Line data Source code
1 : /*-------------------------------------------------------------------------
2 : *
3 : * costsize.c
4 : * Routines to compute (and set) relation sizes and path costs
5 : *
6 : * Path costs are measured in arbitrary units established by these basic
7 : * parameters:
8 : *
9 : * seq_page_cost Cost of a sequential page fetch
10 : * random_page_cost Cost of a non-sequential page fetch
11 : * cpu_tuple_cost Cost of typical CPU time to process a tuple
12 : * cpu_index_tuple_cost Cost of typical CPU time to process an index tuple
13 : * cpu_operator_cost Cost of CPU time to execute an operator or function
14 : * parallel_tuple_cost Cost of CPU time to pass a tuple from worker to leader backend
15 : * parallel_setup_cost Cost of setting up shared memory for parallelism
16 : *
17 : * We expect that the kernel will typically do some amount of read-ahead
18 : * optimization; this in conjunction with seek costs means that seq_page_cost
19 : * is normally considerably less than random_page_cost. (However, if the
20 : * database is fully cached in RAM, it is reasonable to set them equal.)
21 : *
22 : * We also use a rough estimate "effective_cache_size" of the number of
23 : * disk pages in Postgres + OS-level disk cache. (We can't simply use
24 : * NBuffers for this purpose because that would ignore the effects of
25 : * the kernel's disk cache.)
26 : *
27 : * Obviously, taking constants for these values is an oversimplification,
28 : * but it's tough enough to get any useful estimates even at this level of
29 : * detail. Note that all of these parameters are user-settable, in case
30 : * the default values are drastically off for a particular platform.
31 : *
32 : * seq_page_cost and random_page_cost can also be overridden for an individual
33 : * tablespace, in case some data is on a fast disk and other data is on a slow
34 : * disk. Per-tablespace overrides never apply to temporary work files such as
35 : * an external sort or a materialize node that overflows work_mem.
36 : *
37 : * We compute two separate costs for each path:
38 : * total_cost: total estimated cost to fetch all tuples
39 : * startup_cost: cost that is expended before first tuple is fetched
40 : * In some scenarios, such as when there is a LIMIT or we are implementing
41 : * an EXISTS(...) sub-select, it is not necessary to fetch all tuples of the
42 : * path's result. A caller can estimate the cost of fetching a partial
43 : * result by interpolating between startup_cost and total_cost. In detail:
44 : * actual_cost = startup_cost +
45 : * (total_cost - startup_cost) * tuples_to_fetch / path->rows;
46 : * Note that a base relation's rows count (and, by extension, plan_rows for
47 : * plan nodes below the LIMIT node) are set without regard to any LIMIT, so
48 : * that this equation works properly. (Note: while path->rows is never zero
49 : * for ordinary relations, it is zero for paths for provably-empty relations,
50 : * so beware of division-by-zero.) The LIMIT is applied as a top-level
51 : * plan node.
52 : *
53 : * Each path stores the total number of disabled nodes that exist at or
54 : * below that point in the plan tree. This is regarded as a component of
55 : * the cost, and paths with fewer disabled nodes should be regarded as
56 : * cheaper than those with more. Disabled nodes occur when the user sets
57 : * a GUC like enable_seqscan=false. We can't necessarily respect such a
58 : * setting in every part of the plan tree, but we want to respect in as many
59 : * parts of the plan tree as possible. Simpler schemes like storing a Boolean
60 : * here rather than a count fail to do that. We used to disable nodes by
61 : * adding a large constant to the startup cost, but that distorted planning
62 : * in other ways.
63 : *
64 : * For largely historical reasons, most of the routines in this module use
65 : * the passed result Path only to store their results (rows, startup_cost and
66 : * total_cost) into. All the input data they need is passed as separate
67 : * parameters, even though much of it could be extracted from the Path.
68 : * An exception is made for the cost_XXXjoin() routines, which expect all
69 : * the other fields of the passed XXXPath to be filled in, and similarly
70 : * cost_index() assumes the passed IndexPath is valid except for its output
71 : * values.
72 : *
73 : *
74 : * Portions Copyright (c) 1996-2026, PostgreSQL Global Development Group
75 : * Portions Copyright (c) 1994, Regents of the University of California
76 : *
77 : * IDENTIFICATION
78 : * src/backend/optimizer/path/costsize.c
79 : *
80 : *-------------------------------------------------------------------------
81 : */
82 :
83 : #include "postgres.h"
84 :
85 : #include <limits.h>
86 : #include <math.h>
87 :
88 : #include "access/amapi.h"
89 : #include "access/htup_details.h"
90 : #include "access/tsmapi.h"
91 : #include "executor/executor.h"
92 : #include "executor/nodeAgg.h"
93 : #include "executor/nodeHash.h"
94 : #include "executor/nodeMemoize.h"
95 : #include "miscadmin.h"
96 : #include "nodes/makefuncs.h"
97 : #include "nodes/nodeFuncs.h"
98 : #include "optimizer/clauses.h"
99 : #include "optimizer/cost.h"
100 : #include "optimizer/optimizer.h"
101 : #include "optimizer/pathnode.h"
102 : #include "optimizer/paths.h"
103 : #include "optimizer/placeholder.h"
104 : #include "optimizer/plancat.h"
105 : #include "optimizer/restrictinfo.h"
106 : #include "parser/parsetree.h"
107 : #include "utils/lsyscache.h"
108 : #include "utils/selfuncs.h"
109 : #include "utils/spccache.h"
110 : #include "utils/tuplesort.h"
111 :
112 :
113 : #define LOG2(x) (log(x) / 0.693147180559945)
114 :
115 : /*
116 : * Append and MergeAppend nodes are less expensive than some other operations
117 : * which use cpu_tuple_cost; instead of adding a separate GUC, estimate the
118 : * per-tuple cost as cpu_tuple_cost multiplied by this value.
119 : */
120 : #define APPEND_CPU_COST_MULTIPLIER 0.5
121 :
122 : /*
123 : * Maximum value for row estimates. We cap row estimates to this to help
124 : * ensure that costs based on these estimates remain within the range of what
125 : * double can represent. add_path() wouldn't act sanely given infinite or NaN
126 : * cost values.
127 : */
128 : #define MAXIMUM_ROWCOUNT 1e100
129 :
130 : double seq_page_cost = DEFAULT_SEQ_PAGE_COST;
131 : double random_page_cost = DEFAULT_RANDOM_PAGE_COST;
132 : double cpu_tuple_cost = DEFAULT_CPU_TUPLE_COST;
133 : double cpu_index_tuple_cost = DEFAULT_CPU_INDEX_TUPLE_COST;
134 : double cpu_operator_cost = DEFAULT_CPU_OPERATOR_COST;
135 : double parallel_tuple_cost = DEFAULT_PARALLEL_TUPLE_COST;
136 : double parallel_setup_cost = DEFAULT_PARALLEL_SETUP_COST;
137 : double recursive_worktable_factor = DEFAULT_RECURSIVE_WORKTABLE_FACTOR;
138 :
139 : int effective_cache_size = DEFAULT_EFFECTIVE_CACHE_SIZE;
140 :
141 : Cost disable_cost = 1.0e10;
142 :
143 : int max_parallel_workers_per_gather = 2;
144 :
145 : bool enable_seqscan = true;
146 : bool enable_indexscan = true;
147 : bool enable_indexonlyscan = true;
148 : bool enable_bitmapscan = true;
149 : bool enable_tidscan = true;
150 : bool enable_sort = true;
151 : bool enable_incremental_sort = true;
152 : bool enable_hashagg = true;
153 : bool enable_nestloop = true;
154 : bool enable_material = true;
155 : bool enable_memoize = true;
156 : bool enable_mergejoin = true;
157 : bool enable_hashjoin = true;
158 : bool enable_gathermerge = true;
159 : bool enable_partitionwise_join = false;
160 : bool enable_partitionwise_aggregate = false;
161 : bool enable_parallel_append = true;
162 : bool enable_parallel_hash = true;
163 : bool enable_partition_pruning = true;
164 : bool enable_presorted_aggregate = true;
165 : bool enable_async_append = true;
166 :
167 : typedef struct
168 : {
169 : PlannerInfo *root;
170 : QualCost total;
171 : } cost_qual_eval_context;
172 :
173 : static List *extract_nonindex_conditions(List *qual_clauses, List *indexclauses);
174 : static MergeScanSelCache *cached_scansel(PlannerInfo *root,
175 : RestrictInfo *rinfo,
176 : PathKey *pathkey);
177 : static void cost_rescan(PlannerInfo *root, Path *path,
178 : Cost *rescan_startup_cost, Cost *rescan_total_cost);
179 : static bool cost_qual_eval_walker(Node *node, cost_qual_eval_context *context);
180 : static void get_restriction_qual_cost(PlannerInfo *root, RelOptInfo *baserel,
181 : ParamPathInfo *param_info,
182 : QualCost *qpqual_cost);
183 : static bool has_indexed_join_quals(NestPath *path);
184 : static double approx_tuple_count(PlannerInfo *root, JoinPath *path,
185 : List *quals);
186 : static double calc_joinrel_size_estimate(PlannerInfo *root,
187 : RelOptInfo *joinrel,
188 : RelOptInfo *outer_rel,
189 : RelOptInfo *inner_rel,
190 : double outer_rows,
191 : double inner_rows,
192 : SpecialJoinInfo *sjinfo,
193 : List *restrictlist);
194 : static Selectivity get_foreign_key_join_selectivity(PlannerInfo *root,
195 : Relids outer_relids,
196 : Relids inner_relids,
197 : SpecialJoinInfo *sjinfo,
198 : List **restrictlist);
199 : static Cost append_nonpartial_cost(List *subpaths, int numpaths,
200 : int parallel_workers);
201 : static void set_rel_width(PlannerInfo *root, RelOptInfo *rel);
202 : static int32 get_expr_width(PlannerInfo *root, const Node *expr);
203 : static double relation_byte_size(double tuples, int width);
204 : static double page_size(double tuples, int width);
205 : static double get_parallel_divisor(Path *path);
206 :
207 :
208 : /*
209 : * clamp_row_est
210 : * Force a row-count estimate to a sane value.
211 : */
212 : double
213 10779170 : clamp_row_est(double nrows)
214 : {
215 : /*
216 : * Avoid infinite and NaN row estimates. Costs derived from such values
217 : * are going to be useless. Also force the estimate to be at least one
218 : * row, to make explain output look better and to avoid possible
219 : * divide-by-zero when interpolating costs. Make it an integer, too.
220 : */
221 10779170 : if (nrows > MAXIMUM_ROWCOUNT || isnan(nrows))
222 0 : nrows = MAXIMUM_ROWCOUNT;
223 10779170 : else if (nrows <= 1.0)
224 3395402 : nrows = 1.0;
225 : else
226 7383768 : nrows = rint(nrows);
227 :
228 10779170 : return nrows;
229 : }
230 :
231 : /*
232 : * clamp_width_est
233 : * Force a tuple-width estimate to a sane value.
234 : *
235 : * The planner represents datatype width and tuple width estimates as int32.
236 : * When summing column width estimates to create a tuple width estimate,
237 : * it's possible to reach integer overflow in edge cases. To ensure sane
238 : * behavior, we form such sums in int64 arithmetic and then apply this routine
239 : * to clamp to int32 range.
240 : */
241 : int32
242 2022724 : clamp_width_est(int64 tuple_width)
243 : {
244 : /*
245 : * Anything more than MaxAllocSize is clearly bogus, since we could not
246 : * create a tuple that large.
247 : */
248 2022724 : if (tuple_width > MaxAllocSize)
249 0 : return (int32) MaxAllocSize;
250 :
251 : /*
252 : * Unlike clamp_row_est, we just Assert that the value isn't negative,
253 : * rather than masking such errors.
254 : */
255 : Assert(tuple_width >= 0);
256 :
257 2022724 : return (int32) tuple_width;
258 : }
259 :
260 :
261 : /*
262 : * cost_seqscan
263 : * Determines and returns the cost of scanning a relation sequentially.
264 : *
265 : * 'baserel' is the relation to be scanned
266 : * 'param_info' is the ParamPathInfo if this is a parameterized path, else NULL
267 : */
268 : void
269 448482 : cost_seqscan(Path *path, PlannerInfo *root,
270 : RelOptInfo *baserel, ParamPathInfo *param_info)
271 : {
272 448482 : Cost startup_cost = 0;
273 : Cost cpu_run_cost;
274 : Cost disk_run_cost;
275 : double spc_seq_page_cost;
276 : QualCost qpqual_cost;
277 : Cost cpu_per_tuple;
278 448482 : uint64 enable_mask = PGS_SEQSCAN;
279 :
280 : /* Should only be applied to base relations */
281 : Assert(baserel->relid > 0);
282 : Assert(baserel->rtekind == RTE_RELATION);
283 :
284 : /* Mark the path with the correct row estimate */
285 448482 : if (param_info)
286 840 : path->rows = param_info->ppi_rows;
287 : else
288 447642 : path->rows = baserel->rows;
289 :
290 : /* fetch estimated page cost for tablespace containing table */
291 448482 : get_tablespace_page_costs(baserel->reltablespace,
292 : NULL,
293 : &spc_seq_page_cost);
294 :
295 : /*
296 : * disk costs
297 : */
298 448482 : disk_run_cost = spc_seq_page_cost * baserel->pages;
299 :
300 : /* CPU costs */
301 448482 : get_restriction_qual_cost(root, baserel, param_info, &qpqual_cost);
302 :
303 448482 : startup_cost += qpqual_cost.startup;
304 448482 : cpu_per_tuple = cpu_tuple_cost + qpqual_cost.per_tuple;
305 448482 : cpu_run_cost = cpu_per_tuple * baserel->tuples;
306 : /* tlist eval costs are paid per output row, not per tuple scanned */
307 448482 : startup_cost += path->pathtarget->cost.startup;
308 448482 : cpu_run_cost += path->pathtarget->cost.per_tuple * path->rows;
309 :
310 : /* Adjust costing for parallelism, if used. */
311 448482 : if (path->parallel_workers > 0)
312 : {
313 29092 : double parallel_divisor = get_parallel_divisor(path);
314 :
315 : /* The CPU cost is divided among all the workers. */
316 29092 : cpu_run_cost /= parallel_divisor;
317 :
318 : /*
319 : * It may be possible to amortize some of the I/O cost, but probably
320 : * not very much, because most operating systems already do aggressive
321 : * prefetching. For now, we assume that the disk run cost can't be
322 : * amortized at all.
323 : */
324 :
325 : /*
326 : * In the case of a parallel plan, the row count needs to represent
327 : * the number of tuples processed per worker.
328 : */
329 29092 : path->rows = clamp_row_est(path->rows / parallel_divisor);
330 : }
331 : else
332 419390 : enable_mask |= PGS_CONSIDER_NONPARTIAL;
333 :
334 448482 : path->disabled_nodes =
335 448482 : (baserel->pgs_mask & enable_mask) == enable_mask ? 0 : 1;
336 448482 : path->startup_cost = startup_cost;
337 448482 : path->total_cost = startup_cost + cpu_run_cost + disk_run_cost;
338 448482 : }
339 :
340 : /*
341 : * cost_samplescan
342 : * Determines and returns the cost of scanning a relation using sampling.
343 : *
344 : * 'baserel' is the relation to be scanned
345 : * 'param_info' is the ParamPathInfo if this is a parameterized path, else NULL
346 : */
347 : void
348 306 : cost_samplescan(Path *path, PlannerInfo *root,
349 : RelOptInfo *baserel, ParamPathInfo *param_info)
350 : {
351 306 : Cost startup_cost = 0;
352 306 : Cost run_cost = 0;
353 : RangeTblEntry *rte;
354 : TableSampleClause *tsc;
355 : TsmRoutine *tsm;
356 : double spc_seq_page_cost,
357 : spc_random_page_cost,
358 : spc_page_cost;
359 : QualCost qpqual_cost;
360 : Cost cpu_per_tuple;
361 306 : uint64 enable_mask = 0;
362 :
363 : /* Should only be applied to base relations with tablesample clauses */
364 : Assert(baserel->relid > 0);
365 306 : rte = planner_rt_fetch(baserel->relid, root);
366 : Assert(rte->rtekind == RTE_RELATION);
367 306 : tsc = rte->tablesample;
368 : Assert(tsc != NULL);
369 306 : tsm = GetTsmRoutine(tsc->tsmhandler);
370 :
371 : /* Mark the path with the correct row estimate */
372 306 : if (param_info)
373 72 : path->rows = param_info->ppi_rows;
374 : else
375 234 : path->rows = baserel->rows;
376 :
377 : /* fetch estimated page cost for tablespace containing table */
378 306 : get_tablespace_page_costs(baserel->reltablespace,
379 : &spc_random_page_cost,
380 : &spc_seq_page_cost);
381 :
382 : /* if NextSampleBlock is used, assume random access, else sequential */
383 612 : spc_page_cost = (tsm->NextSampleBlock != NULL) ?
384 306 : spc_random_page_cost : spc_seq_page_cost;
385 :
386 : /*
387 : * disk costs (recall that baserel->pages has already been set to the
388 : * number of pages the sampling method will visit)
389 : */
390 306 : run_cost += spc_page_cost * baserel->pages;
391 :
392 : /*
393 : * CPU costs (recall that baserel->tuples has already been set to the
394 : * number of tuples the sampling method will select). Note that we ignore
395 : * execution cost of the TABLESAMPLE parameter expressions; they will be
396 : * evaluated only once per scan, and in most usages they'll likely be
397 : * simple constants anyway. We also don't charge anything for the
398 : * calculations the sampling method might do internally.
399 : */
400 306 : get_restriction_qual_cost(root, baserel, param_info, &qpqual_cost);
401 :
402 306 : startup_cost += qpqual_cost.startup;
403 306 : cpu_per_tuple = cpu_tuple_cost + qpqual_cost.per_tuple;
404 306 : run_cost += cpu_per_tuple * baserel->tuples;
405 : /* tlist eval costs are paid per output row, not per tuple scanned */
406 306 : startup_cost += path->pathtarget->cost.startup;
407 306 : run_cost += path->pathtarget->cost.per_tuple * path->rows;
408 :
409 306 : if (path->parallel_workers == 0)
410 306 : enable_mask |= PGS_CONSIDER_NONPARTIAL;
411 :
412 306 : path->disabled_nodes =
413 306 : (baserel->pgs_mask & enable_mask) == enable_mask ? 0 : 1;
414 306 : path->startup_cost = startup_cost;
415 306 : path->total_cost = startup_cost + run_cost;
416 306 : }
417 :
418 : /*
419 : * cost_gather
420 : * Determines and returns the cost of gather path.
421 : *
422 : * 'rel' is the relation to be operated upon
423 : * 'param_info' is the ParamPathInfo if this is a parameterized path, else NULL
424 : * 'rows' may be used to point to a row estimate; if non-NULL, it overrides
425 : * both 'rel' and 'param_info'. This is useful when the path doesn't exactly
426 : * correspond to any particular RelOptInfo.
427 : */
428 : void
429 26802 : cost_gather(GatherPath *path, PlannerInfo *root,
430 : RelOptInfo *rel, ParamPathInfo *param_info,
431 : double *rows)
432 : {
433 26802 : Cost startup_cost = 0;
434 26802 : Cost run_cost = 0;
435 :
436 : /* Mark the path with the correct row estimate */
437 26802 : if (rows)
438 6946 : path->path.rows = *rows;
439 19856 : else if (param_info)
440 0 : path->path.rows = param_info->ppi_rows;
441 : else
442 19856 : path->path.rows = rel->rows;
443 :
444 26802 : startup_cost = path->subpath->startup_cost;
445 :
446 26802 : run_cost = path->subpath->total_cost - path->subpath->startup_cost;
447 :
448 : /* Parallel setup and communication cost. */
449 26802 : startup_cost += parallel_setup_cost;
450 26802 : run_cost += parallel_tuple_cost * path->path.rows;
451 :
452 26802 : path->path.disabled_nodes = path->subpath->disabled_nodes
453 26802 : + ((rel->pgs_mask & PGS_GATHER) != 0 ? 0 : 1);
454 26802 : path->path.startup_cost = startup_cost;
455 26802 : path->path.total_cost = (startup_cost + run_cost);
456 26802 : }
457 :
458 : /*
459 : * cost_gather_merge
460 : * Determines and returns the cost of gather merge path.
461 : *
462 : * GatherMerge merges several pre-sorted input streams, using a heap that at
463 : * any given instant holds the next tuple from each stream. If there are N
464 : * streams, we need about N*log2(N) tuple comparisons to construct the heap at
465 : * startup, and then for each output tuple, about log2(N) comparisons to
466 : * replace the top heap entry with the next tuple from the same stream.
467 : */
468 : void
469 18944 : cost_gather_merge(GatherMergePath *path, PlannerInfo *root,
470 : RelOptInfo *rel, ParamPathInfo *param_info,
471 : int input_disabled_nodes,
472 : Cost input_startup_cost, Cost input_total_cost,
473 : double *rows)
474 : {
475 18944 : Cost startup_cost = 0;
476 18944 : Cost run_cost = 0;
477 : Cost comparison_cost;
478 : double N;
479 : double logN;
480 :
481 : /* Mark the path with the correct row estimate */
482 18944 : if (rows)
483 11298 : path->path.rows = *rows;
484 7646 : else if (param_info)
485 0 : path->path.rows = param_info->ppi_rows;
486 : else
487 7646 : path->path.rows = rel->rows;
488 :
489 : /*
490 : * Add one to the number of workers to account for the leader. This might
491 : * be overgenerous since the leader will do less work than other workers
492 : * in typical cases, but we'll go with it for now.
493 : */
494 : Assert(path->num_workers > 0);
495 18944 : N = (double) path->num_workers + 1;
496 18944 : logN = LOG2(N);
497 :
498 : /* Assumed cost per tuple comparison */
499 18944 : comparison_cost = 2.0 * cpu_operator_cost;
500 :
501 : /* Heap creation cost */
502 18944 : startup_cost += comparison_cost * N * logN;
503 :
504 : /* Per-tuple heap maintenance cost */
505 18944 : run_cost += path->path.rows * comparison_cost * logN;
506 :
507 : /* small cost for heap management, like cost_merge_append */
508 18944 : run_cost += cpu_operator_cost * path->path.rows;
509 :
510 : /*
511 : * Parallel setup and communication cost. Since Gather Merge, unlike
512 : * Gather, requires us to block until a tuple is available from every
513 : * worker, we bump the IPC cost up a little bit as compared with Gather.
514 : * For lack of a better idea, charge an extra 5%.
515 : */
516 18944 : startup_cost += parallel_setup_cost;
517 18944 : run_cost += parallel_tuple_cost * path->path.rows * 1.05;
518 :
519 18944 : path->path.disabled_nodes = path->subpath->disabled_nodes
520 18944 : + ((rel->pgs_mask & PGS_GATHER_MERGE) != 0 ? 0 : 1);
521 18944 : path->path.startup_cost = startup_cost + input_startup_cost;
522 18944 : path->path.total_cost = (startup_cost + run_cost + input_total_cost);
523 18944 : }
524 :
525 : /*
526 : * cost_index
527 : * Determines and returns the cost of scanning a relation using an index.
528 : *
529 : * 'path' describes the indexscan under consideration, and is complete
530 : * except for the fields to be set by this routine
531 : * 'loop_count' is the number of repetitions of the indexscan to factor into
532 : * estimates of caching behavior
533 : *
534 : * In addition to rows, startup_cost and total_cost, cost_index() sets the
535 : * path's indextotalcost and indexselectivity fields. These values will be
536 : * needed if the IndexPath is used in a BitmapIndexScan.
537 : *
538 : * NOTE: path->indexquals must contain only clauses usable as index
539 : * restrictions. Any additional quals evaluated as qpquals may reduce the
540 : * number of returned tuples, but they won't reduce the number of tuples
541 : * we have to fetch from the table, so they don't reduce the scan cost.
542 : */
543 : void
544 830212 : cost_index(IndexPath *path, PlannerInfo *root, double loop_count,
545 : bool partial_path)
546 : {
547 830212 : IndexOptInfo *index = path->indexinfo;
548 830212 : RelOptInfo *baserel = index->rel;
549 830212 : bool indexonly = (path->path.pathtype == T_IndexOnlyScan);
550 : amcostestimate_function amcostestimate;
551 : List *qpquals;
552 830212 : Cost startup_cost = 0;
553 830212 : Cost run_cost = 0;
554 830212 : Cost cpu_run_cost = 0;
555 : Cost indexStartupCost;
556 : Cost indexTotalCost;
557 : Selectivity indexSelectivity;
558 : double indexCorrelation,
559 : csquared;
560 : double spc_seq_page_cost,
561 : spc_random_page_cost;
562 : Cost min_IO_cost,
563 : max_IO_cost;
564 : QualCost qpqual_cost;
565 : Cost cpu_per_tuple;
566 : double tuples_fetched;
567 : double pages_fetched;
568 : double rand_heap_pages;
569 : double index_pages;
570 : uint64 enable_mask;
571 :
572 : /* Should only be applied to base relations */
573 : Assert(IsA(baserel, RelOptInfo) &&
574 : IsA(index, IndexOptInfo));
575 : Assert(baserel->relid > 0);
576 : Assert(baserel->rtekind == RTE_RELATION);
577 :
578 : /*
579 : * Mark the path with the correct row estimate, and identify which quals
580 : * will need to be enforced as qpquals. We need not check any quals that
581 : * are implied by the index's predicate, so we can use indrestrictinfo not
582 : * baserestrictinfo as the list of relevant restriction clauses for the
583 : * rel.
584 : */
585 830212 : if (path->path.param_info)
586 : {
587 159606 : path->path.rows = path->path.param_info->ppi_rows;
588 : /* qpquals come from the rel's restriction clauses and ppi_clauses */
589 159606 : qpquals = list_concat(extract_nonindex_conditions(path->indexinfo->indrestrictinfo,
590 : path->indexclauses),
591 159606 : extract_nonindex_conditions(path->path.param_info->ppi_clauses,
592 : path->indexclauses));
593 : }
594 : else
595 : {
596 670606 : path->path.rows = baserel->rows;
597 : /* qpquals come from just the rel's restriction clauses */
598 670606 : qpquals = extract_nonindex_conditions(path->indexinfo->indrestrictinfo,
599 : path->indexclauses);
600 : }
601 :
602 : /* is this scan type disabled? */
603 830212 : enable_mask = (indexonly ? PGS_INDEXONLYSCAN : PGS_INDEXSCAN)
604 830212 : | (partial_path ? 0 : PGS_CONSIDER_NONPARTIAL);
605 830212 : path->path.disabled_nodes =
606 830212 : (baserel->pgs_mask & enable_mask) == enable_mask ? 0 : 1;
607 :
608 : /*
609 : * Call index-access-method-specific code to estimate the processing cost
610 : * for scanning the index, as well as the selectivity of the index (ie,
611 : * the fraction of main-table tuples we will have to retrieve) and its
612 : * correlation to the main-table tuple order. We need a cast here because
613 : * pathnodes.h uses a weak function type to avoid including amapi.h.
614 : */
615 830212 : amcostestimate = (amcostestimate_function) index->amcostestimate;
616 830212 : amcostestimate(root, path, loop_count,
617 : &indexStartupCost, &indexTotalCost,
618 : &indexSelectivity, &indexCorrelation,
619 : &index_pages);
620 :
621 : /*
622 : * Save amcostestimate's results for possible use in bitmap scan planning.
623 : * We don't bother to save indexStartupCost or indexCorrelation, because a
624 : * bitmap scan doesn't care about either.
625 : */
626 830212 : path->indextotalcost = indexTotalCost;
627 830212 : path->indexselectivity = indexSelectivity;
628 :
629 : /* all costs for touching index itself included here */
630 830212 : startup_cost += indexStartupCost;
631 830212 : run_cost += indexTotalCost - indexStartupCost;
632 :
633 : /* estimate number of main-table tuples fetched */
634 830212 : tuples_fetched = clamp_row_est(indexSelectivity * baserel->tuples);
635 :
636 : /* fetch estimated page costs for tablespace containing table */
637 830212 : get_tablespace_page_costs(baserel->reltablespace,
638 : &spc_random_page_cost,
639 : &spc_seq_page_cost);
640 :
641 : /*----------
642 : * Estimate number of main-table pages fetched, and compute I/O cost.
643 : *
644 : * When the index ordering is uncorrelated with the table ordering,
645 : * we use an approximation proposed by Mackert and Lohman (see
646 : * index_pages_fetched() for details) to compute the number of pages
647 : * fetched, and then charge spc_random_page_cost per page fetched.
648 : *
649 : * When the index ordering is exactly correlated with the table ordering
650 : * (just after a CLUSTER, for example), the number of pages fetched should
651 : * be exactly selectivity * table_size. What's more, all but the first
652 : * will be sequential fetches, not the random fetches that occur in the
653 : * uncorrelated case. So if the number of pages is more than 1, we
654 : * ought to charge
655 : * spc_random_page_cost + (pages_fetched - 1) * spc_seq_page_cost
656 : * For partially-correlated indexes, we ought to charge somewhere between
657 : * these two estimates. We currently interpolate linearly between the
658 : * estimates based on the correlation squared (XXX is that appropriate?).
659 : *
660 : * If it's an index-only scan, then we will not need to fetch any heap
661 : * pages for which the visibility map shows all tuples are visible.
662 : * Hence, reduce the estimated number of heap fetches accordingly.
663 : * We use the measured fraction of the entire heap that is all-visible,
664 : * which might not be particularly relevant to the subset of the heap
665 : * that this query will fetch; but it's not clear how to do better.
666 : *----------
667 : */
668 830212 : if (loop_count > 1)
669 : {
670 : /*
671 : * For repeated indexscans, the appropriate estimate for the
672 : * uncorrelated case is to scale up the number of tuples fetched in
673 : * the Mackert and Lohman formula by the number of scans, so that we
674 : * estimate the number of pages fetched by all the scans; then
675 : * pro-rate the costs for one scan. In this case we assume all the
676 : * fetches are random accesses.
677 : */
678 91622 : pages_fetched = index_pages_fetched(tuples_fetched * loop_count,
679 : baserel->pages,
680 91622 : (double) index->pages,
681 : root);
682 :
683 91622 : if (indexonly)
684 11390 : pages_fetched = ceil(pages_fetched * (1.0 - baserel->allvisfrac));
685 :
686 91622 : rand_heap_pages = pages_fetched;
687 :
688 91622 : max_IO_cost = (pages_fetched * spc_random_page_cost) / loop_count;
689 :
690 : /*
691 : * In the perfectly correlated case, the number of pages touched by
692 : * each scan is selectivity * table_size, and we can use the Mackert
693 : * and Lohman formula at the page level to estimate how much work is
694 : * saved by caching across scans. We still assume all the fetches are
695 : * random, though, which is an overestimate that's hard to correct for
696 : * without double-counting the cache effects. (But in most cases
697 : * where such a plan is actually interesting, only one page would get
698 : * fetched per scan anyway, so it shouldn't matter much.)
699 : */
700 91622 : pages_fetched = ceil(indexSelectivity * (double) baserel->pages);
701 :
702 91622 : pages_fetched = index_pages_fetched(pages_fetched * loop_count,
703 : baserel->pages,
704 91622 : (double) index->pages,
705 : root);
706 :
707 91622 : if (indexonly)
708 11390 : pages_fetched = ceil(pages_fetched * (1.0 - baserel->allvisfrac));
709 :
710 91622 : min_IO_cost = (pages_fetched * spc_random_page_cost) / loop_count;
711 : }
712 : else
713 : {
714 : /*
715 : * Normal case: apply the Mackert and Lohman formula, and then
716 : * interpolate between that and the correlation-derived result.
717 : */
718 738590 : pages_fetched = index_pages_fetched(tuples_fetched,
719 : baserel->pages,
720 738590 : (double) index->pages,
721 : root);
722 :
723 738590 : if (indexonly)
724 71766 : pages_fetched = ceil(pages_fetched * (1.0 - baserel->allvisfrac));
725 :
726 738590 : rand_heap_pages = pages_fetched;
727 :
728 : /* max_IO_cost is for the perfectly uncorrelated case (csquared=0) */
729 738590 : max_IO_cost = pages_fetched * spc_random_page_cost;
730 :
731 : /* min_IO_cost is for the perfectly correlated case (csquared=1) */
732 738590 : pages_fetched = ceil(indexSelectivity * (double) baserel->pages);
733 :
734 738590 : if (indexonly)
735 71766 : pages_fetched = ceil(pages_fetched * (1.0 - baserel->allvisfrac));
736 :
737 738590 : if (pages_fetched > 0)
738 : {
739 660320 : min_IO_cost = spc_random_page_cost;
740 660320 : if (pages_fetched > 1)
741 204230 : min_IO_cost += (pages_fetched - 1) * spc_seq_page_cost;
742 : }
743 : else
744 78270 : min_IO_cost = 0;
745 : }
746 :
747 830212 : if (partial_path)
748 : {
749 : /*
750 : * For index only scans compute workers based on number of index pages
751 : * fetched; the number of heap pages we fetch might be so small as to
752 : * effectively rule out parallelism, which we don't want to do.
753 : */
754 284714 : if (indexonly)
755 26496 : rand_heap_pages = -1;
756 :
757 : /*
758 : * Estimate the number of parallel workers required to scan index. Use
759 : * the number of heap pages computed considering heap fetches won't be
760 : * sequential as for parallel scans the pages are accessed in random
761 : * order.
762 : */
763 284714 : path->path.parallel_workers = compute_parallel_worker(baserel,
764 : rand_heap_pages,
765 : index_pages,
766 : max_parallel_workers_per_gather);
767 :
768 : /*
769 : * Fall out if workers can't be assigned for parallel scan, because in
770 : * such a case this path will be rejected. So there is no benefit in
771 : * doing extra computation.
772 : */
773 284714 : if (path->path.parallel_workers <= 0)
774 274474 : return;
775 :
776 10240 : path->path.parallel_aware = true;
777 : }
778 :
779 : /*
780 : * Now interpolate based on estimated index order correlation to get total
781 : * disk I/O cost for main table accesses.
782 : */
783 555738 : csquared = indexCorrelation * indexCorrelation;
784 :
785 555738 : run_cost += max_IO_cost + csquared * (min_IO_cost - max_IO_cost);
786 :
787 : /*
788 : * Estimate CPU costs per tuple.
789 : *
790 : * What we want here is cpu_tuple_cost plus the evaluation costs of any
791 : * qual clauses that we have to evaluate as qpquals.
792 : */
793 555738 : cost_qual_eval(&qpqual_cost, qpquals, root);
794 :
795 555738 : startup_cost += qpqual_cost.startup;
796 555738 : cpu_per_tuple = cpu_tuple_cost + qpqual_cost.per_tuple;
797 :
798 555738 : cpu_run_cost += cpu_per_tuple * tuples_fetched;
799 :
800 : /* tlist eval costs are paid per output row, not per tuple scanned */
801 555738 : startup_cost += path->path.pathtarget->cost.startup;
802 555738 : cpu_run_cost += path->path.pathtarget->cost.per_tuple * path->path.rows;
803 :
804 : /* Adjust costing for parallelism, if used. */
805 555738 : if (path->path.parallel_workers > 0)
806 : {
807 10240 : double parallel_divisor = get_parallel_divisor(&path->path);
808 :
809 10240 : path->path.rows = clamp_row_est(path->path.rows / parallel_divisor);
810 :
811 : /* The CPU cost is divided among all the workers. */
812 10240 : cpu_run_cost /= parallel_divisor;
813 : }
814 :
815 555738 : run_cost += cpu_run_cost;
816 :
817 555738 : path->path.startup_cost = startup_cost;
818 555738 : path->path.total_cost = startup_cost + run_cost;
819 : }
820 :
821 : /*
822 : * extract_nonindex_conditions
823 : *
824 : * Given a list of quals to be enforced in an indexscan, extract the ones that
825 : * will have to be applied as qpquals (ie, the index machinery won't handle
826 : * them). Here we detect only whether a qual clause is directly redundant
827 : * with some indexclause. If the index path is chosen for use, createplan.c
828 : * will try a bit harder to get rid of redundant qual conditions; specifically
829 : * it will see if quals can be proven to be implied by the indexquals. But
830 : * it does not seem worth the cycles to try to factor that in at this stage,
831 : * since we're only trying to estimate qual eval costs. Otherwise this must
832 : * match the logic in create_indexscan_plan().
833 : *
834 : * qual_clauses, and the result, are lists of RestrictInfos.
835 : * indexclauses is a list of IndexClauses.
836 : */
837 : static List *
838 989818 : extract_nonindex_conditions(List *qual_clauses, List *indexclauses)
839 : {
840 989818 : List *result = NIL;
841 : ListCell *lc;
842 :
843 2047840 : foreach(lc, qual_clauses)
844 : {
845 1058022 : RestrictInfo *rinfo = lfirst_node(RestrictInfo, lc);
846 :
847 1058022 : if (rinfo->pseudoconstant)
848 10304 : continue; /* we may drop pseudoconstants here */
849 1047718 : if (is_redundant_with_indexclauses(rinfo, indexclauses))
850 622788 : continue; /* dup or derived from same EquivalenceClass */
851 : /* ... skip the predicate proof attempt createplan.c will try ... */
852 424930 : result = lappend(result, rinfo);
853 : }
854 989818 : return result;
855 : }
856 :
857 : /*
858 : * index_pages_fetched
859 : * Estimate the number of pages actually fetched after accounting for
860 : * cache effects.
861 : *
862 : * We use an approximation proposed by Mackert and Lohman, "Index Scans
863 : * Using a Finite LRU Buffer: A Validated I/O Model", ACM Transactions
864 : * on Database Systems, Vol. 14, No. 3, September 1989, Pages 401-424.
865 : * The Mackert and Lohman approximation is that the number of pages
866 : * fetched is
867 : * PF =
868 : * min(2TNs/(2T+Ns), T) when T <= b
869 : * 2TNs/(2T+Ns) when T > b and Ns <= 2Tb/(2T-b)
870 : * b + (Ns - 2Tb/(2T-b))*(T-b)/T when T > b and Ns > 2Tb/(2T-b)
871 : * where
872 : * T = # pages in table
873 : * N = # tuples in table
874 : * s = selectivity = fraction of table to be scanned
875 : * b = # buffer pages available (we include kernel space here)
876 : *
877 : * We assume that effective_cache_size is the total number of buffer pages
878 : * available for the whole query, and pro-rate that space across all the
879 : * tables in the query and the index currently under consideration. (This
880 : * ignores space needed for other indexes used by the query, but since we
881 : * don't know which indexes will get used, we can't estimate that very well;
882 : * and in any case counting all the tables may well be an overestimate, since
883 : * depending on the join plan not all the tables may be scanned concurrently.)
884 : *
885 : * The product Ns is the number of tuples fetched; we pass in that
886 : * product rather than calculating it here. "pages" is the number of pages
887 : * in the object under consideration (either an index or a table).
888 : * "index_pages" is the amount to add to the total table space, which was
889 : * computed for us by make_one_rel.
890 : *
891 : * Caller is expected to have ensured that tuples_fetched is greater than zero
892 : * and rounded to integer (see clamp_row_est). The result will likewise be
893 : * greater than zero and integral.
894 : */
895 : double
896 1169430 : index_pages_fetched(double tuples_fetched, BlockNumber pages,
897 : double index_pages, PlannerInfo *root)
898 : {
899 : double pages_fetched;
900 : double total_pages;
901 : double T,
902 : b;
903 :
904 : /* T is # pages in table, but don't allow it to be zero */
905 1169430 : T = (pages > 1) ? (double) pages : 1.0;
906 :
907 : /* Compute number of pages assumed to be competing for cache space */
908 1169430 : total_pages = root->total_table_pages + index_pages;
909 1169430 : total_pages = Max(total_pages, 1.0);
910 : Assert(T <= total_pages);
911 :
912 : /* b is pro-rated share of effective_cache_size */
913 1169430 : b = (double) effective_cache_size * T / total_pages;
914 :
915 : /* force it positive and integral */
916 1169430 : if (b <= 1.0)
917 0 : b = 1.0;
918 : else
919 1169430 : b = ceil(b);
920 :
921 : /* This part is the Mackert and Lohman formula */
922 1169430 : if (T <= b)
923 : {
924 1169430 : pages_fetched =
925 1169430 : (2.0 * T * tuples_fetched) / (2.0 * T + tuples_fetched);
926 1169430 : if (pages_fetched >= T)
927 686486 : pages_fetched = T;
928 : else
929 482944 : pages_fetched = ceil(pages_fetched);
930 : }
931 : else
932 : {
933 : double lim;
934 :
935 0 : lim = (2.0 * T * b) / (2.0 * T - b);
936 0 : if (tuples_fetched <= lim)
937 : {
938 0 : pages_fetched =
939 0 : (2.0 * T * tuples_fetched) / (2.0 * T + tuples_fetched);
940 : }
941 : else
942 : {
943 0 : pages_fetched =
944 0 : b + (tuples_fetched - lim) * (T - b) / T;
945 : }
946 0 : pages_fetched = ceil(pages_fetched);
947 : }
948 1169430 : return pages_fetched;
949 : }
950 :
951 : /*
952 : * get_indexpath_pages
953 : * Determine the total size of the indexes used in a bitmap index path.
954 : *
955 : * Note: if the same index is used more than once in a bitmap tree, we will
956 : * count it multiple times, which perhaps is the wrong thing ... but it's
957 : * not completely clear, and detecting duplicates is difficult, so ignore it
958 : * for now.
959 : */
960 : static double
961 198528 : get_indexpath_pages(Path *bitmapqual)
962 : {
963 198528 : double result = 0;
964 : ListCell *l;
965 :
966 198528 : if (IsA(bitmapqual, BitmapAndPath))
967 : {
968 23934 : BitmapAndPath *apath = (BitmapAndPath *) bitmapqual;
969 :
970 71802 : foreach(l, apath->bitmapquals)
971 : {
972 47868 : result += get_indexpath_pages((Path *) lfirst(l));
973 : }
974 : }
975 174594 : else if (IsA(bitmapqual, BitmapOrPath))
976 : {
977 74 : BitmapOrPath *opath = (BitmapOrPath *) bitmapqual;
978 :
979 234 : foreach(l, opath->bitmapquals)
980 : {
981 160 : result += get_indexpath_pages((Path *) lfirst(l));
982 : }
983 : }
984 174520 : else if (IsA(bitmapqual, IndexPath))
985 : {
986 174520 : IndexPath *ipath = (IndexPath *) bitmapqual;
987 :
988 174520 : result = (double) ipath->indexinfo->pages;
989 : }
990 : else
991 0 : elog(ERROR, "unrecognized node type: %d", nodeTag(bitmapqual));
992 :
993 198528 : return result;
994 : }
995 :
996 : /*
997 : * cost_bitmap_heap_scan
998 : * Determines and returns the cost of scanning a relation using a bitmap
999 : * index-then-heap plan.
1000 : *
1001 : * 'baserel' is the relation to be scanned
1002 : * 'param_info' is the ParamPathInfo if this is a parameterized path, else NULL
1003 : * 'bitmapqual' is a tree of IndexPaths, BitmapAndPaths, and BitmapOrPaths
1004 : * 'loop_count' is the number of repetitions of the indexscan to factor into
1005 : * estimates of caching behavior
1006 : *
1007 : * Note: the component IndexPaths in bitmapqual should have been costed
1008 : * using the same loop_count.
1009 : */
1010 : void
1011 566564 : cost_bitmap_heap_scan(Path *path, PlannerInfo *root, RelOptInfo *baserel,
1012 : ParamPathInfo *param_info,
1013 : Path *bitmapqual, double loop_count)
1014 : {
1015 566564 : Cost startup_cost = 0;
1016 566564 : Cost run_cost = 0;
1017 : Cost indexTotalCost;
1018 : QualCost qpqual_cost;
1019 : Cost cpu_per_tuple;
1020 : Cost cost_per_page;
1021 : Cost cpu_run_cost;
1022 : double tuples_fetched;
1023 : double pages_fetched;
1024 : double spc_seq_page_cost,
1025 : spc_random_page_cost;
1026 : double T;
1027 566564 : uint64 enable_mask = PGS_BITMAPSCAN;
1028 :
1029 : /* Should only be applied to base relations */
1030 : Assert(IsA(baserel, RelOptInfo));
1031 : Assert(baserel->relid > 0);
1032 : Assert(baserel->rtekind == RTE_RELATION);
1033 :
1034 : /* Mark the path with the correct row estimate */
1035 566564 : if (param_info)
1036 241208 : path->rows = param_info->ppi_rows;
1037 : else
1038 325356 : path->rows = baserel->rows;
1039 :
1040 566564 : pages_fetched = compute_bitmap_pages(root, baserel, bitmapqual,
1041 : loop_count, &indexTotalCost,
1042 : &tuples_fetched);
1043 :
1044 566564 : startup_cost += indexTotalCost;
1045 566564 : T = (baserel->pages > 1) ? (double) baserel->pages : 1.0;
1046 :
1047 : /* Fetch estimated page costs for tablespace containing table. */
1048 566564 : get_tablespace_page_costs(baserel->reltablespace,
1049 : &spc_random_page_cost,
1050 : &spc_seq_page_cost);
1051 :
1052 : /*
1053 : * For small numbers of pages we should charge spc_random_page_cost
1054 : * apiece, while if nearly all the table's pages are being read, it's more
1055 : * appropriate to charge spc_seq_page_cost apiece. The effect is
1056 : * nonlinear, too. For lack of a better idea, interpolate like this to
1057 : * determine the cost per page.
1058 : */
1059 566564 : if (pages_fetched >= 2.0)
1060 119994 : cost_per_page = spc_random_page_cost -
1061 119994 : (spc_random_page_cost - spc_seq_page_cost)
1062 119994 : * sqrt(pages_fetched / T);
1063 : else
1064 446570 : cost_per_page = spc_random_page_cost;
1065 :
1066 566564 : run_cost += pages_fetched * cost_per_page;
1067 :
1068 : /*
1069 : * Estimate CPU costs per tuple.
1070 : *
1071 : * Often the indexquals don't need to be rechecked at each tuple ... but
1072 : * not always, especially not if there are enough tuples involved that the
1073 : * bitmaps become lossy. For the moment, just assume they will be
1074 : * rechecked always. This means we charge the full freight for all the
1075 : * scan clauses.
1076 : */
1077 566564 : get_restriction_qual_cost(root, baserel, param_info, &qpqual_cost);
1078 :
1079 566564 : startup_cost += qpqual_cost.startup;
1080 566564 : cpu_per_tuple = cpu_tuple_cost + qpqual_cost.per_tuple;
1081 566564 : cpu_run_cost = cpu_per_tuple * tuples_fetched;
1082 :
1083 : /* Adjust costing for parallelism, if used. */
1084 566564 : if (path->parallel_workers > 0)
1085 : {
1086 4958 : double parallel_divisor = get_parallel_divisor(path);
1087 :
1088 : /* The CPU cost is divided among all the workers. */
1089 4958 : cpu_run_cost /= parallel_divisor;
1090 :
1091 4958 : path->rows = clamp_row_est(path->rows / parallel_divisor);
1092 : }
1093 : else
1094 561606 : enable_mask |= PGS_CONSIDER_NONPARTIAL;
1095 :
1096 :
1097 566564 : run_cost += cpu_run_cost;
1098 :
1099 : /* tlist eval costs are paid per output row, not per tuple scanned */
1100 566564 : startup_cost += path->pathtarget->cost.startup;
1101 566564 : run_cost += path->pathtarget->cost.per_tuple * path->rows;
1102 :
1103 566564 : path->disabled_nodes =
1104 566564 : (baserel->pgs_mask & enable_mask) == enable_mask ? 0 : 1;
1105 566564 : path->startup_cost = startup_cost;
1106 566564 : path->total_cost = startup_cost + run_cost;
1107 566564 : }
1108 :
1109 : /*
1110 : * cost_bitmap_tree_node
1111 : * Extract cost and selectivity from a bitmap tree node (index/and/or)
1112 : */
1113 : void
1114 1042934 : cost_bitmap_tree_node(Path *path, Cost *cost, Selectivity *selec)
1115 : {
1116 1042934 : if (IsA(path, IndexPath))
1117 : {
1118 987340 : *cost = ((IndexPath *) path)->indextotalcost;
1119 987340 : *selec = ((IndexPath *) path)->indexselectivity;
1120 :
1121 : /*
1122 : * Charge a small amount per retrieved tuple to reflect the costs of
1123 : * manipulating the bitmap. This is mostly to make sure that a bitmap
1124 : * scan doesn't look to be the same cost as an indexscan to retrieve a
1125 : * single tuple.
1126 : */
1127 987340 : *cost += 0.1 * cpu_operator_cost * path->rows;
1128 : }
1129 55594 : else if (IsA(path, BitmapAndPath))
1130 : {
1131 51922 : *cost = path->total_cost;
1132 51922 : *selec = ((BitmapAndPath *) path)->bitmapselectivity;
1133 : }
1134 3672 : else if (IsA(path, BitmapOrPath))
1135 : {
1136 3672 : *cost = path->total_cost;
1137 3672 : *selec = ((BitmapOrPath *) path)->bitmapselectivity;
1138 : }
1139 : else
1140 : {
1141 0 : elog(ERROR, "unrecognized node type: %d", nodeTag(path));
1142 : *cost = *selec = 0; /* keep compiler quiet */
1143 : }
1144 1042934 : }
1145 :
1146 : /*
1147 : * cost_bitmap_and_node
1148 : * Estimate the cost of a BitmapAnd node
1149 : *
1150 : * Note that this considers only the costs of index scanning and bitmap
1151 : * creation, not the eventual heap access. In that sense the object isn't
1152 : * truly a Path, but it has enough path-like properties (costs in particular)
1153 : * to warrant treating it as one. We don't bother to set the path rows field,
1154 : * however.
1155 : */
1156 : void
1157 51704 : cost_bitmap_and_node(BitmapAndPath *path, PlannerInfo *root)
1158 : {
1159 : Cost totalCost;
1160 : Selectivity selec;
1161 : ListCell *l;
1162 :
1163 : /*
1164 : * We estimate AND selectivity on the assumption that the inputs are
1165 : * independent. This is probably often wrong, but we don't have the info
1166 : * to do better.
1167 : *
1168 : * The runtime cost of the BitmapAnd itself is estimated at 100x
1169 : * cpu_operator_cost for each tbm_intersect needed. Probably too small,
1170 : * definitely too simplistic?
1171 : */
1172 51704 : totalCost = 0.0;
1173 51704 : selec = 1.0;
1174 155112 : foreach(l, path->bitmapquals)
1175 : {
1176 103408 : Path *subpath = (Path *) lfirst(l);
1177 : Cost subCost;
1178 : Selectivity subselec;
1179 :
1180 103408 : cost_bitmap_tree_node(subpath, &subCost, &subselec);
1181 :
1182 103408 : selec *= subselec;
1183 :
1184 103408 : totalCost += subCost;
1185 103408 : if (l != list_head(path->bitmapquals))
1186 51704 : totalCost += 100.0 * cpu_operator_cost;
1187 : }
1188 51704 : path->bitmapselectivity = selec;
1189 51704 : path->path.rows = 0; /* per above, not used */
1190 51704 : path->path.disabled_nodes = 0;
1191 51704 : path->path.startup_cost = totalCost;
1192 51704 : path->path.total_cost = totalCost;
1193 51704 : }
1194 :
1195 : /*
1196 : * cost_bitmap_or_node
1197 : * Estimate the cost of a BitmapOr node
1198 : *
1199 : * See comments for cost_bitmap_and_node.
1200 : */
1201 : void
1202 1040 : cost_bitmap_or_node(BitmapOrPath *path, PlannerInfo *root)
1203 : {
1204 : Cost totalCost;
1205 : Selectivity selec;
1206 : ListCell *l;
1207 :
1208 : /*
1209 : * We estimate OR selectivity on the assumption that the inputs are
1210 : * non-overlapping, since that's often the case in "x IN (list)" type
1211 : * situations. Of course, we clamp to 1.0 at the end.
1212 : *
1213 : * The runtime cost of the BitmapOr itself is estimated at 100x
1214 : * cpu_operator_cost for each tbm_union needed. Probably too small,
1215 : * definitely too simplistic? We are aware that the tbm_unions are
1216 : * optimized out when the inputs are BitmapIndexScans.
1217 : */
1218 1040 : totalCost = 0.0;
1219 1040 : selec = 0.0;
1220 2922 : foreach(l, path->bitmapquals)
1221 : {
1222 1882 : Path *subpath = (Path *) lfirst(l);
1223 : Cost subCost;
1224 : Selectivity subselec;
1225 :
1226 1882 : cost_bitmap_tree_node(subpath, &subCost, &subselec);
1227 :
1228 1882 : selec += subselec;
1229 :
1230 1882 : totalCost += subCost;
1231 1882 : if (l != list_head(path->bitmapquals) &&
1232 842 : !IsA(subpath, IndexPath))
1233 0 : totalCost += 100.0 * cpu_operator_cost;
1234 : }
1235 1040 : path->bitmapselectivity = Min(selec, 1.0);
1236 1040 : path->path.rows = 0; /* per above, not used */
1237 1040 : path->path.startup_cost = totalCost;
1238 1040 : path->path.total_cost = totalCost;
1239 1040 : }
1240 :
1241 : /*
1242 : * cost_tidscan
1243 : * Determines and returns the cost of scanning a relation using TIDs.
1244 : *
1245 : * 'baserel' is the relation to be scanned
1246 : * 'tidquals' is the list of TID-checkable quals
1247 : * 'param_info' is the ParamPathInfo if this is a parameterized path, else NULL
1248 : */
1249 : void
1250 872 : cost_tidscan(Path *path, PlannerInfo *root,
1251 : RelOptInfo *baserel, List *tidquals, ParamPathInfo *param_info)
1252 : {
1253 872 : Cost startup_cost = 0;
1254 872 : Cost run_cost = 0;
1255 : QualCost qpqual_cost;
1256 : Cost cpu_per_tuple;
1257 : QualCost tid_qual_cost;
1258 : double ntuples;
1259 : ListCell *l;
1260 : double spc_random_page_cost;
1261 872 : uint64 enable_mask = 0;
1262 :
1263 : /* Should only be applied to base relations */
1264 : Assert(baserel->relid > 0);
1265 : Assert(baserel->rtekind == RTE_RELATION);
1266 : Assert(tidquals != NIL);
1267 :
1268 : /* Mark the path with the correct row estimate */
1269 872 : if (param_info)
1270 144 : path->rows = param_info->ppi_rows;
1271 : else
1272 728 : path->rows = baserel->rows;
1273 :
1274 : /* Count how many tuples we expect to retrieve */
1275 872 : ntuples = 0;
1276 1770 : foreach(l, tidquals)
1277 : {
1278 898 : RestrictInfo *rinfo = lfirst_node(RestrictInfo, l);
1279 898 : Expr *qual = rinfo->clause;
1280 :
1281 : /*
1282 : * We must use a TID scan for CurrentOfExpr; in any other case, we
1283 : * should be generating a TID scan only if TID scans are allowed.
1284 : * Also, if CurrentOfExpr is the qual, there should be only one.
1285 : */
1286 : Assert((baserel->pgs_mask & PGS_TIDSCAN) != 0 || IsA(qual, CurrentOfExpr));
1287 : Assert(list_length(tidquals) == 1 || !IsA(qual, CurrentOfExpr));
1288 :
1289 898 : if (IsA(qual, ScalarArrayOpExpr))
1290 : {
1291 : /* Each element of the array yields 1 tuple */
1292 50 : ScalarArrayOpExpr *saop = (ScalarArrayOpExpr *) qual;
1293 50 : Node *arraynode = (Node *) lsecond(saop->args);
1294 :
1295 50 : ntuples += estimate_array_length(root, arraynode);
1296 : }
1297 848 : else if (IsA(qual, CurrentOfExpr))
1298 : {
1299 : /* CURRENT OF yields 1 tuple */
1300 404 : ntuples++;
1301 : }
1302 : else
1303 : {
1304 : /* It's just CTID = something, count 1 tuple */
1305 444 : ntuples++;
1306 : }
1307 : }
1308 :
1309 : /*
1310 : * The TID qual expressions will be computed once, any other baserestrict
1311 : * quals once per retrieved tuple.
1312 : */
1313 872 : cost_qual_eval(&tid_qual_cost, tidquals, root);
1314 :
1315 : /* fetch estimated page cost for tablespace containing table */
1316 872 : get_tablespace_page_costs(baserel->reltablespace,
1317 : &spc_random_page_cost,
1318 : NULL);
1319 :
1320 : /* disk costs --- assume each tuple on a different page */
1321 872 : run_cost += spc_random_page_cost * ntuples;
1322 :
1323 : /* Add scanning CPU costs */
1324 872 : get_restriction_qual_cost(root, baserel, param_info, &qpqual_cost);
1325 :
1326 : /* XXX currently we assume TID quals are a subset of qpquals */
1327 872 : startup_cost += qpqual_cost.startup + tid_qual_cost.per_tuple;
1328 872 : cpu_per_tuple = cpu_tuple_cost + qpqual_cost.per_tuple -
1329 872 : tid_qual_cost.per_tuple;
1330 872 : run_cost += cpu_per_tuple * ntuples;
1331 :
1332 : /* tlist eval costs are paid per output row, not per tuple scanned */
1333 872 : startup_cost += path->pathtarget->cost.startup;
1334 872 : run_cost += path->pathtarget->cost.per_tuple * path->rows;
1335 :
1336 : /*
1337 : * There are assertions above verifying that we only reach this function
1338 : * either when baserel->pgs_mask includes PGS_TIDSCAN or when the TID scan
1339 : * is the only legal path, so we only need to consider the effects of
1340 : * PGS_CONSIDER_NONPARTIAL here.
1341 : */
1342 872 : if (path->parallel_workers == 0)
1343 872 : enable_mask |= PGS_CONSIDER_NONPARTIAL;
1344 872 : path->disabled_nodes =
1345 872 : (baserel->pgs_mask & enable_mask) != enable_mask ? 1 : 0;
1346 872 : path->startup_cost = startup_cost;
1347 872 : path->total_cost = startup_cost + run_cost;
1348 872 : }
1349 :
1350 : /*
1351 : * cost_tidrangescan
1352 : * Determines and sets the costs of scanning a relation using a range of
1353 : * TIDs for 'path'
1354 : *
1355 : * 'baserel' is the relation to be scanned
1356 : * 'tidrangequals' is the list of TID-checkable range quals
1357 : * 'param_info' is the ParamPathInfo if this is a parameterized path, else NULL
1358 : */
1359 : void
1360 2052 : cost_tidrangescan(Path *path, PlannerInfo *root,
1361 : RelOptInfo *baserel, List *tidrangequals,
1362 : ParamPathInfo *param_info)
1363 : {
1364 : Selectivity selectivity;
1365 : double pages;
1366 : Cost startup_cost;
1367 : Cost cpu_run_cost;
1368 : Cost disk_run_cost;
1369 : QualCost qpqual_cost;
1370 : Cost cpu_per_tuple;
1371 : QualCost tid_qual_cost;
1372 : double ntuples;
1373 : double nseqpages;
1374 : double spc_random_page_cost;
1375 : double spc_seq_page_cost;
1376 2052 : uint64 enable_mask = PGS_TIDSCAN;
1377 :
1378 : /* Should only be applied to base relations */
1379 : Assert(baserel->relid > 0);
1380 : Assert(baserel->rtekind == RTE_RELATION);
1381 :
1382 : /* Mark the path with the correct row estimate */
1383 2052 : if (param_info)
1384 0 : path->rows = param_info->ppi_rows;
1385 : else
1386 2052 : path->rows = baserel->rows;
1387 :
1388 : /* Count how many tuples and pages we expect to scan */
1389 2052 : selectivity = clauselist_selectivity(root, tidrangequals, baserel->relid,
1390 : JOIN_INNER, NULL);
1391 2052 : pages = ceil(selectivity * baserel->pages);
1392 :
1393 2052 : if (pages <= 0.0)
1394 42 : pages = 1.0;
1395 :
1396 : /*
1397 : * The first page in a range requires a random seek, but each subsequent
1398 : * page is just a normal sequential page read. NOTE: it's desirable for
1399 : * TID Range Scans to cost more than the equivalent Sequential Scans,
1400 : * because Seq Scans have some performance advantages such as scan
1401 : * synchronization, and we'd prefer one of them to be picked unless a TID
1402 : * Range Scan really is better.
1403 : */
1404 2052 : ntuples = selectivity * baserel->tuples;
1405 2052 : nseqpages = pages - 1.0;
1406 :
1407 : /*
1408 : * The TID qual expressions will be computed once, any other baserestrict
1409 : * quals once per retrieved tuple.
1410 : */
1411 2052 : cost_qual_eval(&tid_qual_cost, tidrangequals, root);
1412 :
1413 : /* fetch estimated page cost for tablespace containing table */
1414 2052 : get_tablespace_page_costs(baserel->reltablespace,
1415 : &spc_random_page_cost,
1416 : &spc_seq_page_cost);
1417 :
1418 : /* disk costs; 1 random page and the remainder as seq pages */
1419 2052 : disk_run_cost = spc_random_page_cost + spc_seq_page_cost * nseqpages;
1420 :
1421 : /* Add scanning CPU costs */
1422 2052 : get_restriction_qual_cost(root, baserel, param_info, &qpqual_cost);
1423 :
1424 : /*
1425 : * XXX currently we assume TID quals are a subset of qpquals at this
1426 : * point; they will be removed (if possible) when we create the plan, so
1427 : * we subtract their cost from the total qpqual cost. (If the TID quals
1428 : * can't be removed, this is a mistake and we're going to underestimate
1429 : * the CPU cost a bit.)
1430 : */
1431 2052 : startup_cost = qpqual_cost.startup + tid_qual_cost.per_tuple;
1432 2052 : cpu_per_tuple = cpu_tuple_cost + qpqual_cost.per_tuple -
1433 2052 : tid_qual_cost.per_tuple;
1434 2052 : cpu_run_cost = cpu_per_tuple * ntuples;
1435 :
1436 : /* tlist eval costs are paid per output row, not per tuple scanned */
1437 2052 : startup_cost += path->pathtarget->cost.startup;
1438 2052 : cpu_run_cost += path->pathtarget->cost.per_tuple * path->rows;
1439 :
1440 : /* Adjust costing for parallelism, if used. */
1441 2052 : if (path->parallel_workers > 0)
1442 : {
1443 48 : double parallel_divisor = get_parallel_divisor(path);
1444 :
1445 : /* The CPU cost is divided among all the workers. */
1446 48 : cpu_run_cost /= parallel_divisor;
1447 :
1448 : /*
1449 : * In the case of a parallel plan, the row count needs to represent
1450 : * the number of tuples processed per worker.
1451 : */
1452 48 : path->rows = clamp_row_est(path->rows / parallel_divisor);
1453 : }
1454 :
1455 : /*
1456 : * We should not generate this path type when PGS_TIDSCAN is unset, but we
1457 : * might need to disable this path due to PGS_CONSIDER_NONPARTIAL.
1458 : */
1459 : Assert((baserel->pgs_mask & PGS_TIDSCAN) != 0);
1460 2052 : if (path->parallel_workers == 0)
1461 2004 : enable_mask |= PGS_CONSIDER_NONPARTIAL;
1462 2052 : path->disabled_nodes =
1463 2052 : (baserel->pgs_mask & enable_mask) != enable_mask ? 1 : 0;
1464 2052 : path->startup_cost = startup_cost;
1465 2052 : path->total_cost = startup_cost + cpu_run_cost + disk_run_cost;
1466 2052 : }
1467 :
1468 : /*
1469 : * cost_subqueryscan
1470 : * Determines and returns the cost of scanning a subquery RTE.
1471 : *
1472 : * 'baserel' is the relation to be scanned
1473 : * 'param_info' is the ParamPathInfo if this is a parameterized path, else NULL
1474 : * 'trivial_pathtarget' is true if the pathtarget is believed to be trivial.
1475 : */
1476 : void
1477 63190 : cost_subqueryscan(SubqueryScanPath *path, PlannerInfo *root,
1478 : RelOptInfo *baserel, ParamPathInfo *param_info,
1479 : bool trivial_pathtarget)
1480 : {
1481 : Cost startup_cost;
1482 : Cost run_cost;
1483 : List *qpquals;
1484 : QualCost qpqual_cost;
1485 : Cost cpu_per_tuple;
1486 63190 : uint64 enable_mask = 0;
1487 :
1488 : /* Should only be applied to base relations that are subqueries */
1489 : Assert(baserel->relid > 0);
1490 : Assert(baserel->rtekind == RTE_SUBQUERY);
1491 :
1492 : /*
1493 : * We compute the rowcount estimate as the subplan's estimate times the
1494 : * selectivity of relevant restriction clauses. In simple cases this will
1495 : * come out the same as baserel->rows; but when dealing with parallelized
1496 : * paths we must do it like this to get the right answer.
1497 : */
1498 63190 : if (param_info)
1499 940 : qpquals = list_concat_copy(param_info->ppi_clauses,
1500 940 : baserel->baserestrictinfo);
1501 : else
1502 62250 : qpquals = baserel->baserestrictinfo;
1503 :
1504 63190 : path->path.rows = clamp_row_est(path->subpath->rows *
1505 63190 : clauselist_selectivity(root,
1506 : qpquals,
1507 : 0,
1508 : JOIN_INNER,
1509 : NULL));
1510 :
1511 : /*
1512 : * Cost of path is cost of evaluating the subplan, plus cost of evaluating
1513 : * any restriction clauses and tlist that will be attached to the
1514 : * SubqueryScan node, plus cpu_tuple_cost to account for selection and
1515 : * projection overhead.
1516 : */
1517 63190 : if (path->path.parallel_workers == 0)
1518 63124 : enable_mask |= PGS_CONSIDER_NONPARTIAL;
1519 63190 : path->path.disabled_nodes = path->subpath->disabled_nodes
1520 63190 : + (((baserel->pgs_mask & enable_mask) != enable_mask) ? 1 : 0);
1521 63190 : path->path.startup_cost = path->subpath->startup_cost;
1522 63190 : path->path.total_cost = path->subpath->total_cost;
1523 :
1524 : /*
1525 : * However, if there are no relevant restriction clauses and the
1526 : * pathtarget is trivial, then we expect that setrefs.c will optimize away
1527 : * the SubqueryScan plan node altogether, so we should just make its cost
1528 : * and rowcount equal to the input path's.
1529 : *
1530 : * Note: there are some edge cases where createplan.c will apply a
1531 : * different targetlist to the SubqueryScan node, thus falsifying our
1532 : * current estimate of whether the target is trivial, and making the cost
1533 : * estimate (though not the rowcount) wrong. It does not seem worth the
1534 : * extra complication to try to account for that exactly, especially since
1535 : * that behavior falsifies other cost estimates as well.
1536 : */
1537 63190 : if (qpquals == NIL && trivial_pathtarget)
1538 27650 : return;
1539 :
1540 35540 : get_restriction_qual_cost(root, baserel, param_info, &qpqual_cost);
1541 :
1542 35540 : startup_cost = qpqual_cost.startup;
1543 35540 : cpu_per_tuple = cpu_tuple_cost + qpqual_cost.per_tuple;
1544 35540 : run_cost = cpu_per_tuple * path->subpath->rows;
1545 :
1546 : /* tlist eval costs are paid per output row, not per tuple scanned */
1547 35540 : startup_cost += path->path.pathtarget->cost.startup;
1548 35540 : run_cost += path->path.pathtarget->cost.per_tuple * path->path.rows;
1549 :
1550 35540 : path->path.startup_cost += startup_cost;
1551 35540 : path->path.total_cost += startup_cost + run_cost;
1552 : }
1553 :
1554 : /*
1555 : * cost_functionscan
1556 : * Determines and returns the cost of scanning a function RTE.
1557 : *
1558 : * 'baserel' is the relation to be scanned
1559 : * 'param_info' is the ParamPathInfo if this is a parameterized path, else NULL
1560 : */
1561 : void
1562 53556 : cost_functionscan(Path *path, PlannerInfo *root,
1563 : RelOptInfo *baserel, ParamPathInfo *param_info)
1564 : {
1565 53556 : Cost startup_cost = 0;
1566 53556 : Cost run_cost = 0;
1567 : QualCost qpqual_cost;
1568 : Cost cpu_per_tuple;
1569 : RangeTblEntry *rte;
1570 : QualCost exprcost;
1571 53556 : uint64 enable_mask = 0;
1572 :
1573 : /* Should only be applied to base relations that are functions */
1574 : Assert(baserel->relid > 0);
1575 53556 : rte = planner_rt_fetch(baserel->relid, root);
1576 : Assert(rte->rtekind == RTE_FUNCTION);
1577 :
1578 : /* Mark the path with the correct row estimate */
1579 53556 : if (param_info)
1580 8590 : path->rows = param_info->ppi_rows;
1581 : else
1582 44966 : path->rows = baserel->rows;
1583 :
1584 : /*
1585 : * Estimate costs of executing the function expression(s).
1586 : *
1587 : * Currently, nodeFunctionscan.c always executes the functions to
1588 : * completion before returning any rows, and caches the results in a
1589 : * tuplestore. So the function eval cost is all startup cost, and per-row
1590 : * costs are minimal.
1591 : *
1592 : * XXX in principle we ought to charge tuplestore spill costs if the
1593 : * number of rows is large. However, given how phony our rowcount
1594 : * estimates for functions tend to be, there's not a lot of point in that
1595 : * refinement right now.
1596 : */
1597 53556 : cost_qual_eval_node(&exprcost, (Node *) rte->functions, root);
1598 :
1599 53556 : startup_cost += exprcost.startup + exprcost.per_tuple;
1600 :
1601 : /* Add scanning CPU costs */
1602 53556 : get_restriction_qual_cost(root, baserel, param_info, &qpqual_cost);
1603 :
1604 53556 : startup_cost += qpqual_cost.startup;
1605 53556 : cpu_per_tuple = cpu_tuple_cost + qpqual_cost.per_tuple;
1606 53556 : run_cost += cpu_per_tuple * baserel->tuples;
1607 :
1608 : /* tlist eval costs are paid per output row, not per tuple scanned */
1609 53556 : startup_cost += path->pathtarget->cost.startup;
1610 53556 : run_cost += path->pathtarget->cost.per_tuple * path->rows;
1611 :
1612 53556 : if (path->parallel_workers == 0)
1613 53556 : enable_mask |= PGS_CONSIDER_NONPARTIAL;
1614 53556 : path->disabled_nodes =
1615 53556 : (baserel->pgs_mask & enable_mask) != enable_mask ? 1 : 0;
1616 53556 : path->startup_cost = startup_cost;
1617 53556 : path->total_cost = startup_cost + run_cost;
1618 53556 : }
1619 :
1620 : /*
1621 : * cost_tablefuncscan
1622 : * Determines and returns the cost of scanning a table function.
1623 : *
1624 : * 'baserel' is the relation to be scanned
1625 : * 'param_info' is the ParamPathInfo if this is a parameterized path, else NULL
1626 : */
1627 : void
1628 626 : cost_tablefuncscan(Path *path, PlannerInfo *root,
1629 : RelOptInfo *baserel, ParamPathInfo *param_info)
1630 : {
1631 626 : Cost startup_cost = 0;
1632 626 : Cost run_cost = 0;
1633 : QualCost qpqual_cost;
1634 : Cost cpu_per_tuple;
1635 : RangeTblEntry *rte;
1636 : QualCost exprcost;
1637 626 : uint64 enable_mask = 0;
1638 :
1639 : /* Should only be applied to base relations that are functions */
1640 : Assert(baserel->relid > 0);
1641 626 : rte = planner_rt_fetch(baserel->relid, root);
1642 : Assert(rte->rtekind == RTE_TABLEFUNC);
1643 :
1644 : /* Mark the path with the correct row estimate */
1645 626 : if (param_info)
1646 234 : path->rows = param_info->ppi_rows;
1647 : else
1648 392 : path->rows = baserel->rows;
1649 :
1650 : /*
1651 : * Estimate costs of executing the table func expression(s).
1652 : *
1653 : * XXX in principle we ought to charge tuplestore spill costs if the
1654 : * number of rows is large. However, given how phony our rowcount
1655 : * estimates for tablefuncs tend to be, there's not a lot of point in that
1656 : * refinement right now.
1657 : */
1658 626 : cost_qual_eval_node(&exprcost, (Node *) rte->tablefunc, root);
1659 :
1660 626 : startup_cost += exprcost.startup + exprcost.per_tuple;
1661 :
1662 : /* Add scanning CPU costs */
1663 626 : get_restriction_qual_cost(root, baserel, param_info, &qpqual_cost);
1664 :
1665 626 : startup_cost += qpqual_cost.startup;
1666 626 : cpu_per_tuple = cpu_tuple_cost + qpqual_cost.per_tuple;
1667 626 : run_cost += cpu_per_tuple * baserel->tuples;
1668 :
1669 : /* tlist eval costs are paid per output row, not per tuple scanned */
1670 626 : startup_cost += path->pathtarget->cost.startup;
1671 626 : run_cost += path->pathtarget->cost.per_tuple * path->rows;
1672 :
1673 626 : if (path->parallel_workers == 0)
1674 626 : enable_mask |= PGS_CONSIDER_NONPARTIAL;
1675 626 : path->disabled_nodes =
1676 626 : (baserel->pgs_mask & enable_mask) != enable_mask ? 1 : 0;
1677 626 : path->startup_cost = startup_cost;
1678 626 : path->total_cost = startup_cost + run_cost;
1679 626 : }
1680 :
1681 : /*
1682 : * cost_valuesscan
1683 : * Determines and returns the cost of scanning a VALUES RTE.
1684 : *
1685 : * 'baserel' is the relation to be scanned
1686 : * 'param_info' is the ParamPathInfo if this is a parameterized path, else NULL
1687 : */
1688 : void
1689 8520 : cost_valuesscan(Path *path, PlannerInfo *root,
1690 : RelOptInfo *baserel, ParamPathInfo *param_info)
1691 : {
1692 8520 : Cost startup_cost = 0;
1693 8520 : Cost run_cost = 0;
1694 : QualCost qpqual_cost;
1695 : Cost cpu_per_tuple;
1696 8520 : uint64 enable_mask = 0;
1697 :
1698 : /* Should only be applied to base relations that are values lists */
1699 : Assert(baserel->relid > 0);
1700 : Assert(baserel->rtekind == RTE_VALUES);
1701 :
1702 : /* Mark the path with the correct row estimate */
1703 8520 : if (param_info)
1704 66 : path->rows = param_info->ppi_rows;
1705 : else
1706 8454 : path->rows = baserel->rows;
1707 :
1708 : /*
1709 : * For now, estimate list evaluation cost at one operator eval per list
1710 : * (probably pretty bogus, but is it worth being smarter?)
1711 : */
1712 8520 : cpu_per_tuple = cpu_operator_cost;
1713 :
1714 : /* Add scanning CPU costs */
1715 8520 : get_restriction_qual_cost(root, baserel, param_info, &qpqual_cost);
1716 :
1717 8520 : startup_cost += qpqual_cost.startup;
1718 8520 : cpu_per_tuple += cpu_tuple_cost + qpqual_cost.per_tuple;
1719 8520 : run_cost += cpu_per_tuple * baserel->tuples;
1720 :
1721 : /* tlist eval costs are paid per output row, not per tuple scanned */
1722 8520 : startup_cost += path->pathtarget->cost.startup;
1723 8520 : run_cost += path->pathtarget->cost.per_tuple * path->rows;
1724 :
1725 8520 : if (path->parallel_workers == 0)
1726 8520 : enable_mask |= PGS_CONSIDER_NONPARTIAL;
1727 8520 : path->disabled_nodes =
1728 8520 : (baserel->pgs_mask & enable_mask) != enable_mask ? 1 : 0;
1729 8520 : path->startup_cost = startup_cost;
1730 8520 : path->total_cost = startup_cost + run_cost;
1731 8520 : }
1732 :
1733 : /*
1734 : * cost_ctescan
1735 : * Determines and returns the cost of scanning a CTE RTE.
1736 : *
1737 : * Note: this is used for both self-reference and regular CTEs; the
1738 : * possible cost differences are below the threshold of what we could
1739 : * estimate accurately anyway. Note that the costs of evaluating the
1740 : * referenced CTE query are added into the final plan as initplan costs,
1741 : * and should NOT be counted here.
1742 : */
1743 : void
1744 5312 : cost_ctescan(Path *path, PlannerInfo *root,
1745 : RelOptInfo *baserel, ParamPathInfo *param_info)
1746 : {
1747 5312 : Cost startup_cost = 0;
1748 5312 : Cost run_cost = 0;
1749 : QualCost qpqual_cost;
1750 : Cost cpu_per_tuple;
1751 5312 : uint64 enable_mask = 0;
1752 :
1753 : /* Should only be applied to base relations that are CTEs */
1754 : Assert(baserel->relid > 0);
1755 : Assert(baserel->rtekind == RTE_CTE);
1756 :
1757 : /* Mark the path with the correct row estimate */
1758 5312 : if (param_info)
1759 0 : path->rows = param_info->ppi_rows;
1760 : else
1761 5312 : path->rows = baserel->rows;
1762 :
1763 : /* Charge one CPU tuple cost per row for tuplestore manipulation */
1764 5312 : cpu_per_tuple = cpu_tuple_cost;
1765 :
1766 : /* Add scanning CPU costs */
1767 5312 : get_restriction_qual_cost(root, baserel, param_info, &qpqual_cost);
1768 :
1769 5312 : startup_cost += qpqual_cost.startup;
1770 5312 : cpu_per_tuple += cpu_tuple_cost + qpqual_cost.per_tuple;
1771 5312 : run_cost += cpu_per_tuple * baserel->tuples;
1772 :
1773 : /* tlist eval costs are paid per output row, not per tuple scanned */
1774 5312 : startup_cost += path->pathtarget->cost.startup;
1775 5312 : run_cost += path->pathtarget->cost.per_tuple * path->rows;
1776 :
1777 5312 : if (path->parallel_workers == 0)
1778 5312 : enable_mask |= PGS_CONSIDER_NONPARTIAL;
1779 5312 : path->disabled_nodes =
1780 5312 : (baserel->pgs_mask & enable_mask) != enable_mask ? 1 : 0;
1781 5312 : path->startup_cost = startup_cost;
1782 5312 : path->total_cost = startup_cost + run_cost;
1783 5312 : }
1784 :
1785 : /*
1786 : * cost_namedtuplestorescan
1787 : * Determines and returns the cost of scanning a named tuplestore.
1788 : */
1789 : void
1790 474 : cost_namedtuplestorescan(Path *path, PlannerInfo *root,
1791 : RelOptInfo *baserel, ParamPathInfo *param_info)
1792 : {
1793 474 : Cost startup_cost = 0;
1794 474 : Cost run_cost = 0;
1795 : QualCost qpqual_cost;
1796 : Cost cpu_per_tuple;
1797 474 : uint64 enable_mask = 0;
1798 :
1799 : /* Should only be applied to base relations that are Tuplestores */
1800 : Assert(baserel->relid > 0);
1801 : Assert(baserel->rtekind == RTE_NAMEDTUPLESTORE);
1802 :
1803 : /* Mark the path with the correct row estimate */
1804 474 : if (param_info)
1805 0 : path->rows = param_info->ppi_rows;
1806 : else
1807 474 : path->rows = baserel->rows;
1808 :
1809 : /* Charge one CPU tuple cost per row for tuplestore manipulation */
1810 474 : cpu_per_tuple = cpu_tuple_cost;
1811 :
1812 : /* Add scanning CPU costs */
1813 474 : get_restriction_qual_cost(root, baserel, param_info, &qpqual_cost);
1814 :
1815 474 : startup_cost += qpqual_cost.startup;
1816 474 : cpu_per_tuple += cpu_tuple_cost + qpqual_cost.per_tuple;
1817 474 : run_cost += cpu_per_tuple * baserel->tuples;
1818 :
1819 474 : if (path->parallel_workers == 0)
1820 474 : enable_mask |= PGS_CONSIDER_NONPARTIAL;
1821 474 : path->disabled_nodes =
1822 474 : (baserel->pgs_mask & enable_mask) != enable_mask ? 1 : 0;
1823 474 : path->startup_cost = startup_cost;
1824 474 : path->total_cost = startup_cost + run_cost;
1825 474 : }
1826 :
1827 : /*
1828 : * cost_resultscan
1829 : * Determines and returns the cost of scanning an RTE_RESULT relation.
1830 : */
1831 : void
1832 4352 : cost_resultscan(Path *path, PlannerInfo *root,
1833 : RelOptInfo *baserel, ParamPathInfo *param_info)
1834 : {
1835 4352 : Cost startup_cost = 0;
1836 4352 : Cost run_cost = 0;
1837 : QualCost qpqual_cost;
1838 : Cost cpu_per_tuple;
1839 4352 : uint64 enable_mask = 0;
1840 :
1841 : /* Should only be applied to RTE_RESULT base relations */
1842 : Assert(baserel->relid > 0);
1843 : Assert(baserel->rtekind == RTE_RESULT);
1844 :
1845 : /* Mark the path with the correct row estimate */
1846 4352 : if (param_info)
1847 156 : path->rows = param_info->ppi_rows;
1848 : else
1849 4196 : path->rows = baserel->rows;
1850 :
1851 : /* We charge qual cost plus cpu_tuple_cost */
1852 4352 : get_restriction_qual_cost(root, baserel, param_info, &qpqual_cost);
1853 :
1854 4352 : startup_cost += qpqual_cost.startup;
1855 4352 : cpu_per_tuple = cpu_tuple_cost + qpqual_cost.per_tuple;
1856 4352 : run_cost += cpu_per_tuple * baserel->tuples;
1857 :
1858 4352 : if (path->parallel_workers == 0)
1859 4352 : enable_mask |= PGS_CONSIDER_NONPARTIAL;
1860 4352 : path->disabled_nodes =
1861 4352 : (baserel->pgs_mask & enable_mask) != enable_mask ? 1 : 0;
1862 4352 : path->startup_cost = startup_cost;
1863 4352 : path->total_cost = startup_cost + run_cost;
1864 4352 : }
1865 :
1866 : /*
1867 : * cost_recursive_union
1868 : * Determines and returns the cost of performing a recursive union,
1869 : * and also the estimated output size.
1870 : *
1871 : * We are given Paths for the nonrecursive and recursive terms.
1872 : */
1873 : void
1874 936 : cost_recursive_union(Path *runion, Path *nrterm, Path *rterm)
1875 : {
1876 : Cost startup_cost;
1877 : Cost total_cost;
1878 : double total_rows;
1879 936 : uint64 enable_mask = 0;
1880 :
1881 : /* We probably have decent estimates for the non-recursive term */
1882 936 : startup_cost = nrterm->startup_cost;
1883 936 : total_cost = nrterm->total_cost;
1884 936 : total_rows = nrterm->rows;
1885 :
1886 : /*
1887 : * We arbitrarily assume that about 10 recursive iterations will be
1888 : * needed, and that we've managed to get a good fix on the cost and output
1889 : * size of each one of them. These are mighty shaky assumptions but it's
1890 : * hard to see how to do better.
1891 : */
1892 936 : total_cost += 10 * rterm->total_cost;
1893 936 : total_rows += 10 * rterm->rows;
1894 :
1895 : /*
1896 : * Also charge cpu_tuple_cost per row to account for the costs of
1897 : * manipulating the tuplestores. (We don't worry about possible
1898 : * spill-to-disk costs.)
1899 : */
1900 936 : total_cost += cpu_tuple_cost * total_rows;
1901 :
1902 936 : if (runion->parallel_workers == 0)
1903 936 : enable_mask |= PGS_CONSIDER_NONPARTIAL;
1904 936 : runion->disabled_nodes =
1905 936 : (runion->parent->pgs_mask & enable_mask) != enable_mask ? 1 : 0;
1906 936 : runion->startup_cost = startup_cost;
1907 936 : runion->total_cost = total_cost;
1908 936 : runion->rows = total_rows;
1909 936 : runion->pathtarget->width = Max(nrterm->pathtarget->width,
1910 : rterm->pathtarget->width);
1911 936 : }
1912 :
1913 : /*
1914 : * cost_tuplesort
1915 : * Determines and returns the cost of sorting a relation using tuplesort,
1916 : * not including the cost of reading the input data.
1917 : *
1918 : * If the total volume of data to sort is less than sort_mem, we will do
1919 : * an in-memory sort, which requires no I/O and about t*log2(t) tuple
1920 : * comparisons for t tuples.
1921 : *
1922 : * If the total volume exceeds sort_mem, we switch to a tape-style merge
1923 : * algorithm. There will still be about t*log2(t) tuple comparisons in
1924 : * total, but we will also need to write and read each tuple once per
1925 : * merge pass. We expect about ceil(logM(r)) merge passes where r is the
1926 : * number of initial runs formed and M is the merge order used by tuplesort.c.
1927 : * Since the average initial run should be about sort_mem, we have
1928 : * disk traffic = 2 * relsize * ceil(logM(p / sort_mem))
1929 : * cpu = comparison_cost * t * log2(t)
1930 : *
1931 : * If the sort is bounded (i.e., only the first k result tuples are needed)
1932 : * and k tuples can fit into sort_mem, we use a heap method that keeps only
1933 : * k tuples in the heap; this will require about t*log2(k) tuple comparisons.
1934 : *
1935 : * The disk traffic is assumed to be 3/4ths sequential and 1/4th random
1936 : * accesses (XXX can't we refine that guess?)
1937 : *
1938 : * By default, we charge two operator evals per tuple comparison, which should
1939 : * be in the right ballpark in most cases. The caller can tweak this by
1940 : * specifying nonzero comparison_cost; typically that's used for any extra
1941 : * work that has to be done to prepare the inputs to the comparison operators.
1942 : *
1943 : * 'tuples' is the number of tuples in the relation
1944 : * 'width' is the average tuple width in bytes
1945 : * 'comparison_cost' is the extra cost per comparison, if any
1946 : * 'sort_mem' is the number of kilobytes of work memory allowed for the sort
1947 : * 'limit_tuples' is the bound on the number of output tuples; -1 if no bound
1948 : */
1949 : static void
1950 2147642 : cost_tuplesort(Cost *startup_cost, Cost *run_cost,
1951 : double tuples, int width,
1952 : Cost comparison_cost, int sort_mem,
1953 : double limit_tuples)
1954 : {
1955 2147642 : double input_bytes = relation_byte_size(tuples, width);
1956 : double output_bytes;
1957 : double output_tuples;
1958 2147642 : int64 sort_mem_bytes = sort_mem * (int64) 1024;
1959 :
1960 : /*
1961 : * We want to be sure the cost of a sort is never estimated as zero, even
1962 : * if passed-in tuple count is zero. Besides, mustn't do log(0)...
1963 : */
1964 2147642 : if (tuples < 2.0)
1965 581902 : tuples = 2.0;
1966 :
1967 : /* Include the default cost-per-comparison */
1968 2147642 : comparison_cost += 2.0 * cpu_operator_cost;
1969 :
1970 : /* Do we have a useful LIMIT? */
1971 2147642 : if (limit_tuples > 0 && limit_tuples < tuples)
1972 : {
1973 1826 : output_tuples = limit_tuples;
1974 1826 : output_bytes = relation_byte_size(output_tuples, width);
1975 : }
1976 : else
1977 : {
1978 2145816 : output_tuples = tuples;
1979 2145816 : output_bytes = input_bytes;
1980 : }
1981 :
1982 2147642 : if (output_bytes > sort_mem_bytes)
1983 : {
1984 : /*
1985 : * We'll have to use a disk-based sort of all the tuples
1986 : */
1987 17658 : double npages = ceil(input_bytes / BLCKSZ);
1988 17658 : double nruns = input_bytes / sort_mem_bytes;
1989 17658 : double mergeorder = tuplesort_merge_order(sort_mem_bytes);
1990 : double log_runs;
1991 : double npageaccesses;
1992 :
1993 : /*
1994 : * CPU costs
1995 : *
1996 : * Assume about N log2 N comparisons
1997 : */
1998 17658 : *startup_cost = comparison_cost * tuples * LOG2(tuples);
1999 :
2000 : /* Disk costs */
2001 :
2002 : /* Compute logM(r) as log(r) / log(M) */
2003 17658 : if (nruns > mergeorder)
2004 5274 : log_runs = ceil(log(nruns) / log(mergeorder));
2005 : else
2006 12384 : log_runs = 1.0;
2007 17658 : npageaccesses = 2.0 * npages * log_runs;
2008 : /* Assume 3/4ths of accesses are sequential, 1/4th are not */
2009 17658 : *startup_cost += npageaccesses *
2010 17658 : (seq_page_cost * 0.75 + random_page_cost * 0.25);
2011 : }
2012 2129984 : else if (tuples > 2 * output_tuples || input_bytes > sort_mem_bytes)
2013 : {
2014 : /*
2015 : * We'll use a bounded heap-sort keeping just K tuples in memory, for
2016 : * a total number of tuple comparisons of N log2 K; but the constant
2017 : * factor is a bit higher than for quicksort. Tweak it so that the
2018 : * cost curve is continuous at the crossover point.
2019 : */
2020 1346 : *startup_cost = comparison_cost * tuples * LOG2(2.0 * output_tuples);
2021 : }
2022 : else
2023 : {
2024 : /* We'll use plain quicksort on all the input tuples */
2025 2128638 : *startup_cost = comparison_cost * tuples * LOG2(tuples);
2026 : }
2027 :
2028 : /*
2029 : * Also charge a small amount (arbitrarily set equal to operator cost) per
2030 : * extracted tuple. We don't charge cpu_tuple_cost because a Sort node
2031 : * doesn't do qual-checking or projection, so it has less overhead than
2032 : * most plan nodes. Note it's correct to use tuples not output_tuples
2033 : * here --- the upper LIMIT will pro-rate the run cost so we'd be double
2034 : * counting the LIMIT otherwise.
2035 : */
2036 2147642 : *run_cost = cpu_operator_cost * tuples;
2037 2147642 : }
2038 :
2039 : /*
2040 : * cost_incremental_sort
2041 : * Determines and returns the cost of sorting a relation incrementally, when
2042 : * the input path is presorted by a prefix of the pathkeys.
2043 : *
2044 : * 'presorted_keys' is the number of leading pathkeys by which the input path
2045 : * is sorted.
2046 : *
2047 : * We estimate the number of groups into which the relation is divided by the
2048 : * leading pathkeys, and then calculate the cost of sorting a single group
2049 : * with tuplesort using cost_tuplesort().
2050 : */
2051 : void
2052 12890 : cost_incremental_sort(Path *path,
2053 : PlannerInfo *root, List *pathkeys, int presorted_keys,
2054 : int input_disabled_nodes,
2055 : Cost input_startup_cost, Cost input_total_cost,
2056 : double input_tuples, int width, Cost comparison_cost, int sort_mem,
2057 : double limit_tuples)
2058 : {
2059 : Cost startup_cost,
2060 : run_cost,
2061 12890 : input_run_cost = input_total_cost - input_startup_cost;
2062 : double group_tuples,
2063 : input_groups;
2064 : Cost group_startup_cost,
2065 : group_run_cost,
2066 : group_input_run_cost;
2067 12890 : List *presortedExprs = NIL;
2068 : ListCell *l;
2069 12890 : bool unknown_varno = false;
2070 :
2071 : Assert(presorted_keys > 0 && presorted_keys < list_length(pathkeys));
2072 :
2073 : /*
2074 : * We want to be sure the cost of a sort is never estimated as zero, even
2075 : * if passed-in tuple count is zero. Besides, mustn't do log(0)...
2076 : */
2077 12890 : if (input_tuples < 2.0)
2078 6792 : input_tuples = 2.0;
2079 :
2080 : /* Default estimate of number of groups, capped to one group per row. */
2081 12890 : input_groups = Min(input_tuples, DEFAULT_NUM_DISTINCT);
2082 :
2083 : /*
2084 : * Extract presorted keys as list of expressions.
2085 : *
2086 : * We need to be careful about Vars containing "varno 0" which might have
2087 : * been introduced by generate_append_tlist, which would confuse
2088 : * estimate_num_groups (in fact it'd fail for such expressions). See
2089 : * recurse_set_operations which has to deal with the same issue.
2090 : *
2091 : * Unlike recurse_set_operations we can't access the original target list
2092 : * here, and even if we could it's not very clear how useful would that be
2093 : * for a set operation combining multiple tables. So we simply detect if
2094 : * there are any expressions with "varno 0" and use the default
2095 : * DEFAULT_NUM_DISTINCT in that case.
2096 : *
2097 : * We might also use either 1.0 (a single group) or input_tuples (each row
2098 : * being a separate group), pretty much the worst and best case for
2099 : * incremental sort. But those are extreme cases and using something in
2100 : * between seems reasonable. Furthermore, generate_append_tlist is used
2101 : * for set operations, which are likely to produce mostly unique output
2102 : * anyway - from that standpoint the DEFAULT_NUM_DISTINCT is defensive
2103 : * while maintaining lower startup cost.
2104 : */
2105 12986 : foreach(l, pathkeys)
2106 : {
2107 12986 : PathKey *key = (PathKey *) lfirst(l);
2108 12986 : EquivalenceMember *member = (EquivalenceMember *)
2109 12986 : linitial(key->pk_eclass->ec_members);
2110 :
2111 : /*
2112 : * Check if the expression contains Var with "varno 0" so that we
2113 : * don't call estimate_num_groups in that case.
2114 : */
2115 12986 : if (bms_is_member(0, pull_varnos(root, (Node *) member->em_expr)))
2116 : {
2117 10 : unknown_varno = true;
2118 10 : break;
2119 : }
2120 :
2121 : /* expression not containing any Vars with "varno 0" */
2122 12976 : presortedExprs = lappend(presortedExprs, member->em_expr);
2123 :
2124 12976 : if (foreach_current_index(l) + 1 >= presorted_keys)
2125 12880 : break;
2126 : }
2127 :
2128 : /* Estimate the number of groups with equal presorted keys. */
2129 12890 : if (!unknown_varno)
2130 12880 : input_groups = estimate_num_groups(root, presortedExprs, input_tuples,
2131 : NULL, NULL);
2132 :
2133 12890 : group_tuples = input_tuples / input_groups;
2134 12890 : group_input_run_cost = input_run_cost / input_groups;
2135 :
2136 : /*
2137 : * Estimate the average cost of sorting of one group where presorted keys
2138 : * are equal.
2139 : */
2140 12890 : cost_tuplesort(&group_startup_cost, &group_run_cost,
2141 : group_tuples, width, comparison_cost, sort_mem,
2142 : limit_tuples);
2143 :
2144 : /*
2145 : * Startup cost of incremental sort is the startup cost of its first group
2146 : * plus the cost of its input.
2147 : */
2148 12890 : startup_cost = group_startup_cost + input_startup_cost +
2149 : group_input_run_cost;
2150 :
2151 : /*
2152 : * After we started producing tuples from the first group, the cost of
2153 : * producing all the tuples is given by the cost to finish processing this
2154 : * group, plus the total cost to process the remaining groups, plus the
2155 : * remaining cost of input.
2156 : */
2157 12890 : run_cost = group_run_cost + (group_run_cost + group_startup_cost) *
2158 12890 : (input_groups - 1) + group_input_run_cost * (input_groups - 1);
2159 :
2160 : /*
2161 : * Incremental sort adds some overhead by itself. Firstly, it has to
2162 : * detect the sort groups. This is roughly equal to one extra copy and
2163 : * comparison per tuple.
2164 : */
2165 12890 : run_cost += (cpu_tuple_cost + comparison_cost) * input_tuples;
2166 :
2167 : /*
2168 : * Additionally, we charge double cpu_tuple_cost for each input group to
2169 : * account for the tuplesort_reset that's performed after each group.
2170 : */
2171 12890 : run_cost += 2.0 * cpu_tuple_cost * input_groups;
2172 :
2173 12890 : path->rows = input_tuples;
2174 :
2175 : /*
2176 : * We should not generate these paths when enable_incremental_sort=false.
2177 : * We can ignore PGS_CONSIDER_NONPARTIAL here, because if it's relevant,
2178 : * it will have already affected the input path.
2179 : */
2180 : Assert(enable_incremental_sort);
2181 12890 : path->disabled_nodes = input_disabled_nodes;
2182 :
2183 12890 : path->startup_cost = startup_cost;
2184 12890 : path->total_cost = startup_cost + run_cost;
2185 12890 : }
2186 :
2187 : /*
2188 : * cost_sort
2189 : * Determines and returns the cost of sorting a relation, including
2190 : * the cost of reading the input data.
2191 : *
2192 : * NOTE: some callers currently pass NIL for pathkeys because they
2193 : * can't conveniently supply the sort keys. Since this routine doesn't
2194 : * currently do anything with pathkeys anyway, that doesn't matter...
2195 : * but if it ever does, it should react gracefully to lack of key data.
2196 : * (Actually, the thing we'd most likely be interested in is just the number
2197 : * of sort keys, which all callers *could* supply.)
2198 : */
2199 : void
2200 2134752 : cost_sort(Path *path, PlannerInfo *root,
2201 : List *pathkeys, int input_disabled_nodes,
2202 : Cost input_cost, double tuples, int width,
2203 : Cost comparison_cost, int sort_mem,
2204 : double limit_tuples)
2205 :
2206 : {
2207 : Cost startup_cost;
2208 : Cost run_cost;
2209 :
2210 2134752 : cost_tuplesort(&startup_cost, &run_cost,
2211 : tuples, width,
2212 : comparison_cost, sort_mem,
2213 : limit_tuples);
2214 :
2215 2134752 : startup_cost += input_cost;
2216 :
2217 : /*
2218 : * We can ignore PGS_CONSIDER_NONPARTIAL here, because if it's relevant,
2219 : * it will have already affected the input path.
2220 : */
2221 2134752 : path->rows = tuples;
2222 2134752 : path->disabled_nodes = input_disabled_nodes + (enable_sort ? 0 : 1);
2223 2134752 : path->startup_cost = startup_cost;
2224 2134752 : path->total_cost = startup_cost + run_cost;
2225 2134752 : }
2226 :
2227 : /*
2228 : * append_nonpartial_cost
2229 : * Estimate the cost of the non-partial paths in a Parallel Append.
2230 : * The non-partial paths are assumed to be the first "numpaths" paths
2231 : * from the subpaths list, and to be in order of decreasing cost.
2232 : */
2233 : static Cost
2234 26122 : append_nonpartial_cost(List *subpaths, int numpaths, int parallel_workers)
2235 : {
2236 : Cost *costarr;
2237 : int arrlen;
2238 : ListCell *l;
2239 : ListCell *cell;
2240 : int path_index;
2241 : int min_index;
2242 : int max_index;
2243 :
2244 26122 : if (numpaths == 0)
2245 21072 : return 0;
2246 :
2247 : /*
2248 : * Array length is number of workers or number of relevant paths,
2249 : * whichever is less.
2250 : */
2251 5050 : arrlen = Min(parallel_workers, numpaths);
2252 5050 : costarr = palloc_array(Cost, arrlen);
2253 :
2254 : /* The first few paths will each be claimed by a different worker. */
2255 5050 : path_index = 0;
2256 14676 : foreach(cell, subpaths)
2257 : {
2258 10928 : Path *subpath = (Path *) lfirst(cell);
2259 :
2260 10928 : if (path_index == arrlen)
2261 1302 : break;
2262 9626 : costarr[path_index++] = subpath->total_cost;
2263 : }
2264 :
2265 : /*
2266 : * Since subpaths are sorted by decreasing cost, the last one will have
2267 : * the minimum cost.
2268 : */
2269 5050 : min_index = arrlen - 1;
2270 :
2271 : /*
2272 : * For each of the remaining subpaths, add its cost to the array element
2273 : * with minimum cost.
2274 : */
2275 9802 : for_each_cell(l, subpaths, cell)
2276 : {
2277 5298 : Path *subpath = (Path *) lfirst(l);
2278 :
2279 : /* Consider only the non-partial paths */
2280 5298 : if (path_index++ == numpaths)
2281 546 : break;
2282 :
2283 4752 : costarr[min_index] += subpath->total_cost;
2284 :
2285 : /* Update the new min cost array index */
2286 4752 : min_index = 0;
2287 14292 : for (int i = 0; i < arrlen; i++)
2288 : {
2289 9540 : if (costarr[i] < costarr[min_index])
2290 1532 : min_index = i;
2291 : }
2292 : }
2293 :
2294 : /* Return the highest cost from the array */
2295 5050 : max_index = 0;
2296 14676 : for (int i = 0; i < arrlen; i++)
2297 : {
2298 9626 : if (costarr[i] > costarr[max_index])
2299 418 : max_index = i;
2300 : }
2301 :
2302 5050 : return costarr[max_index];
2303 : }
2304 :
2305 : /*
2306 : * cost_append
2307 : * Determines and returns the cost of an Append node.
2308 : */
2309 : void
2310 70896 : cost_append(AppendPath *apath, PlannerInfo *root)
2311 : {
2312 70896 : RelOptInfo *rel = apath->path.parent;
2313 : ListCell *l;
2314 70896 : uint64 enable_mask = PGS_APPEND;
2315 :
2316 70896 : if (apath->path.parallel_workers == 0)
2317 44726 : enable_mask |= PGS_CONSIDER_NONPARTIAL;
2318 :
2319 70896 : apath->path.disabled_nodes =
2320 70896 : (rel->pgs_mask & enable_mask) == enable_mask ? 0 : 1;
2321 70896 : apath->path.startup_cost = 0;
2322 70896 : apath->path.total_cost = 0;
2323 70896 : apath->path.rows = 0;
2324 :
2325 70896 : if (apath->subpaths == NIL)
2326 2084 : return;
2327 :
2328 68812 : if (!apath->path.parallel_aware)
2329 : {
2330 42690 : List *pathkeys = apath->path.pathkeys;
2331 :
2332 42690 : if (pathkeys == NIL)
2333 : {
2334 40546 : Path *firstsubpath = (Path *) linitial(apath->subpaths);
2335 :
2336 : /*
2337 : * For an unordered, non-parallel-aware Append we take the startup
2338 : * cost as the startup cost of the first subpath.
2339 : */
2340 40546 : apath->path.startup_cost = firstsubpath->startup_cost;
2341 :
2342 : /*
2343 : * Compute rows, number of disabled nodes, and total cost as sums
2344 : * of underlying subplan values.
2345 : */
2346 158666 : foreach(l, apath->subpaths)
2347 : {
2348 118120 : Path *subpath = (Path *) lfirst(l);
2349 :
2350 118120 : apath->path.rows += subpath->rows;
2351 118120 : apath->path.disabled_nodes += subpath->disabled_nodes;
2352 118120 : apath->path.total_cost += subpath->total_cost;
2353 : }
2354 : }
2355 : else
2356 : {
2357 : /*
2358 : * For an ordered, non-parallel-aware Append we take the startup
2359 : * cost as the sum of the subpath startup costs. This ensures
2360 : * that we don't underestimate the startup cost when a query's
2361 : * LIMIT is such that several of the children have to be run to
2362 : * satisfy it. This might be overkill --- another plausible hack
2363 : * would be to take the Append's startup cost as the maximum of
2364 : * the child startup costs. But we don't want to risk believing
2365 : * that an ORDER BY LIMIT query can be satisfied at small cost
2366 : * when the first child has small startup cost but later ones
2367 : * don't. (If we had the ability to deal with nonlinear cost
2368 : * interpolation for partial retrievals, we would not need to be
2369 : * so conservative about this.)
2370 : *
2371 : * This case is also different from the above in that we have to
2372 : * account for possibly injecting sorts into subpaths that aren't
2373 : * natively ordered.
2374 : */
2375 8340 : foreach(l, apath->subpaths)
2376 : {
2377 6196 : Path *subpath = (Path *) lfirst(l);
2378 : int presorted_keys;
2379 : Path sort_path; /* dummy for result of
2380 : * cost_sort/cost_incremental_sort */
2381 :
2382 6196 : if (!pathkeys_count_contained_in(pathkeys, subpath->pathkeys,
2383 : &presorted_keys))
2384 : {
2385 : /*
2386 : * We'll need to insert a Sort node, so include costs for
2387 : * that. We choose to use incremental sort if it is
2388 : * enabled and there are presorted keys; otherwise we use
2389 : * full sort.
2390 : *
2391 : * We can use the parent's LIMIT if any, since we
2392 : * certainly won't pull more than that many tuples from
2393 : * any child.
2394 : */
2395 44 : if (enable_incremental_sort && presorted_keys > 0)
2396 : {
2397 12 : cost_incremental_sort(&sort_path,
2398 : root,
2399 : pathkeys,
2400 : presorted_keys,
2401 : subpath->disabled_nodes,
2402 : subpath->startup_cost,
2403 : subpath->total_cost,
2404 : subpath->rows,
2405 12 : subpath->pathtarget->width,
2406 : 0.0,
2407 : work_mem,
2408 : apath->limit_tuples);
2409 : }
2410 : else
2411 : {
2412 32 : cost_sort(&sort_path,
2413 : root,
2414 : pathkeys,
2415 : subpath->disabled_nodes,
2416 : subpath->total_cost,
2417 : subpath->rows,
2418 32 : subpath->pathtarget->width,
2419 : 0.0,
2420 : work_mem,
2421 : apath->limit_tuples);
2422 : }
2423 :
2424 44 : subpath = &sort_path;
2425 : }
2426 :
2427 6196 : apath->path.rows += subpath->rows;
2428 6196 : apath->path.disabled_nodes += subpath->disabled_nodes;
2429 6196 : apath->path.startup_cost += subpath->startup_cost;
2430 6196 : apath->path.total_cost += subpath->total_cost;
2431 : }
2432 : }
2433 : }
2434 : else /* parallel-aware */
2435 : {
2436 26122 : int i = 0;
2437 26122 : double parallel_divisor = get_parallel_divisor(&apath->path);
2438 :
2439 : /* Parallel-aware Append never produces ordered output. */
2440 : Assert(apath->path.pathkeys == NIL);
2441 :
2442 : /* Calculate startup cost. */
2443 103542 : foreach(l, apath->subpaths)
2444 : {
2445 77420 : Path *subpath = (Path *) lfirst(l);
2446 :
2447 : /*
2448 : * Append will start returning tuples when the child node having
2449 : * lowest startup cost is done setting up. We consider only the
2450 : * first few subplans that immediately get a worker assigned.
2451 : */
2452 77420 : if (i == 0)
2453 26122 : apath->path.startup_cost = subpath->startup_cost;
2454 51298 : else if (i < apath->path.parallel_workers)
2455 25546 : apath->path.startup_cost = Min(apath->path.startup_cost,
2456 : subpath->startup_cost);
2457 :
2458 : /*
2459 : * Apply parallel divisor to subpaths. Scale the number of rows
2460 : * for each partial subpath based on the ratio of the parallel
2461 : * divisor originally used for the subpath to the one we adopted.
2462 : * Also add the cost of partial paths to the total cost, but
2463 : * ignore non-partial paths for now.
2464 : */
2465 77420 : if (i < apath->first_partial_path)
2466 14378 : apath->path.rows += subpath->rows / parallel_divisor;
2467 : else
2468 : {
2469 : double subpath_parallel_divisor;
2470 :
2471 63042 : subpath_parallel_divisor = get_parallel_divisor(subpath);
2472 63042 : apath->path.rows += subpath->rows * (subpath_parallel_divisor /
2473 : parallel_divisor);
2474 63042 : apath->path.total_cost += subpath->total_cost;
2475 : }
2476 :
2477 77420 : apath->path.disabled_nodes += subpath->disabled_nodes;
2478 77420 : apath->path.rows = clamp_row_est(apath->path.rows);
2479 :
2480 77420 : i++;
2481 : }
2482 :
2483 : /* Add cost for non-partial subpaths. */
2484 26122 : apath->path.total_cost +=
2485 26122 : append_nonpartial_cost(apath->subpaths,
2486 : apath->first_partial_path,
2487 : apath->path.parallel_workers);
2488 : }
2489 :
2490 : /*
2491 : * Although Append does not do any selection or projection, it's not free;
2492 : * add a small per-tuple overhead.
2493 : */
2494 68812 : apath->path.total_cost +=
2495 68812 : cpu_tuple_cost * APPEND_CPU_COST_MULTIPLIER * apath->path.rows;
2496 : }
2497 :
2498 : /*
2499 : * cost_merge_append
2500 : * Determines and returns the cost of a MergeAppend node.
2501 : *
2502 : * MergeAppend merges several pre-sorted input streams, using a heap that
2503 : * at any given instant holds the next tuple from each stream. If there
2504 : * are N streams, we need about N*log2(N) tuple comparisons to construct
2505 : * the heap at startup, and then for each output tuple, about log2(N)
2506 : * comparisons to replace the top entry.
2507 : *
2508 : * (The effective value of N will drop once some of the input streams are
2509 : * exhausted, but it seems unlikely to be worth trying to account for that.)
2510 : *
2511 : * The heap is never spilled to disk, since we assume N is not very large.
2512 : * So this is much simpler than cost_sort.
2513 : *
2514 : * As in cost_sort, we charge two operator evals per tuple comparison.
2515 : *
2516 : * 'pathkeys' is a list of sort keys
2517 : * 'n_streams' is the number of input streams
2518 : * 'input_disabled_nodes' is the sum of the input streams' disabled node counts
2519 : * 'input_startup_cost' is the sum of the input streams' startup costs
2520 : * 'input_total_cost' is the sum of the input streams' total costs
2521 : * 'tuples' is the number of tuples in all the streams
2522 : */
2523 : void
2524 9900 : cost_merge_append(Path *path, PlannerInfo *root,
2525 : List *pathkeys, int n_streams,
2526 : int input_disabled_nodes,
2527 : Cost input_startup_cost, Cost input_total_cost,
2528 : double tuples)
2529 : {
2530 9900 : RelOptInfo *rel = path->parent;
2531 9900 : Cost startup_cost = 0;
2532 9900 : Cost run_cost = 0;
2533 : Cost comparison_cost;
2534 : double N;
2535 : double logN;
2536 9900 : uint64 enable_mask = PGS_MERGE_APPEND;
2537 :
2538 9900 : if (path->parallel_workers == 0)
2539 9900 : enable_mask |= PGS_CONSIDER_NONPARTIAL;
2540 :
2541 : /*
2542 : * Avoid log(0)...
2543 : */
2544 9900 : N = (n_streams < 2) ? 2.0 : (double) n_streams;
2545 9900 : logN = LOG2(N);
2546 :
2547 : /* Assumed cost per tuple comparison */
2548 9900 : comparison_cost = 2.0 * cpu_operator_cost;
2549 :
2550 : /* Heap creation cost */
2551 9900 : startup_cost += comparison_cost * N * logN;
2552 :
2553 : /* Per-tuple heap maintenance cost */
2554 9900 : run_cost += tuples * comparison_cost * logN;
2555 :
2556 : /*
2557 : * Although MergeAppend does not do any selection or projection, it's not
2558 : * free; add a small per-tuple overhead.
2559 : */
2560 9900 : run_cost += cpu_tuple_cost * APPEND_CPU_COST_MULTIPLIER * tuples;
2561 :
2562 9900 : path->disabled_nodes =
2563 9900 : (rel->pgs_mask & enable_mask) == enable_mask ? 0 : 1;
2564 9900 : path->disabled_nodes += input_disabled_nodes;
2565 9900 : path->startup_cost = startup_cost + input_startup_cost;
2566 9900 : path->total_cost = startup_cost + run_cost + input_total_cost;
2567 9900 : }
2568 :
2569 : /*
2570 : * cost_material
2571 : * Determines and returns the cost of materializing a relation, including
2572 : * the cost of reading the input data.
2573 : *
2574 : * If the total volume of data to materialize exceeds work_mem, we will need
2575 : * to write it to disk, so the cost is much higher in that case.
2576 : *
2577 : * Note that here we are estimating the costs for the first scan of the
2578 : * relation, so the materialization is all overhead --- any savings will
2579 : * occur only on rescan, which is estimated in cost_rescan.
2580 : */
2581 : void
2582 694176 : cost_material(Path *path,
2583 : bool enabled, int input_disabled_nodes,
2584 : Cost input_startup_cost, Cost input_total_cost,
2585 : double tuples, int width)
2586 : {
2587 694176 : Cost startup_cost = input_startup_cost;
2588 694176 : Cost run_cost = input_total_cost - input_startup_cost;
2589 694176 : double nbytes = relation_byte_size(tuples, width);
2590 694176 : double work_mem_bytes = work_mem * (Size) 1024;
2591 :
2592 694176 : if (path->parallel_workers == 0 &&
2593 694176 : path->parent != NULL &&
2594 694102 : (path->parent->pgs_mask & PGS_CONSIDER_NONPARTIAL) == 0)
2595 0 : enabled = false;
2596 :
2597 694176 : path->rows = tuples;
2598 :
2599 : /*
2600 : * Whether spilling or not, charge 2x cpu_operator_cost per tuple to
2601 : * reflect bookkeeping overhead. (This rate must be more than what
2602 : * cost_rescan charges for materialize, ie, cpu_operator_cost per tuple;
2603 : * if it is exactly the same then there will be a cost tie between
2604 : * nestloop with A outer, materialized B inner and nestloop with B outer,
2605 : * materialized A inner. The extra cost ensures we'll prefer
2606 : * materializing the smaller rel.) Note that this is normally a good deal
2607 : * less than cpu_tuple_cost; which is OK because a Material plan node
2608 : * doesn't do qual-checking or projection, so it's got less overhead than
2609 : * most plan nodes.
2610 : */
2611 694176 : run_cost += 2 * cpu_operator_cost * tuples;
2612 :
2613 : /*
2614 : * If we will spill to disk, charge at the rate of seq_page_cost per page.
2615 : * This cost is assumed to be evenly spread through the plan run phase,
2616 : * which isn't exactly accurate but our cost model doesn't allow for
2617 : * nonuniform costs within the run phase.
2618 : */
2619 694176 : if (nbytes > work_mem_bytes)
2620 : {
2621 5082 : double npages = ceil(nbytes / BLCKSZ);
2622 :
2623 5082 : run_cost += seq_page_cost * npages;
2624 : }
2625 :
2626 694176 : path->disabled_nodes = input_disabled_nodes + (enabled ? 0 : 1);
2627 694176 : path->startup_cost = startup_cost;
2628 694176 : path->total_cost = startup_cost + run_cost;
2629 694176 : }
2630 :
2631 : /*
2632 : * cost_memoize_rescan
2633 : * Determines the estimated cost of rescanning a Memoize node.
2634 : *
2635 : * In order to estimate this, we must gain knowledge of how often we expect to
2636 : * be called and how many distinct sets of parameters we are likely to be
2637 : * called with. If we expect a good cache hit ratio, then we can set our
2638 : * costs to account for that hit ratio, plus a little bit of cost for the
2639 : * caching itself. Caching will not work out well if we expect to be called
2640 : * with too many distinct parameter values. The worst-case here is that we
2641 : * never see any parameter value twice, in which case we'd never get a cache
2642 : * hit and caching would be a complete waste of effort.
2643 : */
2644 : static void
2645 297590 : cost_memoize_rescan(PlannerInfo *root, MemoizePath *mpath,
2646 : Cost *rescan_startup_cost, Cost *rescan_total_cost)
2647 : {
2648 : EstimationInfo estinfo;
2649 : ListCell *lc;
2650 297590 : Cost input_startup_cost = mpath->subpath->startup_cost;
2651 297590 : Cost input_total_cost = mpath->subpath->total_cost;
2652 297590 : double tuples = mpath->subpath->rows;
2653 297590 : Cardinality est_calls = mpath->est_calls;
2654 297590 : int width = mpath->subpath->pathtarget->width;
2655 :
2656 : double hash_mem_bytes;
2657 : double est_entry_bytes;
2658 : Cardinality est_cache_entries;
2659 : Cardinality ndistinct;
2660 : double evict_ratio;
2661 : double hit_ratio;
2662 : Cost startup_cost;
2663 : Cost total_cost;
2664 :
2665 : /* available cache space */
2666 297590 : hash_mem_bytes = get_hash_memory_limit();
2667 :
2668 : /*
2669 : * Set the number of bytes each cache entry should consume in the cache.
2670 : * To provide us with better estimations on how many cache entries we can
2671 : * store at once, we make a call to the executor here to ask it what
2672 : * memory overheads there are for a single cache entry.
2673 : */
2674 297590 : est_entry_bytes = relation_byte_size(tuples, width) +
2675 297590 : ExecEstimateCacheEntryOverheadBytes(tuples);
2676 :
2677 : /* include the estimated width for the cache keys */
2678 633854 : foreach(lc, mpath->param_exprs)
2679 336264 : est_entry_bytes += get_expr_width(root, (Node *) lfirst(lc));
2680 :
2681 : /* estimate on the upper limit of cache entries we can hold at once */
2682 297590 : est_cache_entries = floor(hash_mem_bytes / est_entry_bytes);
2683 :
2684 : /* estimate on the distinct number of parameter values */
2685 297590 : ndistinct = estimate_num_groups(root, mpath->param_exprs, est_calls, NULL,
2686 : &estinfo);
2687 :
2688 : /*
2689 : * When the estimation fell back on using a default value, it's a bit too
2690 : * risky to assume that it's ok to use a Memoize node. The use of a
2691 : * default could cause us to use a Memoize node when it's really
2692 : * inappropriate to do so. If we see that this has been done, then we'll
2693 : * assume that every call will have unique parameters, which will almost
2694 : * certainly mean a MemoizePath will never survive add_path().
2695 : */
2696 297590 : if ((estinfo.flags & SELFLAG_USED_DEFAULT) != 0)
2697 17398 : ndistinct = est_calls;
2698 :
2699 : /* Remember the ndistinct estimate for EXPLAIN */
2700 297590 : mpath->est_unique_keys = ndistinct;
2701 :
2702 : /*
2703 : * Since we've already estimated the maximum number of entries we can
2704 : * store at once and know the estimated number of distinct values we'll be
2705 : * called with, we'll take this opportunity to set the path's est_entries.
2706 : * This will ultimately determine the hash table size that the executor
2707 : * will use. If we leave this at zero, the executor will just choose the
2708 : * size itself. Really this is not the right place to do this, but it's
2709 : * convenient since everything is already calculated.
2710 : */
2711 297590 : mpath->est_entries = Min(Min(ndistinct, est_cache_entries),
2712 : PG_UINT32_MAX);
2713 :
2714 : /*
2715 : * When the number of distinct parameter values is above the amount we can
2716 : * store in the cache, then we'll have to evict some entries from the
2717 : * cache. This is not free. Here we estimate how often we'll incur the
2718 : * cost of that eviction.
2719 : */
2720 297590 : evict_ratio = 1.0 - Min(est_cache_entries, ndistinct) / ndistinct;
2721 :
2722 : /*
2723 : * In order to estimate how costly a single scan will be, we need to
2724 : * attempt to estimate what the cache hit ratio will be. To do that we
2725 : * must look at how many scans are estimated in total for this node and
2726 : * how many of those scans we expect to get a cache hit.
2727 : */
2728 595180 : hit_ratio = ((est_calls - ndistinct) / est_calls) *
2729 297590 : (est_cache_entries / Max(ndistinct, est_cache_entries));
2730 :
2731 : /* Remember the hit ratio estimate for EXPLAIN */
2732 297590 : mpath->est_hit_ratio = hit_ratio;
2733 :
2734 : Assert(hit_ratio >= 0 && hit_ratio <= 1.0);
2735 :
2736 : /*
2737 : * Set the total_cost accounting for the expected cache hit ratio. We
2738 : * also add on a cpu_operator_cost to account for a cache lookup. This
2739 : * will happen regardless of whether it's a cache hit or not.
2740 : */
2741 297590 : total_cost = input_total_cost * (1.0 - hit_ratio) + cpu_operator_cost;
2742 :
2743 : /* Now adjust the total cost to account for cache evictions */
2744 :
2745 : /* Charge a cpu_tuple_cost for evicting the actual cache entry */
2746 297590 : total_cost += cpu_tuple_cost * evict_ratio;
2747 :
2748 : /*
2749 : * Charge a 10th of cpu_operator_cost to evict every tuple in that entry.
2750 : * The per-tuple eviction is really just a pfree, so charging a whole
2751 : * cpu_operator_cost seems a little excessive.
2752 : */
2753 297590 : total_cost += cpu_operator_cost / 10.0 * evict_ratio * tuples;
2754 :
2755 : /*
2756 : * Now adjust for storing things in the cache, since that's not free
2757 : * either. Everything must go in the cache. We don't proportion this
2758 : * over any ratio, just apply it once for the scan. We charge a
2759 : * cpu_tuple_cost for the creation of the cache entry and also a
2760 : * cpu_operator_cost for each tuple we expect to cache.
2761 : */
2762 297590 : total_cost += cpu_tuple_cost + cpu_operator_cost * tuples;
2763 :
2764 : /*
2765 : * Getting the first row must be also be proportioned according to the
2766 : * expected cache hit ratio.
2767 : */
2768 297590 : startup_cost = input_startup_cost * (1.0 - hit_ratio);
2769 :
2770 : /*
2771 : * Additionally we charge a cpu_tuple_cost to account for cache lookups,
2772 : * which we'll do regardless of whether it was a cache hit or not.
2773 : */
2774 297590 : startup_cost += cpu_tuple_cost;
2775 :
2776 297590 : *rescan_startup_cost = startup_cost;
2777 297590 : *rescan_total_cost = total_cost;
2778 297590 : }
2779 :
2780 : /*
2781 : * cost_agg
2782 : * Determines and returns the cost of performing an Agg plan node,
2783 : * including the cost of its input.
2784 : *
2785 : * aggcosts can be NULL when there are no actual aggregate functions (i.e.,
2786 : * we are using a hashed Agg node just to do grouping).
2787 : *
2788 : * Note: when aggstrategy == AGG_SORTED, caller must ensure that input costs
2789 : * are for appropriately-sorted input.
2790 : */
2791 : void
2792 95282 : cost_agg(Path *path, PlannerInfo *root,
2793 : AggStrategy aggstrategy, const AggClauseCosts *aggcosts,
2794 : int numGroupCols, double numGroups,
2795 : List *quals,
2796 : int disabled_nodes,
2797 : Cost input_startup_cost, Cost input_total_cost,
2798 : double input_tuples, double input_width)
2799 : {
2800 : double output_tuples;
2801 : Cost startup_cost;
2802 : Cost total_cost;
2803 95282 : const AggClauseCosts dummy_aggcosts = {0};
2804 :
2805 : /* Use all-zero per-aggregate costs if NULL is passed */
2806 95282 : if (aggcosts == NULL)
2807 : {
2808 : Assert(aggstrategy == AGG_HASHED);
2809 19244 : aggcosts = &dummy_aggcosts;
2810 : }
2811 :
2812 : /*
2813 : * The transCost.per_tuple component of aggcosts should be charged once
2814 : * per input tuple, corresponding to the costs of evaluating the aggregate
2815 : * transfns and their input expressions. The finalCost.per_tuple component
2816 : * is charged once per output tuple, corresponding to the costs of
2817 : * evaluating the finalfns. Startup costs are of course charged but once.
2818 : *
2819 : * If we are grouping, we charge an additional cpu_operator_cost per
2820 : * grouping column per input tuple for grouping comparisons.
2821 : *
2822 : * We will produce a single output tuple if not grouping, and a tuple per
2823 : * group otherwise. We charge cpu_tuple_cost for each output tuple.
2824 : *
2825 : * Note: in this cost model, AGG_SORTED and AGG_HASHED have exactly the
2826 : * same total CPU cost, but AGG_SORTED has lower startup cost. If the
2827 : * input path is already sorted appropriately, AGG_SORTED should be
2828 : * preferred (since it has no risk of memory overflow). This will happen
2829 : * as long as the computed total costs are indeed exactly equal --- but if
2830 : * there's roundoff error we might do the wrong thing. So be sure that
2831 : * the computations below form the same intermediate values in the same
2832 : * order.
2833 : */
2834 95282 : if (aggstrategy == AGG_PLAIN)
2835 : {
2836 44980 : startup_cost = input_total_cost;
2837 44980 : startup_cost += aggcosts->transCost.startup;
2838 44980 : startup_cost += aggcosts->transCost.per_tuple * input_tuples;
2839 44980 : startup_cost += aggcosts->finalCost.startup;
2840 44980 : startup_cost += aggcosts->finalCost.per_tuple;
2841 : /* we aren't grouping */
2842 44980 : total_cost = startup_cost + cpu_tuple_cost;
2843 44980 : output_tuples = 1;
2844 : }
2845 50302 : else if (aggstrategy == AGG_SORTED || aggstrategy == AGG_MIXED)
2846 : {
2847 : /* Here we are able to deliver output on-the-fly */
2848 18364 : startup_cost = input_startup_cost;
2849 18364 : total_cost = input_total_cost;
2850 18364 : if (aggstrategy == AGG_MIXED && !enable_hashagg)
2851 540 : ++disabled_nodes;
2852 : /* calcs phrased this way to match HASHED case, see note above */
2853 18364 : total_cost += aggcosts->transCost.startup;
2854 18364 : total_cost += aggcosts->transCost.per_tuple * input_tuples;
2855 18364 : total_cost += (cpu_operator_cost * numGroupCols) * input_tuples;
2856 18364 : total_cost += aggcosts->finalCost.startup;
2857 18364 : total_cost += aggcosts->finalCost.per_tuple * numGroups;
2858 18364 : total_cost += cpu_tuple_cost * numGroups;
2859 18364 : output_tuples = numGroups;
2860 : }
2861 : else
2862 : {
2863 : /* must be AGG_HASHED */
2864 31938 : startup_cost = input_total_cost;
2865 31938 : if (!enable_hashagg)
2866 1890 : ++disabled_nodes;
2867 31938 : startup_cost += aggcosts->transCost.startup;
2868 31938 : startup_cost += aggcosts->transCost.per_tuple * input_tuples;
2869 : /* cost of computing hash value */
2870 31938 : startup_cost += (cpu_operator_cost * numGroupCols) * input_tuples;
2871 31938 : startup_cost += aggcosts->finalCost.startup;
2872 :
2873 31938 : total_cost = startup_cost;
2874 31938 : total_cost += aggcosts->finalCost.per_tuple * numGroups;
2875 : /* cost of retrieving from hash table */
2876 31938 : total_cost += cpu_tuple_cost * numGroups;
2877 31938 : output_tuples = numGroups;
2878 : }
2879 :
2880 : /*
2881 : * Add the disk costs of hash aggregation that spills to disk.
2882 : *
2883 : * Groups that go into the hash table stay in memory until finalized, so
2884 : * spilling and reprocessing tuples doesn't incur additional invocations
2885 : * of transCost or finalCost. Furthermore, the computed hash value is
2886 : * stored with the spilled tuples, so we don't incur extra invocations of
2887 : * the hash function.
2888 : *
2889 : * Hash Agg begins returning tuples after the first batch is complete.
2890 : * Accrue writes (spilled tuples) to startup_cost and to total_cost;
2891 : * accrue reads only to total_cost.
2892 : */
2893 95282 : if (aggstrategy == AGG_HASHED || aggstrategy == AGG_MIXED)
2894 : {
2895 : double pages;
2896 32972 : double pages_written = 0.0;
2897 32972 : double pages_read = 0.0;
2898 : double spill_cost;
2899 : double hashentrysize;
2900 : double nbatches;
2901 : Size mem_limit;
2902 : uint64 ngroups_limit;
2903 : int num_partitions;
2904 : int depth;
2905 :
2906 : /*
2907 : * Estimate number of batches based on the computed limits. If less
2908 : * than or equal to one, all groups are expected to fit in memory;
2909 : * otherwise we expect to spill.
2910 : */
2911 32972 : hashentrysize = hash_agg_entry_size(list_length(root->aggtransinfos),
2912 : input_width,
2913 32972 : aggcosts->transitionSpace);
2914 32972 : hash_agg_set_limits(hashentrysize, numGroups, 0, &mem_limit,
2915 : &ngroups_limit, &num_partitions);
2916 :
2917 32972 : nbatches = Max((numGroups * hashentrysize) / mem_limit,
2918 : numGroups / ngroups_limit);
2919 :
2920 32972 : nbatches = Max(ceil(nbatches), 1.0);
2921 32972 : num_partitions = Max(num_partitions, 2);
2922 :
2923 : /*
2924 : * The number of partitions can change at different levels of
2925 : * recursion; but for the purposes of this calculation assume it stays
2926 : * constant.
2927 : */
2928 32972 : depth = ceil(log(nbatches) / log(num_partitions));
2929 :
2930 : /*
2931 : * Estimate number of pages read and written. For each level of
2932 : * recursion, a tuple must be written and then later read.
2933 : */
2934 32972 : pages = relation_byte_size(input_tuples, input_width) / BLCKSZ;
2935 32972 : pages_written = pages_read = pages * depth;
2936 :
2937 : /*
2938 : * HashAgg has somewhat worse IO behavior than Sort on typical
2939 : * hardware/OS combinations. Account for this with a generic penalty.
2940 : */
2941 32972 : pages_read *= 2.0;
2942 32972 : pages_written *= 2.0;
2943 :
2944 32972 : startup_cost += pages_written * random_page_cost;
2945 32972 : total_cost += pages_written * random_page_cost;
2946 32972 : total_cost += pages_read * seq_page_cost;
2947 :
2948 : /* account for CPU cost of spilling a tuple and reading it back */
2949 32972 : spill_cost = depth * input_tuples * 2.0 * cpu_tuple_cost;
2950 32972 : startup_cost += spill_cost;
2951 32972 : total_cost += spill_cost;
2952 : }
2953 :
2954 : /*
2955 : * If there are quals (HAVING quals), account for their cost and
2956 : * selectivity.
2957 : */
2958 95282 : if (quals)
2959 : {
2960 : QualCost qual_cost;
2961 :
2962 4658 : cost_qual_eval(&qual_cost, quals, root);
2963 4658 : startup_cost += qual_cost.startup;
2964 4658 : total_cost += qual_cost.startup + output_tuples * qual_cost.per_tuple;
2965 :
2966 4658 : output_tuples = clamp_row_est(output_tuples *
2967 4658 : clauselist_selectivity(root,
2968 : quals,
2969 : 0,
2970 : JOIN_INNER,
2971 : NULL));
2972 : }
2973 :
2974 95282 : path->rows = output_tuples;
2975 95282 : path->disabled_nodes = disabled_nodes;
2976 95282 : path->startup_cost = startup_cost;
2977 95282 : path->total_cost = total_cost;
2978 95282 : }
2979 :
2980 : /*
2981 : * get_windowclause_startup_tuples
2982 : * Estimate how many tuples we'll need to fetch from a WindowAgg's
2983 : * subnode before we can output the first WindowAgg tuple.
2984 : *
2985 : * How many tuples need to be read depends on the WindowClause. For example,
2986 : * a WindowClause with no PARTITION BY and no ORDER BY requires that all
2987 : * subnode tuples are read and aggregated before the WindowAgg can output
2988 : * anything. If there's a PARTITION BY, then we only need to look at tuples
2989 : * in the first partition. Here we attempt to estimate just how many
2990 : * 'input_tuples' the WindowAgg will need to read for the given WindowClause
2991 : * before the first tuple can be output.
2992 : */
2993 : static double
2994 2976 : get_windowclause_startup_tuples(PlannerInfo *root, WindowClause *wc,
2995 : double input_tuples)
2996 : {
2997 2976 : int frameOptions = wc->frameOptions;
2998 : double partition_tuples;
2999 : double return_tuples;
3000 : double peer_tuples;
3001 :
3002 : /*
3003 : * First, figure out how many partitions there are likely to be and set
3004 : * partition_tuples according to that estimate.
3005 : */
3006 2976 : if (wc->partitionClause != NIL)
3007 : {
3008 : double num_partitions;
3009 734 : List *partexprs = get_sortgrouplist_exprs(wc->partitionClause,
3010 734 : root->parse->targetList);
3011 :
3012 734 : num_partitions = estimate_num_groups(root, partexprs, input_tuples,
3013 : NULL, NULL);
3014 734 : list_free(partexprs);
3015 :
3016 734 : partition_tuples = input_tuples / num_partitions;
3017 : }
3018 : else
3019 : {
3020 : /* all tuples belong to the same partition */
3021 2242 : partition_tuples = input_tuples;
3022 : }
3023 :
3024 : /* estimate the number of tuples in each peer group */
3025 2976 : if (wc->orderClause != NIL)
3026 : {
3027 : double num_groups;
3028 : List *orderexprs;
3029 :
3030 2370 : orderexprs = get_sortgrouplist_exprs(wc->orderClause,
3031 2370 : root->parse->targetList);
3032 :
3033 : /* estimate out how many peer groups there are in the partition */
3034 2370 : num_groups = estimate_num_groups(root, orderexprs,
3035 : partition_tuples, NULL,
3036 : NULL);
3037 2370 : list_free(orderexprs);
3038 2370 : peer_tuples = partition_tuples / num_groups;
3039 : }
3040 : else
3041 : {
3042 : /* no ORDER BY so only 1 tuple belongs in each peer group */
3043 606 : peer_tuples = 1.0;
3044 : }
3045 :
3046 2976 : if (frameOptions & FRAMEOPTION_END_UNBOUNDED_FOLLOWING)
3047 : {
3048 : /* include all partition rows */
3049 364 : return_tuples = partition_tuples;
3050 : }
3051 2612 : else if (frameOptions & FRAMEOPTION_END_CURRENT_ROW)
3052 : {
3053 1574 : if (frameOptions & FRAMEOPTION_ROWS)
3054 : {
3055 : /* just count the current row */
3056 722 : return_tuples = 1.0;
3057 : }
3058 852 : else if (frameOptions & (FRAMEOPTION_RANGE | FRAMEOPTION_GROUPS))
3059 : {
3060 : /*
3061 : * When in RANGE/GROUPS mode, it's more complex. If there's no
3062 : * ORDER BY, then all rows in the partition are peers, otherwise
3063 : * we'll need to read the first group of peers.
3064 : */
3065 852 : if (wc->orderClause == NIL)
3066 326 : return_tuples = partition_tuples;
3067 : else
3068 526 : return_tuples = peer_tuples;
3069 : }
3070 : else
3071 : {
3072 : /*
3073 : * Something new we don't support yet? This needs attention.
3074 : * We'll just return 1.0 in the meantime.
3075 : */
3076 : Assert(false);
3077 0 : return_tuples = 1.0;
3078 : }
3079 : }
3080 1038 : else if (frameOptions & FRAMEOPTION_END_OFFSET_PRECEDING)
3081 : {
3082 : /*
3083 : * BETWEEN ... AND N PRECEDING will only need to read the WindowAgg's
3084 : * subnode after N ROWS/RANGES/GROUPS. N can be 0, but not negative,
3085 : * so we'll just assume only the current row needs to be read to fetch
3086 : * the first WindowAgg row.
3087 : */
3088 108 : return_tuples = 1.0;
3089 : }
3090 930 : else if (frameOptions & FRAMEOPTION_END_OFFSET_FOLLOWING)
3091 : {
3092 930 : Const *endOffset = (Const *) wc->endOffset;
3093 : double end_offset_value;
3094 :
3095 : /* try and figure out the value specified in the endOffset. */
3096 930 : if (IsA(endOffset, Const))
3097 : {
3098 930 : if (endOffset->constisnull)
3099 : {
3100 : /*
3101 : * NULLs are not allowed, but currently, there's no code to
3102 : * error out if there's a NULL Const. We'll only discover
3103 : * this during execution. For now, just pretend everything is
3104 : * fine and assume that just the first row/range/group will be
3105 : * needed.
3106 : */
3107 0 : end_offset_value = 1.0;
3108 : }
3109 : else
3110 : {
3111 930 : switch (endOffset->consttype)
3112 : {
3113 24 : case INT2OID:
3114 24 : end_offset_value =
3115 24 : (double) DatumGetInt16(endOffset->constvalue);
3116 24 : break;
3117 132 : case INT4OID:
3118 132 : end_offset_value =
3119 132 : (double) DatumGetInt32(endOffset->constvalue);
3120 132 : break;
3121 432 : case INT8OID:
3122 432 : end_offset_value =
3123 432 : (double) DatumGetInt64(endOffset->constvalue);
3124 432 : break;
3125 342 : default:
3126 342 : end_offset_value =
3127 342 : partition_tuples / peer_tuples *
3128 : DEFAULT_INEQ_SEL;
3129 342 : break;
3130 : }
3131 : }
3132 : }
3133 : else
3134 : {
3135 : /*
3136 : * When the end bound is not a Const, we'll just need to guess. We
3137 : * just make use of DEFAULT_INEQ_SEL.
3138 : */
3139 0 : end_offset_value =
3140 0 : partition_tuples / peer_tuples * DEFAULT_INEQ_SEL;
3141 : }
3142 :
3143 930 : if (frameOptions & FRAMEOPTION_ROWS)
3144 : {
3145 : /* include the N FOLLOWING and the current row */
3146 270 : return_tuples = end_offset_value + 1.0;
3147 : }
3148 660 : else if (frameOptions & (FRAMEOPTION_RANGE | FRAMEOPTION_GROUPS))
3149 : {
3150 : /* include N FOLLOWING ranges/group and the initial range/group */
3151 660 : return_tuples = peer_tuples * (end_offset_value + 1.0);
3152 : }
3153 : else
3154 : {
3155 : /*
3156 : * Something new we don't support yet? This needs attention.
3157 : * We'll just return 1.0 in the meantime.
3158 : */
3159 : Assert(false);
3160 0 : return_tuples = 1.0;
3161 : }
3162 : }
3163 : else
3164 : {
3165 : /*
3166 : * Something new we don't support yet? This needs attention. We'll
3167 : * just return 1.0 in the meantime.
3168 : */
3169 : Assert(false);
3170 0 : return_tuples = 1.0;
3171 : }
3172 :
3173 2976 : if (wc->partitionClause != NIL || wc->orderClause != NIL)
3174 : {
3175 : /*
3176 : * Cap the return value to the estimated partition tuples and account
3177 : * for the extra tuple WindowAgg will need to read to confirm the next
3178 : * tuple does not belong to the same partition or peer group.
3179 : */
3180 2582 : return_tuples = Min(return_tuples + 1.0, partition_tuples);
3181 : }
3182 : else
3183 : {
3184 : /*
3185 : * Cap the return value so it's never higher than the expected tuples
3186 : * in the partition.
3187 : */
3188 394 : return_tuples = Min(return_tuples, partition_tuples);
3189 : }
3190 :
3191 : /*
3192 : * We needn't worry about any EXCLUDE options as those only exclude rows
3193 : * from being aggregated, not from being read from the WindowAgg's
3194 : * subnode.
3195 : */
3196 :
3197 2976 : return clamp_row_est(return_tuples);
3198 : }
3199 :
3200 : /*
3201 : * cost_windowagg
3202 : * Determines and returns the cost of performing a WindowAgg plan node,
3203 : * including the cost of its input.
3204 : *
3205 : * Input is assumed already properly sorted.
3206 : */
3207 : void
3208 2976 : cost_windowagg(Path *path, PlannerInfo *root,
3209 : List *windowFuncs, WindowClause *winclause,
3210 : int input_disabled_nodes,
3211 : Cost input_startup_cost, Cost input_total_cost,
3212 : double input_tuples)
3213 : {
3214 : Cost startup_cost;
3215 : Cost total_cost;
3216 : double startup_tuples;
3217 : int numPartCols;
3218 : int numOrderCols;
3219 : ListCell *lc;
3220 :
3221 2976 : numPartCols = list_length(winclause->partitionClause);
3222 2976 : numOrderCols = list_length(winclause->orderClause);
3223 :
3224 2976 : startup_cost = input_startup_cost;
3225 2976 : total_cost = input_total_cost;
3226 :
3227 : /*
3228 : * Window functions are assumed to cost their stated execution cost, plus
3229 : * the cost of evaluating their input expressions, per tuple. Since they
3230 : * may in fact evaluate their inputs at multiple rows during each cycle,
3231 : * this could be a drastic underestimate; but without a way to know how
3232 : * many rows the window function will fetch, it's hard to do better. In
3233 : * any case, it's a good estimate for all the built-in window functions,
3234 : * so we'll just do this for now.
3235 : */
3236 6834 : foreach(lc, windowFuncs)
3237 : {
3238 3858 : WindowFunc *wfunc = lfirst_node(WindowFunc, lc);
3239 : Cost wfunccost;
3240 : QualCost argcosts;
3241 :
3242 3858 : argcosts.startup = argcosts.per_tuple = 0;
3243 3858 : add_function_cost(root, wfunc->winfnoid, (Node *) wfunc,
3244 : &argcosts);
3245 3858 : startup_cost += argcosts.startup;
3246 3858 : wfunccost = argcosts.per_tuple;
3247 :
3248 : /* also add the input expressions' cost to per-input-row costs */
3249 3858 : cost_qual_eval_node(&argcosts, (Node *) wfunc->args, root);
3250 3858 : startup_cost += argcosts.startup;
3251 3858 : wfunccost += argcosts.per_tuple;
3252 :
3253 : /*
3254 : * Add the filter's cost to per-input-row costs. XXX We should reduce
3255 : * input expression costs according to filter selectivity.
3256 : */
3257 3858 : cost_qual_eval_node(&argcosts, (Node *) wfunc->aggfilter, root);
3258 3858 : startup_cost += argcosts.startup;
3259 3858 : wfunccost += argcosts.per_tuple;
3260 :
3261 3858 : total_cost += wfunccost * input_tuples;
3262 : }
3263 :
3264 : /*
3265 : * We also charge cpu_operator_cost per grouping column per tuple for
3266 : * grouping comparisons, plus cpu_tuple_cost per tuple for general
3267 : * overhead.
3268 : *
3269 : * XXX this neglects costs of spooling the data to disk when it overflows
3270 : * work_mem. Sooner or later that should get accounted for.
3271 : */
3272 2976 : total_cost += cpu_operator_cost * (numPartCols + numOrderCols) * input_tuples;
3273 2976 : total_cost += cpu_tuple_cost * input_tuples;
3274 :
3275 2976 : path->rows = input_tuples;
3276 2976 : path->disabled_nodes = input_disabled_nodes;
3277 2976 : path->startup_cost = startup_cost;
3278 2976 : path->total_cost = total_cost;
3279 :
3280 : /*
3281 : * Also, take into account how many tuples we need to read from the
3282 : * subnode in order to produce the first tuple from the WindowAgg. To do
3283 : * this we proportion the run cost (total cost not including startup cost)
3284 : * over the estimated startup tuples. We already included the startup
3285 : * cost of the subnode, so we only need to do this when the estimated
3286 : * startup tuples is above 1.0.
3287 : */
3288 2976 : startup_tuples = get_windowclause_startup_tuples(root, winclause,
3289 : input_tuples);
3290 :
3291 2976 : if (startup_tuples > 1.0)
3292 2568 : path->startup_cost += (total_cost - startup_cost) / input_tuples *
3293 2568 : (startup_tuples - 1.0);
3294 2976 : }
3295 :
3296 : /*
3297 : * cost_group
3298 : * Determines and returns the cost of performing a Group plan node,
3299 : * including the cost of its input.
3300 : *
3301 : * Note: caller must ensure that input costs are for appropriately-sorted
3302 : * input.
3303 : */
3304 : void
3305 1226 : cost_group(Path *path, PlannerInfo *root,
3306 : int numGroupCols, double numGroups,
3307 : List *quals,
3308 : int input_disabled_nodes,
3309 : Cost input_startup_cost, Cost input_total_cost,
3310 : double input_tuples)
3311 : {
3312 : double output_tuples;
3313 : Cost startup_cost;
3314 : Cost total_cost;
3315 :
3316 1226 : output_tuples = numGroups;
3317 1226 : startup_cost = input_startup_cost;
3318 1226 : total_cost = input_total_cost;
3319 :
3320 : /*
3321 : * Charge one cpu_operator_cost per comparison per input tuple. We assume
3322 : * all columns get compared at most of the tuples.
3323 : */
3324 1226 : total_cost += cpu_operator_cost * input_tuples * numGroupCols;
3325 :
3326 : /*
3327 : * If there are quals (HAVING quals), account for their cost and
3328 : * selectivity.
3329 : */
3330 1226 : if (quals)
3331 : {
3332 : QualCost qual_cost;
3333 :
3334 0 : cost_qual_eval(&qual_cost, quals, root);
3335 0 : startup_cost += qual_cost.startup;
3336 0 : total_cost += qual_cost.startup + output_tuples * qual_cost.per_tuple;
3337 :
3338 0 : output_tuples = clamp_row_est(output_tuples *
3339 0 : clauselist_selectivity(root,
3340 : quals,
3341 : 0,
3342 : JOIN_INNER,
3343 : NULL));
3344 : }
3345 :
3346 1226 : path->rows = output_tuples;
3347 1226 : path->disabled_nodes = input_disabled_nodes;
3348 1226 : path->startup_cost = startup_cost;
3349 1226 : path->total_cost = total_cost;
3350 1226 : }
3351 :
3352 : /*
3353 : * initial_cost_nestloop
3354 : * Preliminary estimate of the cost of a nestloop join path.
3355 : *
3356 : * This must quickly produce lower-bound estimates of the path's startup and
3357 : * total costs. If we are unable to eliminate the proposed path from
3358 : * consideration using the lower bounds, final_cost_nestloop will be called
3359 : * to obtain the final estimates.
3360 : *
3361 : * The exact division of labor between this function and final_cost_nestloop
3362 : * is private to them, and represents a tradeoff between speed of the initial
3363 : * estimate and getting a tight lower bound. We choose to not examine the
3364 : * join quals here, since that's by far the most expensive part of the
3365 : * calculations. The end result is that CPU-cost considerations must be
3366 : * left for the second phase; and for SEMI/ANTI joins, we must also postpone
3367 : * incorporation of the inner path's run cost.
3368 : *
3369 : * 'workspace' is to be filled with startup_cost, total_cost, and perhaps
3370 : * other data to be used by final_cost_nestloop
3371 : * 'jointype' is the type of join to be performed
3372 : * 'outer_path' is the outer input to the join
3373 : * 'inner_path' is the inner input to the join
3374 : * 'extra' contains miscellaneous information about the join
3375 : */
3376 : void
3377 3386462 : initial_cost_nestloop(PlannerInfo *root, JoinCostWorkspace *workspace,
3378 : JoinType jointype, uint64 enable_mask,
3379 : Path *outer_path, Path *inner_path,
3380 : JoinPathExtraData *extra)
3381 : {
3382 : int disabled_nodes;
3383 3386462 : Cost startup_cost = 0;
3384 3386462 : Cost run_cost = 0;
3385 3386462 : double outer_path_rows = outer_path->rows;
3386 : Cost inner_rescan_start_cost;
3387 : Cost inner_rescan_total_cost;
3388 : Cost inner_run_cost;
3389 : Cost inner_rescan_run_cost;
3390 :
3391 : /* Count up disabled nodes. */
3392 3386462 : disabled_nodes = (extra->pgs_mask & enable_mask) == enable_mask ? 0 : 1;
3393 3386462 : disabled_nodes += inner_path->disabled_nodes;
3394 3386462 : disabled_nodes += outer_path->disabled_nodes;
3395 :
3396 : /* estimate costs to rescan the inner relation */
3397 3386462 : cost_rescan(root, inner_path,
3398 : &inner_rescan_start_cost,
3399 : &inner_rescan_total_cost);
3400 :
3401 : /* cost of source data */
3402 :
3403 : /*
3404 : * NOTE: clearly, we must pay both outer and inner paths' startup_cost
3405 : * before we can start returning tuples, so the join's startup cost is
3406 : * their sum. We'll also pay the inner path's rescan startup cost
3407 : * multiple times.
3408 : */
3409 3386462 : startup_cost += outer_path->startup_cost + inner_path->startup_cost;
3410 3386462 : run_cost += outer_path->total_cost - outer_path->startup_cost;
3411 3386462 : if (outer_path_rows > 1)
3412 2441920 : run_cost += (outer_path_rows - 1) * inner_rescan_start_cost;
3413 :
3414 3386462 : inner_run_cost = inner_path->total_cost - inner_path->startup_cost;
3415 3386462 : inner_rescan_run_cost = inner_rescan_total_cost - inner_rescan_start_cost;
3416 :
3417 3386462 : if (jointype == JOIN_SEMI || jointype == JOIN_ANTI ||
3418 3324714 : extra->inner_unique)
3419 : {
3420 : /*
3421 : * With a SEMI or ANTI join, or if the innerrel is known unique, the
3422 : * executor will stop after the first match.
3423 : *
3424 : * Getting decent estimates requires inspection of the join quals,
3425 : * which we choose to postpone to final_cost_nestloop.
3426 : */
3427 :
3428 : /* Save private data for final_cost_nestloop */
3429 1385690 : workspace->inner_run_cost = inner_run_cost;
3430 1385690 : workspace->inner_rescan_run_cost = inner_rescan_run_cost;
3431 : }
3432 : else
3433 : {
3434 : /* Normal case; we'll scan whole input rel for each outer row */
3435 2000772 : run_cost += inner_run_cost;
3436 2000772 : if (outer_path_rows > 1)
3437 1534430 : run_cost += (outer_path_rows - 1) * inner_rescan_run_cost;
3438 : }
3439 :
3440 : /* CPU costs left for later */
3441 :
3442 : /* Public result fields */
3443 3386462 : workspace->disabled_nodes = disabled_nodes;
3444 3386462 : workspace->startup_cost = startup_cost;
3445 3386462 : workspace->total_cost = startup_cost + run_cost;
3446 : /* Save private data for final_cost_nestloop */
3447 3386462 : workspace->run_cost = run_cost;
3448 3386462 : }
3449 :
3450 : /*
3451 : * final_cost_nestloop
3452 : * Final estimate of the cost and result size of a nestloop join path.
3453 : *
3454 : * 'path' is already filled in except for the rows and cost fields
3455 : * 'workspace' is the result from initial_cost_nestloop
3456 : * 'extra' contains miscellaneous information about the join
3457 : */
3458 : void
3459 1520154 : final_cost_nestloop(PlannerInfo *root, NestPath *path,
3460 : JoinCostWorkspace *workspace,
3461 : JoinPathExtraData *extra)
3462 : {
3463 1520154 : Path *outer_path = path->jpath.outerjoinpath;
3464 1520154 : Path *inner_path = path->jpath.innerjoinpath;
3465 1520154 : double outer_path_rows = outer_path->rows;
3466 1520154 : double inner_path_rows = inner_path->rows;
3467 1520154 : Cost startup_cost = workspace->startup_cost;
3468 1520154 : Cost run_cost = workspace->run_cost;
3469 : Cost cpu_per_tuple;
3470 : QualCost restrict_qual_cost;
3471 : double ntuples;
3472 :
3473 : /* Set the number of disabled nodes. */
3474 1520154 : path->jpath.path.disabled_nodes = workspace->disabled_nodes;
3475 :
3476 : /* Protect some assumptions below that rowcounts aren't zero */
3477 1520154 : if (outer_path_rows <= 0)
3478 0 : outer_path_rows = 1;
3479 1520154 : if (inner_path_rows <= 0)
3480 726 : inner_path_rows = 1;
3481 : /* Mark the path with the correct row estimate */
3482 1520154 : if (path->jpath.path.param_info)
3483 33238 : path->jpath.path.rows = path->jpath.path.param_info->ppi_rows;
3484 : else
3485 1486916 : path->jpath.path.rows = path->jpath.path.parent->rows;
3486 :
3487 : /* For partial paths, scale row estimate. */
3488 1520154 : if (path->jpath.path.parallel_workers > 0)
3489 : {
3490 44012 : double parallel_divisor = get_parallel_divisor(&path->jpath.path);
3491 :
3492 44012 : path->jpath.path.rows =
3493 44012 : clamp_row_est(path->jpath.path.rows / parallel_divisor);
3494 : }
3495 :
3496 : /* cost of inner-relation source data (we already dealt with outer rel) */
3497 :
3498 1520154 : if (path->jpath.jointype == JOIN_SEMI || path->jpath.jointype == JOIN_ANTI ||
3499 1476706 : extra->inner_unique)
3500 955428 : {
3501 : /*
3502 : * With a SEMI or ANTI join, or if the innerrel is known unique, the
3503 : * executor will stop after the first match.
3504 : */
3505 955428 : Cost inner_run_cost = workspace->inner_run_cost;
3506 955428 : Cost inner_rescan_run_cost = workspace->inner_rescan_run_cost;
3507 : double outer_matched_rows;
3508 : double outer_unmatched_rows;
3509 : Selectivity inner_scan_frac;
3510 :
3511 : /*
3512 : * For an outer-rel row that has at least one match, we can expect the
3513 : * inner scan to stop after a fraction 1/(match_count+1) of the inner
3514 : * rows, if the matches are evenly distributed. Since they probably
3515 : * aren't quite evenly distributed, we apply a fuzz factor of 2.0 to
3516 : * that fraction. (If we used a larger fuzz factor, we'd have to
3517 : * clamp inner_scan_frac to at most 1.0; but since match_count is at
3518 : * least 1, no such clamp is needed now.)
3519 : */
3520 955428 : outer_matched_rows = rint(outer_path_rows * extra->semifactors.outer_match_frac);
3521 955428 : outer_unmatched_rows = outer_path_rows - outer_matched_rows;
3522 955428 : inner_scan_frac = 2.0 / (extra->semifactors.match_count + 1.0);
3523 :
3524 : /*
3525 : * Compute number of tuples processed (not number emitted!). First,
3526 : * account for successfully-matched outer rows.
3527 : */
3528 955428 : ntuples = outer_matched_rows * inner_path_rows * inner_scan_frac;
3529 :
3530 : /*
3531 : * Now we need to estimate the actual costs of scanning the inner
3532 : * relation, which may be quite a bit less than N times inner_run_cost
3533 : * due to early scan stops. We consider two cases. If the inner path
3534 : * is an indexscan using all the joinquals as indexquals, then an
3535 : * unmatched outer row results in an indexscan returning no rows,
3536 : * which is probably quite cheap. Otherwise, the executor will have
3537 : * to scan the whole inner rel for an unmatched row; not so cheap.
3538 : */
3539 955428 : if (has_indexed_join_quals(path))
3540 : {
3541 : /*
3542 : * Successfully-matched outer rows will only require scanning
3543 : * inner_scan_frac of the inner relation. In this case, we don't
3544 : * need to charge the full inner_run_cost even when that's more
3545 : * than inner_rescan_run_cost, because we can assume that none of
3546 : * the inner scans ever scan the whole inner relation. So it's
3547 : * okay to assume that all the inner scan executions can be
3548 : * fractions of the full cost, even if materialization is reducing
3549 : * the rescan cost. At this writing, it's impossible to get here
3550 : * for a materialized inner scan, so inner_run_cost and
3551 : * inner_rescan_run_cost will be the same anyway; but just in
3552 : * case, use inner_run_cost for the first matched tuple and
3553 : * inner_rescan_run_cost for additional ones.
3554 : */
3555 157300 : run_cost += inner_run_cost * inner_scan_frac;
3556 157300 : if (outer_matched_rows > 1)
3557 21210 : run_cost += (outer_matched_rows - 1) * inner_rescan_run_cost * inner_scan_frac;
3558 :
3559 : /*
3560 : * Add the cost of inner-scan executions for unmatched outer rows.
3561 : * We estimate this as the same cost as returning the first tuple
3562 : * of a nonempty scan. We consider that these are all rescans,
3563 : * since we used inner_run_cost once already.
3564 : */
3565 157300 : run_cost += outer_unmatched_rows *
3566 157300 : inner_rescan_run_cost / inner_path_rows;
3567 :
3568 : /*
3569 : * We won't be evaluating any quals at all for unmatched rows, so
3570 : * don't add them to ntuples.
3571 : */
3572 : }
3573 : else
3574 : {
3575 : /*
3576 : * Here, a complicating factor is that rescans may be cheaper than
3577 : * first scans. If we never scan all the way to the end of the
3578 : * inner rel, it might be (depending on the plan type) that we'd
3579 : * never pay the whole inner first-scan run cost. However it is
3580 : * difficult to estimate whether that will happen (and it could
3581 : * not happen if there are any unmatched outer rows!), so be
3582 : * conservative and always charge the whole first-scan cost once.
3583 : * We consider this charge to correspond to the first unmatched
3584 : * outer row, unless there isn't one in our estimate, in which
3585 : * case blame it on the first matched row.
3586 : */
3587 :
3588 : /* First, count all unmatched join tuples as being processed */
3589 798128 : ntuples += outer_unmatched_rows * inner_path_rows;
3590 :
3591 : /* Now add the forced full scan, and decrement appropriate count */
3592 798128 : run_cost += inner_run_cost;
3593 798128 : if (outer_unmatched_rows >= 1)
3594 758170 : outer_unmatched_rows -= 1;
3595 : else
3596 39958 : outer_matched_rows -= 1;
3597 :
3598 : /* Add inner run cost for additional outer tuples having matches */
3599 798128 : if (outer_matched_rows > 0)
3600 282160 : run_cost += outer_matched_rows * inner_rescan_run_cost * inner_scan_frac;
3601 :
3602 : /* Add inner run cost for additional unmatched outer tuples */
3603 798128 : if (outer_unmatched_rows > 0)
3604 496018 : run_cost += outer_unmatched_rows * inner_rescan_run_cost;
3605 : }
3606 : }
3607 : else
3608 : {
3609 : /* Normal-case source costs were included in preliminary estimate */
3610 :
3611 : /* Compute number of tuples processed (not number emitted!) */
3612 564726 : ntuples = outer_path_rows * inner_path_rows;
3613 : }
3614 :
3615 : /* CPU costs */
3616 1520154 : cost_qual_eval(&restrict_qual_cost, path->jpath.joinrestrictinfo, root);
3617 1520154 : startup_cost += restrict_qual_cost.startup;
3618 1520154 : cpu_per_tuple = cpu_tuple_cost + restrict_qual_cost.per_tuple;
3619 1520154 : run_cost += cpu_per_tuple * ntuples;
3620 :
3621 : /* tlist eval costs are paid per output row, not per tuple scanned */
3622 1520154 : startup_cost += path->jpath.path.pathtarget->cost.startup;
3623 1520154 : run_cost += path->jpath.path.pathtarget->cost.per_tuple * path->jpath.path.rows;
3624 :
3625 1520154 : path->jpath.path.startup_cost = startup_cost;
3626 1520154 : path->jpath.path.total_cost = startup_cost + run_cost;
3627 1520154 : }
3628 :
3629 : /*
3630 : * initial_cost_mergejoin
3631 : * Preliminary estimate of the cost of a mergejoin path.
3632 : *
3633 : * This must quickly produce lower-bound estimates of the path's startup and
3634 : * total costs. If we are unable to eliminate the proposed path from
3635 : * consideration using the lower bounds, final_cost_mergejoin will be called
3636 : * to obtain the final estimates.
3637 : *
3638 : * The exact division of labor between this function and final_cost_mergejoin
3639 : * is private to them, and represents a tradeoff between speed of the initial
3640 : * estimate and getting a tight lower bound. We choose to not examine the
3641 : * join quals here, except for obtaining the scan selectivity estimate which
3642 : * is really essential (but fortunately, use of caching keeps the cost of
3643 : * getting that down to something reasonable).
3644 : * We also assume that cost_sort/cost_incremental_sort is cheap enough to use
3645 : * here.
3646 : *
3647 : * 'workspace' is to be filled with startup_cost, total_cost, and perhaps
3648 : * other data to be used by final_cost_mergejoin
3649 : * 'jointype' is the type of join to be performed
3650 : * 'mergeclauses' is the list of joinclauses to be used as merge clauses
3651 : * 'outer_path' is the outer input to the join
3652 : * 'inner_path' is the inner input to the join
3653 : * 'outersortkeys' is the list of sort keys for the outer path
3654 : * 'innersortkeys' is the list of sort keys for the inner path
3655 : * 'outer_presorted_keys' is the number of presorted keys of the outer path
3656 : * 'extra' contains miscellaneous information about the join
3657 : *
3658 : * Note: outersortkeys and innersortkeys should be NIL if no explicit
3659 : * sort is needed because the respective source path is already ordered.
3660 : */
3661 : void
3662 1536716 : initial_cost_mergejoin(PlannerInfo *root, JoinCostWorkspace *workspace,
3663 : JoinType jointype,
3664 : List *mergeclauses,
3665 : Path *outer_path, Path *inner_path,
3666 : List *outersortkeys, List *innersortkeys,
3667 : int outer_presorted_keys,
3668 : JoinPathExtraData *extra)
3669 : {
3670 : int disabled_nodes;
3671 1536716 : Cost startup_cost = 0;
3672 1536716 : Cost run_cost = 0;
3673 1536716 : double outer_path_rows = outer_path->rows;
3674 1536716 : double inner_path_rows = inner_path->rows;
3675 : Cost inner_run_cost;
3676 : double outer_rows,
3677 : inner_rows,
3678 : outer_skip_rows,
3679 : inner_skip_rows;
3680 : Selectivity outerstartsel,
3681 : outerendsel,
3682 : innerstartsel,
3683 : innerendsel;
3684 : Path sort_path; /* dummy for result of
3685 : * cost_sort/cost_incremental_sort */
3686 :
3687 : /* Protect some assumptions below that rowcounts aren't zero */
3688 1536716 : if (outer_path_rows <= 0)
3689 96 : outer_path_rows = 1;
3690 1536716 : if (inner_path_rows <= 0)
3691 126 : inner_path_rows = 1;
3692 :
3693 : /*
3694 : * A merge join will stop as soon as it exhausts either input stream
3695 : * (unless it's an outer join, in which case the outer side has to be
3696 : * scanned all the way anyway). Estimate fraction of the left and right
3697 : * inputs that will actually need to be scanned. Likewise, we can
3698 : * estimate the number of rows that will be skipped before the first join
3699 : * pair is found, which should be factored into startup cost. We use only
3700 : * the first (most significant) merge clause for this purpose. Since
3701 : * mergejoinscansel() is a fairly expensive computation, we cache the
3702 : * results in the merge clause RestrictInfo.
3703 : */
3704 1536716 : if (mergeclauses && jointype != JOIN_FULL)
3705 1530452 : {
3706 1530452 : RestrictInfo *firstclause = (RestrictInfo *) linitial(mergeclauses);
3707 : List *opathkeys;
3708 : List *ipathkeys;
3709 : PathKey *opathkey;
3710 : PathKey *ipathkey;
3711 : MergeScanSelCache *cache;
3712 :
3713 : /* Get the input pathkeys to determine the sort-order details */
3714 1530452 : opathkeys = outersortkeys ? outersortkeys : outer_path->pathkeys;
3715 1530452 : ipathkeys = innersortkeys ? innersortkeys : inner_path->pathkeys;
3716 : Assert(opathkeys);
3717 : Assert(ipathkeys);
3718 1530452 : opathkey = (PathKey *) linitial(opathkeys);
3719 1530452 : ipathkey = (PathKey *) linitial(ipathkeys);
3720 : /* debugging check */
3721 1530452 : if (opathkey->pk_opfamily != ipathkey->pk_opfamily ||
3722 1530452 : opathkey->pk_eclass->ec_collation != ipathkey->pk_eclass->ec_collation ||
3723 1530452 : opathkey->pk_cmptype != ipathkey->pk_cmptype ||
3724 1530452 : opathkey->pk_nulls_first != ipathkey->pk_nulls_first)
3725 0 : elog(ERROR, "left and right pathkeys do not match in mergejoin");
3726 :
3727 : /* Get the selectivity with caching */
3728 1530452 : cache = cached_scansel(root, firstclause, opathkey);
3729 :
3730 1530452 : if (bms_is_subset(firstclause->left_relids,
3731 1530452 : outer_path->parent->relids))
3732 : {
3733 : /* left side of clause is outer */
3734 798028 : outerstartsel = cache->leftstartsel;
3735 798028 : outerendsel = cache->leftendsel;
3736 798028 : innerstartsel = cache->rightstartsel;
3737 798028 : innerendsel = cache->rightendsel;
3738 : }
3739 : else
3740 : {
3741 : /* left side of clause is inner */
3742 732424 : outerstartsel = cache->rightstartsel;
3743 732424 : outerendsel = cache->rightendsel;
3744 732424 : innerstartsel = cache->leftstartsel;
3745 732424 : innerendsel = cache->leftendsel;
3746 : }
3747 1530452 : if (jointype == JOIN_LEFT ||
3748 : jointype == JOIN_ANTI)
3749 : {
3750 200638 : outerstartsel = 0.0;
3751 200638 : outerendsel = 1.0;
3752 : }
3753 1329814 : else if (jointype == JOIN_RIGHT ||
3754 : jointype == JOIN_RIGHT_ANTI)
3755 : {
3756 200294 : innerstartsel = 0.0;
3757 200294 : innerendsel = 1.0;
3758 : }
3759 : }
3760 : else
3761 : {
3762 : /* cope with clauseless or full mergejoin */
3763 6264 : outerstartsel = innerstartsel = 0.0;
3764 6264 : outerendsel = innerendsel = 1.0;
3765 : }
3766 :
3767 : /*
3768 : * Convert selectivities to row counts. We force outer_rows and
3769 : * inner_rows to be at least 1, but the skip_rows estimates can be zero.
3770 : */
3771 1536716 : outer_skip_rows = rint(outer_path_rows * outerstartsel);
3772 1536716 : inner_skip_rows = rint(inner_path_rows * innerstartsel);
3773 1536716 : outer_rows = clamp_row_est(outer_path_rows * outerendsel);
3774 1536716 : inner_rows = clamp_row_est(inner_path_rows * innerendsel);
3775 :
3776 : Assert(outer_skip_rows <= outer_rows);
3777 : Assert(inner_skip_rows <= inner_rows);
3778 :
3779 : /*
3780 : * Readjust scan selectivities to account for above rounding. This is
3781 : * normally an insignificant effect, but when there are only a few rows in
3782 : * the inputs, failing to do this makes for a large percentage error.
3783 : */
3784 1536716 : outerstartsel = outer_skip_rows / outer_path_rows;
3785 1536716 : innerstartsel = inner_skip_rows / inner_path_rows;
3786 1536716 : outerendsel = outer_rows / outer_path_rows;
3787 1536716 : innerendsel = inner_rows / inner_path_rows;
3788 :
3789 : Assert(outerstartsel <= outerendsel);
3790 : Assert(innerstartsel <= innerendsel);
3791 :
3792 : /*
3793 : * We don't decide whether to materialize the inner path until we get to
3794 : * final_cost_mergejoin(), so we don't know whether to check the pgs_mask
3795 : * against PGS_MERGEJOIN_PLAIN or PGS_MERGEJOIN_MATERIALIZE. Instead, we
3796 : * just account for any child nodes here and assume that this node is not
3797 : * itself disabled; we can sort out the details in final_cost_mergejoin().
3798 : *
3799 : * (We could be more precise here by setting disabled_nodes to 1 at this
3800 : * stage if both PGS_MERGEJOIN_PLAIN and PGS_MERGEJOIN_MATERIALIZE are
3801 : * disabled, but that seems to against the idea of making this function
3802 : * produce a quick, optimistic approximation of the final cost.)
3803 : */
3804 1536716 : disabled_nodes = 0;
3805 :
3806 : /* cost of source data */
3807 :
3808 1536716 : if (outersortkeys) /* do we need to sort outer? */
3809 : {
3810 : /*
3811 : * We can assert that the outer path is not already ordered
3812 : * appropriately for the mergejoin; otherwise, outersortkeys would
3813 : * have been set to NIL.
3814 : */
3815 : Assert(!pathkeys_contained_in(outersortkeys, outer_path->pathkeys));
3816 :
3817 : /*
3818 : * We choose to use incremental sort if it is enabled and there are
3819 : * presorted keys; otherwise we use full sort.
3820 : */
3821 776994 : if (enable_incremental_sort && outer_presorted_keys > 0)
3822 : {
3823 2036 : cost_incremental_sort(&sort_path,
3824 : root,
3825 : outersortkeys,
3826 : outer_presorted_keys,
3827 : outer_path->disabled_nodes,
3828 : outer_path->startup_cost,
3829 : outer_path->total_cost,
3830 : outer_path_rows,
3831 2036 : outer_path->pathtarget->width,
3832 : 0.0,
3833 : work_mem,
3834 : -1.0);
3835 : }
3836 : else
3837 : {
3838 774958 : cost_sort(&sort_path,
3839 : root,
3840 : outersortkeys,
3841 : outer_path->disabled_nodes,
3842 : outer_path->total_cost,
3843 : outer_path_rows,
3844 774958 : outer_path->pathtarget->width,
3845 : 0.0,
3846 : work_mem,
3847 : -1.0);
3848 : }
3849 :
3850 776994 : disabled_nodes += sort_path.disabled_nodes;
3851 776994 : startup_cost += sort_path.startup_cost;
3852 776994 : startup_cost += (sort_path.total_cost - sort_path.startup_cost)
3853 776994 : * outerstartsel;
3854 776994 : run_cost += (sort_path.total_cost - sort_path.startup_cost)
3855 776994 : * (outerendsel - outerstartsel);
3856 : }
3857 : else
3858 : {
3859 759722 : disabled_nodes += outer_path->disabled_nodes;
3860 759722 : startup_cost += outer_path->startup_cost;
3861 759722 : startup_cost += (outer_path->total_cost - outer_path->startup_cost)
3862 759722 : * outerstartsel;
3863 759722 : run_cost += (outer_path->total_cost - outer_path->startup_cost)
3864 759722 : * (outerendsel - outerstartsel);
3865 : }
3866 :
3867 1536716 : if (innersortkeys) /* do we need to sort inner? */
3868 : {
3869 : /*
3870 : * We can assert that the inner path is not already ordered
3871 : * appropriately for the mergejoin; otherwise, innersortkeys would
3872 : * have been set to NIL.
3873 : */
3874 : Assert(!pathkeys_contained_in(innersortkeys, inner_path->pathkeys));
3875 :
3876 : /*
3877 : * We do not consider incremental sort for inner path, because
3878 : * incremental sort does not support mark/restore.
3879 : */
3880 :
3881 1229826 : cost_sort(&sort_path,
3882 : root,
3883 : innersortkeys,
3884 : inner_path->disabled_nodes,
3885 : inner_path->total_cost,
3886 : inner_path_rows,
3887 1229826 : inner_path->pathtarget->width,
3888 : 0.0,
3889 : work_mem,
3890 : -1.0);
3891 1229826 : disabled_nodes += sort_path.disabled_nodes;
3892 1229826 : startup_cost += sort_path.startup_cost;
3893 1229826 : startup_cost += (sort_path.total_cost - sort_path.startup_cost)
3894 1229826 : * innerstartsel;
3895 1229826 : inner_run_cost = (sort_path.total_cost - sort_path.startup_cost)
3896 1229826 : * (innerendsel - innerstartsel);
3897 : }
3898 : else
3899 : {
3900 306890 : disabled_nodes += inner_path->disabled_nodes;
3901 306890 : startup_cost += inner_path->startup_cost;
3902 306890 : startup_cost += (inner_path->total_cost - inner_path->startup_cost)
3903 306890 : * innerstartsel;
3904 306890 : inner_run_cost = (inner_path->total_cost - inner_path->startup_cost)
3905 306890 : * (innerendsel - innerstartsel);
3906 : }
3907 :
3908 : /*
3909 : * We can't yet determine whether rescanning occurs, or whether
3910 : * materialization of the inner input should be done. The minimum
3911 : * possible inner input cost, regardless of rescan and materialization
3912 : * considerations, is inner_run_cost. We include that in
3913 : * workspace->total_cost, but not yet in run_cost.
3914 : */
3915 :
3916 : /* CPU costs left for later */
3917 :
3918 : /* Public result fields */
3919 1536716 : workspace->disabled_nodes = disabled_nodes;
3920 1536716 : workspace->startup_cost = startup_cost;
3921 1536716 : workspace->total_cost = startup_cost + run_cost + inner_run_cost;
3922 : /* Save private data for final_cost_mergejoin */
3923 1536716 : workspace->run_cost = run_cost;
3924 1536716 : workspace->inner_run_cost = inner_run_cost;
3925 1536716 : workspace->outer_rows = outer_rows;
3926 1536716 : workspace->inner_rows = inner_rows;
3927 1536716 : workspace->outer_skip_rows = outer_skip_rows;
3928 1536716 : workspace->inner_skip_rows = inner_skip_rows;
3929 1536716 : }
3930 :
3931 : /*
3932 : * final_cost_mergejoin
3933 : * Final estimate of the cost and result size of a mergejoin path.
3934 : *
3935 : * Unlike other costsize functions, this routine makes two actual decisions:
3936 : * whether the executor will need to do mark/restore, and whether we should
3937 : * materialize the inner path. It would be logically cleaner to build
3938 : * separate paths testing these alternatives, but that would require repeating
3939 : * most of the cost calculations, which are not all that cheap. Since the
3940 : * choice will not affect output pathkeys or startup cost, only total cost,
3941 : * there is no possibility of wanting to keep more than one path. So it seems
3942 : * best to make the decisions here and record them in the path's
3943 : * skip_mark_restore and materialize_inner fields.
3944 : *
3945 : * Mark/restore overhead is usually required, but can be skipped if we know
3946 : * that the executor need find only one match per outer tuple, and that the
3947 : * mergeclauses are sufficient to identify a match.
3948 : *
3949 : * We materialize the inner path if we need mark/restore and either the inner
3950 : * path can't support mark/restore, or it's cheaper to use an interposed
3951 : * Material node to handle mark/restore.
3952 : *
3953 : * 'path' is already filled in except for the rows and cost fields and
3954 : * skip_mark_restore and materialize_inner
3955 : * 'workspace' is the result from initial_cost_mergejoin
3956 : * 'extra' contains miscellaneous information about the join
3957 : */
3958 : void
3959 473372 : final_cost_mergejoin(PlannerInfo *root, MergePath *path,
3960 : JoinCostWorkspace *workspace,
3961 : JoinPathExtraData *extra)
3962 : {
3963 473372 : Path *outer_path = path->jpath.outerjoinpath;
3964 473372 : Path *inner_path = path->jpath.innerjoinpath;
3965 473372 : double inner_path_rows = inner_path->rows;
3966 473372 : List *mergeclauses = path->path_mergeclauses;
3967 473372 : List *innersortkeys = path->innersortkeys;
3968 473372 : Cost startup_cost = workspace->startup_cost;
3969 473372 : Cost run_cost = workspace->run_cost;
3970 473372 : Cost inner_run_cost = workspace->inner_run_cost;
3971 473372 : double outer_rows = workspace->outer_rows;
3972 473372 : double inner_rows = workspace->inner_rows;
3973 473372 : double outer_skip_rows = workspace->outer_skip_rows;
3974 473372 : double inner_skip_rows = workspace->inner_skip_rows;
3975 : Cost cpu_per_tuple,
3976 : bare_inner_cost,
3977 : mat_inner_cost;
3978 : QualCost merge_qual_cost;
3979 : QualCost qp_qual_cost;
3980 : double mergejointuples,
3981 : rescannedtuples;
3982 : double rescanratio;
3983 473372 : uint64 enable_mask = 0;
3984 :
3985 : /* Protect some assumptions below that rowcounts aren't zero */
3986 473372 : if (inner_path_rows <= 0)
3987 90 : inner_path_rows = 1;
3988 :
3989 : /* Mark the path with the correct row estimate */
3990 473372 : if (path->jpath.path.param_info)
3991 1624 : path->jpath.path.rows = path->jpath.path.param_info->ppi_rows;
3992 : else
3993 471748 : path->jpath.path.rows = path->jpath.path.parent->rows;
3994 :
3995 : /* For partial paths, scale row estimate. */
3996 473372 : if (path->jpath.path.parallel_workers > 0)
3997 : {
3998 65872 : double parallel_divisor = get_parallel_divisor(&path->jpath.path);
3999 :
4000 65872 : path->jpath.path.rows =
4001 65872 : clamp_row_est(path->jpath.path.rows / parallel_divisor);
4002 : }
4003 :
4004 : /*
4005 : * Compute cost of the mergequals and qpquals (other restriction clauses)
4006 : * separately.
4007 : */
4008 473372 : cost_qual_eval(&merge_qual_cost, mergeclauses, root);
4009 473372 : cost_qual_eval(&qp_qual_cost, path->jpath.joinrestrictinfo, root);
4010 473372 : qp_qual_cost.startup -= merge_qual_cost.startup;
4011 473372 : qp_qual_cost.per_tuple -= merge_qual_cost.per_tuple;
4012 :
4013 : /*
4014 : * With a SEMI or ANTI join, or if the innerrel is known unique, the
4015 : * executor will stop scanning for matches after the first match. When
4016 : * all the joinclauses are merge clauses, this means we don't ever need to
4017 : * back up the merge, and so we can skip mark/restore overhead.
4018 : */
4019 473372 : if ((path->jpath.jointype == JOIN_SEMI ||
4020 466126 : path->jpath.jointype == JOIN_ANTI ||
4021 624292 : extra->inner_unique) &&
4022 165260 : (list_length(path->jpath.joinrestrictinfo) ==
4023 165260 : list_length(path->path_mergeclauses)))
4024 142906 : path->skip_mark_restore = true;
4025 : else
4026 330466 : path->skip_mark_restore = false;
4027 :
4028 : /*
4029 : * Get approx # tuples passing the mergequals. We use approx_tuple_count
4030 : * here because we need an estimate done with JOIN_INNER semantics.
4031 : */
4032 473372 : mergejointuples = approx_tuple_count(root, &path->jpath, mergeclauses);
4033 :
4034 : /*
4035 : * When there are equal merge keys in the outer relation, the mergejoin
4036 : * must rescan any matching tuples in the inner relation. This means
4037 : * re-fetching inner tuples; we have to estimate how often that happens.
4038 : *
4039 : * For regular inner and outer joins, the number of re-fetches can be
4040 : * estimated approximately as size of merge join output minus size of
4041 : * inner relation. Assume that the distinct key values are 1, 2, ..., and
4042 : * denote the number of values of each key in the outer relation as m1,
4043 : * m2, ...; in the inner relation, n1, n2, ... Then we have
4044 : *
4045 : * size of join = m1 * n1 + m2 * n2 + ...
4046 : *
4047 : * number of rescanned tuples = (m1 - 1) * n1 + (m2 - 1) * n2 + ... = m1 *
4048 : * n1 + m2 * n2 + ... - (n1 + n2 + ...) = size of join - size of inner
4049 : * relation
4050 : *
4051 : * This equation works correctly for outer tuples having no inner match
4052 : * (nk = 0), but not for inner tuples having no outer match (mk = 0); we
4053 : * are effectively subtracting those from the number of rescanned tuples,
4054 : * when we should not. Can we do better without expensive selectivity
4055 : * computations?
4056 : *
4057 : * The whole issue is moot if we know we don't need to mark/restore at
4058 : * all, or if we are working from a unique-ified outer input.
4059 : */
4060 473372 : if (path->skip_mark_restore ||
4061 330466 : RELATION_WAS_MADE_UNIQUE(outer_path->parent, extra->sjinfo,
4062 : path->jpath.jointype))
4063 147616 : rescannedtuples = 0;
4064 : else
4065 : {
4066 325756 : rescannedtuples = mergejointuples - inner_path_rows;
4067 : /* Must clamp because of possible underestimate */
4068 325756 : if (rescannedtuples < 0)
4069 80234 : rescannedtuples = 0;
4070 : }
4071 :
4072 : /*
4073 : * We'll inflate various costs this much to account for rescanning. Note
4074 : * that this is to be multiplied by something involving inner_rows, or
4075 : * another number related to the portion of the inner rel we'll scan.
4076 : */
4077 473372 : rescanratio = 1.0 + (rescannedtuples / inner_rows);
4078 :
4079 : /*
4080 : * Decide whether we want to materialize the inner input to shield it from
4081 : * mark/restore and performing re-fetches. Our cost model for regular
4082 : * re-fetches is that a re-fetch costs the same as an original fetch,
4083 : * which is probably an overestimate; but on the other hand we ignore the
4084 : * bookkeeping costs of mark/restore. Not clear if it's worth developing
4085 : * a more refined model. So we just need to inflate the inner run cost by
4086 : * rescanratio.
4087 : */
4088 473372 : bare_inner_cost = inner_run_cost * rescanratio;
4089 :
4090 : /*
4091 : * When we interpose a Material node the re-fetch cost is assumed to be
4092 : * just cpu_operator_cost per tuple, independently of the underlying
4093 : * plan's cost; and we charge an extra cpu_operator_cost per original
4094 : * fetch as well. Note that we're assuming the materialize node will
4095 : * never spill to disk, since it only has to remember tuples back to the
4096 : * last mark. (If there are a huge number of duplicates, our other cost
4097 : * factors will make the path so expensive that it probably won't get
4098 : * chosen anyway.) So we don't use cost_rescan here.
4099 : *
4100 : * Note: keep this estimate in sync with create_mergejoin_plan's labeling
4101 : * of the generated Material node.
4102 : */
4103 473372 : mat_inner_cost = inner_run_cost +
4104 473372 : cpu_operator_cost * inner_rows * rescanratio;
4105 :
4106 : /*
4107 : * If we don't need mark/restore at all, we don't need materialization.
4108 : */
4109 473372 : if (path->skip_mark_restore)
4110 142906 : path->materialize_inner = false;
4111 :
4112 : /*
4113 : * If merge joins with materialization are enabled, then choose
4114 : * materialization if either (a) it looks cheaper or (b) merge joins
4115 : * without materialization are disabled.
4116 : */
4117 330466 : else if ((extra->pgs_mask & PGS_MERGEJOIN_MATERIALIZE) != 0 &&
4118 327064 : (mat_inner_cost < bare_inner_cost ||
4119 327064 : (extra->pgs_mask & PGS_MERGEJOIN_PLAIN) == 0))
4120 3340 : path->materialize_inner = true;
4121 :
4122 : /*
4123 : * Regardless of what plan shapes are enabled and what the costs seem to
4124 : * be, we *must* materialize it if the inner path is to be used directly
4125 : * (without sorting) and it doesn't support mark/restore. Planner failure
4126 : * is not an option!
4127 : *
4128 : * Since the inner side must be ordered, and only Sorts and IndexScans can
4129 : * create order to begin with, and they both support mark/restore, you
4130 : * might think there's no problem --- but you'd be wrong. Nestloop and
4131 : * merge joins can *preserve* the order of their inputs, so they can be
4132 : * selected as the input of a mergejoin, and they don't support
4133 : * mark/restore at present.
4134 : */
4135 327126 : else if (innersortkeys == NIL &&
4136 9084 : !ExecSupportsMarkRestore(inner_path))
4137 1918 : path->materialize_inner = true;
4138 :
4139 : /*
4140 : * Also, force materializing if the inner path is to be sorted and the
4141 : * sort is expected to spill to disk. This is because the final merge
4142 : * pass can be done on-the-fly if it doesn't have to support mark/restore.
4143 : * We don't try to adjust the cost estimates for this consideration,
4144 : * though.
4145 : *
4146 : * Since materialization is a performance optimization in this case,
4147 : * rather than necessary for correctness, we skip it if materialization is
4148 : * switched off.
4149 : */
4150 325208 : else if ((extra->pgs_mask & PGS_MERGEJOIN_MATERIALIZE) != 0 &&
4151 317988 : innersortkeys != NIL &&
4152 317988 : relation_byte_size(inner_path_rows,
4153 317988 : inner_path->pathtarget->width) >
4154 317988 : work_mem * (Size) 1024)
4155 284 : path->materialize_inner = true;
4156 : else
4157 324924 : path->materialize_inner = false;
4158 :
4159 : /* Get the number of disabled nodes, not yet including this one. */
4160 473372 : path->jpath.path.disabled_nodes = workspace->disabled_nodes;
4161 :
4162 : /*
4163 : * Charge the right incremental cost for the chosen case, and update
4164 : * enable_mask as appropriate.
4165 : */
4166 473372 : if (path->materialize_inner)
4167 : {
4168 5542 : run_cost += mat_inner_cost;
4169 5542 : enable_mask |= PGS_MERGEJOIN_MATERIALIZE;
4170 : }
4171 : else
4172 : {
4173 467830 : run_cost += bare_inner_cost;
4174 467830 : enable_mask |= PGS_MERGEJOIN_PLAIN;
4175 : }
4176 :
4177 : /* Incremental count of disabled nodes if this node is disabled. */
4178 473372 : if (path->jpath.path.parallel_workers == 0)
4179 407500 : enable_mask |= PGS_CONSIDER_NONPARTIAL;
4180 473372 : if ((extra->pgs_mask & enable_mask) != enable_mask)
4181 14 : ++path->jpath.path.disabled_nodes;
4182 :
4183 : /* CPU costs */
4184 :
4185 : /*
4186 : * The number of tuple comparisons needed is approximately number of outer
4187 : * rows plus number of inner rows plus number of rescanned tuples (can we
4188 : * refine this?). At each one, we need to evaluate the mergejoin quals.
4189 : */
4190 473372 : startup_cost += merge_qual_cost.startup;
4191 473372 : startup_cost += merge_qual_cost.per_tuple *
4192 473372 : (outer_skip_rows + inner_skip_rows * rescanratio);
4193 473372 : run_cost += merge_qual_cost.per_tuple *
4194 473372 : ((outer_rows - outer_skip_rows) +
4195 473372 : (inner_rows - inner_skip_rows) * rescanratio);
4196 :
4197 : /*
4198 : * For each tuple that gets through the mergejoin proper, we charge
4199 : * cpu_tuple_cost plus the cost of evaluating additional restriction
4200 : * clauses that are to be applied at the join. (This is pessimistic since
4201 : * not all of the quals may get evaluated at each tuple.)
4202 : *
4203 : * Note: we could adjust for SEMI/ANTI joins skipping some qual
4204 : * evaluations here, but it's probably not worth the trouble.
4205 : */
4206 473372 : startup_cost += qp_qual_cost.startup;
4207 473372 : cpu_per_tuple = cpu_tuple_cost + qp_qual_cost.per_tuple;
4208 473372 : run_cost += cpu_per_tuple * mergejointuples;
4209 :
4210 : /* tlist eval costs are paid per output row, not per tuple scanned */
4211 473372 : startup_cost += path->jpath.path.pathtarget->cost.startup;
4212 473372 : run_cost += path->jpath.path.pathtarget->cost.per_tuple * path->jpath.path.rows;
4213 :
4214 473372 : path->jpath.path.startup_cost = startup_cost;
4215 473372 : path->jpath.path.total_cost = startup_cost + run_cost;
4216 473372 : }
4217 :
4218 : /*
4219 : * run mergejoinscansel() with caching
4220 : */
4221 : static MergeScanSelCache *
4222 1530452 : cached_scansel(PlannerInfo *root, RestrictInfo *rinfo, PathKey *pathkey)
4223 : {
4224 : MergeScanSelCache *cache;
4225 : ListCell *lc;
4226 : Selectivity leftstartsel,
4227 : leftendsel,
4228 : rightstartsel,
4229 : rightendsel;
4230 : MemoryContext oldcontext;
4231 :
4232 : /* Do we have this result already? */
4233 1530458 : foreach(lc, rinfo->scansel_cache)
4234 : {
4235 1385704 : cache = (MergeScanSelCache *) lfirst(lc);
4236 1385704 : if (cache->opfamily == pathkey->pk_opfamily &&
4237 1385704 : cache->collation == pathkey->pk_eclass->ec_collation &&
4238 1385704 : cache->cmptype == pathkey->pk_cmptype &&
4239 1385698 : cache->nulls_first == pathkey->pk_nulls_first)
4240 1385698 : return cache;
4241 : }
4242 :
4243 : /* Nope, do the computation */
4244 144754 : mergejoinscansel(root,
4245 144754 : (Node *) rinfo->clause,
4246 : pathkey->pk_opfamily,
4247 : pathkey->pk_cmptype,
4248 144754 : pathkey->pk_nulls_first,
4249 : &leftstartsel,
4250 : &leftendsel,
4251 : &rightstartsel,
4252 : &rightendsel);
4253 :
4254 : /* Cache the result in suitably long-lived workspace */
4255 144754 : oldcontext = MemoryContextSwitchTo(root->planner_cxt);
4256 :
4257 144754 : cache = palloc_object(MergeScanSelCache);
4258 144754 : cache->opfamily = pathkey->pk_opfamily;
4259 144754 : cache->collation = pathkey->pk_eclass->ec_collation;
4260 144754 : cache->cmptype = pathkey->pk_cmptype;
4261 144754 : cache->nulls_first = pathkey->pk_nulls_first;
4262 144754 : cache->leftstartsel = leftstartsel;
4263 144754 : cache->leftendsel = leftendsel;
4264 144754 : cache->rightstartsel = rightstartsel;
4265 144754 : cache->rightendsel = rightendsel;
4266 :
4267 144754 : rinfo->scansel_cache = lappend(rinfo->scansel_cache, cache);
4268 :
4269 144754 : MemoryContextSwitchTo(oldcontext);
4270 :
4271 144754 : return cache;
4272 : }
4273 :
4274 : /*
4275 : * initial_cost_hashjoin
4276 : * Preliminary estimate of the cost of a hashjoin path.
4277 : *
4278 : * This must quickly produce lower-bound estimates of the path's startup and
4279 : * total costs. If we are unable to eliminate the proposed path from
4280 : * consideration using the lower bounds, final_cost_hashjoin will be called
4281 : * to obtain the final estimates.
4282 : *
4283 : * The exact division of labor between this function and final_cost_hashjoin
4284 : * is private to them, and represents a tradeoff between speed of the initial
4285 : * estimate and getting a tight lower bound. We choose to not examine the
4286 : * join quals here (other than by counting the number of hash clauses),
4287 : * so we can't do much with CPU costs. We do assume that
4288 : * ExecChooseHashTableSize is cheap enough to use here.
4289 : *
4290 : * 'workspace' is to be filled with startup_cost, total_cost, and perhaps
4291 : * other data to be used by final_cost_hashjoin
4292 : * 'jointype' is the type of join to be performed
4293 : * 'hashclauses' is the list of joinclauses to be used as hash clauses
4294 : * 'outer_path' is the outer input to the join
4295 : * 'inner_path' is the inner input to the join
4296 : * 'extra' contains miscellaneous information about the join
4297 : * 'parallel_hash' indicates that inner_path is partial and that a shared
4298 : * hash table will be built in parallel
4299 : */
4300 : void
4301 900444 : initial_cost_hashjoin(PlannerInfo *root, JoinCostWorkspace *workspace,
4302 : JoinType jointype,
4303 : List *hashclauses,
4304 : Path *outer_path, Path *inner_path,
4305 : JoinPathExtraData *extra,
4306 : bool parallel_hash)
4307 : {
4308 : int disabled_nodes;
4309 900444 : Cost startup_cost = 0;
4310 900444 : Cost run_cost = 0;
4311 900444 : double outer_path_rows = outer_path->rows;
4312 900444 : double inner_path_rows = inner_path->rows;
4313 900444 : double inner_path_rows_total = inner_path_rows;
4314 900444 : int num_hashclauses = list_length(hashclauses);
4315 : int numbuckets;
4316 : int numbatches;
4317 : int num_skew_mcvs;
4318 : size_t space_allowed; /* unused */
4319 900444 : uint64 enable_mask = PGS_HASHJOIN;
4320 :
4321 900444 : if (outer_path->parallel_workers == 0)
4322 752596 : enable_mask |= PGS_CONSIDER_NONPARTIAL;
4323 :
4324 : /* Count up disabled nodes. */
4325 900444 : disabled_nodes = (extra->pgs_mask & enable_mask) == enable_mask ? 0 : 1;
4326 900444 : disabled_nodes += inner_path->disabled_nodes;
4327 900444 : disabled_nodes += outer_path->disabled_nodes;
4328 :
4329 : /* cost of source data */
4330 900444 : startup_cost += outer_path->startup_cost;
4331 900444 : run_cost += outer_path->total_cost - outer_path->startup_cost;
4332 900444 : startup_cost += inner_path->total_cost;
4333 :
4334 : /*
4335 : * Cost of computing hash function: must do it once per input tuple. We
4336 : * charge one cpu_operator_cost for each column's hash function. Also,
4337 : * tack on one cpu_tuple_cost per inner row, to model the costs of
4338 : * inserting the row into the hashtable.
4339 : *
4340 : * XXX when a hashclause is more complex than a single operator, we really
4341 : * should charge the extra eval costs of the left or right side, as
4342 : * appropriate, here. This seems more work than it's worth at the moment.
4343 : */
4344 900444 : startup_cost += (cpu_operator_cost * num_hashclauses + cpu_tuple_cost)
4345 900444 : * inner_path_rows;
4346 900444 : run_cost += cpu_operator_cost * num_hashclauses * outer_path_rows;
4347 :
4348 : /*
4349 : * If this is a parallel hash build, then the value we have for
4350 : * inner_rows_total currently refers only to the rows returned by each
4351 : * participant. For shared hash table size estimation, we need the total
4352 : * number, so we need to undo the division.
4353 : */
4354 900444 : if (parallel_hash)
4355 75186 : inner_path_rows_total *= get_parallel_divisor(inner_path);
4356 :
4357 : /*
4358 : * Get hash table size that executor would use for inner relation.
4359 : *
4360 : * XXX for the moment, always assume that skew optimization will be
4361 : * performed. As long as SKEW_HASH_MEM_PERCENT is small, it's not worth
4362 : * trying to determine that for sure.
4363 : *
4364 : * XXX at some point it might be interesting to try to account for skew
4365 : * optimization in the cost estimate, but for now, we don't.
4366 : */
4367 900444 : ExecChooseHashTableSize(inner_path_rows_total,
4368 900444 : inner_path->pathtarget->width,
4369 : true, /* useskew */
4370 : parallel_hash, /* try_combined_hash_mem */
4371 : outer_path->parallel_workers,
4372 : &space_allowed,
4373 : &numbuckets,
4374 : &numbatches,
4375 : &num_skew_mcvs);
4376 :
4377 : /*
4378 : * If inner relation is too big then we will need to "batch" the join,
4379 : * which implies writing and reading most of the tuples to disk an extra
4380 : * time. Charge seq_page_cost per page, since the I/O should be nice and
4381 : * sequential. Writing the inner rel counts as startup cost, all the rest
4382 : * as run cost.
4383 : */
4384 900444 : if (numbatches > 1)
4385 : {
4386 4790 : double outerpages = page_size(outer_path_rows,
4387 4790 : outer_path->pathtarget->width);
4388 4790 : double innerpages = page_size(inner_path_rows,
4389 4790 : inner_path->pathtarget->width);
4390 :
4391 4790 : startup_cost += seq_page_cost * innerpages;
4392 4790 : run_cost += seq_page_cost * (innerpages + 2 * outerpages);
4393 : }
4394 :
4395 : /* CPU costs left for later */
4396 :
4397 : /* Public result fields */
4398 900444 : workspace->disabled_nodes = disabled_nodes;
4399 900444 : workspace->startup_cost = startup_cost;
4400 900444 : workspace->total_cost = startup_cost + run_cost;
4401 : /* Save private data for final_cost_hashjoin */
4402 900444 : workspace->run_cost = run_cost;
4403 900444 : workspace->numbuckets = numbuckets;
4404 900444 : workspace->numbatches = numbatches;
4405 900444 : workspace->inner_rows_total = inner_path_rows_total;
4406 900444 : }
4407 :
4408 : /*
4409 : * final_cost_hashjoin
4410 : * Final estimate of the cost and result size of a hashjoin path.
4411 : *
4412 : * Note: the numbatches estimate is also saved into 'path' for use later
4413 : *
4414 : * 'path' is already filled in except for the rows and cost fields and
4415 : * num_batches
4416 : * 'workspace' is the result from initial_cost_hashjoin
4417 : * 'extra' contains miscellaneous information about the join
4418 : */
4419 : void
4420 465420 : final_cost_hashjoin(PlannerInfo *root, HashPath *path,
4421 : JoinCostWorkspace *workspace,
4422 : JoinPathExtraData *extra)
4423 : {
4424 465420 : Path *outer_path = path->jpath.outerjoinpath;
4425 465420 : Path *inner_path = path->jpath.innerjoinpath;
4426 465420 : double outer_path_rows = outer_path->rows;
4427 465420 : double inner_path_rows = inner_path->rows;
4428 465420 : double inner_path_rows_total = workspace->inner_rows_total;
4429 465420 : List *hashclauses = path->path_hashclauses;
4430 465420 : Cost startup_cost = workspace->startup_cost;
4431 465420 : Cost run_cost = workspace->run_cost;
4432 465420 : int numbuckets = workspace->numbuckets;
4433 465420 : int numbatches = workspace->numbatches;
4434 : Cost cpu_per_tuple;
4435 : QualCost hash_qual_cost;
4436 : QualCost qp_qual_cost;
4437 : double hashjointuples;
4438 : double virtualbuckets;
4439 : Selectivity innerbucketsize;
4440 : Selectivity innermcvfreq;
4441 : ListCell *hcl;
4442 :
4443 : /* Set the number of disabled nodes. */
4444 465420 : path->jpath.path.disabled_nodes = workspace->disabled_nodes;
4445 :
4446 : /* Mark the path with the correct row estimate */
4447 465420 : if (path->jpath.path.param_info)
4448 3668 : path->jpath.path.rows = path->jpath.path.param_info->ppi_rows;
4449 : else
4450 461752 : path->jpath.path.rows = path->jpath.path.parent->rows;
4451 :
4452 : /* For partial paths, scale row estimate. */
4453 465420 : if (path->jpath.path.parallel_workers > 0)
4454 : {
4455 107144 : double parallel_divisor = get_parallel_divisor(&path->jpath.path);
4456 :
4457 107144 : path->jpath.path.rows =
4458 107144 : clamp_row_est(path->jpath.path.rows / parallel_divisor);
4459 : }
4460 :
4461 : /* mark the path with estimated # of batches */
4462 465420 : path->num_batches = numbatches;
4463 :
4464 : /* store the total number of tuples (sum of partial row estimates) */
4465 465420 : path->inner_rows_total = inner_path_rows_total;
4466 :
4467 : /* and compute the number of "virtual" buckets in the whole join */
4468 465420 : virtualbuckets = (double) numbuckets * (double) numbatches;
4469 :
4470 : /*
4471 : * Determine bucketsize fraction and MCV frequency for the inner relation.
4472 : * We use the smallest bucketsize or MCV frequency estimated for any
4473 : * individual hashclause; this is undoubtedly conservative.
4474 : *
4475 : * BUT: if inner relation has been unique-ified, we can assume it's good
4476 : * for hashing. This is important both because it's the right answer, and
4477 : * because we avoid contaminating the cache with a value that's wrong for
4478 : * non-unique-ified paths.
4479 : */
4480 465420 : if (RELATION_WAS_MADE_UNIQUE(inner_path->parent, extra->sjinfo,
4481 : path->jpath.jointype))
4482 : {
4483 4476 : innerbucketsize = 1.0 / virtualbuckets;
4484 4476 : innermcvfreq = 1.0 / inner_path_rows_total;
4485 : }
4486 : else
4487 : {
4488 : List *otherclauses;
4489 :
4490 460944 : innerbucketsize = 1.0;
4491 460944 : innermcvfreq = 1.0;
4492 :
4493 : /* At first, try to estimate bucket size using extended statistics. */
4494 460944 : otherclauses = estimate_multivariate_bucketsize(root,
4495 : inner_path->parent,
4496 : hashclauses,
4497 : &innerbucketsize);
4498 :
4499 : /* Pass through the remaining clauses */
4500 959226 : foreach(hcl, otherclauses)
4501 : {
4502 498282 : RestrictInfo *restrictinfo = lfirst_node(RestrictInfo, hcl);
4503 : Selectivity thisbucketsize;
4504 : Selectivity thismcvfreq;
4505 :
4506 : /*
4507 : * First we have to figure out which side of the hashjoin clause
4508 : * is the inner side.
4509 : *
4510 : * Since we tend to visit the same clauses over and over when
4511 : * planning a large query, we cache the bucket stats estimates in
4512 : * the RestrictInfo node to avoid repeated lookups of statistics.
4513 : */
4514 498282 : if (bms_is_subset(restrictinfo->right_relids,
4515 498282 : inner_path->parent->relids))
4516 : {
4517 : /* righthand side is inner */
4518 258760 : thisbucketsize = restrictinfo->right_bucketsize;
4519 258760 : if (thisbucketsize < 0)
4520 : {
4521 : /* not cached yet */
4522 111656 : estimate_hash_bucket_stats(root,
4523 111656 : get_rightop(restrictinfo->clause),
4524 : virtualbuckets,
4525 : &restrictinfo->right_mcvfreq,
4526 : &restrictinfo->right_bucketsize);
4527 111656 : thisbucketsize = restrictinfo->right_bucketsize;
4528 : }
4529 258760 : thismcvfreq = restrictinfo->right_mcvfreq;
4530 : }
4531 : else
4532 : {
4533 : Assert(bms_is_subset(restrictinfo->left_relids,
4534 : inner_path->parent->relids));
4535 : /* lefthand side is inner */
4536 239522 : thisbucketsize = restrictinfo->left_bucketsize;
4537 239522 : if (thisbucketsize < 0)
4538 : {
4539 : /* not cached yet */
4540 95658 : estimate_hash_bucket_stats(root,
4541 95658 : get_leftop(restrictinfo->clause),
4542 : virtualbuckets,
4543 : &restrictinfo->left_mcvfreq,
4544 : &restrictinfo->left_bucketsize);
4545 95658 : thisbucketsize = restrictinfo->left_bucketsize;
4546 : }
4547 239522 : thismcvfreq = restrictinfo->left_mcvfreq;
4548 : }
4549 :
4550 498282 : if (innerbucketsize > thisbucketsize)
4551 365546 : innerbucketsize = thisbucketsize;
4552 : /* Disregard zero for MCV freq, it means we have no data */
4553 498282 : if (thismcvfreq > 0.0 && innermcvfreq > thismcvfreq)
4554 369432 : innermcvfreq = thismcvfreq;
4555 : }
4556 : }
4557 :
4558 : /*
4559 : * If the bucket holding the inner MCV would exceed hash_mem, we don't
4560 : * want to hash unless there is really no other alternative, so apply
4561 : * disable_cost. (The executor normally copes with excessive memory usage
4562 : * by splitting batches, but obviously it cannot separate equal values
4563 : * that way, so it will be unable to drive the batch size below hash_mem
4564 : * when this is true.)
4565 : */
4566 465420 : if (relation_byte_size(clamp_row_est(inner_path_rows * innermcvfreq),
4567 930840 : inner_path->pathtarget->width) > get_hash_memory_limit())
4568 116 : startup_cost += disable_cost;
4569 :
4570 : /*
4571 : * Compute cost of the hashquals and qpquals (other restriction clauses)
4572 : * separately.
4573 : */
4574 465420 : cost_qual_eval(&hash_qual_cost, hashclauses, root);
4575 465420 : cost_qual_eval(&qp_qual_cost, path->jpath.joinrestrictinfo, root);
4576 465420 : qp_qual_cost.startup -= hash_qual_cost.startup;
4577 465420 : qp_qual_cost.per_tuple -= hash_qual_cost.per_tuple;
4578 :
4579 : /* CPU costs */
4580 :
4581 465420 : if (path->jpath.jointype == JOIN_SEMI ||
4582 459056 : path->jpath.jointype == JOIN_ANTI ||
4583 454484 : extra->inner_unique)
4584 127718 : {
4585 : double outer_matched_rows;
4586 : Selectivity inner_scan_frac;
4587 :
4588 : /*
4589 : * With a SEMI or ANTI join, or if the innerrel is known unique, the
4590 : * executor will stop after the first match.
4591 : *
4592 : * For an outer-rel row that has at least one match, we can expect the
4593 : * bucket scan to stop after a fraction 1/(match_count+1) of the
4594 : * bucket's rows, if the matches are evenly distributed. Since they
4595 : * probably aren't quite evenly distributed, we apply a fuzz factor of
4596 : * 2.0 to that fraction. (If we used a larger fuzz factor, we'd have
4597 : * to clamp inner_scan_frac to at most 1.0; but since match_count is
4598 : * at least 1, no such clamp is needed now.)
4599 : */
4600 127718 : outer_matched_rows = rint(outer_path_rows * extra->semifactors.outer_match_frac);
4601 127718 : inner_scan_frac = 2.0 / (extra->semifactors.match_count + 1.0);
4602 :
4603 127718 : startup_cost += hash_qual_cost.startup;
4604 255436 : run_cost += hash_qual_cost.per_tuple * outer_matched_rows *
4605 127718 : clamp_row_est(inner_path_rows * innerbucketsize * inner_scan_frac) * 0.5;
4606 :
4607 : /*
4608 : * For unmatched outer-rel rows, the picture is quite a lot different.
4609 : * In the first place, there is no reason to assume that these rows
4610 : * preferentially hit heavily-populated buckets; instead assume they
4611 : * are uncorrelated with the inner distribution and so they see an
4612 : * average bucket size of inner_path_rows / virtualbuckets. In the
4613 : * second place, it seems likely that they will have few if any exact
4614 : * hash-code matches and so very few of the tuples in the bucket will
4615 : * actually require eval of the hash quals. We don't have any good
4616 : * way to estimate how many will, but for the moment assume that the
4617 : * effective cost per bucket entry is one-tenth what it is for
4618 : * matchable tuples.
4619 : */
4620 255436 : run_cost += hash_qual_cost.per_tuple *
4621 255436 : (outer_path_rows - outer_matched_rows) *
4622 127718 : clamp_row_est(inner_path_rows / virtualbuckets) * 0.05;
4623 :
4624 : /* Get # of tuples that will pass the basic join */
4625 127718 : if (path->jpath.jointype == JOIN_ANTI)
4626 4572 : hashjointuples = outer_path_rows - outer_matched_rows;
4627 : else
4628 123146 : hashjointuples = outer_matched_rows;
4629 : }
4630 : else
4631 : {
4632 : /*
4633 : * The number of tuple comparisons needed is the number of outer
4634 : * tuples times the typical number of tuples in a hash bucket, which
4635 : * is the inner relation size times its bucketsize fraction. At each
4636 : * one, we need to evaluate the hashjoin quals. But actually,
4637 : * charging the full qual eval cost at each tuple is pessimistic,
4638 : * since we don't evaluate the quals unless the hash values match
4639 : * exactly. For lack of a better idea, halve the cost estimate to
4640 : * allow for that.
4641 : */
4642 337702 : startup_cost += hash_qual_cost.startup;
4643 675404 : run_cost += hash_qual_cost.per_tuple * outer_path_rows *
4644 337702 : clamp_row_est(inner_path_rows * innerbucketsize) * 0.5;
4645 :
4646 : /*
4647 : * Get approx # tuples passing the hashquals. We use
4648 : * approx_tuple_count here because we need an estimate done with
4649 : * JOIN_INNER semantics.
4650 : */
4651 337702 : hashjointuples = approx_tuple_count(root, &path->jpath, hashclauses);
4652 : }
4653 :
4654 : /*
4655 : * For each tuple that gets through the hashjoin proper, we charge
4656 : * cpu_tuple_cost plus the cost of evaluating additional restriction
4657 : * clauses that are to be applied at the join. (This is pessimistic since
4658 : * not all of the quals may get evaluated at each tuple.)
4659 : */
4660 465420 : startup_cost += qp_qual_cost.startup;
4661 465420 : cpu_per_tuple = cpu_tuple_cost + qp_qual_cost.per_tuple;
4662 465420 : run_cost += cpu_per_tuple * hashjointuples;
4663 :
4664 : /* tlist eval costs are paid per output row, not per tuple scanned */
4665 465420 : startup_cost += path->jpath.path.pathtarget->cost.startup;
4666 465420 : run_cost += path->jpath.path.pathtarget->cost.per_tuple * path->jpath.path.rows;
4667 :
4668 465420 : path->jpath.path.startup_cost = startup_cost;
4669 465420 : path->jpath.path.total_cost = startup_cost + run_cost;
4670 465420 : }
4671 :
4672 :
4673 : /*
4674 : * cost_subplan
4675 : * Figure the costs for a SubPlan (or initplan).
4676 : *
4677 : * Note: we could dig the subplan's Plan out of the root list, but in practice
4678 : * all callers have it handy already, so we make them pass it.
4679 : */
4680 : void
4681 44336 : cost_subplan(PlannerInfo *root, SubPlan *subplan, Plan *plan)
4682 : {
4683 : QualCost sp_cost;
4684 :
4685 : /*
4686 : * Figure any cost for evaluating the testexpr.
4687 : *
4688 : * Usually, SubPlan nodes are built very early, before we have constructed
4689 : * any RelOptInfos for the parent query level, which means the parent root
4690 : * does not yet contain enough information to safely consult statistics.
4691 : * Therefore, we pass root as NULL here. cost_qual_eval() is already
4692 : * well-equipped to handle a NULL root.
4693 : *
4694 : * One exception is SubPlan nodes built for the initplans of MIN/MAX
4695 : * aggregates from indexes (cf. SS_make_initplan_from_plan). In this
4696 : * case, having a NULL root is safe because testexpr will be NULL.
4697 : * Besides, an initplan will by definition not consult anything from the
4698 : * parent plan.
4699 : */
4700 44336 : cost_qual_eval(&sp_cost,
4701 44336 : make_ands_implicit((Expr *) subplan->testexpr),
4702 : NULL);
4703 :
4704 44336 : if (subplan->useHashTable)
4705 : {
4706 : /*
4707 : * If we are using a hash table for the subquery outputs, then the
4708 : * cost of evaluating the query is a one-time cost. We charge one
4709 : * cpu_operator_cost per tuple for the work of loading the hashtable,
4710 : * too.
4711 : */
4712 2128 : sp_cost.startup += plan->total_cost +
4713 2128 : cpu_operator_cost * plan->plan_rows;
4714 :
4715 : /*
4716 : * The per-tuple costs include the cost of evaluating the lefthand
4717 : * expressions, plus the cost of probing the hashtable. We already
4718 : * accounted for the lefthand expressions as part of the testexpr, and
4719 : * will also have counted one cpu_operator_cost for each comparison
4720 : * operator. That is probably too low for the probing cost, but it's
4721 : * hard to make a better estimate, so live with it for now.
4722 : */
4723 : }
4724 : else
4725 : {
4726 : /*
4727 : * Otherwise we will be rescanning the subplan output on each
4728 : * evaluation. We need to estimate how much of the output we will
4729 : * actually need to scan. NOTE: this logic should agree with the
4730 : * tuple_fraction estimates used by make_subplan() in
4731 : * plan/subselect.c.
4732 : */
4733 42208 : Cost plan_run_cost = plan->total_cost - plan->startup_cost;
4734 :
4735 42208 : if (subplan->subLinkType == EXISTS_SUBLINK)
4736 : {
4737 : /* we only need to fetch 1 tuple; clamp to avoid zero divide */
4738 2576 : sp_cost.per_tuple += plan_run_cost / clamp_row_est(plan->plan_rows);
4739 : }
4740 39632 : else if (subplan->subLinkType == ALL_SUBLINK ||
4741 39614 : subplan->subLinkType == ANY_SUBLINK)
4742 : {
4743 : /* assume we need 50% of the tuples */
4744 146 : sp_cost.per_tuple += 0.50 * plan_run_cost;
4745 : /* also charge a cpu_operator_cost per row examined */
4746 146 : sp_cost.per_tuple += 0.50 * plan->plan_rows * cpu_operator_cost;
4747 : }
4748 : else
4749 : {
4750 : /* assume we need all tuples */
4751 39486 : sp_cost.per_tuple += plan_run_cost;
4752 : }
4753 :
4754 : /*
4755 : * Also account for subplan's startup cost. If the subplan is
4756 : * uncorrelated or undirect correlated, AND its topmost node is one
4757 : * that materializes its output, assume that we'll only need to pay
4758 : * its startup cost once; otherwise assume we pay the startup cost
4759 : * every time.
4760 : */
4761 55598 : if (subplan->parParam == NIL &&
4762 13390 : ExecMaterializesOutput(nodeTag(plan)))
4763 774 : sp_cost.startup += plan->startup_cost;
4764 : else
4765 41434 : sp_cost.per_tuple += plan->startup_cost;
4766 : }
4767 :
4768 44336 : subplan->startup_cost = sp_cost.startup;
4769 44336 : subplan->per_call_cost = sp_cost.per_tuple;
4770 44336 : }
4771 :
4772 :
4773 : /*
4774 : * cost_rescan
4775 : * Given a finished Path, estimate the costs of rescanning it after
4776 : * having done so the first time. For some Path types a rescan is
4777 : * cheaper than an original scan (if no parameters change), and this
4778 : * function embodies knowledge about that. The default is to return
4779 : * the same costs stored in the Path. (Note that the cost estimates
4780 : * actually stored in Paths are always for first scans.)
4781 : *
4782 : * This function is not currently intended to model effects such as rescans
4783 : * being cheaper due to disk block caching; what we are concerned with is
4784 : * plan types wherein the executor caches results explicitly, or doesn't
4785 : * redo startup calculations, etc.
4786 : */
4787 : static void
4788 3386462 : cost_rescan(PlannerInfo *root, Path *path,
4789 : Cost *rescan_startup_cost, /* output parameters */
4790 : Cost *rescan_total_cost)
4791 : {
4792 3386462 : switch (path->pathtype)
4793 : {
4794 53980 : case T_FunctionScan:
4795 :
4796 : /*
4797 : * Currently, nodeFunctionscan.c always executes the function to
4798 : * completion before returning any rows, and caches the results in
4799 : * a tuplestore. So the function eval cost is all startup cost
4800 : * and isn't paid over again on rescans. However, all run costs
4801 : * will be paid over again.
4802 : */
4803 53980 : *rescan_startup_cost = 0;
4804 53980 : *rescan_total_cost = path->total_cost - path->startup_cost;
4805 53980 : break;
4806 133860 : case T_HashJoin:
4807 :
4808 : /*
4809 : * If it's a single-batch join, we don't need to rebuild the hash
4810 : * table during a rescan.
4811 : */
4812 133860 : if (((HashPath *) path)->num_batches == 1)
4813 : {
4814 : /* Startup cost is exactly the cost of hash table building */
4815 133860 : *rescan_startup_cost = 0;
4816 133860 : *rescan_total_cost = path->total_cost - path->startup_cost;
4817 : }
4818 : else
4819 : {
4820 : /* Otherwise, no special treatment */
4821 0 : *rescan_startup_cost = path->startup_cost;
4822 0 : *rescan_total_cost = path->total_cost;
4823 : }
4824 133860 : break;
4825 8142 : case T_CteScan:
4826 : case T_WorkTableScan:
4827 : {
4828 : /*
4829 : * These plan types materialize their final result in a
4830 : * tuplestore or tuplesort object. So the rescan cost is only
4831 : * cpu_tuple_cost per tuple, unless the result is large enough
4832 : * to spill to disk.
4833 : */
4834 8142 : Cost run_cost = cpu_tuple_cost * path->rows;
4835 8142 : double nbytes = relation_byte_size(path->rows,
4836 8142 : path->pathtarget->width);
4837 8142 : double work_mem_bytes = work_mem * (Size) 1024;
4838 :
4839 8142 : if (nbytes > work_mem_bytes)
4840 : {
4841 : /* It will spill, so account for re-read cost */
4842 368 : double npages = ceil(nbytes / BLCKSZ);
4843 :
4844 368 : run_cost += seq_page_cost * npages;
4845 : }
4846 8142 : *rescan_startup_cost = 0;
4847 8142 : *rescan_total_cost = run_cost;
4848 : }
4849 8142 : break;
4850 1222530 : case T_Material:
4851 : case T_Sort:
4852 : {
4853 : /*
4854 : * These plan types not only materialize their results, but do
4855 : * not implement qual filtering or projection. So they are
4856 : * even cheaper to rescan than the ones above. We charge only
4857 : * cpu_operator_cost per tuple. (Note: keep that in sync with
4858 : * the run_cost charge in cost_sort, and also see comments in
4859 : * cost_material before you change it.)
4860 : */
4861 1222530 : Cost run_cost = cpu_operator_cost * path->rows;
4862 1222530 : double nbytes = relation_byte_size(path->rows,
4863 1222530 : path->pathtarget->width);
4864 1222530 : double work_mem_bytes = work_mem * (Size) 1024;
4865 :
4866 1222530 : if (nbytes > work_mem_bytes)
4867 : {
4868 : /* It will spill, so account for re-read cost */
4869 10150 : double npages = ceil(nbytes / BLCKSZ);
4870 :
4871 10150 : run_cost += seq_page_cost * npages;
4872 : }
4873 1222530 : *rescan_startup_cost = 0;
4874 1222530 : *rescan_total_cost = run_cost;
4875 : }
4876 1222530 : break;
4877 297590 : case T_Memoize:
4878 : /* All the hard work is done by cost_memoize_rescan */
4879 297590 : cost_memoize_rescan(root, (MemoizePath *) path,
4880 : rescan_startup_cost, rescan_total_cost);
4881 297590 : break;
4882 1670360 : default:
4883 1670360 : *rescan_startup_cost = path->startup_cost;
4884 1670360 : *rescan_total_cost = path->total_cost;
4885 1670360 : break;
4886 : }
4887 3386462 : }
4888 :
4889 :
4890 : /*
4891 : * cost_qual_eval
4892 : * Estimate the CPU costs of evaluating a WHERE clause.
4893 : * The input can be either an implicitly-ANDed list of boolean
4894 : * expressions, or a list of RestrictInfo nodes. (The latter is
4895 : * preferred since it allows caching of the results.)
4896 : * The result includes both a one-time (startup) component,
4897 : * and a per-evaluation component.
4898 : *
4899 : * Note: in some code paths root can be passed as NULL, resulting in
4900 : * slightly worse estimates.
4901 : */
4902 : void
4903 4801972 : cost_qual_eval(QualCost *cost, List *quals, PlannerInfo *root)
4904 : {
4905 : cost_qual_eval_context context;
4906 : ListCell *l;
4907 :
4908 4801972 : context.root = root;
4909 4801972 : context.total.startup = 0;
4910 4801972 : context.total.per_tuple = 0;
4911 :
4912 : /* We don't charge any cost for the implicit ANDing at top level ... */
4913 :
4914 9144806 : foreach(l, quals)
4915 : {
4916 4342834 : Node *qual = (Node *) lfirst(l);
4917 :
4918 4342834 : cost_qual_eval_walker(qual, &context);
4919 : }
4920 :
4921 4801972 : *cost = context.total;
4922 4801972 : }
4923 :
4924 : /*
4925 : * cost_qual_eval_node
4926 : * As above, for a single RestrictInfo or expression.
4927 : */
4928 : void
4929 1897558 : cost_qual_eval_node(QualCost *cost, Node *qual, PlannerInfo *root)
4930 : {
4931 : cost_qual_eval_context context;
4932 :
4933 1897558 : context.root = root;
4934 1897558 : context.total.startup = 0;
4935 1897558 : context.total.per_tuple = 0;
4936 :
4937 1897558 : cost_qual_eval_walker(qual, &context);
4938 :
4939 1897558 : *cost = context.total;
4940 1897558 : }
4941 :
4942 : static bool
4943 9775244 : cost_qual_eval_walker(Node *node, cost_qual_eval_context *context)
4944 : {
4945 9775244 : if (node == NULL)
4946 95670 : return false;
4947 :
4948 : /*
4949 : * RestrictInfo nodes contain an eval_cost field reserved for this
4950 : * routine's use, so that it's not necessary to evaluate the qual clause's
4951 : * cost more than once. If the clause's cost hasn't been computed yet,
4952 : * the field's startup value will contain -1.
4953 : */
4954 9679574 : if (IsA(node, RestrictInfo))
4955 : {
4956 4534470 : RestrictInfo *rinfo = (RestrictInfo *) node;
4957 :
4958 4534470 : if (rinfo->eval_cost.startup < 0)
4959 : {
4960 : cost_qual_eval_context locContext;
4961 :
4962 615812 : locContext.root = context->root;
4963 615812 : locContext.total.startup = 0;
4964 615812 : locContext.total.per_tuple = 0;
4965 :
4966 : /*
4967 : * For an OR clause, recurse into the marked-up tree so that we
4968 : * set the eval_cost for contained RestrictInfos too.
4969 : */
4970 615812 : if (rinfo->orclause)
4971 10056 : cost_qual_eval_walker((Node *) rinfo->orclause, &locContext);
4972 : else
4973 605756 : cost_qual_eval_walker((Node *) rinfo->clause, &locContext);
4974 :
4975 : /*
4976 : * If the RestrictInfo is marked pseudoconstant, it will be tested
4977 : * only once, so treat its cost as all startup cost.
4978 : */
4979 615812 : if (rinfo->pseudoconstant)
4980 : {
4981 : /* count one execution during startup */
4982 10310 : locContext.total.startup += locContext.total.per_tuple;
4983 10310 : locContext.total.per_tuple = 0;
4984 : }
4985 615812 : rinfo->eval_cost = locContext.total;
4986 : }
4987 4534470 : context->total.startup += rinfo->eval_cost.startup;
4988 4534470 : context->total.per_tuple += rinfo->eval_cost.per_tuple;
4989 : /* do NOT recurse into children */
4990 4534470 : return false;
4991 : }
4992 :
4993 : /*
4994 : * For each operator or function node in the given tree, we charge the
4995 : * estimated execution cost given by pg_proc.procost (remember to multiply
4996 : * this by cpu_operator_cost).
4997 : *
4998 : * Vars and Consts are charged zero, and so are boolean operators (AND,
4999 : * OR, NOT). Simplistic, but a lot better than no model at all.
5000 : *
5001 : * Should we try to account for the possibility of short-circuit
5002 : * evaluation of AND/OR? Probably *not*, because that would make the
5003 : * results depend on the clause ordering, and we are not in any position
5004 : * to expect that the current ordering of the clauses is the one that's
5005 : * going to end up being used. The above per-RestrictInfo caching would
5006 : * not mix well with trying to re-order clauses anyway.
5007 : *
5008 : * Another issue that is entirely ignored here is that if a set-returning
5009 : * function is below top level in the tree, the functions/operators above
5010 : * it will need to be evaluated multiple times. In practical use, such
5011 : * cases arise so seldom as to not be worth the added complexity needed;
5012 : * moreover, since our rowcount estimates for functions tend to be pretty
5013 : * phony, the results would also be pretty phony.
5014 : */
5015 5145104 : if (IsA(node, FuncExpr))
5016 : {
5017 349630 : add_function_cost(context->root, ((FuncExpr *) node)->funcid, node,
5018 : &context->total);
5019 : }
5020 4795474 : else if (IsA(node, OpExpr) ||
5021 4124022 : IsA(node, DistinctExpr) ||
5022 4122938 : IsA(node, NullIfExpr))
5023 : {
5024 : /* rely on struct equivalence to treat these all alike */
5025 672660 : set_opfuncid((OpExpr *) node);
5026 672660 : add_function_cost(context->root, ((OpExpr *) node)->opfuncid, node,
5027 : &context->total);
5028 : }
5029 4122814 : else if (IsA(node, ScalarArrayOpExpr))
5030 : {
5031 45482 : ScalarArrayOpExpr *saop = (ScalarArrayOpExpr *) node;
5032 45482 : Node *arraynode = (Node *) lsecond(saop->args);
5033 : QualCost sacosts;
5034 : QualCost hcosts;
5035 45482 : double estarraylen = estimate_array_length(context->root, arraynode);
5036 :
5037 45482 : set_sa_opfuncid(saop);
5038 45482 : sacosts.startup = sacosts.per_tuple = 0;
5039 45482 : add_function_cost(context->root, saop->opfuncid, NULL,
5040 : &sacosts);
5041 :
5042 45482 : if (OidIsValid(saop->hashfuncid))
5043 : {
5044 : /* Handle costs for hashed ScalarArrayOpExpr */
5045 430 : hcosts.startup = hcosts.per_tuple = 0;
5046 :
5047 430 : add_function_cost(context->root, saop->hashfuncid, NULL, &hcosts);
5048 430 : context->total.startup += sacosts.startup + hcosts.startup;
5049 :
5050 : /* Estimate the cost of building the hashtable. */
5051 430 : context->total.startup += estarraylen * hcosts.per_tuple;
5052 :
5053 : /*
5054 : * XXX should we charge a little bit for sacosts.per_tuple when
5055 : * building the table, or is it ok to assume there will be zero
5056 : * hash collision?
5057 : */
5058 :
5059 : /*
5060 : * Charge for hashtable lookups. Charge a single hash and a
5061 : * single comparison.
5062 : */
5063 430 : context->total.per_tuple += hcosts.per_tuple + sacosts.per_tuple;
5064 : }
5065 : else
5066 : {
5067 : /*
5068 : * Estimate that the operator will be applied to about half of the
5069 : * array elements before the answer is determined.
5070 : */
5071 45052 : context->total.startup += sacosts.startup;
5072 90104 : context->total.per_tuple += sacosts.per_tuple *
5073 45052 : estimate_array_length(context->root, arraynode) * 0.5;
5074 : }
5075 : }
5076 4077332 : else if (IsA(node, Aggref) ||
5077 4004342 : IsA(node, WindowFunc))
5078 : {
5079 : /*
5080 : * Aggref and WindowFunc nodes are (and should be) treated like Vars,
5081 : * ie, zero execution cost in the current model, because they behave
5082 : * essentially like Vars at execution. We disregard the costs of
5083 : * their input expressions for the same reason. The actual execution
5084 : * costs of the aggregate/window functions and their arguments have to
5085 : * be factored into plan-node-specific costing of the Agg or WindowAgg
5086 : * plan node.
5087 : */
5088 76868 : return false; /* don't recurse into children */
5089 : }
5090 4000464 : else if (IsA(node, GroupingFunc))
5091 : {
5092 : /* Treat this as having cost 1 */
5093 430 : context->total.per_tuple += cpu_operator_cost;
5094 430 : return false; /* don't recurse into children */
5095 : }
5096 4000034 : else if (IsA(node, CoerceViaIO))
5097 : {
5098 22808 : CoerceViaIO *iocoerce = (CoerceViaIO *) node;
5099 : Oid iofunc;
5100 : Oid typioparam;
5101 : bool typisvarlena;
5102 :
5103 : /* check the result type's input function */
5104 22808 : getTypeInputInfo(iocoerce->resulttype,
5105 : &iofunc, &typioparam);
5106 22808 : add_function_cost(context->root, iofunc, NULL,
5107 : &context->total);
5108 : /* check the input type's output function */
5109 22808 : getTypeOutputInfo(exprType((Node *) iocoerce->arg),
5110 : &iofunc, &typisvarlena);
5111 22808 : add_function_cost(context->root, iofunc, NULL,
5112 : &context->total);
5113 : }
5114 3977226 : else if (IsA(node, ArrayCoerceExpr))
5115 : {
5116 5210 : ArrayCoerceExpr *acoerce = (ArrayCoerceExpr *) node;
5117 : QualCost perelemcost;
5118 :
5119 5210 : cost_qual_eval_node(&perelemcost, (Node *) acoerce->elemexpr,
5120 : context->root);
5121 5210 : context->total.startup += perelemcost.startup;
5122 5210 : if (perelemcost.per_tuple > 0)
5123 66 : context->total.per_tuple += perelemcost.per_tuple *
5124 66 : estimate_array_length(context->root, (Node *) acoerce->arg);
5125 : }
5126 3972016 : else if (IsA(node, RowCompareExpr))
5127 : {
5128 : /* Conservatively assume we will check all the columns */
5129 252 : RowCompareExpr *rcexpr = (RowCompareExpr *) node;
5130 : ListCell *lc;
5131 :
5132 810 : foreach(lc, rcexpr->opnos)
5133 : {
5134 558 : Oid opid = lfirst_oid(lc);
5135 :
5136 558 : add_function_cost(context->root, get_opcode(opid), NULL,
5137 : &context->total);
5138 : }
5139 : }
5140 3971764 : else if (IsA(node, MinMaxExpr) ||
5141 3971486 : IsA(node, SQLValueFunction) ||
5142 3966692 : IsA(node, XmlExpr) ||
5143 3965990 : IsA(node, CoerceToDomain) ||
5144 3956132 : IsA(node, NextValueExpr) ||
5145 3955736 : IsA(node, JsonExpr))
5146 : {
5147 : /* Treat all these as having cost 1 */
5148 18600 : context->total.per_tuple += cpu_operator_cost;
5149 : }
5150 3953164 : else if (IsA(node, SubLink))
5151 : {
5152 : /* This routine should not be applied to un-planned expressions */
5153 0 : elog(ERROR, "cannot handle unplanned sub-select");
5154 : }
5155 3953164 : else if (IsA(node, SubPlan))
5156 : {
5157 : /*
5158 : * A subplan node in an expression typically indicates that the
5159 : * subplan will be executed on each evaluation, so charge accordingly.
5160 : * (Sub-selects that can be executed as InitPlans have already been
5161 : * removed from the expression.)
5162 : */
5163 43828 : SubPlan *subplan = (SubPlan *) node;
5164 :
5165 43828 : context->total.startup += subplan->startup_cost;
5166 43828 : context->total.per_tuple += subplan->per_call_cost;
5167 :
5168 : /*
5169 : * We don't want to recurse into the testexpr, because it was already
5170 : * counted in the SubPlan node's costs. So we're done.
5171 : */
5172 43828 : return false;
5173 : }
5174 3909336 : else if (IsA(node, AlternativeSubPlan))
5175 : {
5176 : /*
5177 : * Arbitrarily use the first alternative plan for costing. (We should
5178 : * certainly only include one alternative, and we don't yet have
5179 : * enough information to know which one the executor is most likely to
5180 : * use.)
5181 : */
5182 1860 : AlternativeSubPlan *asplan = (AlternativeSubPlan *) node;
5183 :
5184 1860 : return cost_qual_eval_walker((Node *) linitial(asplan->subplans),
5185 : context);
5186 : }
5187 3907476 : else if (IsA(node, PlaceHolderVar))
5188 : {
5189 : /*
5190 : * A PlaceHolderVar should be given cost zero when considering general
5191 : * expression evaluation costs. The expense of doing the contained
5192 : * expression is charged as part of the tlist eval costs of the scan
5193 : * or join where the PHV is first computed (see set_rel_width and
5194 : * add_placeholders_to_joinrel). If we charged it again here, we'd be
5195 : * double-counting the cost for each level of plan that the PHV
5196 : * bubbles up through. Hence, return without recursing into the
5197 : * phexpr.
5198 : */
5199 5636 : return false;
5200 : }
5201 :
5202 : /* recurse into children */
5203 5016482 : return expression_tree_walker(node, cost_qual_eval_walker, context);
5204 : }
5205 :
5206 : /*
5207 : * get_restriction_qual_cost
5208 : * Compute evaluation costs of a baserel's restriction quals, plus any
5209 : * movable join quals that have been pushed down to the scan.
5210 : * Results are returned into *qpqual_cost.
5211 : *
5212 : * This is a convenience subroutine that works for seqscans and other cases
5213 : * where all the given quals will be evaluated the hard way. It's not useful
5214 : * for cost_index(), for example, where the index machinery takes care of
5215 : * some of the quals. We assume baserestrictcost was previously set by
5216 : * set_baserel_size_estimates().
5217 : */
5218 : static void
5219 1126656 : get_restriction_qual_cost(PlannerInfo *root, RelOptInfo *baserel,
5220 : ParamPathInfo *param_info,
5221 : QualCost *qpqual_cost)
5222 : {
5223 1126656 : if (param_info)
5224 : {
5225 : /* Include costs of pushed-down clauses */
5226 251846 : cost_qual_eval(qpqual_cost, param_info->ppi_clauses, root);
5227 :
5228 251846 : qpqual_cost->startup += baserel->baserestrictcost.startup;
5229 251846 : qpqual_cost->per_tuple += baserel->baserestrictcost.per_tuple;
5230 : }
5231 : else
5232 874810 : *qpqual_cost = baserel->baserestrictcost;
5233 1126656 : }
5234 :
5235 :
5236 : /*
5237 : * compute_semi_anti_join_factors
5238 : * Estimate how much of the inner input a SEMI, ANTI, or inner_unique join
5239 : * can be expected to scan.
5240 : *
5241 : * In a hash or nestloop SEMI/ANTI join, the executor will stop scanning
5242 : * inner rows as soon as it finds a match to the current outer row.
5243 : * The same happens if we have detected the inner rel is unique.
5244 : * We should therefore adjust some of the cost components for this effect.
5245 : * This function computes some estimates needed for these adjustments.
5246 : * These estimates will be the same regardless of the particular paths used
5247 : * for the outer and inner relation, so we compute these once and then pass
5248 : * them to all the join cost estimation functions.
5249 : *
5250 : * Input parameters:
5251 : * joinrel: join relation under consideration
5252 : * outerrel: outer relation under consideration
5253 : * innerrel: inner relation under consideration
5254 : * jointype: if not JOIN_SEMI or JOIN_ANTI, we assume it's inner_unique
5255 : * sjinfo: SpecialJoinInfo relevant to this join
5256 : * restrictlist: join quals
5257 : * Output parameters:
5258 : * *semifactors is filled in (see pathnodes.h for field definitions)
5259 : */
5260 : void
5261 223802 : compute_semi_anti_join_factors(PlannerInfo *root,
5262 : RelOptInfo *joinrel,
5263 : RelOptInfo *outerrel,
5264 : RelOptInfo *innerrel,
5265 : JoinType jointype,
5266 : SpecialJoinInfo *sjinfo,
5267 : List *restrictlist,
5268 : SemiAntiJoinFactors *semifactors)
5269 : {
5270 : Selectivity jselec;
5271 : Selectivity nselec;
5272 : Selectivity avgmatch;
5273 : SpecialJoinInfo norm_sjinfo;
5274 : List *joinquals;
5275 : ListCell *l;
5276 :
5277 : /*
5278 : * In an ANTI join, we must ignore clauses that are "pushed down", since
5279 : * those won't affect the match logic. In a SEMI join, we do not
5280 : * distinguish joinquals from "pushed down" quals, so just use the whole
5281 : * restrictinfo list. For other outer join types, we should consider only
5282 : * non-pushed-down quals, so that this devolves to an IS_OUTER_JOIN check.
5283 : */
5284 223802 : if (IS_OUTER_JOIN(jointype))
5285 : {
5286 77484 : joinquals = NIL;
5287 169254 : foreach(l, restrictlist)
5288 : {
5289 91770 : RestrictInfo *rinfo = lfirst_node(RestrictInfo, l);
5290 :
5291 91770 : if (!RINFO_IS_PUSHED_DOWN(rinfo, joinrel->relids))
5292 86876 : joinquals = lappend(joinquals, rinfo);
5293 : }
5294 : }
5295 : else
5296 146318 : joinquals = restrictlist;
5297 :
5298 : /*
5299 : * Get the JOIN_SEMI or JOIN_ANTI selectivity of the join clauses.
5300 : */
5301 223802 : jselec = clauselist_selectivity(root,
5302 : joinquals,
5303 : 0,
5304 : (jointype == JOIN_ANTI) ? JOIN_ANTI : JOIN_SEMI,
5305 : sjinfo);
5306 :
5307 : /*
5308 : * Also get the normal inner-join selectivity of the join clauses.
5309 : */
5310 223802 : init_dummy_sjinfo(&norm_sjinfo, outerrel->relids, innerrel->relids);
5311 :
5312 223802 : nselec = clauselist_selectivity(root,
5313 : joinquals,
5314 : 0,
5315 : JOIN_INNER,
5316 : &norm_sjinfo);
5317 :
5318 : /* Avoid leaking a lot of ListCells */
5319 223802 : if (IS_OUTER_JOIN(jointype))
5320 77484 : list_free(joinquals);
5321 :
5322 : /*
5323 : * jselec can be interpreted as the fraction of outer-rel rows that have
5324 : * any matches (this is true for both SEMI and ANTI cases). And nselec is
5325 : * the fraction of the Cartesian product that matches. So, the average
5326 : * number of matches for each outer-rel row that has at least one match is
5327 : * nselec * inner_rows / jselec.
5328 : *
5329 : * Note: it is correct to use the inner rel's "rows" count here, even
5330 : * though we might later be considering a parameterized inner path with
5331 : * fewer rows. This is because we have included all the join clauses in
5332 : * the selectivity estimate.
5333 : */
5334 223802 : if (jselec > 0) /* protect against zero divide */
5335 : {
5336 223434 : avgmatch = nselec * innerrel->rows / jselec;
5337 : /* Clamp to sane range */
5338 223434 : avgmatch = Max(1.0, avgmatch);
5339 : }
5340 : else
5341 368 : avgmatch = 1.0;
5342 :
5343 223802 : semifactors->outer_match_frac = jselec;
5344 223802 : semifactors->match_count = avgmatch;
5345 223802 : }
5346 :
5347 : /*
5348 : * has_indexed_join_quals
5349 : * Check whether all the joinquals of a nestloop join are used as
5350 : * inner index quals.
5351 : *
5352 : * If the inner path of a SEMI/ANTI join is an indexscan (including bitmap
5353 : * indexscan) that uses all the joinquals as indexquals, we can assume that an
5354 : * unmatched outer tuple is cheap to process, whereas otherwise it's probably
5355 : * expensive.
5356 : */
5357 : static bool
5358 955428 : has_indexed_join_quals(NestPath *path)
5359 : {
5360 955428 : JoinPath *joinpath = &path->jpath;
5361 955428 : Relids joinrelids = joinpath->path.parent->relids;
5362 955428 : Path *innerpath = joinpath->innerjoinpath;
5363 : List *indexclauses;
5364 : bool found_one;
5365 : ListCell *lc;
5366 :
5367 : /* If join still has quals to evaluate, it's not fast */
5368 955428 : if (joinpath->joinrestrictinfo != NIL)
5369 680398 : return false;
5370 : /* Nor if the inner path isn't parameterized at all */
5371 275030 : if (innerpath->param_info == NULL)
5372 3300 : return false;
5373 :
5374 : /* Find the indexclauses list for the inner scan */
5375 271730 : switch (innerpath->pathtype)
5376 : {
5377 166460 : case T_IndexScan:
5378 : case T_IndexOnlyScan:
5379 166460 : indexclauses = ((IndexPath *) innerpath)->indexclauses;
5380 166460 : break;
5381 270 : case T_BitmapHeapScan:
5382 : {
5383 : /* Accept only a simple bitmap scan, not AND/OR cases */
5384 270 : Path *bmqual = ((BitmapHeapPath *) innerpath)->bitmapqual;
5385 :
5386 270 : if (IsA(bmqual, IndexPath))
5387 222 : indexclauses = ((IndexPath *) bmqual)->indexclauses;
5388 : else
5389 48 : return false;
5390 222 : break;
5391 : }
5392 105000 : default:
5393 :
5394 : /*
5395 : * If it's not a simple indexscan, it probably doesn't run quickly
5396 : * for zero rows out, even if it's a parameterized path using all
5397 : * the joinquals.
5398 : */
5399 105000 : return false;
5400 : }
5401 :
5402 : /*
5403 : * Examine the inner path's param clauses. Any that are from the outer
5404 : * path must be found in the indexclauses list, either exactly or in an
5405 : * equivalent form generated by equivclass.c. Also, we must find at least
5406 : * one such clause, else it's a clauseless join which isn't fast.
5407 : */
5408 166682 : found_one = false;
5409 329156 : foreach(lc, innerpath->param_info->ppi_clauses)
5410 : {
5411 171304 : RestrictInfo *rinfo = (RestrictInfo *) lfirst(lc);
5412 :
5413 171304 : if (join_clause_is_movable_into(rinfo,
5414 171304 : innerpath->parent->relids,
5415 : joinrelids))
5416 : {
5417 170752 : if (!is_redundant_with_indexclauses(rinfo, indexclauses))
5418 8830 : return false;
5419 161922 : found_one = true;
5420 : }
5421 : }
5422 157852 : return found_one;
5423 : }
5424 :
5425 :
5426 : /*
5427 : * approx_tuple_count
5428 : * Quick-and-dirty estimation of the number of join rows passing
5429 : * a set of qual conditions.
5430 : *
5431 : * The quals can be either an implicitly-ANDed list of boolean expressions,
5432 : * or a list of RestrictInfo nodes (typically the latter).
5433 : *
5434 : * We intentionally compute the selectivity under JOIN_INNER rules, even
5435 : * if it's some type of outer join. This is appropriate because we are
5436 : * trying to figure out how many tuples pass the initial merge or hash
5437 : * join step.
5438 : *
5439 : * This is quick-and-dirty because we bypass clauselist_selectivity, and
5440 : * simply multiply the independent clause selectivities together. Now
5441 : * clauselist_selectivity often can't do any better than that anyhow, but
5442 : * for some situations (such as range constraints) it is smarter. However,
5443 : * we can't effectively cache the results of clauselist_selectivity, whereas
5444 : * the individual clause selectivities can be and are cached.
5445 : *
5446 : * Since we are only using the results to estimate how many potential
5447 : * output tuples are generated and passed through qpqual checking, it
5448 : * seems OK to live with the approximation.
5449 : */
5450 : static double
5451 811074 : approx_tuple_count(PlannerInfo *root, JoinPath *path, List *quals)
5452 : {
5453 : double tuples;
5454 811074 : double outer_tuples = path->outerjoinpath->rows;
5455 811074 : double inner_tuples = path->innerjoinpath->rows;
5456 : SpecialJoinInfo sjinfo;
5457 811074 : Selectivity selec = 1.0;
5458 : ListCell *l;
5459 :
5460 : /*
5461 : * Make up a SpecialJoinInfo for JOIN_INNER semantics.
5462 : */
5463 811074 : init_dummy_sjinfo(&sjinfo, path->outerjoinpath->parent->relids,
5464 811074 : path->innerjoinpath->parent->relids);
5465 :
5466 : /* Get the approximate selectivity */
5467 1714094 : foreach(l, quals)
5468 : {
5469 903020 : Node *qual = (Node *) lfirst(l);
5470 :
5471 : /* Note that clause_selectivity will be able to cache its result */
5472 903020 : selec *= clause_selectivity(root, qual, 0, JOIN_INNER, &sjinfo);
5473 : }
5474 :
5475 : /* Apply it to the input relation sizes */
5476 811074 : tuples = selec * outer_tuples * inner_tuples;
5477 :
5478 811074 : return clamp_row_est(tuples);
5479 : }
5480 :
5481 :
5482 : /*
5483 : * set_baserel_size_estimates
5484 : * Set the size estimates for the given base relation.
5485 : *
5486 : * The rel's targetlist and restrictinfo list must have been constructed
5487 : * already, and rel->tuples must be set.
5488 : *
5489 : * We set the following fields of the rel node:
5490 : * rows: the estimated number of output tuples (after applying
5491 : * restriction clauses).
5492 : * width: the estimated average output tuple width in bytes.
5493 : * baserestrictcost: estimated cost of evaluating baserestrictinfo clauses.
5494 : */
5495 : void
5496 534534 : set_baserel_size_estimates(PlannerInfo *root, RelOptInfo *rel)
5497 : {
5498 : double nrows;
5499 :
5500 : /* Should only be applied to base relations */
5501 : Assert(rel->relid > 0);
5502 :
5503 1069038 : nrows = rel->tuples *
5504 534534 : clauselist_selectivity(root,
5505 : rel->baserestrictinfo,
5506 : 0,
5507 : JOIN_INNER,
5508 : NULL);
5509 :
5510 534504 : rel->rows = clamp_row_est(nrows);
5511 :
5512 534504 : cost_qual_eval(&rel->baserestrictcost, rel->baserestrictinfo, root);
5513 :
5514 534504 : set_rel_width(root, rel);
5515 534504 : }
5516 :
5517 : /*
5518 : * get_parameterized_baserel_size
5519 : * Make a size estimate for a parameterized scan of a base relation.
5520 : *
5521 : * 'param_clauses' lists the additional join clauses to be used.
5522 : *
5523 : * set_baserel_size_estimates must have been applied already.
5524 : */
5525 : double
5526 165080 : get_parameterized_baserel_size(PlannerInfo *root, RelOptInfo *rel,
5527 : List *param_clauses)
5528 : {
5529 : List *allclauses;
5530 : double nrows;
5531 :
5532 : /*
5533 : * Estimate the number of rows returned by the parameterized scan, knowing
5534 : * that it will apply all the extra join clauses as well as the rel's own
5535 : * restriction clauses. Note that we force the clauses to be treated as
5536 : * non-join clauses during selectivity estimation.
5537 : */
5538 165080 : allclauses = list_concat_copy(param_clauses, rel->baserestrictinfo);
5539 330160 : nrows = rel->tuples *
5540 165080 : clauselist_selectivity(root,
5541 : allclauses,
5542 165080 : rel->relid, /* do not use 0! */
5543 : JOIN_INNER,
5544 : NULL);
5545 165080 : nrows = clamp_row_est(nrows);
5546 : /* For safety, make sure result is not more than the base estimate */
5547 165080 : if (nrows > rel->rows)
5548 0 : nrows = rel->rows;
5549 165080 : return nrows;
5550 : }
5551 :
5552 : /*
5553 : * set_joinrel_size_estimates
5554 : * Set the size estimates for the given join relation.
5555 : *
5556 : * The rel's targetlist must have been constructed already, and a
5557 : * restriction clause list that matches the given component rels must
5558 : * be provided.
5559 : *
5560 : * Since there is more than one way to make a joinrel for more than two
5561 : * base relations, the results we get here could depend on which component
5562 : * rel pair is provided. In theory we should get the same answers no matter
5563 : * which pair is provided; in practice, since the selectivity estimation
5564 : * routines don't handle all cases equally well, we might not. But there's
5565 : * not much to be done about it. (Would it make sense to repeat the
5566 : * calculations for each pair of input rels that's encountered, and somehow
5567 : * average the results? Probably way more trouble than it's worth, and
5568 : * anyway we must keep the rowcount estimate the same for all paths for the
5569 : * joinrel.)
5570 : *
5571 : * We set only the rows field here. The reltarget field was already set by
5572 : * build_joinrel_tlist, and baserestrictcost is not used for join rels.
5573 : */
5574 : void
5575 261768 : set_joinrel_size_estimates(PlannerInfo *root, RelOptInfo *rel,
5576 : RelOptInfo *outer_rel,
5577 : RelOptInfo *inner_rel,
5578 : SpecialJoinInfo *sjinfo,
5579 : List *restrictlist)
5580 : {
5581 261768 : rel->rows = calc_joinrel_size_estimate(root,
5582 : rel,
5583 : outer_rel,
5584 : inner_rel,
5585 : outer_rel->rows,
5586 : inner_rel->rows,
5587 : sjinfo,
5588 : restrictlist);
5589 261768 : }
5590 :
5591 : /*
5592 : * get_parameterized_joinrel_size
5593 : * Make a size estimate for a parameterized scan of a join relation.
5594 : *
5595 : * 'rel' is the joinrel under consideration.
5596 : * 'outer_path', 'inner_path' are (probably also parameterized) Paths that
5597 : * produce the relations being joined.
5598 : * 'sjinfo' is any SpecialJoinInfo relevant to this join.
5599 : * 'restrict_clauses' lists the join clauses that need to be applied at the
5600 : * join node (including any movable clauses that were moved down to this join,
5601 : * and not including any movable clauses that were pushed down into the
5602 : * child paths).
5603 : *
5604 : * set_joinrel_size_estimates must have been applied already.
5605 : */
5606 : double
5607 10238 : get_parameterized_joinrel_size(PlannerInfo *root, RelOptInfo *rel,
5608 : Path *outer_path,
5609 : Path *inner_path,
5610 : SpecialJoinInfo *sjinfo,
5611 : List *restrict_clauses)
5612 : {
5613 : double nrows;
5614 :
5615 : /*
5616 : * Estimate the number of rows returned by the parameterized join as the
5617 : * sizes of the input paths times the selectivity of the clauses that have
5618 : * ended up at this join node.
5619 : *
5620 : * As with set_joinrel_size_estimates, the rowcount estimate could depend
5621 : * on the pair of input paths provided, though ideally we'd get the same
5622 : * estimate for any pair with the same parameterization.
5623 : */
5624 10238 : nrows = calc_joinrel_size_estimate(root,
5625 : rel,
5626 : outer_path->parent,
5627 : inner_path->parent,
5628 : outer_path->rows,
5629 : inner_path->rows,
5630 : sjinfo,
5631 : restrict_clauses);
5632 : /* For safety, make sure result is not more than the base estimate */
5633 10238 : if (nrows > rel->rows)
5634 12 : nrows = rel->rows;
5635 10238 : return nrows;
5636 : }
5637 :
5638 : /*
5639 : * calc_joinrel_size_estimate
5640 : * Workhorse for set_joinrel_size_estimates and
5641 : * get_parameterized_joinrel_size.
5642 : *
5643 : * outer_rel/inner_rel are the relations being joined, but they should be
5644 : * assumed to have sizes outer_rows/inner_rows; those numbers might be less
5645 : * than what rel->rows says, when we are considering parameterized paths.
5646 : */
5647 : static double
5648 272006 : calc_joinrel_size_estimate(PlannerInfo *root,
5649 : RelOptInfo *joinrel,
5650 : RelOptInfo *outer_rel,
5651 : RelOptInfo *inner_rel,
5652 : double outer_rows,
5653 : double inner_rows,
5654 : SpecialJoinInfo *sjinfo,
5655 : List *restrictlist)
5656 : {
5657 272006 : JoinType jointype = sjinfo->jointype;
5658 : Selectivity fkselec;
5659 : Selectivity jselec;
5660 : Selectivity pselec;
5661 : double nrows;
5662 :
5663 : /*
5664 : * Compute joinclause selectivity. Note that we are only considering
5665 : * clauses that become restriction clauses at this join level; we are not
5666 : * double-counting them because they were not considered in estimating the
5667 : * sizes of the component rels.
5668 : *
5669 : * First, see whether any of the joinclauses can be matched to known FK
5670 : * constraints. If so, drop those clauses from the restrictlist, and
5671 : * instead estimate their selectivity using FK semantics. (We do this
5672 : * without regard to whether said clauses are local or "pushed down".
5673 : * Probably, an FK-matching clause could never be seen as pushed down at
5674 : * an outer join, since it would be strict and hence would be grounds for
5675 : * join strength reduction.) fkselec gets the net selectivity for
5676 : * FK-matching clauses, or 1.0 if there are none.
5677 : */
5678 272006 : fkselec = get_foreign_key_join_selectivity(root,
5679 : outer_rel->relids,
5680 : inner_rel->relids,
5681 : sjinfo,
5682 : &restrictlist);
5683 :
5684 : /*
5685 : * For an outer join, we have to distinguish the selectivity of the join's
5686 : * own clauses (JOIN/ON conditions) from any clauses that were "pushed
5687 : * down". For inner joins we just count them all as joinclauses.
5688 : */
5689 272006 : if (IS_OUTER_JOIN(jointype))
5690 : {
5691 82804 : List *joinquals = NIL;
5692 82804 : List *pushedquals = NIL;
5693 : ListCell *l;
5694 :
5695 : /* Grovel through the clauses to separate into two lists */
5696 186182 : foreach(l, restrictlist)
5697 : {
5698 103378 : RestrictInfo *rinfo = lfirst_node(RestrictInfo, l);
5699 :
5700 103378 : if (RINFO_IS_PUSHED_DOWN(rinfo, joinrel->relids))
5701 4308 : pushedquals = lappend(pushedquals, rinfo);
5702 : else
5703 99070 : joinquals = lappend(joinquals, rinfo);
5704 : }
5705 :
5706 : /* Get the separate selectivities */
5707 82804 : jselec = clauselist_selectivity(root,
5708 : joinquals,
5709 : 0,
5710 : jointype,
5711 : sjinfo);
5712 82804 : pselec = clauselist_selectivity(root,
5713 : pushedquals,
5714 : 0,
5715 : jointype,
5716 : sjinfo);
5717 :
5718 : /* Avoid leaking a lot of ListCells */
5719 82804 : list_free(joinquals);
5720 82804 : list_free(pushedquals);
5721 : }
5722 : else
5723 : {
5724 189202 : jselec = clauselist_selectivity(root,
5725 : restrictlist,
5726 : 0,
5727 : jointype,
5728 : sjinfo);
5729 189202 : pselec = 0.0; /* not used, keep compiler quiet */
5730 : }
5731 :
5732 : /*
5733 : * Basically, we multiply size of Cartesian product by selectivity.
5734 : *
5735 : * If we are doing an outer join, take that into account: the joinqual
5736 : * selectivity has to be clamped using the knowledge that the output must
5737 : * be at least as large as the non-nullable input. However, any
5738 : * pushed-down quals are applied after the outer join, so their
5739 : * selectivity applies fully.
5740 : *
5741 : * For JOIN_SEMI and JOIN_ANTI, the selectivity is defined as the fraction
5742 : * of LHS rows that have matches, and we apply that straightforwardly.
5743 : */
5744 272006 : switch (jointype)
5745 : {
5746 180966 : case JOIN_INNER:
5747 180966 : nrows = outer_rows * inner_rows * fkselec * jselec;
5748 : /* pselec not used */
5749 180966 : break;
5750 76008 : case JOIN_LEFT:
5751 76008 : nrows = outer_rows * inner_rows * fkselec * jselec;
5752 76008 : if (nrows < outer_rows)
5753 31198 : nrows = outer_rows;
5754 76008 : nrows *= pselec;
5755 76008 : break;
5756 1720 : case JOIN_FULL:
5757 1720 : nrows = outer_rows * inner_rows * fkselec * jselec;
5758 1720 : if (nrows < outer_rows)
5759 1178 : nrows = outer_rows;
5760 1720 : if (nrows < inner_rows)
5761 120 : nrows = inner_rows;
5762 1720 : nrows *= pselec;
5763 1720 : break;
5764 8236 : case JOIN_SEMI:
5765 8236 : nrows = outer_rows * fkselec * jselec;
5766 : /* pselec not used */
5767 8236 : break;
5768 5076 : case JOIN_ANTI:
5769 5076 : nrows = outer_rows * (1.0 - fkselec * jselec);
5770 5076 : nrows *= pselec;
5771 5076 : break;
5772 0 : default:
5773 : /* other values not expected here */
5774 0 : elog(ERROR, "unrecognized join type: %d", (int) jointype);
5775 : nrows = 0; /* keep compiler quiet */
5776 : break;
5777 : }
5778 :
5779 272006 : return clamp_row_est(nrows);
5780 : }
5781 :
5782 : /*
5783 : * get_foreign_key_join_selectivity
5784 : * Estimate join selectivity for foreign-key-related clauses.
5785 : *
5786 : * Remove any clauses that can be matched to FK constraints from *restrictlist,
5787 : * and return a substitute estimate of their selectivity. 1.0 is returned
5788 : * when there are no such clauses.
5789 : *
5790 : * The reason for treating such clauses specially is that we can get better
5791 : * estimates this way than by relying on clauselist_selectivity(), especially
5792 : * for multi-column FKs where that function's assumption that the clauses are
5793 : * independent falls down badly. But even with single-column FKs, we may be
5794 : * able to get a better answer when the pg_statistic stats are missing or out
5795 : * of date.
5796 : */
5797 : static Selectivity
5798 272006 : get_foreign_key_join_selectivity(PlannerInfo *root,
5799 : Relids outer_relids,
5800 : Relids inner_relids,
5801 : SpecialJoinInfo *sjinfo,
5802 : List **restrictlist)
5803 : {
5804 272006 : Selectivity fkselec = 1.0;
5805 272006 : JoinType jointype = sjinfo->jointype;
5806 272006 : List *worklist = *restrictlist;
5807 : ListCell *lc;
5808 :
5809 : /* Consider each FK constraint that is known to match the query */
5810 273972 : foreach(lc, root->fkey_list)
5811 : {
5812 1966 : ForeignKeyOptInfo *fkinfo = (ForeignKeyOptInfo *) lfirst(lc);
5813 : bool ref_is_outer;
5814 : List *removedlist;
5815 : ListCell *cell;
5816 :
5817 : /*
5818 : * This FK is not relevant unless it connects a baserel on one side of
5819 : * this join to a baserel on the other side.
5820 : */
5821 3580 : if (bms_is_member(fkinfo->con_relid, outer_relids) &&
5822 1614 : bms_is_member(fkinfo->ref_relid, inner_relids))
5823 1440 : ref_is_outer = false;
5824 866 : else if (bms_is_member(fkinfo->ref_relid, outer_relids) &&
5825 340 : bms_is_member(fkinfo->con_relid, inner_relids))
5826 130 : ref_is_outer = true;
5827 : else
5828 396 : continue;
5829 :
5830 : /*
5831 : * If we're dealing with a semi/anti join, and the FK's referenced
5832 : * relation is on the outside, then knowledge of the FK doesn't help
5833 : * us figure out what we need to know (which is the fraction of outer
5834 : * rows that have matches). On the other hand, if the referenced rel
5835 : * is on the inside, then all outer rows must have matches in the
5836 : * referenced table (ignoring nulls). But any restriction or join
5837 : * clauses that filter that table will reduce the fraction of matches.
5838 : * We can account for restriction clauses, but it's too hard to guess
5839 : * how many table rows would get through a join that's inside the RHS.
5840 : * Hence, if either case applies, punt and ignore the FK.
5841 : */
5842 1570 : if ((jointype == JOIN_SEMI || jointype == JOIN_ANTI) &&
5843 1048 : (ref_is_outer || bms_membership(inner_relids) != BMS_SINGLETON))
5844 12 : continue;
5845 :
5846 : /*
5847 : * Modify the restrictlist by removing clauses that match the FK (and
5848 : * putting them into removedlist instead). It seems unsafe to modify
5849 : * the originally-passed List structure, so we make a shallow copy the
5850 : * first time through.
5851 : */
5852 1558 : if (worklist == *restrictlist)
5853 1334 : worklist = list_copy(worklist);
5854 :
5855 1558 : removedlist = NIL;
5856 3252 : foreach(cell, worklist)
5857 : {
5858 1694 : RestrictInfo *rinfo = (RestrictInfo *) lfirst(cell);
5859 1694 : bool remove_it = false;
5860 : int i;
5861 :
5862 : /* Drop this clause if it matches any column of the FK */
5863 2140 : for (i = 0; i < fkinfo->nkeys; i++)
5864 : {
5865 2110 : if (rinfo->parent_ec)
5866 : {
5867 : /*
5868 : * EC-derived clauses can only match by EC. It is okay to
5869 : * consider any clause derived from the same EC as
5870 : * matching the FK: even if equivclass.c chose to generate
5871 : * a clause equating some other pair of Vars, it could
5872 : * have generated one equating the FK's Vars. So for
5873 : * purposes of estimation, we can act as though it did so.
5874 : *
5875 : * Note: checking parent_ec is a bit of a cheat because
5876 : * there are EC-derived clauses that don't have parent_ec
5877 : * set; but such clauses must compare expressions that
5878 : * aren't just Vars, so they cannot match the FK anyway.
5879 : */
5880 304 : if (fkinfo->eclass[i] == rinfo->parent_ec)
5881 : {
5882 298 : remove_it = true;
5883 298 : break;
5884 : }
5885 : }
5886 : else
5887 : {
5888 : /*
5889 : * Otherwise, see if rinfo was previously matched to FK as
5890 : * a "loose" clause.
5891 : */
5892 1806 : if (list_member_ptr(fkinfo->rinfos[i], rinfo))
5893 : {
5894 1366 : remove_it = true;
5895 1366 : break;
5896 : }
5897 : }
5898 : }
5899 1694 : if (remove_it)
5900 : {
5901 1664 : worklist = foreach_delete_current(worklist, cell);
5902 1664 : removedlist = lappend(removedlist, rinfo);
5903 : }
5904 : }
5905 :
5906 : /*
5907 : * If we failed to remove all the matching clauses we expected to
5908 : * find, chicken out and ignore this FK; applying its selectivity
5909 : * might result in double-counting. Put any clauses we did manage to
5910 : * remove back into the worklist.
5911 : *
5912 : * Since the matching clauses are known not outerjoin-delayed, they
5913 : * would normally have appeared in the initial joinclause list. If we
5914 : * didn't find them, there are two possibilities:
5915 : *
5916 : * 1. If the FK match is based on an EC that is ec_has_const, it won't
5917 : * have generated any join clauses at all. We discount such ECs while
5918 : * checking to see if we have "all" the clauses. (Below, we'll adjust
5919 : * the selectivity estimate for this case.)
5920 : *
5921 : * 2. The clauses were matched to some other FK in a previous
5922 : * iteration of this loop, and thus removed from worklist. (A likely
5923 : * case is that two FKs are matched to the same EC; there will be only
5924 : * one EC-derived clause in the initial list, so the first FK will
5925 : * consume it.) Applying both FKs' selectivity independently risks
5926 : * underestimating the join size; in particular, this would undo one
5927 : * of the main things that ECs were invented for, namely to avoid
5928 : * double-counting the selectivity of redundant equality conditions.
5929 : * Later we might think of a reasonable way to combine the estimates,
5930 : * but for now, just punt, since this is a fairly uncommon situation.
5931 : */
5932 1558 : if (removedlist == NIL ||
5933 1272 : list_length(removedlist) !=
5934 1272 : (fkinfo->nmatched_ec - fkinfo->nconst_ec + fkinfo->nmatched_ri))
5935 : {
5936 286 : worklist = list_concat(worklist, removedlist);
5937 286 : continue;
5938 : }
5939 :
5940 : /*
5941 : * Finally we get to the payoff: estimate selectivity using the
5942 : * knowledge that each referencing row will match exactly one row in
5943 : * the referenced table.
5944 : *
5945 : * XXX that's not true in the presence of nulls in the referencing
5946 : * column(s), so in principle we should derate the estimate for those.
5947 : * However (1) if there are any strict restriction clauses for the
5948 : * referencing column(s) elsewhere in the query, derating here would
5949 : * be double-counting the null fraction, and (2) it's not very clear
5950 : * how to combine null fractions for multiple referencing columns. So
5951 : * we do nothing for now about correcting for nulls.
5952 : *
5953 : * XXX another point here is that if either side of an FK constraint
5954 : * is an inheritance parent, we estimate as though the constraint
5955 : * covers all its children as well. This is not an unreasonable
5956 : * assumption for a referencing table, ie the user probably applied
5957 : * identical constraints to all child tables (though perhaps we ought
5958 : * to check that). But it's not possible to have done that for a
5959 : * referenced table. Fortunately, precisely because that doesn't
5960 : * work, it is uncommon in practice to have an FK referencing a parent
5961 : * table. So, at least for now, disregard inheritance here.
5962 : */
5963 1272 : if (jointype == JOIN_SEMI || jointype == JOIN_ANTI)
5964 824 : {
5965 : /*
5966 : * For JOIN_SEMI and JOIN_ANTI, we only get here when the FK's
5967 : * referenced table is exactly the inside of the join. The join
5968 : * selectivity is defined as the fraction of LHS rows that have
5969 : * matches. The FK implies that every LHS row has a match *in the
5970 : * referenced table*; but any restriction clauses on it will
5971 : * reduce the number of matches. Hence we take the join
5972 : * selectivity as equal to the selectivity of the table's
5973 : * restriction clauses, which is rows / tuples; but we must guard
5974 : * against tuples == 0.
5975 : */
5976 824 : RelOptInfo *ref_rel = find_base_rel(root, fkinfo->ref_relid);
5977 824 : double ref_tuples = Max(ref_rel->tuples, 1.0);
5978 :
5979 824 : fkselec *= ref_rel->rows / ref_tuples;
5980 : }
5981 : else
5982 : {
5983 : /*
5984 : * Otherwise, selectivity is exactly 1/referenced-table-size; but
5985 : * guard against tuples == 0. Note we should use the raw table
5986 : * tuple count, not any estimate of its filtered or joined size.
5987 : */
5988 448 : RelOptInfo *ref_rel = find_base_rel(root, fkinfo->ref_relid);
5989 448 : double ref_tuples = Max(ref_rel->tuples, 1.0);
5990 :
5991 448 : fkselec *= 1.0 / ref_tuples;
5992 : }
5993 :
5994 : /*
5995 : * If any of the FK columns participated in ec_has_const ECs, then
5996 : * equivclass.c will have generated "var = const" restrictions for
5997 : * each side of the join, thus reducing the sizes of both input
5998 : * relations. Taking the fkselec at face value would amount to
5999 : * double-counting the selectivity of the constant restriction for the
6000 : * referencing Var. Hence, look for the restriction clause(s) that
6001 : * were applied to the referencing Var(s), and divide out their
6002 : * selectivity to correct for this.
6003 : */
6004 1272 : if (fkinfo->nconst_ec > 0)
6005 : {
6006 24 : for (int i = 0; i < fkinfo->nkeys; i++)
6007 : {
6008 18 : EquivalenceClass *ec = fkinfo->eclass[i];
6009 :
6010 18 : if (ec && ec->ec_has_const)
6011 : {
6012 6 : EquivalenceMember *em = fkinfo->fk_eclass_member[i];
6013 6 : RestrictInfo *rinfo = find_derived_clause_for_ec_member(root,
6014 : ec,
6015 : em);
6016 :
6017 6 : if (rinfo)
6018 : {
6019 : Selectivity s0;
6020 :
6021 6 : s0 = clause_selectivity(root,
6022 : (Node *) rinfo,
6023 : 0,
6024 : jointype,
6025 : sjinfo);
6026 6 : if (s0 > 0)
6027 6 : fkselec /= s0;
6028 : }
6029 : }
6030 : }
6031 : }
6032 : }
6033 :
6034 272006 : *restrictlist = worklist;
6035 272006 : CLAMP_PROBABILITY(fkselec);
6036 272006 : return fkselec;
6037 : }
6038 :
6039 : /*
6040 : * set_subquery_size_estimates
6041 : * Set the size estimates for a base relation that is a subquery.
6042 : *
6043 : * The rel's targetlist and restrictinfo list must have been constructed
6044 : * already, and the Paths for the subquery must have been completed.
6045 : * We look at the subquery's PlannerInfo to extract data.
6046 : *
6047 : * We set the same fields as set_baserel_size_estimates.
6048 : */
6049 : void
6050 40298 : set_subquery_size_estimates(PlannerInfo *root, RelOptInfo *rel)
6051 : {
6052 40298 : PlannerInfo *subroot = rel->subroot;
6053 : RelOptInfo *sub_final_rel;
6054 : ListCell *lc;
6055 :
6056 : /* Should only be applied to base relations that are subqueries */
6057 : Assert(rel->relid > 0);
6058 : Assert(planner_rt_fetch(rel->relid, root)->rtekind == RTE_SUBQUERY);
6059 :
6060 : /*
6061 : * Copy raw number of output rows from subquery. All of its paths should
6062 : * have the same output rowcount, so just look at cheapest-total.
6063 : */
6064 40298 : sub_final_rel = fetch_upper_rel(subroot, UPPERREL_FINAL, NULL);
6065 40298 : rel->tuples = sub_final_rel->cheapest_total_path->rows;
6066 :
6067 : /*
6068 : * Compute per-output-column width estimates by examining the subquery's
6069 : * targetlist. For any output that is a plain Var, get the width estimate
6070 : * that was made while planning the subquery. Otherwise, we leave it to
6071 : * set_rel_width to fill in a datatype-based default estimate.
6072 : */
6073 194592 : foreach(lc, subroot->parse->targetList)
6074 : {
6075 154294 : TargetEntry *te = lfirst_node(TargetEntry, lc);
6076 154294 : Node *texpr = (Node *) te->expr;
6077 154294 : int32 item_width = 0;
6078 :
6079 : /* junk columns aren't visible to upper query */
6080 154294 : if (te->resjunk)
6081 6902 : continue;
6082 :
6083 : /*
6084 : * The subquery could be an expansion of a view that's had columns
6085 : * added to it since the current query was parsed, so that there are
6086 : * non-junk tlist columns in it that don't correspond to any column
6087 : * visible at our query level. Ignore such columns.
6088 : */
6089 147392 : if (te->resno < rel->min_attr || te->resno > rel->max_attr)
6090 0 : continue;
6091 :
6092 : /*
6093 : * XXX This currently doesn't work for subqueries containing set
6094 : * operations, because the Vars in their tlists are bogus references
6095 : * to the first leaf subquery, which wouldn't give the right answer
6096 : * even if we could still get to its PlannerInfo.
6097 : *
6098 : * Also, the subquery could be an appendrel for which all branches are
6099 : * known empty due to constraint exclusion, in which case
6100 : * set_append_rel_pathlist will have left the attr_widths set to zero.
6101 : *
6102 : * In either case, we just leave the width estimate zero until
6103 : * set_rel_width fixes it.
6104 : */
6105 147392 : if (IsA(texpr, Var) &&
6106 64784 : subroot->parse->setOperations == NULL)
6107 : {
6108 62930 : Var *var = (Var *) texpr;
6109 62930 : RelOptInfo *subrel = find_base_rel(subroot, var->varno);
6110 :
6111 62930 : item_width = subrel->attr_widths[var->varattno - subrel->min_attr];
6112 : }
6113 147392 : rel->attr_widths[te->resno - rel->min_attr] = item_width;
6114 : }
6115 :
6116 : /* Now estimate number of output rows, etc */
6117 40298 : set_baserel_size_estimates(root, rel);
6118 40298 : }
6119 :
6120 : /*
6121 : * set_function_size_estimates
6122 : * Set the size estimates for a base relation that is a function call.
6123 : *
6124 : * The rel's targetlist and restrictinfo list must have been constructed
6125 : * already.
6126 : *
6127 : * We set the same fields as set_baserel_size_estimates.
6128 : */
6129 : void
6130 53556 : set_function_size_estimates(PlannerInfo *root, RelOptInfo *rel)
6131 : {
6132 : RangeTblEntry *rte;
6133 : ListCell *lc;
6134 :
6135 : /* Should only be applied to base relations that are functions */
6136 : Assert(rel->relid > 0);
6137 53556 : rte = planner_rt_fetch(rel->relid, root);
6138 : Assert(rte->rtekind == RTE_FUNCTION);
6139 :
6140 : /*
6141 : * Estimate number of rows the functions will return. The rowcount of the
6142 : * node is that of the largest function result.
6143 : */
6144 53556 : rel->tuples = 0;
6145 107616 : foreach(lc, rte->functions)
6146 : {
6147 54060 : RangeTblFunction *rtfunc = (RangeTblFunction *) lfirst(lc);
6148 54060 : double ntup = expression_returns_set_rows(root, rtfunc->funcexpr);
6149 :
6150 54060 : if (ntup > rel->tuples)
6151 53580 : rel->tuples = ntup;
6152 : }
6153 :
6154 : /* Now estimate number of output rows, etc */
6155 53556 : set_baserel_size_estimates(root, rel);
6156 53556 : }
6157 :
6158 : /*
6159 : * set_function_size_estimates
6160 : * Set the size estimates for a base relation that is a function call.
6161 : *
6162 : * The rel's targetlist and restrictinfo list must have been constructed
6163 : * already.
6164 : *
6165 : * We set the same fields as set_tablefunc_size_estimates.
6166 : */
6167 : void
6168 626 : set_tablefunc_size_estimates(PlannerInfo *root, RelOptInfo *rel)
6169 : {
6170 : /* Should only be applied to base relations that are functions */
6171 : Assert(rel->relid > 0);
6172 : Assert(planner_rt_fetch(rel->relid, root)->rtekind == RTE_TABLEFUNC);
6173 :
6174 626 : rel->tuples = 100;
6175 :
6176 : /* Now estimate number of output rows, etc */
6177 626 : set_baserel_size_estimates(root, rel);
6178 626 : }
6179 :
6180 : /*
6181 : * set_values_size_estimates
6182 : * Set the size estimates for a base relation that is a values list.
6183 : *
6184 : * The rel's targetlist and restrictinfo list must have been constructed
6185 : * already.
6186 : *
6187 : * We set the same fields as set_baserel_size_estimates.
6188 : */
6189 : void
6190 8520 : set_values_size_estimates(PlannerInfo *root, RelOptInfo *rel)
6191 : {
6192 : RangeTblEntry *rte;
6193 :
6194 : /* Should only be applied to base relations that are values lists */
6195 : Assert(rel->relid > 0);
6196 8520 : rte = planner_rt_fetch(rel->relid, root);
6197 : Assert(rte->rtekind == RTE_VALUES);
6198 :
6199 : /*
6200 : * Estimate number of rows the values list will return. We know this
6201 : * precisely based on the list length (well, barring set-returning
6202 : * functions in list items, but that's a refinement not catered for
6203 : * anywhere else either).
6204 : */
6205 8520 : rel->tuples = list_length(rte->values_lists);
6206 :
6207 : /* Now estimate number of output rows, etc */
6208 8520 : set_baserel_size_estimates(root, rel);
6209 8520 : }
6210 :
6211 : /*
6212 : * set_cte_size_estimates
6213 : * Set the size estimates for a base relation that is a CTE reference.
6214 : *
6215 : * The rel's targetlist and restrictinfo list must have been constructed
6216 : * already, and we need an estimate of the number of rows returned by the CTE
6217 : * (if a regular CTE) or the non-recursive term (if a self-reference).
6218 : *
6219 : * We set the same fields as set_baserel_size_estimates.
6220 : */
6221 : void
6222 5312 : set_cte_size_estimates(PlannerInfo *root, RelOptInfo *rel, double cte_rows)
6223 : {
6224 : RangeTblEntry *rte;
6225 :
6226 : /* Should only be applied to base relations that are CTE references */
6227 : Assert(rel->relid > 0);
6228 5312 : rte = planner_rt_fetch(rel->relid, root);
6229 : Assert(rte->rtekind == RTE_CTE);
6230 :
6231 5312 : if (rte->self_reference)
6232 : {
6233 : /*
6234 : * In a self-reference, we assume the average worktable size is a
6235 : * multiple of the nonrecursive term's size. The best multiplier will
6236 : * vary depending on query "fan-out", so make its value adjustable.
6237 : */
6238 942 : rel->tuples = clamp_row_est(recursive_worktable_factor * cte_rows);
6239 : }
6240 : else
6241 : {
6242 : /* Otherwise just believe the CTE's rowcount estimate */
6243 4370 : rel->tuples = cte_rows;
6244 : }
6245 :
6246 : /* Now estimate number of output rows, etc */
6247 5312 : set_baserel_size_estimates(root, rel);
6248 5312 : }
6249 :
6250 : /*
6251 : * set_namedtuplestore_size_estimates
6252 : * Set the size estimates for a base relation that is a tuplestore reference.
6253 : *
6254 : * The rel's targetlist and restrictinfo list must have been constructed
6255 : * already.
6256 : *
6257 : * We set the same fields as set_baserel_size_estimates.
6258 : */
6259 : void
6260 474 : set_namedtuplestore_size_estimates(PlannerInfo *root, RelOptInfo *rel)
6261 : {
6262 : RangeTblEntry *rte;
6263 :
6264 : /* Should only be applied to base relations that are tuplestore references */
6265 : Assert(rel->relid > 0);
6266 474 : rte = planner_rt_fetch(rel->relid, root);
6267 : Assert(rte->rtekind == RTE_NAMEDTUPLESTORE);
6268 :
6269 : /*
6270 : * Use the estimate provided by the code which is generating the named
6271 : * tuplestore. In some cases, the actual number might be available; in
6272 : * others the same plan will be re-used, so a "typical" value might be
6273 : * estimated and used.
6274 : */
6275 474 : rel->tuples = rte->enrtuples;
6276 474 : if (rel->tuples < 0)
6277 0 : rel->tuples = 1000;
6278 :
6279 : /* Now estimate number of output rows, etc */
6280 474 : set_baserel_size_estimates(root, rel);
6281 474 : }
6282 :
6283 : /*
6284 : * set_result_size_estimates
6285 : * Set the size estimates for an RTE_RESULT base relation
6286 : *
6287 : * The rel's targetlist and restrictinfo list must have been constructed
6288 : * already.
6289 : *
6290 : * We set the same fields as set_baserel_size_estimates.
6291 : */
6292 : void
6293 4292 : set_result_size_estimates(PlannerInfo *root, RelOptInfo *rel)
6294 : {
6295 : /* Should only be applied to RTE_RESULT base relations */
6296 : Assert(rel->relid > 0);
6297 : Assert(planner_rt_fetch(rel->relid, root)->rtekind == RTE_RESULT);
6298 :
6299 : /* RTE_RESULT always generates a single row, natively */
6300 4292 : rel->tuples = 1;
6301 :
6302 : /* Now estimate number of output rows, etc */
6303 4292 : set_baserel_size_estimates(root, rel);
6304 4292 : }
6305 :
6306 : /*
6307 : * set_foreign_size_estimates
6308 : * Set the size estimates for a base relation that is a foreign table.
6309 : *
6310 : * There is not a whole lot that we can do here; the foreign-data wrapper
6311 : * is responsible for producing useful estimates. We can do a decent job
6312 : * of estimating baserestrictcost, so we set that, and we also set up width
6313 : * using what will be purely datatype-driven estimates from the targetlist.
6314 : * There is no way to do anything sane with the rows value, so we just put
6315 : * a default estimate and hope that the wrapper can improve on it. The
6316 : * wrapper's GetForeignRelSize function will be called momentarily.
6317 : *
6318 : * The rel's targetlist and restrictinfo list must have been constructed
6319 : * already.
6320 : */
6321 : void
6322 2468 : set_foreign_size_estimates(PlannerInfo *root, RelOptInfo *rel)
6323 : {
6324 : /* Should only be applied to base relations */
6325 : Assert(rel->relid > 0);
6326 :
6327 2468 : rel->rows = 1000; /* entirely bogus default estimate */
6328 :
6329 2468 : cost_qual_eval(&rel->baserestrictcost, rel->baserestrictinfo, root);
6330 :
6331 2468 : set_rel_width(root, rel);
6332 2468 : }
6333 :
6334 :
6335 : /*
6336 : * set_rel_width
6337 : * Set the estimated output width of a base relation.
6338 : *
6339 : * The estimated output width is the sum of the per-attribute width estimates
6340 : * for the actually-referenced columns, plus any PHVs or other expressions
6341 : * that have to be calculated at this relation. This is the amount of data
6342 : * we'd need to pass upwards in case of a sort, hash, etc.
6343 : *
6344 : * This function also sets reltarget->cost, so it's a bit misnamed now.
6345 : *
6346 : * NB: this works best on plain relations because it prefers to look at
6347 : * real Vars. For subqueries, set_subquery_size_estimates will already have
6348 : * copied up whatever per-column estimates were made within the subquery,
6349 : * and for other types of rels there isn't much we can do anyway. We fall
6350 : * back on (fairly stupid) datatype-based width estimates if we can't get
6351 : * any better number.
6352 : *
6353 : * The per-attribute width estimates are cached for possible re-use while
6354 : * building join relations or post-scan/join pathtargets.
6355 : */
6356 : static void
6357 536972 : set_rel_width(PlannerInfo *root, RelOptInfo *rel)
6358 : {
6359 536972 : Oid reloid = planner_rt_fetch(rel->relid, root)->relid;
6360 536972 : int64 tuple_width = 0;
6361 536972 : bool have_wholerow_var = false;
6362 : ListCell *lc;
6363 :
6364 : /* Vars are assumed to have cost zero, but other exprs do not */
6365 536972 : rel->reltarget->cost.startup = 0;
6366 536972 : rel->reltarget->cost.per_tuple = 0;
6367 :
6368 1913060 : foreach(lc, rel->reltarget->exprs)
6369 : {
6370 1376088 : Node *node = (Node *) lfirst(lc);
6371 :
6372 : /*
6373 : * Ordinarily, a Var in a rel's targetlist must belong to that rel;
6374 : * but there are corner cases involving LATERAL references where that
6375 : * isn't so. If the Var has the wrong varno, fall through to the
6376 : * generic case (it doesn't seem worth the trouble to be any smarter).
6377 : */
6378 1376088 : if (IsA(node, Var) &&
6379 1351610 : ((Var *) node)->varno == rel->relid)
6380 369018 : {
6381 1351544 : Var *var = (Var *) node;
6382 : int ndx;
6383 : int32 item_width;
6384 :
6385 : Assert(var->varattno >= rel->min_attr);
6386 : Assert(var->varattno <= rel->max_attr);
6387 :
6388 1351544 : ndx = var->varattno - rel->min_attr;
6389 :
6390 : /*
6391 : * If it's a whole-row Var, we'll deal with it below after we have
6392 : * already cached as many attr widths as possible.
6393 : */
6394 1351544 : if (var->varattno == 0)
6395 : {
6396 3062 : have_wholerow_var = true;
6397 3062 : continue;
6398 : }
6399 :
6400 : /*
6401 : * The width may have been cached already (especially if it's a
6402 : * subquery), so don't duplicate effort.
6403 : */
6404 1348482 : if (rel->attr_widths[ndx] > 0)
6405 : {
6406 266026 : tuple_width += rel->attr_widths[ndx];
6407 266026 : continue;
6408 : }
6409 :
6410 : /* Try to get column width from statistics */
6411 1082456 : if (reloid != InvalidOid && var->varattno > 0)
6412 : {
6413 848266 : item_width = get_attavgwidth(reloid, var->varattno);
6414 848266 : if (item_width > 0)
6415 : {
6416 713438 : rel->attr_widths[ndx] = item_width;
6417 713438 : tuple_width += item_width;
6418 713438 : continue;
6419 : }
6420 : }
6421 :
6422 : /*
6423 : * Not a plain relation, or can't find statistics for it. Estimate
6424 : * using just the type info.
6425 : */
6426 369018 : item_width = get_typavgwidth(var->vartype, var->vartypmod);
6427 : Assert(item_width > 0);
6428 369018 : rel->attr_widths[ndx] = item_width;
6429 369018 : tuple_width += item_width;
6430 : }
6431 24544 : else if (IsA(node, PlaceHolderVar))
6432 : {
6433 : /*
6434 : * We will need to evaluate the PHV's contained expression while
6435 : * scanning this rel, so be sure to include it in reltarget->cost.
6436 : */
6437 2134 : PlaceHolderVar *phv = (PlaceHolderVar *) node;
6438 2134 : PlaceHolderInfo *phinfo = find_placeholder_info(root, phv);
6439 : QualCost cost;
6440 :
6441 2134 : tuple_width += phinfo->ph_width;
6442 2134 : cost_qual_eval_node(&cost, (Node *) phv->phexpr, root);
6443 2134 : rel->reltarget->cost.startup += cost.startup;
6444 2134 : rel->reltarget->cost.per_tuple += cost.per_tuple;
6445 : }
6446 : else
6447 : {
6448 : /*
6449 : * We could be looking at an expression pulled up from a subquery,
6450 : * or a ROW() representing a whole-row child Var, etc. Do what we
6451 : * can using the expression type information.
6452 : */
6453 : int32 item_width;
6454 : QualCost cost;
6455 :
6456 22410 : item_width = get_typavgwidth(exprType(node), exprTypmod(node));
6457 : Assert(item_width > 0);
6458 22410 : tuple_width += item_width;
6459 : /* Not entirely clear if we need to account for cost, but do so */
6460 22410 : cost_qual_eval_node(&cost, node, root);
6461 22410 : rel->reltarget->cost.startup += cost.startup;
6462 22410 : rel->reltarget->cost.per_tuple += cost.per_tuple;
6463 : }
6464 : }
6465 :
6466 : /*
6467 : * If we have a whole-row reference, estimate its width as the sum of
6468 : * per-column widths plus heap tuple header overhead.
6469 : */
6470 536972 : if (have_wholerow_var)
6471 : {
6472 3062 : int64 wholerow_width = MAXALIGN(SizeofHeapTupleHeader);
6473 :
6474 3062 : if (reloid != InvalidOid)
6475 : {
6476 : /* Real relation, so estimate true tuple width */
6477 2392 : wholerow_width += get_relation_data_width(reloid,
6478 2392 : rel->attr_widths - rel->min_attr);
6479 : }
6480 : else
6481 : {
6482 : /* Do what we can with info for a phony rel */
6483 : AttrNumber i;
6484 :
6485 1820 : for (i = 1; i <= rel->max_attr; i++)
6486 1150 : wholerow_width += rel->attr_widths[i - rel->min_attr];
6487 : }
6488 :
6489 3062 : rel->attr_widths[0 - rel->min_attr] = clamp_width_est(wholerow_width);
6490 :
6491 : /*
6492 : * Include the whole-row Var as part of the output tuple. Yes, that
6493 : * really is what happens at runtime.
6494 : */
6495 3062 : tuple_width += wholerow_width;
6496 : }
6497 :
6498 536972 : rel->reltarget->width = clamp_width_est(tuple_width);
6499 536972 : }
6500 :
6501 : /*
6502 : * set_pathtarget_cost_width
6503 : * Set the estimated eval cost and output width of a PathTarget tlist.
6504 : *
6505 : * As a notational convenience, returns the same PathTarget pointer passed in.
6506 : *
6507 : * Most, though not quite all, uses of this function occur after we've run
6508 : * set_rel_width() for base relations; so we can usually obtain cached width
6509 : * estimates for Vars. If we can't, fall back on datatype-based width
6510 : * estimates. Present early-planning uses of PathTargets don't need accurate
6511 : * widths badly enough to justify going to the catalogs for better data.
6512 : */
6513 : PathTarget *
6514 641208 : set_pathtarget_cost_width(PlannerInfo *root, PathTarget *target)
6515 : {
6516 641208 : int64 tuple_width = 0;
6517 : ListCell *lc;
6518 :
6519 : /* Vars are assumed to have cost zero, but other exprs do not */
6520 641208 : target->cost.startup = 0;
6521 641208 : target->cost.per_tuple = 0;
6522 :
6523 2217614 : foreach(lc, target->exprs)
6524 : {
6525 1576406 : Node *node = (Node *) lfirst(lc);
6526 :
6527 1576406 : tuple_width += get_expr_width(root, node);
6528 :
6529 : /* For non-Vars, account for evaluation cost */
6530 1576406 : if (!IsA(node, Var))
6531 : {
6532 : QualCost cost;
6533 :
6534 681540 : cost_qual_eval_node(&cost, node, root);
6535 681540 : target->cost.startup += cost.startup;
6536 681540 : target->cost.per_tuple += cost.per_tuple;
6537 : }
6538 : }
6539 :
6540 641208 : target->width = clamp_width_est(tuple_width);
6541 :
6542 641208 : return target;
6543 : }
6544 :
6545 : /*
6546 : * get_expr_width
6547 : * Estimate the width of the given expr attempting to use the width
6548 : * cached in a Var's owning RelOptInfo, else fallback on the type's
6549 : * average width when unable to or when the given Node is not a Var.
6550 : */
6551 : static int32
6552 1912670 : get_expr_width(PlannerInfo *root, const Node *expr)
6553 : {
6554 : int32 width;
6555 :
6556 1912670 : if (IsA(expr, Var))
6557 : {
6558 1218420 : const Var *var = (const Var *) expr;
6559 :
6560 : /* We should not see any upper-level Vars here */
6561 : Assert(var->varlevelsup == 0);
6562 :
6563 : /* Try to get data from RelOptInfo cache */
6564 1218420 : if (!IS_SPECIAL_VARNO(var->varno) &&
6565 1212510 : var->varno < root->simple_rel_array_size)
6566 : {
6567 1212510 : RelOptInfo *rel = root->simple_rel_array[var->varno];
6568 :
6569 1212510 : if (rel != NULL &&
6570 1194486 : var->varattno >= rel->min_attr &&
6571 1194486 : var->varattno <= rel->max_attr)
6572 : {
6573 1194486 : int ndx = var->varattno - rel->min_attr;
6574 :
6575 1194486 : if (rel->attr_widths[ndx] > 0)
6576 1161902 : return rel->attr_widths[ndx];
6577 : }
6578 : }
6579 :
6580 : /*
6581 : * No cached data available, so estimate using just the type info.
6582 : */
6583 56518 : width = get_typavgwidth(var->vartype, var->vartypmod);
6584 : Assert(width > 0);
6585 :
6586 56518 : return width;
6587 : }
6588 :
6589 694250 : width = get_typavgwidth(exprType(expr), exprTypmod(expr));
6590 : Assert(width > 0);
6591 694250 : return width;
6592 : }
6593 :
6594 : /*
6595 : * relation_byte_size
6596 : * Estimate the storage space in bytes for a given number of tuples
6597 : * of a given width (size in bytes).
6598 : */
6599 : static double
6600 5197866 : relation_byte_size(double tuples, int width)
6601 : {
6602 5197866 : return tuples * (MAXALIGN(width) + MAXALIGN(SizeofHeapTupleHeader));
6603 : }
6604 :
6605 : /*
6606 : * page_size
6607 : * Returns an estimate of the number of pages covered by a given
6608 : * number of tuples of a given width (size in bytes).
6609 : */
6610 : static double
6611 9580 : page_size(double tuples, int width)
6612 : {
6613 9580 : return ceil(relation_byte_size(tuples, width) / BLCKSZ);
6614 : }
6615 :
6616 : /*
6617 : * Estimate the fraction of the work that each worker will do given the
6618 : * number of workers budgeted for the path.
6619 : */
6620 : static double
6621 471456 : get_parallel_divisor(Path *path)
6622 : {
6623 471456 : double parallel_divisor = path->parallel_workers;
6624 :
6625 : /*
6626 : * Early experience with parallel query suggests that when there is only
6627 : * one worker, the leader often makes a very substantial contribution to
6628 : * executing the parallel portion of the plan, but as more workers are
6629 : * added, it does less and less, because it's busy reading tuples from the
6630 : * workers and doing whatever non-parallel post-processing is needed. By
6631 : * the time we reach 4 workers, the leader no longer makes a meaningful
6632 : * contribution. Thus, for now, estimate that the leader spends 30% of
6633 : * its time servicing each worker, and the remainder executing the
6634 : * parallel plan.
6635 : */
6636 471456 : if (parallel_leader_participation)
6637 : {
6638 : double leader_contribution;
6639 :
6640 470226 : leader_contribution = 1.0 - (0.3 * path->parallel_workers);
6641 470226 : if (leader_contribution > 0)
6642 467646 : parallel_divisor += leader_contribution;
6643 : }
6644 :
6645 471456 : return parallel_divisor;
6646 : }
6647 :
6648 : /*
6649 : * compute_bitmap_pages
6650 : * Estimate number of pages fetched from heap in a bitmap heap scan.
6651 : *
6652 : * 'baserel' is the relation to be scanned
6653 : * 'bitmapqual' is a tree of IndexPaths, BitmapAndPaths, and BitmapOrPaths
6654 : * 'loop_count' is the number of repetitions of the indexscan to factor into
6655 : * estimates of caching behavior
6656 : *
6657 : * If cost_p isn't NULL, the indexTotalCost estimate is returned in *cost_p.
6658 : * If tuples_p isn't NULL, the tuples_fetched estimate is returned in *tuples_p.
6659 : */
6660 : double
6661 713824 : compute_bitmap_pages(PlannerInfo *root, RelOptInfo *baserel,
6662 : Path *bitmapqual, double loop_count,
6663 : Cost *cost_p, double *tuples_p)
6664 : {
6665 : Cost indexTotalCost;
6666 : Selectivity indexSelectivity;
6667 : double T;
6668 : double pages_fetched;
6669 : double tuples_fetched;
6670 : double heap_pages;
6671 : double maxentries;
6672 :
6673 : /*
6674 : * Fetch total cost of obtaining the bitmap, as well as its total
6675 : * selectivity.
6676 : */
6677 713824 : cost_bitmap_tree_node(bitmapqual, &indexTotalCost, &indexSelectivity);
6678 :
6679 : /*
6680 : * Estimate number of main-table pages fetched.
6681 : */
6682 713824 : tuples_fetched = clamp_row_est(indexSelectivity * baserel->tuples);
6683 :
6684 713824 : T = (baserel->pages > 1) ? (double) baserel->pages : 1.0;
6685 :
6686 : /*
6687 : * For a single scan, the number of heap pages that need to be fetched is
6688 : * the same as the Mackert and Lohman formula for the case T <= b (ie, no
6689 : * re-reads needed).
6690 : */
6691 713824 : pages_fetched = (2.0 * T * tuples_fetched) / (2.0 * T + tuples_fetched);
6692 :
6693 : /*
6694 : * Calculate the number of pages fetched from the heap. Then based on
6695 : * current work_mem estimate get the estimated maxentries in the bitmap.
6696 : * (Note that we always do this calculation based on the number of pages
6697 : * that would be fetched in a single iteration, even if loop_count > 1.
6698 : * That's correct, because only that number of entries will be stored in
6699 : * the bitmap at one time.)
6700 : */
6701 713824 : heap_pages = Min(pages_fetched, baserel->pages);
6702 713824 : maxentries = tbm_calculate_entries(work_mem * (Size) 1024);
6703 :
6704 713824 : if (loop_count > 1)
6705 : {
6706 : /*
6707 : * For repeated bitmap scans, scale up the number of tuples fetched in
6708 : * the Mackert and Lohman formula by the number of scans, so that we
6709 : * estimate the number of pages fetched by all the scans. Then
6710 : * pro-rate for one scan.
6711 : */
6712 150500 : pages_fetched = index_pages_fetched(tuples_fetched * loop_count,
6713 : baserel->pages,
6714 : get_indexpath_pages(bitmapqual),
6715 : root);
6716 150500 : pages_fetched /= loop_count;
6717 : }
6718 :
6719 713824 : if (pages_fetched >= T)
6720 72576 : pages_fetched = T;
6721 : else
6722 641248 : pages_fetched = ceil(pages_fetched);
6723 :
6724 713824 : if (maxentries < heap_pages)
6725 : {
6726 : double exact_pages;
6727 : double lossy_pages;
6728 :
6729 : /*
6730 : * Crude approximation of the number of lossy pages. Because of the
6731 : * way tbm_lossify() is coded, the number of lossy pages increases
6732 : * very sharply as soon as we run short of memory; this formula has
6733 : * that property and seems to perform adequately in testing, but it's
6734 : * possible we could do better somehow.
6735 : */
6736 18 : lossy_pages = Max(0, heap_pages - maxentries / 2);
6737 18 : exact_pages = heap_pages - lossy_pages;
6738 :
6739 : /*
6740 : * If there are lossy pages then recompute the number of tuples
6741 : * processed by the bitmap heap node. We assume here that the chance
6742 : * of a given tuple coming from an exact page is the same as the
6743 : * chance that a given page is exact. This might not be true, but
6744 : * it's not clear how we can do any better.
6745 : */
6746 18 : if (lossy_pages > 0)
6747 : tuples_fetched =
6748 18 : clamp_row_est(indexSelectivity *
6749 18 : (exact_pages / heap_pages) * baserel->tuples +
6750 18 : (lossy_pages / heap_pages) * baserel->tuples);
6751 : }
6752 :
6753 713824 : if (cost_p)
6754 566564 : *cost_p = indexTotalCost;
6755 713824 : if (tuples_p)
6756 566564 : *tuples_p = tuples_fetched;
6757 :
6758 713824 : return pages_fetched;
6759 : }
6760 :
6761 : /*
6762 : * compute_gather_rows
6763 : * Estimate number of rows for gather (merge) nodes.
6764 : *
6765 : * In a parallel plan, each worker's row estimate is determined by dividing the
6766 : * total number of rows by parallel_divisor, which accounts for the leader's
6767 : * contribution in addition to the number of workers. Accordingly, when
6768 : * estimating the number of rows for gather (merge) nodes, we multiply the rows
6769 : * per worker by the same parallel_divisor to undo the division.
6770 : */
6771 : double
6772 45740 : compute_gather_rows(Path *path)
6773 : {
6774 : Assert(path->parallel_workers > 0);
6775 :
6776 45740 : return clamp_row_est(path->rows * get_parallel_divisor(path));
6777 : }
|
