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 10753844 : 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 10753844 : if (nrows > MAXIMUM_ROWCOUNT || isnan(nrows))
222 0 : nrows = MAXIMUM_ROWCOUNT;
223 10753844 : else if (nrows <= 1.0)
224 3411748 : nrows = 1.0;
225 : else
226 7342096 : nrows = rint(nrows);
227 :
228 10753844 : 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 2025032 : 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 2025032 : 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 2025032 : 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 448634 : cost_seqscan(Path *path, PlannerInfo *root,
270 : RelOptInfo *baserel, ParamPathInfo *param_info)
271 : {
272 448634 : 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 448634 : 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 448634 : if (param_info)
286 840 : path->rows = param_info->ppi_rows;
287 : else
288 447794 : path->rows = baserel->rows;
289 :
290 : /* fetch estimated page cost for tablespace containing table */
291 448634 : get_tablespace_page_costs(baserel->reltablespace,
292 : NULL,
293 : &spc_seq_page_cost);
294 :
295 : /*
296 : * disk costs
297 : */
298 448634 : disk_run_cost = spc_seq_page_cost * baserel->pages;
299 :
300 : /* CPU costs */
301 448634 : get_restriction_qual_cost(root, baserel, param_info, &qpqual_cost);
302 :
303 448634 : startup_cost += qpqual_cost.startup;
304 448634 : cpu_per_tuple = cpu_tuple_cost + qpqual_cost.per_tuple;
305 448634 : cpu_run_cost = cpu_per_tuple * baserel->tuples;
306 : /* tlist eval costs are paid per output row, not per tuple scanned */
307 448634 : startup_cost += path->pathtarget->cost.startup;
308 448634 : cpu_run_cost += path->pathtarget->cost.per_tuple * path->rows;
309 :
310 : /* Adjust costing for parallelism, if used. */
311 448634 : if (path->parallel_workers > 0)
312 : {
313 29084 : double parallel_divisor = get_parallel_divisor(path);
314 :
315 : /* The CPU cost is divided among all the workers. */
316 29084 : 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 29084 : path->rows = clamp_row_est(path->rows / parallel_divisor);
330 : }
331 : else
332 419550 : enable_mask |= PGS_CONSIDER_NONPARTIAL;
333 :
334 448634 : path->disabled_nodes =
335 448634 : (baserel->pgs_mask & enable_mask) == enable_mask ? 0 : 1;
336 448634 : path->startup_cost = startup_cost;
337 448634 : path->total_cost = startup_cost + cpu_run_cost + disk_run_cost;
338 448634 : }
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 26812 : cost_gather(GatherPath *path, PlannerInfo *root,
430 : RelOptInfo *rel, ParamPathInfo *param_info,
431 : double *rows)
432 : {
433 26812 : Cost startup_cost = 0;
434 26812 : Cost run_cost = 0;
435 :
436 : /* Mark the path with the correct row estimate */
437 26812 : if (rows)
438 6944 : path->path.rows = *rows;
439 19868 : else if (param_info)
440 0 : path->path.rows = param_info->ppi_rows;
441 : else
442 19868 : path->path.rows = rel->rows;
443 :
444 26812 : startup_cost = path->subpath->startup_cost;
445 :
446 26812 : run_cost = path->subpath->total_cost - path->subpath->startup_cost;
447 :
448 : /* Parallel setup and communication cost. */
449 26812 : startup_cost += parallel_setup_cost;
450 26812 : run_cost += parallel_tuple_cost * path->path.rows;
451 :
452 26812 : path->path.disabled_nodes = path->subpath->disabled_nodes
453 26812 : + ((rel->pgs_mask & PGS_GATHER) != 0 ? 0 : 1);
454 26812 : path->path.startup_cost = startup_cost;
455 26812 : path->path.total_cost = (startup_cost + run_cost);
456 26812 : }
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 18948 : 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 18948 : Cost startup_cost = 0;
476 18948 : 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 18948 : if (rows)
483 11296 : path->path.rows = *rows;
484 7652 : else if (param_info)
485 0 : path->path.rows = param_info->ppi_rows;
486 : else
487 7652 : 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 18948 : N = (double) path->num_workers + 1;
496 18948 : logN = LOG2(N);
497 :
498 : /* Assumed cost per tuple comparison */
499 18948 : comparison_cost = 2.0 * cpu_operator_cost;
500 :
501 : /* Heap creation cost */
502 18948 : startup_cost += comparison_cost * N * logN;
503 :
504 : /* Per-tuple heap maintenance cost */
505 18948 : run_cost += path->path.rows * comparison_cost * logN;
506 :
507 : /* small cost for heap management, like cost_merge_append */
508 18948 : 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 18948 : startup_cost += parallel_setup_cost;
517 18948 : run_cost += parallel_tuple_cost * path->path.rows * 1.05;
518 :
519 18948 : path->path.disabled_nodes = path->subpath->disabled_nodes
520 18948 : + ((rel->pgs_mask & PGS_GATHER_MERGE) != 0 ? 0 : 1);
521 18948 : path->path.startup_cost = startup_cost + input_startup_cost;
522 18948 : path->path.total_cost = (startup_cost + run_cost + input_total_cost);
523 18948 : }
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 830210 : cost_index(IndexPath *path, PlannerInfo *root, double loop_count,
545 : bool partial_path)
546 : {
547 830210 : IndexOptInfo *index = path->indexinfo;
548 830210 : RelOptInfo *baserel = index->rel;
549 830210 : bool indexonly = (path->path.pathtype == T_IndexOnlyScan);
550 : amcostestimate_function amcostestimate;
551 : List *qpquals;
552 830210 : Cost startup_cost = 0;
553 830210 : Cost run_cost = 0;
554 830210 : 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 830210 : if (path->path.param_info)
586 : {
587 159578 : path->path.rows = path->path.param_info->ppi_rows;
588 : /* qpquals come from the rel's restriction clauses and ppi_clauses */
589 159578 : qpquals = list_concat(extract_nonindex_conditions(path->indexinfo->indrestrictinfo,
590 : path->indexclauses),
591 159578 : extract_nonindex_conditions(path->path.param_info->ppi_clauses,
592 : path->indexclauses));
593 : }
594 : else
595 : {
596 670632 : path->path.rows = baserel->rows;
597 : /* qpquals come from just the rel's restriction clauses */
598 670632 : qpquals = extract_nonindex_conditions(path->indexinfo->indrestrictinfo,
599 : path->indexclauses);
600 : }
601 :
602 : /* is this scan type disabled? */
603 830210 : enable_mask = (indexonly ? PGS_INDEXONLYSCAN : PGS_INDEXSCAN)
604 830210 : | (partial_path ? 0 : PGS_CONSIDER_NONPARTIAL);
605 830210 : path->path.disabled_nodes =
606 830210 : (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 830210 : amcostestimate = (amcostestimate_function) index->amcostestimate;
616 830210 : 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 830210 : path->indextotalcost = indexTotalCost;
627 830210 : path->indexselectivity = indexSelectivity;
628 :
629 : /* all costs for touching index itself included here */
630 830210 : startup_cost += indexStartupCost;
631 830210 : run_cost += indexTotalCost - indexStartupCost;
632 :
633 : /* estimate number of main-table tuples fetched */
634 830210 : tuples_fetched = clamp_row_est(indexSelectivity * baserel->tuples);
635 :
636 : /* fetch estimated page costs for tablespace containing table */
637 830210 : 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 830210 : 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 91382 : pages_fetched = index_pages_fetched(tuples_fetched * loop_count,
679 : baserel->pages,
680 91382 : (double) index->pages,
681 : root);
682 :
683 91382 : if (indexonly)
684 11370 : pages_fetched = ceil(pages_fetched * (1.0 - baserel->allvisfrac));
685 :
686 91382 : rand_heap_pages = pages_fetched;
687 :
688 91382 : 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 91382 : pages_fetched = ceil(indexSelectivity * (double) baserel->pages);
701 :
702 91382 : pages_fetched = index_pages_fetched(pages_fetched * loop_count,
703 : baserel->pages,
704 91382 : (double) index->pages,
705 : root);
706 :
707 91382 : if (indexonly)
708 11370 : pages_fetched = ceil(pages_fetched * (1.0 - baserel->allvisfrac));
709 :
710 91382 : 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 738828 : pages_fetched = index_pages_fetched(tuples_fetched,
719 : baserel->pages,
720 738828 : (double) index->pages,
721 : root);
722 :
723 738828 : if (indexonly)
724 71718 : pages_fetched = ceil(pages_fetched * (1.0 - baserel->allvisfrac));
725 :
726 738828 : rand_heap_pages = pages_fetched;
727 :
728 : /* max_IO_cost is for the perfectly uncorrelated case (csquared=0) */
729 738828 : max_IO_cost = pages_fetched * spc_random_page_cost;
730 :
731 : /* min_IO_cost is for the perfectly correlated case (csquared=1) */
732 738828 : pages_fetched = ceil(indexSelectivity * (double) baserel->pages);
733 :
734 738828 : if (indexonly)
735 71718 : pages_fetched = ceil(pages_fetched * (1.0 - baserel->allvisfrac));
736 :
737 738828 : if (pages_fetched > 0)
738 : {
739 663076 : min_IO_cost = spc_random_page_cost;
740 663076 : if (pages_fetched > 1)
741 201780 : min_IO_cost += (pages_fetched - 1) * spc_seq_page_cost;
742 : }
743 : else
744 75752 : min_IO_cost = 0;
745 : }
746 :
747 830210 : 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 284638 : if (indexonly)
755 26472 : 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 284638 : 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 284638 : if (path->path.parallel_workers <= 0)
774 274400 : return;
775 :
776 10238 : 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 555810 : csquared = indexCorrelation * indexCorrelation;
784 :
785 555810 : 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 555810 : cost_qual_eval(&qpqual_cost, qpquals, root);
794 :
795 555810 : startup_cost += qpqual_cost.startup;
796 555810 : cpu_per_tuple = cpu_tuple_cost + qpqual_cost.per_tuple;
797 :
798 555810 : cpu_run_cost += cpu_per_tuple * tuples_fetched;
799 :
800 : /* tlist eval costs are paid per output row, not per tuple scanned */
801 555810 : startup_cost += path->path.pathtarget->cost.startup;
802 555810 : cpu_run_cost += path->path.pathtarget->cost.per_tuple * path->path.rows;
803 :
804 : /* Adjust costing for parallelism, if used. */
805 555810 : if (path->path.parallel_workers > 0)
806 : {
807 10238 : double parallel_divisor = get_parallel_divisor(&path->path);
808 :
809 10238 : path->path.rows = clamp_row_est(path->path.rows / parallel_divisor);
810 :
811 : /* The CPU cost is divided among all the workers. */
812 10238 : cpu_run_cost /= parallel_divisor;
813 : }
814 :
815 555810 : run_cost += cpu_run_cost;
816 :
817 555810 : path->path.startup_cost = startup_cost;
818 555810 : 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 989788 : extract_nonindex_conditions(List *qual_clauses, List *indexclauses)
839 : {
840 989788 : List *result = NIL;
841 : ListCell *lc;
842 :
843 2047818 : foreach(lc, qual_clauses)
844 : {
845 1058030 : RestrictInfo *rinfo = lfirst_node(RestrictInfo, lc);
846 :
847 1058030 : if (rinfo->pseudoconstant)
848 10304 : continue; /* we may drop pseudoconstants here */
849 1047726 : if (is_redundant_with_indexclauses(rinfo, indexclauses))
850 622760 : continue; /* dup or derived from same EquivalenceClass */
851 : /* ... skip the predicate proof attempt createplan.c will try ... */
852 424966 : result = lappend(result, rinfo);
853 : }
854 989788 : 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 1168666 : 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 1168666 : T = (pages > 1) ? (double) pages : 1.0;
906 :
907 : /* Compute number of pages assumed to be competing for cache space */
908 1168666 : total_pages = root->total_table_pages + index_pages;
909 1168666 : total_pages = Max(total_pages, 1.0);
910 : Assert(T <= total_pages);
911 :
912 : /* b is pro-rated share of effective_cache_size */
913 1168666 : b = (double) effective_cache_size * T / total_pages;
914 :
915 : /* force it positive and integral */
916 1168666 : if (b <= 1.0)
917 0 : b = 1.0;
918 : else
919 1168666 : b = ceil(b);
920 :
921 : /* This part is the Mackert and Lohman formula */
922 1168666 : if (T <= b)
923 : {
924 1168666 : pages_fetched =
925 1168666 : (2.0 * T * tuples_fetched) / (2.0 * T + tuples_fetched);
926 1168666 : if (pages_fetched >= T)
927 685622 : pages_fetched = T;
928 : else
929 483044 : 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 1168666 : 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 198244 : get_indexpath_pages(Path *bitmapqual)
962 : {
963 198244 : double result = 0;
964 : ListCell *l;
965 :
966 198244 : if (IsA(bitmapqual, BitmapAndPath))
967 : {
968 23932 : BitmapAndPath *apath = (BitmapAndPath *) bitmapqual;
969 :
970 71796 : foreach(l, apath->bitmapquals)
971 : {
972 47864 : result += get_indexpath_pages((Path *) lfirst(l));
973 : }
974 : }
975 174312 : else if (IsA(bitmapqual, BitmapOrPath))
976 : {
977 70 : BitmapOrPath *opath = (BitmapOrPath *) bitmapqual;
978 :
979 222 : foreach(l, opath->bitmapquals)
980 : {
981 152 : result += get_indexpath_pages((Path *) lfirst(l));
982 : }
983 : }
984 174242 : else if (IsA(bitmapqual, IndexPath))
985 : {
986 174242 : IndexPath *ipath = (IndexPath *) bitmapqual;
987 :
988 174242 : result = (double) ipath->indexinfo->pages;
989 : }
990 : else
991 0 : elog(ERROR, "unrecognized node type: %d", nodeTag(bitmapqual));
992 :
993 198244 : 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 566830 : cost_bitmap_heap_scan(Path *path, PlannerInfo *root, RelOptInfo *baserel,
1012 : ParamPathInfo *param_info,
1013 : Path *bitmapqual, double loop_count)
1014 : {
1015 566830 : Cost startup_cost = 0;
1016 566830 : 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 566830 : 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 566830 : if (param_info)
1036 241350 : path->rows = param_info->ppi_rows;
1037 : else
1038 325480 : path->rows = baserel->rows;
1039 :
1040 566830 : pages_fetched = compute_bitmap_pages(root, baserel, bitmapqual,
1041 : loop_count, &indexTotalCost,
1042 : &tuples_fetched);
1043 :
1044 566830 : startup_cost += indexTotalCost;
1045 566830 : T = (baserel->pages > 1) ? (double) baserel->pages : 1.0;
1046 :
1047 : /* Fetch estimated page costs for tablespace containing table. */
1048 566830 : 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 566830 : if (pages_fetched >= 2.0)
1060 119584 : cost_per_page = spc_random_page_cost -
1061 119584 : (spc_random_page_cost - spc_seq_page_cost)
1062 119584 : * sqrt(pages_fetched / T);
1063 : else
1064 447246 : cost_per_page = spc_random_page_cost;
1065 :
1066 566830 : 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 566830 : get_restriction_qual_cost(root, baserel, param_info, &qpqual_cost);
1078 :
1079 566830 : startup_cost += qpqual_cost.startup;
1080 566830 : cpu_per_tuple = cpu_tuple_cost + qpqual_cost.per_tuple;
1081 566830 : cpu_run_cost = cpu_per_tuple * tuples_fetched;
1082 :
1083 : /* Adjust costing for parallelism, if used. */
1084 566830 : if (path->parallel_workers > 0)
1085 : {
1086 4956 : double parallel_divisor = get_parallel_divisor(path);
1087 :
1088 : /* The CPU cost is divided among all the workers. */
1089 4956 : cpu_run_cost /= parallel_divisor;
1090 :
1091 4956 : path->rows = clamp_row_est(path->rows / parallel_divisor);
1092 : }
1093 : else
1094 561874 : enable_mask |= PGS_CONSIDER_NONPARTIAL;
1095 :
1096 :
1097 566830 : run_cost += cpu_run_cost;
1098 :
1099 : /* tlist eval costs are paid per output row, not per tuple scanned */
1100 566830 : startup_cost += path->pathtarget->cost.startup;
1101 566830 : run_cost += path->pathtarget->cost.per_tuple * path->rows;
1102 :
1103 566830 : path->disabled_nodes =
1104 566830 : (baserel->pgs_mask & enable_mask) == enable_mask ? 0 : 1;
1105 566830 : path->startup_cost = startup_cost;
1106 566830 : path->total_cost = startup_cost + run_cost;
1107 566830 : }
1108 :
1109 : /*
1110 : * cost_bitmap_tree_node
1111 : * Extract cost and selectivity from a bitmap tree node (index/and/or)
1112 : */
1113 : void
1114 1043468 : cost_bitmap_tree_node(Path *path, Cost *cost, Selectivity *selec)
1115 : {
1116 1043468 : if (IsA(path, IndexPath))
1117 : {
1118 987806 : *cost = ((IndexPath *) path)->indextotalcost;
1119 987806 : *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 987806 : *cost += 0.1 * cpu_operator_cost * path->rows;
1128 : }
1129 55662 : else if (IsA(path, BitmapAndPath))
1130 : {
1131 51994 : *cost = path->total_cost;
1132 51994 : *selec = ((BitmapAndPath *) path)->bitmapselectivity;
1133 : }
1134 3668 : else if (IsA(path, BitmapOrPath))
1135 : {
1136 3668 : *cost = path->total_cost;
1137 3668 : *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 1043468 : }
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 51780 : 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 51780 : totalCost = 0.0;
1173 51780 : selec = 1.0;
1174 155340 : foreach(l, path->bitmapquals)
1175 : {
1176 103560 : Path *subpath = (Path *) lfirst(l);
1177 : Cost subCost;
1178 : Selectivity subselec;
1179 :
1180 103560 : cost_bitmap_tree_node(subpath, &subCost, &subselec);
1181 :
1182 103560 : selec *= subselec;
1183 :
1184 103560 : totalCost += subCost;
1185 103560 : if (l != list_head(path->bitmapquals))
1186 51780 : totalCost += 100.0 * cpu_operator_cost;
1187 : }
1188 51780 : path->bitmapselectivity = selec;
1189 51780 : path->path.rows = 0; /* per above, not used */
1190 51780 : path->path.disabled_nodes = 0;
1191 51780 : path->path.startup_cost = totalCost;
1192 51780 : path->path.total_cost = totalCost;
1193 51780 : }
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 63166 : 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 63166 : 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 63166 : if (param_info)
1499 952 : qpquals = list_concat_copy(param_info->ppi_clauses,
1500 952 : baserel->baserestrictinfo);
1501 : else
1502 62214 : qpquals = baserel->baserestrictinfo;
1503 :
1504 63166 : path->path.rows = clamp_row_est(path->subpath->rows *
1505 63166 : 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 63166 : if (path->path.parallel_workers == 0)
1518 63100 : enable_mask |= PGS_CONSIDER_NONPARTIAL;
1519 63166 : path->path.disabled_nodes = path->subpath->disabled_nodes
1520 63166 : + (((baserel->pgs_mask & enable_mask) != enable_mask) ? 1 : 0);
1521 63166 : path->path.startup_cost = path->subpath->startup_cost;
1522 63166 : 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 63166 : if (qpquals == NIL && trivial_pathtarget)
1538 27646 : return;
1539 :
1540 35520 : get_restriction_qual_cost(root, baserel, param_info, &qpqual_cost);
1541 :
1542 35520 : startup_cost = qpqual_cost.startup;
1543 35520 : cpu_per_tuple = cpu_tuple_cost + qpqual_cost.per_tuple;
1544 35520 : run_cost = cpu_per_tuple * path->subpath->rows;
1545 :
1546 : /* tlist eval costs are paid per output row, not per tuple scanned */
1547 35520 : startup_cost += path->path.pathtarget->cost.startup;
1548 35520 : run_cost += path->path.pathtarget->cost.per_tuple * path->path.rows;
1549 :
1550 35520 : path->path.startup_cost += startup_cost;
1551 35520 : 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 53134 : cost_functionscan(Path *path, PlannerInfo *root,
1563 : RelOptInfo *baserel, ParamPathInfo *param_info)
1564 : {
1565 53134 : Cost startup_cost = 0;
1566 53134 : Cost run_cost = 0;
1567 : QualCost qpqual_cost;
1568 : Cost cpu_per_tuple;
1569 : RangeTblEntry *rte;
1570 : QualCost exprcost;
1571 53134 : uint64 enable_mask = 0;
1572 :
1573 : /* Should only be applied to base relations that are functions */
1574 : Assert(baserel->relid > 0);
1575 53134 : rte = planner_rt_fetch(baserel->relid, root);
1576 : Assert(rte->rtekind == RTE_FUNCTION);
1577 :
1578 : /* Mark the path with the correct row estimate */
1579 53134 : if (param_info)
1580 8602 : path->rows = param_info->ppi_rows;
1581 : else
1582 44532 : 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 53134 : cost_qual_eval_node(&exprcost, (Node *) rte->functions, root);
1598 :
1599 53134 : startup_cost += exprcost.startup + exprcost.per_tuple;
1600 :
1601 : /* Add scanning CPU costs */
1602 53134 : get_restriction_qual_cost(root, baserel, param_info, &qpqual_cost);
1603 :
1604 53134 : startup_cost += qpqual_cost.startup;
1605 53134 : cpu_per_tuple = cpu_tuple_cost + qpqual_cost.per_tuple;
1606 53134 : run_cost += cpu_per_tuple * baserel->tuples;
1607 :
1608 : /* tlist eval costs are paid per output row, not per tuple scanned */
1609 53134 : startup_cost += path->pathtarget->cost.startup;
1610 53134 : run_cost += path->pathtarget->cost.per_tuple * path->rows;
1611 :
1612 53134 : if (path->parallel_workers == 0)
1613 53134 : enable_mask |= PGS_CONSIDER_NONPARTIAL;
1614 53134 : path->disabled_nodes =
1615 53134 : (baserel->pgs_mask & enable_mask) != enable_mask ? 1 : 0;
1616 53134 : path->startup_cost = startup_cost;
1617 53134 : path->total_cost = startup_cost + run_cost;
1618 53134 : }
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 8536 : cost_valuesscan(Path *path, PlannerInfo *root,
1690 : RelOptInfo *baserel, ParamPathInfo *param_info)
1691 : {
1692 8536 : Cost startup_cost = 0;
1693 8536 : Cost run_cost = 0;
1694 : QualCost qpqual_cost;
1695 : Cost cpu_per_tuple;
1696 8536 : 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 8536 : if (param_info)
1704 66 : path->rows = param_info->ppi_rows;
1705 : else
1706 8470 : 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 8536 : cpu_per_tuple = cpu_operator_cost;
1713 :
1714 : /* Add scanning CPU costs */
1715 8536 : get_restriction_qual_cost(root, baserel, param_info, &qpqual_cost);
1716 :
1717 8536 : startup_cost += qpqual_cost.startup;
1718 8536 : cpu_per_tuple += cpu_tuple_cost + qpqual_cost.per_tuple;
1719 8536 : run_cost += cpu_per_tuple * baserel->tuples;
1720 :
1721 : /* tlist eval costs are paid per output row, not per tuple scanned */
1722 8536 : startup_cost += path->pathtarget->cost.startup;
1723 8536 : run_cost += path->pathtarget->cost.per_tuple * path->rows;
1724 :
1725 8536 : if (path->parallel_workers == 0)
1726 8536 : enable_mask |= PGS_CONSIDER_NONPARTIAL;
1727 8536 : path->disabled_nodes =
1728 8536 : (baserel->pgs_mask & enable_mask) != enable_mask ? 1 : 0;
1729 8536 : path->startup_cost = startup_cost;
1730 8536 : path->total_cost = startup_cost + run_cost;
1731 8536 : }
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 5388 : cost_ctescan(Path *path, PlannerInfo *root,
1745 : RelOptInfo *baserel, ParamPathInfo *param_info)
1746 : {
1747 5388 : Cost startup_cost = 0;
1748 5388 : Cost run_cost = 0;
1749 : QualCost qpqual_cost;
1750 : Cost cpu_per_tuple;
1751 5388 : 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 5388 : if (param_info)
1759 0 : path->rows = param_info->ppi_rows;
1760 : else
1761 5388 : path->rows = baserel->rows;
1762 :
1763 : /* Charge one CPU tuple cost per row for tuplestore manipulation */
1764 5388 : cpu_per_tuple = cpu_tuple_cost;
1765 :
1766 : /* Add scanning CPU costs */
1767 5388 : get_restriction_qual_cost(root, baserel, param_info, &qpqual_cost);
1768 :
1769 5388 : startup_cost += qpqual_cost.startup;
1770 5388 : cpu_per_tuple += cpu_tuple_cost + qpqual_cost.per_tuple;
1771 5388 : run_cost += cpu_per_tuple * baserel->tuples;
1772 :
1773 : /* tlist eval costs are paid per output row, not per tuple scanned */
1774 5388 : startup_cost += path->pathtarget->cost.startup;
1775 5388 : run_cost += path->pathtarget->cost.per_tuple * path->rows;
1776 :
1777 5388 : if (path->parallel_workers == 0)
1778 5388 : enable_mask |= PGS_CONSIDER_NONPARTIAL;
1779 5388 : path->disabled_nodes =
1780 5388 : (baserel->pgs_mask & enable_mask) != enable_mask ? 1 : 0;
1781 5388 : path->startup_cost = startup_cost;
1782 5388 : path->total_cost = startup_cost + run_cost;
1783 5388 : }
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 932 : cost_recursive_union(Path *runion, Path *nrterm, Path *rterm)
1875 : {
1876 : Cost startup_cost;
1877 : Cost total_cost;
1878 : double total_rows;
1879 932 : uint64 enable_mask = 0;
1880 :
1881 : /* We probably have decent estimates for the non-recursive term */
1882 932 : startup_cost = nrterm->startup_cost;
1883 932 : total_cost = nrterm->total_cost;
1884 932 : 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 932 : total_cost += 10 * rterm->total_cost;
1893 932 : 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 932 : total_cost += cpu_tuple_cost * total_rows;
1901 :
1902 932 : if (runion->parallel_workers == 0)
1903 932 : enable_mask |= PGS_CONSIDER_NONPARTIAL;
1904 932 : runion->disabled_nodes =
1905 932 : (runion->parent->pgs_mask & enable_mask) != enable_mask ? 1 : 0;
1906 932 : runion->startup_cost = startup_cost;
1907 932 : runion->total_cost = total_cost;
1908 932 : runion->rows = total_rows;
1909 932 : runion->pathtarget->width = Max(nrterm->pathtarget->width,
1910 : rterm->pathtarget->width);
1911 932 : }
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 2145946 : 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 2145946 : double input_bytes = relation_byte_size(tuples, width);
1956 : double output_bytes;
1957 : double output_tuples;
1958 2145946 : 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 2145946 : if (tuples < 2.0)
1965 587714 : tuples = 2.0;
1966 :
1967 : /* Include the default cost-per-comparison */
1968 2145946 : comparison_cost += 2.0 * cpu_operator_cost;
1969 :
1970 : /* Do we have a useful LIMIT? */
1971 2145946 : if (limit_tuples > 0 && limit_tuples < tuples)
1972 : {
1973 1930 : output_tuples = limit_tuples;
1974 1930 : output_bytes = relation_byte_size(output_tuples, width);
1975 : }
1976 : else
1977 : {
1978 2144016 : output_tuples = tuples;
1979 2144016 : output_bytes = input_bytes;
1980 : }
1981 :
1982 2145946 : if (output_bytes > sort_mem_bytes)
1983 : {
1984 : /*
1985 : * We'll have to use a disk-based sort of all the tuples
1986 : */
1987 17720 : double npages = ceil(input_bytes / BLCKSZ);
1988 17720 : double nruns = input_bytes / sort_mem_bytes;
1989 17720 : 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 17720 : *startup_cost = comparison_cost * tuples * LOG2(tuples);
1999 :
2000 : /* Disk costs */
2001 :
2002 : /* Compute logM(r) as log(r) / log(M) */
2003 17720 : if (nruns > mergeorder)
2004 5312 : log_runs = ceil(log(nruns) / log(mergeorder));
2005 : else
2006 12408 : log_runs = 1.0;
2007 17720 : npageaccesses = 2.0 * npages * log_runs;
2008 : /* Assume 3/4ths of accesses are sequential, 1/4th are not */
2009 17720 : *startup_cost += npageaccesses *
2010 17720 : (seq_page_cost * 0.75 + random_page_cost * 0.25);
2011 : }
2012 2128226 : 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 1350 : *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 2126876 : *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 2145946 : *run_cost = cpu_operator_cost * tuples;
2037 2145946 : }
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 12828 : 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 12828 : 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 12828 : List *presortedExprs = NIL;
2068 : ListCell *l;
2069 12828 : 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 12828 : if (input_tuples < 2.0)
2078 6794 : input_tuples = 2.0;
2079 :
2080 : /* Default estimate of number of groups, capped to one group per row. */
2081 12828 : 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 12924 : foreach(l, pathkeys)
2106 : {
2107 12924 : PathKey *key = (PathKey *) lfirst(l);
2108 12924 : EquivalenceMember *member = (EquivalenceMember *)
2109 12924 : 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 12924 : 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 12914 : presortedExprs = lappend(presortedExprs, member->em_expr);
2123 :
2124 12914 : if (foreach_current_index(l) + 1 >= presorted_keys)
2125 12818 : break;
2126 : }
2127 :
2128 : /* Estimate the number of groups with equal presorted keys. */
2129 12828 : if (!unknown_varno)
2130 12818 : input_groups = estimate_num_groups(root, presortedExprs, input_tuples,
2131 : NULL, NULL);
2132 :
2133 12828 : group_tuples = input_tuples / input_groups;
2134 12828 : 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 12828 : 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 12828 : 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 12828 : run_cost = group_run_cost + (group_run_cost + group_startup_cost) *
2158 12828 : (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 12828 : 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 12828 : run_cost += 2.0 * cpu_tuple_cost * input_groups;
2172 :
2173 12828 : 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 12828 : path->disabled_nodes = input_disabled_nodes;
2182 :
2183 12828 : path->startup_cost = startup_cost;
2184 12828 : path->total_cost = startup_cost + run_cost;
2185 12828 : }
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 2133118 : 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 2133118 : cost_tuplesort(&startup_cost, &run_cost,
2211 : tuples, width,
2212 : comparison_cost, sort_mem,
2213 : limit_tuples);
2214 :
2215 2133118 : 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 2133118 : path->rows = tuples;
2222 2133118 : path->disabled_nodes = input_disabled_nodes + (enable_sort ? 0 : 1);
2223 2133118 : path->startup_cost = startup_cost;
2224 2133118 : path->total_cost = startup_cost + run_cost;
2225 2133118 : }
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 26108 : 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 26108 : if (numpaths == 0)
2245 21060 : return 0;
2246 :
2247 : /*
2248 : * Array length is number of workers or number of relevant paths,
2249 : * whichever is less.
2250 : */
2251 5048 : arrlen = Min(parallel_workers, numpaths);
2252 5048 : costarr = palloc_array(Cost, arrlen);
2253 :
2254 : /* The first few paths will each be claimed by a different worker. */
2255 5048 : path_index = 0;
2256 14670 : foreach(cell, subpaths)
2257 : {
2258 10922 : Path *subpath = (Path *) lfirst(cell);
2259 :
2260 10922 : if (path_index == arrlen)
2261 1300 : break;
2262 9622 : 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 5048 : 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 9778 : for_each_cell(l, subpaths, cell)
2276 : {
2277 5276 : Path *subpath = (Path *) lfirst(l);
2278 :
2279 : /* Consider only the non-partial paths */
2280 5276 : if (path_index++ == numpaths)
2281 546 : break;
2282 :
2283 4730 : costarr[min_index] += subpath->total_cost;
2284 :
2285 : /* Update the new min cost array index */
2286 4730 : min_index = 0;
2287 14226 : for (int i = 0; i < arrlen; i++)
2288 : {
2289 9496 : if (costarr[i] < costarr[min_index])
2290 1530 : min_index = i;
2291 : }
2292 : }
2293 :
2294 : /* Return the highest cost from the array */
2295 5048 : max_index = 0;
2296 14670 : for (int i = 0; i < arrlen; i++)
2297 : {
2298 9622 : if (costarr[i] > costarr[max_index])
2299 416 : max_index = i;
2300 : }
2301 :
2302 5048 : return costarr[max_index];
2303 : }
2304 :
2305 : /*
2306 : * cost_append
2307 : * Determines and returns the cost of an Append node.
2308 : */
2309 : void
2310 70888 : cost_append(AppendPath *apath, PlannerInfo *root)
2311 : {
2312 70888 : RelOptInfo *rel = apath->path.parent;
2313 : ListCell *l;
2314 70888 : uint64 enable_mask = PGS_APPEND;
2315 :
2316 70888 : if (apath->path.parallel_workers == 0)
2317 44732 : enable_mask |= PGS_CONSIDER_NONPARTIAL;
2318 :
2319 70888 : apath->path.disabled_nodes =
2320 70888 : (rel->pgs_mask & enable_mask) == enable_mask ? 0 : 1;
2321 70888 : apath->path.startup_cost = 0;
2322 70888 : apath->path.total_cost = 0;
2323 70888 : apath->path.rows = 0;
2324 :
2325 70888 : if (apath->subpaths == NIL)
2326 2108 : return;
2327 :
2328 68780 : if (!apath->path.parallel_aware)
2329 : {
2330 42672 : List *pathkeys = apath->path.pathkeys;
2331 :
2332 42672 : if (pathkeys == NIL)
2333 : {
2334 40528 : 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 40528 : 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 158542 : foreach(l, apath->subpaths)
2347 : {
2348 118014 : Path *subpath = (Path *) lfirst(l);
2349 :
2350 118014 : apath->path.rows += subpath->rows;
2351 118014 : apath->path.disabled_nodes += subpath->disabled_nodes;
2352 118014 : 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 26108 : int i = 0;
2437 26108 : 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 103438 : foreach(l, apath->subpaths)
2444 : {
2445 77330 : 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 77330 : if (i == 0)
2453 26108 : apath->path.startup_cost = subpath->startup_cost;
2454 51222 : else if (i < apath->path.parallel_workers)
2455 25532 : 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 77330 : if (i < apath->first_partial_path)
2466 14352 : apath->path.rows += subpath->rows / parallel_divisor;
2467 : else
2468 : {
2469 : double subpath_parallel_divisor;
2470 :
2471 62978 : subpath_parallel_divisor = get_parallel_divisor(subpath);
2472 62978 : apath->path.rows += subpath->rows * (subpath_parallel_divisor /
2473 : parallel_divisor);
2474 62978 : apath->path.total_cost += subpath->total_cost;
2475 : }
2476 :
2477 77330 : apath->path.disabled_nodes += subpath->disabled_nodes;
2478 77330 : apath->path.rows = clamp_row_est(apath->path.rows);
2479 :
2480 77330 : i++;
2481 : }
2482 :
2483 : /* Add cost for non-partial subpaths. */
2484 26108 : apath->path.total_cost +=
2485 26108 : 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 68780 : apath->path.total_cost +=
2495 68780 : 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 694140 : 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 694140 : Cost startup_cost = input_startup_cost;
2588 694140 : Cost run_cost = input_total_cost - input_startup_cost;
2589 694140 : double nbytes = relation_byte_size(tuples, width);
2590 694140 : double work_mem_bytes = work_mem * (Size) 1024;
2591 :
2592 694140 : path->rows = tuples;
2593 :
2594 : /*
2595 : * Whether spilling or not, charge 2x cpu_operator_cost per tuple to
2596 : * reflect bookkeeping overhead. (This rate must be more than what
2597 : * cost_rescan charges for materialize, ie, cpu_operator_cost per tuple;
2598 : * if it is exactly the same then there will be a cost tie between
2599 : * nestloop with A outer, materialized B inner and nestloop with B outer,
2600 : * materialized A inner. The extra cost ensures we'll prefer
2601 : * materializing the smaller rel.) Note that this is normally a good deal
2602 : * less than cpu_tuple_cost; which is OK because a Material plan node
2603 : * doesn't do qual-checking or projection, so it's got less overhead than
2604 : * most plan nodes.
2605 : */
2606 694140 : run_cost += 2 * cpu_operator_cost * tuples;
2607 :
2608 : /*
2609 : * If we will spill to disk, charge at the rate of seq_page_cost per page.
2610 : * This cost is assumed to be evenly spread through the plan run phase,
2611 : * which isn't exactly accurate but our cost model doesn't allow for
2612 : * nonuniform costs within the run phase.
2613 : */
2614 694140 : if (nbytes > work_mem_bytes)
2615 : {
2616 5056 : double npages = ceil(nbytes / BLCKSZ);
2617 :
2618 5056 : run_cost += seq_page_cost * npages;
2619 : }
2620 :
2621 694140 : path->disabled_nodes = input_disabled_nodes + (enabled ? 0 : 1);
2622 694140 : path->startup_cost = startup_cost;
2623 694140 : path->total_cost = startup_cost + run_cost;
2624 694140 : }
2625 :
2626 : /*
2627 : * cost_memoize_rescan
2628 : * Determines the estimated cost of rescanning a Memoize node.
2629 : *
2630 : * In order to estimate this, we must gain knowledge of how often we expect to
2631 : * be called and how many distinct sets of parameters we are likely to be
2632 : * called with. If we expect a good cache hit ratio, then we can set our
2633 : * costs to account for that hit ratio, plus a little bit of cost for the
2634 : * caching itself. Caching will not work out well if we expect to be called
2635 : * with too many distinct parameter values. The worst-case here is that we
2636 : * never see any parameter value twice, in which case we'd never get a cache
2637 : * hit and caching would be a complete waste of effort.
2638 : */
2639 : static void
2640 296716 : cost_memoize_rescan(PlannerInfo *root, MemoizePath *mpath,
2641 : Cost *rescan_startup_cost, Cost *rescan_total_cost)
2642 : {
2643 : EstimationInfo estinfo;
2644 : ListCell *lc;
2645 296716 : Cost input_startup_cost = mpath->subpath->startup_cost;
2646 296716 : Cost input_total_cost = mpath->subpath->total_cost;
2647 296716 : double tuples = mpath->subpath->rows;
2648 296716 : Cardinality est_calls = mpath->est_calls;
2649 296716 : int width = mpath->subpath->pathtarget->width;
2650 :
2651 : double hash_mem_bytes;
2652 : double est_entry_bytes;
2653 : Cardinality est_cache_entries;
2654 : Cardinality ndistinct;
2655 : double evict_ratio;
2656 : double hit_ratio;
2657 : Cost startup_cost;
2658 : Cost total_cost;
2659 :
2660 : /* available cache space */
2661 296716 : hash_mem_bytes = get_hash_memory_limit();
2662 :
2663 : /*
2664 : * Set the number of bytes each cache entry should consume in the cache.
2665 : * To provide us with better estimations on how many cache entries we can
2666 : * store at once, we make a call to the executor here to ask it what
2667 : * memory overheads there are for a single cache entry.
2668 : */
2669 296716 : est_entry_bytes = relation_byte_size(tuples, width) +
2670 296716 : ExecEstimateCacheEntryOverheadBytes(tuples);
2671 :
2672 : /* include the estimated width for the cache keys */
2673 631892 : foreach(lc, mpath->param_exprs)
2674 335176 : est_entry_bytes += get_expr_width(root, (Node *) lfirst(lc));
2675 :
2676 : /* estimate on the upper limit of cache entries we can hold at once */
2677 296716 : est_cache_entries = floor(hash_mem_bytes / est_entry_bytes);
2678 :
2679 : /* estimate on the distinct number of parameter values */
2680 296716 : ndistinct = estimate_num_groups(root, mpath->param_exprs, est_calls, NULL,
2681 : &estinfo);
2682 :
2683 : /*
2684 : * When the estimation fell back on using a default value, it's a bit too
2685 : * risky to assume that it's ok to use a Memoize node. The use of a
2686 : * default could cause us to use a Memoize node when it's really
2687 : * inappropriate to do so. If we see that this has been done, then we'll
2688 : * assume that every call will have unique parameters, which will almost
2689 : * certainly mean a MemoizePath will never survive add_path().
2690 : */
2691 296716 : if ((estinfo.flags & SELFLAG_USED_DEFAULT) != 0)
2692 17726 : ndistinct = est_calls;
2693 :
2694 : /* Remember the ndistinct estimate for EXPLAIN */
2695 296716 : mpath->est_unique_keys = ndistinct;
2696 :
2697 : /*
2698 : * Since we've already estimated the maximum number of entries we can
2699 : * store at once and know the estimated number of distinct values we'll be
2700 : * called with, we'll take this opportunity to set the path's est_entries.
2701 : * This will ultimately determine the hash table size that the executor
2702 : * will use. If we leave this at zero, the executor will just choose the
2703 : * size itself. Really this is not the right place to do this, but it's
2704 : * convenient since everything is already calculated.
2705 : */
2706 296716 : mpath->est_entries = Min(Min(ndistinct, est_cache_entries),
2707 : PG_UINT32_MAX);
2708 :
2709 : /*
2710 : * When the number of distinct parameter values is above the amount we can
2711 : * store in the cache, then we'll have to evict some entries from the
2712 : * cache. This is not free. Here we estimate how often we'll incur the
2713 : * cost of that eviction.
2714 : */
2715 296716 : evict_ratio = 1.0 - Min(est_cache_entries, ndistinct) / ndistinct;
2716 :
2717 : /*
2718 : * In order to estimate how costly a single scan will be, we need to
2719 : * attempt to estimate what the cache hit ratio will be. To do that we
2720 : * must look at how many scans are estimated in total for this node and
2721 : * how many of those scans we expect to get a cache hit.
2722 : */
2723 593432 : hit_ratio = ((est_calls - ndistinct) / est_calls) *
2724 296716 : (est_cache_entries / Max(ndistinct, est_cache_entries));
2725 :
2726 : /* Remember the hit ratio estimate for EXPLAIN */
2727 296716 : mpath->est_hit_ratio = hit_ratio;
2728 :
2729 : Assert(hit_ratio >= 0 && hit_ratio <= 1.0);
2730 :
2731 : /*
2732 : * Set the total_cost accounting for the expected cache hit ratio. We
2733 : * also add on a cpu_operator_cost to account for a cache lookup. This
2734 : * will happen regardless of whether it's a cache hit or not.
2735 : */
2736 296716 : total_cost = input_total_cost * (1.0 - hit_ratio) + cpu_operator_cost;
2737 :
2738 : /* Now adjust the total cost to account for cache evictions */
2739 :
2740 : /* Charge a cpu_tuple_cost for evicting the actual cache entry */
2741 296716 : total_cost += cpu_tuple_cost * evict_ratio;
2742 :
2743 : /*
2744 : * Charge a 10th of cpu_operator_cost to evict every tuple in that entry.
2745 : * The per-tuple eviction is really just a pfree, so charging a whole
2746 : * cpu_operator_cost seems a little excessive.
2747 : */
2748 296716 : total_cost += cpu_operator_cost / 10.0 * evict_ratio * tuples;
2749 :
2750 : /*
2751 : * Now adjust for storing things in the cache, since that's not free
2752 : * either. Everything must go in the cache. We don't proportion this
2753 : * over any ratio, just apply it once for the scan. We charge a
2754 : * cpu_tuple_cost for the creation of the cache entry and also a
2755 : * cpu_operator_cost for each tuple we expect to cache.
2756 : */
2757 296716 : total_cost += cpu_tuple_cost + cpu_operator_cost * tuples;
2758 :
2759 : /*
2760 : * Getting the first row must be also be proportioned according to the
2761 : * expected cache hit ratio.
2762 : */
2763 296716 : startup_cost = input_startup_cost * (1.0 - hit_ratio);
2764 :
2765 : /*
2766 : * Additionally we charge a cpu_tuple_cost to account for cache lookups,
2767 : * which we'll do regardless of whether it was a cache hit or not.
2768 : */
2769 296716 : startup_cost += cpu_tuple_cost;
2770 :
2771 296716 : *rescan_startup_cost = startup_cost;
2772 296716 : *rescan_total_cost = total_cost;
2773 296716 : }
2774 :
2775 : /*
2776 : * cost_agg
2777 : * Determines and returns the cost of performing an Agg plan node,
2778 : * including the cost of its input.
2779 : *
2780 : * aggcosts can be NULL when there are no actual aggregate functions (i.e.,
2781 : * we are using a hashed Agg node just to do grouping).
2782 : *
2783 : * Note: when aggstrategy == AGG_SORTED, caller must ensure that input costs
2784 : * are for appropriately-sorted input.
2785 : */
2786 : void
2787 95260 : cost_agg(Path *path, PlannerInfo *root,
2788 : AggStrategy aggstrategy, const AggClauseCosts *aggcosts,
2789 : int numGroupCols, double numGroups,
2790 : List *quals,
2791 : int disabled_nodes,
2792 : Cost input_startup_cost, Cost input_total_cost,
2793 : double input_tuples, double input_width)
2794 : {
2795 : double output_tuples;
2796 : Cost startup_cost;
2797 : Cost total_cost;
2798 95260 : const AggClauseCosts dummy_aggcosts = {0};
2799 :
2800 : /* Use all-zero per-aggregate costs if NULL is passed */
2801 95260 : if (aggcosts == NULL)
2802 : {
2803 : Assert(aggstrategy == AGG_HASHED);
2804 19240 : aggcosts = &dummy_aggcosts;
2805 : }
2806 :
2807 : /*
2808 : * The transCost.per_tuple component of aggcosts should be charged once
2809 : * per input tuple, corresponding to the costs of evaluating the aggregate
2810 : * transfns and their input expressions. The finalCost.per_tuple component
2811 : * is charged once per output tuple, corresponding to the costs of
2812 : * evaluating the finalfns. Startup costs are of course charged but once.
2813 : *
2814 : * If we are grouping, we charge an additional cpu_operator_cost per
2815 : * grouping column per input tuple for grouping comparisons.
2816 : *
2817 : * We will produce a single output tuple if not grouping, and a tuple per
2818 : * group otherwise. We charge cpu_tuple_cost for each output tuple.
2819 : *
2820 : * Note: in this cost model, AGG_SORTED and AGG_HASHED have exactly the
2821 : * same total CPU cost, but AGG_SORTED has lower startup cost. If the
2822 : * input path is already sorted appropriately, AGG_SORTED should be
2823 : * preferred (since it has no risk of memory overflow). This will happen
2824 : * as long as the computed total costs are indeed exactly equal --- but if
2825 : * there's roundoff error we might do the wrong thing. So be sure that
2826 : * the computations below form the same intermediate values in the same
2827 : * order.
2828 : */
2829 95260 : if (aggstrategy == AGG_PLAIN)
2830 : {
2831 44958 : startup_cost = input_total_cost;
2832 44958 : startup_cost += aggcosts->transCost.startup;
2833 44958 : startup_cost += aggcosts->transCost.per_tuple * input_tuples;
2834 44958 : startup_cost += aggcosts->finalCost.startup;
2835 44958 : startup_cost += aggcosts->finalCost.per_tuple;
2836 : /* we aren't grouping */
2837 44958 : total_cost = startup_cost + cpu_tuple_cost;
2838 44958 : output_tuples = 1;
2839 : }
2840 50302 : else if (aggstrategy == AGG_SORTED || aggstrategy == AGG_MIXED)
2841 : {
2842 : /* Here we are able to deliver output on-the-fly */
2843 18366 : startup_cost = input_startup_cost;
2844 18366 : total_cost = input_total_cost;
2845 18366 : if (aggstrategy == AGG_MIXED && !enable_hashagg)
2846 540 : ++disabled_nodes;
2847 : /* calcs phrased this way to match HASHED case, see note above */
2848 18366 : total_cost += aggcosts->transCost.startup;
2849 18366 : total_cost += aggcosts->transCost.per_tuple * input_tuples;
2850 18366 : total_cost += (cpu_operator_cost * numGroupCols) * input_tuples;
2851 18366 : total_cost += aggcosts->finalCost.startup;
2852 18366 : total_cost += aggcosts->finalCost.per_tuple * numGroups;
2853 18366 : total_cost += cpu_tuple_cost * numGroups;
2854 18366 : output_tuples = numGroups;
2855 : }
2856 : else
2857 : {
2858 : /* must be AGG_HASHED */
2859 31936 : startup_cost = input_total_cost;
2860 31936 : if (!enable_hashagg)
2861 1890 : ++disabled_nodes;
2862 31936 : startup_cost += aggcosts->transCost.startup;
2863 31936 : startup_cost += aggcosts->transCost.per_tuple * input_tuples;
2864 : /* cost of computing hash value */
2865 31936 : startup_cost += (cpu_operator_cost * numGroupCols) * input_tuples;
2866 31936 : startup_cost += aggcosts->finalCost.startup;
2867 :
2868 31936 : total_cost = startup_cost;
2869 31936 : total_cost += aggcosts->finalCost.per_tuple * numGroups;
2870 : /* cost of retrieving from hash table */
2871 31936 : total_cost += cpu_tuple_cost * numGroups;
2872 31936 : output_tuples = numGroups;
2873 : }
2874 :
2875 : /*
2876 : * Add the disk costs of hash aggregation that spills to disk.
2877 : *
2878 : * Groups that go into the hash table stay in memory until finalized, so
2879 : * spilling and reprocessing tuples doesn't incur additional invocations
2880 : * of transCost or finalCost. Furthermore, the computed hash value is
2881 : * stored with the spilled tuples, so we don't incur extra invocations of
2882 : * the hash function.
2883 : *
2884 : * Hash Agg begins returning tuples after the first batch is complete.
2885 : * Accrue writes (spilled tuples) to startup_cost and to total_cost;
2886 : * accrue reads only to total_cost.
2887 : */
2888 95260 : if (aggstrategy == AGG_HASHED || aggstrategy == AGG_MIXED)
2889 : {
2890 : double pages;
2891 32970 : double pages_written = 0.0;
2892 32970 : double pages_read = 0.0;
2893 : double spill_cost;
2894 : double hashentrysize;
2895 : double nbatches;
2896 : Size mem_limit;
2897 : uint64 ngroups_limit;
2898 : int num_partitions;
2899 : int depth;
2900 :
2901 : /*
2902 : * Estimate number of batches based on the computed limits. If less
2903 : * than or equal to one, all groups are expected to fit in memory;
2904 : * otherwise we expect to spill.
2905 : */
2906 32970 : hashentrysize = hash_agg_entry_size(list_length(root->aggtransinfos),
2907 : input_width,
2908 32970 : aggcosts->transitionSpace);
2909 32970 : hash_agg_set_limits(hashentrysize, numGroups, 0, &mem_limit,
2910 : &ngroups_limit, &num_partitions);
2911 :
2912 32970 : nbatches = Max((numGroups * hashentrysize) / mem_limit,
2913 : numGroups / ngroups_limit);
2914 :
2915 32970 : nbatches = Max(ceil(nbatches), 1.0);
2916 32970 : num_partitions = Max(num_partitions, 2);
2917 :
2918 : /*
2919 : * The number of partitions can change at different levels of
2920 : * recursion; but for the purposes of this calculation assume it stays
2921 : * constant.
2922 : */
2923 32970 : depth = ceil(log(nbatches) / log(num_partitions));
2924 :
2925 : /*
2926 : * Estimate number of pages read and written. For each level of
2927 : * recursion, a tuple must be written and then later read.
2928 : */
2929 32970 : pages = relation_byte_size(input_tuples, input_width) / BLCKSZ;
2930 32970 : pages_written = pages_read = pages * depth;
2931 :
2932 : /*
2933 : * HashAgg has somewhat worse IO behavior than Sort on typical
2934 : * hardware/OS combinations. Account for this with a generic penalty.
2935 : */
2936 32970 : pages_read *= 2.0;
2937 32970 : pages_written *= 2.0;
2938 :
2939 32970 : startup_cost += pages_written * random_page_cost;
2940 32970 : total_cost += pages_written * random_page_cost;
2941 32970 : total_cost += pages_read * seq_page_cost;
2942 :
2943 : /* account for CPU cost of spilling a tuple and reading it back */
2944 32970 : spill_cost = depth * input_tuples * 2.0 * cpu_tuple_cost;
2945 32970 : startup_cost += spill_cost;
2946 32970 : total_cost += spill_cost;
2947 : }
2948 :
2949 : /*
2950 : * If there are quals (HAVING quals), account for their cost and
2951 : * selectivity.
2952 : */
2953 95260 : if (quals)
2954 : {
2955 : QualCost qual_cost;
2956 :
2957 4656 : cost_qual_eval(&qual_cost, quals, root);
2958 4656 : startup_cost += qual_cost.startup;
2959 4656 : total_cost += qual_cost.startup + output_tuples * qual_cost.per_tuple;
2960 :
2961 4656 : output_tuples = clamp_row_est(output_tuples *
2962 4656 : clauselist_selectivity(root,
2963 : quals,
2964 : 0,
2965 : JOIN_INNER,
2966 : NULL));
2967 : }
2968 :
2969 95260 : path->rows = output_tuples;
2970 95260 : path->disabled_nodes = disabled_nodes;
2971 95260 : path->startup_cost = startup_cost;
2972 95260 : path->total_cost = total_cost;
2973 95260 : }
2974 :
2975 : /*
2976 : * get_windowclause_startup_tuples
2977 : * Estimate how many tuples we'll need to fetch from a WindowAgg's
2978 : * subnode before we can output the first WindowAgg tuple.
2979 : *
2980 : * How many tuples need to be read depends on the WindowClause. For example,
2981 : * a WindowClause with no PARTITION BY and no ORDER BY requires that all
2982 : * subnode tuples are read and aggregated before the WindowAgg can output
2983 : * anything. If there's a PARTITION BY, then we only need to look at tuples
2984 : * in the first partition. Here we attempt to estimate just how many
2985 : * 'input_tuples' the WindowAgg will need to read for the given WindowClause
2986 : * before the first tuple can be output.
2987 : */
2988 : static double
2989 3078 : get_windowclause_startup_tuples(PlannerInfo *root, WindowClause *wc,
2990 : double input_tuples)
2991 : {
2992 3078 : int frameOptions = wc->frameOptions;
2993 : double partition_tuples;
2994 : double return_tuples;
2995 : double peer_tuples;
2996 :
2997 : /*
2998 : * First, figure out how many partitions there are likely to be and set
2999 : * partition_tuples according to that estimate.
3000 : */
3001 3078 : if (wc->partitionClause != NIL)
3002 : {
3003 : double num_partitions;
3004 734 : List *partexprs = get_sortgrouplist_exprs(wc->partitionClause,
3005 734 : root->parse->targetList);
3006 :
3007 734 : num_partitions = estimate_num_groups(root, partexprs, input_tuples,
3008 : NULL, NULL);
3009 734 : list_free(partexprs);
3010 :
3011 734 : partition_tuples = input_tuples / num_partitions;
3012 : }
3013 : else
3014 : {
3015 : /* all tuples belong to the same partition */
3016 2344 : partition_tuples = input_tuples;
3017 : }
3018 :
3019 : /* estimate the number of tuples in each peer group */
3020 3078 : if (wc->orderClause != NIL)
3021 : {
3022 : double num_groups;
3023 : List *orderexprs;
3024 :
3025 2370 : orderexprs = get_sortgrouplist_exprs(wc->orderClause,
3026 2370 : root->parse->targetList);
3027 :
3028 : /* estimate out how many peer groups there are in the partition */
3029 2370 : num_groups = estimate_num_groups(root, orderexprs,
3030 : partition_tuples, NULL,
3031 : NULL);
3032 2370 : list_free(orderexprs);
3033 2370 : peer_tuples = partition_tuples / num_groups;
3034 : }
3035 : else
3036 : {
3037 : /* no ORDER BY so only 1 tuple belongs in each peer group */
3038 708 : peer_tuples = 1.0;
3039 : }
3040 :
3041 3078 : if (frameOptions & FRAMEOPTION_END_UNBOUNDED_FOLLOWING)
3042 : {
3043 : /* include all partition rows */
3044 364 : return_tuples = partition_tuples;
3045 : }
3046 2714 : else if (frameOptions & FRAMEOPTION_END_CURRENT_ROW)
3047 : {
3048 1676 : if (frameOptions & FRAMEOPTION_ROWS)
3049 : {
3050 : /* just count the current row */
3051 724 : return_tuples = 1.0;
3052 : }
3053 952 : else if (frameOptions & (FRAMEOPTION_RANGE | FRAMEOPTION_GROUPS))
3054 : {
3055 : /*
3056 : * When in RANGE/GROUPS mode, it's more complex. If there's no
3057 : * ORDER BY, then all rows in the partition are peers, otherwise
3058 : * we'll need to read the first group of peers.
3059 : */
3060 952 : if (wc->orderClause == NIL)
3061 426 : return_tuples = partition_tuples;
3062 : else
3063 526 : return_tuples = peer_tuples;
3064 : }
3065 : else
3066 : {
3067 : /*
3068 : * Something new we don't support yet? This needs attention.
3069 : * We'll just return 1.0 in the meantime.
3070 : */
3071 : Assert(false);
3072 0 : return_tuples = 1.0;
3073 : }
3074 : }
3075 1038 : else if (frameOptions & FRAMEOPTION_END_OFFSET_PRECEDING)
3076 : {
3077 : /*
3078 : * BETWEEN ... AND N PRECEDING will only need to read the WindowAgg's
3079 : * subnode after N ROWS/RANGES/GROUPS. N can be 0, but not negative,
3080 : * so we'll just assume only the current row needs to be read to fetch
3081 : * the first WindowAgg row.
3082 : */
3083 108 : return_tuples = 1.0;
3084 : }
3085 930 : else if (frameOptions & FRAMEOPTION_END_OFFSET_FOLLOWING)
3086 : {
3087 930 : Const *endOffset = (Const *) wc->endOffset;
3088 : double end_offset_value;
3089 :
3090 : /* try and figure out the value specified in the endOffset. */
3091 930 : if (IsA(endOffset, Const))
3092 : {
3093 930 : if (endOffset->constisnull)
3094 : {
3095 : /*
3096 : * NULLs are not allowed, but currently, there's no code to
3097 : * error out if there's a NULL Const. We'll only discover
3098 : * this during execution. For now, just pretend everything is
3099 : * fine and assume that just the first row/range/group will be
3100 : * needed.
3101 : */
3102 0 : end_offset_value = 1.0;
3103 : }
3104 : else
3105 : {
3106 930 : switch (endOffset->consttype)
3107 : {
3108 24 : case INT2OID:
3109 24 : end_offset_value =
3110 24 : (double) DatumGetInt16(endOffset->constvalue);
3111 24 : break;
3112 132 : case INT4OID:
3113 132 : end_offset_value =
3114 132 : (double) DatumGetInt32(endOffset->constvalue);
3115 132 : break;
3116 432 : case INT8OID:
3117 432 : end_offset_value =
3118 432 : (double) DatumGetInt64(endOffset->constvalue);
3119 432 : break;
3120 342 : default:
3121 342 : end_offset_value =
3122 342 : partition_tuples / peer_tuples *
3123 : DEFAULT_INEQ_SEL;
3124 342 : break;
3125 : }
3126 : }
3127 : }
3128 : else
3129 : {
3130 : /*
3131 : * When the end bound is not a Const, we'll just need to guess. We
3132 : * just make use of DEFAULT_INEQ_SEL.
3133 : */
3134 0 : end_offset_value =
3135 0 : partition_tuples / peer_tuples * DEFAULT_INEQ_SEL;
3136 : }
3137 :
3138 930 : if (frameOptions & FRAMEOPTION_ROWS)
3139 : {
3140 : /* include the N FOLLOWING and the current row */
3141 270 : return_tuples = end_offset_value + 1.0;
3142 : }
3143 660 : else if (frameOptions & (FRAMEOPTION_RANGE | FRAMEOPTION_GROUPS))
3144 : {
3145 : /* include N FOLLOWING ranges/group and the initial range/group */
3146 660 : return_tuples = peer_tuples * (end_offset_value + 1.0);
3147 : }
3148 : else
3149 : {
3150 : /*
3151 : * Something new we don't support yet? This needs attention.
3152 : * We'll just return 1.0 in the meantime.
3153 : */
3154 : Assert(false);
3155 0 : return_tuples = 1.0;
3156 : }
3157 : }
3158 : else
3159 : {
3160 : /*
3161 : * Something new we don't support yet? This needs attention. We'll
3162 : * just return 1.0 in the meantime.
3163 : */
3164 : Assert(false);
3165 0 : return_tuples = 1.0;
3166 : }
3167 :
3168 3078 : if (wc->partitionClause != NIL || wc->orderClause != NIL)
3169 : {
3170 : /*
3171 : * Cap the return value to the estimated partition tuples and account
3172 : * for the extra tuple WindowAgg will need to read to confirm the next
3173 : * tuple does not belong to the same partition or peer group.
3174 : */
3175 2582 : return_tuples = Min(return_tuples + 1.0, partition_tuples);
3176 : }
3177 : else
3178 : {
3179 : /*
3180 : * Cap the return value so it's never higher than the expected tuples
3181 : * in the partition.
3182 : */
3183 496 : return_tuples = Min(return_tuples, partition_tuples);
3184 : }
3185 :
3186 : /*
3187 : * We needn't worry about any EXCLUDE options as those only exclude rows
3188 : * from being aggregated, not from being read from the WindowAgg's
3189 : * subnode.
3190 : */
3191 :
3192 3078 : return clamp_row_est(return_tuples);
3193 : }
3194 :
3195 : /*
3196 : * cost_windowagg
3197 : * Determines and returns the cost of performing a WindowAgg plan node,
3198 : * including the cost of its input.
3199 : *
3200 : * Input is assumed already properly sorted.
3201 : */
3202 : void
3203 3078 : cost_windowagg(Path *path, PlannerInfo *root,
3204 : List *windowFuncs, WindowClause *winclause,
3205 : int input_disabled_nodes,
3206 : Cost input_startup_cost, Cost input_total_cost,
3207 : double input_tuples)
3208 : {
3209 : Cost startup_cost;
3210 : Cost total_cost;
3211 : double startup_tuples;
3212 : int numPartCols;
3213 : int numOrderCols;
3214 : ListCell *lc;
3215 :
3216 3078 : numPartCols = list_length(winclause->partitionClause);
3217 3078 : numOrderCols = list_length(winclause->orderClause);
3218 :
3219 3078 : startup_cost = input_startup_cost;
3220 3078 : total_cost = input_total_cost;
3221 :
3222 : /*
3223 : * Window functions are assumed to cost their stated execution cost, plus
3224 : * the cost of evaluating their input expressions, per tuple. Since they
3225 : * may in fact evaluate their inputs at multiple rows during each cycle,
3226 : * this could be a drastic underestimate; but without a way to know how
3227 : * many rows the window function will fetch, it's hard to do better. In
3228 : * any case, it's a good estimate for all the built-in window functions,
3229 : * so we'll just do this for now.
3230 : */
3231 7038 : foreach(lc, windowFuncs)
3232 : {
3233 3960 : WindowFunc *wfunc = lfirst_node(WindowFunc, lc);
3234 : Cost wfunccost;
3235 : QualCost argcosts;
3236 :
3237 3960 : argcosts.startup = argcosts.per_tuple = 0;
3238 3960 : add_function_cost(root, wfunc->winfnoid, (Node *) wfunc,
3239 : &argcosts);
3240 3960 : startup_cost += argcosts.startup;
3241 3960 : wfunccost = argcosts.per_tuple;
3242 :
3243 : /* also add the input expressions' cost to per-input-row costs */
3244 3960 : cost_qual_eval_node(&argcosts, (Node *) wfunc->args, root);
3245 3960 : startup_cost += argcosts.startup;
3246 3960 : wfunccost += argcosts.per_tuple;
3247 :
3248 : /*
3249 : * Add the filter's cost to per-input-row costs. XXX We should reduce
3250 : * input expression costs according to filter selectivity.
3251 : */
3252 3960 : cost_qual_eval_node(&argcosts, (Node *) wfunc->aggfilter, root);
3253 3960 : startup_cost += argcosts.startup;
3254 3960 : wfunccost += argcosts.per_tuple;
3255 :
3256 3960 : total_cost += wfunccost * input_tuples;
3257 : }
3258 :
3259 : /*
3260 : * We also charge cpu_operator_cost per grouping column per tuple for
3261 : * grouping comparisons, plus cpu_tuple_cost per tuple for general
3262 : * overhead.
3263 : *
3264 : * XXX this neglects costs of spooling the data to disk when it overflows
3265 : * work_mem. Sooner or later that should get accounted for.
3266 : */
3267 3078 : total_cost += cpu_operator_cost * (numPartCols + numOrderCols) * input_tuples;
3268 3078 : total_cost += cpu_tuple_cost * input_tuples;
3269 :
3270 3078 : path->rows = input_tuples;
3271 3078 : path->disabled_nodes = input_disabled_nodes;
3272 3078 : path->startup_cost = startup_cost;
3273 3078 : path->total_cost = total_cost;
3274 :
3275 : /*
3276 : * Also, take into account how many tuples we need to read from the
3277 : * subnode in order to produce the first tuple from the WindowAgg. To do
3278 : * this we proportion the run cost (total cost not including startup cost)
3279 : * over the estimated startup tuples. We already included the startup
3280 : * cost of the subnode, so we only need to do this when the estimated
3281 : * startup tuples is above 1.0.
3282 : */
3283 3078 : startup_tuples = get_windowclause_startup_tuples(root, winclause,
3284 : input_tuples);
3285 :
3286 3078 : if (startup_tuples > 1.0)
3287 2568 : path->startup_cost += (total_cost - startup_cost) / input_tuples *
3288 2568 : (startup_tuples - 1.0);
3289 3078 : }
3290 :
3291 : /*
3292 : * cost_group
3293 : * Determines and returns the cost of performing a Group plan node,
3294 : * including the cost of its input.
3295 : *
3296 : * Note: caller must ensure that input costs are for appropriately-sorted
3297 : * input.
3298 : */
3299 : void
3300 1226 : cost_group(Path *path, PlannerInfo *root,
3301 : int numGroupCols, double numGroups,
3302 : List *quals,
3303 : int input_disabled_nodes,
3304 : Cost input_startup_cost, Cost input_total_cost,
3305 : double input_tuples)
3306 : {
3307 : double output_tuples;
3308 : Cost startup_cost;
3309 : Cost total_cost;
3310 :
3311 1226 : output_tuples = numGroups;
3312 1226 : startup_cost = input_startup_cost;
3313 1226 : total_cost = input_total_cost;
3314 :
3315 : /*
3316 : * Charge one cpu_operator_cost per comparison per input tuple. We assume
3317 : * all columns get compared at most of the tuples.
3318 : */
3319 1226 : total_cost += cpu_operator_cost * input_tuples * numGroupCols;
3320 :
3321 : /*
3322 : * If there are quals (HAVING quals), account for their cost and
3323 : * selectivity.
3324 : */
3325 1226 : if (quals)
3326 : {
3327 : QualCost qual_cost;
3328 :
3329 0 : cost_qual_eval(&qual_cost, quals, root);
3330 0 : startup_cost += qual_cost.startup;
3331 0 : total_cost += qual_cost.startup + output_tuples * qual_cost.per_tuple;
3332 :
3333 0 : output_tuples = clamp_row_est(output_tuples *
3334 0 : clauselist_selectivity(root,
3335 : quals,
3336 : 0,
3337 : JOIN_INNER,
3338 : NULL));
3339 : }
3340 :
3341 1226 : path->rows = output_tuples;
3342 1226 : path->disabled_nodes = input_disabled_nodes;
3343 1226 : path->startup_cost = startup_cost;
3344 1226 : path->total_cost = total_cost;
3345 1226 : }
3346 :
3347 : /*
3348 : * initial_cost_nestloop
3349 : * Preliminary estimate of the cost of a nestloop join path.
3350 : *
3351 : * This must quickly produce lower-bound estimates of the path's startup and
3352 : * total costs. If we are unable to eliminate the proposed path from
3353 : * consideration using the lower bounds, final_cost_nestloop will be called
3354 : * to obtain the final estimates.
3355 : *
3356 : * The exact division of labor between this function and final_cost_nestloop
3357 : * is private to them, and represents a tradeoff between speed of the initial
3358 : * estimate and getting a tight lower bound. We choose to not examine the
3359 : * join quals here, since that's by far the most expensive part of the
3360 : * calculations. The end result is that CPU-cost considerations must be
3361 : * left for the second phase; and for SEMI/ANTI joins, we must also postpone
3362 : * incorporation of the inner path's run cost.
3363 : *
3364 : * 'workspace' is to be filled with startup_cost, total_cost, and perhaps
3365 : * other data to be used by final_cost_nestloop
3366 : * 'jointype' is the type of join to be performed
3367 : * 'outer_path' is the outer input to the join
3368 : * 'inner_path' is the inner input to the join
3369 : * 'extra' contains miscellaneous information about the join
3370 : */
3371 : void
3372 3382080 : initial_cost_nestloop(PlannerInfo *root, JoinCostWorkspace *workspace,
3373 : JoinType jointype, uint64 enable_mask,
3374 : Path *outer_path, Path *inner_path,
3375 : JoinPathExtraData *extra)
3376 : {
3377 : int disabled_nodes;
3378 3382080 : Cost startup_cost = 0;
3379 3382080 : Cost run_cost = 0;
3380 3382080 : double outer_path_rows = outer_path->rows;
3381 : Cost inner_rescan_start_cost;
3382 : Cost inner_rescan_total_cost;
3383 : Cost inner_run_cost;
3384 : Cost inner_rescan_run_cost;
3385 :
3386 : /* Count up disabled nodes. */
3387 3382080 : disabled_nodes = (extra->pgs_mask & enable_mask) == enable_mask ? 0 : 1;
3388 3382080 : disabled_nodes += inner_path->disabled_nodes;
3389 3382080 : disabled_nodes += outer_path->disabled_nodes;
3390 :
3391 : /* estimate costs to rescan the inner relation */
3392 3382080 : cost_rescan(root, inner_path,
3393 : &inner_rescan_start_cost,
3394 : &inner_rescan_total_cost);
3395 :
3396 : /* cost of source data */
3397 :
3398 : /*
3399 : * NOTE: clearly, we must pay both outer and inner paths' startup_cost
3400 : * before we can start returning tuples, so the join's startup cost is
3401 : * their sum. We'll also pay the inner path's rescan startup cost
3402 : * multiple times.
3403 : */
3404 3382080 : startup_cost += outer_path->startup_cost + inner_path->startup_cost;
3405 3382080 : run_cost += outer_path->total_cost - outer_path->startup_cost;
3406 3382080 : if (outer_path_rows > 1)
3407 2428176 : run_cost += (outer_path_rows - 1) * inner_rescan_start_cost;
3408 :
3409 3382080 : inner_run_cost = inner_path->total_cost - inner_path->startup_cost;
3410 3382080 : inner_rescan_run_cost = inner_rescan_total_cost - inner_rescan_start_cost;
3411 :
3412 3382080 : if (jointype == JOIN_SEMI || jointype == JOIN_ANTI ||
3413 3320360 : extra->inner_unique)
3414 : {
3415 : /*
3416 : * With a SEMI or ANTI join, or if the innerrel is known unique, the
3417 : * executor will stop after the first match.
3418 : *
3419 : * Getting decent estimates requires inspection of the join quals,
3420 : * which we choose to postpone to final_cost_nestloop.
3421 : */
3422 :
3423 : /* Save private data for final_cost_nestloop */
3424 1384466 : workspace->inner_run_cost = inner_run_cost;
3425 1384466 : workspace->inner_rescan_run_cost = inner_rescan_run_cost;
3426 : }
3427 : else
3428 : {
3429 : /* Normal case; we'll scan whole input rel for each outer row */
3430 1997614 : run_cost += inner_run_cost;
3431 1997614 : if (outer_path_rows > 1)
3432 1527588 : run_cost += (outer_path_rows - 1) * inner_rescan_run_cost;
3433 : }
3434 :
3435 : /* CPU costs left for later */
3436 :
3437 : /* Public result fields */
3438 3382080 : workspace->disabled_nodes = disabled_nodes;
3439 3382080 : workspace->startup_cost = startup_cost;
3440 3382080 : workspace->total_cost = startup_cost + run_cost;
3441 : /* Save private data for final_cost_nestloop */
3442 3382080 : workspace->run_cost = run_cost;
3443 3382080 : }
3444 :
3445 : /*
3446 : * final_cost_nestloop
3447 : * Final estimate of the cost and result size of a nestloop join path.
3448 : *
3449 : * 'path' is already filled in except for the rows and cost fields
3450 : * 'workspace' is the result from initial_cost_nestloop
3451 : * 'extra' contains miscellaneous information about the join
3452 : */
3453 : void
3454 1516240 : final_cost_nestloop(PlannerInfo *root, NestPath *path,
3455 : JoinCostWorkspace *workspace,
3456 : JoinPathExtraData *extra)
3457 : {
3458 1516240 : Path *outer_path = path->jpath.outerjoinpath;
3459 1516240 : Path *inner_path = path->jpath.innerjoinpath;
3460 1516240 : double outer_path_rows = outer_path->rows;
3461 1516240 : double inner_path_rows = inner_path->rows;
3462 1516240 : Cost startup_cost = workspace->startup_cost;
3463 1516240 : Cost run_cost = workspace->run_cost;
3464 : Cost cpu_per_tuple;
3465 : QualCost restrict_qual_cost;
3466 : double ntuples;
3467 :
3468 : /* Set the number of disabled nodes. */
3469 1516240 : path->jpath.path.disabled_nodes = workspace->disabled_nodes;
3470 :
3471 : /* Protect some assumptions below that rowcounts aren't zero */
3472 1516240 : if (outer_path_rows <= 0)
3473 0 : outer_path_rows = 1;
3474 1516240 : if (inner_path_rows <= 0)
3475 726 : inner_path_rows = 1;
3476 : /* Mark the path with the correct row estimate */
3477 1516240 : if (path->jpath.path.param_info)
3478 32178 : path->jpath.path.rows = path->jpath.path.param_info->ppi_rows;
3479 : else
3480 1484062 : path->jpath.path.rows = path->jpath.path.parent->rows;
3481 :
3482 : /* For partial paths, scale row estimate. */
3483 1516240 : if (path->jpath.path.parallel_workers > 0)
3484 : {
3485 44092 : double parallel_divisor = get_parallel_divisor(&path->jpath.path);
3486 :
3487 44092 : path->jpath.path.rows =
3488 44092 : clamp_row_est(path->jpath.path.rows / parallel_divisor);
3489 : }
3490 :
3491 : /* cost of inner-relation source data (we already dealt with outer rel) */
3492 :
3493 1516240 : if (path->jpath.jointype == JOIN_SEMI || path->jpath.jointype == JOIN_ANTI ||
3494 1472978 : extra->inner_unique)
3495 952342 : {
3496 : /*
3497 : * With a SEMI or ANTI join, or if the innerrel is known unique, the
3498 : * executor will stop after the first match.
3499 : */
3500 952342 : Cost inner_run_cost = workspace->inner_run_cost;
3501 952342 : Cost inner_rescan_run_cost = workspace->inner_rescan_run_cost;
3502 : double outer_matched_rows;
3503 : double outer_unmatched_rows;
3504 : Selectivity inner_scan_frac;
3505 :
3506 : /*
3507 : * For an outer-rel row that has at least one match, we can expect the
3508 : * inner scan to stop after a fraction 1/(match_count+1) of the inner
3509 : * rows, if the matches are evenly distributed. Since they probably
3510 : * aren't quite evenly distributed, we apply a fuzz factor of 2.0 to
3511 : * that fraction. (If we used a larger fuzz factor, we'd have to
3512 : * clamp inner_scan_frac to at most 1.0; but since match_count is at
3513 : * least 1, no such clamp is needed now.)
3514 : */
3515 952342 : outer_matched_rows = rint(outer_path_rows * extra->semifactors.outer_match_frac);
3516 952342 : outer_unmatched_rows = outer_path_rows - outer_matched_rows;
3517 952342 : inner_scan_frac = 2.0 / (extra->semifactors.match_count + 1.0);
3518 :
3519 : /*
3520 : * Compute number of tuples processed (not number emitted!). First,
3521 : * account for successfully-matched outer rows.
3522 : */
3523 952342 : ntuples = outer_matched_rows * inner_path_rows * inner_scan_frac;
3524 :
3525 : /*
3526 : * Now we need to estimate the actual costs of scanning the inner
3527 : * relation, which may be quite a bit less than N times inner_run_cost
3528 : * due to early scan stops. We consider two cases. If the inner path
3529 : * is an indexscan using all the joinquals as indexquals, then an
3530 : * unmatched outer row results in an indexscan returning no rows,
3531 : * which is probably quite cheap. Otherwise, the executor will have
3532 : * to scan the whole inner rel for an unmatched row; not so cheap.
3533 : */
3534 952342 : if (has_indexed_join_quals(path))
3535 : {
3536 : /*
3537 : * Successfully-matched outer rows will only require scanning
3538 : * inner_scan_frac of the inner relation. In this case, we don't
3539 : * need to charge the full inner_run_cost even when that's more
3540 : * than inner_rescan_run_cost, because we can assume that none of
3541 : * the inner scans ever scan the whole inner relation. So it's
3542 : * okay to assume that all the inner scan executions can be
3543 : * fractions of the full cost, even if materialization is reducing
3544 : * the rescan cost. At this writing, it's impossible to get here
3545 : * for a materialized inner scan, so inner_run_cost and
3546 : * inner_rescan_run_cost will be the same anyway; but just in
3547 : * case, use inner_run_cost for the first matched tuple and
3548 : * inner_rescan_run_cost for additional ones.
3549 : */
3550 156900 : run_cost += inner_run_cost * inner_scan_frac;
3551 156900 : if (outer_matched_rows > 1)
3552 21180 : run_cost += (outer_matched_rows - 1) * inner_rescan_run_cost * inner_scan_frac;
3553 :
3554 : /*
3555 : * Add the cost of inner-scan executions for unmatched outer rows.
3556 : * We estimate this as the same cost as returning the first tuple
3557 : * of a nonempty scan. We consider that these are all rescans,
3558 : * since we used inner_run_cost once already.
3559 : */
3560 156900 : run_cost += outer_unmatched_rows *
3561 156900 : inner_rescan_run_cost / inner_path_rows;
3562 :
3563 : /*
3564 : * We won't be evaluating any quals at all for unmatched rows, so
3565 : * don't add them to ntuples.
3566 : */
3567 : }
3568 : else
3569 : {
3570 : /*
3571 : * Here, a complicating factor is that rescans may be cheaper than
3572 : * first scans. If we never scan all the way to the end of the
3573 : * inner rel, it might be (depending on the plan type) that we'd
3574 : * never pay the whole inner first-scan run cost. However it is
3575 : * difficult to estimate whether that will happen (and it could
3576 : * not happen if there are any unmatched outer rows!), so be
3577 : * conservative and always charge the whole first-scan cost once.
3578 : * We consider this charge to correspond to the first unmatched
3579 : * outer row, unless there isn't one in our estimate, in which
3580 : * case blame it on the first matched row.
3581 : */
3582 :
3583 : /* First, count all unmatched join tuples as being processed */
3584 795442 : ntuples += outer_unmatched_rows * inner_path_rows;
3585 :
3586 : /* Now add the forced full scan, and decrement appropriate count */
3587 795442 : run_cost += inner_run_cost;
3588 795442 : if (outer_unmatched_rows >= 1)
3589 753516 : outer_unmatched_rows -= 1;
3590 : else
3591 41926 : outer_matched_rows -= 1;
3592 :
3593 : /* Add inner run cost for additional outer tuples having matches */
3594 795442 : if (outer_matched_rows > 0)
3595 278912 : run_cost += outer_matched_rows * inner_rescan_run_cost * inner_scan_frac;
3596 :
3597 : /* Add inner run cost for additional unmatched outer tuples */
3598 795442 : if (outer_unmatched_rows > 0)
3599 489982 : run_cost += outer_unmatched_rows * inner_rescan_run_cost;
3600 : }
3601 : }
3602 : else
3603 : {
3604 : /* Normal-case source costs were included in preliminary estimate */
3605 :
3606 : /* Compute number of tuples processed (not number emitted!) */
3607 563898 : ntuples = outer_path_rows * inner_path_rows;
3608 : }
3609 :
3610 : /* CPU costs */
3611 1516240 : cost_qual_eval(&restrict_qual_cost, path->jpath.joinrestrictinfo, root);
3612 1516240 : startup_cost += restrict_qual_cost.startup;
3613 1516240 : cpu_per_tuple = cpu_tuple_cost + restrict_qual_cost.per_tuple;
3614 1516240 : run_cost += cpu_per_tuple * ntuples;
3615 :
3616 : /* tlist eval costs are paid per output row, not per tuple scanned */
3617 1516240 : startup_cost += path->jpath.path.pathtarget->cost.startup;
3618 1516240 : run_cost += path->jpath.path.pathtarget->cost.per_tuple * path->jpath.path.rows;
3619 :
3620 1516240 : path->jpath.path.startup_cost = startup_cost;
3621 1516240 : path->jpath.path.total_cost = startup_cost + run_cost;
3622 1516240 : }
3623 :
3624 : /*
3625 : * initial_cost_mergejoin
3626 : * Preliminary estimate of the cost of a mergejoin path.
3627 : *
3628 : * This must quickly produce lower-bound estimates of the path's startup and
3629 : * total costs. If we are unable to eliminate the proposed path from
3630 : * consideration using the lower bounds, final_cost_mergejoin will be called
3631 : * to obtain the final estimates.
3632 : *
3633 : * The exact division of labor between this function and final_cost_mergejoin
3634 : * is private to them, and represents a tradeoff between speed of the initial
3635 : * estimate and getting a tight lower bound. We choose to not examine the
3636 : * join quals here, except for obtaining the scan selectivity estimate which
3637 : * is really essential (but fortunately, use of caching keeps the cost of
3638 : * getting that down to something reasonable).
3639 : * We also assume that cost_sort/cost_incremental_sort is cheap enough to use
3640 : * here.
3641 : *
3642 : * 'workspace' is to be filled with startup_cost, total_cost, and perhaps
3643 : * other data to be used by final_cost_mergejoin
3644 : * 'jointype' is the type of join to be performed
3645 : * 'mergeclauses' is the list of joinclauses to be used as merge clauses
3646 : * 'outer_path' is the outer input to the join
3647 : * 'inner_path' is the inner input to the join
3648 : * 'outersortkeys' is the list of sort keys for the outer path
3649 : * 'innersortkeys' is the list of sort keys for the inner path
3650 : * 'outer_presorted_keys' is the number of presorted keys of the outer path
3651 : * 'extra' contains miscellaneous information about the join
3652 : *
3653 : * Note: outersortkeys and innersortkeys should be NIL if no explicit
3654 : * sort is needed because the respective source path is already ordered.
3655 : */
3656 : void
3657 1534904 : initial_cost_mergejoin(PlannerInfo *root, JoinCostWorkspace *workspace,
3658 : JoinType jointype,
3659 : List *mergeclauses,
3660 : Path *outer_path, Path *inner_path,
3661 : List *outersortkeys, List *innersortkeys,
3662 : int outer_presorted_keys,
3663 : JoinPathExtraData *extra)
3664 : {
3665 : int disabled_nodes;
3666 1534904 : Cost startup_cost = 0;
3667 1534904 : Cost run_cost = 0;
3668 1534904 : double outer_path_rows = outer_path->rows;
3669 1534904 : double inner_path_rows = inner_path->rows;
3670 : Cost inner_run_cost;
3671 : double outer_rows,
3672 : inner_rows,
3673 : outer_skip_rows,
3674 : inner_skip_rows;
3675 : Selectivity outerstartsel,
3676 : outerendsel,
3677 : innerstartsel,
3678 : innerendsel;
3679 : Path sort_path; /* dummy for result of
3680 : * cost_sort/cost_incremental_sort */
3681 :
3682 : /* Protect some assumptions below that rowcounts aren't zero */
3683 1534904 : if (outer_path_rows <= 0)
3684 96 : outer_path_rows = 1;
3685 1534904 : if (inner_path_rows <= 0)
3686 126 : inner_path_rows = 1;
3687 :
3688 : /*
3689 : * A merge join will stop as soon as it exhausts either input stream
3690 : * (unless it's an outer join, in which case the outer side has to be
3691 : * scanned all the way anyway). Estimate fraction of the left and right
3692 : * inputs that will actually need to be scanned. Likewise, we can
3693 : * estimate the number of rows that will be skipped before the first join
3694 : * pair is found, which should be factored into startup cost. We use only
3695 : * the first (most significant) merge clause for this purpose. Since
3696 : * mergejoinscansel() is a fairly expensive computation, we cache the
3697 : * results in the merge clause RestrictInfo.
3698 : */
3699 1534904 : if (mergeclauses && jointype != JOIN_FULL)
3700 1528640 : {
3701 1528640 : RestrictInfo *firstclause = (RestrictInfo *) linitial(mergeclauses);
3702 : List *opathkeys;
3703 : List *ipathkeys;
3704 : PathKey *opathkey;
3705 : PathKey *ipathkey;
3706 : MergeScanSelCache *cache;
3707 :
3708 : /* Get the input pathkeys to determine the sort-order details */
3709 1528640 : opathkeys = outersortkeys ? outersortkeys : outer_path->pathkeys;
3710 1528640 : ipathkeys = innersortkeys ? innersortkeys : inner_path->pathkeys;
3711 : Assert(opathkeys);
3712 : Assert(ipathkeys);
3713 1528640 : opathkey = (PathKey *) linitial(opathkeys);
3714 1528640 : ipathkey = (PathKey *) linitial(ipathkeys);
3715 : /* debugging check */
3716 1528640 : if (opathkey->pk_opfamily != ipathkey->pk_opfamily ||
3717 1528640 : opathkey->pk_eclass->ec_collation != ipathkey->pk_eclass->ec_collation ||
3718 1528640 : opathkey->pk_cmptype != ipathkey->pk_cmptype ||
3719 1528640 : opathkey->pk_nulls_first != ipathkey->pk_nulls_first)
3720 0 : elog(ERROR, "left and right pathkeys do not match in mergejoin");
3721 :
3722 : /* Get the selectivity with caching */
3723 1528640 : cache = cached_scansel(root, firstclause, opathkey);
3724 :
3725 1528640 : if (bms_is_subset(firstclause->left_relids,
3726 1528640 : outer_path->parent->relids))
3727 : {
3728 : /* left side of clause is outer */
3729 797216 : outerstartsel = cache->leftstartsel;
3730 797216 : outerendsel = cache->leftendsel;
3731 797216 : innerstartsel = cache->rightstartsel;
3732 797216 : innerendsel = cache->rightendsel;
3733 : }
3734 : else
3735 : {
3736 : /* left side of clause is inner */
3737 731424 : outerstartsel = cache->rightstartsel;
3738 731424 : outerendsel = cache->rightendsel;
3739 731424 : innerstartsel = cache->leftstartsel;
3740 731424 : innerendsel = cache->leftendsel;
3741 : }
3742 1528640 : if (jointype == JOIN_LEFT ||
3743 : jointype == JOIN_ANTI)
3744 : {
3745 200206 : outerstartsel = 0.0;
3746 200206 : outerendsel = 1.0;
3747 : }
3748 1328434 : else if (jointype == JOIN_RIGHT ||
3749 : jointype == JOIN_RIGHT_ANTI)
3750 : {
3751 200054 : innerstartsel = 0.0;
3752 200054 : innerendsel = 1.0;
3753 : }
3754 : }
3755 : else
3756 : {
3757 : /* cope with clauseless or full mergejoin */
3758 6264 : outerstartsel = innerstartsel = 0.0;
3759 6264 : outerendsel = innerendsel = 1.0;
3760 : }
3761 :
3762 : /*
3763 : * Convert selectivities to row counts. We force outer_rows and
3764 : * inner_rows to be at least 1, but the skip_rows estimates can be zero.
3765 : */
3766 1534904 : outer_skip_rows = rint(outer_path_rows * outerstartsel);
3767 1534904 : inner_skip_rows = rint(inner_path_rows * innerstartsel);
3768 1534904 : outer_rows = clamp_row_est(outer_path_rows * outerendsel);
3769 1534904 : inner_rows = clamp_row_est(inner_path_rows * innerendsel);
3770 :
3771 : Assert(outer_skip_rows <= outer_rows);
3772 : Assert(inner_skip_rows <= inner_rows);
3773 :
3774 : /*
3775 : * Readjust scan selectivities to account for above rounding. This is
3776 : * normally an insignificant effect, but when there are only a few rows in
3777 : * the inputs, failing to do this makes for a large percentage error.
3778 : */
3779 1534904 : outerstartsel = outer_skip_rows / outer_path_rows;
3780 1534904 : innerstartsel = inner_skip_rows / inner_path_rows;
3781 1534904 : outerendsel = outer_rows / outer_path_rows;
3782 1534904 : innerendsel = inner_rows / inner_path_rows;
3783 :
3784 : Assert(outerstartsel <= outerendsel);
3785 : Assert(innerstartsel <= innerendsel);
3786 :
3787 : /*
3788 : * We don't decide whether to materialize the inner path until we get to
3789 : * final_cost_mergejoin(), so we don't know whether to check the pgs_mask
3790 : * against PGS_MERGEJOIN_PLAIN or PGS_MERGEJOIN_MATERIALIZE. Instead, we
3791 : * just account for any child nodes here and assume that this node is not
3792 : * itself disabled; we can sort out the details in final_cost_mergejoin().
3793 : *
3794 : * (We could be more precise here by setting disabled_nodes to 1 at this
3795 : * stage if both PGS_MERGEJOIN_PLAIN and PGS_MERGEJOIN_MATERIALIZE are
3796 : * disabled, but that seems to against the idea of making this function
3797 : * produce a quick, optimistic approximation of the final cost.)
3798 : */
3799 1534904 : disabled_nodes = 0;
3800 :
3801 : /* cost of source data */
3802 :
3803 1534904 : if (outersortkeys) /* do we need to sort outer? */
3804 : {
3805 : /*
3806 : * We can assert that the outer path is not already ordered
3807 : * appropriately for the mergejoin; otherwise, outersortkeys would
3808 : * have been set to NIL.
3809 : */
3810 : Assert(!pathkeys_contained_in(outersortkeys, outer_path->pathkeys));
3811 :
3812 : /*
3813 : * We choose to use incremental sort if it is enabled and there are
3814 : * presorted keys; otherwise we use full sort.
3815 : */
3816 775948 : if (enable_incremental_sort && outer_presorted_keys > 0)
3817 : {
3818 1984 : cost_incremental_sort(&sort_path,
3819 : root,
3820 : outersortkeys,
3821 : outer_presorted_keys,
3822 : outer_path->disabled_nodes,
3823 : outer_path->startup_cost,
3824 : outer_path->total_cost,
3825 : outer_path_rows,
3826 1984 : outer_path->pathtarget->width,
3827 : 0.0,
3828 : work_mem,
3829 : -1.0);
3830 : }
3831 : else
3832 : {
3833 773964 : cost_sort(&sort_path,
3834 : root,
3835 : outersortkeys,
3836 : outer_path->disabled_nodes,
3837 : outer_path->total_cost,
3838 : outer_path_rows,
3839 773964 : outer_path->pathtarget->width,
3840 : 0.0,
3841 : work_mem,
3842 : -1.0);
3843 : }
3844 :
3845 775948 : disabled_nodes += sort_path.disabled_nodes;
3846 775948 : startup_cost += sort_path.startup_cost;
3847 775948 : startup_cost += (sort_path.total_cost - sort_path.startup_cost)
3848 775948 : * outerstartsel;
3849 775948 : run_cost += (sort_path.total_cost - sort_path.startup_cost)
3850 775948 : * (outerendsel - outerstartsel);
3851 : }
3852 : else
3853 : {
3854 758956 : disabled_nodes += outer_path->disabled_nodes;
3855 758956 : startup_cost += outer_path->startup_cost;
3856 758956 : startup_cost += (outer_path->total_cost - outer_path->startup_cost)
3857 758956 : * outerstartsel;
3858 758956 : run_cost += (outer_path->total_cost - outer_path->startup_cost)
3859 758956 : * (outerendsel - outerstartsel);
3860 : }
3861 :
3862 1534904 : if (innersortkeys) /* do we need to sort inner? */
3863 : {
3864 : /*
3865 : * We can assert that the inner path is not already ordered
3866 : * appropriately for the mergejoin; otherwise, innersortkeys would
3867 : * have been set to NIL.
3868 : */
3869 : Assert(!pathkeys_contained_in(innersortkeys, inner_path->pathkeys));
3870 :
3871 : /*
3872 : * We do not consider incremental sort for inner path, because
3873 : * incremental sort does not support mark/restore.
3874 : */
3875 :
3876 1228560 : cost_sort(&sort_path,
3877 : root,
3878 : innersortkeys,
3879 : inner_path->disabled_nodes,
3880 : inner_path->total_cost,
3881 : inner_path_rows,
3882 1228560 : inner_path->pathtarget->width,
3883 : 0.0,
3884 : work_mem,
3885 : -1.0);
3886 1228560 : disabled_nodes += sort_path.disabled_nodes;
3887 1228560 : startup_cost += sort_path.startup_cost;
3888 1228560 : startup_cost += (sort_path.total_cost - sort_path.startup_cost)
3889 1228560 : * innerstartsel;
3890 1228560 : inner_run_cost = (sort_path.total_cost - sort_path.startup_cost)
3891 1228560 : * (innerendsel - innerstartsel);
3892 : }
3893 : else
3894 : {
3895 306344 : disabled_nodes += inner_path->disabled_nodes;
3896 306344 : startup_cost += inner_path->startup_cost;
3897 306344 : startup_cost += (inner_path->total_cost - inner_path->startup_cost)
3898 306344 : * innerstartsel;
3899 306344 : inner_run_cost = (inner_path->total_cost - inner_path->startup_cost)
3900 306344 : * (innerendsel - innerstartsel);
3901 : }
3902 :
3903 : /*
3904 : * We can't yet determine whether rescanning occurs, or whether
3905 : * materialization of the inner input should be done. The minimum
3906 : * possible inner input cost, regardless of rescan and materialization
3907 : * considerations, is inner_run_cost. We include that in
3908 : * workspace->total_cost, but not yet in run_cost.
3909 : */
3910 :
3911 : /* CPU costs left for later */
3912 :
3913 : /* Public result fields */
3914 1534904 : workspace->disabled_nodes = disabled_nodes;
3915 1534904 : workspace->startup_cost = startup_cost;
3916 1534904 : workspace->total_cost = startup_cost + run_cost + inner_run_cost;
3917 : /* Save private data for final_cost_mergejoin */
3918 1534904 : workspace->run_cost = run_cost;
3919 1534904 : workspace->inner_run_cost = inner_run_cost;
3920 1534904 : workspace->outer_rows = outer_rows;
3921 1534904 : workspace->inner_rows = inner_rows;
3922 1534904 : workspace->outer_skip_rows = outer_skip_rows;
3923 1534904 : workspace->inner_skip_rows = inner_skip_rows;
3924 1534904 : }
3925 :
3926 : /*
3927 : * final_cost_mergejoin
3928 : * Final estimate of the cost and result size of a mergejoin path.
3929 : *
3930 : * Unlike other costsize functions, this routine makes two actual decisions:
3931 : * whether the executor will need to do mark/restore, and whether we should
3932 : * materialize the inner path. It would be logically cleaner to build
3933 : * separate paths testing these alternatives, but that would require repeating
3934 : * most of the cost calculations, which are not all that cheap. Since the
3935 : * choice will not affect output pathkeys or startup cost, only total cost,
3936 : * there is no possibility of wanting to keep more than one path. So it seems
3937 : * best to make the decisions here and record them in the path's
3938 : * skip_mark_restore and materialize_inner fields.
3939 : *
3940 : * Mark/restore overhead is usually required, but can be skipped if we know
3941 : * that the executor need find only one match per outer tuple, and that the
3942 : * mergeclauses are sufficient to identify a match.
3943 : *
3944 : * We materialize the inner path if we need mark/restore and either the inner
3945 : * path can't support mark/restore, or it's cheaper to use an interposed
3946 : * Material node to handle mark/restore.
3947 : *
3948 : * 'path' is already filled in except for the rows and cost fields and
3949 : * skip_mark_restore and materialize_inner
3950 : * 'workspace' is the result from initial_cost_mergejoin
3951 : * 'extra' contains miscellaneous information about the join
3952 : */
3953 : void
3954 470292 : final_cost_mergejoin(PlannerInfo *root, MergePath *path,
3955 : JoinCostWorkspace *workspace,
3956 : JoinPathExtraData *extra)
3957 : {
3958 470292 : Path *outer_path = path->jpath.outerjoinpath;
3959 470292 : Path *inner_path = path->jpath.innerjoinpath;
3960 470292 : double inner_path_rows = inner_path->rows;
3961 470292 : List *mergeclauses = path->path_mergeclauses;
3962 470292 : List *innersortkeys = path->innersortkeys;
3963 470292 : Cost startup_cost = workspace->startup_cost;
3964 470292 : Cost run_cost = workspace->run_cost;
3965 470292 : Cost inner_run_cost = workspace->inner_run_cost;
3966 470292 : double outer_rows = workspace->outer_rows;
3967 470292 : double inner_rows = workspace->inner_rows;
3968 470292 : double outer_skip_rows = workspace->outer_skip_rows;
3969 470292 : double inner_skip_rows = workspace->inner_skip_rows;
3970 : Cost cpu_per_tuple,
3971 : bare_inner_cost,
3972 : mat_inner_cost;
3973 : QualCost merge_qual_cost;
3974 : QualCost qp_qual_cost;
3975 : double mergejointuples,
3976 : rescannedtuples;
3977 : double rescanratio;
3978 470292 : uint64 enable_mask = 0;
3979 :
3980 : /* Protect some assumptions below that rowcounts aren't zero */
3981 470292 : if (inner_path_rows <= 0)
3982 90 : inner_path_rows = 1;
3983 :
3984 : /* Mark the path with the correct row estimate */
3985 470292 : if (path->jpath.path.param_info)
3986 1624 : path->jpath.path.rows = path->jpath.path.param_info->ppi_rows;
3987 : else
3988 468668 : path->jpath.path.rows = path->jpath.path.parent->rows;
3989 :
3990 : /* For partial paths, scale row estimate. */
3991 470292 : if (path->jpath.path.parallel_workers > 0)
3992 : {
3993 65904 : double parallel_divisor = get_parallel_divisor(&path->jpath.path);
3994 :
3995 65904 : path->jpath.path.rows =
3996 65904 : clamp_row_est(path->jpath.path.rows / parallel_divisor);
3997 : }
3998 :
3999 : /*
4000 : * Compute cost of the mergequals and qpquals (other restriction clauses)
4001 : * separately.
4002 : */
4003 470292 : cost_qual_eval(&merge_qual_cost, mergeclauses, root);
4004 470292 : cost_qual_eval(&qp_qual_cost, path->jpath.joinrestrictinfo, root);
4005 470292 : qp_qual_cost.startup -= merge_qual_cost.startup;
4006 470292 : qp_qual_cost.per_tuple -= merge_qual_cost.per_tuple;
4007 :
4008 : /*
4009 : * With a SEMI or ANTI join, or if the innerrel is known unique, the
4010 : * executor will stop scanning for matches after the first match. When
4011 : * all the joinclauses are merge clauses, this means we don't ever need to
4012 : * back up the merge, and so we can skip mark/restore overhead.
4013 : */
4014 470292 : if ((path->jpath.jointype == JOIN_SEMI ||
4015 463046 : path->jpath.jointype == JOIN_ANTI ||
4016 620090 : extra->inner_unique) &&
4017 164096 : (list_length(path->jpath.joinrestrictinfo) ==
4018 164096 : list_length(path->path_mergeclauses)))
4019 142580 : path->skip_mark_restore = true;
4020 : else
4021 327712 : path->skip_mark_restore = false;
4022 :
4023 : /*
4024 : * Get approx # tuples passing the mergequals. We use approx_tuple_count
4025 : * here because we need an estimate done with JOIN_INNER semantics.
4026 : */
4027 470292 : mergejointuples = approx_tuple_count(root, &path->jpath, mergeclauses);
4028 :
4029 : /*
4030 : * When there are equal merge keys in the outer relation, the mergejoin
4031 : * must rescan any matching tuples in the inner relation. This means
4032 : * re-fetching inner tuples; we have to estimate how often that happens.
4033 : *
4034 : * For regular inner and outer joins, the number of re-fetches can be
4035 : * estimated approximately as size of merge join output minus size of
4036 : * inner relation. Assume that the distinct key values are 1, 2, ..., and
4037 : * denote the number of values of each key in the outer relation as m1,
4038 : * m2, ...; in the inner relation, n1, n2, ... Then we have
4039 : *
4040 : * size of join = m1 * n1 + m2 * n2 + ...
4041 : *
4042 : * number of rescanned tuples = (m1 - 1) * n1 + (m2 - 1) * n2 + ... = m1 *
4043 : * n1 + m2 * n2 + ... - (n1 + n2 + ...) = size of join - size of inner
4044 : * relation
4045 : *
4046 : * This equation works correctly for outer tuples having no inner match
4047 : * (nk = 0), but not for inner tuples having no outer match (mk = 0); we
4048 : * are effectively subtracting those from the number of rescanned tuples,
4049 : * when we should not. Can we do better without expensive selectivity
4050 : * computations?
4051 : *
4052 : * The whole issue is moot if we know we don't need to mark/restore at
4053 : * all, or if we are working from a unique-ified outer input.
4054 : */
4055 470292 : if (path->skip_mark_restore ||
4056 327712 : RELATION_WAS_MADE_UNIQUE(outer_path->parent, extra->sjinfo,
4057 : path->jpath.jointype))
4058 147868 : rescannedtuples = 0;
4059 : else
4060 : {
4061 322424 : rescannedtuples = mergejointuples - inner_path_rows;
4062 : /* Must clamp because of possible underestimate */
4063 322424 : if (rescannedtuples < 0)
4064 77886 : rescannedtuples = 0;
4065 : }
4066 :
4067 : /*
4068 : * We'll inflate various costs this much to account for rescanning. Note
4069 : * that this is to be multiplied by something involving inner_rows, or
4070 : * another number related to the portion of the inner rel we'll scan.
4071 : */
4072 470292 : rescanratio = 1.0 + (rescannedtuples / inner_rows);
4073 :
4074 : /*
4075 : * Decide whether we want to materialize the inner input to shield it from
4076 : * mark/restore and performing re-fetches. Our cost model for regular
4077 : * re-fetches is that a re-fetch costs the same as an original fetch,
4078 : * which is probably an overestimate; but on the other hand we ignore the
4079 : * bookkeeping costs of mark/restore. Not clear if it's worth developing
4080 : * a more refined model. So we just need to inflate the inner run cost by
4081 : * rescanratio.
4082 : */
4083 470292 : bare_inner_cost = inner_run_cost * rescanratio;
4084 :
4085 : /*
4086 : * When we interpose a Material node the re-fetch cost is assumed to be
4087 : * just cpu_operator_cost per tuple, independently of the underlying
4088 : * plan's cost; and we charge an extra cpu_operator_cost per original
4089 : * fetch as well. Note that we're assuming the materialize node will
4090 : * never spill to disk, since it only has to remember tuples back to the
4091 : * last mark. (If there are a huge number of duplicates, our other cost
4092 : * factors will make the path so expensive that it probably won't get
4093 : * chosen anyway.) So we don't use cost_rescan here.
4094 : *
4095 : * Note: keep this estimate in sync with create_mergejoin_plan's labeling
4096 : * of the generated Material node.
4097 : */
4098 470292 : mat_inner_cost = inner_run_cost +
4099 470292 : cpu_operator_cost * inner_rows * rescanratio;
4100 :
4101 : /*
4102 : * If we don't need mark/restore at all, we don't need materialization.
4103 : */
4104 470292 : if (path->skip_mark_restore)
4105 142580 : path->materialize_inner = false;
4106 :
4107 : /*
4108 : * If merge joins with materialization are enabled, then choose
4109 : * materialization if either (a) it looks cheaper or (b) merge joins
4110 : * without materialization are disabled.
4111 : */
4112 327712 : else if ((extra->pgs_mask & PGS_MERGEJOIN_MATERIALIZE) != 0 &&
4113 324242 : (mat_inner_cost < bare_inner_cost ||
4114 324242 : (extra->pgs_mask & PGS_MERGEJOIN_PLAIN) == 0))
4115 3408 : path->materialize_inner = true;
4116 :
4117 : /*
4118 : * Regardless of what plan shapes are enabled and what the costs seem to
4119 : * be, we *must* materialize it if the inner path is to be used directly
4120 : * (without sorting) and it doesn't support mark/restore. Planner failure
4121 : * is not an option!
4122 : *
4123 : * Since the inner side must be ordered, and only Sorts and IndexScans can
4124 : * create order to begin with, and they both support mark/restore, you
4125 : * might think there's no problem --- but you'd be wrong. Nestloop and
4126 : * merge joins can *preserve* the order of their inputs, so they can be
4127 : * selected as the input of a mergejoin, and they don't support
4128 : * mark/restore at present.
4129 : */
4130 324304 : else if (innersortkeys == NIL &&
4131 8942 : !ExecSupportsMarkRestore(inner_path))
4132 1914 : path->materialize_inner = true;
4133 :
4134 : /*
4135 : * Also, force materializing if the inner path is to be sorted and the
4136 : * sort is expected to spill to disk. This is because the final merge
4137 : * pass can be done on-the-fly if it doesn't have to support mark/restore.
4138 : * We don't try to adjust the cost estimates for this consideration,
4139 : * though.
4140 : *
4141 : * Since materialization is a performance optimization in this case,
4142 : * rather than necessary for correctness, we skip it if materialization is
4143 : * switched off.
4144 : */
4145 322390 : else if ((extra->pgs_mask & PGS_MERGEJOIN_MATERIALIZE) != 0 &&
4146 315308 : innersortkeys != NIL &&
4147 315308 : relation_byte_size(inner_path_rows,
4148 315308 : inner_path->pathtarget->width) >
4149 315308 : work_mem * (Size) 1024)
4150 284 : path->materialize_inner = true;
4151 : else
4152 322106 : path->materialize_inner = false;
4153 :
4154 : /* Get the number of disabled nodes, not yet including this one. */
4155 470292 : path->jpath.path.disabled_nodes = workspace->disabled_nodes;
4156 :
4157 : /*
4158 : * Charge the right incremental cost for the chosen case, and update
4159 : * enable_mask as appropriate.
4160 : */
4161 470292 : if (path->materialize_inner)
4162 : {
4163 5606 : run_cost += mat_inner_cost;
4164 5606 : enable_mask |= PGS_MERGEJOIN_MATERIALIZE;
4165 : }
4166 : else
4167 : {
4168 464686 : run_cost += bare_inner_cost;
4169 464686 : enable_mask |= PGS_MERGEJOIN_PLAIN;
4170 : }
4171 :
4172 : /* Incremental count of disabled nodes if this node is disabled. */
4173 470292 : if (path->jpath.path.parallel_workers == 0)
4174 404388 : enable_mask |= PGS_CONSIDER_NONPARTIAL;
4175 470292 : if ((extra->pgs_mask & enable_mask) != enable_mask)
4176 14 : ++path->jpath.path.disabled_nodes;
4177 :
4178 : /* CPU costs */
4179 :
4180 : /*
4181 : * The number of tuple comparisons needed is approximately number of outer
4182 : * rows plus number of inner rows plus number of rescanned tuples (can we
4183 : * refine this?). At each one, we need to evaluate the mergejoin quals.
4184 : */
4185 470292 : startup_cost += merge_qual_cost.startup;
4186 470292 : startup_cost += merge_qual_cost.per_tuple *
4187 470292 : (outer_skip_rows + inner_skip_rows * rescanratio);
4188 470292 : run_cost += merge_qual_cost.per_tuple *
4189 470292 : ((outer_rows - outer_skip_rows) +
4190 470292 : (inner_rows - inner_skip_rows) * rescanratio);
4191 :
4192 : /*
4193 : * For each tuple that gets through the mergejoin proper, we charge
4194 : * cpu_tuple_cost plus the cost of evaluating additional restriction
4195 : * clauses that are to be applied at the join. (This is pessimistic since
4196 : * not all of the quals may get evaluated at each tuple.)
4197 : *
4198 : * Note: we could adjust for SEMI/ANTI joins skipping some qual
4199 : * evaluations here, but it's probably not worth the trouble.
4200 : */
4201 470292 : startup_cost += qp_qual_cost.startup;
4202 470292 : cpu_per_tuple = cpu_tuple_cost + qp_qual_cost.per_tuple;
4203 470292 : run_cost += cpu_per_tuple * mergejointuples;
4204 :
4205 : /* tlist eval costs are paid per output row, not per tuple scanned */
4206 470292 : startup_cost += path->jpath.path.pathtarget->cost.startup;
4207 470292 : run_cost += path->jpath.path.pathtarget->cost.per_tuple * path->jpath.path.rows;
4208 :
4209 470292 : path->jpath.path.startup_cost = startup_cost;
4210 470292 : path->jpath.path.total_cost = startup_cost + run_cost;
4211 470292 : }
4212 :
4213 : /*
4214 : * run mergejoinscansel() with caching
4215 : */
4216 : static MergeScanSelCache *
4217 1528640 : cached_scansel(PlannerInfo *root, RestrictInfo *rinfo, PathKey *pathkey)
4218 : {
4219 : MergeScanSelCache *cache;
4220 : ListCell *lc;
4221 : Selectivity leftstartsel,
4222 : leftendsel,
4223 : rightstartsel,
4224 : rightendsel;
4225 : MemoryContext oldcontext;
4226 :
4227 : /* Do we have this result already? */
4228 1528646 : foreach(lc, rinfo->scansel_cache)
4229 : {
4230 1384208 : cache = (MergeScanSelCache *) lfirst(lc);
4231 1384208 : if (cache->opfamily == pathkey->pk_opfamily &&
4232 1384208 : cache->collation == pathkey->pk_eclass->ec_collation &&
4233 1384208 : cache->cmptype == pathkey->pk_cmptype &&
4234 1384202 : cache->nulls_first == pathkey->pk_nulls_first)
4235 1384202 : return cache;
4236 : }
4237 :
4238 : /* Nope, do the computation */
4239 144438 : mergejoinscansel(root,
4240 144438 : (Node *) rinfo->clause,
4241 : pathkey->pk_opfamily,
4242 : pathkey->pk_cmptype,
4243 144438 : pathkey->pk_nulls_first,
4244 : &leftstartsel,
4245 : &leftendsel,
4246 : &rightstartsel,
4247 : &rightendsel);
4248 :
4249 : /* Cache the result in suitably long-lived workspace */
4250 144438 : oldcontext = MemoryContextSwitchTo(root->planner_cxt);
4251 :
4252 144438 : cache = palloc_object(MergeScanSelCache);
4253 144438 : cache->opfamily = pathkey->pk_opfamily;
4254 144438 : cache->collation = pathkey->pk_eclass->ec_collation;
4255 144438 : cache->cmptype = pathkey->pk_cmptype;
4256 144438 : cache->nulls_first = pathkey->pk_nulls_first;
4257 144438 : cache->leftstartsel = leftstartsel;
4258 144438 : cache->leftendsel = leftendsel;
4259 144438 : cache->rightstartsel = rightstartsel;
4260 144438 : cache->rightendsel = rightendsel;
4261 :
4262 144438 : rinfo->scansel_cache = lappend(rinfo->scansel_cache, cache);
4263 :
4264 144438 : MemoryContextSwitchTo(oldcontext);
4265 :
4266 144438 : return cache;
4267 : }
4268 :
4269 : /*
4270 : * initial_cost_hashjoin
4271 : * Preliminary estimate of the cost of a hashjoin path.
4272 : *
4273 : * This must quickly produce lower-bound estimates of the path's startup and
4274 : * total costs. If we are unable to eliminate the proposed path from
4275 : * consideration using the lower bounds, final_cost_hashjoin will be called
4276 : * to obtain the final estimates.
4277 : *
4278 : * The exact division of labor between this function and final_cost_hashjoin
4279 : * is private to them, and represents a tradeoff between speed of the initial
4280 : * estimate and getting a tight lower bound. We choose to not examine the
4281 : * join quals here (other than by counting the number of hash clauses),
4282 : * so we can't do much with CPU costs. We do assume that
4283 : * ExecChooseHashTableSize is cheap enough to use here.
4284 : *
4285 : * 'workspace' is to be filled with startup_cost, total_cost, and perhaps
4286 : * other data to be used by final_cost_hashjoin
4287 : * 'jointype' is the type of join to be performed
4288 : * 'hashclauses' is the list of joinclauses to be used as hash clauses
4289 : * 'outer_path' is the outer input to the join
4290 : * 'inner_path' is the inner input to the join
4291 : * 'extra' contains miscellaneous information about the join
4292 : * 'parallel_hash' indicates that inner_path is partial and that a shared
4293 : * hash table will be built in parallel
4294 : */
4295 : void
4296 899086 : initial_cost_hashjoin(PlannerInfo *root, JoinCostWorkspace *workspace,
4297 : JoinType jointype,
4298 : List *hashclauses,
4299 : Path *outer_path, Path *inner_path,
4300 : JoinPathExtraData *extra,
4301 : bool parallel_hash)
4302 : {
4303 : int disabled_nodes;
4304 899086 : Cost startup_cost = 0;
4305 899086 : Cost run_cost = 0;
4306 899086 : double outer_path_rows = outer_path->rows;
4307 899086 : double inner_path_rows = inner_path->rows;
4308 899086 : double inner_path_rows_total = inner_path_rows;
4309 899086 : int num_hashclauses = list_length(hashclauses);
4310 : int numbuckets;
4311 : int numbatches;
4312 : int num_skew_mcvs;
4313 : size_t space_allowed; /* unused */
4314 899086 : uint64 enable_mask = PGS_HASHJOIN;
4315 :
4316 899086 : if (outer_path->parallel_workers == 0)
4317 751216 : enable_mask |= PGS_CONSIDER_NONPARTIAL;
4318 :
4319 : /* Count up disabled nodes. */
4320 899086 : disabled_nodes = (extra->pgs_mask & enable_mask) == enable_mask ? 0 : 1;
4321 899086 : disabled_nodes += inner_path->disabled_nodes;
4322 899086 : disabled_nodes += outer_path->disabled_nodes;
4323 :
4324 : /* cost of source data */
4325 899086 : startup_cost += outer_path->startup_cost;
4326 899086 : run_cost += outer_path->total_cost - outer_path->startup_cost;
4327 899086 : startup_cost += inner_path->total_cost;
4328 :
4329 : /*
4330 : * Cost of computing hash function: must do it once per input tuple. We
4331 : * charge one cpu_operator_cost for each column's hash function. Also,
4332 : * tack on one cpu_tuple_cost per inner row, to model the costs of
4333 : * inserting the row into the hashtable.
4334 : *
4335 : * XXX when a hashclause is more complex than a single operator, we really
4336 : * should charge the extra eval costs of the left or right side, as
4337 : * appropriate, here. This seems more work than it's worth at the moment.
4338 : */
4339 899086 : startup_cost += (cpu_operator_cost * num_hashclauses + cpu_tuple_cost)
4340 899086 : * inner_path_rows;
4341 899086 : run_cost += cpu_operator_cost * num_hashclauses * outer_path_rows;
4342 :
4343 : /*
4344 : * If this is a parallel hash build, then the value we have for
4345 : * inner_rows_total currently refers only to the rows returned by each
4346 : * participant. For shared hash table size estimation, we need the total
4347 : * number, so we need to undo the division.
4348 : */
4349 899086 : if (parallel_hash)
4350 75186 : inner_path_rows_total *= get_parallel_divisor(inner_path);
4351 :
4352 : /*
4353 : * Get hash table size that executor would use for inner relation.
4354 : *
4355 : * XXX for the moment, always assume that skew optimization will be
4356 : * performed. As long as SKEW_HASH_MEM_PERCENT is small, it's not worth
4357 : * trying to determine that for sure.
4358 : *
4359 : * XXX at some point it might be interesting to try to account for skew
4360 : * optimization in the cost estimate, but for now, we don't.
4361 : */
4362 899086 : ExecChooseHashTableSize(inner_path_rows_total,
4363 899086 : inner_path->pathtarget->width,
4364 : true, /* useskew */
4365 : parallel_hash, /* try_combined_hash_mem */
4366 : outer_path->parallel_workers,
4367 : &space_allowed,
4368 : &numbuckets,
4369 : &numbatches,
4370 : &num_skew_mcvs);
4371 :
4372 : /*
4373 : * If inner relation is too big then we will need to "batch" the join,
4374 : * which implies writing and reading most of the tuples to disk an extra
4375 : * time. Charge seq_page_cost per page, since the I/O should be nice and
4376 : * sequential. Writing the inner rel counts as startup cost, all the rest
4377 : * as run cost.
4378 : */
4379 899086 : if (numbatches > 1)
4380 : {
4381 4756 : double outerpages = page_size(outer_path_rows,
4382 4756 : outer_path->pathtarget->width);
4383 4756 : double innerpages = page_size(inner_path_rows,
4384 4756 : inner_path->pathtarget->width);
4385 :
4386 4756 : startup_cost += seq_page_cost * innerpages;
4387 4756 : run_cost += seq_page_cost * (innerpages + 2 * outerpages);
4388 : }
4389 :
4390 : /* CPU costs left for later */
4391 :
4392 : /* Public result fields */
4393 899086 : workspace->disabled_nodes = disabled_nodes;
4394 899086 : workspace->startup_cost = startup_cost;
4395 899086 : workspace->total_cost = startup_cost + run_cost;
4396 : /* Save private data for final_cost_hashjoin */
4397 899086 : workspace->run_cost = run_cost;
4398 899086 : workspace->numbuckets = numbuckets;
4399 899086 : workspace->numbatches = numbatches;
4400 899086 : workspace->inner_rows_total = inner_path_rows_total;
4401 899086 : }
4402 :
4403 : /*
4404 : * final_cost_hashjoin
4405 : * Final estimate of the cost and result size of a hashjoin path.
4406 : *
4407 : * Note: the numbatches estimate is also saved into 'path' for use later
4408 : *
4409 : * 'path' is already filled in except for the rows and cost fields and
4410 : * num_batches
4411 : * 'workspace' is the result from initial_cost_hashjoin
4412 : * 'extra' contains miscellaneous information about the join
4413 : */
4414 : void
4415 463572 : final_cost_hashjoin(PlannerInfo *root, HashPath *path,
4416 : JoinCostWorkspace *workspace,
4417 : JoinPathExtraData *extra)
4418 : {
4419 463572 : Path *outer_path = path->jpath.outerjoinpath;
4420 463572 : Path *inner_path = path->jpath.innerjoinpath;
4421 463572 : double outer_path_rows = outer_path->rows;
4422 463572 : double inner_path_rows = inner_path->rows;
4423 463572 : double inner_path_rows_total = workspace->inner_rows_total;
4424 463572 : List *hashclauses = path->path_hashclauses;
4425 463572 : Cost startup_cost = workspace->startup_cost;
4426 463572 : Cost run_cost = workspace->run_cost;
4427 463572 : int numbuckets = workspace->numbuckets;
4428 463572 : int numbatches = workspace->numbatches;
4429 : Cost cpu_per_tuple;
4430 : QualCost hash_qual_cost;
4431 : QualCost qp_qual_cost;
4432 : double hashjointuples;
4433 : double virtualbuckets;
4434 : Selectivity innerbucketsize;
4435 : Selectivity innermcvfreq;
4436 : ListCell *hcl;
4437 :
4438 : /* Set the number of disabled nodes. */
4439 463572 : path->jpath.path.disabled_nodes = workspace->disabled_nodes;
4440 :
4441 : /* Mark the path with the correct row estimate */
4442 463572 : if (path->jpath.path.param_info)
4443 3654 : path->jpath.path.rows = path->jpath.path.param_info->ppi_rows;
4444 : else
4445 459918 : path->jpath.path.rows = path->jpath.path.parent->rows;
4446 :
4447 : /* For partial paths, scale row estimate. */
4448 463572 : if (path->jpath.path.parallel_workers > 0)
4449 : {
4450 107164 : double parallel_divisor = get_parallel_divisor(&path->jpath.path);
4451 :
4452 107164 : path->jpath.path.rows =
4453 107164 : clamp_row_est(path->jpath.path.rows / parallel_divisor);
4454 : }
4455 :
4456 : /* mark the path with estimated # of batches */
4457 463572 : path->num_batches = numbatches;
4458 :
4459 : /* store the total number of tuples (sum of partial row estimates) */
4460 463572 : path->inner_rows_total = inner_path_rows_total;
4461 :
4462 : /* and compute the number of "virtual" buckets in the whole join */
4463 463572 : virtualbuckets = (double) numbuckets * (double) numbatches;
4464 :
4465 : /*
4466 : * Determine bucketsize fraction and MCV frequency for the inner relation.
4467 : * We use the smallest bucketsize or MCV frequency estimated for any
4468 : * individual hashclause; this is undoubtedly conservative.
4469 : *
4470 : * BUT: if inner relation has been unique-ified, we can assume it's good
4471 : * for hashing. This is important both because it's the right answer, and
4472 : * because we avoid contaminating the cache with a value that's wrong for
4473 : * non-unique-ified paths.
4474 : */
4475 463572 : if (RELATION_WAS_MADE_UNIQUE(inner_path->parent, extra->sjinfo,
4476 : path->jpath.jointype))
4477 : {
4478 4374 : innerbucketsize = 1.0 / virtualbuckets;
4479 4374 : innermcvfreq = 1.0 / inner_path_rows_total;
4480 : }
4481 : else
4482 : {
4483 : List *otherclauses;
4484 :
4485 459198 : innerbucketsize = 1.0;
4486 459198 : innermcvfreq = 1.0;
4487 :
4488 : /* At first, try to estimate bucket size using extended statistics. */
4489 459198 : otherclauses = estimate_multivariate_bucketsize(root,
4490 : inner_path->parent,
4491 : hashclauses,
4492 : &innerbucketsize);
4493 :
4494 : /* Pass through the remaining clauses */
4495 955144 : foreach(hcl, otherclauses)
4496 : {
4497 495946 : RestrictInfo *restrictinfo = lfirst_node(RestrictInfo, hcl);
4498 : Selectivity thisbucketsize;
4499 : Selectivity thismcvfreq;
4500 :
4501 : /*
4502 : * First we have to figure out which side of the hashjoin clause
4503 : * is the inner side.
4504 : *
4505 : * Since we tend to visit the same clauses over and over when
4506 : * planning a large query, we cache the bucket stats estimates in
4507 : * the RestrictInfo node to avoid repeated lookups of statistics.
4508 : */
4509 495946 : if (bms_is_subset(restrictinfo->right_relids,
4510 495946 : inner_path->parent->relids))
4511 : {
4512 : /* righthand side is inner */
4513 257834 : thisbucketsize = restrictinfo->right_bucketsize;
4514 257834 : if (thisbucketsize < 0)
4515 : {
4516 : /* not cached yet */
4517 111418 : estimate_hash_bucket_stats(root,
4518 111418 : get_rightop(restrictinfo->clause),
4519 : virtualbuckets,
4520 : &restrictinfo->right_mcvfreq,
4521 : &restrictinfo->right_bucketsize);
4522 111418 : thisbucketsize = restrictinfo->right_bucketsize;
4523 : }
4524 257834 : thismcvfreq = restrictinfo->right_mcvfreq;
4525 : }
4526 : else
4527 : {
4528 : Assert(bms_is_subset(restrictinfo->left_relids,
4529 : inner_path->parent->relids));
4530 : /* lefthand side is inner */
4531 238112 : thisbucketsize = restrictinfo->left_bucketsize;
4532 238112 : if (thisbucketsize < 0)
4533 : {
4534 : /* not cached yet */
4535 95032 : estimate_hash_bucket_stats(root,
4536 95032 : get_leftop(restrictinfo->clause),
4537 : virtualbuckets,
4538 : &restrictinfo->left_mcvfreq,
4539 : &restrictinfo->left_bucketsize);
4540 95032 : thisbucketsize = restrictinfo->left_bucketsize;
4541 : }
4542 238112 : thismcvfreq = restrictinfo->left_mcvfreq;
4543 : }
4544 :
4545 495946 : if (innerbucketsize > thisbucketsize)
4546 362648 : innerbucketsize = thisbucketsize;
4547 : /* Disregard zero for MCV freq, it means we have no data */
4548 495946 : if (thismcvfreq > 0.0 && innermcvfreq > thismcvfreq)
4549 364916 : innermcvfreq = thismcvfreq;
4550 : }
4551 : }
4552 :
4553 : /*
4554 : * If the bucket holding the inner MCV would exceed hash_mem, we don't
4555 : * want to hash unless there is really no other alternative, so apply
4556 : * disable_cost. (The executor normally copes with excessive memory usage
4557 : * by splitting batches, but obviously it cannot separate equal values
4558 : * that way, so it will be unable to drive the batch size below hash_mem
4559 : * when this is true.)
4560 : */
4561 463572 : if (relation_byte_size(clamp_row_est(inner_path_rows * innermcvfreq),
4562 927144 : inner_path->pathtarget->width) > get_hash_memory_limit())
4563 116 : startup_cost += disable_cost;
4564 :
4565 : /*
4566 : * Compute cost of the hashquals and qpquals (other restriction clauses)
4567 : * separately.
4568 : */
4569 463572 : cost_qual_eval(&hash_qual_cost, hashclauses, root);
4570 463572 : cost_qual_eval(&qp_qual_cost, path->jpath.joinrestrictinfo, root);
4571 463572 : qp_qual_cost.startup -= hash_qual_cost.startup;
4572 463572 : qp_qual_cost.per_tuple -= hash_qual_cost.per_tuple;
4573 :
4574 : /* CPU costs */
4575 :
4576 463572 : if (path->jpath.jointype == JOIN_SEMI ||
4577 457214 : path->jpath.jointype == JOIN_ANTI ||
4578 452632 : extra->inner_unique)
4579 126914 : {
4580 : double outer_matched_rows;
4581 : Selectivity inner_scan_frac;
4582 :
4583 : /*
4584 : * With a SEMI or ANTI join, or if the innerrel is known unique, the
4585 : * executor will stop after the first match.
4586 : *
4587 : * For an outer-rel row that has at least one match, we can expect the
4588 : * bucket scan to stop after a fraction 1/(match_count+1) of the
4589 : * bucket's rows, if the matches are evenly distributed. Since they
4590 : * probably aren't quite evenly distributed, we apply a fuzz factor of
4591 : * 2.0 to that fraction. (If we used a larger fuzz factor, we'd have
4592 : * to clamp inner_scan_frac to at most 1.0; but since match_count is
4593 : * at least 1, no such clamp is needed now.)
4594 : */
4595 126914 : outer_matched_rows = rint(outer_path_rows * extra->semifactors.outer_match_frac);
4596 126914 : inner_scan_frac = 2.0 / (extra->semifactors.match_count + 1.0);
4597 :
4598 126914 : startup_cost += hash_qual_cost.startup;
4599 253828 : run_cost += hash_qual_cost.per_tuple * outer_matched_rows *
4600 126914 : clamp_row_est(inner_path_rows * innerbucketsize * inner_scan_frac) * 0.5;
4601 :
4602 : /*
4603 : * For unmatched outer-rel rows, the picture is quite a lot different.
4604 : * In the first place, there is no reason to assume that these rows
4605 : * preferentially hit heavily-populated buckets; instead assume they
4606 : * are uncorrelated with the inner distribution and so they see an
4607 : * average bucket size of inner_path_rows / virtualbuckets. In the
4608 : * second place, it seems likely that they will have few if any exact
4609 : * hash-code matches and so very few of the tuples in the bucket will
4610 : * actually require eval of the hash quals. We don't have any good
4611 : * way to estimate how many will, but for the moment assume that the
4612 : * effective cost per bucket entry is one-tenth what it is for
4613 : * matchable tuples.
4614 : */
4615 253828 : run_cost += hash_qual_cost.per_tuple *
4616 253828 : (outer_path_rows - outer_matched_rows) *
4617 126914 : clamp_row_est(inner_path_rows / virtualbuckets) * 0.05;
4618 :
4619 : /* Get # of tuples that will pass the basic join */
4620 126914 : if (path->jpath.jointype == JOIN_ANTI)
4621 4582 : hashjointuples = outer_path_rows - outer_matched_rows;
4622 : else
4623 122332 : hashjointuples = outer_matched_rows;
4624 : }
4625 : else
4626 : {
4627 : /*
4628 : * The number of tuple comparisons needed is the number of outer
4629 : * tuples times the typical number of tuples in a hash bucket, which
4630 : * is the inner relation size times its bucketsize fraction. At each
4631 : * one, we need to evaluate the hashjoin quals. But actually,
4632 : * charging the full qual eval cost at each tuple is pessimistic,
4633 : * since we don't evaluate the quals unless the hash values match
4634 : * exactly. For lack of a better idea, halve the cost estimate to
4635 : * allow for that.
4636 : */
4637 336658 : startup_cost += hash_qual_cost.startup;
4638 673316 : run_cost += hash_qual_cost.per_tuple * outer_path_rows *
4639 336658 : clamp_row_est(inner_path_rows * innerbucketsize) * 0.5;
4640 :
4641 : /*
4642 : * Get approx # tuples passing the hashquals. We use
4643 : * approx_tuple_count here because we need an estimate done with
4644 : * JOIN_INNER semantics.
4645 : */
4646 336658 : hashjointuples = approx_tuple_count(root, &path->jpath, hashclauses);
4647 : }
4648 :
4649 : /*
4650 : * For each tuple that gets through the hashjoin proper, we charge
4651 : * cpu_tuple_cost plus the cost of evaluating additional restriction
4652 : * clauses that are to be applied at the join. (This is pessimistic since
4653 : * not all of the quals may get evaluated at each tuple.)
4654 : */
4655 463572 : startup_cost += qp_qual_cost.startup;
4656 463572 : cpu_per_tuple = cpu_tuple_cost + qp_qual_cost.per_tuple;
4657 463572 : run_cost += cpu_per_tuple * hashjointuples;
4658 :
4659 : /* tlist eval costs are paid per output row, not per tuple scanned */
4660 463572 : startup_cost += path->jpath.path.pathtarget->cost.startup;
4661 463572 : run_cost += path->jpath.path.pathtarget->cost.per_tuple * path->jpath.path.rows;
4662 :
4663 463572 : path->jpath.path.startup_cost = startup_cost;
4664 463572 : path->jpath.path.total_cost = startup_cost + run_cost;
4665 463572 : }
4666 :
4667 :
4668 : /*
4669 : * cost_subplan
4670 : * Figure the costs for a SubPlan (or initplan).
4671 : *
4672 : * Note: we could dig the subplan's Plan out of the root list, but in practice
4673 : * all callers have it handy already, so we make them pass it.
4674 : */
4675 : void
4676 44298 : cost_subplan(PlannerInfo *root, SubPlan *subplan, Plan *plan)
4677 : {
4678 : QualCost sp_cost;
4679 :
4680 : /*
4681 : * Figure any cost for evaluating the testexpr.
4682 : *
4683 : * Usually, SubPlan nodes are built very early, before we have constructed
4684 : * any RelOptInfos for the parent query level, which means the parent root
4685 : * does not yet contain enough information to safely consult statistics.
4686 : * Therefore, we pass root as NULL here. cost_qual_eval() is already
4687 : * well-equipped to handle a NULL root.
4688 : *
4689 : * One exception is SubPlan nodes built for the initplans of MIN/MAX
4690 : * aggregates from indexes (cf. SS_make_initplan_from_plan). In this
4691 : * case, having a NULL root is safe because testexpr will be NULL.
4692 : * Besides, an initplan will by definition not consult anything from the
4693 : * parent plan.
4694 : */
4695 44298 : cost_qual_eval(&sp_cost,
4696 44298 : make_ands_implicit((Expr *) subplan->testexpr),
4697 : NULL);
4698 :
4699 44298 : if (subplan->useHashTable)
4700 : {
4701 : /*
4702 : * If we are using a hash table for the subquery outputs, then the
4703 : * cost of evaluating the query is a one-time cost. We charge one
4704 : * cpu_operator_cost per tuple for the work of loading the hashtable,
4705 : * too.
4706 : */
4707 2126 : sp_cost.startup += plan->total_cost +
4708 2126 : cpu_operator_cost * plan->plan_rows;
4709 :
4710 : /*
4711 : * The per-tuple costs include the cost of evaluating the lefthand
4712 : * expressions, plus the cost of probing the hashtable. We already
4713 : * accounted for the lefthand expressions as part of the testexpr, and
4714 : * will also have counted one cpu_operator_cost for each comparison
4715 : * operator. That is probably too low for the probing cost, but it's
4716 : * hard to make a better estimate, so live with it for now.
4717 : */
4718 : }
4719 : else
4720 : {
4721 : /*
4722 : * Otherwise we will be rescanning the subplan output on each
4723 : * evaluation. We need to estimate how much of the output we will
4724 : * actually need to scan. NOTE: this logic should agree with the
4725 : * tuple_fraction estimates used by make_subplan() in
4726 : * plan/subselect.c.
4727 : */
4728 42172 : Cost plan_run_cost = plan->total_cost - plan->startup_cost;
4729 :
4730 42172 : if (subplan->subLinkType == EXISTS_SUBLINK)
4731 : {
4732 : /* we only need to fetch 1 tuple; clamp to avoid zero divide */
4733 2572 : sp_cost.per_tuple += plan_run_cost / clamp_row_est(plan->plan_rows);
4734 : }
4735 39600 : else if (subplan->subLinkType == ALL_SUBLINK ||
4736 39582 : subplan->subLinkType == ANY_SUBLINK)
4737 : {
4738 : /* assume we need 50% of the tuples */
4739 146 : sp_cost.per_tuple += 0.50 * plan_run_cost;
4740 : /* also charge a cpu_operator_cost per row examined */
4741 146 : sp_cost.per_tuple += 0.50 * plan->plan_rows * cpu_operator_cost;
4742 : }
4743 : else
4744 : {
4745 : /* assume we need all tuples */
4746 39454 : sp_cost.per_tuple += plan_run_cost;
4747 : }
4748 :
4749 : /*
4750 : * Also account for subplan's startup cost. If the subplan is
4751 : * uncorrelated or undirect correlated, AND its topmost node is one
4752 : * that materializes its output, assume that we'll only need to pay
4753 : * its startup cost once; otherwise assume we pay the startup cost
4754 : * every time.
4755 : */
4756 55652 : if (subplan->parParam == NIL &&
4757 13480 : ExecMaterializesOutput(nodeTag(plan)))
4758 780 : sp_cost.startup += plan->startup_cost;
4759 : else
4760 41392 : sp_cost.per_tuple += plan->startup_cost;
4761 : }
4762 :
4763 44298 : subplan->startup_cost = sp_cost.startup;
4764 44298 : subplan->per_call_cost = sp_cost.per_tuple;
4765 44298 : }
4766 :
4767 :
4768 : /*
4769 : * cost_rescan
4770 : * Given a finished Path, estimate the costs of rescanning it after
4771 : * having done so the first time. For some Path types a rescan is
4772 : * cheaper than an original scan (if no parameters change), and this
4773 : * function embodies knowledge about that. The default is to return
4774 : * the same costs stored in the Path. (Note that the cost estimates
4775 : * actually stored in Paths are always for first scans.)
4776 : *
4777 : * This function is not currently intended to model effects such as rescans
4778 : * being cheaper due to disk block caching; what we are concerned with is
4779 : * plan types wherein the executor caches results explicitly, or doesn't
4780 : * redo startup calculations, etc.
4781 : */
4782 : static void
4783 3382080 : cost_rescan(PlannerInfo *root, Path *path,
4784 : Cost *rescan_startup_cost, /* output parameters */
4785 : Cost *rescan_total_cost)
4786 : {
4787 3382080 : switch (path->pathtype)
4788 : {
4789 53742 : case T_FunctionScan:
4790 :
4791 : /*
4792 : * Currently, nodeFunctionscan.c always executes the function to
4793 : * completion before returning any rows, and caches the results in
4794 : * a tuplestore. So the function eval cost is all startup cost
4795 : * and isn't paid over again on rescans. However, all run costs
4796 : * will be paid over again.
4797 : */
4798 53742 : *rescan_startup_cost = 0;
4799 53742 : *rescan_total_cost = path->total_cost - path->startup_cost;
4800 53742 : break;
4801 134326 : case T_HashJoin:
4802 :
4803 : /*
4804 : * If it's a single-batch join, we don't need to rebuild the hash
4805 : * table during a rescan.
4806 : */
4807 134326 : if (((HashPath *) path)->num_batches == 1)
4808 : {
4809 : /* Startup cost is exactly the cost of hash table building */
4810 134326 : *rescan_startup_cost = 0;
4811 134326 : *rescan_total_cost = path->total_cost - path->startup_cost;
4812 : }
4813 : else
4814 : {
4815 : /* Otherwise, no special treatment */
4816 0 : *rescan_startup_cost = path->startup_cost;
4817 0 : *rescan_total_cost = path->total_cost;
4818 : }
4819 134326 : break;
4820 8456 : case T_CteScan:
4821 : case T_WorkTableScan:
4822 : {
4823 : /*
4824 : * These plan types materialize their final result in a
4825 : * tuplestore or tuplesort object. So the rescan cost is only
4826 : * cpu_tuple_cost per tuple, unless the result is large enough
4827 : * to spill to disk.
4828 : */
4829 8456 : Cost run_cost = cpu_tuple_cost * path->rows;
4830 8456 : double nbytes = relation_byte_size(path->rows,
4831 8456 : path->pathtarget->width);
4832 8456 : double work_mem_bytes = work_mem * (Size) 1024;
4833 :
4834 8456 : if (nbytes > work_mem_bytes)
4835 : {
4836 : /* It will spill, so account for re-read cost */
4837 368 : double npages = ceil(nbytes / BLCKSZ);
4838 :
4839 368 : run_cost += seq_page_cost * npages;
4840 : }
4841 8456 : *rescan_startup_cost = 0;
4842 8456 : *rescan_total_cost = run_cost;
4843 : }
4844 8456 : break;
4845 1221728 : case T_Material:
4846 : case T_Sort:
4847 : {
4848 : /*
4849 : * These plan types not only materialize their results, but do
4850 : * not implement qual filtering or projection. So they are
4851 : * even cheaper to rescan than the ones above. We charge only
4852 : * cpu_operator_cost per tuple. (Note: keep that in sync with
4853 : * the run_cost charge in cost_sort, and also see comments in
4854 : * cost_material before you change it.)
4855 : */
4856 1221728 : Cost run_cost = cpu_operator_cost * path->rows;
4857 1221728 : double nbytes = relation_byte_size(path->rows,
4858 1221728 : path->pathtarget->width);
4859 1221728 : double work_mem_bytes = work_mem * (Size) 1024;
4860 :
4861 1221728 : if (nbytes > work_mem_bytes)
4862 : {
4863 : /* It will spill, so account for re-read cost */
4864 10098 : double npages = ceil(nbytes / BLCKSZ);
4865 :
4866 10098 : run_cost += seq_page_cost * npages;
4867 : }
4868 1221728 : *rescan_startup_cost = 0;
4869 1221728 : *rescan_total_cost = run_cost;
4870 : }
4871 1221728 : break;
4872 296716 : case T_Memoize:
4873 : /* All the hard work is done by cost_memoize_rescan */
4874 296716 : cost_memoize_rescan(root, (MemoizePath *) path,
4875 : rescan_startup_cost, rescan_total_cost);
4876 296716 : break;
4877 1667112 : default:
4878 1667112 : *rescan_startup_cost = path->startup_cost;
4879 1667112 : *rescan_total_cost = path->total_cost;
4880 1667112 : break;
4881 : }
4882 3382080 : }
4883 :
4884 :
4885 : /*
4886 : * cost_qual_eval
4887 : * Estimate the CPU costs of evaluating a WHERE clause.
4888 : * The input can be either an implicitly-ANDed list of boolean
4889 : * expressions, or a list of RestrictInfo nodes. (The latter is
4890 : * preferred since it allows caching of the results.)
4891 : * The result includes both a one-time (startup) component,
4892 : * and a per-evaluation component.
4893 : *
4894 : * Note: in some code paths root can be passed as NULL, resulting in
4895 : * slightly worse estimates.
4896 : */
4897 : void
4898 4788206 : cost_qual_eval(QualCost *cost, List *quals, PlannerInfo *root)
4899 : {
4900 : cost_qual_eval_context context;
4901 : ListCell *l;
4902 :
4903 4788206 : context.root = root;
4904 4788206 : context.total.startup = 0;
4905 4788206 : context.total.per_tuple = 0;
4906 :
4907 : /* We don't charge any cost for the implicit ANDing at top level ... */
4908 :
4909 9109100 : foreach(l, quals)
4910 : {
4911 4320894 : Node *qual = (Node *) lfirst(l);
4912 :
4913 4320894 : cost_qual_eval_walker(qual, &context);
4914 : }
4915 :
4916 4788206 : *cost = context.total;
4917 4788206 : }
4918 :
4919 : /*
4920 : * cost_qual_eval_node
4921 : * As above, for a single RestrictInfo or expression.
4922 : */
4923 : void
4924 1896420 : cost_qual_eval_node(QualCost *cost, Node *qual, PlannerInfo *root)
4925 : {
4926 : cost_qual_eval_context context;
4927 :
4928 1896420 : context.root = root;
4929 1896420 : context.total.startup = 0;
4930 1896420 : context.total.per_tuple = 0;
4931 :
4932 1896420 : cost_qual_eval_walker(qual, &context);
4933 :
4934 1896420 : *cost = context.total;
4935 1896420 : }
4936 :
4937 : static bool
4938 9751660 : cost_qual_eval_walker(Node *node, cost_qual_eval_context *context)
4939 : {
4940 9751660 : if (node == NULL)
4941 95810 : return false;
4942 :
4943 : /*
4944 : * RestrictInfo nodes contain an eval_cost field reserved for this
4945 : * routine's use, so that it's not necessary to evaluate the qual clause's
4946 : * cost more than once. If the clause's cost hasn't been computed yet,
4947 : * the field's startup value will contain -1.
4948 : */
4949 9655850 : if (IsA(node, RestrictInfo))
4950 : {
4951 4511654 : RestrictInfo *rinfo = (RestrictInfo *) node;
4952 :
4953 4511654 : if (rinfo->eval_cost.startup < 0)
4954 : {
4955 : cost_qual_eval_context locContext;
4956 :
4957 615470 : locContext.root = context->root;
4958 615470 : locContext.total.startup = 0;
4959 615470 : locContext.total.per_tuple = 0;
4960 :
4961 : /*
4962 : * For an OR clause, recurse into the marked-up tree so that we
4963 : * set the eval_cost for contained RestrictInfos too.
4964 : */
4965 615470 : if (rinfo->orclause)
4966 10020 : cost_qual_eval_walker((Node *) rinfo->orclause, &locContext);
4967 : else
4968 605450 : cost_qual_eval_walker((Node *) rinfo->clause, &locContext);
4969 :
4970 : /*
4971 : * If the RestrictInfo is marked pseudoconstant, it will be tested
4972 : * only once, so treat its cost as all startup cost.
4973 : */
4974 615470 : if (rinfo->pseudoconstant)
4975 : {
4976 : /* count one execution during startup */
4977 10310 : locContext.total.startup += locContext.total.per_tuple;
4978 10310 : locContext.total.per_tuple = 0;
4979 : }
4980 615470 : rinfo->eval_cost = locContext.total;
4981 : }
4982 4511654 : context->total.startup += rinfo->eval_cost.startup;
4983 4511654 : context->total.per_tuple += rinfo->eval_cost.per_tuple;
4984 : /* do NOT recurse into children */
4985 4511654 : return false;
4986 : }
4987 :
4988 : /*
4989 : * For each operator or function node in the given tree, we charge the
4990 : * estimated execution cost given by pg_proc.procost (remember to multiply
4991 : * this by cpu_operator_cost).
4992 : *
4993 : * Vars and Consts are charged zero, and so are boolean operators (AND,
4994 : * OR, NOT). Simplistic, but a lot better than no model at all.
4995 : *
4996 : * Should we try to account for the possibility of short-circuit
4997 : * evaluation of AND/OR? Probably *not*, because that would make the
4998 : * results depend on the clause ordering, and we are not in any position
4999 : * to expect that the current ordering of the clauses is the one that's
5000 : * going to end up being used. The above per-RestrictInfo caching would
5001 : * not mix well with trying to re-order clauses anyway.
5002 : *
5003 : * Another issue that is entirely ignored here is that if a set-returning
5004 : * function is below top level in the tree, the functions/operators above
5005 : * it will need to be evaluated multiple times. In practical use, such
5006 : * cases arise so seldom as to not be worth the added complexity needed;
5007 : * moreover, since our rowcount estimates for functions tend to be pretty
5008 : * phony, the results would also be pretty phony.
5009 : */
5010 5144196 : if (IsA(node, FuncExpr))
5011 : {
5012 349876 : add_function_cost(context->root, ((FuncExpr *) node)->funcid, node,
5013 : &context->total);
5014 : }
5015 4794320 : else if (IsA(node, OpExpr) ||
5016 4123182 : IsA(node, DistinctExpr) ||
5017 4122090 : IsA(node, NullIfExpr))
5018 : {
5019 : /* rely on struct equivalence to treat these all alike */
5020 672354 : set_opfuncid((OpExpr *) node);
5021 672354 : add_function_cost(context->root, ((OpExpr *) node)->opfuncid, node,
5022 : &context->total);
5023 : }
5024 4121966 : else if (IsA(node, ScalarArrayOpExpr))
5025 : {
5026 45122 : ScalarArrayOpExpr *saop = (ScalarArrayOpExpr *) node;
5027 45122 : Node *arraynode = (Node *) lsecond(saop->args);
5028 : QualCost sacosts;
5029 : QualCost hcosts;
5030 45122 : double estarraylen = estimate_array_length(context->root, arraynode);
5031 :
5032 45122 : set_sa_opfuncid(saop);
5033 45122 : sacosts.startup = sacosts.per_tuple = 0;
5034 45122 : add_function_cost(context->root, saop->opfuncid, NULL,
5035 : &sacosts);
5036 :
5037 45122 : if (OidIsValid(saop->hashfuncid))
5038 : {
5039 : /* Handle costs for hashed ScalarArrayOpExpr */
5040 430 : hcosts.startup = hcosts.per_tuple = 0;
5041 :
5042 430 : add_function_cost(context->root, saop->hashfuncid, NULL, &hcosts);
5043 430 : context->total.startup += sacosts.startup + hcosts.startup;
5044 :
5045 : /* Estimate the cost of building the hashtable. */
5046 430 : context->total.startup += estarraylen * hcosts.per_tuple;
5047 :
5048 : /*
5049 : * XXX should we charge a little bit for sacosts.per_tuple when
5050 : * building the table, or is it ok to assume there will be zero
5051 : * hash collision?
5052 : */
5053 :
5054 : /*
5055 : * Charge for hashtable lookups. Charge a single hash and a
5056 : * single comparison.
5057 : */
5058 430 : context->total.per_tuple += hcosts.per_tuple + sacosts.per_tuple;
5059 : }
5060 : else
5061 : {
5062 : /*
5063 : * Estimate that the operator will be applied to about half of the
5064 : * array elements before the answer is determined.
5065 : */
5066 44692 : context->total.startup += sacosts.startup;
5067 89384 : context->total.per_tuple += sacosts.per_tuple *
5068 44692 : estimate_array_length(context->root, arraynode) * 0.5;
5069 : }
5070 : }
5071 4076844 : else if (IsA(node, Aggref) ||
5072 4003876 : IsA(node, WindowFunc))
5073 : {
5074 : /*
5075 : * Aggref and WindowFunc nodes are (and should be) treated like Vars,
5076 : * ie, zero execution cost in the current model, because they behave
5077 : * essentially like Vars at execution. We disregard the costs of
5078 : * their input expressions for the same reason. The actual execution
5079 : * costs of the aggregate/window functions and their arguments have to
5080 : * be factored into plan-node-specific costing of the Agg or WindowAgg
5081 : * plan node.
5082 : */
5083 77048 : return false; /* don't recurse into children */
5084 : }
5085 3999796 : else if (IsA(node, GroupingFunc))
5086 : {
5087 : /* Treat this as having cost 1 */
5088 430 : context->total.per_tuple += cpu_operator_cost;
5089 430 : return false; /* don't recurse into children */
5090 : }
5091 3999366 : else if (IsA(node, CoerceViaIO))
5092 : {
5093 22824 : CoerceViaIO *iocoerce = (CoerceViaIO *) node;
5094 : Oid iofunc;
5095 : Oid typioparam;
5096 : bool typisvarlena;
5097 :
5098 : /* check the result type's input function */
5099 22824 : getTypeInputInfo(iocoerce->resulttype,
5100 : &iofunc, &typioparam);
5101 22824 : add_function_cost(context->root, iofunc, NULL,
5102 : &context->total);
5103 : /* check the input type's output function */
5104 22824 : getTypeOutputInfo(exprType((Node *) iocoerce->arg),
5105 : &iofunc, &typisvarlena);
5106 22824 : add_function_cost(context->root, iofunc, NULL,
5107 : &context->total);
5108 : }
5109 3976542 : else if (IsA(node, ArrayCoerceExpr))
5110 : {
5111 5206 : ArrayCoerceExpr *acoerce = (ArrayCoerceExpr *) node;
5112 : QualCost perelemcost;
5113 :
5114 5206 : cost_qual_eval_node(&perelemcost, (Node *) acoerce->elemexpr,
5115 : context->root);
5116 5206 : context->total.startup += perelemcost.startup;
5117 5206 : if (perelemcost.per_tuple > 0)
5118 66 : context->total.per_tuple += perelemcost.per_tuple *
5119 66 : estimate_array_length(context->root, (Node *) acoerce->arg);
5120 : }
5121 3971336 : else if (IsA(node, RowCompareExpr))
5122 : {
5123 : /* Conservatively assume we will check all the columns */
5124 252 : RowCompareExpr *rcexpr = (RowCompareExpr *) node;
5125 : ListCell *lc;
5126 :
5127 810 : foreach(lc, rcexpr->opnos)
5128 : {
5129 558 : Oid opid = lfirst_oid(lc);
5130 :
5131 558 : add_function_cost(context->root, get_opcode(opid), NULL,
5132 : &context->total);
5133 : }
5134 : }
5135 3971084 : else if (IsA(node, MinMaxExpr) ||
5136 3970806 : IsA(node, SQLValueFunction) ||
5137 3966030 : IsA(node, XmlExpr) ||
5138 3965328 : IsA(node, CoerceToDomain) ||
5139 3955470 : IsA(node, NextValueExpr) ||
5140 3955072 : IsA(node, JsonExpr))
5141 : {
5142 : /* Treat all these as having cost 1 */
5143 18584 : context->total.per_tuple += cpu_operator_cost;
5144 : }
5145 3952500 : else if (IsA(node, SubLink))
5146 : {
5147 : /* This routine should not be applied to un-planned expressions */
5148 0 : elog(ERROR, "cannot handle unplanned sub-select");
5149 : }
5150 3952500 : else if (IsA(node, SubPlan))
5151 : {
5152 : /*
5153 : * A subplan node in an expression typically indicates that the
5154 : * subplan will be executed on each evaluation, so charge accordingly.
5155 : * (Sub-selects that can be executed as InitPlans have already been
5156 : * removed from the expression.)
5157 : */
5158 43686 : SubPlan *subplan = (SubPlan *) node;
5159 :
5160 43686 : context->total.startup += subplan->startup_cost;
5161 43686 : context->total.per_tuple += subplan->per_call_cost;
5162 :
5163 : /*
5164 : * We don't want to recurse into the testexpr, because it was already
5165 : * counted in the SubPlan node's costs. So we're done.
5166 : */
5167 43686 : return false;
5168 : }
5169 3908814 : else if (IsA(node, AlternativeSubPlan))
5170 : {
5171 : /*
5172 : * Arbitrarily use the first alternative plan for costing. (We should
5173 : * certainly only include one alternative, and we don't yet have
5174 : * enough information to know which one the executor is most likely to
5175 : * use.)
5176 : */
5177 1858 : AlternativeSubPlan *asplan = (AlternativeSubPlan *) node;
5178 :
5179 1858 : return cost_qual_eval_walker((Node *) linitial(asplan->subplans),
5180 : context);
5181 : }
5182 3906956 : else if (IsA(node, PlaceHolderVar))
5183 : {
5184 : /*
5185 : * A PlaceHolderVar should be given cost zero when considering general
5186 : * expression evaluation costs. The expense of doing the contained
5187 : * expression is charged as part of the tlist eval costs of the scan
5188 : * or join where the PHV is first computed (see set_rel_width and
5189 : * add_placeholders_to_joinrel). If we charged it again here, we'd be
5190 : * double-counting the cost for each level of plan that the PHV
5191 : * bubbles up through. Hence, return without recursing into the
5192 : * phexpr.
5193 : */
5194 5636 : return false;
5195 : }
5196 :
5197 : /* recurse into children */
5198 5015538 : return expression_tree_walker(node, cost_qual_eval_walker, context);
5199 : }
5200 :
5201 : /*
5202 : * get_restriction_qual_cost
5203 : * Compute evaluation costs of a baserel's restriction quals, plus any
5204 : * movable join quals that have been pushed down to the scan.
5205 : * Results are returned into *qpqual_cost.
5206 : *
5207 : * This is a convenience subroutine that works for seqscans and other cases
5208 : * where all the given quals will be evaluated the hard way. It's not useful
5209 : * for cost_index(), for example, where the index machinery takes care of
5210 : * some of the quals. We assume baserestrictcost was previously set by
5211 : * set_baserel_size_estimates().
5212 : */
5213 : static void
5214 1126724 : get_restriction_qual_cost(PlannerInfo *root, RelOptInfo *baserel,
5215 : ParamPathInfo *param_info,
5216 : QualCost *qpqual_cost)
5217 : {
5218 1126724 : if (param_info)
5219 : {
5220 : /* Include costs of pushed-down clauses */
5221 252012 : cost_qual_eval(qpqual_cost, param_info->ppi_clauses, root);
5222 :
5223 252012 : qpqual_cost->startup += baserel->baserestrictcost.startup;
5224 252012 : qpqual_cost->per_tuple += baserel->baserestrictcost.per_tuple;
5225 : }
5226 : else
5227 874712 : *qpqual_cost = baserel->baserestrictcost;
5228 1126724 : }
5229 :
5230 :
5231 : /*
5232 : * compute_semi_anti_join_factors
5233 : * Estimate how much of the inner input a SEMI, ANTI, or inner_unique join
5234 : * can be expected to scan.
5235 : *
5236 : * In a hash or nestloop SEMI/ANTI join, the executor will stop scanning
5237 : * inner rows as soon as it finds a match to the current outer row.
5238 : * The same happens if we have detected the inner rel is unique.
5239 : * We should therefore adjust some of the cost components for this effect.
5240 : * This function computes some estimates needed for these adjustments.
5241 : * These estimates will be the same regardless of the particular paths used
5242 : * for the outer and inner relation, so we compute these once and then pass
5243 : * them to all the join cost estimation functions.
5244 : *
5245 : * Input parameters:
5246 : * joinrel: join relation under consideration
5247 : * outerrel: outer relation under consideration
5248 : * innerrel: inner relation under consideration
5249 : * jointype: if not JOIN_SEMI or JOIN_ANTI, we assume it's inner_unique
5250 : * sjinfo: SpecialJoinInfo relevant to this join
5251 : * restrictlist: join quals
5252 : * Output parameters:
5253 : * *semifactors is filled in (see pathnodes.h for field definitions)
5254 : */
5255 : void
5256 223782 : compute_semi_anti_join_factors(PlannerInfo *root,
5257 : RelOptInfo *joinrel,
5258 : RelOptInfo *outerrel,
5259 : RelOptInfo *innerrel,
5260 : JoinType jointype,
5261 : SpecialJoinInfo *sjinfo,
5262 : List *restrictlist,
5263 : SemiAntiJoinFactors *semifactors)
5264 : {
5265 : Selectivity jselec;
5266 : Selectivity nselec;
5267 : Selectivity avgmatch;
5268 : SpecialJoinInfo norm_sjinfo;
5269 : List *joinquals;
5270 : ListCell *l;
5271 :
5272 : /*
5273 : * In an ANTI join, we must ignore clauses that are "pushed down", since
5274 : * those won't affect the match logic. In a SEMI join, we do not
5275 : * distinguish joinquals from "pushed down" quals, so just use the whole
5276 : * restrictinfo list. For other outer join types, we should consider only
5277 : * non-pushed-down quals, so that this devolves to an IS_OUTER_JOIN check.
5278 : */
5279 223782 : if (IS_OUTER_JOIN(jointype))
5280 : {
5281 77434 : joinquals = NIL;
5282 169106 : foreach(l, restrictlist)
5283 : {
5284 91672 : RestrictInfo *rinfo = lfirst_node(RestrictInfo, l);
5285 :
5286 91672 : if (!RINFO_IS_PUSHED_DOWN(rinfo, joinrel->relids))
5287 86796 : joinquals = lappend(joinquals, rinfo);
5288 : }
5289 : }
5290 : else
5291 146348 : joinquals = restrictlist;
5292 :
5293 : /*
5294 : * Get the JOIN_SEMI or JOIN_ANTI selectivity of the join clauses.
5295 : */
5296 223782 : jselec = clauselist_selectivity(root,
5297 : joinquals,
5298 : 0,
5299 : (jointype == JOIN_ANTI) ? JOIN_ANTI : JOIN_SEMI,
5300 : sjinfo);
5301 :
5302 : /*
5303 : * Also get the normal inner-join selectivity of the join clauses.
5304 : */
5305 223782 : init_dummy_sjinfo(&norm_sjinfo, outerrel->relids, innerrel->relids);
5306 :
5307 223782 : nselec = clauselist_selectivity(root,
5308 : joinquals,
5309 : 0,
5310 : JOIN_INNER,
5311 : &norm_sjinfo);
5312 :
5313 : /* Avoid leaking a lot of ListCells */
5314 223782 : if (IS_OUTER_JOIN(jointype))
5315 77434 : list_free(joinquals);
5316 :
5317 : /*
5318 : * jselec can be interpreted as the fraction of outer-rel rows that have
5319 : * any matches (this is true for both SEMI and ANTI cases). And nselec is
5320 : * the fraction of the Cartesian product that matches. So, the average
5321 : * number of matches for each outer-rel row that has at least one match is
5322 : * nselec * inner_rows / jselec.
5323 : *
5324 : * Note: it is correct to use the inner rel's "rows" count here, even
5325 : * though we might later be considering a parameterized inner path with
5326 : * fewer rows. This is because we have included all the join clauses in
5327 : * the selectivity estimate.
5328 : */
5329 223782 : if (jselec > 0) /* protect against zero divide */
5330 : {
5331 223406 : avgmatch = nselec * innerrel->rows / jselec;
5332 : /* Clamp to sane range */
5333 223406 : avgmatch = Max(1.0, avgmatch);
5334 : }
5335 : else
5336 376 : avgmatch = 1.0;
5337 :
5338 223782 : semifactors->outer_match_frac = jselec;
5339 223782 : semifactors->match_count = avgmatch;
5340 223782 : }
5341 :
5342 : /*
5343 : * has_indexed_join_quals
5344 : * Check whether all the joinquals of a nestloop join are used as
5345 : * inner index quals.
5346 : *
5347 : * If the inner path of a SEMI/ANTI join is an indexscan (including bitmap
5348 : * indexscan) that uses all the joinquals as indexquals, we can assume that an
5349 : * unmatched outer tuple is cheap to process, whereas otherwise it's probably
5350 : * expensive.
5351 : */
5352 : static bool
5353 952342 : has_indexed_join_quals(NestPath *path)
5354 : {
5355 952342 : JoinPath *joinpath = &path->jpath;
5356 952342 : Relids joinrelids = joinpath->path.parent->relids;
5357 952342 : Path *innerpath = joinpath->innerjoinpath;
5358 : List *indexclauses;
5359 : bool found_one;
5360 : ListCell *lc;
5361 :
5362 : /* If join still has quals to evaluate, it's not fast */
5363 952342 : if (joinpath->joinrestrictinfo != NIL)
5364 678566 : return false;
5365 : /* Nor if the inner path isn't parameterized at all */
5366 273776 : if (innerpath->param_info == NULL)
5367 3300 : return false;
5368 :
5369 : /* Find the indexclauses list for the inner scan */
5370 270476 : switch (innerpath->pathtype)
5371 : {
5372 166002 : case T_IndexScan:
5373 : case T_IndexOnlyScan:
5374 166002 : indexclauses = ((IndexPath *) innerpath)->indexclauses;
5375 166002 : break;
5376 270 : case T_BitmapHeapScan:
5377 : {
5378 : /* Accept only a simple bitmap scan, not AND/OR cases */
5379 270 : Path *bmqual = ((BitmapHeapPath *) innerpath)->bitmapqual;
5380 :
5381 270 : if (IsA(bmqual, IndexPath))
5382 222 : indexclauses = ((IndexPath *) bmqual)->indexclauses;
5383 : else
5384 48 : return false;
5385 222 : break;
5386 : }
5387 104204 : default:
5388 :
5389 : /*
5390 : * If it's not a simple indexscan, it probably doesn't run quickly
5391 : * for zero rows out, even if it's a parameterized path using all
5392 : * the joinquals.
5393 : */
5394 104204 : return false;
5395 : }
5396 :
5397 : /*
5398 : * Examine the inner path's param clauses. Any that are from the outer
5399 : * path must be found in the indexclauses list, either exactly or in an
5400 : * equivalent form generated by equivclass.c. Also, we must find at least
5401 : * one such clause, else it's a clauseless join which isn't fast.
5402 : */
5403 166224 : found_one = false;
5404 328274 : foreach(lc, innerpath->param_info->ppi_clauses)
5405 : {
5406 170822 : RestrictInfo *rinfo = (RestrictInfo *) lfirst(lc);
5407 :
5408 170822 : if (join_clause_is_movable_into(rinfo,
5409 170822 : innerpath->parent->relids,
5410 : joinrelids))
5411 : {
5412 170270 : if (!is_redundant_with_indexclauses(rinfo, indexclauses))
5413 8772 : return false;
5414 161498 : found_one = true;
5415 : }
5416 : }
5417 157452 : return found_one;
5418 : }
5419 :
5420 :
5421 : /*
5422 : * approx_tuple_count
5423 : * Quick-and-dirty estimation of the number of join rows passing
5424 : * a set of qual conditions.
5425 : *
5426 : * The quals can be either an implicitly-ANDed list of boolean expressions,
5427 : * or a list of RestrictInfo nodes (typically the latter).
5428 : *
5429 : * We intentionally compute the selectivity under JOIN_INNER rules, even
5430 : * if it's some type of outer join. This is appropriate because we are
5431 : * trying to figure out how many tuples pass the initial merge or hash
5432 : * join step.
5433 : *
5434 : * This is quick-and-dirty because we bypass clauselist_selectivity, and
5435 : * simply multiply the independent clause selectivities together. Now
5436 : * clauselist_selectivity often can't do any better than that anyhow, but
5437 : * for some situations (such as range constraints) it is smarter. However,
5438 : * we can't effectively cache the results of clauselist_selectivity, whereas
5439 : * the individual clause selectivities can be and are cached.
5440 : *
5441 : * Since we are only using the results to estimate how many potential
5442 : * output tuples are generated and passed through qpqual checking, it
5443 : * seems OK to live with the approximation.
5444 : */
5445 : static double
5446 806950 : approx_tuple_count(PlannerInfo *root, JoinPath *path, List *quals)
5447 : {
5448 : double tuples;
5449 806950 : double outer_tuples = path->outerjoinpath->rows;
5450 806950 : double inner_tuples = path->innerjoinpath->rows;
5451 : SpecialJoinInfo sjinfo;
5452 806950 : Selectivity selec = 1.0;
5453 : ListCell *l;
5454 :
5455 : /*
5456 : * Make up a SpecialJoinInfo for JOIN_INNER semantics.
5457 : */
5458 806950 : init_dummy_sjinfo(&sjinfo, path->outerjoinpath->parent->relids,
5459 806950 : path->innerjoinpath->parent->relids);
5460 :
5461 : /* Get the approximate selectivity */
5462 1704266 : foreach(l, quals)
5463 : {
5464 897316 : Node *qual = (Node *) lfirst(l);
5465 :
5466 : /* Note that clause_selectivity will be able to cache its result */
5467 897316 : selec *= clause_selectivity(root, qual, 0, JOIN_INNER, &sjinfo);
5468 : }
5469 :
5470 : /* Apply it to the input relation sizes */
5471 806950 : tuples = selec * outer_tuples * inner_tuples;
5472 :
5473 806950 : return clamp_row_est(tuples);
5474 : }
5475 :
5476 :
5477 : /*
5478 : * set_baserel_size_estimates
5479 : * Set the size estimates for the given base relation.
5480 : *
5481 : * The rel's targetlist and restrictinfo list must have been constructed
5482 : * already, and rel->tuples must be set.
5483 : *
5484 : * We set the following fields of the rel node:
5485 : * rows: the estimated number of output tuples (after applying
5486 : * restriction clauses).
5487 : * width: the estimated average output tuple width in bytes.
5488 : * baserestrictcost: estimated cost of evaluating baserestrictinfo clauses.
5489 : */
5490 : void
5491 534342 : set_baserel_size_estimates(PlannerInfo *root, RelOptInfo *rel)
5492 : {
5493 : double nrows;
5494 :
5495 : /* Should only be applied to base relations */
5496 : Assert(rel->relid > 0);
5497 :
5498 1068654 : nrows = rel->tuples *
5499 534342 : clauselist_selectivity(root,
5500 : rel->baserestrictinfo,
5501 : 0,
5502 : JOIN_INNER,
5503 : NULL);
5504 :
5505 534312 : rel->rows = clamp_row_est(nrows);
5506 :
5507 534312 : cost_qual_eval(&rel->baserestrictcost, rel->baserestrictinfo, root);
5508 :
5509 534312 : set_rel_width(root, rel);
5510 534312 : }
5511 :
5512 : /*
5513 : * get_parameterized_baserel_size
5514 : * Make a size estimate for a parameterized scan of a base relation.
5515 : *
5516 : * 'param_clauses' lists the additional join clauses to be used.
5517 : *
5518 : * set_baserel_size_estimates must have been applied already.
5519 : */
5520 : double
5521 165084 : get_parameterized_baserel_size(PlannerInfo *root, RelOptInfo *rel,
5522 : List *param_clauses)
5523 : {
5524 : List *allclauses;
5525 : double nrows;
5526 :
5527 : /*
5528 : * Estimate the number of rows returned by the parameterized scan, knowing
5529 : * that it will apply all the extra join clauses as well as the rel's own
5530 : * restriction clauses. Note that we force the clauses to be treated as
5531 : * non-join clauses during selectivity estimation.
5532 : */
5533 165084 : allclauses = list_concat_copy(param_clauses, rel->baserestrictinfo);
5534 330168 : nrows = rel->tuples *
5535 165084 : clauselist_selectivity(root,
5536 : allclauses,
5537 165084 : rel->relid, /* do not use 0! */
5538 : JOIN_INNER,
5539 : NULL);
5540 165084 : nrows = clamp_row_est(nrows);
5541 : /* For safety, make sure result is not more than the base estimate */
5542 165084 : if (nrows > rel->rows)
5543 0 : nrows = rel->rows;
5544 165084 : return nrows;
5545 : }
5546 :
5547 : /*
5548 : * set_joinrel_size_estimates
5549 : * Set the size estimates for the given join relation.
5550 : *
5551 : * The rel's targetlist must have been constructed already, and a
5552 : * restriction clause list that matches the given component rels must
5553 : * be provided.
5554 : *
5555 : * Since there is more than one way to make a joinrel for more than two
5556 : * base relations, the results we get here could depend on which component
5557 : * rel pair is provided. In theory we should get the same answers no matter
5558 : * which pair is provided; in practice, since the selectivity estimation
5559 : * routines don't handle all cases equally well, we might not. But there's
5560 : * not much to be done about it. (Would it make sense to repeat the
5561 : * calculations for each pair of input rels that's encountered, and somehow
5562 : * average the results? Probably way more trouble than it's worth, and
5563 : * anyway we must keep the rowcount estimate the same for all paths for the
5564 : * joinrel.)
5565 : *
5566 : * We set only the rows field here. The reltarget field was already set by
5567 : * build_joinrel_tlist, and baserestrictcost is not used for join rels.
5568 : */
5569 : void
5570 261688 : set_joinrel_size_estimates(PlannerInfo *root, RelOptInfo *rel,
5571 : RelOptInfo *outer_rel,
5572 : RelOptInfo *inner_rel,
5573 : SpecialJoinInfo *sjinfo,
5574 : List *restrictlist)
5575 : {
5576 261688 : rel->rows = calc_joinrel_size_estimate(root,
5577 : rel,
5578 : outer_rel,
5579 : inner_rel,
5580 : outer_rel->rows,
5581 : inner_rel->rows,
5582 : sjinfo,
5583 : restrictlist);
5584 261688 : }
5585 :
5586 : /*
5587 : * get_parameterized_joinrel_size
5588 : * Make a size estimate for a parameterized scan of a join relation.
5589 : *
5590 : * 'rel' is the joinrel under consideration.
5591 : * 'outer_path', 'inner_path' are (probably also parameterized) Paths that
5592 : * produce the relations being joined.
5593 : * 'sjinfo' is any SpecialJoinInfo relevant to this join.
5594 : * 'restrict_clauses' lists the join clauses that need to be applied at the
5595 : * join node (including any movable clauses that were moved down to this join,
5596 : * and not including any movable clauses that were pushed down into the
5597 : * child paths).
5598 : *
5599 : * set_joinrel_size_estimates must have been applied already.
5600 : */
5601 : double
5602 9718 : get_parameterized_joinrel_size(PlannerInfo *root, RelOptInfo *rel,
5603 : Path *outer_path,
5604 : Path *inner_path,
5605 : SpecialJoinInfo *sjinfo,
5606 : List *restrict_clauses)
5607 : {
5608 : double nrows;
5609 :
5610 : /*
5611 : * Estimate the number of rows returned by the parameterized join as the
5612 : * sizes of the input paths times the selectivity of the clauses that have
5613 : * ended up at this join node.
5614 : *
5615 : * As with set_joinrel_size_estimates, the rowcount estimate could depend
5616 : * on the pair of input paths provided, though ideally we'd get the same
5617 : * estimate for any pair with the same parameterization.
5618 : */
5619 9718 : nrows = calc_joinrel_size_estimate(root,
5620 : rel,
5621 : outer_path->parent,
5622 : inner_path->parent,
5623 : outer_path->rows,
5624 : inner_path->rows,
5625 : sjinfo,
5626 : restrict_clauses);
5627 : /* For safety, make sure result is not more than the base estimate */
5628 9718 : if (nrows > rel->rows)
5629 12 : nrows = rel->rows;
5630 9718 : return nrows;
5631 : }
5632 :
5633 : /*
5634 : * calc_joinrel_size_estimate
5635 : * Workhorse for set_joinrel_size_estimates and
5636 : * get_parameterized_joinrel_size.
5637 : *
5638 : * outer_rel/inner_rel are the relations being joined, but they should be
5639 : * assumed to have sizes outer_rows/inner_rows; those numbers might be less
5640 : * than what rel->rows says, when we are considering parameterized paths.
5641 : */
5642 : static double
5643 271406 : calc_joinrel_size_estimate(PlannerInfo *root,
5644 : RelOptInfo *joinrel,
5645 : RelOptInfo *outer_rel,
5646 : RelOptInfo *inner_rel,
5647 : double outer_rows,
5648 : double inner_rows,
5649 : SpecialJoinInfo *sjinfo,
5650 : List *restrictlist)
5651 : {
5652 271406 : JoinType jointype = sjinfo->jointype;
5653 : Selectivity fkselec;
5654 : Selectivity jselec;
5655 : Selectivity pselec;
5656 : double nrows;
5657 :
5658 : /*
5659 : * Compute joinclause selectivity. Note that we are only considering
5660 : * clauses that become restriction clauses at this join level; we are not
5661 : * double-counting them because they were not considered in estimating the
5662 : * sizes of the component rels.
5663 : *
5664 : * First, see whether any of the joinclauses can be matched to known FK
5665 : * constraints. If so, drop those clauses from the restrictlist, and
5666 : * instead estimate their selectivity using FK semantics. (We do this
5667 : * without regard to whether said clauses are local or "pushed down".
5668 : * Probably, an FK-matching clause could never be seen as pushed down at
5669 : * an outer join, since it would be strict and hence would be grounds for
5670 : * join strength reduction.) fkselec gets the net selectivity for
5671 : * FK-matching clauses, or 1.0 if there are none.
5672 : */
5673 271406 : fkselec = get_foreign_key_join_selectivity(root,
5674 : outer_rel->relids,
5675 : inner_rel->relids,
5676 : sjinfo,
5677 : &restrictlist);
5678 :
5679 : /*
5680 : * For an outer join, we have to distinguish the selectivity of the join's
5681 : * own clauses (JOIN/ON conditions) from any clauses that were "pushed
5682 : * down". For inner joins we just count them all as joinclauses.
5683 : */
5684 271406 : if (IS_OUTER_JOIN(jointype))
5685 : {
5686 82718 : List *joinquals = NIL;
5687 82718 : List *pushedquals = NIL;
5688 : ListCell *l;
5689 :
5690 : /* Grovel through the clauses to separate into two lists */
5691 185910 : foreach(l, restrictlist)
5692 : {
5693 103192 : RestrictInfo *rinfo = lfirst_node(RestrictInfo, l);
5694 :
5695 103192 : if (RINFO_IS_PUSHED_DOWN(rinfo, joinrel->relids))
5696 4284 : pushedquals = lappend(pushedquals, rinfo);
5697 : else
5698 98908 : joinquals = lappend(joinquals, rinfo);
5699 : }
5700 :
5701 : /* Get the separate selectivities */
5702 82718 : jselec = clauselist_selectivity(root,
5703 : joinquals,
5704 : 0,
5705 : jointype,
5706 : sjinfo);
5707 82718 : pselec = clauselist_selectivity(root,
5708 : pushedquals,
5709 : 0,
5710 : jointype,
5711 : sjinfo);
5712 :
5713 : /* Avoid leaking a lot of ListCells */
5714 82718 : list_free(joinquals);
5715 82718 : list_free(pushedquals);
5716 : }
5717 : else
5718 : {
5719 188688 : jselec = clauselist_selectivity(root,
5720 : restrictlist,
5721 : 0,
5722 : jointype,
5723 : sjinfo);
5724 188688 : pselec = 0.0; /* not used, keep compiler quiet */
5725 : }
5726 :
5727 : /*
5728 : * Basically, we multiply size of Cartesian product by selectivity.
5729 : *
5730 : * If we are doing an outer join, take that into account: the joinqual
5731 : * selectivity has to be clamped using the knowledge that the output must
5732 : * be at least as large as the non-nullable input. However, any
5733 : * pushed-down quals are applied after the outer join, so their
5734 : * selectivity applies fully.
5735 : *
5736 : * For JOIN_SEMI and JOIN_ANTI, the selectivity is defined as the fraction
5737 : * of LHS rows that have matches, and we apply that straightforwardly.
5738 : */
5739 271406 : switch (jointype)
5740 : {
5741 180452 : case JOIN_INNER:
5742 180452 : nrows = outer_rows * inner_rows * fkselec * jselec;
5743 : /* pselec not used */
5744 180452 : break;
5745 75928 : case JOIN_LEFT:
5746 75928 : nrows = outer_rows * inner_rows * fkselec * jselec;
5747 75928 : if (nrows < outer_rows)
5748 30462 : nrows = outer_rows;
5749 75928 : nrows *= pselec;
5750 75928 : break;
5751 1720 : case JOIN_FULL:
5752 1720 : nrows = outer_rows * inner_rows * fkselec * jselec;
5753 1720 : if (nrows < outer_rows)
5754 1178 : nrows = outer_rows;
5755 1720 : if (nrows < inner_rows)
5756 120 : nrows = inner_rows;
5757 1720 : nrows *= pselec;
5758 1720 : break;
5759 8236 : case JOIN_SEMI:
5760 8236 : nrows = outer_rows * fkselec * jselec;
5761 : /* pselec not used */
5762 8236 : break;
5763 5070 : case JOIN_ANTI:
5764 5070 : nrows = outer_rows * (1.0 - fkselec * jselec);
5765 5070 : nrows *= pselec;
5766 5070 : break;
5767 0 : default:
5768 : /* other values not expected here */
5769 0 : elog(ERROR, "unrecognized join type: %d", (int) jointype);
5770 : nrows = 0; /* keep compiler quiet */
5771 : break;
5772 : }
5773 :
5774 271406 : return clamp_row_est(nrows);
5775 : }
5776 :
5777 : /*
5778 : * get_foreign_key_join_selectivity
5779 : * Estimate join selectivity for foreign-key-related clauses.
5780 : *
5781 : * Remove any clauses that can be matched to FK constraints from *restrictlist,
5782 : * and return a substitute estimate of their selectivity. 1.0 is returned
5783 : * when there are no such clauses.
5784 : *
5785 : * The reason for treating such clauses specially is that we can get better
5786 : * estimates this way than by relying on clauselist_selectivity(), especially
5787 : * for multi-column FKs where that function's assumption that the clauses are
5788 : * independent falls down badly. But even with single-column FKs, we may be
5789 : * able to get a better answer when the pg_statistic stats are missing or out
5790 : * of date.
5791 : */
5792 : static Selectivity
5793 271406 : get_foreign_key_join_selectivity(PlannerInfo *root,
5794 : Relids outer_relids,
5795 : Relids inner_relids,
5796 : SpecialJoinInfo *sjinfo,
5797 : List **restrictlist)
5798 : {
5799 271406 : Selectivity fkselec = 1.0;
5800 271406 : JoinType jointype = sjinfo->jointype;
5801 271406 : List *worklist = *restrictlist;
5802 : ListCell *lc;
5803 :
5804 : /* Consider each FK constraint that is known to match the query */
5805 273372 : foreach(lc, root->fkey_list)
5806 : {
5807 1966 : ForeignKeyOptInfo *fkinfo = (ForeignKeyOptInfo *) lfirst(lc);
5808 : bool ref_is_outer;
5809 : List *removedlist;
5810 : ListCell *cell;
5811 :
5812 : /*
5813 : * This FK is not relevant unless it connects a baserel on one side of
5814 : * this join to a baserel on the other side.
5815 : */
5816 3580 : if (bms_is_member(fkinfo->con_relid, outer_relids) &&
5817 1614 : bms_is_member(fkinfo->ref_relid, inner_relids))
5818 1440 : ref_is_outer = false;
5819 866 : else if (bms_is_member(fkinfo->ref_relid, outer_relids) &&
5820 340 : bms_is_member(fkinfo->con_relid, inner_relids))
5821 130 : ref_is_outer = true;
5822 : else
5823 396 : continue;
5824 :
5825 : /*
5826 : * If we're dealing with a semi/anti join, and the FK's referenced
5827 : * relation is on the outside, then knowledge of the FK doesn't help
5828 : * us figure out what we need to know (which is the fraction of outer
5829 : * rows that have matches). On the other hand, if the referenced rel
5830 : * is on the inside, then all outer rows must have matches in the
5831 : * referenced table (ignoring nulls). But any restriction or join
5832 : * clauses that filter that table will reduce the fraction of matches.
5833 : * We can account for restriction clauses, but it's too hard to guess
5834 : * how many table rows would get through a join that's inside the RHS.
5835 : * Hence, if either case applies, punt and ignore the FK.
5836 : */
5837 1570 : if ((jointype == JOIN_SEMI || jointype == JOIN_ANTI) &&
5838 1048 : (ref_is_outer || bms_membership(inner_relids) != BMS_SINGLETON))
5839 12 : continue;
5840 :
5841 : /*
5842 : * Modify the restrictlist by removing clauses that match the FK (and
5843 : * putting them into removedlist instead). It seems unsafe to modify
5844 : * the originally-passed List structure, so we make a shallow copy the
5845 : * first time through.
5846 : */
5847 1558 : if (worklist == *restrictlist)
5848 1334 : worklist = list_copy(worklist);
5849 :
5850 1558 : removedlist = NIL;
5851 3252 : foreach(cell, worklist)
5852 : {
5853 1694 : RestrictInfo *rinfo = (RestrictInfo *) lfirst(cell);
5854 1694 : bool remove_it = false;
5855 : int i;
5856 :
5857 : /* Drop this clause if it matches any column of the FK */
5858 2140 : for (i = 0; i < fkinfo->nkeys; i++)
5859 : {
5860 2110 : if (rinfo->parent_ec)
5861 : {
5862 : /*
5863 : * EC-derived clauses can only match by EC. It is okay to
5864 : * consider any clause derived from the same EC as
5865 : * matching the FK: even if equivclass.c chose to generate
5866 : * a clause equating some other pair of Vars, it could
5867 : * have generated one equating the FK's Vars. So for
5868 : * purposes of estimation, we can act as though it did so.
5869 : *
5870 : * Note: checking parent_ec is a bit of a cheat because
5871 : * there are EC-derived clauses that don't have parent_ec
5872 : * set; but such clauses must compare expressions that
5873 : * aren't just Vars, so they cannot match the FK anyway.
5874 : */
5875 304 : if (fkinfo->eclass[i] == rinfo->parent_ec)
5876 : {
5877 298 : remove_it = true;
5878 298 : break;
5879 : }
5880 : }
5881 : else
5882 : {
5883 : /*
5884 : * Otherwise, see if rinfo was previously matched to FK as
5885 : * a "loose" clause.
5886 : */
5887 1806 : if (list_member_ptr(fkinfo->rinfos[i], rinfo))
5888 : {
5889 1366 : remove_it = true;
5890 1366 : break;
5891 : }
5892 : }
5893 : }
5894 1694 : if (remove_it)
5895 : {
5896 1664 : worklist = foreach_delete_current(worklist, cell);
5897 1664 : removedlist = lappend(removedlist, rinfo);
5898 : }
5899 : }
5900 :
5901 : /*
5902 : * If we failed to remove all the matching clauses we expected to
5903 : * find, chicken out and ignore this FK; applying its selectivity
5904 : * might result in double-counting. Put any clauses we did manage to
5905 : * remove back into the worklist.
5906 : *
5907 : * Since the matching clauses are known not outerjoin-delayed, they
5908 : * would normally have appeared in the initial joinclause list. If we
5909 : * didn't find them, there are two possibilities:
5910 : *
5911 : * 1. If the FK match is based on an EC that is ec_has_const, it won't
5912 : * have generated any join clauses at all. We discount such ECs while
5913 : * checking to see if we have "all" the clauses. (Below, we'll adjust
5914 : * the selectivity estimate for this case.)
5915 : *
5916 : * 2. The clauses were matched to some other FK in a previous
5917 : * iteration of this loop, and thus removed from worklist. (A likely
5918 : * case is that two FKs are matched to the same EC; there will be only
5919 : * one EC-derived clause in the initial list, so the first FK will
5920 : * consume it.) Applying both FKs' selectivity independently risks
5921 : * underestimating the join size; in particular, this would undo one
5922 : * of the main things that ECs were invented for, namely to avoid
5923 : * double-counting the selectivity of redundant equality conditions.
5924 : * Later we might think of a reasonable way to combine the estimates,
5925 : * but for now, just punt, since this is a fairly uncommon situation.
5926 : */
5927 1558 : if (removedlist == NIL ||
5928 1272 : list_length(removedlist) !=
5929 1272 : (fkinfo->nmatched_ec - fkinfo->nconst_ec + fkinfo->nmatched_ri))
5930 : {
5931 286 : worklist = list_concat(worklist, removedlist);
5932 286 : continue;
5933 : }
5934 :
5935 : /*
5936 : * Finally we get to the payoff: estimate selectivity using the
5937 : * knowledge that each referencing row will match exactly one row in
5938 : * the referenced table.
5939 : *
5940 : * XXX that's not true in the presence of nulls in the referencing
5941 : * column(s), so in principle we should derate the estimate for those.
5942 : * However (1) if there are any strict restriction clauses for the
5943 : * referencing column(s) elsewhere in the query, derating here would
5944 : * be double-counting the null fraction, and (2) it's not very clear
5945 : * how to combine null fractions for multiple referencing columns. So
5946 : * we do nothing for now about correcting for nulls.
5947 : *
5948 : * XXX another point here is that if either side of an FK constraint
5949 : * is an inheritance parent, we estimate as though the constraint
5950 : * covers all its children as well. This is not an unreasonable
5951 : * assumption for a referencing table, ie the user probably applied
5952 : * identical constraints to all child tables (though perhaps we ought
5953 : * to check that). But it's not possible to have done that for a
5954 : * referenced table. Fortunately, precisely because that doesn't
5955 : * work, it is uncommon in practice to have an FK referencing a parent
5956 : * table. So, at least for now, disregard inheritance here.
5957 : */
5958 1272 : if (jointype == JOIN_SEMI || jointype == JOIN_ANTI)
5959 824 : {
5960 : /*
5961 : * For JOIN_SEMI and JOIN_ANTI, we only get here when the FK's
5962 : * referenced table is exactly the inside of the join. The join
5963 : * selectivity is defined as the fraction of LHS rows that have
5964 : * matches. The FK implies that every LHS row has a match *in the
5965 : * referenced table*; but any restriction clauses on it will
5966 : * reduce the number of matches. Hence we take the join
5967 : * selectivity as equal to the selectivity of the table's
5968 : * restriction clauses, which is rows / tuples; but we must guard
5969 : * against tuples == 0.
5970 : */
5971 824 : RelOptInfo *ref_rel = find_base_rel(root, fkinfo->ref_relid);
5972 824 : double ref_tuples = Max(ref_rel->tuples, 1.0);
5973 :
5974 824 : fkselec *= ref_rel->rows / ref_tuples;
5975 : }
5976 : else
5977 : {
5978 : /*
5979 : * Otherwise, selectivity is exactly 1/referenced-table-size; but
5980 : * guard against tuples == 0. Note we should use the raw table
5981 : * tuple count, not any estimate of its filtered or joined size.
5982 : */
5983 448 : RelOptInfo *ref_rel = find_base_rel(root, fkinfo->ref_relid);
5984 448 : double ref_tuples = Max(ref_rel->tuples, 1.0);
5985 :
5986 448 : fkselec *= 1.0 / ref_tuples;
5987 : }
5988 :
5989 : /*
5990 : * If any of the FK columns participated in ec_has_const ECs, then
5991 : * equivclass.c will have generated "var = const" restrictions for
5992 : * each side of the join, thus reducing the sizes of both input
5993 : * relations. Taking the fkselec at face value would amount to
5994 : * double-counting the selectivity of the constant restriction for the
5995 : * referencing Var. Hence, look for the restriction clause(s) that
5996 : * were applied to the referencing Var(s), and divide out their
5997 : * selectivity to correct for this.
5998 : */
5999 1272 : if (fkinfo->nconst_ec > 0)
6000 : {
6001 24 : for (int i = 0; i < fkinfo->nkeys; i++)
6002 : {
6003 18 : EquivalenceClass *ec = fkinfo->eclass[i];
6004 :
6005 18 : if (ec && ec->ec_has_const)
6006 : {
6007 6 : EquivalenceMember *em = fkinfo->fk_eclass_member[i];
6008 6 : RestrictInfo *rinfo = find_derived_clause_for_ec_member(root,
6009 : ec,
6010 : em);
6011 :
6012 6 : if (rinfo)
6013 : {
6014 : Selectivity s0;
6015 :
6016 6 : s0 = clause_selectivity(root,
6017 : (Node *) rinfo,
6018 : 0,
6019 : jointype,
6020 : sjinfo);
6021 6 : if (s0 > 0)
6022 6 : fkselec /= s0;
6023 : }
6024 : }
6025 : }
6026 : }
6027 : }
6028 :
6029 271406 : *restrictlist = worklist;
6030 271406 : CLAMP_PROBABILITY(fkselec);
6031 271406 : return fkselec;
6032 : }
6033 :
6034 : /*
6035 : * set_subquery_size_estimates
6036 : * Set the size estimates for a base relation that is a subquery.
6037 : *
6038 : * The rel's targetlist and restrictinfo list must have been constructed
6039 : * already, and the Paths for the subquery must have been completed.
6040 : * We look at the subquery's PlannerInfo to extract data.
6041 : *
6042 : * We set the same fields as set_baserel_size_estimates.
6043 : */
6044 : void
6045 40274 : set_subquery_size_estimates(PlannerInfo *root, RelOptInfo *rel)
6046 : {
6047 40274 : PlannerInfo *subroot = rel->subroot;
6048 : RelOptInfo *sub_final_rel;
6049 : ListCell *lc;
6050 :
6051 : /* Should only be applied to base relations that are subqueries */
6052 : Assert(rel->relid > 0);
6053 : Assert(planner_rt_fetch(rel->relid, root)->rtekind == RTE_SUBQUERY);
6054 :
6055 : /*
6056 : * Copy raw number of output rows from subquery. All of its paths should
6057 : * have the same output rowcount, so just look at cheapest-total.
6058 : */
6059 40274 : sub_final_rel = fetch_upper_rel(subroot, UPPERREL_FINAL, NULL);
6060 40274 : rel->tuples = sub_final_rel->cheapest_total_path->rows;
6061 :
6062 : /*
6063 : * Compute per-output-column width estimates by examining the subquery's
6064 : * targetlist. For any output that is a plain Var, get the width estimate
6065 : * that was made while planning the subquery. Otherwise, we leave it to
6066 : * set_rel_width to fill in a datatype-based default estimate.
6067 : */
6068 194560 : foreach(lc, subroot->parse->targetList)
6069 : {
6070 154286 : TargetEntry *te = lfirst_node(TargetEntry, lc);
6071 154286 : Node *texpr = (Node *) te->expr;
6072 154286 : int32 item_width = 0;
6073 :
6074 : /* junk columns aren't visible to upper query */
6075 154286 : if (te->resjunk)
6076 6900 : continue;
6077 :
6078 : /*
6079 : * The subquery could be an expansion of a view that's had columns
6080 : * added to it since the current query was parsed, so that there are
6081 : * non-junk tlist columns in it that don't correspond to any column
6082 : * visible at our query level. Ignore such columns.
6083 : */
6084 147386 : if (te->resno < rel->min_attr || te->resno > rel->max_attr)
6085 0 : continue;
6086 :
6087 : /*
6088 : * XXX This currently doesn't work for subqueries containing set
6089 : * operations, because the Vars in their tlists are bogus references
6090 : * to the first leaf subquery, which wouldn't give the right answer
6091 : * even if we could still get to its PlannerInfo.
6092 : *
6093 : * Also, the subquery could be an appendrel for which all branches are
6094 : * known empty due to constraint exclusion, in which case
6095 : * set_append_rel_pathlist will have left the attr_widths set to zero.
6096 : *
6097 : * In either case, we just leave the width estimate zero until
6098 : * set_rel_width fixes it.
6099 : */
6100 147386 : if (IsA(texpr, Var) &&
6101 64584 : subroot->parse->setOperations == NULL)
6102 : {
6103 62744 : Var *var = (Var *) texpr;
6104 62744 : RelOptInfo *subrel = find_base_rel(subroot, var->varno);
6105 :
6106 62744 : item_width = subrel->attr_widths[var->varattno - subrel->min_attr];
6107 : }
6108 147386 : rel->attr_widths[te->resno - rel->min_attr] = item_width;
6109 : }
6110 :
6111 : /* Now estimate number of output rows, etc */
6112 40274 : set_baserel_size_estimates(root, rel);
6113 40274 : }
6114 :
6115 : /*
6116 : * set_function_size_estimates
6117 : * Set the size estimates for a base relation that is a function call.
6118 : *
6119 : * The rel's targetlist and restrictinfo list must have been constructed
6120 : * already.
6121 : *
6122 : * We set the same fields as set_baserel_size_estimates.
6123 : */
6124 : void
6125 53134 : set_function_size_estimates(PlannerInfo *root, RelOptInfo *rel)
6126 : {
6127 : RangeTblEntry *rte;
6128 : ListCell *lc;
6129 :
6130 : /* Should only be applied to base relations that are functions */
6131 : Assert(rel->relid > 0);
6132 53134 : rte = planner_rt_fetch(rel->relid, root);
6133 : Assert(rte->rtekind == RTE_FUNCTION);
6134 :
6135 : /*
6136 : * Estimate number of rows the functions will return. The rowcount of the
6137 : * node is that of the largest function result.
6138 : */
6139 53134 : rel->tuples = 0;
6140 106772 : foreach(lc, rte->functions)
6141 : {
6142 53638 : RangeTblFunction *rtfunc = (RangeTblFunction *) lfirst(lc);
6143 53638 : double ntup = expression_returns_set_rows(root, rtfunc->funcexpr);
6144 :
6145 53638 : if (ntup > rel->tuples)
6146 53158 : rel->tuples = ntup;
6147 : }
6148 :
6149 : /* Now estimate number of output rows, etc */
6150 53134 : set_baserel_size_estimates(root, rel);
6151 53134 : }
6152 :
6153 : /*
6154 : * set_function_size_estimates
6155 : * Set the size estimates for a base relation that is a function call.
6156 : *
6157 : * The rel's targetlist and restrictinfo list must have been constructed
6158 : * already.
6159 : *
6160 : * We set the same fields as set_tablefunc_size_estimates.
6161 : */
6162 : void
6163 626 : set_tablefunc_size_estimates(PlannerInfo *root, RelOptInfo *rel)
6164 : {
6165 : /* Should only be applied to base relations that are functions */
6166 : Assert(rel->relid > 0);
6167 : Assert(planner_rt_fetch(rel->relid, root)->rtekind == RTE_TABLEFUNC);
6168 :
6169 626 : rel->tuples = 100;
6170 :
6171 : /* Now estimate number of output rows, etc */
6172 626 : set_baserel_size_estimates(root, rel);
6173 626 : }
6174 :
6175 : /*
6176 : * set_values_size_estimates
6177 : * Set the size estimates for a base relation that is a values list.
6178 : *
6179 : * The rel's targetlist and restrictinfo list must have been constructed
6180 : * already.
6181 : *
6182 : * We set the same fields as set_baserel_size_estimates.
6183 : */
6184 : void
6185 8536 : set_values_size_estimates(PlannerInfo *root, RelOptInfo *rel)
6186 : {
6187 : RangeTblEntry *rte;
6188 :
6189 : /* Should only be applied to base relations that are values lists */
6190 : Assert(rel->relid > 0);
6191 8536 : rte = planner_rt_fetch(rel->relid, root);
6192 : Assert(rte->rtekind == RTE_VALUES);
6193 :
6194 : /*
6195 : * Estimate number of rows the values list will return. We know this
6196 : * precisely based on the list length (well, barring set-returning
6197 : * functions in list items, but that's a refinement not catered for
6198 : * anywhere else either).
6199 : */
6200 8536 : rel->tuples = list_length(rte->values_lists);
6201 :
6202 : /* Now estimate number of output rows, etc */
6203 8536 : set_baserel_size_estimates(root, rel);
6204 8536 : }
6205 :
6206 : /*
6207 : * set_cte_size_estimates
6208 : * Set the size estimates for a base relation that is a CTE reference.
6209 : *
6210 : * The rel's targetlist and restrictinfo list must have been constructed
6211 : * already, and we need an estimate of the number of rows returned by the CTE
6212 : * (if a regular CTE) or the non-recursive term (if a self-reference).
6213 : *
6214 : * We set the same fields as set_baserel_size_estimates.
6215 : */
6216 : void
6217 5388 : set_cte_size_estimates(PlannerInfo *root, RelOptInfo *rel, double cte_rows)
6218 : {
6219 : RangeTblEntry *rte;
6220 :
6221 : /* Should only be applied to base relations that are CTE references */
6222 : Assert(rel->relid > 0);
6223 5388 : rte = planner_rt_fetch(rel->relid, root);
6224 : Assert(rte->rtekind == RTE_CTE);
6225 :
6226 5388 : if (rte->self_reference)
6227 : {
6228 : /*
6229 : * In a self-reference, we assume the average worktable size is a
6230 : * multiple of the nonrecursive term's size. The best multiplier will
6231 : * vary depending on query "fan-out", so make its value adjustable.
6232 : */
6233 938 : rel->tuples = clamp_row_est(recursive_worktable_factor * cte_rows);
6234 : }
6235 : else
6236 : {
6237 : /* Otherwise just believe the CTE's rowcount estimate */
6238 4450 : rel->tuples = cte_rows;
6239 : }
6240 :
6241 : /* Now estimate number of output rows, etc */
6242 5388 : set_baserel_size_estimates(root, rel);
6243 5388 : }
6244 :
6245 : /*
6246 : * set_namedtuplestore_size_estimates
6247 : * Set the size estimates for a base relation that is a tuplestore reference.
6248 : *
6249 : * The rel's targetlist and restrictinfo list must have been constructed
6250 : * already.
6251 : *
6252 : * We set the same fields as set_baserel_size_estimates.
6253 : */
6254 : void
6255 474 : set_namedtuplestore_size_estimates(PlannerInfo *root, RelOptInfo *rel)
6256 : {
6257 : RangeTblEntry *rte;
6258 :
6259 : /* Should only be applied to base relations that are tuplestore references */
6260 : Assert(rel->relid > 0);
6261 474 : rte = planner_rt_fetch(rel->relid, root);
6262 : Assert(rte->rtekind == RTE_NAMEDTUPLESTORE);
6263 :
6264 : /*
6265 : * Use the estimate provided by the code which is generating the named
6266 : * tuplestore. In some cases, the actual number might be available; in
6267 : * others the same plan will be re-used, so a "typical" value might be
6268 : * estimated and used.
6269 : */
6270 474 : rel->tuples = rte->enrtuples;
6271 474 : if (rel->tuples < 0)
6272 0 : rel->tuples = 1000;
6273 :
6274 : /* Now estimate number of output rows, etc */
6275 474 : set_baserel_size_estimates(root, rel);
6276 474 : }
6277 :
6278 : /*
6279 : * set_result_size_estimates
6280 : * Set the size estimates for an RTE_RESULT base relation
6281 : *
6282 : * The rel's targetlist and restrictinfo list must have been constructed
6283 : * already.
6284 : *
6285 : * We set the same fields as set_baserel_size_estimates.
6286 : */
6287 : void
6288 4292 : set_result_size_estimates(PlannerInfo *root, RelOptInfo *rel)
6289 : {
6290 : /* Should only be applied to RTE_RESULT base relations */
6291 : Assert(rel->relid > 0);
6292 : Assert(planner_rt_fetch(rel->relid, root)->rtekind == RTE_RESULT);
6293 :
6294 : /* RTE_RESULT always generates a single row, natively */
6295 4292 : rel->tuples = 1;
6296 :
6297 : /* Now estimate number of output rows, etc */
6298 4292 : set_baserel_size_estimates(root, rel);
6299 4292 : }
6300 :
6301 : /*
6302 : * set_foreign_size_estimates
6303 : * Set the size estimates for a base relation that is a foreign table.
6304 : *
6305 : * There is not a whole lot that we can do here; the foreign-data wrapper
6306 : * is responsible for producing useful estimates. We can do a decent job
6307 : * of estimating baserestrictcost, so we set that, and we also set up width
6308 : * using what will be purely datatype-driven estimates from the targetlist.
6309 : * There is no way to do anything sane with the rows value, so we just put
6310 : * a default estimate and hope that the wrapper can improve on it. The
6311 : * wrapper's GetForeignRelSize function will be called momentarily.
6312 : *
6313 : * The rel's targetlist and restrictinfo list must have been constructed
6314 : * already.
6315 : */
6316 : void
6317 2468 : set_foreign_size_estimates(PlannerInfo *root, RelOptInfo *rel)
6318 : {
6319 : /* Should only be applied to base relations */
6320 : Assert(rel->relid > 0);
6321 :
6322 2468 : rel->rows = 1000; /* entirely bogus default estimate */
6323 :
6324 2468 : cost_qual_eval(&rel->baserestrictcost, rel->baserestrictinfo, root);
6325 :
6326 2468 : set_rel_width(root, rel);
6327 2468 : }
6328 :
6329 :
6330 : /*
6331 : * set_rel_width
6332 : * Set the estimated output width of a base relation.
6333 : *
6334 : * The estimated output width is the sum of the per-attribute width estimates
6335 : * for the actually-referenced columns, plus any PHVs or other expressions
6336 : * that have to be calculated at this relation. This is the amount of data
6337 : * we'd need to pass upwards in case of a sort, hash, etc.
6338 : *
6339 : * This function also sets reltarget->cost, so it's a bit misnamed now.
6340 : *
6341 : * NB: this works best on plain relations because it prefers to look at
6342 : * real Vars. For subqueries, set_subquery_size_estimates will already have
6343 : * copied up whatever per-column estimates were made within the subquery,
6344 : * and for other types of rels there isn't much we can do anyway. We fall
6345 : * back on (fairly stupid) datatype-based width estimates if we can't get
6346 : * any better number.
6347 : *
6348 : * The per-attribute width estimates are cached for possible re-use while
6349 : * building join relations or post-scan/join pathtargets.
6350 : */
6351 : static void
6352 536780 : set_rel_width(PlannerInfo *root, RelOptInfo *rel)
6353 : {
6354 536780 : Oid reloid = planner_rt_fetch(rel->relid, root)->relid;
6355 536780 : int64 tuple_width = 0;
6356 536780 : bool have_wholerow_var = false;
6357 : ListCell *lc;
6358 :
6359 : /* Vars are assumed to have cost zero, but other exprs do not */
6360 536780 : rel->reltarget->cost.startup = 0;
6361 536780 : rel->reltarget->cost.per_tuple = 0;
6362 :
6363 1911064 : foreach(lc, rel->reltarget->exprs)
6364 : {
6365 1374284 : Node *node = (Node *) lfirst(lc);
6366 :
6367 : /*
6368 : * Ordinarily, a Var in a rel's targetlist must belong to that rel;
6369 : * but there are corner cases involving LATERAL references where that
6370 : * isn't so. If the Var has the wrong varno, fall through to the
6371 : * generic case (it doesn't seem worth the trouble to be any smarter).
6372 : */
6373 1374284 : if (IsA(node, Var) &&
6374 1349854 : ((Var *) node)->varno == rel->relid)
6375 371498 : {
6376 1349788 : Var *var = (Var *) node;
6377 : int ndx;
6378 : int32 item_width;
6379 :
6380 : Assert(var->varattno >= rel->min_attr);
6381 : Assert(var->varattno <= rel->max_attr);
6382 :
6383 1349788 : ndx = var->varattno - rel->min_attr;
6384 :
6385 : /*
6386 : * If it's a whole-row Var, we'll deal with it below after we have
6387 : * already cached as many attr widths as possible.
6388 : */
6389 1349788 : if (var->varattno == 0)
6390 : {
6391 3062 : have_wholerow_var = true;
6392 3062 : continue;
6393 : }
6394 :
6395 : /*
6396 : * The width may have been cached already (especially if it's a
6397 : * subquery), so don't duplicate effort.
6398 : */
6399 1346726 : if (rel->attr_widths[ndx] > 0)
6400 : {
6401 274720 : tuple_width += rel->attr_widths[ndx];
6402 274720 : continue;
6403 : }
6404 :
6405 : /* Try to get column width from statistics */
6406 1072006 : if (reloid != InvalidOid && var->varattno > 0)
6407 : {
6408 838788 : item_width = get_attavgwidth(reloid, var->varattno);
6409 838788 : if (item_width > 0)
6410 : {
6411 700508 : rel->attr_widths[ndx] = item_width;
6412 700508 : tuple_width += item_width;
6413 700508 : continue;
6414 : }
6415 : }
6416 :
6417 : /*
6418 : * Not a plain relation, or can't find statistics for it. Estimate
6419 : * using just the type info.
6420 : */
6421 371498 : item_width = get_typavgwidth(var->vartype, var->vartypmod);
6422 : Assert(item_width > 0);
6423 371498 : rel->attr_widths[ndx] = item_width;
6424 371498 : tuple_width += item_width;
6425 : }
6426 24496 : else if (IsA(node, PlaceHolderVar))
6427 : {
6428 : /*
6429 : * We will need to evaluate the PHV's contained expression while
6430 : * scanning this rel, so be sure to include it in reltarget->cost.
6431 : */
6432 2134 : PlaceHolderVar *phv = (PlaceHolderVar *) node;
6433 2134 : PlaceHolderInfo *phinfo = find_placeholder_info(root, phv);
6434 : QualCost cost;
6435 :
6436 2134 : tuple_width += phinfo->ph_width;
6437 2134 : cost_qual_eval_node(&cost, (Node *) phv->phexpr, root);
6438 2134 : rel->reltarget->cost.startup += cost.startup;
6439 2134 : rel->reltarget->cost.per_tuple += cost.per_tuple;
6440 : }
6441 : else
6442 : {
6443 : /*
6444 : * We could be looking at an expression pulled up from a subquery,
6445 : * or a ROW() representing a whole-row child Var, etc. Do what we
6446 : * can using the expression type information.
6447 : */
6448 : int32 item_width;
6449 : QualCost cost;
6450 :
6451 22362 : item_width = get_typavgwidth(exprType(node), exprTypmod(node));
6452 : Assert(item_width > 0);
6453 22362 : tuple_width += item_width;
6454 : /* Not entirely clear if we need to account for cost, but do so */
6455 22362 : cost_qual_eval_node(&cost, node, root);
6456 22362 : rel->reltarget->cost.startup += cost.startup;
6457 22362 : rel->reltarget->cost.per_tuple += cost.per_tuple;
6458 : }
6459 : }
6460 :
6461 : /*
6462 : * If we have a whole-row reference, estimate its width as the sum of
6463 : * per-column widths plus heap tuple header overhead.
6464 : */
6465 536780 : if (have_wholerow_var)
6466 : {
6467 3062 : int64 wholerow_width = MAXALIGN(SizeofHeapTupleHeader);
6468 :
6469 3062 : if (reloid != InvalidOid)
6470 : {
6471 : /* Real relation, so estimate true tuple width */
6472 2392 : wholerow_width += get_relation_data_width(reloid,
6473 2392 : rel->attr_widths - rel->min_attr);
6474 : }
6475 : else
6476 : {
6477 : /* Do what we can with info for a phony rel */
6478 : AttrNumber i;
6479 :
6480 1820 : for (i = 1; i <= rel->max_attr; i++)
6481 1150 : wholerow_width += rel->attr_widths[i - rel->min_attr];
6482 : }
6483 :
6484 3062 : rel->attr_widths[0 - rel->min_attr] = clamp_width_est(wholerow_width);
6485 :
6486 : /*
6487 : * Include the whole-row Var as part of the output tuple. Yes, that
6488 : * really is what happens at runtime.
6489 : */
6490 3062 : tuple_width += wholerow_width;
6491 : }
6492 :
6493 536780 : rel->reltarget->width = clamp_width_est(tuple_width);
6494 536780 : }
6495 :
6496 : /*
6497 : * set_pathtarget_cost_width
6498 : * Set the estimated eval cost and output width of a PathTarget tlist.
6499 : *
6500 : * As a notational convenience, returns the same PathTarget pointer passed in.
6501 : *
6502 : * Most, though not quite all, uses of this function occur after we've run
6503 : * set_rel_width() for base relations; so we can usually obtain cached width
6504 : * estimates for Vars. If we can't, fall back on datatype-based width
6505 : * estimates. Present early-planning uses of PathTargets don't need accurate
6506 : * widths badly enough to justify going to the catalogs for better data.
6507 : */
6508 : PathTarget *
6509 641506 : set_pathtarget_cost_width(PlannerInfo *root, PathTarget *target)
6510 : {
6511 641506 : int64 tuple_width = 0;
6512 : ListCell *lc;
6513 :
6514 : /* Vars are assumed to have cost zero, but other exprs do not */
6515 641506 : target->cost.startup = 0;
6516 641506 : target->cost.per_tuple = 0;
6517 :
6518 2217926 : foreach(lc, target->exprs)
6519 : {
6520 1576420 : Node *node = (Node *) lfirst(lc);
6521 :
6522 1576420 : tuple_width += get_expr_width(root, node);
6523 :
6524 : /* For non-Vars, account for evaluation cost */
6525 1576420 : if (!IsA(node, Var))
6526 : {
6527 : QualCost cost;
6528 :
6529 681844 : cost_qual_eval_node(&cost, node, root);
6530 681844 : target->cost.startup += cost.startup;
6531 681844 : target->cost.per_tuple += cost.per_tuple;
6532 : }
6533 : }
6534 :
6535 641506 : target->width = clamp_width_est(tuple_width);
6536 :
6537 641506 : return target;
6538 : }
6539 :
6540 : /*
6541 : * get_expr_width
6542 : * Estimate the width of the given expr attempting to use the width
6543 : * cached in a Var's owning RelOptInfo, else fallback on the type's
6544 : * average width when unable to or when the given Node is not a Var.
6545 : */
6546 : static int32
6547 1911596 : get_expr_width(PlannerInfo *root, const Node *expr)
6548 : {
6549 : int32 width;
6550 :
6551 1911596 : if (IsA(expr, Var))
6552 : {
6553 1217088 : const Var *var = (const Var *) expr;
6554 :
6555 : /* We should not see any upper-level Vars here */
6556 : Assert(var->varlevelsup == 0);
6557 :
6558 : /* Try to get data from RelOptInfo cache */
6559 1217088 : if (!IS_SPECIAL_VARNO(var->varno) &&
6560 1211178 : var->varno < root->simple_rel_array_size)
6561 : {
6562 1211178 : RelOptInfo *rel = root->simple_rel_array[var->varno];
6563 :
6564 1211178 : if (rel != NULL &&
6565 1193186 : var->varattno >= rel->min_attr &&
6566 1193186 : var->varattno <= rel->max_attr)
6567 : {
6568 1193186 : int ndx = var->varattno - rel->min_attr;
6569 :
6570 1193186 : if (rel->attr_widths[ndx] > 0)
6571 1160646 : return rel->attr_widths[ndx];
6572 : }
6573 : }
6574 :
6575 : /*
6576 : * No cached data available, so estimate using just the type info.
6577 : */
6578 56442 : width = get_typavgwidth(var->vartype, var->vartypmod);
6579 : Assert(width > 0);
6580 :
6581 56442 : return width;
6582 : }
6583 :
6584 694508 : width = get_typavgwidth(exprType(expr), exprTypmod(expr));
6585 : Assert(width > 0);
6586 694508 : return width;
6587 : }
6588 :
6589 : /*
6590 : * relation_byte_size
6591 : * Estimate the storage space in bytes for a given number of tuples
6592 : * of a given width (size in bytes).
6593 : */
6594 : static double
6595 5190278 : relation_byte_size(double tuples, int width)
6596 : {
6597 5190278 : return tuples * (MAXALIGN(width) + MAXALIGN(SizeofHeapTupleHeader));
6598 : }
6599 :
6600 : /*
6601 : * page_size
6602 : * Returns an estimate of the number of pages covered by a given
6603 : * number of tuples of a given width (size in bytes).
6604 : */
6605 : static double
6606 9512 : page_size(double tuples, int width)
6607 : {
6608 9512 : return ceil(relation_byte_size(tuples, width) / BLCKSZ);
6609 : }
6610 :
6611 : /*
6612 : * Estimate the fraction of the work that each worker will do given the
6613 : * number of workers budgeted for the path.
6614 : */
6615 : static double
6616 471512 : get_parallel_divisor(Path *path)
6617 : {
6618 471512 : double parallel_divisor = path->parallel_workers;
6619 :
6620 : /*
6621 : * Early experience with parallel query suggests that when there is only
6622 : * one worker, the leader often makes a very substantial contribution to
6623 : * executing the parallel portion of the plan, but as more workers are
6624 : * added, it does less and less, because it's busy reading tuples from the
6625 : * workers and doing whatever non-parallel post-processing is needed. By
6626 : * the time we reach 4 workers, the leader no longer makes a meaningful
6627 : * contribution. Thus, for now, estimate that the leader spends 30% of
6628 : * its time servicing each worker, and the remainder executing the
6629 : * parallel plan.
6630 : */
6631 471512 : if (parallel_leader_participation)
6632 : {
6633 : double leader_contribution;
6634 :
6635 470282 : leader_contribution = 1.0 - (0.3 * path->parallel_workers);
6636 470282 : if (leader_contribution > 0)
6637 467702 : parallel_divisor += leader_contribution;
6638 : }
6639 :
6640 471512 : return parallel_divisor;
6641 : }
6642 :
6643 : /*
6644 : * compute_bitmap_pages
6645 : * Estimate number of pages fetched from heap in a bitmap heap scan.
6646 : *
6647 : * 'baserel' is the relation to be scanned
6648 : * 'bitmapqual' is a tree of IndexPaths, BitmapAndPaths, and BitmapOrPaths
6649 : * 'loop_count' is the number of repetitions of the indexscan to factor into
6650 : * estimates of caching behavior
6651 : *
6652 : * If cost_p isn't NULL, the indexTotalCost estimate is returned in *cost_p.
6653 : * If tuples_p isn't NULL, the tuples_fetched estimate is returned in *tuples_p.
6654 : */
6655 : double
6656 714106 : compute_bitmap_pages(PlannerInfo *root, RelOptInfo *baserel,
6657 : Path *bitmapqual, double loop_count,
6658 : Cost *cost_p, double *tuples_p)
6659 : {
6660 : Cost indexTotalCost;
6661 : Selectivity indexSelectivity;
6662 : double T;
6663 : double pages_fetched;
6664 : double tuples_fetched;
6665 : double heap_pages;
6666 : double maxentries;
6667 :
6668 : /*
6669 : * Fetch total cost of obtaining the bitmap, as well as its total
6670 : * selectivity.
6671 : */
6672 714106 : cost_bitmap_tree_node(bitmapqual, &indexTotalCost, &indexSelectivity);
6673 :
6674 : /*
6675 : * Estimate number of main-table pages fetched.
6676 : */
6677 714106 : tuples_fetched = clamp_row_est(indexSelectivity * baserel->tuples);
6678 :
6679 714106 : T = (baserel->pages > 1) ? (double) baserel->pages : 1.0;
6680 :
6681 : /*
6682 : * For a single scan, the number of heap pages that need to be fetched is
6683 : * the same as the Mackert and Lohman formula for the case T <= b (ie, no
6684 : * re-reads needed).
6685 : */
6686 714106 : pages_fetched = (2.0 * T * tuples_fetched) / (2.0 * T + tuples_fetched);
6687 :
6688 : /*
6689 : * Calculate the number of pages fetched from the heap. Then based on
6690 : * current work_mem estimate get the estimated maxentries in the bitmap.
6691 : * (Note that we always do this calculation based on the number of pages
6692 : * that would be fetched in a single iteration, even if loop_count > 1.
6693 : * That's correct, because only that number of entries will be stored in
6694 : * the bitmap at one time.)
6695 : */
6696 714106 : heap_pages = Min(pages_fetched, baserel->pages);
6697 714106 : maxentries = tbm_calculate_entries(work_mem * (Size) 1024);
6698 :
6699 714106 : if (loop_count > 1)
6700 : {
6701 : /*
6702 : * For repeated bitmap scans, scale up the number of tuples fetched in
6703 : * the Mackert and Lohman formula by the number of scans, so that we
6704 : * estimate the number of pages fetched by all the scans. Then
6705 : * pro-rate for one scan.
6706 : */
6707 150228 : pages_fetched = index_pages_fetched(tuples_fetched * loop_count,
6708 : baserel->pages,
6709 : get_indexpath_pages(bitmapqual),
6710 : root);
6711 150228 : pages_fetched /= loop_count;
6712 : }
6713 :
6714 714106 : if (pages_fetched >= T)
6715 71970 : pages_fetched = T;
6716 : else
6717 642136 : pages_fetched = ceil(pages_fetched);
6718 :
6719 714106 : if (maxentries < heap_pages)
6720 : {
6721 : double exact_pages;
6722 : double lossy_pages;
6723 :
6724 : /*
6725 : * Crude approximation of the number of lossy pages. Because of the
6726 : * way tbm_lossify() is coded, the number of lossy pages increases
6727 : * very sharply as soon as we run short of memory; this formula has
6728 : * that property and seems to perform adequately in testing, but it's
6729 : * possible we could do better somehow.
6730 : */
6731 18 : lossy_pages = Max(0, heap_pages - maxentries / 2);
6732 18 : exact_pages = heap_pages - lossy_pages;
6733 :
6734 : /*
6735 : * If there are lossy pages then recompute the number of tuples
6736 : * processed by the bitmap heap node. We assume here that the chance
6737 : * of a given tuple coming from an exact page is the same as the
6738 : * chance that a given page is exact. This might not be true, but
6739 : * it's not clear how we can do any better.
6740 : */
6741 18 : if (lossy_pages > 0)
6742 : tuples_fetched =
6743 18 : clamp_row_est(indexSelectivity *
6744 18 : (exact_pages / heap_pages) * baserel->tuples +
6745 18 : (lossy_pages / heap_pages) * baserel->tuples);
6746 : }
6747 :
6748 714106 : if (cost_p)
6749 566830 : *cost_p = indexTotalCost;
6750 714106 : if (tuples_p)
6751 566830 : *tuples_p = tuples_fetched;
6752 :
6753 714106 : return pages_fetched;
6754 : }
6755 :
6756 : /*
6757 : * compute_gather_rows
6758 : * Estimate number of rows for gather (merge) nodes.
6759 : *
6760 : * In a parallel plan, each worker's row estimate is determined by dividing the
6761 : * total number of rows by parallel_divisor, which accounts for the leader's
6762 : * contribution in addition to the number of workers. Accordingly, when
6763 : * estimating the number of rows for gather (merge) nodes, we multiply the rows
6764 : * per worker by the same parallel_divisor to undo the division.
6765 : */
6766 : double
6767 45754 : compute_gather_rows(Path *path)
6768 : {
6769 : Assert(path->parallel_workers > 0);
6770 :
6771 45754 : return clamp_row_est(path->rows * get_parallel_divisor(path));
6772 : }
|
