| 1 | /* |
| 2 | * Copyright 2011 INRIA Saclay |
| 3 | * Copyright 2012 Ecole Normale Superieure |
| 4 | * Copyright 2020 Cerebras Systems |
| 5 | * |
| 6 | * Use of this software is governed by the MIT license |
| 7 | * |
| 8 | * Written by Sven Verdoolaege, INRIA Saclay - Ile-de-France, |
| 9 | * Parc Club Orsay Universite, ZAC des vignes, 4 rue Jacques Monod, |
| 10 | * 91893 Orsay, France |
| 11 | * and Ecole Normale Superieure, 45 rue d'Ulm, 75230 Paris, France |
| 12 | * and Cerebras Systems, 175 S San Antonio Rd, Los Altos, CA, USA |
| 13 | */ |
| 14 | |
| 15 | #include <isl_pw_macro.h> |
| 16 | |
| 17 | /* Given a function "cmp" that returns the set of elements where |
| 18 | * "el1" is "better" than "el2", return this set. |
| 19 | */ |
| 20 | static __isl_give isl_set *FN(PW,better)(__isl_keep EL *el1, __isl_keep EL *el2, |
| 21 | __isl_give isl_set *(*cmp)(__isl_take EL *el1, __isl_take EL *el2)) |
| 22 | { |
| 23 | return cmp(FN(EL,copy)(aff: el1), FN(EL,copy)(aff: el2)); |
| 24 | } |
| 25 | |
| 26 | /* Return a list containing the domains of the pieces of "pw". |
| 27 | */ |
| 28 | static __isl_give isl_set_list *FN(PW,extract_domains)(__isl_keep PW *pw) |
| 29 | { |
| 30 | int i; |
| 31 | isl_ctx *ctx; |
| 32 | isl_set_list *list; |
| 33 | |
| 34 | if (!pw) |
| 35 | return NULL; |
| 36 | ctx = FN(PW,get_ctx)(pw); |
| 37 | list = isl_set_list_alloc(ctx, n: pw->n); |
| 38 | for (i = 0; i < pw->n; ++i) |
| 39 | list = isl_set_list_add(list, el: isl_set_copy(set: pw->p[i].set)); |
| 40 | |
| 41 | return list; |
| 42 | } |
| 43 | |
| 44 | /* Given sets B ("set"), C ("better") and A' ("out"), return |
| 45 | * |
| 46 | * (B \cap C) \cup ((B \setminus C) \setminus A') |
| 47 | */ |
| 48 | static __isl_give isl_set *FN(PW,better_or_out)(__isl_take isl_set *set, |
| 49 | __isl_take isl_set *better, __isl_take isl_set *out) |
| 50 | { |
| 51 | isl_set *set_better, *set_out; |
| 52 | |
| 53 | set_better = isl_set_intersect(set1: isl_set_copy(set), set2: isl_set_copy(set: better)); |
| 54 | set_out = isl_set_subtract(set1: isl_set_subtract(set1: set, set2: better), set2: out); |
| 55 | |
| 56 | return isl_set_union(set1: set_better, set2: set_out); |
| 57 | } |
| 58 | |
| 59 | /* Given sets A ("set"), C ("better") and B' ("out"), return |
| 60 | * |
| 61 | * (A \setminus C) \cup ((A \cap C) \setminus B') |
| 62 | */ |
| 63 | static __isl_give isl_set *FN(PW,worse_or_out)(__isl_take isl_set *set, |
| 64 | __isl_take isl_set *better, __isl_take isl_set *out) |
| 65 | { |
| 66 | isl_set *set_worse, *set_out; |
| 67 | |
| 68 | set_worse = isl_set_subtract(set1: isl_set_copy(set), set2: isl_set_copy(set: better)); |
| 69 | set_out = isl_set_subtract(set1: isl_set_intersect(set1: set, set2: better), set2: out); |
| 70 | |
| 71 | return isl_set_union(set1: set_worse, set2: set_out); |
| 72 | } |
| 73 | |
| 74 | /* Internal data structure used by isl_pw_*_union_opt_cmp |
| 75 | * that keeps track of a piecewise expression with updated cells. |
| 76 | * "pw" holds the original piecewise expression. |
| 77 | * "list" holds the updated cells. |
| 78 | */ |
| 79 | S(PW,union_opt_cmp_data) { |
| 80 | PW *pw; |
| 81 | isl_set_list *cell; |
| 82 | }; |
| 83 | |
| 84 | /* Free all memory allocated for "data". |
| 85 | */ |
| 86 | static void FN(PW,union_opt_cmp_data_clear)(S(PW,union_opt_cmp_data) *data) |
| 87 | { |
| 88 | isl_set_list_free(list: data->cell); |
| 89 | FN(PW,free)(pw: data->pw); |
| 90 | } |
| 91 | |
| 92 | /* Given (potentially) updated cells "i" of data_i->pw and "j" of data_j->pw and |
| 93 | * a set "better" where the piece from data_j->pw is better |
| 94 | * than the piece from data_i->pw, |
| 95 | * (further) update the specified cells such that only the better elements |
| 96 | * remain on the (non-empty) intersection. |
| 97 | * |
| 98 | * Let C be the set "better". |
| 99 | * Let A be the cell data_i->cell[i] and B the cell data_j->cell[j]. |
| 100 | * |
| 101 | * The elements in C need to be removed from A, except for those parts |
| 102 | * that lie outside of B. That is, |
| 103 | * |
| 104 | * A <- (A \setminus C) \cup ((A \cap C) \setminus B') |
| 105 | * |
| 106 | * Conversely, the elements in B need to be restricted to C, except |
| 107 | * for those parts that lie outside of A. That is |
| 108 | * |
| 109 | * B <- (B \cap C) \cup ((B \setminus C) \setminus A') |
| 110 | * |
| 111 | * Since all pairs of pieces are considered, the domains are updated |
| 112 | * several times. A and B refer to these updated domains |
| 113 | * (kept track of in data_i->cell[i] and data_j->cell[j]), while A' and B' refer |
| 114 | * to the original domains of the pieces. It is safe to use these |
| 115 | * original domains because the difference between, say, A' and A is |
| 116 | * the domains of pw2-pieces that have been removed before and |
| 117 | * those domains are disjoint from B. A' is used instead of A |
| 118 | * because the continued updating of A may result in this domain |
| 119 | * getting broken up into more disjuncts. |
| 120 | */ |
| 121 | static isl_stat FN(PW,union_opt_cmp_split)(S(PW,union_opt_cmp_data) *data_i, |
| 122 | int i, S(PW,union_opt_cmp_data) *data_j, int j, |
| 123 | __isl_take isl_set *better) |
| 124 | { |
| 125 | isl_set *set_i, *set_j; |
| 126 | |
| 127 | set_i = isl_set_list_get_set(list: data_i->cell, index: i); |
| 128 | set_j = FN(PW,get_domain_at)(pw: data_j->pw, pos: j); |
| 129 | set_i = FN(PW,worse_or_out)(set: set_i, better: isl_set_copy(set: better), out: set_j); |
| 130 | data_i->cell = isl_set_list_set_set(list: data_i->cell, index: i, el: set_i); |
| 131 | set_i = FN(PW,get_domain_at)(pw: data_i->pw, pos: i); |
| 132 | set_j = isl_set_list_get_set(list: data_j->cell, index: j); |
| 133 | set_j = FN(PW,better_or_out)(set: set_j, better, out: set_i); |
| 134 | data_j->cell = isl_set_list_set_set(list: data_j->cell, index: j, el: set_j); |
| 135 | |
| 136 | return isl_stat_ok; |
| 137 | } |
| 138 | |
| 139 | /* Given (potentially) updated cells "i" of data_i->pw and "j" of data_j->pw and |
| 140 | * a function "cmp" that returns the set of elements where |
| 141 | * "el1" is "better" than "el2", |
| 142 | * (further) update the specified cells such that only the "better" elements |
| 143 | * remain on the (non-empty) intersection. |
| 144 | */ |
| 145 | static isl_stat FN(PW,union_opt_cmp_pair)(S(PW,union_opt_cmp_data) *data_i, |
| 146 | int i, S(PW,union_opt_cmp_data) *data_j, int j, |
| 147 | __isl_give isl_set *(*cmp)(__isl_take EL *el1, __isl_take EL *el2)) |
| 148 | { |
| 149 | isl_set *better; |
| 150 | EL *el_i, *el_j; |
| 151 | |
| 152 | el_i = FN(PW,peek_base_at)(pw: data_i->pw, pos: i); |
| 153 | el_j = FN(PW,peek_base_at)(pw: data_j->pw, pos: j); |
| 154 | better = FN(PW,better)(el1: el_j, el2: el_i, cmp); |
| 155 | return FN(PW,union_opt_cmp_split)(data_i, i, data_j, j, better); |
| 156 | } |
| 157 | |
| 158 | /* Given (potentially) updated cells "i" of data_i->pw and "j" of data_j->pw and |
| 159 | * a function "cmp" that returns the set of elements where |
| 160 | * "el1" is "better" than "el2", |
| 161 | * (further) update the specified cells such that only the "better" elements |
| 162 | * remain on the (non-empty) intersection. |
| 163 | * |
| 164 | * The base computation is performed by isl_pw_*_union_opt_cmp_pair, |
| 165 | * which splits the cells according to the set of elements |
| 166 | * where the piece from data_j->pw is better than the piece from data_i->pw. |
| 167 | * |
| 168 | * In some cases, there may be a subset of the intersection |
| 169 | * where both pieces have the same value and can therefore |
| 170 | * both be considered to be "better" than the other. |
| 171 | * This can result in unnecessary splitting on this subset. |
| 172 | * Avoid some of these cases by checking whether |
| 173 | * data_i->pw is always better than data_j->pw on the intersection. |
| 174 | * In particular, do this for the special case where this intersection |
| 175 | * is equal to the cell "j" and data_i->pw is better on its entire cell. |
| 176 | * |
| 177 | * Similarly, if data_i->pw is never better than data_j->pw, |
| 178 | * then no splitting will occur and there is no need to check |
| 179 | * where data_j->pw is better than data_i->pw. |
| 180 | */ |
| 181 | static isl_stat FN(PW,union_opt_cmp_two)(S(PW,union_opt_cmp_data) *data_i, |
| 182 | int i, S(PW,union_opt_cmp_data) *data_j, int j, |
| 183 | __isl_give isl_set *(*cmp)(__isl_take EL *el1, __isl_take EL *el2)) |
| 184 | { |
| 185 | isl_bool is_subset, is_empty; |
| 186 | isl_set *better, *set_i, *set_j; |
| 187 | EL *el_i, *el_j; |
| 188 | |
| 189 | set_i = FN(PW,peek_domain_at)(pw: data_i->pw, pos: i); |
| 190 | set_j = FN(PW,peek_domain_at)(pw: data_j->pw, pos: j); |
| 191 | is_subset = isl_set_is_subset(set1: set_j, set2: set_i); |
| 192 | if (is_subset < 0) |
| 193 | return isl_stat_error; |
| 194 | if (!is_subset) |
| 195 | return FN(PW,union_opt_cmp_pair)(data_i, i, data_j, j, cmp); |
| 196 | |
| 197 | el_i = FN(PW,peek_base_at)(pw: data_i->pw, pos: i); |
| 198 | el_j = FN(PW,peek_base_at)(pw: data_j->pw, pos: j); |
| 199 | better = FN(PW,better)(el1: el_i, el2: el_j, cmp); |
| 200 | is_empty = isl_set_is_empty(set: better); |
| 201 | if (is_empty >= 0 && is_empty) |
| 202 | return FN(PW,union_opt_cmp_split)(data_i: data_j, i: j, data_j: data_i, j: i, better); |
| 203 | is_subset = isl_set_is_subset(set1: set_i, set2: better); |
| 204 | if (is_subset >= 0 && is_subset) |
| 205 | return FN(PW,union_opt_cmp_split)(data_i: data_j, i: j, data_j: data_i, j: i, better); |
| 206 | isl_set_free(set: better); |
| 207 | if (is_empty < 0 || is_subset < 0) |
| 208 | return isl_stat_error; |
| 209 | |
| 210 | return FN(PW,union_opt_cmp_pair)(data_i, i, data_j, j, cmp); |
| 211 | } |
| 212 | |
| 213 | /* Given two piecewise expressions data1->pw and data2->pw, replace |
| 214 | * their domains |
| 215 | * by the sets in data1->cell and data2->cell and combine the results into |
| 216 | * a single piecewise expression. |
| 217 | * The pieces of data1->pw and data2->pw are assumed to have been sorted |
| 218 | * according to the function value expressions. |
| 219 | * The pieces of the result are also sorted in this way. |
| 220 | * |
| 221 | * Run through the pieces of data1->pw and data2->pw in order until they |
| 222 | * have both been exhausted, picking the piece from data1->pw or data2->pw |
| 223 | * depending on which should come first, together with the corresponding |
| 224 | * domain from data1->cell or data2->cell. In cases where the next pieces |
| 225 | * in both data1->pw and data2->pw have the same function value expression, |
| 226 | * construct only a single piece in the result with as domain |
| 227 | * the union of the domains in data1->cell and data2->cell. |
| 228 | */ |
| 229 | static __isl_give PW *FN(PW,merge)(S(PW,union_opt_cmp_data) *data1, |
| 230 | S(PW,union_opt_cmp_data) *data2) |
| 231 | { |
| 232 | int i, j; |
| 233 | PW *res; |
| 234 | PW *pw1 = data1->pw; |
| 235 | PW *pw2 = data2->pw; |
| 236 | isl_set_list *list1 = data1->cell; |
| 237 | isl_set_list *list2 = data2->cell; |
| 238 | |
| 239 | if (!pw1 || !pw2) |
| 240 | return NULL; |
| 241 | |
| 242 | res = FN(PW,alloc_size)(space: isl_space_copy(space: pw1->dim), n: pw1->n + pw2->n); |
| 243 | |
| 244 | i = 0; j = 0; |
| 245 | while (i < pw1->n || j < pw2->n) { |
| 246 | int cmp; |
| 247 | isl_set *set; |
| 248 | EL *el; |
| 249 | |
| 250 | if (i < pw1->n && j < pw2->n) |
| 251 | cmp = FN(EL,plain_cmp)(aff1: pw1->p[i].FIELD, |
| 252 | aff2: pw2->p[j].FIELD); |
| 253 | else |
| 254 | cmp = i < pw1->n ? -1 : 1; |
| 255 | |
| 256 | if (cmp < 0) { |
| 257 | set = isl_set_list_get_set(list: list1, index: i); |
| 258 | el = FN(EL,copy)(aff: pw1->p[i].FIELD); |
| 259 | ++i; |
| 260 | } else if (cmp > 0) { |
| 261 | set = isl_set_list_get_set(list: list2, index: j); |
| 262 | el = FN(EL,copy)(aff: pw2->p[j].FIELD); |
| 263 | ++j; |
| 264 | } else { |
| 265 | set = isl_set_union(set1: isl_set_list_get_set(list: list1, index: i), |
| 266 | set2: isl_set_list_get_set(list: list2, index: j)); |
| 267 | el = FN(EL,copy)(aff: pw1->p[i].FIELD); |
| 268 | ++i; |
| 269 | ++j; |
| 270 | } |
| 271 | res = FN(PW,add_piece)(pw: res, set, el); |
| 272 | } |
| 273 | |
| 274 | return res; |
| 275 | } |
| 276 | |
| 277 | /* Given a function "cmp" that returns the set of elements where |
| 278 | * "el1" is "better" than "el2", return a piecewise |
| 279 | * expression defined on the union of the definition domains |
| 280 | * of "pw1" and "pw2" that maps to the "best" of "pw1" and |
| 281 | * "pw2" on each cell. If only one of the two input functions |
| 282 | * is defined on a given cell, then it is considered the best. |
| 283 | * |
| 284 | * Run through all pairs of pieces in "pw1" and "pw2". |
| 285 | * If the domains of these pieces intersect, then the intersection |
| 286 | * needs to be distributed over the two pieces based on "cmp". |
| 287 | * |
| 288 | * After the updated domains have been computed, the result is constructed |
| 289 | * from "pw1", "pw2", data[0].cell and data[1].cell. If there are any pieces |
| 290 | * in "pw1" and "pw2" with the same function value expression, then |
| 291 | * they are combined into a single piece in the result. |
| 292 | * In order to be able to do this efficiently, the pieces of "pw1" and |
| 293 | * "pw2" are first sorted according to their function value expressions. |
| 294 | */ |
| 295 | static __isl_give PW *FN(PW,union_opt_cmp)( |
| 296 | __isl_take PW *pw1, __isl_take PW *pw2, |
| 297 | __isl_give isl_set *(*cmp)(__isl_take EL *el1, __isl_take EL *el2)) |
| 298 | { |
| 299 | S(PW,union_opt_cmp_data) data[2] = { { pw1, NULL }, { pw2, NULL } }; |
| 300 | int i, j; |
| 301 | isl_size n1, n2; |
| 302 | PW *res = NULL; |
| 303 | isl_ctx *ctx; |
| 304 | |
| 305 | if (!pw1 || !pw2) |
| 306 | goto error; |
| 307 | |
| 308 | ctx = isl_space_get_ctx(space: pw1->dim); |
| 309 | if (!isl_space_is_equal(space1: pw1->dim, space2: pw2->dim)) |
| 310 | isl_die(ctx, isl_error_invalid, |
| 311 | "arguments should live in the same space" , goto error); |
| 312 | |
| 313 | if (FN(PW,is_empty)(pw: pw1)) { |
| 314 | FN(PW,free)(pw: pw1); |
| 315 | return pw2; |
| 316 | } |
| 317 | |
| 318 | if (FN(PW,is_empty)(pw: pw2)) { |
| 319 | FN(PW,free)(pw: pw2); |
| 320 | return pw1; |
| 321 | } |
| 322 | |
| 323 | for (i = 0; i < 2; ++i) { |
| 324 | data[i].pw = FN(PW,sort_unique)(pw: data[i].pw); |
| 325 | data[i].cell = FN(PW,extract_domains)(pw: data[i].pw); |
| 326 | } |
| 327 | |
| 328 | n1 = FN(PW,n_piece)(pw: data[0].pw); |
| 329 | n2 = FN(PW,n_piece)(pw: data[1].pw); |
| 330 | if (n1 < 0 || n2 < 0) |
| 331 | goto error; |
| 332 | for (i = 0; i < n1; ++i) { |
| 333 | for (j = 0; j < n2; ++j) { |
| 334 | isl_bool disjoint; |
| 335 | isl_set *set_i, *set_j; |
| 336 | |
| 337 | set_i = FN(PW,peek_domain_at)(pw: data[0].pw, pos: i); |
| 338 | set_j = FN(PW,peek_domain_at)(pw: data[1].pw, pos: j); |
| 339 | disjoint = isl_set_is_disjoint(set1: set_i, set2: set_j); |
| 340 | if (disjoint < 0) |
| 341 | goto error; |
| 342 | if (disjoint) |
| 343 | continue; |
| 344 | if (FN(PW,union_opt_cmp_two)(data_i: &data[0], i, |
| 345 | data_j: &data[1], j, cmp) < 0) |
| 346 | goto error; |
| 347 | } |
| 348 | } |
| 349 | |
| 350 | res = FN(PW,merge)(data1: &data[0], data2: &data[1]); |
| 351 | for (i = 0; i < 2; ++i) |
| 352 | FN(PW,union_opt_cmp_data_clear)(data: &data[i]); |
| 353 | |
| 354 | return res; |
| 355 | error: |
| 356 | for (i = 0; i < 2; ++i) |
| 357 | FN(PW,union_opt_cmp_data_clear)(data: &data[i]); |
| 358 | return FN(PW,free)(pw: res); |
| 359 | } |
| 360 | |