| 1 | /*------------------------------------------------------------------------- |
| 2 | * |
| 3 | * clausesel.c |
| 4 | * Routines to compute clause selectivities |
| 5 | * |
| 6 | * Portions Copyright (c) 1996-2019, PostgreSQL Global Development Group |
| 7 | * Portions Copyright (c) 1994, Regents of the University of California |
| 8 | * |
| 9 | * |
| 10 | * IDENTIFICATION |
| 11 | * src/backend/optimizer/path/clausesel.c |
| 12 | * |
| 13 | *------------------------------------------------------------------------- |
| 14 | */ |
| 15 | #include "postgres.h" |
| 16 | |
| 17 | #include "nodes/makefuncs.h" |
| 18 | #include "nodes/nodeFuncs.h" |
| 19 | #include "optimizer/clauses.h" |
| 20 | #include "optimizer/cost.h" |
| 21 | #include "optimizer/optimizer.h" |
| 22 | #include "optimizer/pathnode.h" |
| 23 | #include "optimizer/plancat.h" |
| 24 | #include "utils/fmgroids.h" |
| 25 | #include "utils/lsyscache.h" |
| 26 | #include "utils/selfuncs.h" |
| 27 | #include "statistics/statistics.h" |
| 28 | |
| 29 | |
| 30 | /* |
| 31 | * Data structure for accumulating info about possible range-query |
| 32 | * clause pairs in clauselist_selectivity. |
| 33 | */ |
| 34 | typedef struct RangeQueryClause |
| 35 | { |
| 36 | struct RangeQueryClause *next; /* next in linked list */ |
| 37 | Node *var; /* The common variable of the clauses */ |
| 38 | bool have_lobound; /* found a low-bound clause yet? */ |
| 39 | bool have_hibound; /* found a high-bound clause yet? */ |
| 40 | Selectivity lobound; /* Selectivity of a var > something clause */ |
| 41 | Selectivity hibound; /* Selectivity of a var < something clause */ |
| 42 | } RangeQueryClause; |
| 43 | |
| 44 | static void addRangeClause(RangeQueryClause **rqlist, Node *clause, |
| 45 | bool varonleft, bool isLTsel, Selectivity s2); |
| 46 | static RelOptInfo *find_single_rel_for_clauses(PlannerInfo *root, |
| 47 | List *clauses); |
| 48 | |
| 49 | /**************************************************************************** |
| 50 | * ROUTINES TO COMPUTE SELECTIVITIES |
| 51 | ****************************************************************************/ |
| 52 | |
| 53 | /* |
| 54 | * clauselist_selectivity - |
| 55 | * Compute the selectivity of an implicitly-ANDed list of boolean |
| 56 | * expression clauses. The list can be empty, in which case 1.0 |
| 57 | * must be returned. List elements may be either RestrictInfos |
| 58 | * or bare expression clauses --- the former is preferred since |
| 59 | * it allows caching of results. |
| 60 | * |
| 61 | * See clause_selectivity() for the meaning of the additional parameters. |
| 62 | * |
| 63 | * The basic approach is to apply extended statistics first, on as many |
| 64 | * clauses as possible, in order to capture cross-column dependencies etc. |
| 65 | * The remaining clauses are then estimated using regular statistics tracked |
| 66 | * for individual columns. This is done by simply passing the clauses to |
| 67 | * clauselist_selectivity_simple. |
| 68 | */ |
| 69 | Selectivity |
| 70 | clauselist_selectivity(PlannerInfo *root, |
| 71 | List *clauses, |
| 72 | int varRelid, |
| 73 | JoinType jointype, |
| 74 | SpecialJoinInfo *sjinfo) |
| 75 | { |
| 76 | Selectivity s1 = 1.0; |
| 77 | RelOptInfo *rel; |
| 78 | Bitmapset *estimatedclauses = NULL; |
| 79 | |
| 80 | /* |
| 81 | * Determine if these clauses reference a single relation. If so, and if |
| 82 | * it has extended statistics, try to apply those. |
| 83 | */ |
| 84 | rel = find_single_rel_for_clauses(root, clauses); |
| 85 | if (rel && rel->rtekind == RTE_RELATION && rel->statlist != NIL) |
| 86 | { |
| 87 | /* |
| 88 | * Estimate as many clauses as possible using extended statistics. |
| 89 | * |
| 90 | * 'estimatedclauses' tracks the 0-based list position index of |
| 91 | * clauses that we've estimated using extended statistics, and that |
| 92 | * should be ignored. |
| 93 | */ |
| 94 | s1 *= statext_clauselist_selectivity(root, clauses, varRelid, |
| 95 | jointype, sjinfo, rel, |
| 96 | &estimatedclauses); |
| 97 | } |
| 98 | |
| 99 | /* |
| 100 | * Apply normal selectivity estimates for the remaining clauses, passing |
| 101 | * 'estimatedclauses' so that it skips already estimated ones. |
| 102 | */ |
| 103 | return s1 * clauselist_selectivity_simple(root, clauses, varRelid, |
| 104 | jointype, sjinfo, |
| 105 | estimatedclauses); |
| 106 | } |
| 107 | |
| 108 | /* |
| 109 | * clauselist_selectivity_simple - |
| 110 | * Compute the selectivity of an implicitly-ANDed list of boolean |
| 111 | * expression clauses. The list can be empty, in which case 1.0 |
| 112 | * must be returned. List elements may be either RestrictInfos |
| 113 | * or bare expression clauses --- the former is preferred since |
| 114 | * it allows caching of results. The estimatedclauses bitmap tracks |
| 115 | * clauses that have already been estimated by other means. |
| 116 | * |
| 117 | * See clause_selectivity() for the meaning of the additional parameters. |
| 118 | * |
| 119 | * Our basic approach is to take the product of the selectivities of the |
| 120 | * subclauses. However, that's only right if the subclauses have independent |
| 121 | * probabilities, and in reality they are often NOT independent. So, |
| 122 | * we want to be smarter where we can. |
| 123 | * |
| 124 | * We also recognize "range queries", such as "x > 34 AND x < 42". Clauses |
| 125 | * are recognized as possible range query components if they are restriction |
| 126 | * opclauses whose operators have scalarltsel or a related function as their |
| 127 | * restriction selectivity estimator. We pair up clauses of this form that |
| 128 | * refer to the same variable. An unpairable clause of this kind is simply |
| 129 | * multiplied into the selectivity product in the normal way. But when we |
| 130 | * find a pair, we know that the selectivities represent the relative |
| 131 | * positions of the low and high bounds within the column's range, so instead |
| 132 | * of figuring the selectivity as hisel * losel, we can figure it as hisel + |
| 133 | * losel - 1. (To visualize this, see that hisel is the fraction of the range |
| 134 | * below the high bound, while losel is the fraction above the low bound; so |
| 135 | * hisel can be interpreted directly as a 0..1 value but we need to convert |
| 136 | * losel to 1-losel before interpreting it as a value. Then the available |
| 137 | * range is 1-losel to hisel. However, this calculation double-excludes |
| 138 | * nulls, so really we need hisel + losel + null_frac - 1.) |
| 139 | * |
| 140 | * If either selectivity is exactly DEFAULT_INEQ_SEL, we forget this equation |
| 141 | * and instead use DEFAULT_RANGE_INEQ_SEL. The same applies if the equation |
| 142 | * yields an impossible (negative) result. |
| 143 | * |
| 144 | * A free side-effect is that we can recognize redundant inequalities such |
| 145 | * as "x < 4 AND x < 5"; only the tighter constraint will be counted. |
| 146 | * |
| 147 | * Of course this is all very dependent on the behavior of the inequality |
| 148 | * selectivity functions; perhaps some day we can generalize the approach. |
| 149 | */ |
| 150 | Selectivity |
| 151 | clauselist_selectivity_simple(PlannerInfo *root, |
| 152 | List *clauses, |
| 153 | int varRelid, |
| 154 | JoinType jointype, |
| 155 | SpecialJoinInfo *sjinfo, |
| 156 | Bitmapset *estimatedclauses) |
| 157 | { |
| 158 | Selectivity s1 = 1.0; |
| 159 | RangeQueryClause *rqlist = NULL; |
| 160 | ListCell *l; |
| 161 | int listidx; |
| 162 | |
| 163 | /* |
| 164 | * If there's exactly one clause (and it was not estimated yet), just go |
| 165 | * directly to clause_selectivity(). None of what we might do below is |
| 166 | * relevant. |
| 167 | */ |
| 168 | if ((list_length(clauses) == 1) && |
| 169 | bms_num_members(estimatedclauses) == 0) |
| 170 | return clause_selectivity(root, (Node *) linitial(clauses), |
| 171 | varRelid, jointype, sjinfo); |
| 172 | |
| 173 | /* |
| 174 | * Anything that doesn't look like a potential rangequery clause gets |
| 175 | * multiplied into s1 and forgotten. Anything that does gets inserted into |
| 176 | * an rqlist entry. |
| 177 | */ |
| 178 | listidx = -1; |
| 179 | foreach(l, clauses) |
| 180 | { |
| 181 | Node *clause = (Node *) lfirst(l); |
| 182 | RestrictInfo *rinfo; |
| 183 | Selectivity s2; |
| 184 | |
| 185 | listidx++; |
| 186 | |
| 187 | /* |
| 188 | * Skip this clause if it's already been estimated by some other |
| 189 | * statistics above. |
| 190 | */ |
| 191 | if (bms_is_member(listidx, estimatedclauses)) |
| 192 | continue; |
| 193 | |
| 194 | /* Always compute the selectivity using clause_selectivity */ |
| 195 | s2 = clause_selectivity(root, clause, varRelid, jointype, sjinfo); |
| 196 | |
| 197 | /* |
| 198 | * Check for being passed a RestrictInfo. |
| 199 | * |
| 200 | * If it's a pseudoconstant RestrictInfo, then s2 is either 1.0 or |
| 201 | * 0.0; just use that rather than looking for range pairs. |
| 202 | */ |
| 203 | if (IsA(clause, RestrictInfo)) |
| 204 | { |
| 205 | rinfo = (RestrictInfo *) clause; |
| 206 | if (rinfo->pseudoconstant) |
| 207 | { |
| 208 | s1 = s1 * s2; |
| 209 | continue; |
| 210 | } |
| 211 | clause = (Node *) rinfo->clause; |
| 212 | } |
| 213 | else |
| 214 | rinfo = NULL; |
| 215 | |
| 216 | /* |
| 217 | * See if it looks like a restriction clause with a pseudoconstant on |
| 218 | * one side. (Anything more complicated than that might not behave in |
| 219 | * the simple way we are expecting.) Most of the tests here can be |
| 220 | * done more efficiently with rinfo than without. |
| 221 | */ |
| 222 | if (is_opclause(clause) && list_length(((OpExpr *) clause)->args) == 2) |
| 223 | { |
| 224 | OpExpr *expr = (OpExpr *) clause; |
| 225 | bool varonleft = true; |
| 226 | bool ok; |
| 227 | |
| 228 | if (rinfo) |
| 229 | { |
| 230 | ok = (bms_membership(rinfo->clause_relids) == BMS_SINGLETON) && |
| 231 | (is_pseudo_constant_clause_relids(lsecond(expr->args), |
| 232 | rinfo->right_relids) || |
| 233 | (varonleft = false, |
| 234 | is_pseudo_constant_clause_relids(linitial(expr->args), |
| 235 | rinfo->left_relids))); |
| 236 | } |
| 237 | else |
| 238 | { |
| 239 | ok = (NumRelids(clause) == 1) && |
| 240 | (is_pseudo_constant_clause(lsecond(expr->args)) || |
| 241 | (varonleft = false, |
| 242 | is_pseudo_constant_clause(linitial(expr->args)))); |
| 243 | } |
| 244 | |
| 245 | if (ok) |
| 246 | { |
| 247 | /* |
| 248 | * If it's not a "<"/"<="/">"/">=" operator, just merge the |
| 249 | * selectivity in generically. But if it's the right oprrest, |
| 250 | * add the clause to rqlist for later processing. |
| 251 | */ |
| 252 | switch (get_oprrest(expr->opno)) |
| 253 | { |
| 254 | case F_SCALARLTSEL: |
| 255 | case F_SCALARLESEL: |
| 256 | addRangeClause(&rqlist, clause, |
| 257 | varonleft, true, s2); |
| 258 | break; |
| 259 | case F_SCALARGTSEL: |
| 260 | case F_SCALARGESEL: |
| 261 | addRangeClause(&rqlist, clause, |
| 262 | varonleft, false, s2); |
| 263 | break; |
| 264 | default: |
| 265 | /* Just merge the selectivity in generically */ |
| 266 | s1 = s1 * s2; |
| 267 | break; |
| 268 | } |
| 269 | continue; /* drop to loop bottom */ |
| 270 | } |
| 271 | } |
| 272 | |
| 273 | /* Not the right form, so treat it generically. */ |
| 274 | s1 = s1 * s2; |
| 275 | } |
| 276 | |
| 277 | /* |
| 278 | * Now scan the rangequery pair list. |
| 279 | */ |
| 280 | while (rqlist != NULL) |
| 281 | { |
| 282 | RangeQueryClause *rqnext; |
| 283 | |
| 284 | if (rqlist->have_lobound && rqlist->have_hibound) |
| 285 | { |
| 286 | /* Successfully matched a pair of range clauses */ |
| 287 | Selectivity s2; |
| 288 | |
| 289 | /* |
| 290 | * Exact equality to the default value probably means the |
| 291 | * selectivity function punted. This is not airtight but should |
| 292 | * be good enough. |
| 293 | */ |
| 294 | if (rqlist->hibound == DEFAULT_INEQ_SEL || |
| 295 | rqlist->lobound == DEFAULT_INEQ_SEL) |
| 296 | { |
| 297 | s2 = DEFAULT_RANGE_INEQ_SEL; |
| 298 | } |
| 299 | else |
| 300 | { |
| 301 | s2 = rqlist->hibound + rqlist->lobound - 1.0; |
| 302 | |
| 303 | /* Adjust for double-exclusion of NULLs */ |
| 304 | s2 += nulltestsel(root, IS_NULL, rqlist->var, |
| 305 | varRelid, jointype, sjinfo); |
| 306 | |
| 307 | /* |
| 308 | * A zero or slightly negative s2 should be converted into a |
| 309 | * small positive value; we probably are dealing with a very |
| 310 | * tight range and got a bogus result due to roundoff errors. |
| 311 | * However, if s2 is very negative, then we probably have |
| 312 | * default selectivity estimates on one or both sides of the |
| 313 | * range that we failed to recognize above for some reason. |
| 314 | */ |
| 315 | if (s2 <= 0.0) |
| 316 | { |
| 317 | if (s2 < -0.01) |
| 318 | { |
| 319 | /* |
| 320 | * No data available --- use a default estimate that |
| 321 | * is small, but not real small. |
| 322 | */ |
| 323 | s2 = DEFAULT_RANGE_INEQ_SEL; |
| 324 | } |
| 325 | else |
| 326 | { |
| 327 | /* |
| 328 | * It's just roundoff error; use a small positive |
| 329 | * value |
| 330 | */ |
| 331 | s2 = 1.0e-10; |
| 332 | } |
| 333 | } |
| 334 | } |
| 335 | /* Merge in the selectivity of the pair of clauses */ |
| 336 | s1 *= s2; |
| 337 | } |
| 338 | else |
| 339 | { |
| 340 | /* Only found one of a pair, merge it in generically */ |
| 341 | if (rqlist->have_lobound) |
| 342 | s1 *= rqlist->lobound; |
| 343 | else |
| 344 | s1 *= rqlist->hibound; |
| 345 | } |
| 346 | /* release storage and advance */ |
| 347 | rqnext = rqlist->next; |
| 348 | pfree(rqlist); |
| 349 | rqlist = rqnext; |
| 350 | } |
| 351 | |
| 352 | return s1; |
| 353 | } |
| 354 | |
| 355 | /* |
| 356 | * addRangeClause --- add a new range clause for clauselist_selectivity |
| 357 | * |
| 358 | * Here is where we try to match up pairs of range-query clauses |
| 359 | */ |
| 360 | static void |
| 361 | addRangeClause(RangeQueryClause **rqlist, Node *clause, |
| 362 | bool varonleft, bool isLTsel, Selectivity s2) |
| 363 | { |
| 364 | RangeQueryClause *rqelem; |
| 365 | Node *var; |
| 366 | bool is_lobound; |
| 367 | |
| 368 | if (varonleft) |
| 369 | { |
| 370 | var = get_leftop((Expr *) clause); |
| 371 | is_lobound = !isLTsel; /* x < something is high bound */ |
| 372 | } |
| 373 | else |
| 374 | { |
| 375 | var = get_rightop((Expr *) clause); |
| 376 | is_lobound = isLTsel; /* something < x is low bound */ |
| 377 | } |
| 378 | |
| 379 | for (rqelem = *rqlist; rqelem; rqelem = rqelem->next) |
| 380 | { |
| 381 | /* |
| 382 | * We use full equal() here because the "var" might be a function of |
| 383 | * one or more attributes of the same relation... |
| 384 | */ |
| 385 | if (!equal(var, rqelem->var)) |
| 386 | continue; |
| 387 | /* Found the right group to put this clause in */ |
| 388 | if (is_lobound) |
| 389 | { |
| 390 | if (!rqelem->have_lobound) |
| 391 | { |
| 392 | rqelem->have_lobound = true; |
| 393 | rqelem->lobound = s2; |
| 394 | } |
| 395 | else |
| 396 | { |
| 397 | |
| 398 | /*------ |
| 399 | * We have found two similar clauses, such as |
| 400 | * x < y AND x <= z. |
| 401 | * Keep only the more restrictive one. |
| 402 | *------ |
| 403 | */ |
| 404 | if (rqelem->lobound > s2) |
| 405 | rqelem->lobound = s2; |
| 406 | } |
| 407 | } |
| 408 | else |
| 409 | { |
| 410 | if (!rqelem->have_hibound) |
| 411 | { |
| 412 | rqelem->have_hibound = true; |
| 413 | rqelem->hibound = s2; |
| 414 | } |
| 415 | else |
| 416 | { |
| 417 | |
| 418 | /*------ |
| 419 | * We have found two similar clauses, such as |
| 420 | * x > y AND x >= z. |
| 421 | * Keep only the more restrictive one. |
| 422 | *------ |
| 423 | */ |
| 424 | if (rqelem->hibound > s2) |
| 425 | rqelem->hibound = s2; |
| 426 | } |
| 427 | } |
| 428 | return; |
| 429 | } |
| 430 | |
| 431 | /* No matching var found, so make a new clause-pair data structure */ |
| 432 | rqelem = (RangeQueryClause *) palloc(sizeof(RangeQueryClause)); |
| 433 | rqelem->var = var; |
| 434 | if (is_lobound) |
| 435 | { |
| 436 | rqelem->have_lobound = true; |
| 437 | rqelem->have_hibound = false; |
| 438 | rqelem->lobound = s2; |
| 439 | } |
| 440 | else |
| 441 | { |
| 442 | rqelem->have_lobound = false; |
| 443 | rqelem->have_hibound = true; |
| 444 | rqelem->hibound = s2; |
| 445 | } |
| 446 | rqelem->next = *rqlist; |
| 447 | *rqlist = rqelem; |
| 448 | } |
| 449 | |
| 450 | /* |
| 451 | * find_single_rel_for_clauses |
| 452 | * Examine each clause in 'clauses' and determine if all clauses |
| 453 | * reference only a single relation. If so return that relation, |
| 454 | * otherwise return NULL. |
| 455 | */ |
| 456 | static RelOptInfo * |
| 457 | find_single_rel_for_clauses(PlannerInfo *root, List *clauses) |
| 458 | { |
| 459 | int lastrelid = 0; |
| 460 | ListCell *l; |
| 461 | |
| 462 | foreach(l, clauses) |
| 463 | { |
| 464 | RestrictInfo *rinfo = (RestrictInfo *) lfirst(l); |
| 465 | int relid; |
| 466 | |
| 467 | /* |
| 468 | * If we have a list of bare clauses rather than RestrictInfos, we |
| 469 | * could pull out their relids the hard way with pull_varnos(). |
| 470 | * However, currently the extended-stats machinery won't do anything |
| 471 | * with non-RestrictInfo clauses anyway, so there's no point in |
| 472 | * spending extra cycles; just fail if that's what we have. |
| 473 | */ |
| 474 | if (!IsA(rinfo, RestrictInfo)) |
| 475 | return NULL; |
| 476 | |
| 477 | if (bms_is_empty(rinfo->clause_relids)) |
| 478 | continue; /* we can ignore variable-free clauses */ |
| 479 | if (!bms_get_singleton_member(rinfo->clause_relids, &relid)) |
| 480 | return NULL; /* multiple relations in this clause */ |
| 481 | if (lastrelid == 0) |
| 482 | lastrelid = relid; /* first clause referencing a relation */ |
| 483 | else if (relid != lastrelid) |
| 484 | return NULL; /* relation not same as last one */ |
| 485 | } |
| 486 | |
| 487 | if (lastrelid != 0) |
| 488 | return find_base_rel(root, lastrelid); |
| 489 | |
| 490 | return NULL; /* no clauses */ |
| 491 | } |
| 492 | |
| 493 | /* |
| 494 | * bms_is_subset_singleton |
| 495 | * |
| 496 | * Same result as bms_is_subset(s, bms_make_singleton(x)), |
| 497 | * but a little faster and doesn't leak memory. |
| 498 | * |
| 499 | * Is this of use anywhere else? If so move to bitmapset.c ... |
| 500 | */ |
| 501 | static bool |
| 502 | bms_is_subset_singleton(const Bitmapset *s, int x) |
| 503 | { |
| 504 | switch (bms_membership(s)) |
| 505 | { |
| 506 | case BMS_EMPTY_SET: |
| 507 | return true; |
| 508 | case BMS_SINGLETON: |
| 509 | return bms_is_member(x, s); |
| 510 | case BMS_MULTIPLE: |
| 511 | return false; |
| 512 | } |
| 513 | /* can't get here... */ |
| 514 | return false; |
| 515 | } |
| 516 | |
| 517 | /* |
| 518 | * treat_as_join_clause - |
| 519 | * Decide whether an operator clause is to be handled by the |
| 520 | * restriction or join estimator. Subroutine for clause_selectivity(). |
| 521 | */ |
| 522 | static inline bool |
| 523 | treat_as_join_clause(Node *clause, RestrictInfo *rinfo, |
| 524 | int varRelid, SpecialJoinInfo *sjinfo) |
| 525 | { |
| 526 | if (varRelid != 0) |
| 527 | { |
| 528 | /* |
| 529 | * Caller is forcing restriction mode (eg, because we are examining an |
| 530 | * inner indexscan qual). |
| 531 | */ |
| 532 | return false; |
| 533 | } |
| 534 | else if (sjinfo == NULL) |
| 535 | { |
| 536 | /* |
| 537 | * It must be a restriction clause, since it's being evaluated at a |
| 538 | * scan node. |
| 539 | */ |
| 540 | return false; |
| 541 | } |
| 542 | else |
| 543 | { |
| 544 | /* |
| 545 | * Otherwise, it's a join if there's more than one relation used. We |
| 546 | * can optimize this calculation if an rinfo was passed. |
| 547 | * |
| 548 | * XXX Since we know the clause is being evaluated at a join, the |
| 549 | * only way it could be single-relation is if it was delayed by outer |
| 550 | * joins. Although we can make use of the restriction qual estimators |
| 551 | * anyway, it seems likely that we ought to account for the |
| 552 | * probability of injected nulls somehow. |
| 553 | */ |
| 554 | if (rinfo) |
| 555 | return (bms_membership(rinfo->clause_relids) == BMS_MULTIPLE); |
| 556 | else |
| 557 | return (NumRelids(clause) > 1); |
| 558 | } |
| 559 | } |
| 560 | |
| 561 | |
| 562 | /* |
| 563 | * clause_selectivity - |
| 564 | * Compute the selectivity of a general boolean expression clause. |
| 565 | * |
| 566 | * The clause can be either a RestrictInfo or a plain expression. If it's |
| 567 | * a RestrictInfo, we try to cache the selectivity for possible re-use, |
| 568 | * so passing RestrictInfos is preferred. |
| 569 | * |
| 570 | * varRelid is either 0 or a rangetable index. |
| 571 | * |
| 572 | * When varRelid is not 0, only variables belonging to that relation are |
| 573 | * considered in computing selectivity; other vars are treated as constants |
| 574 | * of unknown values. This is appropriate for estimating the selectivity of |
| 575 | * a join clause that is being used as a restriction clause in a scan of a |
| 576 | * nestloop join's inner relation --- varRelid should then be the ID of the |
| 577 | * inner relation. |
| 578 | * |
| 579 | * When varRelid is 0, all variables are treated as variables. This |
| 580 | * is appropriate for ordinary join clauses and restriction clauses. |
| 581 | * |
| 582 | * jointype is the join type, if the clause is a join clause. Pass JOIN_INNER |
| 583 | * if the clause isn't a join clause. |
| 584 | * |
| 585 | * sjinfo is NULL for a non-join clause, otherwise it provides additional |
| 586 | * context information about the join being performed. There are some |
| 587 | * special cases: |
| 588 | * 1. For a special (not INNER) join, sjinfo is always a member of |
| 589 | * root->join_info_list. |
| 590 | * 2. For an INNER join, sjinfo is just a transient struct, and only the |
| 591 | * relids and jointype fields in it can be trusted. |
| 592 | * It is possible for jointype to be different from sjinfo->jointype. |
| 593 | * This indicates we are considering a variant join: either with |
| 594 | * the LHS and RHS switched, or with one input unique-ified. |
| 595 | * |
| 596 | * Note: when passing nonzero varRelid, it's normally appropriate to set |
| 597 | * jointype == JOIN_INNER, sjinfo == NULL, even if the clause is really a |
| 598 | * join clause; because we aren't treating it as a join clause. |
| 599 | */ |
| 600 | Selectivity |
| 601 | clause_selectivity(PlannerInfo *root, |
| 602 | Node *clause, |
| 603 | int varRelid, |
| 604 | JoinType jointype, |
| 605 | SpecialJoinInfo *sjinfo) |
| 606 | { |
| 607 | Selectivity s1 = 0.5; /* default for any unhandled clause type */ |
| 608 | RestrictInfo *rinfo = NULL; |
| 609 | bool cacheable = false; |
| 610 | |
| 611 | if (clause == NULL) /* can this still happen? */ |
| 612 | return s1; |
| 613 | |
| 614 | if (IsA(clause, RestrictInfo)) |
| 615 | { |
| 616 | rinfo = (RestrictInfo *) clause; |
| 617 | |
| 618 | /* |
| 619 | * If the clause is marked pseudoconstant, then it will be used as a |
| 620 | * gating qual and should not affect selectivity estimates; hence |
| 621 | * return 1.0. The only exception is that a constant FALSE may be |
| 622 | * taken as having selectivity 0.0, since it will surely mean no rows |
| 623 | * out of the plan. This case is simple enough that we need not |
| 624 | * bother caching the result. |
| 625 | */ |
| 626 | if (rinfo->pseudoconstant) |
| 627 | { |
| 628 | if (!IsA(rinfo->clause, Const)) |
| 629 | return (Selectivity) 1.0; |
| 630 | } |
| 631 | |
| 632 | /* |
| 633 | * If the clause is marked redundant, always return 1.0. |
| 634 | */ |
| 635 | if (rinfo->norm_selec > 1) |
| 636 | return (Selectivity) 1.0; |
| 637 | |
| 638 | /* |
| 639 | * If possible, cache the result of the selectivity calculation for |
| 640 | * the clause. We can cache if varRelid is zero or the clause |
| 641 | * contains only vars of that relid --- otherwise varRelid will affect |
| 642 | * the result, so mustn't cache. Outer join quals might be examined |
| 643 | * with either their join's actual jointype or JOIN_INNER, so we need |
| 644 | * two cache variables to remember both cases. Note: we assume the |
| 645 | * result won't change if we are switching the input relations or |
| 646 | * considering a unique-ified case, so we only need one cache variable |
| 647 | * for all non-JOIN_INNER cases. |
| 648 | */ |
| 649 | if (varRelid == 0 || |
| 650 | bms_is_subset_singleton(rinfo->clause_relids, varRelid)) |
| 651 | { |
| 652 | /* Cacheable --- do we already have the result? */ |
| 653 | if (jointype == JOIN_INNER) |
| 654 | { |
| 655 | if (rinfo->norm_selec >= 0) |
| 656 | return rinfo->norm_selec; |
| 657 | } |
| 658 | else |
| 659 | { |
| 660 | if (rinfo->outer_selec >= 0) |
| 661 | return rinfo->outer_selec; |
| 662 | } |
| 663 | cacheable = true; |
| 664 | } |
| 665 | |
| 666 | /* |
| 667 | * Proceed with examination of contained clause. If the clause is an |
| 668 | * OR-clause, we want to look at the variant with sub-RestrictInfos, |
| 669 | * so that per-subclause selectivities can be cached. |
| 670 | */ |
| 671 | if (rinfo->orclause) |
| 672 | clause = (Node *) rinfo->orclause; |
| 673 | else |
| 674 | clause = (Node *) rinfo->clause; |
| 675 | } |
| 676 | |
| 677 | if (IsA(clause, Var)) |
| 678 | { |
| 679 | Var *var = (Var *) clause; |
| 680 | |
| 681 | /* |
| 682 | * We probably shouldn't ever see an uplevel Var here, but if we do, |
| 683 | * return the default selectivity... |
| 684 | */ |
| 685 | if (var->varlevelsup == 0 && |
| 686 | (varRelid == 0 || varRelid == (int) var->varno)) |
| 687 | { |
| 688 | /* Use the restriction selectivity function for a bool Var */ |
| 689 | s1 = boolvarsel(root, (Node *) var, varRelid); |
| 690 | } |
| 691 | } |
| 692 | else if (IsA(clause, Const)) |
| 693 | { |
| 694 | /* bool constant is pretty easy... */ |
| 695 | Const *con = (Const *) clause; |
| 696 | |
| 697 | s1 = con->constisnull ? 0.0 : |
| 698 | DatumGetBool(con->constvalue) ? 1.0 : 0.0; |
| 699 | } |
| 700 | else if (IsA(clause, Param)) |
| 701 | { |
| 702 | /* see if we can replace the Param */ |
| 703 | Node *subst = estimate_expression_value(root, clause); |
| 704 | |
| 705 | if (IsA(subst, Const)) |
| 706 | { |
| 707 | /* bool constant is pretty easy... */ |
| 708 | Const *con = (Const *) subst; |
| 709 | |
| 710 | s1 = con->constisnull ? 0.0 : |
| 711 | DatumGetBool(con->constvalue) ? 1.0 : 0.0; |
| 712 | } |
| 713 | else |
| 714 | { |
| 715 | /* XXX any way to do better than default? */ |
| 716 | } |
| 717 | } |
| 718 | else if (is_notclause(clause)) |
| 719 | { |
| 720 | /* inverse of the selectivity of the underlying clause */ |
| 721 | s1 = 1.0 - clause_selectivity(root, |
| 722 | (Node *) get_notclausearg((Expr *) clause), |
| 723 | varRelid, |
| 724 | jointype, |
| 725 | sjinfo); |
| 726 | } |
| 727 | else if (is_andclause(clause)) |
| 728 | { |
| 729 | /* share code with clauselist_selectivity() */ |
| 730 | s1 = clauselist_selectivity(root, |
| 731 | ((BoolExpr *) clause)->args, |
| 732 | varRelid, |
| 733 | jointype, |
| 734 | sjinfo); |
| 735 | } |
| 736 | else if (is_orclause(clause)) |
| 737 | { |
| 738 | /* |
| 739 | * Selectivities for an OR clause are computed as s1+s2 - s1*s2 to |
| 740 | * account for the probable overlap of selected tuple sets. |
| 741 | * |
| 742 | * XXX is this too conservative? |
| 743 | */ |
| 744 | ListCell *arg; |
| 745 | |
| 746 | s1 = 0.0; |
| 747 | foreach(arg, ((BoolExpr *) clause)->args) |
| 748 | { |
| 749 | Selectivity s2 = clause_selectivity(root, |
| 750 | (Node *) lfirst(arg), |
| 751 | varRelid, |
| 752 | jointype, |
| 753 | sjinfo); |
| 754 | |
| 755 | s1 = s1 + s2 - s1 * s2; |
| 756 | } |
| 757 | } |
| 758 | else if (is_opclause(clause) || IsA(clause, DistinctExpr)) |
| 759 | { |
| 760 | OpExpr *opclause = (OpExpr *) clause; |
| 761 | Oid opno = opclause->opno; |
| 762 | |
| 763 | if (treat_as_join_clause(clause, rinfo, varRelid, sjinfo)) |
| 764 | { |
| 765 | /* Estimate selectivity for a join clause. */ |
| 766 | s1 = join_selectivity(root, opno, |
| 767 | opclause->args, |
| 768 | opclause->inputcollid, |
| 769 | jointype, |
| 770 | sjinfo); |
| 771 | } |
| 772 | else |
| 773 | { |
| 774 | /* Estimate selectivity for a restriction clause. */ |
| 775 | s1 = restriction_selectivity(root, opno, |
| 776 | opclause->args, |
| 777 | opclause->inputcollid, |
| 778 | varRelid); |
| 779 | } |
| 780 | |
| 781 | /* |
| 782 | * DistinctExpr has the same representation as OpExpr, but the |
| 783 | * contained operator is "=" not "<>", so we must negate the result. |
| 784 | * This estimation method doesn't give the right behavior for nulls, |
| 785 | * but it's better than doing nothing. |
| 786 | */ |
| 787 | if (IsA(clause, DistinctExpr)) |
| 788 | s1 = 1.0 - s1; |
| 789 | } |
| 790 | else if (is_funcclause(clause)) |
| 791 | { |
| 792 | FuncExpr *funcclause = (FuncExpr *) clause; |
| 793 | |
| 794 | /* Try to get an estimate from the support function, if any */ |
| 795 | s1 = function_selectivity(root, |
| 796 | funcclause->funcid, |
| 797 | funcclause->args, |
| 798 | funcclause->inputcollid, |
| 799 | treat_as_join_clause(clause, rinfo, |
| 800 | varRelid, sjinfo), |
| 801 | varRelid, |
| 802 | jointype, |
| 803 | sjinfo); |
| 804 | } |
| 805 | else if (IsA(clause, ScalarArrayOpExpr)) |
| 806 | { |
| 807 | /* Use node specific selectivity calculation function */ |
| 808 | s1 = scalararraysel(root, |
| 809 | (ScalarArrayOpExpr *) clause, |
| 810 | treat_as_join_clause(clause, rinfo, |
| 811 | varRelid, sjinfo), |
| 812 | varRelid, |
| 813 | jointype, |
| 814 | sjinfo); |
| 815 | } |
| 816 | else if (IsA(clause, RowCompareExpr)) |
| 817 | { |
| 818 | /* Use node specific selectivity calculation function */ |
| 819 | s1 = rowcomparesel(root, |
| 820 | (RowCompareExpr *) clause, |
| 821 | varRelid, |
| 822 | jointype, |
| 823 | sjinfo); |
| 824 | } |
| 825 | else if (IsA(clause, NullTest)) |
| 826 | { |
| 827 | /* Use node specific selectivity calculation function */ |
| 828 | s1 = nulltestsel(root, |
| 829 | ((NullTest *) clause)->nulltesttype, |
| 830 | (Node *) ((NullTest *) clause)->arg, |
| 831 | varRelid, |
| 832 | jointype, |
| 833 | sjinfo); |
| 834 | } |
| 835 | else if (IsA(clause, BooleanTest)) |
| 836 | { |
| 837 | /* Use node specific selectivity calculation function */ |
| 838 | s1 = booltestsel(root, |
| 839 | ((BooleanTest *) clause)->booltesttype, |
| 840 | (Node *) ((BooleanTest *) clause)->arg, |
| 841 | varRelid, |
| 842 | jointype, |
| 843 | sjinfo); |
| 844 | } |
| 845 | else if (IsA(clause, CurrentOfExpr)) |
| 846 | { |
| 847 | /* CURRENT OF selects at most one row of its table */ |
| 848 | CurrentOfExpr *cexpr = (CurrentOfExpr *) clause; |
| 849 | RelOptInfo *crel = find_base_rel(root, cexpr->cvarno); |
| 850 | |
| 851 | if (crel->tuples > 0) |
| 852 | s1 = 1.0 / crel->tuples; |
| 853 | } |
| 854 | else if (IsA(clause, RelabelType)) |
| 855 | { |
| 856 | /* Not sure this case is needed, but it can't hurt */ |
| 857 | s1 = clause_selectivity(root, |
| 858 | (Node *) ((RelabelType *) clause)->arg, |
| 859 | varRelid, |
| 860 | jointype, |
| 861 | sjinfo); |
| 862 | } |
| 863 | else if (IsA(clause, CoerceToDomain)) |
| 864 | { |
| 865 | /* Not sure this case is needed, but it can't hurt */ |
| 866 | s1 = clause_selectivity(root, |
| 867 | (Node *) ((CoerceToDomain *) clause)->arg, |
| 868 | varRelid, |
| 869 | jointype, |
| 870 | sjinfo); |
| 871 | } |
| 872 | else |
| 873 | { |
| 874 | /* |
| 875 | * For anything else, see if we can consider it as a boolean variable. |
| 876 | * This only works if it's an immutable expression in Vars of a single |
| 877 | * relation; but there's no point in us checking that here because |
| 878 | * boolvarsel() will do it internally, and return a suitable default |
| 879 | * selectivity if not. |
| 880 | */ |
| 881 | s1 = boolvarsel(root, clause, varRelid); |
| 882 | } |
| 883 | |
| 884 | /* Cache the result if possible */ |
| 885 | if (cacheable) |
| 886 | { |
| 887 | if (jointype == JOIN_INNER) |
| 888 | rinfo->norm_selec = s1; |
| 889 | else |
| 890 | rinfo->outer_selec = s1; |
| 891 | } |
| 892 | |
| 893 | #ifdef SELECTIVITY_DEBUG |
| 894 | elog(DEBUG4, "clause_selectivity: s1 %f" , s1); |
| 895 | #endif /* SELECTIVITY_DEBUG */ |
| 896 | |
| 897 | return s1; |
| 898 | } |
| 899 | |