 1515 1.1 mrg * the purpose is to wrap "set" more tightly. In this case, 1516 1.1 mrg * wrapping doesn't fail, although it is possible that no constraint 1517 1.1 mrg * gets wrapped. 1518 1.1 mrg * 1519 1.1 mrg * If "add_valid" is not set, then some of the ridges are cut constraints 1520 1.1 mrg * and only those are wrapped around "set". 1521 1.1 mrg * 1522 1.1 mrg * Since the wrapped constraints are not guaranteed to contain the whole 1523 1.1 mrg * of info->bmap, we check them in check_wraps. 1524 1.1 mrg * If any of the wrapped constraints turn out to be invalid, then 1525 1.1 mrg * check_wraps will mark "wraps" as failed if "add_valid" is not set. 1526 1.1 mrg * If "add_valid" is set, then the offending constraints are 1527 1.1 mrg * simply removed. 1528 1.1 mrg * 1529 1.1 mrg * If the facet turns out to be empty, then no wrapping can be performed. 1530 1.1 mrg * This is considered a failure, unless "add_valid" is set. 1531 1.1 mrg * 1532 1.1 mrg * If any of the cut constraints of info->bmap turn out 1533 1.1 mrg * to be redundant with respect to other constraints 1534 1.1 mrg * then these will neither be wrapped nor added directly to the result. 1535 1.1 mrg * The result may therefore not be correct. 1536 1.1 mrg * Skip wrapping and mark "wraps" as failed in this case. 1537 1.1 mrg */ 1538 1.1 mrg static isl_stat add_selected_wraps_around_facet(struct isl_wraps *wraps, 1539 1.1 mrg struct isl_coalesce_info *info, int k, isl_int *bound, 1540 1.1 mrg __isl_keep isl_set *set, int add_valid) 1541 1.1 mrg { 1542 1.1 mrg isl_bool nowrap; 1543 1.1 mrg struct isl_tab_undo *snap; 1544 1.1 mrg int n; 1545 1.1 mrg isl_size total = isl_basic_map_dim(info->bmap, isl_dim_all); 1546 1.1 mrg 1547 1.1 mrg if (total < 0) 1548 1.1 mrg return isl_stat_error; 1549 1.1 mrg 1550 1.1 mrg snap = isl_tab_snap(info->tab); 1551 1.1 mrg 1552 1.1 mrg if (isl_tab_select_facet(info->tab, info->bmap->n_eq + k) < 0) 1553 1.1 mrg return isl_stat_error; 1554 1.1 mrg if (isl_tab_detect_redundant(info->tab) < 0) 1555 1.1 mrg return isl_stat_error; 1556 1.1 mrg if (info->tab->empty) { 1557 1.1 mrg if (isl_tab_rollback(info->tab, snap) < 0) 1558 1.1 mrg return isl_stat_error; 1559 1.1 mrg if (!add_valid) 1560 1.1 mrg return wraps_mark_failed(wraps); 1561 1.1 mrg return isl_stat_ok; 1562 1.1 mrg } 1563 1.1 mrg nowrap = has_redundant_cuts(info); 1564 1.1 mrg if (nowrap < 0) 1565 1.1 mrg return isl_stat_error; 1566 1.1 mrg 1567 1.1 mrg n = wraps->mat->n_row; 1568 1.1 mrg if (!nowrap) { 1569 1.1 mrg isl_seq_neg(bound, info->bmap->ineq[k], 1 + total); 1570 1.1 mrg 1571 1.1 mrg if (add_selected_wraps(wraps, info, bound, set, add_valid) < 0) 1572 1.1 mrg return isl_stat_error; 1573 1.1 mrg } 1574 1.1 mrg 1575 1.1 mrg if (isl_tab_rollback(info->tab, snap) < 0) 1576 1.1 mrg return isl_stat_error; 1577 1.1 mrg if (nowrap) 1578 1.1 mrg return wraps_mark_failed(wraps); 1579 1.1 mrg if (check_wraps(wraps, n, info->tab, add_valid) < 0) 1580 1.1 mrg return isl_stat_error; 1581 1.1 mrg 1582 1.1 mrg return isl_stat_ok; 1583 1.1 mrg } 1584 1.1 mrg 1585 1.1 mrg /* Wrap the constraints of info->bmap that bound the facet defined 1586 1.1 mrg * by inequality "k" around (the opposite of) this inequality to 1587 1.1 mrg * include "set". "bound" may be used to store the negated inequality. 1588 1.1 mrg * If any of the wrapped constraints turn out to be invalid for info->bmap 1589 1.1 mrg * itself, then mark "wraps" as failed. 1590 1.1 mrg */ 1591 1.1 mrg static isl_stat add_wraps_around_facet(struct isl_wraps *wraps, 1592 1.1 mrg struct isl_coalesce_info *info, int k, isl_int *bound, 1593 1.1 mrg __isl_keep isl_set *set) 1594 1.1 mrg { 1595 1.1 mrg return add_selected_wraps_around_facet(wraps, info, k, bound, set, 0); 1596 1.1 mrg } 1597 1.1 mrg 1598 1.1 mrg /* Wrap the (valid) constraints of info->bmap that bound the facet defined 1599 1.1 mrg * by inequality "k" around (the opposite of) this inequality to 1600 1.1 mrg * include "set" more tightly. 1601 1.1 mrg * "bound" may be used to store the negated inequality. 1602 1.1 mrg * Remove any wrapping constraints that turn out to be invalid 1603 1.1 mrg * for info->bmap itself. 1604 1.1 mrg */ 1605 1.1 mrg static isl_stat add_valid_wraps_around_facet(struct isl_wraps *wraps, 1606 1.1 mrg struct isl_coalesce_info *info, int k, isl_int *bound, 1607 1.1 mrg __isl_keep isl_set *set) 1608 1.1 mrg { 1609 1.1 mrg return add_selected_wraps_around_facet(wraps, info, k, bound, set, 1); 1610 1.1 mrg } 1611 1.1 mrg 1612 1.1 mrg /* Basic map "i" has an inequality (say "k") that is adjacent 1613 1.1 mrg * to some inequality of basic map "j". All the other inequalities 1614 1.1 mrg * are valid for "j". 1615 1.1 mrg * Check if basic map "j" forms an extension of basic map "i". 1616 1.1 mrg * 1617 1.1 mrg * Note that this function is only called if some of the equalities or 1618 1.1 mrg * inequalities of basic map "j" do cut basic map "i". The function is 1619 1.1 mrg * correct even if there are no such cut constraints, but in that case 1620 1.1 mrg * the additional checks performed by this function are overkill. 1621 1.1 mrg * 1622 1.1 mrg * First try and wrap the ridges of "k" around "j". 1623 1.1 mrg * Note that those ridges are already valid for "j", 1624 1.1 mrg * but the wrapped versions may wrap "j" more tightly, 1625 1.1 mrg * increasing the chances of "j" being detected as an extension of "i" 1626 1.1 mrg */ 1627 1.1 mrg static enum isl_change is_adj_ineq_extension(int i, int j, 1628 1.1 mrg struct isl_coalesce_info *info) 1629 1.1 mrg { 1630 1.1 mrg int k; 1631 1.1 mrg enum isl_change change; 1632 1.1 mrg isl_size total; 1633 1.1 mrg isl_size n_eq_i, n_ineq_i; 1634 1.1 mrg struct isl_wraps wraps; 1635 1.1 mrg isl_ctx *ctx; 1636 1.1 mrg isl_mat *mat; 1637 1.1 mrg isl_vec *bound; 1638 1.1 mrg isl_set *set_j; 1639 1.1 mrg isl_stat r; 1640 1.1 mrg 1641 1.1 mrg k = find_ineq(&info[i], STATUS_ADJ_INEQ); 1642 1.1 mrg if (k < 0) 1643 1.1 mrg isl_die(isl_basic_map_get_ctx(info[i].bmap), isl_error_internal, 1644 1.1 mrg "info[i].ineq should have exactly one STATUS_ADJ_INEQ", 1645 1.1 mrg return isl_change_error); 1646 1.1 mrg 1647 1.1 mrg total = isl_basic_map_dim(info[i].bmap, isl_dim_all); 1648 1.1 mrg n_eq_i = isl_basic_map_n_equality(info[i].bmap); 1649 1.1 mrg n_ineq_i = isl_basic_map_n_inequality(info[i].bmap); 1650 1.1 mrg if (total < 0 || n_eq_i < 0 || n_ineq_i < 0) 1651 1.1 mrg return isl_change_error; 1652 1.1 mrg 1653 1.1 mrg set_j = set_from_updated_bmap(info[j].bmap, info[j].tab); 1654 1.1 mrg ctx = isl_basic_map_get_ctx(info[i].bmap); 1655 1.1 mrg bound = isl_vec_alloc(ctx, 1 + total); 1656 1.1 mrg mat = isl_mat_alloc(ctx, 2 * n_eq_i + n_ineq_i, 1 + total); 1657 1.1 mrg if (wraps_init(&wraps, mat, info, i, j) < 0) 1658 1.1 mrg goto error; 1659 1.1 mrg if (!bound || !set_j) 1660 1.1 mrg goto error; 1661 1.1 mrg r = add_valid_wraps_around_facet(&wraps, &info[i], k, bound->el, set_j); 1662 1.1 mrg if (r < 0) 1663 1.1 mrg goto error; 1664 1.1 mrg 1665 1.1 mrg change = is_adj_ineq_extension_with_wraps(i, j, k, info, wraps.mat); 1666 1.1 mrg 1667 1.1 mrg wraps_free(&wraps); 1668 1.1 mrg isl_vec_free(bound); 1669 1.1 mrg isl_set_free(set_j); 1670 1.1 mrg 1671 1.1 mrg return change; 1672 1.1 mrg error: 1673 1.1 mrg wraps_free(&wraps); 1674 1.1 mrg isl_vec_free(bound); 1675 1.1 mrg isl_set_free(set_j); 1676 1.1 mrg return isl_change_error; 1677 1.1 mrg } 1678 1.1 mrg 1679 1.1 mrg /* Both basic maps have at least one inequality with and adjacent 1680 1.1 mrg * (but opposite) inequality in the other basic map. 1681 1.1 mrg * Check that there are no cut constraints and that there is only 1682 1.1 mrg * a single pair of adjacent inequalities. 1683 1.1 mrg * If so, we can replace the pair by a single basic map described 1684 1.1 mrg * by all but the pair of adjacent inequalities. 1685 1.1 mrg * Any additional points introduced lie strictly between the two 1686 1.1 mrg * adjacent hyperplanes and can therefore be integral. 1687 1.1 mrg * 1688 1.1 mrg * ____ _____ 1689 1.1 mrg * / ||\ / \ 1690 1.1 mrg * / || \ / \ 1691 1.1 mrg * \ || \ => \ \ 1692 1.1 mrg * \ || / \ / 1693 1.1 mrg * \___||_/ \_____/ 1694 1.1 mrg * 1695 1.1 mrg * The test for a single pair of adjacent inequalities is important 1696 1.1 mrg * for avoiding the combination of two basic maps like the following 1697 1.1 mrg * 1698 1.1 mrg * /| 1699 1.1 mrg * / | 1700 1.1 mrg * /__| 1701 1.1 mrg * _____ 1702 1.1 mrg * | | 1703 1.1 mrg * | | 1704 1.1 mrg * |___| 1705 1.1 mrg * 1706 1.1 mrg * If there are some cut constraints on one side, then we may 1707 1.1 mrg * still be able to fuse the two basic maps, but we need to perform 1708 1.1 mrg * some additional checks in is_adj_ineq_extension. 1709 1.1 mrg */ 1710 1.1 mrg static enum isl_change check_adj_ineq(int i, int j, 1711 1.1 mrg struct isl_coalesce_info *info) 1712 1.1 mrg { 1713 1.1 mrg int count_i, count_j; 1714 1.1 mrg int cut_i, cut_j; 1715 1.1 mrg 1716 1.1 mrg count_i = count_ineq(&info[i], STATUS_ADJ_INEQ); 1717 1.1 mrg count_j = count_ineq(&info[j], STATUS_ADJ_INEQ); 1718 1.1 mrg 1719 1.1 mrg if (count_i != 1 && count_j != 1) 1720 1.1 mrg return isl_change_none; 1721 1.1 mrg 1722 1.1 mrg cut_i = any_eq(&info[i], STATUS_CUT) || any_ineq(&info[i], STATUS_CUT); 1723 1.1 mrg cut_j = any_eq(&info[j], STATUS_CUT) || any_ineq(&info[j], STATUS_CUT); 1724 1.1 mrg 1725 1.1 mrg if (!cut_i && !cut_j && count_i == 1 && count_j == 1) 1726 1.1 mrg return fuse(i, j, info, NULL, 0, 0); 1727 1.1 mrg 1728 1.1 mrg if (count_i == 1 && !cut_i) 1729 1.1 mrg return is_adj_ineq_extension(i, j, info); 1730 1.1 mrg 1731 1.1 mrg if (count_j == 1 && !cut_j) 1732 1.1 mrg return is_adj_ineq_extension(j, i, info); 1733 1.1 mrg 1734 1.1 mrg return isl_change_none; 1735 1.1 mrg } 1736 1.1 mrg 1737 1.1 mrg /* Given a basic set i with a constraint k that is adjacent to 1738 1.1 mrg * basic set j, check if we can wrap 1739 1.1 mrg * both the facet corresponding to k (if "wrap_facet" is set) and basic map j 1740 1.1 mrg * (always) around their ridges to include the other set. 1741 1.1 mrg * If so, replace the pair of basic sets by their union. 1742 1.1 mrg * 1743 1.1 mrg * All constraints of i (except k) are assumed to be valid or 1744 1.1 mrg * cut constraints for j. 1745 1.1 mrg * Wrapping the cut constraints to include basic map j may result 1746 1.1 mrg * in constraints that are no longer valid of basic map i 1747 1.1 mrg * we have to check that the resulting wrapping constraints are valid for i. 1748 1.1 mrg * If "wrap_facet" is not set, then all constraints of i (except k) 1749 1.1 mrg * are assumed to be valid for j. 1750 1.1 mrg * ____ _____ 1751 1.1 mrg * / | / \ 1752 1.1 mrg * / || / | 1753 1.1 mrg * \ || => \ | 1754 1.1 mrg * \ || \ | 1755 1.1 mrg * \___|| \____| 1756 1.1 mrg * 1757 1.1 mrg */ 1758 1.1 mrg static enum isl_change can_wrap_in_facet(int i, int j, int k, 1759 1.1 mrg struct isl_coalesce_info *info, int wrap_facet) 1760 1.1 mrg { 1761 1.1 mrg enum isl_change change = isl_change_none; 1762 1.1 mrg struct isl_wraps wraps; 1763 1.1 mrg isl_ctx *ctx; 1764 1.1 mrg isl_mat *mat; 1765 1.1 mrg struct isl_set *set_i = NULL; 1766 1.1 mrg struct isl_set *set_j = NULL; 1767 1.1 mrg struct isl_vec *bound = NULL; 1768 1.1 mrg isl_size total = isl_basic_map_dim(info[i].bmap, isl_dim_all); 1769 1.1 mrg 1770 1.1 mrg if (total < 0) 1771 1.1 mrg return isl_change_error; 1772 1.1 mrg set_i = set_from_updated_bmap(info[i].bmap, info[i].tab); 1773 1.1 mrg set_j = set_from_updated_bmap(info[j].bmap, info[j].tab); 1774 1.1 mrg ctx = isl_basic_map_get_ctx(info[i].bmap); 1775 1.1 mrg mat = isl_mat_alloc(ctx, 2 * (info[i].bmap->n_eq + info[j].bmap->n_eq) + 1776 1.1 mrg info[i].bmap->n_ineq + info[j].bmap->n_ineq, 1777 1.1 mrg 1 + total); 1778 1.1 mrg if (wraps_init(&wraps, mat, info, i, j) < 0) 1779 1.1 mrg goto error; 1780 1.1 mrg bound = isl_vec_alloc(ctx, 1 + total); 1781 1.1 mrg if (!set_i || !set_j || !bound) 1782 1.1 mrg goto error; 1783 1.1 mrg 1784 1.1 mrg isl_seq_cpy(bound->el, info[i].bmap->ineq[k], 1 + total); 1785 1.1 mrg isl_int_add_ui(bound->el[0], bound->el[0], 1); 1786 1.1 mrg isl_seq_normalize(ctx, bound->el, 1 + total); 1787 1.1 mrg 1788 1.1 mrg isl_seq_cpy(wraps.mat->row[0], bound->el, 1 + total); 1789 1.1 mrg wraps.mat->n_row = 1; 1790 1.1 mrg 1791 1.1 mrg if (add_wraps(&wraps, &info[j], bound->el, set_i) < 0) 1792 1.1 mrg goto error; 1793 1.1 mrg if (wraps.failed) 1794 1.1 mrg goto unbounded; 1795 1.1 mrg 1796 1.1 mrg if (wrap_facet) { 1797 1.1 mrg if (add_wraps_around_facet(&wraps, &info[i], k, 1798 1.1 mrg bound->el, set_j) < 0) 1799 1.1 mrg goto error; 1800 1.1 mrg if (wraps.failed) 1801 1.1 mrg goto unbounded; 1802 1.1 mrg } 1803 1.1 mrg 1804 1.1 mrg change = fuse(i, j, info, wraps.mat, 0, 0); 1805 1.1 mrg 1806 1.1 mrg unbounded: 1807 1.1 mrg wraps_free(&wraps); 1808 1.1 mrg 1809 1.1 mrg isl_set_free(set_i); 1810 1.1 mrg isl_set_free(set_j); 1811 1.1 mrg 1812 1.1 mrg isl_vec_free(bound); 1813 1.1 mrg 1814 1.1 mrg return change; 1815 1.1 mrg error: 1816 1.1 mrg wraps_free(&wraps); 1817 1.1 mrg isl_vec_free(bound); 1818 1.1 mrg isl_set_free(set_i); 1819 1.1 mrg isl_set_free(set_j); 1820 1.1 mrg return isl_change_error; 1821 1.1 mrg } 1822 1.1 mrg 1823 1.1 mrg /* Given a cut constraint t(x) >= 0 of basic map i, stored in row "w" 1824 1.1 mrg * of wrap.mat, replace it by its relaxed version t(x) + 1 >= 0, and 1825 1.1 mrg * add wrapping constraints to wrap.mat for all constraints 1826 1.1 mrg * of basic map j that bound the part of basic map j that sticks out 1827 1.1 mrg * of the cut constraint. 1828 1.1 mrg * "set_i" is the underlying set of basic map i. 1829 1.1 mrg * If any wrapping fails, then wraps->mat.n_row is reset to zero. 1830 1.1 mrg * 1831 1.1 mrg * In particular, we first intersect basic map j with t(x) + 1 = 0. 1832 1.1 mrg * If the result is empty, then t(x) >= 0 was actually a valid constraint 1833 1.1 mrg * (with respect to the integer points), so we add t(x) >= 0 instead. 1834 1.1 mrg * Otherwise, we wrap the constraints of basic map j that are not 1835 1.1 mrg * redundant in this intersection and that are not already valid 1836 1.1 mrg * for basic map i over basic map i. 1837 1.1 mrg * Note that it is sufficient to wrap the constraints to include 1838 1.1 mrg * basic map i, because we will only wrap the constraints that do 1839 1.1 mrg * not include basic map i already. The wrapped constraint will 1840 1.1 mrg * therefore be more relaxed compared to the original constraint. 1841 1.1 mrg * Since the original constraint is valid for basic map j, so is 1842 1.1 mrg * the wrapped constraint. 1843 1.1 mrg */ 1844 1.1 mrg static isl_stat wrap_in_facet(struct isl_wraps *wraps, int w, 1845 1.1 mrg struct isl_coalesce_info *info_j, __isl_keep isl_set *set_i, 1846 1.1 mrg struct isl_tab_undo *snap) 1847 1.1 mrg { 1848 1.1 mrg isl_int_add_ui(wraps->mat->row[w][0], wraps->mat->row[w][0], 1); 1849 1.1 mrg if (isl_tab_add_eq(info_j->tab, wraps->mat->row[w]) < 0) 1850 1.1 mrg return isl_stat_error; 1851 1.1 mrg if (isl_tab_detect_redundant(info_j->tab) < 0) 1852 1.1 mrg return isl_stat_error; 1853 1.1 mrg 1854 1.1 mrg if (info_j->tab->empty) 1855 1.1 mrg isl_int_sub_ui(wraps->mat->row[w][0], wraps->mat->row[w][0], 1); 1856 1.1 mrg else if (add_wraps(wraps, info_j, wraps->mat->row[w], set_i) < 0) 1857 1.1 mrg return isl_stat_error; 1858 1.1 mrg 1859 1.1 mrg if (isl_tab_rollback(info_j->tab, snap) < 0) 1860 1.1 mrg return isl_stat_error; 1861 1.1 mrg 1862 1.1 mrg return isl_stat_ok; 1863 1.1 mrg } 1864 1.1 mrg 1865 1.1 mrg /* Given a pair of basic maps i and j such that j sticks out 1866 1.1 mrg * of i at n cut constraints, each time by at most one, 1867 1.1 mrg * try to compute wrapping constraints and replace the two 1868 1.1 mrg * basic maps by a single basic map. 1869 1.1 mrg * The other constraints of i are assumed to be valid for j. 1870 1.1 mrg * "set_i" is the underlying set of basic map i. 1871 1.1 mrg * "wraps" has been initialized to be of the right size. 1872 1.1 mrg * 1873 1.1 mrg * For each cut constraint t(x) >= 0 of i, we add the relaxed version 1874 1.1 mrg * t(x) + 1 >= 0, along with wrapping constraints for all constraints 1875 1.1 mrg * of basic map j that bound the part of basic map j that sticks out 1876 1.1 mrg * of the cut constraint. 1877 1.1 mrg * 1878 1.1 mrg * If any wrapping fails, i.e., if we cannot wrap to touch 1879 1.1 mrg * the union, then we give up. 1880 1.1 mrg * Otherwise, the pair of basic maps is replaced by their union. 1881 1.1 mrg */ 1882 1.1 mrg static enum isl_change try_wrap_in_facets(int i, int j, 1883 1.1 mrg struct isl_coalesce_info *info, struct isl_wraps *wraps, 1884 1.1 mrg __isl_keep isl_set *set_i) 1885 1.1 mrg { 1886 1.1 mrg int k, l, w; 1887 1.1 mrg isl_size total; 1888 1.1 mrg struct isl_tab_undo *snap; 1889 1.1 mrg 1890 1.1 mrg total = isl_basic_map_dim(info[i].bmap, isl_dim_all); 1891 1.1 mrg if (total < 0) 1892 1.1 mrg return isl_change_error; 1893 1.1 mrg 1894 1.1 mrg snap = isl_tab_snap(info[j].tab); 1895 1.1 mrg 1896 1.1 mrg for (k = 0; k < info[i].bmap->n_eq; ++k) { 1897 1.1 mrg for (l = 0; l < 2; ++l) { 1898 1.1 mrg if (info[i].eq[2 * k + l] != STATUS_CUT) 1899 1.1 mrg continue; 1900 1.1 mrg w = wraps->mat->n_row++; 1901 1.1 mrg if (l == 0) 1902 1.1 mrg isl_seq_neg(wraps->mat->row[w], 1903 1.1 mrg info[i].bmap->eq[k], 1 + total); 1904 1.1 mrg else 1905 1.1 mrg isl_seq_cpy(wraps->mat->row[w], 1906 1.1 mrg info[i].bmap->eq[k], 1 + total); 1907 1.1 mrg if (wrap_in_facet(wraps, w, &info[j], set_i, snap) < 0) 1908 1.1 mrg return isl_change_error; 1909 1.1 mrg 1910 1.1 mrg if (wraps->failed) 1911 1.1 mrg return isl_change_none; 1912 1.1 mrg } 1913 1.1 mrg } 1914 1.1 mrg 1915 1.1 mrg for (k = 0; k < info[i].bmap->n_ineq; ++k) { 1916 1.1 mrg if (info[i].ineq[k] != STATUS_CUT) 1917 1.1 mrg continue; 1918 1.1 mrg w = wraps->mat->n_row++; 1919 1.1 mrg isl_seq_cpy(wraps->mat->row[w], 1920 1.1 mrg info[i].bmap->ineq[k], 1 + total); 1921 1.1 mrg if (wrap_in_facet(wraps, w, &info[j], set_i, snap) < 0) 1922 1.1 mrg return isl_change_error; 1923 1.1 mrg 1924 1.1 mrg if (wraps->failed) 1925 1.1 mrg return isl_change_none; 1926 1.1 mrg } 1927 1.1 mrg 1928 1.1 mrg return fuse(i, j, info, wraps->mat, 0, 1); 1929 1.1 mrg } 1930 1.1 mrg 1931 1.1 mrg /* Given a pair of basic maps i and j such that j sticks out 1932 1.1 mrg * of i at n cut constraints, each time by at most one, 1933 1.1 mrg * try to compute wrapping constraints and replace the two 1934 1.1 mrg * basic maps by a single basic map. 1935 1.1 mrg * The other constraints of i are assumed to be valid for j. 1936 1.1 mrg * 1937 1.1 mrg * The core computation is performed by try_wrap_in_facets. 1938 1.1 mrg * This function simply extracts an underlying set representation 1939 1.1 mrg * of basic map i and initializes the data structure for keeping 1940 1.1 mrg * track of wrapping constraints. 1941 1.1 mrg */ 1942 1.1 mrg static enum isl_change wrap_in_facets(int i, int j, int n, 1943 1.1 mrg struct isl_coalesce_info *info) 1944 1.1 mrg { 1945 1.1 mrg enum isl_change change = isl_change_none; 1946 1.1 mrg struct isl_wraps wraps; 1947 1.1 mrg isl_ctx *ctx; 1948 1.1 mrg isl_mat *mat; 1949 1.1 mrg isl_set *set_i = NULL; 1950 1.1 mrg isl_size total = isl_basic_map_dim(info[i].bmap, isl_dim_all); 1951 1.1 mrg int max_wrap; 1952 1.1 mrg 1953 1.1 mrg if (total < 0) 1954 1.1 mrg return isl_change_error; 1955 1.1 mrg if (isl_tab_extend_cons(info[j].tab, 1) < 0) 1956 1.1 mrg return isl_change_error; 1957 1.1 mrg 1958 1.1 mrg max_wrap = 1 + 2 * info[j].bmap->n_eq + info[j].bmap->n_ineq; 1959 1.1 mrg max_wrap *= n; 1960 1.1 mrg 1961 1.1 mrg set_i = set_from_updated_bmap(info[i].bmap, info[i].tab); 1962 1.1 mrg ctx = isl_basic_map_get_ctx(info[i].bmap); 1963 1.1 mrg mat = isl_mat_alloc(ctx, max_wrap, 1 + total); 1964 1.1 mrg if (wraps_init(&wraps, mat, info, i, j) < 0) 1965 1.1 mrg goto error; 1966 1.1 mrg if (!set_i) 1967 1.1 mrg goto error; 1968 1.1 mrg 1969 1.1 mrg change = try_wrap_in_facets(i, j, info, &wraps, set_i); 1970 1.1 mrg 1971 1.1 mrg wraps_free(&wraps); 1972 1.1 mrg isl_set_free(set_i); 1973 1.1 mrg 1974 1.1 mrg return change; 1975 1.1 mrg error: 1976 1.1 mrg wraps_free(&wraps); 1977 1.1 mrg isl_set_free(set_i); 1978 1.1 mrg return isl_change_error; 1979 1.1 mrg } 1980 1.1 mrg 1981 1.1 mrg /* Return the effect of inequality "ineq" on the tableau "tab", 1982 1.1 mrg * after relaxing the constant term of "ineq" by one. 1983 1.1 mrg */ 1984 1.1 mrg static enum isl_ineq_type type_of_relaxed(struct isl_tab *tab, isl_int *ineq) 1985 1.1 mrg { 1986 1.1 mrg enum isl_ineq_type type; 1987 1.1 mrg 1988 1.1 mrg isl_int_add_ui(ineq[0], ineq[0], 1); 1989 1.1 mrg type = isl_tab_ineq_type(tab, ineq); 1990 1.1 mrg isl_int_sub_ui(ineq[0], ineq[0], 1); 1991 1.1 mrg 1992 1.1 mrg return type; 1993 1.1 mrg } 1994 1.1 mrg 1995 1.1 mrg /* Given two basic sets i and j, 1996 1.1 mrg * check if relaxing all the cut constraints of i by one turns 1997 1.1 mrg * them into valid constraint for j and check if we can wrap in 1998 1.1 mrg * the bits that are sticking out. 1999 1.1 mrg * If so, replace the pair by their union. 2000 1.1 mrg * 2001 1.1 mrg * We first check if all relaxed cut inequalities of i are valid for j 2002 1.1 mrg * and then try to wrap in the intersections of the relaxed cut inequalities 2003 1.1 mrg * with j. 2004 1.1 mrg * 2005 1.1 mrg * During this wrapping, we consider the points of j that lie at a distance 2006 1.1 mrg * of exactly 1 from i. In particular, we ignore the points that lie in 2007 1.1 mrg * between this lower-dimensional space and the basic map i. 2008 1.1 mrg * We can therefore only apply this to integer maps. 2009 1.1 mrg * ____ _____ 2010 1.1 mrg * / ___|_ / \ 2011 1.1 mrg * / | | / | 2012 1.1 mrg * \ | | => \ | 2013 1.1 mrg * \|____| \ | 2014 1.1 mrg * \___| \____/ 2015 1.1 mrg * 2016 1.1 mrg * _____ ______ 2017 1.1 mrg * | ____|_ | \ 2018 1.1 mrg * | | | | | 2019 1.1 mrg * | | | => | | 2020 1.1 mrg * |_| | | | 2021 1.1 mrg * |_____| \______| 2022 1.1 mrg * 2023 1.1 mrg * _______ 2024 1.1 mrg * | | 2025 1.1 mrg * | |\ | 2026 1.1 mrg * | | \ | 2027 1.1 mrg * | | \ | 2028 1.1 mrg * | | \| 2029 1.1 mrg * | | \ 2030 1.1 mrg * | |_____\ 2031 1.1 mrg * | | 2032 1.1 mrg * |_______| 2033 1.1 mrg * 2034 1.1 mrg * Wrapping can fail if the result of wrapping one of the facets 2035 1.1 mrg * around its edges does not produce any new facet constraint. 2036 1.1 mrg * In particular, this happens when we try to wrap in unbounded sets. 2037 1.1 mrg * 2038 1.1 mrg * _______________________________________________________________________ 2039 1.1 mrg * | 2040 1.1 mrg * | ___ 2041 1.1 mrg * | | | 2042 1.1 mrg * |_| |_________________________________________________________________ 2043 1.1 mrg * |___| 2044 1.1 mrg * 2045 1.1 mrg * The following is not an acceptable result of coalescing the above two 2046 1.1 mrg * sets as it includes extra integer points. 2047 1.1 mrg * _______________________________________________________________________ 2048 1.1 mrg * | 2049 1.1 mrg * | 2050 1.1 mrg * | 2051 1.1 mrg * | 2052 1.1 mrg * \______________________________________________________________________ 2053 1.1 mrg */ 2054 1.1 mrg static enum isl_change can_wrap_in_set(int i, int j, 2055 1.1 mrg struct isl_coalesce_info *info) 2056 1.1 mrg { 2057 1.1 mrg int k, l; 2058 1.1 mrg int n; 2059 1.1 mrg isl_size total; 2060 1.1 mrg 2061 1.1 mrg if (ISL_F_ISSET(info[i].bmap, ISL_BASIC_MAP_RATIONAL) || 2062 1.1 mrg ISL_F_ISSET(info[j].bmap, ISL_BASIC_MAP_RATIONAL)) 2063 1.1 mrg return isl_change_none; 2064 1.1 mrg 2065 1.1 mrg n = count_eq(&info[i], STATUS_CUT) + count_ineq(&info[i], STATUS_CUT); 2066 1.1 mrg if (n == 0) 2067 1.1 mrg return isl_change_none; 2068 1.1 mrg 2069 1.1 mrg total = isl_basic_map_dim(info[i].bmap, isl_dim_all); 2070 1.1 mrg if (total < 0) 2071 1.1 mrg return isl_change_error; 2072 1.1 mrg for (k = 0; k < info[i].bmap->n_eq; ++k) { 2073 1.1 mrg for (l = 0; l < 2; ++l) { 2074 1.1 mrg enum isl_ineq_type type; 2075 1.1 mrg 2076 1.1 mrg if (info[i].eq[2 * k + l] != STATUS_CUT) 2077 1.1 mrg continue; 2078 1.1 mrg 2079 1.1 mrg if (l == 0) 2080 1.1 mrg isl_seq_neg(info[i].bmap->eq[k], 2081 1.1 mrg info[i].bmap->eq[k], 1 + total); 2082 1.1 mrg type = type_of_relaxed(info[j].tab, 2083 1.1 mrg info[i].bmap->eq[k]); 2084 1.1 mrg if (l == 0) 2085 1.1 mrg isl_seq_neg(info[i].bmap->eq[k], 2086 1.1 mrg info[i].bmap->eq[k], 1 + total); 2087 1.1 mrg if (type == isl_ineq_error) 2088 1.1 mrg return isl_change_error; 2089 1.1 mrg if (type != isl_ineq_redundant) 2090 1.1 mrg return isl_change_none; 2091 1.1 mrg } 2092 1.1 mrg } 2093 1.1 mrg 2094 1.1 mrg for (k = 0; k < info[i].bmap->n_ineq; ++k) { 2095 1.1 mrg enum isl_ineq_type type; 2096 1.1 mrg 2097 1.1 mrg if (info[i].ineq[k] != STATUS_CUT) 2098 1.1 mrg continue; 2099 1.1 mrg 2100 1.1 mrg type = type_of_relaxed(info[j].tab, info[i].bmap->ineq[k]); 2101 1.1 mrg if (type == isl_ineq_error) 2102 1.1 mrg return isl_change_error; 2103 1.1 mrg if (type != isl_ineq_redundant) 2104 1.1 mrg return isl_change_none; 2105 1.1 mrg } 2106 1.1 mrg 2107 1.1 mrg return wrap_in_facets(i, j, n, info); 2108 1.1 mrg } 2109 1.1 mrg 2110 1.1 mrg /* Check if either i or j has only cut constraints that can 2111 1.1 mrg * be used to wrap in (a facet of) the other basic set. 2112 1.1 mrg * if so, replace the pair by their union. 2113 1.1 mrg */ 2114 1.1 mrg static enum isl_change check_wrap(int i, int j, struct isl_coalesce_info *info) 2115 1.1 mrg { 2116 1.1 mrg enum isl_change change = isl_change_none; 2117 1.1 mrg 2118 1.1 mrg change = can_wrap_in_set(i, j, info); 2119 1.1 mrg if (change != isl_change_none) 2120 1.1 mrg return change; 2121 1.1 mrg 2122 1.1 mrg change = can_wrap_in_set(j, i, info); 2123 1.1 mrg return change; 2124 1.1 mrg } 2125 1.1 mrg 2126 1.1 mrg /* Check if all inequality constraints of "i" that cut "j" cease 2127 1.1 mrg * to be cut constraints if they are relaxed by one. 2128 1.1 mrg * If so, collect the cut constraints in "list". 2129 1.1 mrg * The caller is responsible for allocating "list". 2130 1.1 mrg */ 2131 1.1 mrg static isl_bool all_cut_by_one(int i, int j, struct isl_coalesce_info *info, 2132 1.1 mrg int *list) 2133 1.1 mrg { 2134 1.1 mrg int l, n; 2135 1.1 mrg 2136 1.1 mrg n = 0; 2137 1.1 mrg for (l = 0; l < info[i].bmap->n_ineq; ++l) { 2138 1.1 mrg enum isl_ineq_type type; 2139 1.1 mrg 2140 1.1 mrg if (info[i].ineq[l] != STATUS_CUT) 2141 1.1 mrg continue; 2142 1.1 mrg type = type_of_relaxed(info[j].tab, info[i].bmap->ineq[l]); 2143 1.1 mrg if (type == isl_ineq_error) 2144 1.1 mrg return isl_bool_error; 2145 1.1 mrg if (type != isl_ineq_redundant) 2146 1.1 mrg return isl_bool_false; 2147 1.1 mrg list[n++] = l; 2148 1.1 mrg } 2149 1.1 mrg 2150 1.1 mrg return isl_bool_true; 2151 1.1 mrg } 2152 1.1 mrg 2153 1.1 mrg /* Given two basic maps such that "j" has at least one equality constraint 2154 1.1 mrg * that is adjacent to an inequality constraint of "i" and such that "i" has 2155 1.1 mrg * exactly one inequality constraint that is adjacent to an equality 2156 1.1 mrg * constraint of "j", check whether "i" can be extended to include "j" or 2157 1.1 mrg * whether "j" can be wrapped into "i". 2158 1.1 mrg * All remaining constraints of "i" and "j" are assumed to be valid 2159 1.1 mrg * or cut constraints of the other basic map. 2160 1.1 mrg * However, none of the equality constraints of "i" are cut constraints. 2161 1.1 mrg * 2162 1.1 mrg * If "i" has any "cut" inequality constraints, then check if relaxing 2163 1.1 mrg * each of them by one is sufficient for them to become valid. 2164 1.1 mrg * If so, check if the inequality constraint adjacent to an equality 2165 1.1 mrg * constraint of "j" along with all these cut constraints 2166 1.1 mrg * can be relaxed by one to contain exactly "j". 2167 1.1 mrg * Otherwise, or if this fails, check if "j" can be wrapped into "i". 2168 1.1 mrg */ 2169 1.1 mrg static enum isl_change check_single_adj_eq(int i, int j, 2170 1.1 mrg struct isl_coalesce_info *info) 2171 1.1 mrg { 2172 1.1 mrg enum isl_change change = isl_change_none; 2173 1.1 mrg int k; 2174 1.1 mrg int n_cut; 2175 1.1 mrg int *relax; 2176 1.1 mrg isl_ctx *ctx; 2177 1.1 mrg isl_bool try_relax; 2178 1.1 mrg 2179 1.1 mrg n_cut = count_ineq(&info[i], STATUS_CUT); 2180 1.1 mrg 2181 1.1 mrg k = find_ineq(&info[i], STATUS_ADJ_EQ); 2182 1.1 mrg 2183 1.1 mrg if (n_cut > 0) { 2184 1.1 mrg ctx = isl_basic_map_get_ctx(info[i].bmap); 2185 1.1 mrg relax = isl_calloc_array(ctx, int, 1 + n_cut); 2186 1.1 mrg if (!relax) 2187 1.1 mrg return isl_change_error; 2188 1.1 mrg relax[0] = k; 2189 1.1 mrg try_relax = all_cut_by_one(i, j, info, relax + 1); 2190 1.1 mrg if (try_relax < 0) 2191 1.1 mrg change = isl_change_error; 2192 1.1 mrg } else { 2193 1.1 mrg try_relax = isl_bool_true; 2194 1.1 mrg relax = &k; 2195 1.1 mrg } 2196 1.1 mrg if (try_relax && change == isl_change_none) 2197 1.1 mrg change = is_relaxed_extension(i, j, 1 + n_cut, relax, info); 2198 1.1 mrg if (n_cut > 0) 2199 1.1 mrg free(relax); 2200 1.1 mrg if (change != isl_change_none) 2201 1.1 mrg return change; 2202 1.1 mrg 2203 1.1 mrg change = can_wrap_in_facet(i, j, k, info, n_cut > 0); 2204 1.1 mrg 2205 1.1 mrg return change; 2206 1.1 mrg } 2207 1.1 mrg 2208 1.1 mrg /* At least one of the basic maps has an equality that is adjacent 2209 1.1 mrg * to an inequality. Make sure that only one of the basic maps has 2210 1.1 mrg * such an equality and that the other basic map has exactly one 2211 1.1 mrg * inequality adjacent to an equality. 2212 1.1 mrg * If the other basic map does not have such an inequality, then 2213 1.1 mrg * check if all its constraints are either valid or cut constraints 2214 1.1 mrg * and, if so, try wrapping in the first map into the second. 2215 1.1 mrg * Otherwise, try to extend one basic map with the other or 2216 1.1 mrg * wrap one basic map in the other. 2217 1.1 mrg */ 2218 1.1 mrg static enum isl_change check_adj_eq(int i, int j, 2219 1.1 mrg struct isl_coalesce_info *info) 2220 1.1 mrg { 2221 1.1 mrg if (any_eq(&info[i], STATUS_ADJ_INEQ) && 2222 1.1 mrg any_eq(&info[j], STATUS_ADJ_INEQ)) 2223 1.1 mrg /* ADJ EQ TOO MANY */ 2224 1.1 mrg return isl_change_none; 2225 1.1 mrg 2226 1.1 mrg if (any_eq(&info[i], STATUS_ADJ_INEQ)) 2227 1.1 mrg return check_adj_eq(j, i, info); 2228 1.1 mrg 2229 1.1 mrg /* j has an equality adjacent to an inequality in i */ 2230 1.1 mrg 2231 1.1 mrg if (count_ineq(&info[i], STATUS_ADJ_EQ) != 1) { 2232 1.1 mrg if (all_valid_or_cut(&info[i])) 2233 1.1 mrg return can_wrap_in_set(i, j, info); 2234 1.1 mrg return isl_change_none; 2235 1.1 mrg } 2236 1.1 mrg if (any_eq(&info[i], STATUS_CUT)) 2237 1.1 mrg return isl_change_none; 2238 1.1 mrg if (any_ineq(&info[j], STATUS_ADJ_EQ) || 2239 1.1 mrg any_ineq(&info[i], STATUS_ADJ_INEQ) || 2240 1.1 mrg any_ineq(&info[j], STATUS_ADJ_INEQ)) 2241 1.1 mrg /* ADJ EQ TOO MANY */ 2242 1.1 mrg return isl_change_none; 2243 1.1 mrg 2244 1.1 mrg return check_single_adj_eq(i, j, info); 2245 1.1 mrg } 2246 1.1 mrg 2247 1.1 mrg /* Disjunct "j" lies on a hyperplane that is adjacent to disjunct "i". 2248 1.1 mrg * In particular, disjunct "i" has an inequality constraint that is adjacent 2249 1.1 mrg * to a (combination of) equality constraint(s) of disjunct "j", 2250 1.1 mrg * but disjunct "j" has no explicit equality constraint adjacent 2251 1.1 mrg * to an inequality constraint of disjunct "i". 2252 1.1 mrg * 2253 1.1 mrg * Disjunct "i" is already known not to have any equality constraints 2254 1.1 mrg * that are adjacent to an equality or inequality constraint. 2255 1.1 mrg * Check that, other than the inequality constraint mentioned above, 2256 1.1 mrg * all other constraints of disjunct "i" are valid for disjunct "j". 2257 1.1 mrg * If so, try and wrap in disjunct "j". 2258 1.1 mrg */ 2259 1.1 mrg static enum isl_change check_ineq_adj_eq(int i, int j, 2260 1.1 mrg struct isl_coalesce_info *info) 2261 1.1 mrg { 2262 1.1 mrg int k; 2263 1.1 mrg 2264 1.1 mrg if (any_eq(&info[i], STATUS_CUT)) 2265 1.1 mrg return isl_change_none; 2266 1.1 mrg if (any_ineq(&info[i], STATUS_CUT)) 2267 1.1 mrg return isl_change_none; 2268 1.1 mrg if (any_ineq(&info[i], STATUS_ADJ_INEQ)) 2269 1.1 mrg return isl_change_none; 2270 1.1 mrg if (count_ineq(&info[i], STATUS_ADJ_EQ) != 1) 2271 1.1 mrg return isl_change_none; 2272 1.1 mrg 2273 1.1 mrg k = find_ineq(&info[i], STATUS_ADJ_EQ); 2274 1.1 mrg 2275 1.1 mrg return can_wrap_in_facet(i, j, k, info, 0); 2276 1.1 mrg } 2277 1.1 mrg 2278 1.1 mrg /* The two basic maps lie on adjacent hyperplanes. In particular, 2279 1.1 mrg * basic map "i" has an equality that lies parallel to basic map "j". 2280 1.1 mrg * Check if we can wrap the facets around the parallel hyperplanes 2281 1.1 mrg * to include the other set. 2282 1.1 mrg * 2283 1.1 mrg * We perform basically the same operations as can_wrap_in_facet, 2284 1.1 mrg * except that we don't need to select a facet of one of the sets. 2285 1.1 mrg * _ 2286 1.1 mrg * \\ \\ 2287 1.1 mrg * \\ => \\ 2288 1.1 mrg * \ \| 2289 1.1 mrg * 2290 1.1 mrg * If there is more than one equality of "i" adjacent to an equality of "j", 2291 1.1 mrg * then the result will satisfy one or more equalities that are a linear 2292 1.1 mrg * combination of these equalities. These will be encoded as pairs 2293 1.1 mrg * of inequalities in the wrapping constraints and need to be made 2294 1.1 mrg * explicit. 2295 1.1 mrg */ 2296 1.1 mrg static enum isl_change check_eq_adj_eq(int i, int j, 2297 1.1 mrg struct isl_coalesce_info *info) 2298 1.1 mrg { 2299 1.1 mrg int k; 2300 1.1 mrg enum isl_change change = isl_change_none; 2301 1.1 mrg int detect_equalities = 0; 2302 1.1 mrg struct isl_wraps wraps; 2303 1.1 mrg isl_ctx *ctx; 2304 1.1 mrg isl_mat *mat; 2305 1.1 mrg struct isl_set *set_i = NULL; 2306 1.1 mrg struct isl_set *set_j = NULL; 2307 1.1 mrg struct isl_vec *bound = NULL; 2308 1.1 mrg isl_size total = isl_basic_map_dim(info[i].bmap, isl_dim_all); 2309 1.1 mrg 2310 1.1 mrg if (total < 0) 2311 1.1 mrg return isl_change_error; 2312 1.1 mrg if (count_eq(&info[i], STATUS_ADJ_EQ) != 1) 2313 1.1 mrg detect_equalities = 1; 2314 1.1 mrg 2315 1.1 mrg k = find_eq(&info[i], STATUS_ADJ_EQ); 2316 1.1 mrg 2317 1.1 mrg set_i = set_from_updated_bmap(info[i].bmap, info[i].tab); 2318 1.1 mrg set_j = set_from_updated_bmap(info[j].bmap, info[j].tab); 2319 1.1 mrg ctx = isl_basic_map_get_ctx(info[i].bmap); 2320 1.1 mrg mat = isl_mat_alloc(ctx, 2 * (info[i].bmap->n_eq + info[j].bmap->n_eq) + 2321 1.1 mrg info[i].bmap->n_ineq + info[j].bmap->n_ineq, 2322 1.1 mrg 1 + total); 2323 1.1 mrg if (wraps_init(&wraps, mat, info, i, j) < 0) 2324 1.1 mrg goto error; 2325 1.1 mrg bound = isl_vec_alloc(ctx, 1 + total); 2326 1.1 mrg if (!set_i || !set_j || !bound) 2327 1.1 mrg goto error; 2328 1.1 mrg 2329 1.1 mrg if (k % 2 == 0) 2330 1.1 mrg isl_seq_neg(bound->el, info[i].bmap->eq[k / 2], 1 + total); 2331 1.1 mrg else 2332 1.1 mrg isl_seq_cpy(bound->el, info[i].bmap->eq[k / 2], 1 + total); 2333 1.1 mrg isl_int_add_ui(bound->el[0], bound->el[0], 1); 2334 1.1 mrg 2335 1.1 mrg isl_seq_cpy(wraps.mat->row[0], bound->el, 1 + total); 2336 1.1 mrg wraps.mat->n_row = 1; 2337 1.1 mrg 2338 1.1 mrg if (add_wraps(&wraps, &info[j], bound->el, set_i) < 0) 2339 1.1 mrg goto error; 2340 1.1 mrg if (wraps.failed) 2341 1.1 mrg goto unbounded; 2342 1.1 mrg 2343 1.1 mrg isl_int_sub_ui(bound->el[0], bound->el[0], 1); 2344 1.1 mrg isl_seq_neg(bound->el, bound->el, 1 + total); 2345 1.1 mrg 2346 1.1 mrg isl_seq_cpy(wraps.mat->row[wraps.mat->n_row], bound->el, 1 + total); 2347 1.1 mrg wraps.mat->n_row++; 2348 1.1 mrg 2349 1.1 mrg if (add_wraps(&wraps, &info[i], bound->el, set_j) < 0) 2350 1.1 mrg goto error; 2351 1.1 mrg if (wraps.failed) 2352 1.1 mrg goto unbounded; 2353 1.1 mrg 2354 1.1 mrg change = fuse(i, j, info, wraps.mat, detect_equalities, 0); 2355 1.1 mrg 2356 1.1 mrg if (0) { 2357 1.1 mrg error: change = isl_change_error; 2358 1.1 mrg } 2359 1.1 mrg unbounded: 2360 1.1 mrg 2361 1.1 mrg wraps_free(&wraps); 2362 1.1 mrg isl_set_free(set_i); 2363 1.1 mrg isl_set_free(set_j); 2364 1.1 mrg isl_vec_free(bound); 2365 1.1 mrg 2366 1.1 mrg return change; 2367 1.1 mrg } 2368 1.1 mrg 2369 1.1 mrg /* Initialize the "eq" and "ineq" fields of "info". 2370 1.1 mrg */ 2371 1.1 mrg static void init_status(struct isl_coalesce_info *info) 2372 1.1 mrg { 2373 1.1 mrg info->eq = info->ineq = NULL; 2374 1.1 mrg } 2375 1.1 mrg 2376 1.1 mrg /* Set info->eq to the positions of the equalities of info->bmap 2377 1.1 mrg * with respect to the basic map represented by "tab". 2378 1.1 mrg * If info->eq has already been computed, then do not compute it again. 2379 1.1 mrg */ 2380 1.1 mrg static void set_eq_status_in(struct isl_coalesce_info *info, 2381 1.1 mrg struct isl_tab *tab) 2382 1.1 mrg { 2383 1.1 mrg if (info->eq) 2384 1.1 mrg return; 2385 1.1 mrg info->eq = eq_status_in(info->bmap, tab); 2386 1.1 mrg } 2387 1.1 mrg 2388 1.1 mrg /* Set info->ineq to the positions of the inequalities of info->bmap 2389 1.1 mrg * with respect to the basic map represented by "tab". 2390 1.1 mrg * If info->ineq has already been computed, then do not compute it again. 2391 1.1 mrg */ 2392 1.1 mrg static void set_ineq_status_in(struct isl_coalesce_info *info, 2393 1.1 mrg struct isl_tab *tab) 2394 1.1 mrg { 2395 1.1 mrg if (info->ineq) 2396 1.1 mrg return; 2397 1.1 mrg info->ineq = ineq_status_in(info->bmap, info->tab, tab); 2398 1.1 mrg } 2399 1.1 mrg 2400 1.1 mrg /* Free the memory allocated by the "eq" and "ineq" fields of "info". 2401 1.1 mrg * This function assumes that init_status has been called on "info" first, 2402 1.1 mrg * after which the "eq" and "ineq" fields may or may not have been 2403 1.1 mrg * assigned a newly allocated array. 2404 1.1 mrg */ 2405 1.1 mrg static void clear_status(struct isl_coalesce_info *info) 2406 1.1 mrg { 2407 1.1 mrg free(info->eq); 2408 1.1 mrg free(info->ineq); 2409 1.1 mrg } 2410 1.1 mrg 2411 1.1 mrg /* Are all inequality constraints of the basic map represented by "info" 2412 1.1 mrg * valid for the other basic map, except for a single constraint 2413 1.1 mrg * that is adjacent to an inequality constraint of the other basic map? 2414 1.1 mrg */ 2415 1.1 mrg static int all_ineq_valid_or_single_adj_ineq(struct isl_coalesce_info *info) 2416 1.1 mrg { 2417 1.1 mrg int i; 2418 1.1 mrg int k = -1; 2419 1.1 mrg 2420 1.1 mrg for (i = 0; i < info->bmap->n_ineq; ++i) { 2421 1.1 mrg if (info->ineq[i] == STATUS_REDUNDANT) 2422 1.1 mrg continue; 2423 1.1 mrg if (info->ineq[i] == STATUS_VALID) 2424 1.1 mrg continue; 2425 1.1 mrg if (info->ineq[i] != STATUS_ADJ_INEQ) 2426 1.1 mrg return 0; 2427 1.1 mrg if (k != -1) 2428 1.1 mrg return 0; 2429 1.1 mrg k = i; 2430 1.1 mrg } 2431 1.1 mrg 2432 1.1 mrg return k != -1; 2433 1.1 mrg } 2434 1.1 mrg 2435 1.1 mrg /* Basic map "i" has one or more equality constraints that separate it 2436 1.1 mrg * from basic map "j". Check if it happens to be an extension 2437 1.1 mrg * of basic map "j". 2438 1.1 mrg * In particular, check that all constraints of "j" are valid for "i", 2439 1.1 mrg * except for one inequality constraint that is adjacent 2440 1.1 mrg * to an inequality constraints of "i". 2441 1.1 mrg * If so, check for "i" being an extension of "j" by calling 2442 1.1 mrg * is_adj_ineq_extension. 2443 1.1 mrg * 2444 1.1 mrg * Clean up the memory allocated for keeping track of the status 2445 1.1 mrg * of the constraints before returning. 2446 1.1 mrg */ 2447 1.1 mrg static enum isl_change separating_equality(int i, int j, 2448 1.1 mrg struct isl_coalesce_info *info) 2449 1.1 mrg { 2450 1.1 mrg enum isl_change change = isl_change_none; 2451 1.1 mrg 2452 1.1 mrg if (all(info[j].eq, 2 * info[j].bmap->n_eq, STATUS_VALID) && 2453 1.1 mrg all_ineq_valid_or_single_adj_ineq(&info[j])) 2454 1.1 mrg change = is_adj_ineq_extension(j, i, info); 2455 1.1 mrg 2456 1.1 mrg clear_status(&info[i]); 2457 1.1 mrg clear_status(&info[j]); 2458 1.1 mrg return change; 2459 1.1 mrg } 2460 1.1 mrg 2461 1.1 mrg /* Check if the union of the given pair of basic maps 2462 1.1 mrg * can be represented by a single basic map. 2463 1.1 mrg * If so, replace the pair by the single basic map and return 2464 1.1 mrg * isl_change_drop_first, isl_change_drop_second or isl_change_fuse. 2465 1.1 mrg * Otherwise, return isl_change_none. 2466 1.1 mrg * The two basic maps are assumed to live in the same local space. 2467 1.1 mrg * The "eq" and "ineq" fields of info[i] and info[j] are assumed 2468 1.1 mrg * to have been initialized by the caller, either to NULL or 2469 1.1 mrg * to valid information. 2470 1.1 mrg * 2471 1.1 mrg * We first check the effect of each constraint of one basic map 2472 1.1 mrg * on the other basic map. 2473 1.1 mrg * The constraint may be 2474 1.1 mrg * redundant the constraint is redundant in its own 2475 1.1 mrg * basic map and should be ignore and removed 2476 1.1 mrg * in the end 2477 1.1 mrg * valid all (integer) points of the other basic map 2478 1.1 mrg * satisfy the constraint 2479 1.1 mrg * separate no (integer) point of the other basic map 2480 1.1 mrg * satisfies the constraint 2481 1.1 mrg * cut some but not all points of the other basic map 2482 1.1 mrg * satisfy the constraint 2483 1.1 mrg * adj_eq the given constraint is adjacent (on the outside) 2484 1.1 mrg * to an equality of the other basic map 2485 1.1 mrg * adj_ineq the given constraint is adjacent (on the outside) 2486 1.1 mrg * to an inequality of the other basic map 2487 1.1 mrg * 2488 1.1 mrg * We consider seven cases in which we can replace the pair by a single 2489 1.1 mrg * basic map. We ignore all "redundant" constraints. 2490 1.1 mrg * 2491 1.1 mrg * 1. all constraints of one basic map are valid 2492 1.1 mrg * => the other basic map is a subset and can be removed 2493 1.1 mrg * 2494 1.1 mrg * 2. all constraints of both basic maps are either "valid" or "cut" 2495 1.1 mrg * and the facets corresponding to the "cut" constraints 2496 1.1 mrg * of one of the basic maps lies entirely inside the other basic map 2497 1.1 mrg * => the pair can be replaced by a basic map consisting 2498 1.1 mrg * of the valid constraints in both basic maps 2499 1.1 mrg * 2500 1.1 mrg * 3. there is a single pair of adjacent inequalities 2501 1.1 mrg * (all other constraints are "valid") 2502 1.1 mrg * => the pair can be replaced by a basic map consisting 2503 1.1 mrg * of the valid constraints in both basic maps 2504 1.1 mrg * 2505 1.1 mrg * 4. one basic map has a single adjacent inequality, while the other 2506 1.1 mrg * constraints are "valid". The other basic map has some 2507 1.1 mrg * "cut" constraints, but replacing the adjacent inequality by 2508 1.1 mrg * its opposite and adding the valid constraints of the other 2509 1.1 mrg * basic map results in a subset of the other basic map 2510 1.1 mrg * => the pair can be replaced by a basic map consisting 2511 1.1 mrg * of the valid constraints in both basic maps 2512 1.1 mrg * 2513 1.1 mrg * 5. there is a single adjacent pair of an inequality and an equality, 2514 1.1 mrg * the other constraints of the basic map containing the inequality are 2515 1.1 mrg * "valid". Moreover, if the inequality the basic map is relaxed 2516 1.1 mrg * and then turned into an equality, then resulting facet lies 2517 1.1 mrg * entirely inside the other basic map 2518 1.1 mrg * => the pair can be replaced by the basic map containing 2519 1.1 mrg * the inequality, with the inequality relaxed. 2520 1.1 mrg * 2521 1.1 mrg * 6. there is a single inequality adjacent to an equality, 2522 1.1 mrg * the other constraints of the basic map containing the inequality are 2523 1.1 mrg * "valid". Moreover, the facets corresponding to both 2524 1.1 mrg * the inequality and the equality can be wrapped around their 2525 1.1 mrg * ridges to include the other basic map 2526 1.1 mrg * => the pair can be replaced by a basic map consisting 2527 1.1 mrg * of the valid constraints in both basic maps together 2528 1.1 mrg * with all wrapping constraints 2529 1.1 mrg * 2530 1.1 mrg * 7. one of the basic maps extends beyond the other by at most one. 2531 1.1 mrg * Moreover, the facets corresponding to the cut constraints and 2532 1.1 mrg * the pieces of the other basic map at offset one from these cut 2533 1.1 mrg * constraints can be wrapped around their ridges to include 2534 1.1 mrg * the union of the two basic maps 2535 1.1 mrg * => the pair can be replaced by a basic map consisting 2536 1.1 mrg * of the valid constraints in both basic maps together 2537 1.1 mrg * with all wrapping constraints 2538 1.1 mrg * 2539 1.1 mrg * 8. the two basic maps live in adjacent hyperplanes. In principle 2540 1.1 mrg * such sets can always be combined through wrapping, but we impose 2541 1.1 mrg * that there is only one such pair, to avoid overeager coalescing. 2542 1.1 mrg * 2543 1.1 mrg * Throughout the computation, we maintain a collection of tableaus 2544 1.1 mrg * corresponding to the basic maps. When the basic maps are dropped 2545 1.1 mrg * or combined, the tableaus are modified accordingly. 2546 1.1 mrg */ 2547 1.1 mrg static enum isl_change coalesce_local_pair_reuse(int i, int j, 2548 1.1 mrg struct isl_coalesce_info *info) 2549 1.1 mrg { 2550 1.1 mrg enum isl_change change = isl_change_none; 2551 1.1 mrg 2552 1.1 mrg set_ineq_status_in(&info[i], info[j].tab); 2553 1.1 mrg if (info[i].bmap->n_ineq && !info[i].ineq) 2554 1.1 mrg goto error; 2555 1.1 mrg if (any_ineq(&info[i], STATUS_ERROR)) 2556 1.1 mrg goto error; 2557 1.1 mrg if (any_ineq(&info[i], STATUS_SEPARATE)) 2558 1.1 mrg goto done; 2559 1.1 mrg 2560 1.1 mrg set_ineq_status_in(&info[j], info[i].tab); 2561 1.1 mrg if (info[j].bmap->n_ineq && !info[j].ineq) 2562 1.1 mrg goto error; 2563 1.1 mrg if (any_ineq(&info[j], STATUS_ERROR)) 2564 1.1 mrg goto error; 2565 1.1 mrg if (any_ineq(&info[j], STATUS_SEPARATE)) 2566 1.1 mrg goto done; 2567 1.1 mrg 2568 1.1 mrg set_eq_status_in(&info[i], info[j].tab); 2569 1.1 mrg if (info[i].bmap->n_eq && !info[i].eq) 2570 1.1 mrg goto error; 2571 1.1 mrg if (any_eq(&info[i], STATUS_ERROR)) 2572 1.1 mrg goto error; 2573 1.1 mrg 2574 1.1 mrg set_eq_status_in(&info[j], info[i].tab); 2575 1.1 mrg if (info[j].bmap->n_eq && !info[j].eq) 2576 1.1 mrg goto error; 2577 1.1 mrg if (any_eq(&info[j], STATUS_ERROR)) 2578 1.1 mrg goto error; 2579 1.1 mrg 2580 1.1 mrg if (any_eq(&info[i], STATUS_SEPARATE)) 2581 1.1 mrg return separating_equality(i, j, info); 2582 1.1 mrg if (any_eq(&info[j], STATUS_SEPARATE)) 2583 1.1 mrg return separating_equality(j, i, info); 2584 1.1 mrg 2585 1.1 mrg if (all(info[i].eq, 2 * info[i].bmap->n_eq, STATUS_VALID) && 2586 1.1 mrg all(info[i].ineq, info[i].bmap->n_ineq, STATUS_VALID)) { 2587 1.1 mrg drop(&info[j]); 2588 1.1 mrg change = isl_change_drop_second; 2589 1.1 mrg } else if (all(info[j].eq, 2 * info[j].bmap->n_eq, STATUS_VALID) && 2590 1.1 mrg all(info[j].ineq, info[j].bmap->n_ineq, STATUS_VALID)) { 2591 1.1 mrg drop(&info[i]); 2592 1.1 mrg change = isl_change_drop_first; 2593 1.1 mrg } else if (any_eq(&info[i], STATUS_ADJ_EQ)) { 2594 1.1 mrg change = check_eq_adj_eq(i, j, info); 2595 1.1 mrg } else if (any_eq(&info[j], STATUS_ADJ_EQ)) { 2596 1.1 mrg change = check_eq_adj_eq(j, i, info); 2597 1.1 mrg } else if (any_eq(&info[i], STATUS_ADJ_INEQ) || 2598 1.1 mrg any_eq(&info[j], STATUS_ADJ_INEQ)) { 2599 1.1 mrg change = check_adj_eq(i, j, info); 2600 1.1 mrg } else if (any_ineq(&info[i], STATUS_ADJ_EQ)) { 2601 1.1 mrg change = check_ineq_adj_eq(i, j, info); 2602 1.1 mrg } else if (any_ineq(&info[j], STATUS_ADJ_EQ)) { 2603 1.1 mrg change = check_ineq_adj_eq(j, i, info); 2604 1.1 mrg } else if (any_ineq(&info[i], STATUS_ADJ_INEQ) || 2605 1.1 mrg any_ineq(&info[j], STATUS_ADJ_INEQ)) { 2606 1.1 mrg change = check_adj_ineq(i, j, info); 2607 1.1 mrg } else { 2608 1.1 mrg if (!any_eq(&info[i], STATUS_CUT) && 2609 1.1 mrg !any_eq(&info[j], STATUS_CUT)) 2610 1.1 mrg change = check_facets(i, j, info); 2611 1.1 mrg if (change == isl_change_none) 2612 1.1 mrg change = check_wrap(i, j, info); 2613 1.1 mrg } 2614 1.1 mrg 2615 1.1 mrg done: 2616 1.1 mrg clear_status(&info[i]); 2617 1.1 mrg clear_status(&info[j]); 2618 1.1 mrg return change; 2619 1.1 mrg error: 2620 1.1 mrg clear_status(&info[i]); 2621 1.1 mrg clear_status(&info[j]); 2622 1.1 mrg return isl_change_error; 2623 1.1 mrg } 2624 1.1 mrg 2625 1.1 mrg /* Check if the union of the given pair of basic maps 2626 1.1 mrg * can be represented by a single basic map. 2627 1.1 mrg * If so, replace the pair by the single basic map and return 2628 1.1 mrg * isl_change_drop_first, isl_change_drop_second or isl_change_fuse. 2629 1.1 mrg * Otherwise, return isl_change_none. 2630 1.1 mrg * The two basic maps are assumed to live in the same local space. 2631 1.1 mrg */ 2632 1.1 mrg static enum isl_change coalesce_local_pair(int i, int j, 2633 1.1 mrg struct isl_coalesce_info *info) 2634 1.1 mrg { 2635 1.1 mrg init_status(&info[i]); 2636 1.1 mrg init_status(&info[j]); 2637 1.1 mrg return coalesce_local_pair_reuse(i, j, info); 2638 1.1 mrg } 2639 1.1 mrg 2640 1.1 mrg /* Shift the integer division at position "div" of the basic map 2641 1.1 mrg * represented by "info" by "shift". 2642 1.1 mrg * 2643 1.1 mrg * That is, if the integer division has the form 2644 1.1 mrg * 2645 1.1 mrg * floor(f(x)/d) 2646 1.1 mrg * 2647 1.1 mrg * then replace it by 2648 1.1 mrg * 2649 1.1 mrg * floor((f(x) + shift * d)/d) - shift 2650 1.1 mrg */ 2651 1.1 mrg static isl_stat shift_div(struct isl_coalesce_info *info, int div, 2652 1.1 mrg isl_int shift) 2653 1.1 mrg { 2654 1.1 mrg isl_size total, n_div; 2655 1.1 mrg 2656 1.1 mrg info->bmap = isl_basic_map_shift_div(info->bmap, div, 0, shift); 2657 1.1 mrg if (!info->bmap) 2658 1.1 mrg return isl_stat_error; 2659 1.1 mrg 2660 1.1 mrg total = isl_basic_map_dim(info->bmap, isl_dim_all); 2661 1.1 mrg n_div = isl_basic_map_dim(info->bmap, isl_dim_div); 2662 1.1 mrg if (total < 0 || n_div < 0) 2663 1.1 mrg return isl_stat_error; 2664 1.1 mrg total -= n_div; 2665 1.1 mrg if (isl_tab_shift_var(info->tab, total + div, shift) < 0) 2666 1.1 mrg return isl_stat_error; 2667 1.1 mrg 2668 1.1 mrg return isl_stat_ok; 2669 1.1 mrg } 2670 1.1 mrg 2671 1.1 mrg /* If the integer division at position "div" is defined by an equality, 2672 1.1 mrg * i.e., a stride constraint, then change the integer division expression 2673 1.1 mrg * to have a constant term equal to zero. 2674 1.1 mrg * 2675 1.1 mrg * Let the equality constraint be 2676 1.1 mrg * 2677 1.1 mrg * c + f + m a = 0 2678 1.1 mrg * 2679 1.1 mrg * The integer division expression is then typically of the form 2680 1.1 mrg * 2681 1.1 mrg * a = floor((-f - c')/m) 2682 1.1 mrg * 2683 1.1 mrg * The integer division is first shifted by t = floor(c/m), 2684 1.1 mrg * turning the equality constraint into 2685 1.1 mrg * 2686 1.1 mrg * c - m floor(c/m) + f + m a' = 0 2687 1.1 mrg * 2688 1.1 mrg * i.e., 2689 1.1 mrg * 2690 1.1 mrg * (c mod m) + f + m a' = 0 2691 1.1 mrg * 2692 1.1 mrg * That is, 2693 1.1 mrg * 2694 1.1 mrg * a' = (-f - (c mod m))/m = floor((-f)/m) 2695 1.1 mrg * 2696 1.1 mrg * because a' is an integer and 0 <= (c mod m) < m. 2697 1.1 mrg * The constant term of a' can therefore be zeroed out, 2698 1.1 mrg * but only if the integer division expression is of the expected form. 2699 1.1 mrg */ 2700 1.1 mrg static isl_stat normalize_stride_div(struct isl_coalesce_info *info, int div) 2701 1.1 mrg { 2702 1.1 mrg isl_bool defined, valid; 2703 1.1 mrg isl_stat r; 2704 1.1 mrg isl_constraint *c; 2705 1.1 mrg isl_int shift, stride; 2706 1.1 mrg 2707 1.1 mrg defined = isl_basic_map_has_defining_equality(info->bmap, isl_dim_div, 2708 1.1 mrg div, &c); 2709 1.1 mrg if (defined < 0) 2710 1.1 mrg return isl_stat_error; 2711 1.1 mrg if (!defined) 2712 1.1 mrg return isl_stat_ok; 2713 1.1 mrg if (!c) 2714 1.1 mrg return isl_stat_error; 2715 1.1 mrg valid = isl_constraint_is_div_equality(c, div); 2716 1.1 mrg isl_int_init(shift); 2717 1.1 mrg isl_int_init(stride); 2718 1.1 mrg isl_constraint_get_constant(c, &shift); 2719 1.1 mrg isl_constraint_get_coefficient(c, isl_dim_div, div, &stride); 2720 1.1 mrg isl_int_fdiv_q(shift, shift, stride); 2721 1.1 mrg r = shift_div(info, div, shift); 2722 1.1 mrg isl_int_clear(stride); 2723 1.1 mrg isl_int_clear(shift); 2724 1.1 mrg isl_constraint_free(c); 2725 1.1 mrg if (r < 0 || valid < 0) 2726 1.1 mrg return isl_stat_error; 2727 1.1 mrg if (!valid) 2728 1.1 mrg return isl_stat_ok; 2729 1.1 mrg info->bmap = isl_basic_map_set_div_expr_constant_num_si_inplace( 2730 1.1 mrg info->bmap, div, 0); 2731 1.1 mrg if (!info->bmap) 2732 1.1 mrg return isl_stat_error; 2733 1.1 mrg return isl_stat_ok; 2734 1.1 mrg } 2735 1.1 mrg 2736 1.1 mrg /* The basic maps represented by "info1" and "info2" are known 2737 1.1 mrg * to have the same number of integer divisions. 2738 1.1 mrg * Check if pairs of integer divisions are equal to each other 2739 1.1 mrg * despite the fact that they differ by a rational constant. 2740 1.1 mrg * 2741 1.1 mrg * In particular, look for any pair of integer divisions that 2742 1.1 mrg * only differ in their constant terms. 2743 1.1 mrg * If either of these integer divisions is defined 2744 1.1 mrg * by stride constraints, then modify it to have a zero constant term. 2745 1.1 mrg * If both are defined by stride constraints then in the end they will have 2746 1.1 mrg * the same (zero) constant term. 2747 1.1 mrg */ 2748 1.1 mrg static isl_stat harmonize_stride_divs(struct isl_coalesce_info *info1, 2749 1.1 mrg struct isl_coalesce_info *info2) 2750 1.1 mrg { 2751 1.1 mrg int i; 2752 1.1 mrg isl_size n; 2753 1.1 mrg 2754 1.1 mrg n = isl_basic_map_dim(info1->bmap, isl_dim_div); 2755 1.1 mrg if (n < 0) 2756 1.1 mrg return isl_stat_error; 2757 1.1 mrg for (i = 0; i < n; ++i) { 2758 1.1 mrg isl_bool known, harmonize; 2759 1.1 mrg 2760 1.1 mrg known = isl_basic_map_div_is_known(info1->bmap, i); 2761 1.1 mrg if (known >= 0 && known) 2762 1.1 mrg known = isl_basic_map_div_is_known(info2->bmap, i); 2763 1.1 mrg if (known < 0) 2764 1.1 mrg return isl_stat_error; 2765 1.1 mrg if (!known) 2766 1.1 mrg continue; 2767 1.1 mrg harmonize = isl_basic_map_equal_div_expr_except_constant( 2768 1.1 mrg info1->bmap, i, info2->bmap, i); 2769 1.1 mrg if (harmonize < 0) 2770 1.1 mrg return isl_stat_error; 2771 1.1 mrg if (!harmonize) 2772 1.1 mrg continue; 2773 1.1 mrg if (normalize_stride_div(info1, i) < 0) 2774 1.1 mrg return isl_stat_error; 2775 1.1 mrg if (normalize_stride_div(info2, i) < 0) 2776 1.1 mrg return isl_stat_error; 2777 1.1 mrg } 2778 1.1 mrg 2779 1.1 mrg return isl_stat_ok; 2780 1.1 mrg } 2781 1.1 mrg 2782 1.1 mrg /* If "shift" is an integer constant, then shift the integer division 2783 1.1 mrg * at position "div" of the basic map represented by "info" by "shift". 2784 1.1 mrg * If "shift" is not an integer constant, then do nothing. 2785 1.1 mrg * If "shift" is equal to zero, then no shift needs to be performed either. 2786 1.1 mrg * 2787 1.1 mrg * That is, if the integer division has the form 2788 1.1 mrg * 2789 1.1 mrg * floor(f(x)/d) 2790 1.1 mrg * 2791 1.1 mrg * then replace it by 2792 1.1 mrg * 2793 1.1 mrg * floor((f(x) + shift * d)/d) - shift 2794 1.1 mrg */ 2795 1.1 mrg static isl_stat shift_if_cst_int(struct isl_coalesce_info *info, int div, 2796 1.1 mrg __isl_keep isl_aff *shift) 2797 1.1 mrg { 2798 1.1 mrg isl_bool cst; 2799 1.1 mrg isl_stat r; 2800 1.1 mrg isl_int d; 2801 1.1 mrg isl_val *c; 2802 1.1 mrg 2803 1.1 mrg cst = isl_aff_is_cst(shift); 2804 1.1 mrg if (cst < 0 || !cst) 2805 1.1 mrg return cst < 0 ? isl_stat_error : isl_stat_ok; 2806 1.1 mrg 2807 1.1 mrg c = isl_aff_get_constant_val(shift); 2808 1.1 mrg cst = isl_val_is_int(c); 2809 1.1 mrg if (cst >= 0 && cst) 2810 1.1 mrg cst = isl_bool_not(isl_val_is_zero(c)); 2811 1.1 mrg if (cst < 0 || !cst) { 2812 1.1 mrg isl_val_free(c); 2813 1.1 mrg return cst < 0 ? isl_stat_error : isl_stat_ok; 2814 1.1 mrg } 2815 1.1 mrg 2816 1.1 mrg isl_int_init(d); 2817 1.1 mrg r = isl_val_get_num_isl_int(c, &d); 2818 1.1 mrg if (r >= 0) 2819 1.1 mrg r = shift_div(info, div, d); 2820 1.1 mrg isl_int_clear(d); 2821 1.1 mrg 2822 1.1 mrg isl_val_free(c); 2823 1.1 mrg 2824 1.1 mrg return r; 2825 1.1 mrg } 2826 1.1 mrg 2827 1.1 mrg /* Check if some of the divs in the basic map represented by "info1" 2828 1.1 mrg * are shifts of the corresponding divs in the basic map represented 2829 1.1 mrg * by "info2", taking into account the equality constraints "eq1" of "info1" 2830 1.1 mrg * and "eq2" of "info2". If so, align them with those of "info2". 2831 1.1 mrg * "info1" and "info2" are assumed to have the same number 2832 1.1 mrg * of integer divisions. 2833 1.1 mrg * 2834 1.1 mrg * An integer division is considered to be a shift of another integer 2835 1.1 mrg * division if, after simplification with respect to the equality 2836 1.1 mrg * constraints of the other basic map, one is equal to the other 2837 1.1 mrg * plus a constant. 2838 1.1 mrg * 2839 1.1 mrg * In particular, for each pair of integer divisions, if both are known, 2840 1.1 mrg * have the same denominator and are not already equal to each other, 2841 1.1 mrg * simplify each with respect to the equality constraints 2842 1.1 mrg * of the other basic map. If the difference is an integer constant, 2843 1.1 mrg * then move this difference outside. 2844 1.1 mrg * That is, if, after simplification, one integer division is of the form 2845 1.1 mrg * 2846 1.1 mrg * floor((f(x) + c_1)/d) 2847 1.1 mrg * 2848 1.1 mrg * while the other is of the form 2849 1.1 mrg * 2850 1.1 mrg * floor((f(x) + c_2)/d) 2851 1.1 mrg * 2852 1.1 mrg * and n = (c_2 - c_1)/d is an integer, then replace the first 2853 1.1 mrg * integer division by 2854 1.1 mrg * 2855 1.1 mrg * floor((f_1(x) + c_1 + n * d)/d) - n, 2856 1.1 mrg * 2857 1.1 mrg * where floor((f_1(x) + c_1 + n * d)/d) = floor((f2(x) + c_2)/d) 2858 1.1 mrg * after simplification with respect to the equality constraints. 2859 1.1 mrg */ 2860 1.1 mrg static isl_stat harmonize_divs_with_hulls(struct isl_coalesce_info *info1, 2861 1.1 mrg struct isl_coalesce_info *info2, __isl_keep isl_basic_set *eq1, 2862 1.1 mrg __isl_keep isl_basic_set *eq2) 2863 1.1 mrg { 2864 1.1 mrg int i; 2865 1.1 mrg isl_size total; 2866 1.1 mrg isl_local_space *ls1, *ls2; 2867 1.1 mrg 2868 1.1 mrg total = isl_basic_map_dim(info1->bmap, isl_dim_all); 2869 1.1 mrg if (total < 0) 2870 1.1 mrg return isl_stat_error; 2871 1.1 mrg ls1 = isl_local_space_wrap(isl_basic_map_get_local_space(info1->bmap)); 2872 1.1 mrg ls2 = isl_local_space_wrap(isl_basic_map_get_local_space(info2->bmap)); 2873 1.1 mrg for (i = 0; i < info1->bmap->n_div; ++i) { 2874 1.1 mrg isl_stat r; 2875 1.1 mrg isl_aff *div1, *div2; 2876 1.1 mrg 2877 1.1 mrg if (!isl_local_space_div_is_known(ls1, i) || 2878 1.1 mrg !isl_local_space_div_is_known(ls2, i)) 2879 1.1 mrg continue; 2880 1.1 mrg if (isl_int_ne(info1->bmap->div[i][0], info2->bmap->div[i][0])) 2881 1.1 mrg continue; 2882 1.1 mrg if (isl_seq_eq(info1->bmap->div[i] + 1, 2883 1.1 mrg info2->bmap->div[i] + 1, 1 + total)) 2884 1.1 mrg continue; 2885 1.1 mrg div1 = isl_local_space_get_div(ls1, i); 2886 1.1 mrg div2 = isl_local_space_get_div(ls2, i); 2887 1.1 mrg div1 = isl_aff_substitute_equalities(div1, 2888 1.1 mrg isl_basic_set_copy(eq2)); 2889 1.1 mrg div2 = isl_aff_substitute_equalities(div2, 2890 1.1 mrg isl_basic_set_copy(eq1)); 2891 1.1 mrg div2 = isl_aff_sub(div2, div1); 2892 1.1 mrg r = shift_if_cst_int(info1, i, div2); 2893 1.1 mrg isl_aff_free(div2); 2894 1.1 mrg if (r < 0) 2895 1.1 mrg break; 2896 1.1 mrg } 2897 1.1 mrg isl_local_space_free(ls1); 2898 1.1 mrg isl_local_space_free(ls2); 2899 1.1 mrg 2900 1.1 mrg if (i < info1->bmap->n_div) 2901 1.1 mrg return isl_stat_error; 2902 1.1 mrg return isl_stat_ok; 2903 1.1 mrg } 2904 1.1 mrg 2905 1.1 mrg /* Check if some of the divs in the basic map represented by "info1" 2906 1.1 mrg * are shifts of the corresponding divs in the basic map represented 2907 1.1 mrg * by "info2". If so, align them with those of "info2". 2908 1.1 mrg * Only do this if "info1" and "info2" have the same number 2909 1.1 mrg * of integer divisions. 2910 1.1 mrg * 2911 1.1 mrg * An integer division is considered to be a shift of another integer 2912 1.1 mrg * division if, after simplification with respect to the equality 2913 1.1 mrg * constraints of the other basic map, one is equal to the other 2914 1.1 mrg * plus a constant. 2915 1.1 mrg * 2916 1.1 mrg * First check if pairs of integer divisions are equal to each other 2917 1.1 mrg * despite the fact that they differ by a rational constant. 2918 1.1 mrg * If so, try and arrange for them to have the same constant term. 2919 1.1 mrg * 2920 1.1 mrg * Then, extract the equality constraints and continue with 2921 1.1 mrg * harmonize_divs_with_hulls. 2922 1.1 mrg * 2923 1.1 mrg * If the equality constraints of both basic maps are the same, 2924 1.1 mrg * then there is no need to perform any shifting since 2925 1.1 mrg * the coefficients of the integer divisions should have been 2926 1.1 mrg * reduced in the same way. 2927 1.1 mrg */ 2928 1.1 mrg static isl_stat harmonize_divs(struct isl_coalesce_info *info1, 2929 1.1 mrg struct isl_coalesce_info *info2) 2930 1.1 mrg { 2931 1.1 mrg isl_bool equal; 2932 1.1 mrg isl_basic_map *bmap1, *bmap2; 2933 1.1 mrg isl_basic_set *eq1, *eq2; 2934 1.1 mrg isl_stat r; 2935 1.1 mrg 2936 1.1 mrg if (!info1->bmap || !info2->bmap) 2937 1.1 mrg return isl_stat_error; 2938 1.1 mrg 2939 1.1 mrg if (info1->bmap->n_div != info2->bmap->n_div) 2940 1.1 mrg return isl_stat_ok; 2941 1.1 mrg if (info1->bmap->n_div == 0) 2942 1.1 mrg return isl_stat_ok; 2943 1.1 mrg 2944 1.1 mrg if (harmonize_stride_divs(info1, info2) < 0) 2945 1.1 mrg return isl_stat_error; 2946 1.1 mrg 2947 1.1 mrg bmap1 = isl_basic_map_copy(info1->bmap); 2948 1.1 mrg bmap2 = isl_basic_map_copy(info2->bmap); 2949 1.1 mrg eq1 = isl_basic_map_wrap(isl_basic_map_plain_affine_hull(bmap1)); 2950 1.1 mrg eq2 = isl_basic_map_wrap(isl_basic_map_plain_affine_hull(bmap2)); 2951 1.1 mrg equal = isl_basic_set_plain_is_equal(eq1, eq2); 2952 1.1 mrg if (equal < 0) 2953 1.1 mrg r = isl_stat_error; 2954 1.1 mrg else if (equal) 2955 1.1 mrg r = isl_stat_ok; 2956 1.1 mrg else 2957 1.1 mrg r = harmonize_divs_with_hulls(info1, info2, eq1, eq2); 2958 1.1 mrg isl_basic_set_free(eq1); 2959 1.1 mrg isl_basic_set_free(eq2); 2960 1.1 mrg 2961 1.1 mrg return r; 2962 1.1 mrg } 2963 1.1 mrg 2964 1.1 mrg /* Do the two basic maps live in the same local space, i.e., 2965 1.1 mrg * do they have the same (known) divs? 2966 1.1 mrg * If either basic map has any unknown divs, then we can only assume 2967 1.1 mrg * that they do not live in the same local space. 2968 1.1 mrg */ 2969 1.1 mrg static isl_bool same_divs(__isl_keep isl_basic_map *bmap1, 2970 1.1 mrg __isl_keep isl_basic_map *bmap2) 2971 1.1 mrg { 2972 1.1 mrg int i; 2973 1.1 mrg isl_bool known; 2974 1.1 mrg isl_size total; 2975 1.1 mrg 2976 1.1 mrg if (!bmap1 || !bmap2) 2977 1.1 mrg return isl_bool_error; 2978 1.1 mrg if (bmap1->n_div != bmap2->n_div) 2979 1.1 mrg return isl_bool_false; 2980 1.1 mrg 2981 1.1 mrg if (bmap1->n_div == 0) 2982 1.1 mrg return isl_bool_true; 2983 1.1 mrg 2984 1.1 mrg known = isl_basic_map_divs_known(bmap1); 2985 1.1 mrg if (known < 0 || !known) 2986 1.1 mrg return known; 2987 1.1 mrg known = isl_basic_map_divs_known(bmap2); 2988 1.1 mrg if (known < 0 || !known) 2989 1.1 mrg return known; 2990 1.1 mrg 2991 1.1 mrg total = isl_basic_map_dim(bmap1, isl_dim_all); 2992 1.1 mrg if (total < 0) 2993 1.1 mrg return isl_bool_error; 2994 1.1 mrg for (i = 0; i < bmap1->n_div; ++i) 2995 1.1 mrg if (!isl_seq_eq(bmap1->div[i], bmap2->div[i], 2 + total)) 2996 1.1 mrg return isl_bool_false; 2997 1.1 mrg 2998 1.1 mrg return isl_bool_true; 2999 1.1 mrg } 3000 1.1 mrg 3001 1.1 mrg /* Assuming that "tab" contains the equality constraints and 3002 1.1 mrg * the initial inequality constraints of "bmap", copy the remaining 3003 1.1 mrg * inequality constraints of "bmap" to "Tab". 3004 1.1 mrg */ 3005 1.1 mrg static isl_stat copy_ineq(struct isl_tab *tab, __isl_keep isl_basic_map *bmap) 3006 1.1 mrg { 3007 1.1 mrg int i, n_ineq; 3008 1.1 mrg 3009 1.1 mrg if (!bmap) 3010 1.1 mrg return isl_stat_error; 3011 1.1 mrg 3012 1.1 mrg n_ineq = tab->n_con - tab->n_eq; 3013 1.1 mrg for (i = n_ineq; i < bmap->n_ineq; ++i) 3014 1.1 mrg if (isl_tab_add_ineq(tab, bmap->ineq[i]) < 0) 3015 1.1 mrg return isl_stat_error; 3016 1.1 mrg 3017 1.1 mrg return isl_stat_ok; 3018 1.1 mrg } 3019 1.1 mrg 3020 1.1 mrg /* Description of an integer division that is added 3021 1.1 mrg * during an expansion. 3022 1.1 mrg * "pos" is the position of the corresponding variable. 3023 1.1 mrg * "cst" indicates whether this integer division has a fixed value. 3024 1.1 mrg * "val" contains the fixed value, if the value is fixed. 3025 1.1 mrg */ 3026 1.1 mrg struct isl_expanded { 3027 1.1 mrg int pos; 3028 1.1 mrg isl_bool cst; 3029 1.1 mrg isl_int val; 3030 1.1 mrg }; 3031 1.1 mrg 3032 1.1 mrg /* For each of the "n" integer division variables "expanded", 3033 1.1 mrg * if the variable has a fixed value, then add two inequality 3034 1.1 mrg * constraints expressing the fixed value. 3035 1.1 mrg * Otherwise, add the corresponding div constraints. 3036 1.1 mrg * The caller is responsible for removing the div constraints 3037 1.1 mrg * that it added for all these "n" integer divisions. 3038 1.1 mrg * 3039 1.1 mrg * The div constraints and the pair of inequality constraints 3040 1.1 mrg * forcing the fixed value cannot both be added for a given variable 3041 1.1 mrg * as the combination may render some of the original constraints redundant. 3042 1.1 mrg * These would then be ignored during the coalescing detection, 3043 1.1 mrg * while they could remain in the fused result. 3044 1.1 mrg * 3045 1.1 mrg * The two added inequality constraints are 3046 1.1 mrg * 3047 1.1 mrg * -a + v >= 0 3048 1.1 mrg * a - v >= 0 3049 1.1 mrg * 3050 1.1 mrg * with "a" the variable and "v" its fixed value. 3051 1.1 mrg * The facet corresponding to one of these two constraints is selected 3052 1.1 mrg * in the tableau to ensure that the pair of inequality constraints 3053 1.1 mrg * is treated as an equality constraint. 3054 1.1 mrg * 3055 1.1 mrg * The information in info->ineq is thrown away because it was 3056 1.1 mrg * computed in terms of div constraints, while some of those 3057 1.1 mrg * have now been replaced by these pairs of inequality constraints. 3058 1.1 mrg */ 3059 1.1 mrg static isl_stat fix_constant_divs(struct isl_coalesce_info *info, 3060 1.1 mrg int n, struct isl_expanded *expanded) 3061 1.1 mrg { 3062 1.1 mrg unsigned o_div; 3063 1.1 mrg int i; 3064 1.1 mrg isl_vec *ineq; 3065 1.1 mrg 3066 1.1 mrg o_div = isl_basic_map_offset(info->bmap, isl_dim_div) - 1; 3067 1.1 mrg ineq = isl_vec_alloc(isl_tab_get_ctx(info->tab), 1 + info->tab->n_var); 3068 1.1 mrg if (!ineq) 3069 1.1 mrg return isl_stat_error; 3070 1.1 mrg isl_seq_clr(ineq->el + 1, info->tab->n_var); 3071 1.1 mrg 3072 1.1 mrg for (i = 0; i < n; ++i) { 3073 1.1 mrg if (!expanded[i].cst) { 3074 1.1 mrg info->bmap = isl_basic_map_extend_constraints( 3075 1.1 mrg info->bmap, 0, 2); 3076 1.1 mrg info->bmap = isl_basic_map_add_div_constraints( 3077 1.1 mrg info->bmap, expanded[i].pos - o_div); 3078 1.1 mrg } else { 3079 1.1 mrg isl_int_set_si(ineq->el[1 + expanded[i].pos], -1); 3080 1.1 mrg isl_int_set(ineq->el[0], expanded[i].val); 3081 1.1 mrg info->bmap = isl_basic_map_add_ineq(info->bmap, 3082 1.1 mrg ineq->el); 3083 1.1 mrg isl_int_set_si(ineq->el[1 + expanded[i].pos], 1); 3084 1.1 mrg isl_int_neg(ineq->el[0], expanded[i].val); 3085 1.1 mrg info->bmap = isl_basic_map_add_ineq(info->bmap, 3086 1.1 mrg ineq->el); 3087 1.1 mrg isl_int_set_si(ineq->el[1 + expanded[i].pos], 0); 3088 1.1 mrg } 3089 1.1 mrg if (copy_ineq(info->tab, info->bmap) < 0) 3090 1.1 mrg break; 3091 1.1 mrg if (expanded[i].cst && 3092 1.1 mrg isl_tab_select_facet(info->tab, info->tab->n_con - 1) < 0) 3093 1.1 mrg break; 3094 1.1 mrg } 3095 1.1 mrg 3096 1.1 mrg isl_vec_free(ineq); 3097 1.1 mrg 3098 1.1 mrg clear_status(info); 3099 1.1 mrg init_status(info); 3100 1.1 mrg 3101 1.1 mrg return i < n ? isl_stat_error : isl_stat_ok; 3102 1.1 mrg } 3103 1.1 mrg 3104 1.1 mrg /* Insert the "n" integer division variables "expanded" 3105 1.1 mrg * into info->tab and info->bmap and 3106 1.1 mrg * update info->ineq with respect to the redundant constraints 3107 1.1 mrg * in the resulting tableau. 3108 1.1 mrg * "bmap" contains the result of this insertion in info->bmap, 3109 1.1 mrg * while info->bmap is the original version 3110 1.1 mrg * of "bmap", i.e., the one that corresponds to the current 3111 1.1 mrg * state of info->tab. The number of constraints in info->bmap 3112 1.1 mrg * is assumed to be the same as the number of constraints 3113 1.1 mrg * in info->tab. This is required to be able to detect 3114 1.1 mrg * the extra constraints in "bmap". 3115 1.1 mrg * 3116 1.1 mrg * In particular, introduce extra variables corresponding 3117 1.1 mrg * to the extra integer divisions and add the div constraints 3118 1.1 mrg * that were added to "bmap" after info->tab was created 3119 1.1 mrg * from info->bmap. 3120 1.1 mrg * Furthermore, check if these extra integer divisions happen 3121 1.1 mrg * to attain a fixed integer value in info->tab. 3122 1.1 mrg * If so, replace the corresponding div constraints by pairs 3123 1.1 mrg * of inequality constraints that fix these 3124 1.1 mrg * integer divisions to their single integer values. 3125 1.1 mrg * Replace info->bmap by "bmap" to match the changes to info->tab. 3126 1.1 mrg * info->ineq was computed without a tableau and therefore 3127 1.1 mrg * does not take into account the redundant constraints 3128 1.1 mrg * in the tableau. Mark them here. 3129 1.1 mrg * There is no need to check the newly added div constraints 3130 1.1 mrg * since they cannot be redundant. 3131 1.1 mrg * The redundancy check is not performed when constants have been discovered 3132 1.1 mrg * since info->ineq is completely thrown away in this case. 3133 1.1 mrg */ 3134 1.1 mrg static isl_stat tab_insert_divs(struct isl_coalesce_info *info, 3135 1.1 mrg int n, struct isl_expanded *expanded, __isl_take isl_basic_map *bmap) 3136 1.1 mrg { 3137 1.1 mrg int i, n_ineq; 3138 1.1 mrg unsigned n_eq; 3139 1.1 mrg struct isl_tab_undo *snap; 3140 1.1 mrg int any; 3141 1.1 mrg 3142 1.1 mrg if (!bmap) 3143 1.1 mrg return isl_stat_error; 3144 1.1 mrg if (info->bmap->n_eq + info->bmap->n_ineq != info->tab->n_con) 3145 1.1 mrg isl_die(isl_basic_map_get_ctx(bmap), isl_error_internal, 3146 1.1 mrg "original tableau does not correspond " 3147 1.1 mrg "to original basic map", goto error); 3148 1.1 mrg 3149 1.1 mrg if (isl_tab_extend_vars(info->tab, n) < 0) 3150 1.1 mrg goto error; 3151 1.1 mrg if (isl_tab_extend_cons(info->tab, 2 * n) < 0) 3152 1.1 mrg goto error; 3153 1.1 mrg 3154 1.1 mrg for (i = 0; i < n; ++i) { 3155 1.1 mrg if (isl_tab_insert_var(info->tab, expanded[i].pos) < 0) 3156 1.1 mrg goto error; 3157 1.1 mrg } 3158 1.1 mrg 3159 1.1 mrg snap = isl_tab_snap(info->tab); 3160 1.1 mrg 3161 1.1 mrg n_ineq = info->tab->n_con - info->tab->n_eq; 3162 1.1 mrg if (copy_ineq(info->tab, bmap) < 0) 3163 1.1 mrg goto error; 3164 1.1 mrg 3165 1.1 mrg isl_basic_map_free(info->bmap); 3166 1.1 mrg info->bmap = bmap; 3167 1.1 mrg 3168 1.1 mrg any = 0; 3169 1.1 mrg for (i = 0; i < n; ++i) { 3170 1.1 mrg expanded[i].cst = isl_tab_is_constant(info->tab, 3171 1.1 mrg expanded[i].pos, &expanded[i].val); 3172 1.1 mrg if (expanded[i].cst < 0) 3173 1.1 mrg return isl_stat_error; 3174 1.1 mrg if (expanded[i].cst) 3175 1.1 mrg any = 1; 3176 1.1 mrg } 3177 1.1 mrg 3178 1.1 mrg if (any) { 3179 1.1 mrg if (isl_tab_rollback(info->tab, snap) < 0) 3180 1.1 mrg return isl_stat_error; 3181 1.1 mrg info->bmap = isl_basic_map_cow(info->bmap); 3182 1.1 mrg info->bmap = isl_basic_map_free_inequality(info->bmap, 2 * n); 3183 1.1 mrg if (!info->bmap) 3184 1.1 mrg return isl_stat_error; 3185 1.1 mrg 3186 1.1 mrg return fix_constant_divs(info, n, expanded); 3187 1.1 mrg } 3188 1.1 mrg 3189 1.1 mrg n_eq = info->bmap->n_eq; 3190 1.1 mrg for (i = 0; i < n_ineq; ++i) { 3191 1.1 mrg if (isl_tab_is_redundant(info->tab, n_eq + i)) 3192 1.1 mrg info->ineq[i] = STATUS_REDUNDANT; 3193 1.1 mrg } 3194 1.1 mrg 3195 1.1 mrg return isl_stat_ok; 3196 1.1 mrg error: 3197 1.1 mrg isl_basic_map_free(bmap); 3198 1.1 mrg return isl_stat_error; 3199 1.1 mrg } 3200 1.1 mrg 3201 1.1 mrg /* Expand info->tab and info->bmap in the same way "bmap" was expanded 3202 1.1 mrg * in isl_basic_map_expand_divs using the expansion "exp" and 3203 1.1 mrg * update info->ineq with respect to the redundant constraints 3204 1.1 mrg * in the resulting tableau. info->bmap is the original version 3205 1.1 mrg * of "bmap", i.e., the one that corresponds to the current 3206 1.1 mrg * state of info->tab. The number of constraints in info->bmap 3207 1.1 mrg * is assumed to be the same as the number of constraints 3208 1.1 mrg * in info->tab. This is required to be able to detect 3209 1.1 mrg * the extra constraints in "bmap". 3210 1.1 mrg * 3211 1.1 mrg * Extract the positions where extra local variables are introduced 3212 1.1 mrg * from "exp" and call tab_insert_divs. 3213 1.1 mrg */ 3214 1.1 mrg static isl_stat expand_tab(struct isl_coalesce_info *info, int *exp, 3215 1.1 mrg __isl_take isl_basic_map *bmap) 3216 1.1 mrg { 3217 1.1 mrg isl_ctx *ctx; 3218 1.1 mrg struct isl_expanded *expanded; 3219 1.1 mrg int i, j, k, n; 3220 1.1 mrg int extra_var; 3221 1.1 mrg isl_size total, n_div; 3222 1.1 mrg unsigned pos; 3223 1.1 mrg isl_stat r; 3224 1.1 mrg 3225 1.1 mrg total = isl_basic_map_dim(bmap, isl_dim_all); 3226 1.1 mrg n_div = isl_basic_map_dim(bmap, isl_dim_div); 3227 1.1 mrg if (total < 0 || n_div < 0) 3228 1.1 mrg return isl_stat_error; 3229 1.1 mrg pos = total - n_div; 3230 1.1 mrg extra_var = total - info->tab->n_var; 3231 1.1 mrg n = n_div - extra_var; 3232 1.1 mrg 3233 1.1 mrg ctx = isl_basic_map_get_ctx(bmap); 3234 1.1 mrg expanded = isl_calloc_array(ctx, struct isl_expanded, extra_var); 3235 1.1 mrg if (extra_var && !expanded) 3236 1.1 mrg goto error; 3237 1.1 mrg 3238 1.1 mrg i = 0; 3239 1.1 mrg k = 0; 3240 1.1 mrg for (j = 0; j < n_div; ++j) { 3241 1.1 mrg if (i < n && exp[i] == j) { 3242 1.1 mrg ++i; 3243 1.1 mrg continue; 3244 1.1 mrg } 3245 1.1 mrg expanded[k++].pos = pos + j; 3246 1.1 mrg } 3247 1.1 mrg 3248 1.1 mrg for (k = 0; k < extra_var; ++k) 3249 1.1 mrg isl_int_init(expanded[k].val); 3250 1.1 mrg 3251 1.1 mrg r = tab_insert_divs(info, extra_var, expanded, bmap); 3252 1.1 mrg 3253 1.1 mrg for (k = 0; k < extra_var; ++k) 3254 1.1 mrg isl_int_clear(expanded[k].val); 3255 1.1 mrg free(expanded); 3256 1.1 mrg 3257 1.1 mrg return r; 3258 1.1 mrg error: 3259 1.1 mrg isl_basic_map_free(bmap); 3260 1.1 mrg return isl_stat_error; 3261 1.1 mrg } 3262 1.1 mrg 3263 1.1 mrg /* Check if the union of the basic maps represented by info[i] and info[j] 3264 1.1 mrg * can be represented by a single basic map, 3265 1.1 mrg * after expanding the divs of info[i] to match those of info[j]. 3266 1.1 mrg * If so, replace the pair by the single basic map and return 3267 1.1 mrg * isl_change_drop_first, isl_change_drop_second or isl_change_fuse. 3268 1.1 mrg * Otherwise, return isl_change_none. 3269 1.1 mrg * 3270 1.1 mrg * The caller has already checked for info[j] being a subset of info[i]. 3271 1.1 mrg * If some of the divs of info[j] are unknown, then the expanded info[i] 3272 1.1 mrg * will not have the corresponding div constraints. The other patterns 3273 1.1 mrg * therefore cannot apply. Skip the computation in this case. 3274 1.1 mrg * 3275 1.1 mrg * The expansion is performed using the divs "div" and expansion "exp" 3276 1.1 mrg * computed by the caller. 3277 1.1 mrg * info[i].bmap has already been expanded and the result is passed in 3278 1.1 mrg * as "bmap". 3279 1.1 mrg * The "eq" and "ineq" fields of info[i] reflect the status of 3280 1.1 mrg * the constraints of the expanded "bmap" with respect to info[j].tab. 3281 1.1 mrg * However, inequality constraints that are redundant in info[i].tab 3282 1.1 mrg * have not yet been marked as such because no tableau was available. 3283 1.1 mrg * 3284 1.1 mrg * Replace info[i].bmap by "bmap" and expand info[i].tab as well, 3285 1.1 mrg * updating info[i].ineq with respect to the redundant constraints. 3286 1.1 mrg * Then try and coalesce the expanded info[i] with info[j], 3287 1.1 mrg * reusing the information in info[i].eq and info[i].ineq. 3288 1.1 mrg * If this does not result in any coalescing or if it results in info[j] 3289 1.1 mrg * getting dropped (which should not happen in practice, since the case 3290 1.1 mrg * of info[j] being a subset of info[i] has already been checked by 3291 1.1 mrg * the caller), then revert info[i] to its original state. 3292 1.1 mrg */ 3293 1.1 mrg static enum isl_change coalesce_expand_tab_divs(__isl_take isl_basic_map *bmap, 3294 1.1 mrg int i, int j, struct isl_coalesce_info *info, __isl_keep isl_mat *div, 3295 1.1 mrg int *exp) 3296 1.1 mrg { 3297 1.1 mrg isl_bool known; 3298 1.1 mrg isl_basic_map *bmap_i; 3299 1.1 mrg struct isl_tab_undo *snap; 3300 1.1 mrg enum isl_change change = isl_change_none; 3301 1.1 mrg 3302 1.1 mrg known = isl_basic_map_divs_known(info[j].bmap); 3303 1.1 mrg if (known < 0 || !known) { 3304 1.1 mrg clear_status(&info[i]); 3305 1.1 mrg isl_basic_map_free(bmap); 3306 1.1 mrg return known < 0 ? isl_change_error : isl_change_none; 3307 1.1 mrg } 3308 1.1 mrg 3309 1.1 mrg bmap_i = isl_basic_map_copy(info[i].bmap); 3310 1.1 mrg snap = isl_tab_snap(info[i].tab); 3311 1.1 mrg if (expand_tab(&info[i], exp, bmap) < 0) 3312 1.1 mrg change = isl_change_error; 3313 1.1 mrg 3314 1.1 mrg init_status(&info[j]); 3315 1.1 mrg if (change == isl_change_none) 3316 1.1 mrg change = coalesce_local_pair_reuse(i, j, info); 3317 1.1 mrg else 3318 1.1 mrg clear_status(&info[i]); 3319 1.1 mrg if (change != isl_change_none && change != isl_change_drop_second) { 3320 1.1 mrg isl_basic_map_free(bmap_i); 3321 1.1 mrg } else { 3322 1.1 mrg isl_basic_map_free(info[i].bmap); 3323 1.1 mrg info[i].bmap = bmap_i; 3324 1.1 mrg 3325 1.1 mrg if (isl_tab_rollback(info[i].tab, snap) < 0) 3326 1.1 mrg change = isl_change_error; 3327 1.1 mrg } 3328 1.1 mrg 3329 1.1 mrg return change; 3330 1.1 mrg } 3331 1.1 mrg 3332 1.1 mrg /* Check if the union of "bmap" and the basic map represented by info[j] 3333 1.1 mrg * can be represented by a single basic map, 3334 1.1 mrg * after expanding the divs of "bmap" to match those of info[j]. 3335 1.1 mrg * If so, replace the pair by the single basic map and return 3336 1.1 mrg * isl_change_drop_first, isl_change_drop_second or isl_change_fuse. 3337 1.1 mrg * Otherwise, return isl_change_none. 3338 1.1 mrg * 3339 1.1 mrg * In particular, check if the expanded "bmap" contains the basic map 3340 1.1 mrg * represented by the tableau info[j].tab. 3341 1.1 mrg * The expansion is performed using the divs "div" and expansion "exp" 3342 1.1 mrg * computed by the caller. 3343 1.1 mrg * Then we check if all constraints of the expanded "bmap" are valid for 3344 1.1 mrg * info[j].tab. 3345 1.1 mrg * 3346 1.1 mrg * If "i" is not equal to -1, then "bmap" is equal to info[i].bmap. 3347 1.1 mrg * In this case, the positions of the constraints of info[i].bmap 3348 1.1 mrg * with respect to the basic map represented by info[j] are stored 3349 1.1 mrg * in info[i]. 3350 1.1 mrg * 3351 1.1 mrg * If the expanded "bmap" does not contain the basic map 3352 1.1 mrg * represented by the tableau info[j].tab and if "i" is not -1, 3353 1.1 mrg * i.e., if the original "bmap" is info[i].bmap, then expand info[i].tab 3354 1.1 mrg * as well and check if that results in coalescing. 3355 1.1 mrg */ 3356 1.1 mrg static enum isl_change coalesce_with_expanded_divs( 3357 1.1 mrg __isl_keep isl_basic_map *bmap, int i, int j, 3358 1.1 mrg struct isl_coalesce_info *info, __isl_keep isl_mat *div, int *exp) 3359 1.1 mrg { 3360 1.1 mrg enum isl_change change = isl_change_none; 3361 1.1 mrg struct isl_coalesce_info info_local, *info_i; 3362 1.1 mrg 3363 1.1 mrg info_i = i >= 0 ? &info[i] : &info_local; 3364 1.1 mrg init_status(info_i); 3365 1.1 mrg bmap = isl_basic_map_copy(bmap); 3366 1.1 mrg bmap = isl_basic_map_expand_divs(bmap, isl_mat_copy(div), exp); 3367 1.1 mrg bmap = isl_basic_map_mark_final(bmap); 3368 1.1 mrg 3369 1.1 mrg if (!bmap) 3370 1.1 mrg goto error; 3371 1.1 mrg 3372 1.1 mrg info_local.bmap = bmap; 3373 1.1 mrg info_i->eq = eq_status_in(bmap, info[j].tab); 3374 1.1 mrg if (bmap->n_eq && !info_i->eq) 3375 1.1 mrg goto error; 3376 1.1 mrg if (any_eq(info_i, STATUS_ERROR)) 3377 1.1 mrg goto error; 3378 1.1 mrg if (any_eq(info_i, STATUS_SEPARATE)) 3379 1.1 mrg goto done; 3380 1.1 mrg 3381 1.1 mrg info_i->ineq = ineq_status_in(bmap, NULL, info[j].tab); 3382 1.1 mrg if (bmap->n_ineq && !info_i->ineq) 3383 1.1 mrg goto error; 3384 1.1 mrg if (any_ineq(info_i, STATUS_ERROR)) 3385 1.1 mrg goto error; 3386 1.1 mrg if (any_ineq(info_i, STATUS_SEPARATE)) 3387 1.1 mrg goto done; 3388 1.1 mrg 3389 1.1 mrg if (all(info_i->eq, 2 * bmap->n_eq, STATUS_VALID) && 3390 1.1 mrg all(info_i->ineq, bmap->n_ineq, STATUS_VALID)) { 3391 1.1 mrg drop(&info[j]); 3392 1.1 mrg change = isl_change_drop_second; 3393 1.1 mrg } 3394 1.1 mrg 3395 1.1 mrg if (change == isl_change_none && i != -1) 3396 1.1 mrg return coalesce_expand_tab_divs(bmap, i, j, info, div, exp); 3397 1.1 mrg 3398 1.1 mrg done: 3399 1.1 mrg isl_basic_map_free(bmap); 3400 1.1 mrg clear_status(info_i); 3401 1.1 mrg return change; 3402 1.1 mrg error: 3403 1.1 mrg isl_basic_map_free(bmap); 3404 1.1 mrg clear_status(info_i); 3405 1.1 mrg return isl_change_error; 3406 1.1 mrg } 3407 1.1 mrg 3408 1.1 mrg /* Check if the union of "bmap_i" and the basic map represented by info[j] 3409 1.1 mrg * can be represented by a single basic map, 3410 1.1 mrg * after aligning the divs of "bmap_i" to match those of info[j]. 3411 1.1 mrg * If so, replace the pair by the single basic map and return 3412 1.1 mrg * isl_change_drop_first, isl_change_drop_second or isl_change_fuse. 3413 1.1 mrg * Otherwise, return isl_change_none. 3414 1.1 mrg * 3415 1.1 mrg * In particular, check if "bmap_i" contains the basic map represented by 3416 1.1 mrg * info[j] after aligning the divs of "bmap_i" to those of info[j]. 3417 1.1 mrg * Note that this can only succeed if the number of divs of "bmap_i" 3418 1.1 mrg * is smaller than (or equal to) the number of divs of info[j]. 3419 1.1 mrg * 3420 1.1 mrg * We first check if the divs of "bmap_i" are all known and form a subset 3421 1.1 mrg * of those of info[j].bmap. If so, we pass control over to 3422 1.1 mrg * coalesce_with_expanded_divs. 3423 1.1 mrg * 3424 1.1 mrg * If "i" is not equal to -1, then "bmap" is equal to info[i].bmap. 3425 1.1 mrg */ 3426 1.1 mrg static enum isl_change coalesce_after_aligning_divs( 3427 1.1 mrg __isl_keep isl_basic_map *bmap_i, int i, int j, 3428 1.1 mrg struct isl_coalesce_info *info) 3429 1.1 mrg { 3430 1.1 mrg isl_bool known; 3431 1.1 mrg isl_mat *div_i, *div_j, *div; 3432 1.1 mrg int *exp1 = NULL; 3433 1.1 mrg int *exp2 = NULL; 3434 1.1 mrg isl_ctx *ctx; 3435 1.1 mrg enum isl_change change; 3436 1.1 mrg 3437 1.1 mrg known = isl_basic_map_divs_known(bmap_i); 3438 1.1 mrg if (known < 0) 3439 1.1 mrg return isl_change_error; 3440 1.1 mrg if (!known) 3441 1.1 mrg return isl_change_none; 3442 1.1 mrg 3443 1.1 mrg ctx = isl_basic_map_get_ctx(bmap_i); 3444 1.1 mrg 3445 1.1 mrg div_i = isl_basic_map_get_divs(bmap_i); 3446 1.1 mrg div_j = isl_basic_map_get_divs(info[j].bmap); 3447 1.1 mrg 3448 1.1 mrg if (!div_i || !div_j) 3449 1.1 mrg goto error; 3450 1.1 mrg 3451 1.1 mrg exp1 = isl_alloc_array(ctx, int, div_i->n_row); 3452 1.1 mrg exp2 = isl_alloc_array(ctx, int, div_j->n_row); 3453 1.1 mrg if ((div_i->n_row && !exp1) || (div_j->n_row && !exp2)) 3454 1.1 mrg goto error; 3455 1.1 mrg 3456 1.1 mrg div = isl_merge_divs(div_i, div_j, exp1, exp2); 3457 1.1 mrg if (!div) 3458 1.1 mrg goto error; 3459 1.1 mrg 3460 1.1 mrg if (div->n_row == div_j->n_row) 3461 1.1 mrg change = coalesce_with_expanded_divs(bmap_i, 3462 1.1 mrg i, j, info, div, exp1); 3463 1.1 mrg else 3464 1.1 mrg change = isl_change_none; 3465 1.1 mrg 3466 1.1 mrg isl_mat_free(div); 3467 1.1 mrg 3468 1.1 mrg isl_mat_free(div_i); 3469 1.1 mrg isl_mat_free(div_j); 3470 1.1 mrg 3471 1.1 mrg free(exp2); 3472 1.1 mrg free(exp1); 3473 1.1 mrg 3474 1.1 mrg return change; 3475 1.1 mrg error: 3476 1.1 mrg isl_mat_free(div_i); 3477 1.1 mrg isl_mat_free(div_j); 3478 1.1 mrg free(exp1); 3479 1.1 mrg free(exp2); 3480 1.1 mrg return isl_change_error; 3481 1.1 mrg } 3482 1.1 mrg 3483 1.1 mrg /* Check if basic map "j" is a subset of basic map "i" after 3484 1.1 mrg * exploiting the extra equalities of "j" to simplify the divs of "i". 3485 1.1 mrg * If so, remove basic map "j" and return isl_change_drop_second. 3486 1.1 mrg * 3487 1.1 mrg * If "j" does not have any equalities or if they are the same 3488 1.1 mrg * as those of "i", then we cannot exploit them to simplify the divs. 3489 1.1 mrg * Similarly, if there are no divs in "i", then they cannot be simplified. 3490 1.1 mrg * If, on the other hand, the affine hulls of "i" and "j" do not intersect, 3491 1.1 mrg * then "j" cannot be a subset of "i". 3492 1.1 mrg * 3493 1.1 mrg * Otherwise, we intersect "i" with the affine hull of "j" and then 3494 1.1 mrg * check if "j" is a subset of the result after aligning the divs. 3495 1.1 mrg * If so, then "j" is definitely a subset of "i" and can be removed. 3496 1.1 mrg * Note that if after intersection with the affine hull of "j". 3497 1.1 mrg * "i" still has more divs than "j", then there is no way we can 3498 1.1 mrg * align the divs of "i" to those of "j". 3499 1.1 mrg */ 3500 1.1 mrg static enum isl_change coalesce_subset_with_equalities(int i, int j, 3501 1.1 mrg struct isl_coalesce_info *info) 3502 1.1 mrg { 3503 1.1 mrg isl_basic_map *hull_i, *hull_j, *bmap_i; 3504 1.1 mrg int equal, empty; 3505 1.1 mrg enum isl_change change; 3506 1.1 mrg 3507 1.1 mrg if (info[j].bmap->n_eq == 0) 3508 1.1 mrg return isl_change_none; 3509 1.1 mrg if (info[i].bmap->n_div == 0) 3510 1.1 mrg return isl_change_none; 3511 1.1 mrg 3512 1.1 mrg hull_i = isl_basic_map_copy(info[i].bmap); 3513 1.1 mrg hull_i = isl_basic_map_plain_affine_hull(hull_i); 3514 1.1 mrg hull_j = isl_basic_map_copy(info[j].bmap); 3515 1.1 mrg hull_j = isl_basic_map_plain_affine_hull(hull_j); 3516 1.1 mrg 3517 1.1 mrg hull_j = isl_basic_map_intersect(hull_j, isl_basic_map_copy(hull_i)); 3518 1.1 mrg equal = isl_basic_map_plain_is_equal(hull_i, hull_j); 3519 1.1 mrg empty = isl_basic_map_plain_is_empty(hull_j); 3520 1.1 mrg isl_basic_map_free(hull_i); 3521 1.1 mrg 3522 1.1 mrg if (equal < 0 || equal || empty < 0 || empty) { 3523 1.1 mrg isl_basic_map_free(hull_j); 3524 1.1 mrg if (equal < 0 || empty < 0) 3525 1.1 mrg return isl_change_error; 3526 1.1 mrg return isl_change_none; 3527 1.1 mrg } 3528 1.1 mrg 3529 1.1 mrg bmap_i = isl_basic_map_copy(info[i].bmap); 3530 1.1 mrg bmap_i = isl_basic_map_intersect(bmap_i, hull_j); 3531 1.1 mrg if (!bmap_i) 3532 1.1 mrg return isl_change_error; 3533 1.1 mrg 3534 1.1 mrg if (bmap_i->n_div > info[j].bmap->n_div) { 3535 1.1 mrg isl_basic_map_free(bmap_i); 3536 1.1 mrg return isl_change_none; 3537 1.1 mrg } 3538 1.1 mrg 3539 1.1 mrg change = coalesce_after_aligning_divs(bmap_i, -1, j, info); 3540 1.1 mrg 3541 1.1 mrg isl_basic_map_free(bmap_i); 3542 1.1 mrg 3543 1.1 mrg return change; 3544 1.1 mrg } 3545 1.1 mrg 3546 1.1 mrg /* Check if the union of the basic maps represented by info[i] and info[j] 3547 1.1 mrg * can be represented by a single basic map, by aligning or equating 3548 1.1 mrg * their integer divisions. 3549 1.1 mrg * If so, replace the pair by the single basic map and return 3550 1.1 mrg * isl_change_drop_first, isl_change_drop_second or isl_change_fuse. 3551 1.1 mrg * Otherwise, return isl_change_none. 3552 1.1 mrg * 3553 1.1 mrg * Note that we only perform any test if the number of divs is different 3554 1.1 mrg * in the two basic maps. In case the number of divs is the same, 3555 1.1 mrg * we have already established that the divs are different 3556 1.1 mrg * in the two basic maps. 3557 1.1 mrg * In particular, if the number of divs of basic map i is smaller than 3558 1.1 mrg * the number of divs of basic map j, then we check if j is a subset of i 3559 1.1 mrg * and vice versa. 3560 1.1 mrg */ 3561 1.1 mrg static enum isl_change coalesce_divs(int i, int j, 3562 1.1 mrg struct isl_coalesce_info *info) 3563 1.1 mrg { 3564 1.1 mrg enum isl_change change = isl_change_none; 3565 1.1 mrg 3566 1.1 mrg if (info[i].bmap->n_div < info[j].bmap->n_div) 3567 1.1 mrg change = coalesce_after_aligning_divs(info[i].bmap, i, j, info); 3568 1.1 mrg if (change != isl_change_none) 3569 1.1 mrg return change; 3570 1.1 mrg 3571 1.1 mrg if (info[j].bmap->n_div < info[i].bmap->n_div) 3572 1.1 mrg change = coalesce_after_aligning_divs(info[j].bmap, j, i, info); 3573 1.1 mrg if (change != isl_change_none) 3574 1.1 mrg return invert_change(change); 3575 1.1 mrg 3576 1.1 mrg change = coalesce_subset_with_equalities(i, j, info); 3577 1.1 mrg if (change != isl_change_none) 3578 1.1 mrg return change; 3579 1.1 mrg 3580 1.1 mrg change = coalesce_subset_with_equalities(j, i, info); 3581 1.1 mrg if (change != isl_change_none) 3582 1.1 mrg return invert_change(change); 3583 1.1 mrg 3584 1.1 mrg return isl_change_none; 3585 1.1 mrg } 3586 1.1 mrg 3587 1.1 mrg /* Does "bmap" involve any divs that themselves refer to divs? 3588 1.1 mrg */ 3589 1.1 mrg static isl_bool has_nested_div(__isl_keep isl_basic_map *bmap) 3590 1.1 mrg { 3591 1.1 mrg int i; 3592 1.1 mrg isl_size total; 3593 1.1 mrg isl_size n_div; 3594 1.1 mrg 3595 1.1 mrg total = isl_basic_map_dim(bmap, isl_dim_all); 3596 1.1 mrg n_div = isl_basic_map_dim(bmap, isl_dim_div); 3597 1.1 mrg if (total < 0 || n_div < 0) 3598 1.1 mrg return isl_bool_error; 3599 1.1 mrg total -= n_div; 3600 1.1 mrg 3601 1.1 mrg for (i = 0; i < n_div; ++i) 3602 1.1 mrg if (isl_seq_first_non_zero(bmap->div[i] + 2 + total, 3603 1.1 mrg n_div) != -1) 3604 1.1 mrg return isl_bool_true; 3605 1.1 mrg 3606 1.1 mrg return isl_bool_false; 3607 1.1 mrg } 3608 1.1 mrg 3609 1.1 mrg /* Return a list of affine expressions, one for each integer division 3610 1.1 mrg * in "bmap_i". For each integer division that also appears in "bmap_j", 3611 1.1 mrg * the affine expression is set to NaN. The number of NaNs in the list 3612 1.1 mrg * is equal to the number of integer divisions in "bmap_j". 3613 1.1 mrg * For the other integer divisions of "bmap_i", the corresponding 3614 1.1 mrg * element in the list is a purely affine expression equal to the integer 3615 1.1 mrg * division in "hull". 3616 1.1 mrg * If no such list can be constructed, then the number of elements 3617 1.1 mrg * in the returned list is smaller than the number of integer divisions 3618 1.1 mrg * in "bmap_i". 3619 1.1 mrg * The integer division of "bmap_i" and "bmap_j" are assumed to be known and 3620 1.1 mrg * not contain any nested divs. 3621 1.1 mrg */ 3622 1.1 mrg static __isl_give isl_aff_list *set_up_substitutions( 3623 1.1 mrg __isl_keep isl_basic_map *bmap_i, __isl_keep isl_basic_map *bmap_j, 3624 1.1 mrg __isl_take isl_basic_map *hull) 3625 1.1 mrg { 3626 1.1 mrg isl_size n_div_i, n_div_j, total; 3627 1.1 mrg isl_ctx *ctx; 3628 1.1 mrg isl_local_space *ls; 3629 1.1 mrg isl_basic_set *wrap_hull; 3630 1.1 mrg isl_aff *aff_nan; 3631 1.1 mrg isl_aff_list *list; 3632 1.1 mrg int i, j; 3633 1.1 mrg 3634 1.1 mrg n_div_i = isl_basic_map_dim(bmap_i, isl_dim_div); 3635 1.1 mrg n_div_j = isl_basic_map_dim(bmap_j, isl_dim_div); 3636 1.1 mrg total = isl_basic_map_dim(bmap_i, isl_dim_all); 3637 1.1 mrg if (!hull || n_div_i < 0 || n_div_j < 0 || total < 0) 3638 1.1 mrg return NULL; 3639 1.1 mrg 3640 1.1 mrg ctx = isl_basic_map_get_ctx(hull); 3641 1.1 mrg total -= n_div_i; 3642 1.1 mrg 3643 1.1 mrg ls = isl_basic_map_get_local_space(bmap_i); 3644 1.1 mrg ls = isl_local_space_wrap(ls); 3645 1.1 mrg wrap_hull = isl_basic_map_wrap(hull); 3646 1.1 mrg 3647 1.1 mrg aff_nan = isl_aff_nan_on_domain(isl_local_space_copy(ls)); 3648 1.1 mrg list = isl_aff_list_alloc(ctx, n_div_i); 3649 1.1 mrg 3650 1.1 mrg j = 0; 3651 1.1 mrg for (i = 0; i < n_div_i; ++i) { 3652 1.1 mrg isl_aff *aff; 3653 1.1 mrg isl_size n_div; 3654 1.1 mrg 3655 1.1 mrg if (j < n_div_j && 3656 1.1 mrg isl_basic_map_equal_div_expr_part(bmap_i, i, bmap_j, j, 3657 1.1 mrg 0, 2 + total)) { 3658 1.1 mrg ++j; 3659 1.1 mrg list = isl_aff_list_add(list, isl_aff_copy(aff_nan)); 3660 1.1 mrg continue; 3661 1.1 mrg } 3662 1.1 mrg if (n_div_i - i <= n_div_j - j) 3663 1.1 mrg break; 3664 1.1 mrg 3665 1.1 mrg aff = isl_local_space_get_div(ls, i); 3666 1.1 mrg aff = isl_aff_substitute_equalities(aff, 3667 1.1 mrg isl_basic_set_copy(wrap_hull)); 3668 1.1 mrg aff = isl_aff_floor(aff); 3669 1.1 mrg n_div = isl_aff_dim(aff, isl_dim_div); 3670 1.1 mrg if (n_div < 0) 3671 1.1 mrg goto error; 3672 1.1 mrg if (n_div != 0) { 3673 1.1 mrg isl_aff_free(aff); 3674 1.1 mrg break; 3675 1.1 mrg } 3676 1.1 mrg 3677 1.1 mrg list = isl_aff_list_add(list, aff); 3678 1.1 mrg } 3679 1.1 mrg 3680 1.1 mrg isl_aff_free(aff_nan); 3681 1.1 mrg isl_local_space_free(ls); 3682 1.1 mrg isl_basic_set_free(wrap_hull); 3683 1.1 mrg 3684 1.1 mrg return list; 3685 1.1 mrg error: 3686 1.1 mrg isl_aff_free(aff_nan); 3687 1.1 mrg isl_local_space_free(ls); 3688 1.1 mrg isl_basic_set_free(wrap_hull); 3689 1.1 mrg isl_aff_list_free(list); 3690 1.1 mrg return NULL; 3691 1.1 mrg } 3692 1.1 mrg 3693 1.1 mrg /* Add variables to info->bmap and info->tab corresponding to the elements 3694 1.1 mrg * in "list" that are not set to NaN. 3695 1.1 mrg * "extra_var" is the number of these elements. 3696 1.1 mrg * "dim" is the offset in the variables of "tab" where we should 3697 1.1 mrg * start considering the elements in "list". 3698 1.1 mrg * When this function returns, the total number of variables in "tab" 3699 1.1 mrg * is equal to "dim" plus the number of elements in "list". 3700 1.1 mrg * 3701 1.1 mrg * The newly added existentially quantified variables are not given 3702 1.1 mrg * an explicit representation because the corresponding div constraints 3703 1.1 mrg * do not appear in info->bmap. These constraints are not added 3704 1.1 mrg * to info->bmap because for internal consistency, they would need to 3705 1.1 mrg * be added to info->tab as well, where they could combine with the equality 3706 1.1 mrg * that is added later to result in constraints that do not hold 3707 1.1 mrg * in the original input. 3708 1.1 mrg */ 3709 1.1 mrg static isl_stat add_sub_vars(struct isl_coalesce_info *info, 3710 1.1 mrg __isl_keep isl_aff_list *list, int dim, int extra_var) 3711 1.1 mrg { 3712 1.1 mrg int i, j, d; 3713 1.1 mrg isl_size n; 3714 1.1 mrg 3715 1.1 mrg info->bmap = isl_basic_map_cow(info->bmap); 3716 1.1 mrg info->bmap = isl_basic_map_extend(info->bmap, extra_var, 0, 0); 3717 1.1 mrg n = isl_aff_list_n_aff(list); 3718 1.1 mrg if (!info->bmap || n < 0) 3719 1.1 mrg return isl_stat_error; 3720 1.1 mrg for (i = 0; i < n; ++i) { 3721 1.1 mrg int is_nan; 3722 1.1 mrg isl_aff *aff; 3723 1.1 mrg 3724 1.1 mrg aff = isl_aff_list_get_aff(list, i); 3725 1.1 mrg is_nan = isl_aff_is_nan(aff); 3726 1.1 mrg isl_aff_free(aff); 3727 1.1 mrg if (is_nan < 0) 3728 1.1 mrg return isl_stat_error; 3729 1.1 mrg if (is_nan) 3730 1.1 mrg continue; 3731 1.1 mrg 3732 1.1 mrg if (isl_tab_insert_var(info->tab, dim + i) < 0) 3733 1.1 mrg return isl_stat_error; 3734 1.1 mrg d = isl_basic_map_alloc_div(info->bmap); 3735 1.1 mrg if (d < 0) 3736 1.1 mrg return isl_stat_error; 3737 1.1 mrg info->bmap = isl_basic_map_mark_div_unknown(info->bmap, d); 3738 1.1 mrg for (j = d; j > i; --j) 3739 1.1 mrg info->bmap = isl_basic_map_swap_div(info->bmap, 3740 1.1 mrg j - 1, j); 3741 1.1 mrg if (!info->bmap) 3742 1.1 mrg return isl_stat_error; 3743 1.1 mrg } 3744 1.1 mrg 3745 1.1 mrg return isl_stat_ok; 3746 1.1 mrg } 3747 1.1 mrg 3748 1.1 mrg /* For each element in "list" that is not set to NaN, fix the corresponding 3749 1.1 mrg * variable in "tab" to the purely affine expression defined by the element. 3750 1.1 mrg * "dim" is the offset in the variables of "tab" where we should 3751 1.1 mrg * start considering the elements in "list". 3752 1.1 mrg * 3753 1.1 mrg * This function assumes that a sufficient number of rows and 3754 1.1 mrg * elements in the constraint array are available in the tableau. 3755 1.1 mrg */ 3756 1.1 mrg static isl_stat add_sub_equalities(struct isl_tab *tab, 3757 1.1 mrg __isl_keep isl_aff_list *list, int dim) 3758 1.1 mrg { 3759 1.1 mrg int i; 3760 1.1 mrg isl_size n; 3761 1.1 mrg isl_ctx *ctx; 3762 1.1 mrg isl_vec *sub; 3763 1.1 mrg isl_aff *aff; 3764 1.1 mrg 3765 1.1 mrg n = isl_aff_list_n_aff(list); 3766 1.1 mrg if (n < 0) 3767 1.1 mrg return isl_stat_error; 3768 1.1 mrg 3769 1.1 mrg ctx = isl_tab_get_ctx(tab); 3770 1.1 mrg sub = isl_vec_alloc(ctx, 1 + dim + n); 3771 1.1 mrg if (!sub) 3772 1.1 mrg return isl_stat_error; 3773 1.1 mrg isl_seq_clr(sub->el + 1 + dim, n); 3774 1.1 mrg 3775 1.1 mrg for (i = 0; i < n; ++i) { 3776 1.1 mrg aff = isl_aff_list_get_aff(list, i); 3777 1.1 mrg if (!aff) 3778 1.1 mrg goto error; 3779 1.1 mrg if (isl_aff_is_nan(aff)) { 3780 1.1 mrg isl_aff_free(aff); 3781 1.1 mrg continue; 3782 1.1 mrg } 3783 1.1 mrg isl_seq_cpy(sub->el, aff->v->el + 1, 1 + dim); 3784 1.1 mrg isl_int_neg(sub->el[1 + dim + i], aff->v->el[0]); 3785 1.1 mrg if (isl_tab_add_eq(tab, sub->el) < 0) 3786 1.1 mrg goto error; 3787 1.1 mrg isl_int_set_si(sub->el[1 + dim + i], 0); 3788 1.1 mrg isl_aff_free(aff); 3789 1.1 mrg } 3790 1.1 mrg 3791 1.1 mrg isl_vec_free(sub); 3792 1.1 mrg return isl_stat_ok; 3793 1.1 mrg error: 3794 1.1 mrg isl_aff_free(aff); 3795 1.1 mrg isl_vec_free(sub); 3796 1.1 mrg return isl_stat_error; 3797 1.1 mrg } 3798 1.1 mrg 3799 1.1 mrg /* Add variables to info->tab and info->bmap corresponding to the elements 3800 1.1 mrg * in "list" that are not set to NaN. The value of the added variable 3801 1.1 mrg * in info->tab is fixed to the purely affine expression defined by the element. 3802 1.1 mrg * "dim" is the offset in the variables of info->tab where we should 3803 1.1 mrg * start considering the elements in "list". 3804 1.1 mrg * When this function returns, the total number of variables in info->tab 3805 1.1 mrg * is equal to "dim" plus the number of elements in "list". 3806 1.1 mrg */ 3807 1.1 mrg static isl_stat add_subs(struct isl_coalesce_info *info, 3808 1.1 mrg __isl_keep isl_aff_list *list, int dim) 3809 1.1 mrg { 3810 1.1 mrg int extra_var; 3811 1.1 mrg isl_size n; 3812 1.1 mrg 3813 1.1 mrg n = isl_aff_list_n_aff(list); 3814 1.1 mrg if (n < 0) 3815 1.1 mrg return isl_stat_error; 3816 1.1 mrg 3817 1.1 mrg extra_var = n - (info->tab->n_var - dim); 3818 1.1 mrg 3819 1.1 mrg if (isl_tab_extend_vars(info->tab, extra_var) < 0) 3820 1.1 mrg return isl_stat_error; 3821 1.1 mrg if (isl_tab_extend_cons(info->tab, 2 * extra_var) < 0) 3822 1.1 mrg return isl_stat_error; 3823 1.1 mrg if (add_sub_vars(info, list, dim, extra_var) < 0) 3824 1.1 mrg return isl_stat_error; 3825 1.1 mrg 3826 1.1 mrg return add_sub_equalities(info->tab, list, dim); 3827 1.1 mrg } 3828 1.1 mrg 3829 1.1 mrg /* Coalesce basic map "j" into basic map "i" after adding the extra integer 3830 1.1 mrg * divisions in "i" but not in "j" to basic map "j", with values 3831 1.1 mrg * specified by "list". The total number of elements in "list" 3832 1.1 mrg * is equal to the number of integer divisions in "i", while the number 3833 1.1 mrg * of NaN elements in the list is equal to the number of integer divisions 3834 1.1 mrg * in "j". 3835 1.1 mrg * 3836 1.1 mrg * If no coalescing can be performed, then we need to revert basic map "j" 3837 1.1 mrg * to its original state. We do the same if basic map "i" gets dropped 3838 1.1 mrg * during the coalescing, even though this should not happen in practice 3839 1.1 mrg * since we have already checked for "j" being a subset of "i" 3840 1.1 mrg * before we reach this stage. 3841 1.1 mrg */ 3842 1.1 mrg static enum isl_change coalesce_with_subs(int i, int j, 3843 1.1 mrg struct isl_coalesce_info *info, __isl_keep isl_aff_list *list) 3844 1.1 mrg { 3845 1.1 mrg isl_basic_map *bmap_j; 3846 1.1 mrg struct isl_tab_undo *snap; 3847 1.1 mrg isl_size dim, n_div; 3848 1.1 mrg enum isl_change change; 3849 1.1 mrg 3850 1.1 mrg bmap_j = isl_basic_map_copy(info[j].bmap); 3851 1.1 mrg snap = isl_tab_snap(info[j].tab); 3852 1.1 mrg 3853 1.1 mrg dim = isl_basic_map_dim(bmap_j, isl_dim_all); 3854 1.1 mrg n_div = isl_basic_map_dim(bmap_j, isl_dim_div); 3855 1.1 mrg if (dim < 0 || n_div < 0) 3856 1.1 mrg goto error; 3857 1.1 mrg dim -= n_div; 3858 1.1 mrg if (add_subs(&info[j], list, dim) < 0) 3859 1.1 mrg goto error; 3860 1.1 mrg 3861 1.1 mrg change = coalesce_local_pair(i, j, info); 3862 1.1 mrg if (change != isl_change_none && change != isl_change_drop_first) { 3863 1.1 mrg isl_basic_map_free(bmap_j); 3864 1.1 mrg } else { 3865 1.1 mrg isl_basic_map_free(info[j].bmap); 3866 1.1 mrg info[j].bmap = bmap_j; 3867 1.1 mrg 3868 1.1 mrg if (isl_tab_rollback(info[j].tab, snap) < 0) 3869 1.1 mrg return isl_change_error; 3870 1.1 mrg } 3871 1.1 mrg 3872 1.1 mrg return change; 3873 1.1 mrg error: 3874 1.1 mrg isl_basic_map_free(bmap_j); 3875 1.1 mrg return isl_change_error; 3876 1.1 mrg } 3877 1.1 mrg 3878 1.1 mrg /* Check if we can coalesce basic map "j" into basic map "i" after copying 3879 1.1 mrg * those extra integer divisions in "i" that can be simplified away 3880 1.1 mrg * using the extra equalities in "j". 3881 1.1 mrg * All divs are assumed to be known and not contain any nested divs. 3882 1.1 mrg * 3883 1.1 mrg * We first check if there are any extra equalities in "j" that we 3884 1.1 mrg * can exploit. Then we check if every integer division in "i" 3885 1.1 mrg * either already appears in "j" or can be simplified using the 3886 1.1 mrg * extra equalities to a purely affine expression. 3887 1.1 mrg * If these tests succeed, then we try to coalesce the two basic maps 3888 1.1 mrg * by introducing extra dimensions in "j" corresponding to 3889 1.1 mrg * the extra integer divisions "i" fixed to the corresponding 3890 1.1 mrg * purely affine expression. 3891 1.1 mrg */ 3892 1.1 mrg static enum isl_change check_coalesce_into_eq(int i, int j, 3893 1.1 mrg struct isl_coalesce_info *info) 3894 1.1 mrg { 3895 1.1 mrg isl_size n_div_i, n_div_j, n; 3896 1.1 mrg isl_basic_map *hull_i, *hull_j; 3897 1.1 mrg isl_bool equal, empty; 3898 1.1 mrg isl_aff_list *list; 3899 1.1 mrg enum isl_change change; 3900 1.1 mrg 3901 1.1 mrg n_div_i = isl_basic_map_dim(info[i].bmap, isl_dim_div); 3902 1.1 mrg n_div_j = isl_basic_map_dim(info[j].bmap, isl_dim_div); 3903 1.1 mrg if (n_div_i < 0 || n_div_j < 0) 3904 1.1 mrg return isl_change_error; 3905 1.1 mrg if (n_div_i <= n_div_j) 3906 1.1 mrg return isl_change_none; 3907 1.1 mrg if (info[j].bmap->n_eq == 0) 3908 1.1 mrg return isl_change_none; 3909 1.1 mrg 3910 1.1 mrg hull_i = isl_basic_map_copy(info[i].bmap); 3911 1.1 mrg hull_i = isl_basic_map_plain_affine_hull(hull_i); 3912 1.1 mrg hull_j = isl_basic_map_copy(info[j].bmap); 3913 1.1 mrg hull_j = isl_basic_map_plain_affine_hull(hull_j); 3914 1.1 mrg 3915 1.1 mrg hull_j = isl_basic_map_intersect(hull_j, isl_basic_map_copy(hull_i)); 3916 1.1 mrg equal = isl_basic_map_plain_is_equal(hull_i, hull_j); 3917 1.1 mrg empty = isl_basic_map_plain_is_empty(hull_j); 3918 1.1 mrg isl_basic_map_free(hull_i); 3919 1.1 mrg 3920 1.1 mrg if (equal < 0 || empty < 0) 3921 1.1 mrg goto error; 3922 1.1 mrg if (equal || empty) { 3923 1.1 mrg isl_basic_map_free(hull_j); 3924 1.1 mrg return isl_change_none; 3925 1.1 mrg } 3926 1.1 mrg 3927 1.1 mrg list = set_up_substitutions(info[i].bmap, info[j].bmap, hull_j); 3928 1.1 mrg if (!list) 3929 1.1 mrg return isl_change_error; 3930 1.1 mrg n = isl_aff_list_n_aff(list); 3931 1.1 mrg if (n < 0) 3932 1.1 mrg change = isl_change_error; 3933 1.1 mrg else if (n < n_div_i) 3934 1.1 mrg change = isl_change_none; 3935 1.1 mrg else 3936 1.1 mrg change = coalesce_with_subs(i, j, info, list); 3937 1.1 mrg 3938 1.1 mrg isl_aff_list_free(list); 3939 1.1 mrg 3940 1.1 mrg return change; 3941 1.1 mrg error: 3942 1.1 mrg isl_basic_map_free(hull_j); 3943 1.1 mrg return isl_change_error; 3944 1.1 mrg } 3945 1.1 mrg 3946 1.1 mrg /* Check if we can coalesce basic maps "i" and "j" after copying 3947 1.1 mrg * those extra integer divisions in one of the basic maps that can 3948 1.1 mrg * be simplified away using the extra equalities in the other basic map. 3949 1.1 mrg * We require all divs to be known in both basic maps. 3950 1.1 mrg * Furthermore, to simplify the comparison of div expressions, 3951 1.1 mrg * we do not allow any nested integer divisions. 3952 1.1 mrg */ 3953 1.1 mrg static enum isl_change check_coalesce_eq(int i, int j, 3954 1.1 mrg struct isl_coalesce_info *info) 3955 1.1 mrg { 3956 1.1 mrg isl_bool known, nested; 3957 1.1 mrg enum isl_change change; 3958 1.1 mrg 3959 1.1 mrg known = isl_basic_map_divs_known(info[i].bmap); 3960 1.1 mrg if (known < 0 || !known) 3961 1.1 mrg return known < 0 ? isl_change_error : isl_change_none; 3962 1.1 mrg known = isl_basic_map_divs_known(info[j].bmap); 3963 1.1 mrg if (known < 0 || !known) 3964 1.1 mrg return known < 0 ? isl_change_error : isl_change_none; 3965 1.1 mrg nested = has_nested_div(info[i].bmap); 3966 1.1 mrg if (nested < 0 || nested) 3967 1.1 mrg return nested < 0 ? isl_change_error : isl_change_none; 3968 1.1 mrg nested = has_nested_div(info[j].bmap); 3969 1.1 mrg if (nested < 0 || nested) 3970 1.1 mrg return nested < 0 ? isl_change_error : isl_change_none; 3971 1.1 mrg 3972 1.1 mrg change = check_coalesce_into_eq(i, j, info); 3973 1.1 mrg if (change != isl_change_none) 3974 1.1 mrg return change; 3975 1.1 mrg change = check_coalesce_into_eq(j, i, info); 3976 1.1 mrg if (change != isl_change_none) 3977 1.1 mrg return invert_change(change); 3978 1.1 mrg 3979 1.1 mrg return isl_change_none; 3980 1.1 mrg } 3981 1.1 mrg 3982 1.1 mrg /* Check if the union of the given pair of basic maps 3983 1.1 mrg * can be represented by a single basic map. 3984 1.1 mrg * If so, replace the pair by the single basic map and return 3985 1.1 mrg * isl_change_drop_first, isl_change_drop_second or isl_change_fuse. 3986 1.1 mrg * Otherwise, return isl_change_none. 3987 1.1 mrg * 3988 1.1 mrg * We first check if the two basic maps live in the same local space, 3989 1.1 mrg * after aligning the divs that differ by only an integer constant. 3990 1.1 mrg * If so, we do the complete check. Otherwise, we check if they have 3991 1.1 mrg * the same number of integer divisions and can be coalesced, if one is 3992 1.1 mrg * an obvious subset of the other or if the extra integer divisions 3993 1.1 mrg * of one basic map can be simplified away using the extra equalities 3994 1.1 mrg * of the other basic map. 3995 1.1 mrg * 3996 1.1 mrg * Note that trying to coalesce pairs of disjuncts with the same 3997 1.1 mrg * number, but different local variables may drop the explicit 3998 1.1 mrg * representation of some of these local variables. 3999 1.1 mrg * This operation is therefore not performed when 4000 1.1 mrg * the "coalesce_preserve_locals" option is set. 4001 1.1 mrg */ 4002 1.1 mrg static enum isl_change coalesce_pair(int i, int j, 4003 1.1 mrg struct isl_coalesce_info *info) 4004 1.1 mrg { 4005 1.1 mrg int preserve; 4006 1.1 mrg isl_bool same; 4007 1.1 mrg enum isl_change change; 4008 1.1 mrg isl_ctx *ctx; 4009 1.1 mrg 4010 1.1 mrg if (harmonize_divs(&info[i], &info[j]) < 0) 4011 1.1 mrg return isl_change_error; 4012 1.1 mrg same = same_divs(info[i].bmap, info[j].bmap); 4013 1.1 mrg if (same < 0) 4014 1.1 mrg return isl_change_error; 4015 1.1 mrg if (same) 4016 1.1 mrg return coalesce_local_pair(i, j, info); 4017 1.1 mrg 4018 1.1 mrg ctx = isl_basic_map_get_ctx(info[i].bmap); 4019 1.1 mrg preserve = isl_options_get_coalesce_preserve_locals(ctx); 4020 1.1 mrg if (!preserve && info[i].bmap->n_div == info[j].bmap->n_div) { 4021 1.1 mrg change = coalesce_local_pair(i, j, info); 4022 1.1 mrg if (change != isl_change_none) 4023 1.1 mrg return change; 4024 1.1 mrg } 4025 1.1 mrg 4026 1.1 mrg change = coalesce_divs(i, j, info); 4027 1.1 mrg if (change != isl_change_none) 4028 1.1 mrg return change; 4029 1.1 mrg 4030 1.1 mrg return check_coalesce_eq(i, j, info); 4031 1.1 mrg } 4032 1.1 mrg 4033 1.1 mrg /* Return the maximum of "a" and "b". 4034 1.1 mrg */ 4035 1.1 mrg static int isl_max(int a, int b) 4036 1.1 mrg { 4037 1.1 mrg return a > b ? a : b; 4038 1.1 mrg } 4039 1.1 mrg 4040 1.1 mrg /* Pairwise coalesce the basic maps in the range [start1, end1[ of "info" 4041 1.1 mrg * with those in the range [start2, end2[, skipping basic maps 4042 1.1 mrg * that have been removed (either before or within this function). 4043 1.1 mrg * 4044 1.1 mrg * For each basic map i in the first range, we check if it can be coalesced 4045 1.1 mrg * with respect to any previously considered basic map j in the second range. 4046 1.1 mrg * If i gets dropped (because it was a subset of some j), then 4047 1.1 mrg * we can move on to the next basic map. 4048 1.1 mrg * If j gets dropped, we need to continue checking against the other 4049 1.1 mrg * previously considered basic maps. 4050 1.1 mrg * If the two basic maps got fused, then we recheck the fused basic map 4051 1.1 mrg * against the previously considered basic maps, starting at i + 1 4052 1.1 mrg * (even if start2 is greater than i + 1). 4053 1.1 mrg */ 4054 1.1 mrg static int coalesce_range(isl_ctx *ctx, struct isl_coalesce_info *info, 4055 1.1 mrg int start1, int end1, int start2, int end2) 4056 1.1 mrg { 4057 1.1 mrg int i, j; 4058 1.1 mrg 4059 1.1 mrg for (i = end1 - 1; i >= start1; --i) { 4060 1.1 mrg if (info[i].removed) 4061 1.1 mrg continue; 4062 1.1 mrg for (j = isl_max(i + 1, start2); j < end2; ++j) { 4063 1.1 mrg enum isl_change changed; 4064 1.1 mrg 4065 1.1 mrg if (info[j].removed) 4066 1.1 mrg continue; 4067 1.1 mrg if (info[i].removed) 4068 1.1 mrg isl_die(ctx, isl_error_internal, 4069 1.1 mrg "basic map unexpectedly removed", 4070 1.1 mrg return -1); 4071 1.1 mrg changed = coalesce_pair(i, j, info); 4072 1.1 mrg switch (changed) { 4073 1.1 mrg case isl_change_error: 4074 1.1 mrg return -1; 4075 1.1 mrg case isl_change_none: 4076 1.1 mrg case isl_change_drop_second: 4077 1.1 mrg continue; 4078 1.1 mrg case isl_change_drop_first: 4079 1.1 mrg j = end2; 4080 1.1 mrg break; 4081 1.1 mrg case isl_change_fuse: 4082 1.1 mrg j = i; 4083 1.1 mrg break; 4084 1.1 mrg } 4085 1.1 mrg } 4086 1.1 mrg } 4087 1.1 mrg 4088 1.1 mrg return 0; 4089 1.1 mrg } 4090 1.1 mrg 4091 1.1 mrg /* Pairwise coalesce the basic maps described by the "n" elements of "info". 4092 1.1 mrg * 4093 1.1 mrg * We consider groups of basic maps that live in the same apparent 4094 1.1 mrg * affine hull and we first coalesce within such a group before we 4095 1.1 mrg * coalesce the elements in the group with elements of previously 4096 1.1 mrg * considered groups. If a fuse happens during the second phase, 4097 1.1 mrg * then we also reconsider the elements within the group. 4098 1.1 mrg */ 4099 1.1 mrg static int coalesce(isl_ctx *ctx, int n, struct isl_coalesce_info *info) 4100 1.1 mrg { 4101 1.1 mrg int start, end; 4102 1.1 mrg 4103 1.1 mrg for (end = n; end > 0; end = start) { 4104 1.1 mrg start = end - 1; 4105 1.1 mrg while (start >= 1 && 4106 1.1 mrg info[start - 1].hull_hash == info[start].hull_hash) 4107 1.1 mrg start--; 4108 1.1 mrg if (coalesce_range(ctx, info, start, end, start, end) < 0) 4109 1.1 mrg return -1; 4110 1.1 mrg if (coalesce_range(ctx, info, start, end, end, n) < 0) 4111 1.1 mrg return -1; 4112 1.1 mrg } 4113 1.1 mrg 4114 1.1 mrg return 0; 4115 1.1 mrg } 4116 1.1 mrg 4117 1.1 mrg /* Update the basic maps in "map" based on the information in "info". 4118 1.1 mrg * In particular, remove the basic maps that have been marked removed and 4119 1.1 mrg * update the others based on the information in the corresponding tableau. 4120 1.1 mrg * Since we detected implicit equalities without calling 4121 1.1 mrg * isl_basic_map_gauss, we need to do it now. 4122 1.1 mrg * Also call isl_basic_map_simplify if we may have lost the definition 4123 1.1 mrg * of one or more integer divisions. 4124 1.1 mrg * If a basic map is still equal to the one from which the corresponding "info" 4125 1.1 mrg * entry was created, then redundant constraint and 4126 1.1 mrg * implicit equality constraint detection have been performed 4127 1.1 mrg * on the corresponding tableau and the basic map can be marked as such. 4128 1.1 mrg */ 4129 1.1 mrg static __isl_give isl_map *update_basic_maps(__isl_take isl_map *map, 4130 1.1 mrg int n, struct isl_coalesce_info *info) 4131 1.1 mrg { 4132 1.1 mrg int i; 4133 1.1 mrg 4134 1.1 mrg if (!map) 4135 1.1 mrg return NULL; 4136 1.1 mrg 4137 1.1 mrg for (i = n - 1; i >= 0; --i) { 4138 1.1 mrg if (info[i].removed) { 4139 1.1 mrg isl_basic_map_free(map->p[i]); 4140 1.1 mrg if (i != map->n - 1) 4141 1.1 mrg map->p[i] = map->p[map->n - 1]; 4142 1.1 mrg map->n--; 4143 1.1 mrg continue; 4144 1.1 mrg } 4145 1.1 mrg 4146 1.1 mrg info[i].bmap = isl_basic_map_update_from_tab(info[i].bmap, 4147 1.1 mrg info[i].tab); 4148 1.1 mrg info[i].bmap = isl_basic_map_gauss(info[i].bmap, NULL); 4149 1.1 mrg if (info[i].simplify) 4150 1.1 mrg info[i].bmap = isl_basic_map_simplify(info[i].bmap); 4151 1.1 mrg info[i].bmap = isl_basic_map_finalize(info[i].bmap); 4152 1.1 mrg if (!info[i].bmap) 4153 1.1 mrg return isl_map_free(map); 4154 1.1 mrg if (!info[i].modified) { 4155 1.1 mrg ISL_F_SET(info[i].bmap, ISL_BASIC_MAP_NO_IMPLICIT); 4156 1.1 mrg ISL_F_SET(info[i].bmap, ISL_BASIC_MAP_NO_REDUNDANT); 4157 1.1 mrg } 4158 1.1 mrg isl_basic_map_free(map->p[i]); 4159 1.1 mrg map->p[i] = info[i].bmap; 4160 1.1 mrg info[i].bmap = NULL; 4161 1.1 mrg } 4162 1.1 mrg 4163 1.1 mrg return map; 4164 1.1 mrg } 4165 1.1 mrg 4166 1.1 mrg /* For each pair of basic maps in the map, check if the union of the two 4167 1.1 mrg * can be represented by a single basic map. 4168 1.1 mrg * If so, replace the pair by the single basic map and start over. 4169 1.1 mrg * 4170 1.1 mrg * We factor out any (hidden) common factor from the constraint 4171 1.1 mrg * coefficients to improve the detection of adjacent constraints. 4172 1.1 mrg * Note that this function does not call isl_basic_map_gauss, 4173 1.1 mrg * but it does make sure that only a single copy of the basic map 4174 1.1 mrg * is affected. This means that isl_basic_map_gauss may have 4175 1.1 mrg * to be called at the end of the computation (in update_basic_maps) 4176 1.1 mrg * on this single copy to ensure that 4177 1.1 mrg * the basic maps are not left in an unexpected state. 4178 1.1 mrg * 4179 1.1 mrg * Since we are constructing the tableaus of the basic maps anyway, 4180 1.1 mrg * we exploit them to detect implicit equalities and redundant constraints. 4181 1.1 mrg * This also helps the coalescing as it can ignore the redundant constraints. 4182 1.1 mrg * In order to avoid confusion, we make all implicit equalities explicit 4183 1.1 mrg * in the basic maps. If the basic map only has a single reference 4184 1.1 mrg * (this happens in particular if it was modified by 4185 1.1 mrg * isl_basic_map_reduce_coefficients), then isl_basic_map_gauss 4186 1.1 mrg * does not get called on the result. The call to 4187 1.1 mrg * isl_basic_map_gauss in update_basic_maps resolves this as well. 4188 1.1 mrg * For each basic map, we also compute the hash of the apparent affine hull 4189 1.1 mrg * for use in coalesce. 4190 1.1 mrg */ 4191 1.1 mrg __isl_give isl_map *isl_map_coalesce(__isl_take isl_map *map) 4192 1.1 mrg { 4193 1.1 mrg int i; 4194 1.1 mrg unsigned n; 4195 1.1 mrg isl_ctx *ctx; 4196 1.1 mrg struct isl_coalesce_info *info = NULL; 4197 1.1 mrg 4198 1.1 mrg map = isl_map_remove_empty_parts(map); 4199 1.1 mrg if (!map) 4200 1.1 mrg return NULL; 4201 1.1 mrg 4202 1.1 mrg if (map->n <= 1) 4203 1.1 mrg return map; 4204 1.1 mrg 4205 1.1 mrg ctx = isl_map_get_ctx(map); 4206 1.1 mrg map = isl_map_sort_divs(map); 4207 1.1 mrg map = isl_map_cow(map); 4208 1.1 mrg 4209 1.1 mrg if (!map) 4210 1.1 mrg return NULL; 4211 1.1 mrg 4212 1.1 mrg n = map->n; 4213 1.1 mrg 4214 1.1 mrg info = isl_calloc_array(map->ctx, struct isl_coalesce_info, n); 4215 1.1 mrg if (!info) 4216 1.1 mrg goto error; 4217 1.1 mrg 4218 1.1 mrg for (i = 0; i < map->n; ++i) { 4219 1.1 mrg map->p[i] = isl_basic_map_reduce_coefficients(map->p[i]); 4220 1.1 mrg if (!map->p[i]) 4221 1.1 mrg goto error; 4222 1.1 mrg info[i].bmap = isl_basic_map_copy(map->p[i]); 4223 1.1 mrg info[i].tab = isl_tab_from_basic_map(info[i].bmap, 0); 4224 1.1 mrg if (!info[i].tab) 4225 1.1 mrg goto error; 4226 1.1 mrg if (!ISL_F_ISSET(info[i].bmap, ISL_BASIC_MAP_NO_IMPLICIT)) 4227 1.1 mrg if (isl_tab_detect_implicit_equalities(info[i].tab) < 0) 4228 1.1 mrg goto error; 4229 1.1 mrg info[i].bmap = isl_tab_make_equalities_explicit(info[i].tab, 4230 1.1 mrg info[i].bmap); 4231 1.1 mrg if (!info[i].bmap) 4232 1.1 mrg goto error; 4233 1.1 mrg if (!ISL_F_ISSET(info[i].bmap, ISL_BASIC_MAP_NO_REDUNDANT)) 4234 1.1 mrg if (isl_tab_detect_redundant(info[i].tab) < 0) 4235 1.1 mrg goto error; 4236 1.1 mrg if (coalesce_info_set_hull_hash(&info[i]) < 0) 4237 1.1 mrg goto error; 4238 1.1 mrg } 4239 1.1 mrg for (i = map->n - 1; i >= 0; --i) 4240 1.1 mrg if (info[i].tab->empty) 4241 1.1 mrg drop(&info[i]); 4242 1.1 mrg 4243 1.1 mrg if (coalesce(ctx, n, info) < 0) 4244 1.1 mrg goto error; 4245 1.1 mrg 4246 1.1 mrg map = update_basic_maps(map, n, info); 4247 1.1 mrg 4248 1.1 mrg clear_coalesce_info(n, info); 4249 1.1 mrg 4250 1.1 mrg return map; 4251 1.1 mrg error: 4252 1.1 mrg clear_coalesce_info(n, info); 4253 1.1 mrg isl_map_free(map); 4254 1.1 mrg return NULL; 4255 1.1 mrg } 4256 1.1 mrg 4257 1.1 mrg /* For each pair of basic sets in the set, check if the union of the two 4258 1.1 mrg * can be represented by a single basic set. 4259 1.1 mrg * If so, replace the pair by the single basic set and start over. 4260 1.1 mrg */ 4261 1.1 mrg __isl_give isl_set *isl_set_coalesce(__isl_take isl_set *set) 4262 1.1 mrg { 4263 1.1 mrg return set_from_map(isl_map_coalesce(set_to_map(set))); 4264 1.1 mrg } 4265