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) 1996-2025, 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 10436928 : 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 10436928 : if (nrows > MAXIMUM_ROWCOUNT || isnan(nrows))
222 0 : nrows = MAXIMUM_ROWCOUNT;
223 10436928 : else if (nrows <= 1.0)
224 3336088 : nrows = 1.0;
225 : else
226 7100840 : nrows = rint(nrows);
227 :
228 10436928 : 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 1941796 : 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 1941796 : 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 1941796 : 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 432962 : cost_seqscan(Path *path, PlannerInfo *root,
270 : RelOptInfo *baserel, ParamPathInfo *param_info)
271 : {
272 432962 : 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 :
279 : /* Should only be applied to base relations */
280 : Assert(baserel->relid > 0);
281 : Assert(baserel->rtekind == RTE_RELATION);
282 :
283 : /* Mark the path with the correct row estimate */
284 432962 : if (param_info)
285 840 : path->rows = param_info->ppi_rows;
286 : else
287 432122 : path->rows = baserel->rows;
288 :
289 : /* fetch estimated page cost for tablespace containing table */
290 432962 : get_tablespace_page_costs(baserel->reltablespace,
291 : NULL,
292 : &spc_seq_page_cost);
293 :
294 : /*
295 : * disk costs
296 : */
297 432962 : disk_run_cost = spc_seq_page_cost * baserel->pages;
298 :
299 : /* CPU costs */
300 432962 : get_restriction_qual_cost(root, baserel, param_info, &qpqual_cost);
301 :
302 432962 : startup_cost += qpqual_cost.startup;
303 432962 : cpu_per_tuple = cpu_tuple_cost + qpqual_cost.per_tuple;
304 432962 : cpu_run_cost = cpu_per_tuple * baserel->tuples;
305 : /* tlist eval costs are paid per output row, not per tuple scanned */
306 432962 : startup_cost += path->pathtarget->cost.startup;
307 432962 : cpu_run_cost += path->pathtarget->cost.per_tuple * path->rows;
308 :
309 : /* Adjust costing for parallelism, if used. */
310 432962 : if (path->parallel_workers > 0)
311 : {
312 27398 : double parallel_divisor = get_parallel_divisor(path);
313 :
314 : /* The CPU cost is divided among all the workers. */
315 27398 : cpu_run_cost /= parallel_divisor;
316 :
317 : /*
318 : * It may be possible to amortize some of the I/O cost, but probably
319 : * not very much, because most operating systems already do aggressive
320 : * prefetching. For now, we assume that the disk run cost can't be
321 : * amortized at all.
322 : */
323 :
324 : /*
325 : * In the case of a parallel plan, the row count needs to represent
326 : * the number of tuples processed per worker.
327 : */
328 27398 : path->rows = clamp_row_est(path->rows / parallel_divisor);
329 : }
330 :
331 432962 : path->disabled_nodes = enable_seqscan ? 0 : 1;
332 432962 : path->startup_cost = startup_cost;
333 432962 : path->total_cost = startup_cost + cpu_run_cost + disk_run_cost;
334 432962 : }
335 :
336 : /*
337 : * cost_samplescan
338 : * Determines and returns the cost of scanning a relation using sampling.
339 : *
340 : * 'baserel' is the relation to be scanned
341 : * 'param_info' is the ParamPathInfo if this is a parameterized path, else NULL
342 : */
343 : void
344 306 : cost_samplescan(Path *path, PlannerInfo *root,
345 : RelOptInfo *baserel, ParamPathInfo *param_info)
346 : {
347 306 : Cost startup_cost = 0;
348 306 : Cost run_cost = 0;
349 : RangeTblEntry *rte;
350 : TableSampleClause *tsc;
351 : TsmRoutine *tsm;
352 : double spc_seq_page_cost,
353 : spc_random_page_cost,
354 : spc_page_cost;
355 : QualCost qpqual_cost;
356 : Cost cpu_per_tuple;
357 :
358 : /* Should only be applied to base relations with tablesample clauses */
359 : Assert(baserel->relid > 0);
360 306 : rte = planner_rt_fetch(baserel->relid, root);
361 : Assert(rte->rtekind == RTE_RELATION);
362 306 : tsc = rte->tablesample;
363 : Assert(tsc != NULL);
364 306 : tsm = GetTsmRoutine(tsc->tsmhandler);
365 :
366 : /* Mark the path with the correct row estimate */
367 306 : if (param_info)
368 72 : path->rows = param_info->ppi_rows;
369 : else
370 234 : path->rows = baserel->rows;
371 :
372 : /* fetch estimated page cost for tablespace containing table */
373 306 : get_tablespace_page_costs(baserel->reltablespace,
374 : &spc_random_page_cost,
375 : &spc_seq_page_cost);
376 :
377 : /* if NextSampleBlock is used, assume random access, else sequential */
378 612 : spc_page_cost = (tsm->NextSampleBlock != NULL) ?
379 306 : spc_random_page_cost : spc_seq_page_cost;
380 :
381 : /*
382 : * disk costs (recall that baserel->pages has already been set to the
383 : * number of pages the sampling method will visit)
384 : */
385 306 : run_cost += spc_page_cost * baserel->pages;
386 :
387 : /*
388 : * CPU costs (recall that baserel->tuples has already been set to the
389 : * number of tuples the sampling method will select). Note that we ignore
390 : * execution cost of the TABLESAMPLE parameter expressions; they will be
391 : * evaluated only once per scan, and in most usages they'll likely be
392 : * simple constants anyway. We also don't charge anything for the
393 : * calculations the sampling method might do internally.
394 : */
395 306 : get_restriction_qual_cost(root, baserel, param_info, &qpqual_cost);
396 :
397 306 : startup_cost += qpqual_cost.startup;
398 306 : cpu_per_tuple = cpu_tuple_cost + qpqual_cost.per_tuple;
399 306 : run_cost += cpu_per_tuple * baserel->tuples;
400 : /* tlist eval costs are paid per output row, not per tuple scanned */
401 306 : startup_cost += path->pathtarget->cost.startup;
402 306 : run_cost += path->pathtarget->cost.per_tuple * path->rows;
403 :
404 306 : path->disabled_nodes = 0;
405 306 : path->startup_cost = startup_cost;
406 306 : path->total_cost = startup_cost + run_cost;
407 306 : }
408 :
409 : /*
410 : * cost_gather
411 : * Determines and returns the cost of gather path.
412 : *
413 : * 'rel' is the relation to be operated upon
414 : * 'param_info' is the ParamPathInfo if this is a parameterized path, else NULL
415 : * 'rows' may be used to point to a row estimate; if non-NULL, it overrides
416 : * both 'rel' and 'param_info'. This is useful when the path doesn't exactly
417 : * correspond to any particular RelOptInfo.
418 : */
419 : void
420 24778 : cost_gather(GatherPath *path, PlannerInfo *root,
421 : RelOptInfo *rel, ParamPathInfo *param_info,
422 : double *rows)
423 : {
424 24778 : Cost startup_cost = 0;
425 24778 : Cost run_cost = 0;
426 :
427 : /* Mark the path with the correct row estimate */
428 24778 : if (rows)
429 6078 : path->path.rows = *rows;
430 18700 : else if (param_info)
431 0 : path->path.rows = param_info->ppi_rows;
432 : else
433 18700 : path->path.rows = rel->rows;
434 :
435 24778 : startup_cost = path->subpath->startup_cost;
436 :
437 24778 : run_cost = path->subpath->total_cost - path->subpath->startup_cost;
438 :
439 : /* Parallel setup and communication cost. */
440 24778 : startup_cost += parallel_setup_cost;
441 24778 : run_cost += parallel_tuple_cost * path->path.rows;
442 :
443 24778 : path->path.disabled_nodes = path->subpath->disabled_nodes;
444 24778 : path->path.startup_cost = startup_cost;
445 24778 : path->path.total_cost = (startup_cost + run_cost);
446 24778 : }
447 :
448 : /*
449 : * cost_gather_merge
450 : * Determines and returns the cost of gather merge path.
451 : *
452 : * GatherMerge merges several pre-sorted input streams, using a heap that at
453 : * any given instant holds the next tuple from each stream. If there are N
454 : * streams, we need about N*log2(N) tuple comparisons to construct the heap at
455 : * startup, and then for each output tuple, about log2(N) comparisons to
456 : * replace the top heap entry with the next tuple from the same stream.
457 : */
458 : void
459 17552 : cost_gather_merge(GatherMergePath *path, PlannerInfo *root,
460 : RelOptInfo *rel, ParamPathInfo *param_info,
461 : int input_disabled_nodes,
462 : Cost input_startup_cost, Cost input_total_cost,
463 : double *rows)
464 : {
465 17552 : Cost startup_cost = 0;
466 17552 : Cost run_cost = 0;
467 : Cost comparison_cost;
468 : double N;
469 : double logN;
470 :
471 : /* Mark the path with the correct row estimate */
472 17552 : if (rows)
473 10920 : path->path.rows = *rows;
474 6632 : else if (param_info)
475 0 : path->path.rows = param_info->ppi_rows;
476 : else
477 6632 : path->path.rows = rel->rows;
478 :
479 : /*
480 : * Add one to the number of workers to account for the leader. This might
481 : * be overgenerous since the leader will do less work than other workers
482 : * in typical cases, but we'll go with it for now.
483 : */
484 : Assert(path->num_workers > 0);
485 17552 : N = (double) path->num_workers + 1;
486 17552 : logN = LOG2(N);
487 :
488 : /* Assumed cost per tuple comparison */
489 17552 : comparison_cost = 2.0 * cpu_operator_cost;
490 :
491 : /* Heap creation cost */
492 17552 : startup_cost += comparison_cost * N * logN;
493 :
494 : /* Per-tuple heap maintenance cost */
495 17552 : run_cost += path->path.rows * comparison_cost * logN;
496 :
497 : /* small cost for heap management, like cost_merge_append */
498 17552 : run_cost += cpu_operator_cost * path->path.rows;
499 :
500 : /*
501 : * Parallel setup and communication cost. Since Gather Merge, unlike
502 : * Gather, requires us to block until a tuple is available from every
503 : * worker, we bump the IPC cost up a little bit as compared with Gather.
504 : * For lack of a better idea, charge an extra 5%.
505 : */
506 17552 : startup_cost += parallel_setup_cost;
507 17552 : run_cost += parallel_tuple_cost * path->path.rows * 1.05;
508 :
509 17552 : path->path.disabled_nodes = input_disabled_nodes
510 17552 : + (enable_gathermerge ? 0 : 1);
511 17552 : path->path.startup_cost = startup_cost + input_startup_cost;
512 17552 : path->path.total_cost = (startup_cost + run_cost + input_total_cost);
513 17552 : }
514 :
515 : /*
516 : * cost_index
517 : * Determines and returns the cost of scanning a relation using an index.
518 : *
519 : * 'path' describes the indexscan under consideration, and is complete
520 : * except for the fields to be set by this routine
521 : * 'loop_count' is the number of repetitions of the indexscan to factor into
522 : * estimates of caching behavior
523 : *
524 : * In addition to rows, startup_cost and total_cost, cost_index() sets the
525 : * path's indextotalcost and indexselectivity fields. These values will be
526 : * needed if the IndexPath is used in a BitmapIndexScan.
527 : *
528 : * NOTE: path->indexquals must contain only clauses usable as index
529 : * restrictions. Any additional quals evaluated as qpquals may reduce the
530 : * number of returned tuples, but they won't reduce the number of tuples
531 : * we have to fetch from the table, so they don't reduce the scan cost.
532 : */
533 : void
534 802418 : cost_index(IndexPath *path, PlannerInfo *root, double loop_count,
535 : bool partial_path)
536 : {
537 802418 : IndexOptInfo *index = path->indexinfo;
538 802418 : RelOptInfo *baserel = index->rel;
539 802418 : bool indexonly = (path->path.pathtype == T_IndexOnlyScan);
540 : amcostestimate_function amcostestimate;
541 : List *qpquals;
542 802418 : Cost startup_cost = 0;
543 802418 : Cost run_cost = 0;
544 802418 : Cost cpu_run_cost = 0;
545 : Cost indexStartupCost;
546 : Cost indexTotalCost;
547 : Selectivity indexSelectivity;
548 : double indexCorrelation,
549 : csquared;
550 : double spc_seq_page_cost,
551 : spc_random_page_cost;
552 : Cost min_IO_cost,
553 : max_IO_cost;
554 : QualCost qpqual_cost;
555 : Cost cpu_per_tuple;
556 : double tuples_fetched;
557 : double pages_fetched;
558 : double rand_heap_pages;
559 : double index_pages;
560 :
561 : /* Should only be applied to base relations */
562 : Assert(IsA(baserel, RelOptInfo) &&
563 : IsA(index, IndexOptInfo));
564 : Assert(baserel->relid > 0);
565 : Assert(baserel->rtekind == RTE_RELATION);
566 :
567 : /*
568 : * Mark the path with the correct row estimate, and identify which quals
569 : * will need to be enforced as qpquals. We need not check any quals that
570 : * are implied by the index's predicate, so we can use indrestrictinfo not
571 : * baserestrictinfo as the list of relevant restriction clauses for the
572 : * rel.
573 : */
574 802418 : if (path->path.param_info)
575 : {
576 154436 : path->path.rows = path->path.param_info->ppi_rows;
577 : /* qpquals come from the rel's restriction clauses and ppi_clauses */
578 154436 : qpquals = list_concat(extract_nonindex_conditions(path->indexinfo->indrestrictinfo,
579 : path->indexclauses),
580 154436 : extract_nonindex_conditions(path->path.param_info->ppi_clauses,
581 : path->indexclauses));
582 : }
583 : else
584 : {
585 647982 : path->path.rows = baserel->rows;
586 : /* qpquals come from just the rel's restriction clauses */
587 647982 : qpquals = extract_nonindex_conditions(path->indexinfo->indrestrictinfo,
588 : path->indexclauses);
589 : }
590 :
591 : /* we don't need to check enable_indexonlyscan; indxpath.c does that */
592 802418 : path->path.disabled_nodes = enable_indexscan ? 0 : 1;
593 :
594 : /*
595 : * Call index-access-method-specific code to estimate the processing cost
596 : * for scanning the index, as well as the selectivity of the index (ie,
597 : * the fraction of main-table tuples we will have to retrieve) and its
598 : * correlation to the main-table tuple order. We need a cast here because
599 : * pathnodes.h uses a weak function type to avoid including amapi.h.
600 : */
601 802418 : amcostestimate = (amcostestimate_function) index->amcostestimate;
602 802418 : amcostestimate(root, path, loop_count,
603 : &indexStartupCost, &indexTotalCost,
604 : &indexSelectivity, &indexCorrelation,
605 : &index_pages);
606 :
607 : /*
608 : * Save amcostestimate's results for possible use in bitmap scan planning.
609 : * We don't bother to save indexStartupCost or indexCorrelation, because a
610 : * bitmap scan doesn't care about either.
611 : */
612 802418 : path->indextotalcost = indexTotalCost;
613 802418 : path->indexselectivity = indexSelectivity;
614 :
615 : /* all costs for touching index itself included here */
616 802418 : startup_cost += indexStartupCost;
617 802418 : run_cost += indexTotalCost - indexStartupCost;
618 :
619 : /* estimate number of main-table tuples fetched */
620 802418 : tuples_fetched = clamp_row_est(indexSelectivity * baserel->tuples);
621 :
622 : /* fetch estimated page costs for tablespace containing table */
623 802418 : get_tablespace_page_costs(baserel->reltablespace,
624 : &spc_random_page_cost,
625 : &spc_seq_page_cost);
626 :
627 : /*----------
628 : * Estimate number of main-table pages fetched, and compute I/O cost.
629 : *
630 : * When the index ordering is uncorrelated with the table ordering,
631 : * we use an approximation proposed by Mackert and Lohman (see
632 : * index_pages_fetched() for details) to compute the number of pages
633 : * fetched, and then charge spc_random_page_cost per page fetched.
634 : *
635 : * When the index ordering is exactly correlated with the table ordering
636 : * (just after a CLUSTER, for example), the number of pages fetched should
637 : * be exactly selectivity * table_size. What's more, all but the first
638 : * will be sequential fetches, not the random fetches that occur in the
639 : * uncorrelated case. So if the number of pages is more than 1, we
640 : * ought to charge
641 : * spc_random_page_cost + (pages_fetched - 1) * spc_seq_page_cost
642 : * For partially-correlated indexes, we ought to charge somewhere between
643 : * these two estimates. We currently interpolate linearly between the
644 : * estimates based on the correlation squared (XXX is that appropriate?).
645 : *
646 : * If it's an index-only scan, then we will not need to fetch any heap
647 : * pages for which the visibility map shows all tuples are visible.
648 : * Hence, reduce the estimated number of heap fetches accordingly.
649 : * We use the measured fraction of the entire heap that is all-visible,
650 : * which might not be particularly relevant to the subset of the heap
651 : * that this query will fetch; but it's not clear how to do better.
652 : *----------
653 : */
654 802418 : if (loop_count > 1)
655 : {
656 : /*
657 : * For repeated indexscans, the appropriate estimate for the
658 : * uncorrelated case is to scale up the number of tuples fetched in
659 : * the Mackert and Lohman formula by the number of scans, so that we
660 : * estimate the number of pages fetched by all the scans; then
661 : * pro-rate the costs for one scan. In this case we assume all the
662 : * fetches are random accesses.
663 : */
664 88988 : pages_fetched = index_pages_fetched(tuples_fetched * loop_count,
665 : baserel->pages,
666 88988 : (double) index->pages,
667 : root);
668 :
669 88988 : if (indexonly)
670 11208 : pages_fetched = ceil(pages_fetched * (1.0 - baserel->allvisfrac));
671 :
672 88988 : rand_heap_pages = pages_fetched;
673 :
674 88988 : max_IO_cost = (pages_fetched * spc_random_page_cost) / loop_count;
675 :
676 : /*
677 : * In the perfectly correlated case, the number of pages touched by
678 : * each scan is selectivity * table_size, and we can use the Mackert
679 : * and Lohman formula at the page level to estimate how much work is
680 : * saved by caching across scans. We still assume all the fetches are
681 : * random, though, which is an overestimate that's hard to correct for
682 : * without double-counting the cache effects. (But in most cases
683 : * where such a plan is actually interesting, only one page would get
684 : * fetched per scan anyway, so it shouldn't matter much.)
685 : */
686 88988 : pages_fetched = ceil(indexSelectivity * (double) baserel->pages);
687 :
688 88988 : pages_fetched = index_pages_fetched(pages_fetched * loop_count,
689 : baserel->pages,
690 88988 : (double) index->pages,
691 : root);
692 :
693 88988 : if (indexonly)
694 11208 : pages_fetched = ceil(pages_fetched * (1.0 - baserel->allvisfrac));
695 :
696 88988 : min_IO_cost = (pages_fetched * spc_random_page_cost) / loop_count;
697 : }
698 : else
699 : {
700 : /*
701 : * Normal case: apply the Mackert and Lohman formula, and then
702 : * interpolate between that and the correlation-derived result.
703 : */
704 713430 : pages_fetched = index_pages_fetched(tuples_fetched,
705 : baserel->pages,
706 713430 : (double) index->pages,
707 : root);
708 :
709 713430 : if (indexonly)
710 70820 : pages_fetched = ceil(pages_fetched * (1.0 - baserel->allvisfrac));
711 :
712 713430 : rand_heap_pages = pages_fetched;
713 :
714 : /* max_IO_cost is for the perfectly uncorrelated case (csquared=0) */
715 713430 : max_IO_cost = pages_fetched * spc_random_page_cost;
716 :
717 : /* min_IO_cost is for the perfectly correlated case (csquared=1) */
718 713430 : pages_fetched = ceil(indexSelectivity * (double) baserel->pages);
719 :
720 713430 : if (indexonly)
721 70820 : pages_fetched = ceil(pages_fetched * (1.0 - baserel->allvisfrac));
722 :
723 713430 : if (pages_fetched > 0)
724 : {
725 634994 : min_IO_cost = spc_random_page_cost;
726 634994 : if (pages_fetched > 1)
727 186490 : min_IO_cost += (pages_fetched - 1) * spc_seq_page_cost;
728 : }
729 : else
730 78436 : min_IO_cost = 0;
731 : }
732 :
733 802418 : if (partial_path)
734 : {
735 : /*
736 : * For index only scans compute workers based on number of index pages
737 : * fetched; the number of heap pages we fetch might be so small as to
738 : * effectively rule out parallelism, which we don't want to do.
739 : */
740 276976 : if (indexonly)
741 26136 : rand_heap_pages = -1;
742 :
743 : /*
744 : * Estimate the number of parallel workers required to scan index. Use
745 : * the number of heap pages computed considering heap fetches won't be
746 : * sequential as for parallel scans the pages are accessed in random
747 : * order.
748 : */
749 276976 : path->path.parallel_workers = compute_parallel_worker(baserel,
750 : rand_heap_pages,
751 : index_pages,
752 : max_parallel_workers_per_gather);
753 :
754 : /*
755 : * Fall out if workers can't be assigned for parallel scan, because in
756 : * such a case this path will be rejected. So there is no benefit in
757 : * doing extra computation.
758 : */
759 276976 : if (path->path.parallel_workers <= 0)
760 266874 : return;
761 :
762 10102 : path->path.parallel_aware = true;
763 : }
764 :
765 : /*
766 : * Now interpolate based on estimated index order correlation to get total
767 : * disk I/O cost for main table accesses.
768 : */
769 535544 : csquared = indexCorrelation * indexCorrelation;
770 :
771 535544 : run_cost += max_IO_cost + csquared * (min_IO_cost - max_IO_cost);
772 :
773 : /*
774 : * Estimate CPU costs per tuple.
775 : *
776 : * What we want here is cpu_tuple_cost plus the evaluation costs of any
777 : * qual clauses that we have to evaluate as qpquals.
778 : */
779 535544 : cost_qual_eval(&qpqual_cost, qpquals, root);
780 :
781 535544 : startup_cost += qpqual_cost.startup;
782 535544 : cpu_per_tuple = cpu_tuple_cost + qpqual_cost.per_tuple;
783 :
784 535544 : cpu_run_cost += cpu_per_tuple * tuples_fetched;
785 :
786 : /* tlist eval costs are paid per output row, not per tuple scanned */
787 535544 : startup_cost += path->path.pathtarget->cost.startup;
788 535544 : cpu_run_cost += path->path.pathtarget->cost.per_tuple * path->path.rows;
789 :
790 : /* Adjust costing for parallelism, if used. */
791 535544 : if (path->path.parallel_workers > 0)
792 : {
793 10102 : double parallel_divisor = get_parallel_divisor(&path->path);
794 :
795 10102 : path->path.rows = clamp_row_est(path->path.rows / parallel_divisor);
796 :
797 : /* The CPU cost is divided among all the workers. */
798 10102 : cpu_run_cost /= parallel_divisor;
799 : }
800 :
801 535544 : run_cost += cpu_run_cost;
802 :
803 535544 : path->path.startup_cost = startup_cost;
804 535544 : path->path.total_cost = startup_cost + run_cost;
805 : }
806 :
807 : /*
808 : * extract_nonindex_conditions
809 : *
810 : * Given a list of quals to be enforced in an indexscan, extract the ones that
811 : * will have to be applied as qpquals (ie, the index machinery won't handle
812 : * them). Here we detect only whether a qual clause is directly redundant
813 : * with some indexclause. If the index path is chosen for use, createplan.c
814 : * will try a bit harder to get rid of redundant qual conditions; specifically
815 : * it will see if quals can be proven to be implied by the indexquals. But
816 : * it does not seem worth the cycles to try to factor that in at this stage,
817 : * since we're only trying to estimate qual eval costs. Otherwise this must
818 : * match the logic in create_indexscan_plan().
819 : *
820 : * qual_clauses, and the result, are lists of RestrictInfos.
821 : * indexclauses is a list of IndexClauses.
822 : */
823 : static List *
824 956854 : extract_nonindex_conditions(List *qual_clauses, List *indexclauses)
825 : {
826 956854 : List *result = NIL;
827 : ListCell *lc;
828 :
829 1982252 : foreach(lc, qual_clauses)
830 : {
831 1025398 : RestrictInfo *rinfo = lfirst_node(RestrictInfo, lc);
832 :
833 1025398 : if (rinfo->pseudoconstant)
834 10070 : continue; /* we may drop pseudoconstants here */
835 1015328 : if (is_redundant_with_indexclauses(rinfo, indexclauses))
836 602404 : continue; /* dup or derived from same EquivalenceClass */
837 : /* ... skip the predicate proof attempt createplan.c will try ... */
838 412924 : result = lappend(result, rinfo);
839 : }
840 956854 : return result;
841 : }
842 :
843 : /*
844 : * index_pages_fetched
845 : * Estimate the number of pages actually fetched after accounting for
846 : * cache effects.
847 : *
848 : * We use an approximation proposed by Mackert and Lohman, "Index Scans
849 : * Using a Finite LRU Buffer: A Validated I/O Model", ACM Transactions
850 : * on Database Systems, Vol. 14, No. 3, September 1989, Pages 401-424.
851 : * The Mackert and Lohman approximation is that the number of pages
852 : * fetched is
853 : * PF =
854 : * min(2TNs/(2T+Ns), T) when T <= b
855 : * 2TNs/(2T+Ns) when T > b and Ns <= 2Tb/(2T-b)
856 : * b + (Ns - 2Tb/(2T-b))*(T-b)/T when T > b and Ns > 2Tb/(2T-b)
857 : * where
858 : * T = # pages in table
859 : * N = # tuples in table
860 : * s = selectivity = fraction of table to be scanned
861 : * b = # buffer pages available (we include kernel space here)
862 : *
863 : * We assume that effective_cache_size is the total number of buffer pages
864 : * available for the whole query, and pro-rate that space across all the
865 : * tables in the query and the index currently under consideration. (This
866 : * ignores space needed for other indexes used by the query, but since we
867 : * don't know which indexes will get used, we can't estimate that very well;
868 : * and in any case counting all the tables may well be an overestimate, since
869 : * depending on the join plan not all the tables may be scanned concurrently.)
870 : *
871 : * The product Ns is the number of tuples fetched; we pass in that
872 : * product rather than calculating it here. "pages" is the number of pages
873 : * in the object under consideration (either an index or a table).
874 : * "index_pages" is the amount to add to the total table space, which was
875 : * computed for us by make_one_rel.
876 : *
877 : * Caller is expected to have ensured that tuples_fetched is greater than zero
878 : * and rounded to integer (see clamp_row_est). The result will likewise be
879 : * greater than zero and integral.
880 : */
881 : double
882 1130686 : index_pages_fetched(double tuples_fetched, BlockNumber pages,
883 : double index_pages, PlannerInfo *root)
884 : {
885 : double pages_fetched;
886 : double total_pages;
887 : double T,
888 : b;
889 :
890 : /* T is # pages in table, but don't allow it to be zero */
891 1130686 : T = (pages > 1) ? (double) pages : 1.0;
892 :
893 : /* Compute number of pages assumed to be competing for cache space */
894 1130686 : total_pages = root->total_table_pages + index_pages;
895 1130686 : total_pages = Max(total_pages, 1.0);
896 : Assert(T <= total_pages);
897 :
898 : /* b is pro-rated share of effective_cache_size */
899 1130686 : b = (double) effective_cache_size * T / total_pages;
900 :
901 : /* force it positive and integral */
902 1130686 : if (b <= 1.0)
903 0 : b = 1.0;
904 : else
905 1130686 : b = ceil(b);
906 :
907 : /* This part is the Mackert and Lohman formula */
908 1130686 : if (T <= b)
909 : {
910 1130686 : pages_fetched =
911 1130686 : (2.0 * T * tuples_fetched) / (2.0 * T + tuples_fetched);
912 1130686 : if (pages_fetched >= T)
913 659674 : pages_fetched = T;
914 : else
915 471012 : pages_fetched = ceil(pages_fetched);
916 : }
917 : else
918 : {
919 : double lim;
920 :
921 0 : lim = (2.0 * T * b) / (2.0 * T - b);
922 0 : if (tuples_fetched <= lim)
923 : {
924 0 : pages_fetched =
925 0 : (2.0 * T * tuples_fetched) / (2.0 * T + tuples_fetched);
926 : }
927 : else
928 : {
929 0 : pages_fetched =
930 0 : b + (tuples_fetched - lim) * (T - b) / T;
931 : }
932 0 : pages_fetched = ceil(pages_fetched);
933 : }
934 1130686 : return pages_fetched;
935 : }
936 :
937 : /*
938 : * get_indexpath_pages
939 : * Determine the total size of the indexes used in a bitmap index path.
940 : *
941 : * Note: if the same index is used more than once in a bitmap tree, we will
942 : * count it multiple times, which perhaps is the wrong thing ... but it's
943 : * not completely clear, and detecting duplicates is difficult, so ignore it
944 : * for now.
945 : */
946 : static double
947 190388 : get_indexpath_pages(Path *bitmapqual)
948 : {
949 190388 : double result = 0;
950 : ListCell *l;
951 :
952 190388 : if (IsA(bitmapqual, BitmapAndPath))
953 : {
954 22606 : BitmapAndPath *apath = (BitmapAndPath *) bitmapqual;
955 :
956 67818 : foreach(l, apath->bitmapquals)
957 : {
958 45212 : result += get_indexpath_pages((Path *) lfirst(l));
959 : }
960 : }
961 167782 : else if (IsA(bitmapqual, BitmapOrPath))
962 : {
963 70 : BitmapOrPath *opath = (BitmapOrPath *) bitmapqual;
964 :
965 222 : foreach(l, opath->bitmapquals)
966 : {
967 152 : result += get_indexpath_pages((Path *) lfirst(l));
968 : }
969 : }
970 167712 : else if (IsA(bitmapqual, IndexPath))
971 : {
972 167712 : IndexPath *ipath = (IndexPath *) bitmapqual;
973 :
974 167712 : result = (double) ipath->indexinfo->pages;
975 : }
976 : else
977 0 : elog(ERROR, "unrecognized node type: %d", nodeTag(bitmapqual));
978 :
979 190388 : return result;
980 : }
981 :
982 : /*
983 : * cost_bitmap_heap_scan
984 : * Determines and returns the cost of scanning a relation using a bitmap
985 : * index-then-heap plan.
986 : *
987 : * 'baserel' is the relation to be scanned
988 : * 'param_info' is the ParamPathInfo if this is a parameterized path, else NULL
989 : * 'bitmapqual' is a tree of IndexPaths, BitmapAndPaths, and BitmapOrPaths
990 : * 'loop_count' is the number of repetitions of the indexscan to factor into
991 : * estimates of caching behavior
992 : *
993 : * Note: the component IndexPaths in bitmapqual should have been costed
994 : * using the same loop_count.
995 : */
996 : void
997 542954 : cost_bitmap_heap_scan(Path *path, PlannerInfo *root, RelOptInfo *baserel,
998 : ParamPathInfo *param_info,
999 : Path *bitmapqual, double loop_count)
1000 : {
1001 542954 : Cost startup_cost = 0;
1002 542954 : Cost run_cost = 0;
1003 : Cost indexTotalCost;
1004 : QualCost qpqual_cost;
1005 : Cost cpu_per_tuple;
1006 : Cost cost_per_page;
1007 : Cost cpu_run_cost;
1008 : double tuples_fetched;
1009 : double pages_fetched;
1010 : double spc_seq_page_cost,
1011 : spc_random_page_cost;
1012 : double T;
1013 :
1014 : /* Should only be applied to base relations */
1015 : Assert(IsA(baserel, RelOptInfo));
1016 : Assert(baserel->relid > 0);
1017 : Assert(baserel->rtekind == RTE_RELATION);
1018 :
1019 : /* Mark the path with the correct row estimate */
1020 542954 : if (param_info)
1021 232376 : path->rows = param_info->ppi_rows;
1022 : else
1023 310578 : path->rows = baserel->rows;
1024 :
1025 542954 : pages_fetched = compute_bitmap_pages(root, baserel, bitmapqual,
1026 : loop_count, &indexTotalCost,
1027 : &tuples_fetched);
1028 :
1029 542954 : startup_cost += indexTotalCost;
1030 542954 : T = (baserel->pages > 1) ? (double) baserel->pages : 1.0;
1031 :
1032 : /* Fetch estimated page costs for tablespace containing table. */
1033 542954 : get_tablespace_page_costs(baserel->reltablespace,
1034 : &spc_random_page_cost,
1035 : &spc_seq_page_cost);
1036 :
1037 : /*
1038 : * For small numbers of pages we should charge spc_random_page_cost
1039 : * apiece, while if nearly all the table's pages are being read, it's more
1040 : * appropriate to charge spc_seq_page_cost apiece. The effect is
1041 : * nonlinear, too. For lack of a better idea, interpolate like this to
1042 : * determine the cost per page.
1043 : */
1044 542954 : if (pages_fetched >= 2.0)
1045 111960 : cost_per_page = spc_random_page_cost -
1046 111960 : (spc_random_page_cost - spc_seq_page_cost)
1047 111960 : * sqrt(pages_fetched / T);
1048 : else
1049 430994 : cost_per_page = spc_random_page_cost;
1050 :
1051 542954 : run_cost += pages_fetched * cost_per_page;
1052 :
1053 : /*
1054 : * Estimate CPU costs per tuple.
1055 : *
1056 : * Often the indexquals don't need to be rechecked at each tuple ... but
1057 : * not always, especially not if there are enough tuples involved that the
1058 : * bitmaps become lossy. For the moment, just assume they will be
1059 : * rechecked always. This means we charge the full freight for all the
1060 : * scan clauses.
1061 : */
1062 542954 : get_restriction_qual_cost(root, baserel, param_info, &qpqual_cost);
1063 :
1064 542954 : startup_cost += qpqual_cost.startup;
1065 542954 : cpu_per_tuple = cpu_tuple_cost + qpqual_cost.per_tuple;
1066 542954 : cpu_run_cost = cpu_per_tuple * tuples_fetched;
1067 :
1068 : /* Adjust costing for parallelism, if used. */
1069 542954 : if (path->parallel_workers > 0)
1070 : {
1071 4166 : double parallel_divisor = get_parallel_divisor(path);
1072 :
1073 : /* The CPU cost is divided among all the workers. */
1074 4166 : cpu_run_cost /= parallel_divisor;
1075 :
1076 4166 : path->rows = clamp_row_est(path->rows / parallel_divisor);
1077 : }
1078 :
1079 :
1080 542954 : run_cost += cpu_run_cost;
1081 :
1082 : /* tlist eval costs are paid per output row, not per tuple scanned */
1083 542954 : startup_cost += path->pathtarget->cost.startup;
1084 542954 : run_cost += path->pathtarget->cost.per_tuple * path->rows;
1085 :
1086 542954 : path->disabled_nodes = enable_bitmapscan ? 0 : 1;
1087 542954 : path->startup_cost = startup_cost;
1088 542954 : path->total_cost = startup_cost + run_cost;
1089 542954 : }
1090 :
1091 : /*
1092 : * cost_bitmap_tree_node
1093 : * Extract cost and selectivity from a bitmap tree node (index/and/or)
1094 : */
1095 : void
1096 1005158 : cost_bitmap_tree_node(Path *path, Cost *cost, Selectivity *selec)
1097 : {
1098 1005158 : if (IsA(path, IndexPath))
1099 : {
1100 951772 : *cost = ((IndexPath *) path)->indextotalcost;
1101 951772 : *selec = ((IndexPath *) path)->indexselectivity;
1102 :
1103 : /*
1104 : * Charge a small amount per retrieved tuple to reflect the costs of
1105 : * manipulating the bitmap. This is mostly to make sure that a bitmap
1106 : * scan doesn't look to be the same cost as an indexscan to retrieve a
1107 : * single tuple.
1108 : */
1109 951772 : *cost += 0.1 * cpu_operator_cost * path->rows;
1110 : }
1111 53386 : else if (IsA(path, BitmapAndPath))
1112 : {
1113 49850 : *cost = path->total_cost;
1114 49850 : *selec = ((BitmapAndPath *) path)->bitmapselectivity;
1115 : }
1116 3536 : else if (IsA(path, BitmapOrPath))
1117 : {
1118 3536 : *cost = path->total_cost;
1119 3536 : *selec = ((BitmapOrPath *) path)->bitmapselectivity;
1120 : }
1121 : else
1122 : {
1123 0 : elog(ERROR, "unrecognized node type: %d", nodeTag(path));
1124 : *cost = *selec = 0; /* keep compiler quiet */
1125 : }
1126 1005158 : }
1127 :
1128 : /*
1129 : * cost_bitmap_and_node
1130 : * Estimate the cost of a BitmapAnd node
1131 : *
1132 : * Note that this considers only the costs of index scanning and bitmap
1133 : * creation, not the eventual heap access. In that sense the object isn't
1134 : * truly a Path, but it has enough path-like properties (costs in particular)
1135 : * to warrant treating it as one. We don't bother to set the path rows field,
1136 : * however.
1137 : */
1138 : void
1139 49644 : cost_bitmap_and_node(BitmapAndPath *path, PlannerInfo *root)
1140 : {
1141 : Cost totalCost;
1142 : Selectivity selec;
1143 : ListCell *l;
1144 :
1145 : /*
1146 : * We estimate AND selectivity on the assumption that the inputs are
1147 : * independent. This is probably often wrong, but we don't have the info
1148 : * to do better.
1149 : *
1150 : * The runtime cost of the BitmapAnd itself is estimated at 100x
1151 : * cpu_operator_cost for each tbm_intersect needed. Probably too small,
1152 : * definitely too simplistic?
1153 : */
1154 49644 : totalCost = 0.0;
1155 49644 : selec = 1.0;
1156 148932 : foreach(l, path->bitmapquals)
1157 : {
1158 99288 : Path *subpath = (Path *) lfirst(l);
1159 : Cost subCost;
1160 : Selectivity subselec;
1161 :
1162 99288 : cost_bitmap_tree_node(subpath, &subCost, &subselec);
1163 :
1164 99288 : selec *= subselec;
1165 :
1166 99288 : totalCost += subCost;
1167 99288 : if (l != list_head(path->bitmapquals))
1168 49644 : totalCost += 100.0 * cpu_operator_cost;
1169 : }
1170 49644 : path->bitmapselectivity = selec;
1171 49644 : path->path.rows = 0; /* per above, not used */
1172 49644 : path->path.disabled_nodes = 0;
1173 49644 : path->path.startup_cost = totalCost;
1174 49644 : path->path.total_cost = totalCost;
1175 49644 : }
1176 :
1177 : /*
1178 : * cost_bitmap_or_node
1179 : * Estimate the cost of a BitmapOr node
1180 : *
1181 : * See comments for cost_bitmap_and_node.
1182 : */
1183 : void
1184 1016 : cost_bitmap_or_node(BitmapOrPath *path, PlannerInfo *root)
1185 : {
1186 : Cost totalCost;
1187 : Selectivity selec;
1188 : ListCell *l;
1189 :
1190 : /*
1191 : * We estimate OR selectivity on the assumption that the inputs are
1192 : * non-overlapping, since that's often the case in "x IN (list)" type
1193 : * situations. Of course, we clamp to 1.0 at the end.
1194 : *
1195 : * The runtime cost of the BitmapOr itself is estimated at 100x
1196 : * cpu_operator_cost for each tbm_union needed. Probably too small,
1197 : * definitely too simplistic? We are aware that the tbm_unions are
1198 : * optimized out when the inputs are BitmapIndexScans.
1199 : */
1200 1016 : totalCost = 0.0;
1201 1016 : selec = 0.0;
1202 2850 : foreach(l, path->bitmapquals)
1203 : {
1204 1834 : Path *subpath = (Path *) lfirst(l);
1205 : Cost subCost;
1206 : Selectivity subselec;
1207 :
1208 1834 : cost_bitmap_tree_node(subpath, &subCost, &subselec);
1209 :
1210 1834 : selec += subselec;
1211 :
1212 1834 : totalCost += subCost;
1213 1834 : if (l != list_head(path->bitmapquals) &&
1214 818 : !IsA(subpath, IndexPath))
1215 0 : totalCost += 100.0 * cpu_operator_cost;
1216 : }
1217 1016 : path->bitmapselectivity = Min(selec, 1.0);
1218 1016 : path->path.rows = 0; /* per above, not used */
1219 1016 : path->path.startup_cost = totalCost;
1220 1016 : path->path.total_cost = totalCost;
1221 1016 : }
1222 :
1223 : /*
1224 : * cost_tidscan
1225 : * Determines and returns the cost of scanning a relation using TIDs.
1226 : *
1227 : * 'baserel' is the relation to be scanned
1228 : * 'tidquals' is the list of TID-checkable quals
1229 : * 'param_info' is the ParamPathInfo if this is a parameterized path, else NULL
1230 : */
1231 : void
1232 872 : cost_tidscan(Path *path, PlannerInfo *root,
1233 : RelOptInfo *baserel, List *tidquals, ParamPathInfo *param_info)
1234 : {
1235 872 : Cost startup_cost = 0;
1236 872 : Cost run_cost = 0;
1237 : QualCost qpqual_cost;
1238 : Cost cpu_per_tuple;
1239 : QualCost tid_qual_cost;
1240 : double ntuples;
1241 : ListCell *l;
1242 : double spc_random_page_cost;
1243 :
1244 : /* Should only be applied to base relations */
1245 : Assert(baserel->relid > 0);
1246 : Assert(baserel->rtekind == RTE_RELATION);
1247 : Assert(tidquals != NIL);
1248 :
1249 : /* Mark the path with the correct row estimate */
1250 872 : if (param_info)
1251 144 : path->rows = param_info->ppi_rows;
1252 : else
1253 728 : path->rows = baserel->rows;
1254 :
1255 : /* Count how many tuples we expect to retrieve */
1256 872 : ntuples = 0;
1257 1770 : foreach(l, tidquals)
1258 : {
1259 898 : RestrictInfo *rinfo = lfirst_node(RestrictInfo, l);
1260 898 : Expr *qual = rinfo->clause;
1261 :
1262 : /*
1263 : * We must use a TID scan for CurrentOfExpr; in any other case, we
1264 : * should be generating a TID scan only if enable_tidscan=true. Also,
1265 : * if CurrentOfExpr is the qual, there should be only one.
1266 : */
1267 : Assert(enable_tidscan || IsA(qual, CurrentOfExpr));
1268 : Assert(list_length(tidquals) == 1 || !IsA(qual, CurrentOfExpr));
1269 :
1270 898 : if (IsA(qual, ScalarArrayOpExpr))
1271 : {
1272 : /* Each element of the array yields 1 tuple */
1273 50 : ScalarArrayOpExpr *saop = (ScalarArrayOpExpr *) qual;
1274 50 : Node *arraynode = (Node *) lsecond(saop->args);
1275 :
1276 50 : ntuples += estimate_array_length(root, arraynode);
1277 : }
1278 848 : else if (IsA(qual, CurrentOfExpr))
1279 : {
1280 : /* CURRENT OF yields 1 tuple */
1281 404 : ntuples++;
1282 : }
1283 : else
1284 : {
1285 : /* It's just CTID = something, count 1 tuple */
1286 444 : ntuples++;
1287 : }
1288 : }
1289 :
1290 : /*
1291 : * The TID qual expressions will be computed once, any other baserestrict
1292 : * quals once per retrieved tuple.
1293 : */
1294 872 : cost_qual_eval(&tid_qual_cost, tidquals, root);
1295 :
1296 : /* fetch estimated page cost for tablespace containing table */
1297 872 : get_tablespace_page_costs(baserel->reltablespace,
1298 : &spc_random_page_cost,
1299 : NULL);
1300 :
1301 : /* disk costs --- assume each tuple on a different page */
1302 872 : run_cost += spc_random_page_cost * ntuples;
1303 :
1304 : /* Add scanning CPU costs */
1305 872 : get_restriction_qual_cost(root, baserel, param_info, &qpqual_cost);
1306 :
1307 : /* XXX currently we assume TID quals are a subset of qpquals */
1308 872 : startup_cost += qpqual_cost.startup + tid_qual_cost.per_tuple;
1309 872 : cpu_per_tuple = cpu_tuple_cost + qpqual_cost.per_tuple -
1310 872 : tid_qual_cost.per_tuple;
1311 872 : run_cost += cpu_per_tuple * ntuples;
1312 :
1313 : /* tlist eval costs are paid per output row, not per tuple scanned */
1314 872 : startup_cost += path->pathtarget->cost.startup;
1315 872 : run_cost += path->pathtarget->cost.per_tuple * path->rows;
1316 :
1317 : /*
1318 : * There are assertions above verifying that we only reach this function
1319 : * either when enable_tidscan=true or when the TID scan is the only legal
1320 : * path, so it's safe to set disabled_nodes to zero here.
1321 : */
1322 872 : path->disabled_nodes = 0;
1323 872 : path->startup_cost = startup_cost;
1324 872 : path->total_cost = startup_cost + run_cost;
1325 872 : }
1326 :
1327 : /*
1328 : * cost_tidrangescan
1329 : * Determines and sets the costs of scanning a relation using a range of
1330 : * TIDs for 'path'
1331 : *
1332 : * 'baserel' is the relation to be scanned
1333 : * 'tidrangequals' is the list of TID-checkable range quals
1334 : * 'param_info' is the ParamPathInfo if this is a parameterized path, else NULL
1335 : */
1336 : void
1337 1944 : cost_tidrangescan(Path *path, PlannerInfo *root,
1338 : RelOptInfo *baserel, List *tidrangequals,
1339 : ParamPathInfo *param_info)
1340 : {
1341 : Selectivity selectivity;
1342 : double pages;
1343 1944 : Cost startup_cost = 0;
1344 1944 : Cost run_cost = 0;
1345 : QualCost qpqual_cost;
1346 : Cost cpu_per_tuple;
1347 : QualCost tid_qual_cost;
1348 : double ntuples;
1349 : double nseqpages;
1350 : double spc_random_page_cost;
1351 : double spc_seq_page_cost;
1352 :
1353 : /* Should only be applied to base relations */
1354 : Assert(baserel->relid > 0);
1355 : Assert(baserel->rtekind == RTE_RELATION);
1356 :
1357 : /* Mark the path with the correct row estimate */
1358 1944 : if (param_info)
1359 0 : path->rows = param_info->ppi_rows;
1360 : else
1361 1944 : path->rows = baserel->rows;
1362 :
1363 : /* Count how many tuples and pages we expect to scan */
1364 1944 : selectivity = clauselist_selectivity(root, tidrangequals, baserel->relid,
1365 : JOIN_INNER, NULL);
1366 1944 : pages = ceil(selectivity * baserel->pages);
1367 :
1368 1944 : if (pages <= 0.0)
1369 42 : pages = 1.0;
1370 :
1371 : /*
1372 : * The first page in a range requires a random seek, but each subsequent
1373 : * page is just a normal sequential page read. NOTE: it's desirable for
1374 : * TID Range Scans to cost more than the equivalent Sequential Scans,
1375 : * because Seq Scans have some performance advantages such as scan
1376 : * synchronization and parallelizability, and we'd prefer one of them to
1377 : * be picked unless a TID Range Scan really is better.
1378 : */
1379 1944 : ntuples = selectivity * baserel->tuples;
1380 1944 : nseqpages = pages - 1.0;
1381 :
1382 : /*
1383 : * The TID qual expressions will be computed once, any other baserestrict
1384 : * quals once per retrieved tuple.
1385 : */
1386 1944 : cost_qual_eval(&tid_qual_cost, tidrangequals, root);
1387 :
1388 : /* fetch estimated page cost for tablespace containing table */
1389 1944 : get_tablespace_page_costs(baserel->reltablespace,
1390 : &spc_random_page_cost,
1391 : &spc_seq_page_cost);
1392 :
1393 : /* disk costs; 1 random page and the remainder as seq pages */
1394 1944 : run_cost += spc_random_page_cost + spc_seq_page_cost * nseqpages;
1395 :
1396 : /* Add scanning CPU costs */
1397 1944 : get_restriction_qual_cost(root, baserel, param_info, &qpqual_cost);
1398 :
1399 : /*
1400 : * XXX currently we assume TID quals are a subset of qpquals at this
1401 : * point; they will be removed (if possible) when we create the plan, so
1402 : * we subtract their cost from the total qpqual cost. (If the TID quals
1403 : * can't be removed, this is a mistake and we're going to underestimate
1404 : * the CPU cost a bit.)
1405 : */
1406 1944 : startup_cost += qpqual_cost.startup + tid_qual_cost.per_tuple;
1407 1944 : cpu_per_tuple = cpu_tuple_cost + qpqual_cost.per_tuple -
1408 1944 : tid_qual_cost.per_tuple;
1409 1944 : run_cost += cpu_per_tuple * ntuples;
1410 :
1411 : /* tlist eval costs are paid per output row, not per tuple scanned */
1412 1944 : startup_cost += path->pathtarget->cost.startup;
1413 1944 : run_cost += path->pathtarget->cost.per_tuple * path->rows;
1414 :
1415 : /* we should not generate this path type when enable_tidscan=false */
1416 : Assert(enable_tidscan);
1417 1944 : path->disabled_nodes = 0;
1418 1944 : path->startup_cost = startup_cost;
1419 1944 : path->total_cost = startup_cost + run_cost;
1420 1944 : }
1421 :
1422 : /*
1423 : * cost_subqueryscan
1424 : * Determines and returns the cost of scanning a subquery RTE.
1425 : *
1426 : * 'baserel' is the relation to be scanned
1427 : * 'param_info' is the ParamPathInfo if this is a parameterized path, else NULL
1428 : * 'trivial_pathtarget' is true if the pathtarget is believed to be trivial.
1429 : */
1430 : void
1431 55828 : cost_subqueryscan(SubqueryScanPath *path, PlannerInfo *root,
1432 : RelOptInfo *baserel, ParamPathInfo *param_info,
1433 : bool trivial_pathtarget)
1434 : {
1435 : Cost startup_cost;
1436 : Cost run_cost;
1437 : List *qpquals;
1438 : QualCost qpqual_cost;
1439 : Cost cpu_per_tuple;
1440 :
1441 : /* Should only be applied to base relations that are subqueries */
1442 : Assert(baserel->relid > 0);
1443 : Assert(baserel->rtekind == RTE_SUBQUERY);
1444 :
1445 : /*
1446 : * We compute the rowcount estimate as the subplan's estimate times the
1447 : * selectivity of relevant restriction clauses. In simple cases this will
1448 : * come out the same as baserel->rows; but when dealing with parallelized
1449 : * paths we must do it like this to get the right answer.
1450 : */
1451 55828 : if (param_info)
1452 606 : qpquals = list_concat_copy(param_info->ppi_clauses,
1453 606 : baserel->baserestrictinfo);
1454 : else
1455 55222 : qpquals = baserel->baserestrictinfo;
1456 :
1457 55828 : path->path.rows = clamp_row_est(path->subpath->rows *
1458 55828 : clauselist_selectivity(root,
1459 : qpquals,
1460 : 0,
1461 : JOIN_INNER,
1462 : NULL));
1463 :
1464 : /*
1465 : * Cost of path is cost of evaluating the subplan, plus cost of evaluating
1466 : * any restriction clauses and tlist that will be attached to the
1467 : * SubqueryScan node, plus cpu_tuple_cost to account for selection and
1468 : * projection overhead.
1469 : */
1470 55828 : path->path.disabled_nodes = path->subpath->disabled_nodes;
1471 55828 : path->path.startup_cost = path->subpath->startup_cost;
1472 55828 : path->path.total_cost = path->subpath->total_cost;
1473 :
1474 : /*
1475 : * However, if there are no relevant restriction clauses and the
1476 : * pathtarget is trivial, then we expect that setrefs.c will optimize away
1477 : * the SubqueryScan plan node altogether, so we should just make its cost
1478 : * and rowcount equal to the input path's.
1479 : *
1480 : * Note: there are some edge cases where createplan.c will apply a
1481 : * different targetlist to the SubqueryScan node, thus falsifying our
1482 : * current estimate of whether the target is trivial, and making the cost
1483 : * estimate (though not the rowcount) wrong. It does not seem worth the
1484 : * extra complication to try to account for that exactly, especially since
1485 : * that behavior falsifies other cost estimates as well.
1486 : */
1487 55828 : if (qpquals == NIL && trivial_pathtarget)
1488 27088 : return;
1489 :
1490 28740 : get_restriction_qual_cost(root, baserel, param_info, &qpqual_cost);
1491 :
1492 28740 : startup_cost = qpqual_cost.startup;
1493 28740 : cpu_per_tuple = cpu_tuple_cost + qpqual_cost.per_tuple;
1494 28740 : run_cost = cpu_per_tuple * path->subpath->rows;
1495 :
1496 : /* tlist eval costs are paid per output row, not per tuple scanned */
1497 28740 : startup_cost += path->path.pathtarget->cost.startup;
1498 28740 : run_cost += path->path.pathtarget->cost.per_tuple * path->path.rows;
1499 :
1500 28740 : path->path.startup_cost += startup_cost;
1501 28740 : path->path.total_cost += startup_cost + run_cost;
1502 : }
1503 :
1504 : /*
1505 : * cost_functionscan
1506 : * Determines and returns the cost of scanning a function RTE.
1507 : *
1508 : * 'baserel' is the relation to be scanned
1509 : * 'param_info' is the ParamPathInfo if this is a parameterized path, else NULL
1510 : */
1511 : void
1512 52258 : cost_functionscan(Path *path, PlannerInfo *root,
1513 : RelOptInfo *baserel, ParamPathInfo *param_info)
1514 : {
1515 52258 : Cost startup_cost = 0;
1516 52258 : Cost run_cost = 0;
1517 : QualCost qpqual_cost;
1518 : Cost cpu_per_tuple;
1519 : RangeTblEntry *rte;
1520 : QualCost exprcost;
1521 :
1522 : /* Should only be applied to base relations that are functions */
1523 : Assert(baserel->relid > 0);
1524 52258 : rte = planner_rt_fetch(baserel->relid, root);
1525 : Assert(rte->rtekind == RTE_FUNCTION);
1526 :
1527 : /* Mark the path with the correct row estimate */
1528 52258 : if (param_info)
1529 8582 : path->rows = param_info->ppi_rows;
1530 : else
1531 43676 : path->rows = baserel->rows;
1532 :
1533 : /*
1534 : * Estimate costs of executing the function expression(s).
1535 : *
1536 : * Currently, nodeFunctionscan.c always executes the functions to
1537 : * completion before returning any rows, and caches the results in a
1538 : * tuplestore. So the function eval cost is all startup cost, and per-row
1539 : * costs are minimal.
1540 : *
1541 : * XXX in principle we ought to charge tuplestore spill costs if the
1542 : * number of rows is large. However, given how phony our rowcount
1543 : * estimates for functions tend to be, there's not a lot of point in that
1544 : * refinement right now.
1545 : */
1546 52258 : cost_qual_eval_node(&exprcost, (Node *) rte->functions, root);
1547 :
1548 52258 : startup_cost += exprcost.startup + exprcost.per_tuple;
1549 :
1550 : /* Add scanning CPU costs */
1551 52258 : get_restriction_qual_cost(root, baserel, param_info, &qpqual_cost);
1552 :
1553 52258 : startup_cost += qpqual_cost.startup;
1554 52258 : cpu_per_tuple = cpu_tuple_cost + qpqual_cost.per_tuple;
1555 52258 : run_cost += cpu_per_tuple * baserel->tuples;
1556 :
1557 : /* tlist eval costs are paid per output row, not per tuple scanned */
1558 52258 : startup_cost += path->pathtarget->cost.startup;
1559 52258 : run_cost += path->pathtarget->cost.per_tuple * path->rows;
1560 :
1561 52258 : path->disabled_nodes = 0;
1562 52258 : path->startup_cost = startup_cost;
1563 52258 : path->total_cost = startup_cost + run_cost;
1564 52258 : }
1565 :
1566 : /*
1567 : * cost_tablefuncscan
1568 : * Determines and returns the cost of scanning a table function.
1569 : *
1570 : * 'baserel' is the relation to be scanned
1571 : * 'param_info' is the ParamPathInfo if this is a parameterized path, else NULL
1572 : */
1573 : void
1574 626 : cost_tablefuncscan(Path *path, PlannerInfo *root,
1575 : RelOptInfo *baserel, ParamPathInfo *param_info)
1576 : {
1577 626 : Cost startup_cost = 0;
1578 626 : Cost run_cost = 0;
1579 : QualCost qpqual_cost;
1580 : Cost cpu_per_tuple;
1581 : RangeTblEntry *rte;
1582 : QualCost exprcost;
1583 :
1584 : /* Should only be applied to base relations that are functions */
1585 : Assert(baserel->relid > 0);
1586 626 : rte = planner_rt_fetch(baserel->relid, root);
1587 : Assert(rte->rtekind == RTE_TABLEFUNC);
1588 :
1589 : /* Mark the path with the correct row estimate */
1590 626 : if (param_info)
1591 234 : path->rows = param_info->ppi_rows;
1592 : else
1593 392 : path->rows = baserel->rows;
1594 :
1595 : /*
1596 : * Estimate costs of executing the table func expression(s).
1597 : *
1598 : * XXX in principle we ought to charge tuplestore spill costs if the
1599 : * number of rows is large. However, given how phony our rowcount
1600 : * estimates for tablefuncs tend to be, there's not a lot of point in that
1601 : * refinement right now.
1602 : */
1603 626 : cost_qual_eval_node(&exprcost, (Node *) rte->tablefunc, root);
1604 :
1605 626 : startup_cost += exprcost.startup + exprcost.per_tuple;
1606 :
1607 : /* Add scanning CPU costs */
1608 626 : get_restriction_qual_cost(root, baserel, param_info, &qpqual_cost);
1609 :
1610 626 : startup_cost += qpqual_cost.startup;
1611 626 : cpu_per_tuple = cpu_tuple_cost + qpqual_cost.per_tuple;
1612 626 : run_cost += cpu_per_tuple * baserel->tuples;
1613 :
1614 : /* tlist eval costs are paid per output row, not per tuple scanned */
1615 626 : startup_cost += path->pathtarget->cost.startup;
1616 626 : run_cost += path->pathtarget->cost.per_tuple * path->rows;
1617 :
1618 626 : path->disabled_nodes = 0;
1619 626 : path->startup_cost = startup_cost;
1620 626 : path->total_cost = startup_cost + run_cost;
1621 626 : }
1622 :
1623 : /*
1624 : * cost_valuesscan
1625 : * Determines and returns the cost of scanning a VALUES RTE.
1626 : *
1627 : * 'baserel' is the relation to be scanned
1628 : * 'param_info' is the ParamPathInfo if this is a parameterized path, else NULL
1629 : */
1630 : void
1631 8286 : cost_valuesscan(Path *path, PlannerInfo *root,
1632 : RelOptInfo *baserel, ParamPathInfo *param_info)
1633 : {
1634 8286 : Cost startup_cost = 0;
1635 8286 : Cost run_cost = 0;
1636 : QualCost qpqual_cost;
1637 : Cost cpu_per_tuple;
1638 :
1639 : /* Should only be applied to base relations that are values lists */
1640 : Assert(baserel->relid > 0);
1641 : Assert(baserel->rtekind == RTE_VALUES);
1642 :
1643 : /* Mark the path with the correct row estimate */
1644 8286 : if (param_info)
1645 66 : path->rows = param_info->ppi_rows;
1646 : else
1647 8220 : path->rows = baserel->rows;
1648 :
1649 : /*
1650 : * For now, estimate list evaluation cost at one operator eval per list
1651 : * (probably pretty bogus, but is it worth being smarter?)
1652 : */
1653 8286 : cpu_per_tuple = cpu_operator_cost;
1654 :
1655 : /* Add scanning CPU costs */
1656 8286 : get_restriction_qual_cost(root, baserel, param_info, &qpqual_cost);
1657 :
1658 8286 : startup_cost += qpqual_cost.startup;
1659 8286 : cpu_per_tuple += cpu_tuple_cost + qpqual_cost.per_tuple;
1660 8286 : run_cost += cpu_per_tuple * baserel->tuples;
1661 :
1662 : /* tlist eval costs are paid per output row, not per tuple scanned */
1663 8286 : startup_cost += path->pathtarget->cost.startup;
1664 8286 : run_cost += path->pathtarget->cost.per_tuple * path->rows;
1665 :
1666 8286 : path->disabled_nodes = 0;
1667 8286 : path->startup_cost = startup_cost;
1668 8286 : path->total_cost = startup_cost + run_cost;
1669 8286 : }
1670 :
1671 : /*
1672 : * cost_ctescan
1673 : * Determines and returns the cost of scanning a CTE RTE.
1674 : *
1675 : * Note: this is used for both self-reference and regular CTEs; the
1676 : * possible cost differences are below the threshold of what we could
1677 : * estimate accurately anyway. Note that the costs of evaluating the
1678 : * referenced CTE query are added into the final plan as initplan costs,
1679 : * and should NOT be counted here.
1680 : */
1681 : void
1682 5194 : cost_ctescan(Path *path, PlannerInfo *root,
1683 : RelOptInfo *baserel, ParamPathInfo *param_info)
1684 : {
1685 5194 : Cost startup_cost = 0;
1686 5194 : Cost run_cost = 0;
1687 : QualCost qpqual_cost;
1688 : Cost cpu_per_tuple;
1689 :
1690 : /* Should only be applied to base relations that are CTEs */
1691 : Assert(baserel->relid > 0);
1692 : Assert(baserel->rtekind == RTE_CTE);
1693 :
1694 : /* Mark the path with the correct row estimate */
1695 5194 : if (param_info)
1696 0 : path->rows = param_info->ppi_rows;
1697 : else
1698 5194 : path->rows = baserel->rows;
1699 :
1700 : /* Charge one CPU tuple cost per row for tuplestore manipulation */
1701 5194 : cpu_per_tuple = cpu_tuple_cost;
1702 :
1703 : /* Add scanning CPU costs */
1704 5194 : get_restriction_qual_cost(root, baserel, param_info, &qpqual_cost);
1705 :
1706 5194 : startup_cost += qpqual_cost.startup;
1707 5194 : cpu_per_tuple += cpu_tuple_cost + qpqual_cost.per_tuple;
1708 5194 : run_cost += cpu_per_tuple * baserel->tuples;
1709 :
1710 : /* tlist eval costs are paid per output row, not per tuple scanned */
1711 5194 : startup_cost += path->pathtarget->cost.startup;
1712 5194 : run_cost += path->pathtarget->cost.per_tuple * path->rows;
1713 :
1714 5194 : path->disabled_nodes = 0;
1715 5194 : path->startup_cost = startup_cost;
1716 5194 : path->total_cost = startup_cost + run_cost;
1717 5194 : }
1718 :
1719 : /*
1720 : * cost_namedtuplestorescan
1721 : * Determines and returns the cost of scanning a named tuplestore.
1722 : */
1723 : void
1724 474 : cost_namedtuplestorescan(Path *path, PlannerInfo *root,
1725 : RelOptInfo *baserel, ParamPathInfo *param_info)
1726 : {
1727 474 : Cost startup_cost = 0;
1728 474 : Cost run_cost = 0;
1729 : QualCost qpqual_cost;
1730 : Cost cpu_per_tuple;
1731 :
1732 : /* Should only be applied to base relations that are Tuplestores */
1733 : Assert(baserel->relid > 0);
1734 : Assert(baserel->rtekind == RTE_NAMEDTUPLESTORE);
1735 :
1736 : /* Mark the path with the correct row estimate */
1737 474 : if (param_info)
1738 0 : path->rows = param_info->ppi_rows;
1739 : else
1740 474 : path->rows = baserel->rows;
1741 :
1742 : /* Charge one CPU tuple cost per row for tuplestore manipulation */
1743 474 : cpu_per_tuple = cpu_tuple_cost;
1744 :
1745 : /* Add scanning CPU costs */
1746 474 : get_restriction_qual_cost(root, baserel, param_info, &qpqual_cost);
1747 :
1748 474 : startup_cost += qpqual_cost.startup;
1749 474 : cpu_per_tuple += cpu_tuple_cost + qpqual_cost.per_tuple;
1750 474 : run_cost += cpu_per_tuple * baserel->tuples;
1751 :
1752 474 : path->disabled_nodes = 0;
1753 474 : path->startup_cost = startup_cost;
1754 474 : path->total_cost = startup_cost + run_cost;
1755 474 : }
1756 :
1757 : /*
1758 : * cost_resultscan
1759 : * Determines and returns the cost of scanning an RTE_RESULT relation.
1760 : */
1761 : void
1762 4268 : cost_resultscan(Path *path, PlannerInfo *root,
1763 : RelOptInfo *baserel, ParamPathInfo *param_info)
1764 : {
1765 4268 : Cost startup_cost = 0;
1766 4268 : Cost run_cost = 0;
1767 : QualCost qpqual_cost;
1768 : Cost cpu_per_tuple;
1769 :
1770 : /* Should only be applied to RTE_RESULT base relations */
1771 : Assert(baserel->relid > 0);
1772 : Assert(baserel->rtekind == RTE_RESULT);
1773 :
1774 : /* Mark the path with the correct row estimate */
1775 4268 : if (param_info)
1776 156 : path->rows = param_info->ppi_rows;
1777 : else
1778 4112 : path->rows = baserel->rows;
1779 :
1780 : /* We charge qual cost plus cpu_tuple_cost */
1781 4268 : get_restriction_qual_cost(root, baserel, param_info, &qpqual_cost);
1782 :
1783 4268 : startup_cost += qpqual_cost.startup;
1784 4268 : cpu_per_tuple = cpu_tuple_cost + qpqual_cost.per_tuple;
1785 4268 : run_cost += cpu_per_tuple * baserel->tuples;
1786 :
1787 4268 : path->disabled_nodes = 0;
1788 4268 : path->startup_cost = startup_cost;
1789 4268 : path->total_cost = startup_cost + run_cost;
1790 4268 : }
1791 :
1792 : /*
1793 : * cost_recursive_union
1794 : * Determines and returns the cost of performing a recursive union,
1795 : * and also the estimated output size.
1796 : *
1797 : * We are given Paths for the nonrecursive and recursive terms.
1798 : */
1799 : void
1800 928 : cost_recursive_union(Path *runion, Path *nrterm, Path *rterm)
1801 : {
1802 : Cost startup_cost;
1803 : Cost total_cost;
1804 : double total_rows;
1805 :
1806 : /* We probably have decent estimates for the non-recursive term */
1807 928 : startup_cost = nrterm->startup_cost;
1808 928 : total_cost = nrterm->total_cost;
1809 928 : total_rows = nrterm->rows;
1810 :
1811 : /*
1812 : * We arbitrarily assume that about 10 recursive iterations will be
1813 : * needed, and that we've managed to get a good fix on the cost and output
1814 : * size of each one of them. These are mighty shaky assumptions but it's
1815 : * hard to see how to do better.
1816 : */
1817 928 : total_cost += 10 * rterm->total_cost;
1818 928 : total_rows += 10 * rterm->rows;
1819 :
1820 : /*
1821 : * Also charge cpu_tuple_cost per row to account for the costs of
1822 : * manipulating the tuplestores. (We don't worry about possible
1823 : * spill-to-disk costs.)
1824 : */
1825 928 : total_cost += cpu_tuple_cost * total_rows;
1826 :
1827 928 : runion->disabled_nodes = nrterm->disabled_nodes + rterm->disabled_nodes;
1828 928 : runion->startup_cost = startup_cost;
1829 928 : runion->total_cost = total_cost;
1830 928 : runion->rows = total_rows;
1831 928 : runion->pathtarget->width = Max(nrterm->pathtarget->width,
1832 : rterm->pathtarget->width);
1833 928 : }
1834 :
1835 : /*
1836 : * cost_tuplesort
1837 : * Determines and returns the cost of sorting a relation using tuplesort,
1838 : * not including the cost of reading the input data.
1839 : *
1840 : * If the total volume of data to sort is less than sort_mem, we will do
1841 : * an in-memory sort, which requires no I/O and about t*log2(t) tuple
1842 : * comparisons for t tuples.
1843 : *
1844 : * If the total volume exceeds sort_mem, we switch to a tape-style merge
1845 : * algorithm. There will still be about t*log2(t) tuple comparisons in
1846 : * total, but we will also need to write and read each tuple once per
1847 : * merge pass. We expect about ceil(logM(r)) merge passes where r is the
1848 : * number of initial runs formed and M is the merge order used by tuplesort.c.
1849 : * Since the average initial run should be about sort_mem, we have
1850 : * disk traffic = 2 * relsize * ceil(logM(p / sort_mem))
1851 : * cpu = comparison_cost * t * log2(t)
1852 : *
1853 : * If the sort is bounded (i.e., only the first k result tuples are needed)
1854 : * and k tuples can fit into sort_mem, we use a heap method that keeps only
1855 : * k tuples in the heap; this will require about t*log2(k) tuple comparisons.
1856 : *
1857 : * The disk traffic is assumed to be 3/4ths sequential and 1/4th random
1858 : * accesses (XXX can't we refine that guess?)
1859 : *
1860 : * By default, we charge two operator evals per tuple comparison, which should
1861 : * be in the right ballpark in most cases. The caller can tweak this by
1862 : * specifying nonzero comparison_cost; typically that's used for any extra
1863 : * work that has to be done to prepare the inputs to the comparison operators.
1864 : *
1865 : * 'tuples' is the number of tuples in the relation
1866 : * 'width' is the average tuple width in bytes
1867 : * 'comparison_cost' is the extra cost per comparison, if any
1868 : * 'sort_mem' is the number of kilobytes of work memory allowed for the sort
1869 : * 'limit_tuples' is the bound on the number of output tuples; -1 if no bound
1870 : */
1871 : static void
1872 2080776 : cost_tuplesort(Cost *startup_cost, Cost *run_cost,
1873 : double tuples, int width,
1874 : Cost comparison_cost, int sort_mem,
1875 : double limit_tuples)
1876 : {
1877 2080776 : double input_bytes = relation_byte_size(tuples, width);
1878 : double output_bytes;
1879 : double output_tuples;
1880 2080776 : int64 sort_mem_bytes = sort_mem * (int64) 1024;
1881 :
1882 : /*
1883 : * We want to be sure the cost of a sort is never estimated as zero, even
1884 : * if passed-in tuple count is zero. Besides, mustn't do log(0)...
1885 : */
1886 2080776 : if (tuples < 2.0)
1887 548998 : tuples = 2.0;
1888 :
1889 : /* Include the default cost-per-comparison */
1890 2080776 : comparison_cost += 2.0 * cpu_operator_cost;
1891 :
1892 : /* Do we have a useful LIMIT? */
1893 2080776 : if (limit_tuples > 0 && limit_tuples < tuples)
1894 : {
1895 1830 : output_tuples = limit_tuples;
1896 1830 : output_bytes = relation_byte_size(output_tuples, width);
1897 : }
1898 : else
1899 : {
1900 2078946 : output_tuples = tuples;
1901 2078946 : output_bytes = input_bytes;
1902 : }
1903 :
1904 2080776 : if (output_bytes > sort_mem_bytes)
1905 : {
1906 : /*
1907 : * We'll have to use a disk-based sort of all the tuples
1908 : */
1909 18448 : double npages = ceil(input_bytes / BLCKSZ);
1910 18448 : double nruns = input_bytes / sort_mem_bytes;
1911 18448 : double mergeorder = tuplesort_merge_order(sort_mem_bytes);
1912 : double log_runs;
1913 : double npageaccesses;
1914 :
1915 : /*
1916 : * CPU costs
1917 : *
1918 : * Assume about N log2 N comparisons
1919 : */
1920 18448 : *startup_cost = comparison_cost * tuples * LOG2(tuples);
1921 :
1922 : /* Disk costs */
1923 :
1924 : /* Compute logM(r) as log(r) / log(M) */
1925 18448 : if (nruns > mergeorder)
1926 4676 : log_runs = ceil(log(nruns) / log(mergeorder));
1927 : else
1928 13772 : log_runs = 1.0;
1929 18448 : npageaccesses = 2.0 * npages * log_runs;
1930 : /* Assume 3/4ths of accesses are sequential, 1/4th are not */
1931 18448 : *startup_cost += npageaccesses *
1932 18448 : (seq_page_cost * 0.75 + random_page_cost * 0.25);
1933 : }
1934 2062328 : else if (tuples > 2 * output_tuples || input_bytes > sort_mem_bytes)
1935 : {
1936 : /*
1937 : * We'll use a bounded heap-sort keeping just K tuples in memory, for
1938 : * a total number of tuple comparisons of N log2 K; but the constant
1939 : * factor is a bit higher than for quicksort. Tweak it so that the
1940 : * cost curve is continuous at the crossover point.
1941 : */
1942 1352 : *startup_cost = comparison_cost * tuples * LOG2(2.0 * output_tuples);
1943 : }
1944 : else
1945 : {
1946 : /* We'll use plain quicksort on all the input tuples */
1947 2060976 : *startup_cost = comparison_cost * tuples * LOG2(tuples);
1948 : }
1949 :
1950 : /*
1951 : * Also charge a small amount (arbitrarily set equal to operator cost) per
1952 : * extracted tuple. We don't charge cpu_tuple_cost because a Sort node
1953 : * doesn't do qual-checking or projection, so it has less overhead than
1954 : * most plan nodes. Note it's correct to use tuples not output_tuples
1955 : * here --- the upper LIMIT will pro-rate the run cost so we'd be double
1956 : * counting the LIMIT otherwise.
1957 : */
1958 2080776 : *run_cost = cpu_operator_cost * tuples;
1959 2080776 : }
1960 :
1961 : /*
1962 : * cost_incremental_sort
1963 : * Determines and returns the cost of sorting a relation incrementally, when
1964 : * the input path is presorted by a prefix of the pathkeys.
1965 : *
1966 : * 'presorted_keys' is the number of leading pathkeys by which the input path
1967 : * is sorted.
1968 : *
1969 : * We estimate the number of groups into which the relation is divided by the
1970 : * leading pathkeys, and then calculate the cost of sorting a single group
1971 : * with tuplesort using cost_tuplesort().
1972 : */
1973 : void
1974 12124 : cost_incremental_sort(Path *path,
1975 : PlannerInfo *root, List *pathkeys, int presorted_keys,
1976 : int input_disabled_nodes,
1977 : Cost input_startup_cost, Cost input_total_cost,
1978 : double input_tuples, int width, Cost comparison_cost, int sort_mem,
1979 : double limit_tuples)
1980 : {
1981 : Cost startup_cost,
1982 : run_cost,
1983 12124 : input_run_cost = input_total_cost - input_startup_cost;
1984 : double group_tuples,
1985 : input_groups;
1986 : Cost group_startup_cost,
1987 : group_run_cost,
1988 : group_input_run_cost;
1989 12124 : List *presortedExprs = NIL;
1990 : ListCell *l;
1991 12124 : bool unknown_varno = false;
1992 :
1993 : Assert(presorted_keys > 0 && presorted_keys < list_length(pathkeys));
1994 :
1995 : /*
1996 : * We want to be sure the cost of a sort is never estimated as zero, even
1997 : * if passed-in tuple count is zero. Besides, mustn't do log(0)...
1998 : */
1999 12124 : if (input_tuples < 2.0)
2000 6798 : input_tuples = 2.0;
2001 :
2002 : /* Default estimate of number of groups, capped to one group per row. */
2003 12124 : input_groups = Min(input_tuples, DEFAULT_NUM_DISTINCT);
2004 :
2005 : /*
2006 : * Extract presorted keys as list of expressions.
2007 : *
2008 : * We need to be careful about Vars containing "varno 0" which might have
2009 : * been introduced by generate_append_tlist, which would confuse
2010 : * estimate_num_groups (in fact it'd fail for such expressions). See
2011 : * recurse_set_operations which has to deal with the same issue.
2012 : *
2013 : * Unlike recurse_set_operations we can't access the original target list
2014 : * here, and even if we could it's not very clear how useful would that be
2015 : * for a set operation combining multiple tables. So we simply detect if
2016 : * there are any expressions with "varno 0" and use the default
2017 : * DEFAULT_NUM_DISTINCT in that case.
2018 : *
2019 : * We might also use either 1.0 (a single group) or input_tuples (each row
2020 : * being a separate group), pretty much the worst and best case for
2021 : * incremental sort. But those are extreme cases and using something in
2022 : * between seems reasonable. Furthermore, generate_append_tlist is used
2023 : * for set operations, which are likely to produce mostly unique output
2024 : * anyway - from that standpoint the DEFAULT_NUM_DISTINCT is defensive
2025 : * while maintaining lower startup cost.
2026 : */
2027 12220 : foreach(l, pathkeys)
2028 : {
2029 12220 : PathKey *key = (PathKey *) lfirst(l);
2030 12220 : EquivalenceMember *member = (EquivalenceMember *)
2031 12220 : linitial(key->pk_eclass->ec_members);
2032 :
2033 : /*
2034 : * Check if the expression contains Var with "varno 0" so that we
2035 : * don't call estimate_num_groups in that case.
2036 : */
2037 12220 : if (bms_is_member(0, pull_varnos(root, (Node *) member->em_expr)))
2038 : {
2039 10 : unknown_varno = true;
2040 10 : break;
2041 : }
2042 :
2043 : /* expression not containing any Vars with "varno 0" */
2044 12210 : presortedExprs = lappend(presortedExprs, member->em_expr);
2045 :
2046 12210 : if (foreach_current_index(l) + 1 >= presorted_keys)
2047 12114 : break;
2048 : }
2049 :
2050 : /* Estimate the number of groups with equal presorted keys. */
2051 12124 : if (!unknown_varno)
2052 12114 : input_groups = estimate_num_groups(root, presortedExprs, input_tuples,
2053 : NULL, NULL);
2054 :
2055 12124 : group_tuples = input_tuples / input_groups;
2056 12124 : group_input_run_cost = input_run_cost / input_groups;
2057 :
2058 : /*
2059 : * Estimate the average cost of sorting of one group where presorted keys
2060 : * are equal.
2061 : */
2062 12124 : cost_tuplesort(&group_startup_cost, &group_run_cost,
2063 : group_tuples, width, comparison_cost, sort_mem,
2064 : limit_tuples);
2065 :
2066 : /*
2067 : * Startup cost of incremental sort is the startup cost of its first group
2068 : * plus the cost of its input.
2069 : */
2070 12124 : startup_cost = group_startup_cost + input_startup_cost +
2071 : group_input_run_cost;
2072 :
2073 : /*
2074 : * After we started producing tuples from the first group, the cost of
2075 : * producing all the tuples is given by the cost to finish processing this
2076 : * group, plus the total cost to process the remaining groups, plus the
2077 : * remaining cost of input.
2078 : */
2079 12124 : run_cost = group_run_cost + (group_run_cost + group_startup_cost) *
2080 12124 : (input_groups - 1) + group_input_run_cost * (input_groups - 1);
2081 :
2082 : /*
2083 : * Incremental sort adds some overhead by itself. Firstly, it has to
2084 : * detect the sort groups. This is roughly equal to one extra copy and
2085 : * comparison per tuple.
2086 : */
2087 12124 : run_cost += (cpu_tuple_cost + comparison_cost) * input_tuples;
2088 :
2089 : /*
2090 : * Additionally, we charge double cpu_tuple_cost for each input group to
2091 : * account for the tuplesort_reset that's performed after each group.
2092 : */
2093 12124 : run_cost += 2.0 * cpu_tuple_cost * input_groups;
2094 :
2095 12124 : path->rows = input_tuples;
2096 :
2097 : /* should not generate these paths when enable_incremental_sort=false */
2098 : Assert(enable_incremental_sort);
2099 12124 : path->disabled_nodes = input_disabled_nodes;
2100 :
2101 12124 : path->startup_cost = startup_cost;
2102 12124 : path->total_cost = startup_cost + run_cost;
2103 12124 : }
2104 :
2105 : /*
2106 : * cost_sort
2107 : * Determines and returns the cost of sorting a relation, including
2108 : * the cost of reading the input data.
2109 : *
2110 : * NOTE: some callers currently pass NIL for pathkeys because they
2111 : * can't conveniently supply the sort keys. Since this routine doesn't
2112 : * currently do anything with pathkeys anyway, that doesn't matter...
2113 : * but if it ever does, it should react gracefully to lack of key data.
2114 : * (Actually, the thing we'd most likely be interested in is just the number
2115 : * of sort keys, which all callers *could* supply.)
2116 : */
2117 : void
2118 2068652 : cost_sort(Path *path, PlannerInfo *root,
2119 : List *pathkeys, int input_disabled_nodes,
2120 : Cost input_cost, double tuples, int width,
2121 : Cost comparison_cost, int sort_mem,
2122 : double limit_tuples)
2123 :
2124 : {
2125 : Cost startup_cost;
2126 : Cost run_cost;
2127 :
2128 2068652 : cost_tuplesort(&startup_cost, &run_cost,
2129 : tuples, width,
2130 : comparison_cost, sort_mem,
2131 : limit_tuples);
2132 :
2133 2068652 : startup_cost += input_cost;
2134 :
2135 2068652 : path->rows = tuples;
2136 2068652 : path->disabled_nodes = input_disabled_nodes + (enable_sort ? 0 : 1);
2137 2068652 : path->startup_cost = startup_cost;
2138 2068652 : path->total_cost = startup_cost + run_cost;
2139 2068652 : }
2140 :
2141 : /*
2142 : * append_nonpartial_cost
2143 : * Estimate the cost of the non-partial paths in a Parallel Append.
2144 : * The non-partial paths are assumed to be the first "numpaths" paths
2145 : * from the subpaths list, and to be in order of decreasing cost.
2146 : */
2147 : static Cost
2148 25564 : append_nonpartial_cost(List *subpaths, int numpaths, int parallel_workers)
2149 : {
2150 : Cost *costarr;
2151 : int arrlen;
2152 : ListCell *l;
2153 : ListCell *cell;
2154 : int path_index;
2155 : int min_index;
2156 : int max_index;
2157 :
2158 25564 : if (numpaths == 0)
2159 20598 : return 0;
2160 :
2161 : /*
2162 : * Array length is number of workers or number of relevant paths,
2163 : * whichever is less.
2164 : */
2165 4966 : arrlen = Min(parallel_workers, numpaths);
2166 4966 : costarr = (Cost *) palloc(sizeof(Cost) * arrlen);
2167 :
2168 : /* The first few paths will each be claimed by a different worker. */
2169 4966 : path_index = 0;
2170 14424 : foreach(cell, subpaths)
2171 : {
2172 10756 : Path *subpath = (Path *) lfirst(cell);
2173 :
2174 10756 : if (path_index == arrlen)
2175 1298 : break;
2176 9458 : costarr[path_index++] = subpath->total_cost;
2177 : }
2178 :
2179 : /*
2180 : * Since subpaths are sorted by decreasing cost, the last one will have
2181 : * the minimum cost.
2182 : */
2183 4966 : min_index = arrlen - 1;
2184 :
2185 : /*
2186 : * For each of the remaining subpaths, add its cost to the array element
2187 : * with minimum cost.
2188 : */
2189 9674 : for_each_cell(l, subpaths, cell)
2190 : {
2191 5254 : Path *subpath = (Path *) lfirst(l);
2192 :
2193 : /* Consider only the non-partial paths */
2194 5254 : if (path_index++ == numpaths)
2195 546 : break;
2196 :
2197 4708 : costarr[min_index] += subpath->total_cost;
2198 :
2199 : /* Update the new min cost array index */
2200 4708 : min_index = 0;
2201 14160 : for (int i = 0; i < arrlen; i++)
2202 : {
2203 9452 : if (costarr[i] < costarr[min_index])
2204 1514 : min_index = i;
2205 : }
2206 : }
2207 :
2208 : /* Return the highest cost from the array */
2209 4966 : max_index = 0;
2210 14424 : for (int i = 0; i < arrlen; i++)
2211 : {
2212 9458 : if (costarr[i] > costarr[max_index])
2213 414 : max_index = i;
2214 : }
2215 :
2216 4966 : return costarr[max_index];
2217 : }
2218 :
2219 : /*
2220 : * cost_append
2221 : * Determines and returns the cost of an Append node.
2222 : */
2223 : void
2224 69582 : cost_append(AppendPath *apath, PlannerInfo *root)
2225 : {
2226 : ListCell *l;
2227 :
2228 69582 : apath->path.disabled_nodes = 0;
2229 69582 : apath->path.startup_cost = 0;
2230 69582 : apath->path.total_cost = 0;
2231 69582 : apath->path.rows = 0;
2232 :
2233 69582 : if (apath->subpaths == NIL)
2234 2012 : return;
2235 :
2236 67570 : if (!apath->path.parallel_aware)
2237 : {
2238 42006 : List *pathkeys = apath->path.pathkeys;
2239 :
2240 42006 : if (pathkeys == NIL)
2241 : {
2242 39862 : Path *firstsubpath = (Path *) linitial(apath->subpaths);
2243 :
2244 : /*
2245 : * For an unordered, non-parallel-aware Append we take the startup
2246 : * cost as the startup cost of the first subpath.
2247 : */
2248 39862 : apath->path.startup_cost = firstsubpath->startup_cost;
2249 :
2250 : /*
2251 : * Compute rows, number of disabled nodes, and total cost as sums
2252 : * of underlying subplan values.
2253 : */
2254 155712 : foreach(l, apath->subpaths)
2255 : {
2256 115850 : Path *subpath = (Path *) lfirst(l);
2257 :
2258 115850 : apath->path.rows += subpath->rows;
2259 115850 : apath->path.disabled_nodes += subpath->disabled_nodes;
2260 115850 : apath->path.total_cost += subpath->total_cost;
2261 : }
2262 : }
2263 : else
2264 : {
2265 : /*
2266 : * For an ordered, non-parallel-aware Append we take the startup
2267 : * cost as the sum of the subpath startup costs. This ensures
2268 : * that we don't underestimate the startup cost when a query's
2269 : * LIMIT is such that several of the children have to be run to
2270 : * satisfy it. This might be overkill --- another plausible hack
2271 : * would be to take the Append's startup cost as the maximum of
2272 : * the child startup costs. But we don't want to risk believing
2273 : * that an ORDER BY LIMIT query can be satisfied at small cost
2274 : * when the first child has small startup cost but later ones
2275 : * don't. (If we had the ability to deal with nonlinear cost
2276 : * interpolation for partial retrievals, we would not need to be
2277 : * so conservative about this.)
2278 : *
2279 : * This case is also different from the above in that we have to
2280 : * account for possibly injecting sorts into subpaths that aren't
2281 : * natively ordered.
2282 : */
2283 8340 : foreach(l, apath->subpaths)
2284 : {
2285 6196 : Path *subpath = (Path *) lfirst(l);
2286 : int presorted_keys;
2287 : Path sort_path; /* dummy for result of
2288 : * cost_sort/cost_incremental_sort */
2289 :
2290 6196 : if (!pathkeys_count_contained_in(pathkeys, subpath->pathkeys,
2291 : &presorted_keys))
2292 : {
2293 : /*
2294 : * We'll need to insert a Sort node, so include costs for
2295 : * that. We choose to use incremental sort if it is
2296 : * enabled and there are presorted keys; otherwise we use
2297 : * full sort.
2298 : *
2299 : * We can use the parent's LIMIT if any, since we
2300 : * certainly won't pull more than that many tuples from
2301 : * any child.
2302 : */
2303 44 : if (enable_incremental_sort && presorted_keys > 0)
2304 : {
2305 12 : cost_incremental_sort(&sort_path,
2306 : root,
2307 : pathkeys,
2308 : presorted_keys,
2309 : subpath->disabled_nodes,
2310 : subpath->startup_cost,
2311 : subpath->total_cost,
2312 : subpath->rows,
2313 12 : subpath->pathtarget->width,
2314 : 0.0,
2315 : work_mem,
2316 : apath->limit_tuples);
2317 : }
2318 : else
2319 : {
2320 32 : cost_sort(&sort_path,
2321 : root,
2322 : pathkeys,
2323 : subpath->disabled_nodes,
2324 : subpath->total_cost,
2325 : subpath->rows,
2326 32 : subpath->pathtarget->width,
2327 : 0.0,
2328 : work_mem,
2329 : apath->limit_tuples);
2330 : }
2331 :
2332 44 : subpath = &sort_path;
2333 : }
2334 :
2335 6196 : apath->path.rows += subpath->rows;
2336 6196 : apath->path.disabled_nodes += subpath->disabled_nodes;
2337 6196 : apath->path.startup_cost += subpath->startup_cost;
2338 6196 : apath->path.total_cost += subpath->total_cost;
2339 : }
2340 : }
2341 : }
2342 : else /* parallel-aware */
2343 : {
2344 25564 : int i = 0;
2345 25564 : double parallel_divisor = get_parallel_divisor(&apath->path);
2346 :
2347 : /* Parallel-aware Append never produces ordered output. */
2348 : Assert(apath->path.pathkeys == NIL);
2349 :
2350 : /* Calculate startup cost. */
2351 101118 : foreach(l, apath->subpaths)
2352 : {
2353 75554 : Path *subpath = (Path *) lfirst(l);
2354 :
2355 : /*
2356 : * Append will start returning tuples when the child node having
2357 : * lowest startup cost is done setting up. We consider only the
2358 : * first few subplans that immediately get a worker assigned.
2359 : */
2360 75554 : if (i == 0)
2361 25564 : apath->path.startup_cost = subpath->startup_cost;
2362 49990 : else if (i < apath->path.parallel_workers)
2363 25006 : apath->path.startup_cost = Min(apath->path.startup_cost,
2364 : subpath->startup_cost);
2365 :
2366 : /*
2367 : * Apply parallel divisor to subpaths. Scale the number of rows
2368 : * for each partial subpath based on the ratio of the parallel
2369 : * divisor originally used for the subpath to the one we adopted.
2370 : * Also add the cost of partial paths to the total cost, but
2371 : * ignore non-partial paths for now.
2372 : */
2373 75554 : if (i < apath->first_partial_path)
2374 14166 : apath->path.rows += subpath->rows / parallel_divisor;
2375 : else
2376 : {
2377 : double subpath_parallel_divisor;
2378 :
2379 61388 : subpath_parallel_divisor = get_parallel_divisor(subpath);
2380 61388 : apath->path.rows += subpath->rows * (subpath_parallel_divisor /
2381 : parallel_divisor);
2382 61388 : apath->path.total_cost += subpath->total_cost;
2383 : }
2384 :
2385 75554 : apath->path.disabled_nodes += subpath->disabled_nodes;
2386 75554 : apath->path.rows = clamp_row_est(apath->path.rows);
2387 :
2388 75554 : i++;
2389 : }
2390 :
2391 : /* Add cost for non-partial subpaths. */
2392 25564 : apath->path.total_cost +=
2393 25564 : append_nonpartial_cost(apath->subpaths,
2394 : apath->first_partial_path,
2395 : apath->path.parallel_workers);
2396 : }
2397 :
2398 : /*
2399 : * Although Append does not do any selection or projection, it's not free;
2400 : * add a small per-tuple overhead.
2401 : */
2402 67570 : apath->path.total_cost +=
2403 67570 : cpu_tuple_cost * APPEND_CPU_COST_MULTIPLIER * apath->path.rows;
2404 : }
2405 :
2406 : /*
2407 : * cost_merge_append
2408 : * Determines and returns the cost of a MergeAppend node.
2409 : *
2410 : * MergeAppend merges several pre-sorted input streams, using a heap that
2411 : * at any given instant holds the next tuple from each stream. If there
2412 : * are N streams, we need about N*log2(N) tuple comparisons to construct
2413 : * the heap at startup, and then for each output tuple, about log2(N)
2414 : * comparisons to replace the top entry.
2415 : *
2416 : * (The effective value of N will drop once some of the input streams are
2417 : * exhausted, but it seems unlikely to be worth trying to account for that.)
2418 : *
2419 : * The heap is never spilled to disk, since we assume N is not very large.
2420 : * So this is much simpler than cost_sort.
2421 : *
2422 : * As in cost_sort, we charge two operator evals per tuple comparison.
2423 : *
2424 : * 'pathkeys' is a list of sort keys
2425 : * 'n_streams' is the number of input streams
2426 : * 'input_disabled_nodes' is the sum of the input streams' disabled node counts
2427 : * 'input_startup_cost' is the sum of the input streams' startup costs
2428 : * 'input_total_cost' is the sum of the input streams' total costs
2429 : * 'tuples' is the number of tuples in all the streams
2430 : */
2431 : void
2432 9908 : cost_merge_append(Path *path, PlannerInfo *root,
2433 : List *pathkeys, int n_streams,
2434 : int input_disabled_nodes,
2435 : Cost input_startup_cost, Cost input_total_cost,
2436 : double tuples)
2437 : {
2438 9908 : Cost startup_cost = 0;
2439 9908 : Cost run_cost = 0;
2440 : Cost comparison_cost;
2441 : double N;
2442 : double logN;
2443 :
2444 : /*
2445 : * Avoid log(0)...
2446 : */
2447 9908 : N = (n_streams < 2) ? 2.0 : (double) n_streams;
2448 9908 : logN = LOG2(N);
2449 :
2450 : /* Assumed cost per tuple comparison */
2451 9908 : comparison_cost = 2.0 * cpu_operator_cost;
2452 :
2453 : /* Heap creation cost */
2454 9908 : startup_cost += comparison_cost * N * logN;
2455 :
2456 : /* Per-tuple heap maintenance cost */
2457 9908 : run_cost += tuples * comparison_cost * logN;
2458 :
2459 : /*
2460 : * Although MergeAppend does not do any selection or projection, it's not
2461 : * free; add a small per-tuple overhead.
2462 : */
2463 9908 : run_cost += cpu_tuple_cost * APPEND_CPU_COST_MULTIPLIER * tuples;
2464 :
2465 9908 : path->disabled_nodes = input_disabled_nodes;
2466 9908 : path->startup_cost = startup_cost + input_startup_cost;
2467 9908 : path->total_cost = startup_cost + run_cost + input_total_cost;
2468 9908 : }
2469 :
2470 : /*
2471 : * cost_material
2472 : * Determines and returns the cost of materializing a relation, including
2473 : * the cost of reading the input data.
2474 : *
2475 : * If the total volume of data to materialize exceeds work_mem, we will need
2476 : * to write it to disk, so the cost is much higher in that case.
2477 : *
2478 : * Note that here we are estimating the costs for the first scan of the
2479 : * relation, so the materialization is all overhead --- any savings will
2480 : * occur only on rescan, which is estimated in cost_rescan.
2481 : */
2482 : void
2483 673568 : cost_material(Path *path,
2484 : int input_disabled_nodes,
2485 : Cost input_startup_cost, Cost input_total_cost,
2486 : double tuples, int width)
2487 : {
2488 673568 : Cost startup_cost = input_startup_cost;
2489 673568 : Cost run_cost = input_total_cost - input_startup_cost;
2490 673568 : double nbytes = relation_byte_size(tuples, width);
2491 673568 : double work_mem_bytes = work_mem * (Size) 1024;
2492 :
2493 673568 : path->rows = tuples;
2494 :
2495 : /*
2496 : * Whether spilling or not, charge 2x cpu_operator_cost per tuple to
2497 : * reflect bookkeeping overhead. (This rate must be more than what
2498 : * cost_rescan charges for materialize, ie, cpu_operator_cost per tuple;
2499 : * if it is exactly the same then there will be a cost tie between
2500 : * nestloop with A outer, materialized B inner and nestloop with B outer,
2501 : * materialized A inner. The extra cost ensures we'll prefer
2502 : * materializing the smaller rel.) Note that this is normally a good deal
2503 : * less than cpu_tuple_cost; which is OK because a Material plan node
2504 : * doesn't do qual-checking or projection, so it's got less overhead than
2505 : * most plan nodes.
2506 : */
2507 673568 : run_cost += 2 * cpu_operator_cost * tuples;
2508 :
2509 : /*
2510 : * If we will spill to disk, charge at the rate of seq_page_cost per page.
2511 : * This cost is assumed to be evenly spread through the plan run phase,
2512 : * which isn't exactly accurate but our cost model doesn't allow for
2513 : * nonuniform costs within the run phase.
2514 : */
2515 673568 : if (nbytes > work_mem_bytes)
2516 : {
2517 4992 : double npages = ceil(nbytes / BLCKSZ);
2518 :
2519 4992 : run_cost += seq_page_cost * npages;
2520 : }
2521 :
2522 673568 : path->disabled_nodes = input_disabled_nodes + (enable_material ? 0 : 1);
2523 673568 : path->startup_cost = startup_cost;
2524 673568 : path->total_cost = startup_cost + run_cost;
2525 673568 : }
2526 :
2527 : /*
2528 : * cost_memoize_rescan
2529 : * Determines the estimated cost of rescanning a Memoize node.
2530 : *
2531 : * In order to estimate this, we must gain knowledge of how often we expect to
2532 : * be called and how many distinct sets of parameters we are likely to be
2533 : * called with. If we expect a good cache hit ratio, then we can set our
2534 : * costs to account for that hit ratio, plus a little bit of cost for the
2535 : * caching itself. Caching will not work out well if we expect to be called
2536 : * with too many distinct parameter values. The worst-case here is that we
2537 : * never see any parameter value twice, in which case we'd never get a cache
2538 : * hit and caching would be a complete waste of effort.
2539 : */
2540 : static void
2541 291440 : cost_memoize_rescan(PlannerInfo *root, MemoizePath *mpath,
2542 : Cost *rescan_startup_cost, Cost *rescan_total_cost)
2543 : {
2544 : EstimationInfo estinfo;
2545 : ListCell *lc;
2546 291440 : Cost input_startup_cost = mpath->subpath->startup_cost;
2547 291440 : Cost input_total_cost = mpath->subpath->total_cost;
2548 291440 : double tuples = mpath->subpath->rows;
2549 291440 : Cardinality est_calls = mpath->est_calls;
2550 291440 : int width = mpath->subpath->pathtarget->width;
2551 :
2552 : double hash_mem_bytes;
2553 : double est_entry_bytes;
2554 : Cardinality est_cache_entries;
2555 : Cardinality ndistinct;
2556 : double evict_ratio;
2557 : double hit_ratio;
2558 : Cost startup_cost;
2559 : Cost total_cost;
2560 :
2561 : /* available cache space */
2562 291440 : hash_mem_bytes = get_hash_memory_limit();
2563 :
2564 : /*
2565 : * Set the number of bytes each cache entry should consume in the cache.
2566 : * To provide us with better estimations on how many cache entries we can
2567 : * store at once, we make a call to the executor here to ask it what
2568 : * memory overheads there are for a single cache entry.
2569 : */
2570 291440 : est_entry_bytes = relation_byte_size(tuples, width) +
2571 291440 : ExecEstimateCacheEntryOverheadBytes(tuples);
2572 :
2573 : /* include the estimated width for the cache keys */
2574 621238 : foreach(lc, mpath->param_exprs)
2575 329798 : est_entry_bytes += get_expr_width(root, (Node *) lfirst(lc));
2576 :
2577 : /* estimate on the upper limit of cache entries we can hold at once */
2578 291440 : est_cache_entries = floor(hash_mem_bytes / est_entry_bytes);
2579 :
2580 : /* estimate on the distinct number of parameter values */
2581 291440 : ndistinct = estimate_num_groups(root, mpath->param_exprs, est_calls, NULL,
2582 : &estinfo);
2583 :
2584 : /*
2585 : * When the estimation fell back on using a default value, it's a bit too
2586 : * risky to assume that it's ok to use a Memoize node. The use of a
2587 : * default could cause us to use a Memoize node when it's really
2588 : * inappropriate to do so. If we see that this has been done, then we'll
2589 : * assume that every call will have unique parameters, which will almost
2590 : * certainly mean a MemoizePath will never survive add_path().
2591 : */
2592 291440 : if ((estinfo.flags & SELFLAG_USED_DEFAULT) != 0)
2593 17172 : ndistinct = est_calls;
2594 :
2595 : /* Remember the ndistinct estimate for EXPLAIN */
2596 291440 : mpath->est_unique_keys = ndistinct;
2597 :
2598 : /*
2599 : * Since we've already estimated the maximum number of entries we can
2600 : * store at once and know the estimated number of distinct values we'll be
2601 : * called with, we'll take this opportunity to set the path's est_entries.
2602 : * This will ultimately determine the hash table size that the executor
2603 : * will use. If we leave this at zero, the executor will just choose the
2604 : * size itself. Really this is not the right place to do this, but it's
2605 : * convenient since everything is already calculated.
2606 : */
2607 291440 : mpath->est_entries = Min(Min(ndistinct, est_cache_entries),
2608 : PG_UINT32_MAX);
2609 :
2610 : /*
2611 : * When the number of distinct parameter values is above the amount we can
2612 : * store in the cache, then we'll have to evict some entries from the
2613 : * cache. This is not free. Here we estimate how often we'll incur the
2614 : * cost of that eviction.
2615 : */
2616 291440 : evict_ratio = 1.0 - Min(est_cache_entries, ndistinct) / ndistinct;
2617 :
2618 : /*
2619 : * In order to estimate how costly a single scan will be, we need to
2620 : * attempt to estimate what the cache hit ratio will be. To do that we
2621 : * must look at how many scans are estimated in total for this node and
2622 : * how many of those scans we expect to get a cache hit.
2623 : */
2624 582880 : hit_ratio = ((est_calls - ndistinct) / est_calls) *
2625 291440 : (est_cache_entries / Max(ndistinct, est_cache_entries));
2626 :
2627 : /* Remember the hit ratio estimate for EXPLAIN */
2628 291440 : mpath->est_hit_ratio = hit_ratio;
2629 :
2630 : Assert(hit_ratio >= 0 && hit_ratio <= 1.0);
2631 :
2632 : /*
2633 : * Set the total_cost accounting for the expected cache hit ratio. We
2634 : * also add on a cpu_operator_cost to account for a cache lookup. This
2635 : * will happen regardless of whether it's a cache hit or not.
2636 : */
2637 291440 : total_cost = input_total_cost * (1.0 - hit_ratio) + cpu_operator_cost;
2638 :
2639 : /* Now adjust the total cost to account for cache evictions */
2640 :
2641 : /* Charge a cpu_tuple_cost for evicting the actual cache entry */
2642 291440 : total_cost += cpu_tuple_cost * evict_ratio;
2643 :
2644 : /*
2645 : * Charge a 10th of cpu_operator_cost to evict every tuple in that entry.
2646 : * The per-tuple eviction is really just a pfree, so charging a whole
2647 : * cpu_operator_cost seems a little excessive.
2648 : */
2649 291440 : total_cost += cpu_operator_cost / 10.0 * evict_ratio * tuples;
2650 :
2651 : /*
2652 : * Now adjust for storing things in the cache, since that's not free
2653 : * either. Everything must go in the cache. We don't proportion this
2654 : * over any ratio, just apply it once for the scan. We charge a
2655 : * cpu_tuple_cost for the creation of the cache entry and also a
2656 : * cpu_operator_cost for each tuple we expect to cache.
2657 : */
2658 291440 : total_cost += cpu_tuple_cost + cpu_operator_cost * tuples;
2659 :
2660 : /*
2661 : * Getting the first row must be also be proportioned according to the
2662 : * expected cache hit ratio.
2663 : */
2664 291440 : startup_cost = input_startup_cost * (1.0 - hit_ratio);
2665 :
2666 : /*
2667 : * Additionally we charge a cpu_tuple_cost to account for cache lookups,
2668 : * which we'll do regardless of whether it was a cache hit or not.
2669 : */
2670 291440 : startup_cost += cpu_tuple_cost;
2671 :
2672 291440 : *rescan_startup_cost = startup_cost;
2673 291440 : *rescan_total_cost = total_cost;
2674 291440 : }
2675 :
2676 : /*
2677 : * cost_agg
2678 : * Determines and returns the cost of performing an Agg plan node,
2679 : * including the cost of its input.
2680 : *
2681 : * aggcosts can be NULL when there are no actual aggregate functions (i.e.,
2682 : * we are using a hashed Agg node just to do grouping).
2683 : *
2684 : * Note: when aggstrategy == AGG_SORTED, caller must ensure that input costs
2685 : * are for appropriately-sorted input.
2686 : */
2687 : void
2688 86294 : cost_agg(Path *path, PlannerInfo *root,
2689 : AggStrategy aggstrategy, const AggClauseCosts *aggcosts,
2690 : int numGroupCols, double numGroups,
2691 : List *quals,
2692 : int disabled_nodes,
2693 : Cost input_startup_cost, Cost input_total_cost,
2694 : double input_tuples, double input_width)
2695 : {
2696 : double output_tuples;
2697 : Cost startup_cost;
2698 : Cost total_cost;
2699 86294 : const AggClauseCosts dummy_aggcosts = {0};
2700 :
2701 : /* Use all-zero per-aggregate costs if NULL is passed */
2702 86294 : if (aggcosts == NULL)
2703 : {
2704 : Assert(aggstrategy == AGG_HASHED);
2705 18784 : aggcosts = &dummy_aggcosts;
2706 : }
2707 :
2708 : /*
2709 : * The transCost.per_tuple component of aggcosts should be charged once
2710 : * per input tuple, corresponding to the costs of evaluating the aggregate
2711 : * transfns and their input expressions. The finalCost.per_tuple component
2712 : * is charged once per output tuple, corresponding to the costs of
2713 : * evaluating the finalfns. Startup costs are of course charged but once.
2714 : *
2715 : * If we are grouping, we charge an additional cpu_operator_cost per
2716 : * grouping column per input tuple for grouping comparisons.
2717 : *
2718 : * We will produce a single output tuple if not grouping, and a tuple per
2719 : * group otherwise. We charge cpu_tuple_cost for each output tuple.
2720 : *
2721 : * Note: in this cost model, AGG_SORTED and AGG_HASHED have exactly the
2722 : * same total CPU cost, but AGG_SORTED has lower startup cost. If the
2723 : * input path is already sorted appropriately, AGG_SORTED should be
2724 : * preferred (since it has no risk of memory overflow). This will happen
2725 : * as long as the computed total costs are indeed exactly equal --- but if
2726 : * there's roundoff error we might do the wrong thing. So be sure that
2727 : * the computations below form the same intermediate values in the same
2728 : * order.
2729 : */
2730 86294 : if (aggstrategy == AGG_PLAIN)
2731 : {
2732 37226 : startup_cost = input_total_cost;
2733 37226 : startup_cost += aggcosts->transCost.startup;
2734 37226 : startup_cost += aggcosts->transCost.per_tuple * input_tuples;
2735 37226 : startup_cost += aggcosts->finalCost.startup;
2736 37226 : startup_cost += aggcosts->finalCost.per_tuple;
2737 : /* we aren't grouping */
2738 37226 : total_cost = startup_cost + cpu_tuple_cost;
2739 37226 : output_tuples = 1;
2740 : }
2741 49068 : else if (aggstrategy == AGG_SORTED || aggstrategy == AGG_MIXED)
2742 : {
2743 : /* Here we are able to deliver output on-the-fly */
2744 17758 : startup_cost = input_startup_cost;
2745 17758 : total_cost = input_total_cost;
2746 17758 : if (aggstrategy == AGG_MIXED && !enable_hashagg)
2747 480 : ++disabled_nodes;
2748 : /* calcs phrased this way to match HASHED case, see note above */
2749 17758 : total_cost += aggcosts->transCost.startup;
2750 17758 : total_cost += aggcosts->transCost.per_tuple * input_tuples;
2751 17758 : total_cost += (cpu_operator_cost * numGroupCols) * input_tuples;
2752 17758 : total_cost += aggcosts->finalCost.startup;
2753 17758 : total_cost += aggcosts->finalCost.per_tuple * numGroups;
2754 17758 : total_cost += cpu_tuple_cost * numGroups;
2755 17758 : output_tuples = numGroups;
2756 : }
2757 : else
2758 : {
2759 : /* must be AGG_HASHED */
2760 31310 : startup_cost = input_total_cost;
2761 31310 : if (!enable_hashagg)
2762 1866 : ++disabled_nodes;
2763 31310 : startup_cost += aggcosts->transCost.startup;
2764 31310 : startup_cost += aggcosts->transCost.per_tuple * input_tuples;
2765 : /* cost of computing hash value */
2766 31310 : startup_cost += (cpu_operator_cost * numGroupCols) * input_tuples;
2767 31310 : startup_cost += aggcosts->finalCost.startup;
2768 :
2769 31310 : total_cost = startup_cost;
2770 31310 : total_cost += aggcosts->finalCost.per_tuple * numGroups;
2771 : /* cost of retrieving from hash table */
2772 31310 : total_cost += cpu_tuple_cost * numGroups;
2773 31310 : output_tuples = numGroups;
2774 : }
2775 :
2776 : /*
2777 : * Add the disk costs of hash aggregation that spills to disk.
2778 : *
2779 : * Groups that go into the hash table stay in memory until finalized, so
2780 : * spilling and reprocessing tuples doesn't incur additional invocations
2781 : * of transCost or finalCost. Furthermore, the computed hash value is
2782 : * stored with the spilled tuples, so we don't incur extra invocations of
2783 : * the hash function.
2784 : *
2785 : * Hash Agg begins returning tuples after the first batch is complete.
2786 : * Accrue writes (spilled tuples) to startup_cost and to total_cost;
2787 : * accrue reads only to total_cost.
2788 : */
2789 86294 : if (aggstrategy == AGG_HASHED || aggstrategy == AGG_MIXED)
2790 : {
2791 : double pages;
2792 32250 : double pages_written = 0.0;
2793 32250 : double pages_read = 0.0;
2794 : double spill_cost;
2795 : double hashentrysize;
2796 : double nbatches;
2797 : Size mem_limit;
2798 : uint64 ngroups_limit;
2799 : int num_partitions;
2800 : int depth;
2801 :
2802 : /*
2803 : * Estimate number of batches based on the computed limits. If less
2804 : * than or equal to one, all groups are expected to fit in memory;
2805 : * otherwise we expect to spill.
2806 : */
2807 32250 : hashentrysize = hash_agg_entry_size(list_length(root->aggtransinfos),
2808 : input_width,
2809 32250 : aggcosts->transitionSpace);
2810 32250 : hash_agg_set_limits(hashentrysize, numGroups, 0, &mem_limit,
2811 : &ngroups_limit, &num_partitions);
2812 :
2813 32250 : nbatches = Max((numGroups * hashentrysize) / mem_limit,
2814 : numGroups / ngroups_limit);
2815 :
2816 32250 : nbatches = Max(ceil(nbatches), 1.0);
2817 32250 : num_partitions = Max(num_partitions, 2);
2818 :
2819 : /*
2820 : * The number of partitions can change at different levels of
2821 : * recursion; but for the purposes of this calculation assume it stays
2822 : * constant.
2823 : */
2824 32250 : depth = ceil(log(nbatches) / log(num_partitions));
2825 :
2826 : /*
2827 : * Estimate number of pages read and written. For each level of
2828 : * recursion, a tuple must be written and then later read.
2829 : */
2830 32250 : pages = relation_byte_size(input_tuples, input_width) / BLCKSZ;
2831 32250 : pages_written = pages_read = pages * depth;
2832 :
2833 : /*
2834 : * HashAgg has somewhat worse IO behavior than Sort on typical
2835 : * hardware/OS combinations. Account for this with a generic penalty.
2836 : */
2837 32250 : pages_read *= 2.0;
2838 32250 : pages_written *= 2.0;
2839 :
2840 32250 : startup_cost += pages_written * random_page_cost;
2841 32250 : total_cost += pages_written * random_page_cost;
2842 32250 : total_cost += pages_read * seq_page_cost;
2843 :
2844 : /* account for CPU cost of spilling a tuple and reading it back */
2845 32250 : spill_cost = depth * input_tuples * 2.0 * cpu_tuple_cost;
2846 32250 : startup_cost += spill_cost;
2847 32250 : total_cost += spill_cost;
2848 : }
2849 :
2850 : /*
2851 : * If there are quals (HAVING quals), account for their cost and
2852 : * selectivity.
2853 : */
2854 86294 : if (quals)
2855 : {
2856 : QualCost qual_cost;
2857 :
2858 4620 : cost_qual_eval(&qual_cost, quals, root);
2859 4620 : startup_cost += qual_cost.startup;
2860 4620 : total_cost += qual_cost.startup + output_tuples * qual_cost.per_tuple;
2861 :
2862 4620 : output_tuples = clamp_row_est(output_tuples *
2863 4620 : clauselist_selectivity(root,
2864 : quals,
2865 : 0,
2866 : JOIN_INNER,
2867 : NULL));
2868 : }
2869 :
2870 86294 : path->rows = output_tuples;
2871 86294 : path->disabled_nodes = disabled_nodes;
2872 86294 : path->startup_cost = startup_cost;
2873 86294 : path->total_cost = total_cost;
2874 86294 : }
2875 :
2876 : /*
2877 : * get_windowclause_startup_tuples
2878 : * Estimate how many tuples we'll need to fetch from a WindowAgg's
2879 : * subnode before we can output the first WindowAgg tuple.
2880 : *
2881 : * How many tuples need to be read depends on the WindowClause. For example,
2882 : * a WindowClause with no PARTITION BY and no ORDER BY requires that all
2883 : * subnode tuples are read and aggregated before the WindowAgg can output
2884 : * anything. If there's a PARTITION BY, then we only need to look at tuples
2885 : * in the first partition. Here we attempt to estimate just how many
2886 : * 'input_tuples' the WindowAgg will need to read for the given WindowClause
2887 : * before the first tuple can be output.
2888 : */
2889 : static double
2890 2964 : get_windowclause_startup_tuples(PlannerInfo *root, WindowClause *wc,
2891 : double input_tuples)
2892 : {
2893 2964 : int frameOptions = wc->frameOptions;
2894 : double partition_tuples;
2895 : double return_tuples;
2896 : double peer_tuples;
2897 :
2898 : /*
2899 : * First, figure out how many partitions there are likely to be and set
2900 : * partition_tuples according to that estimate.
2901 : */
2902 2964 : if (wc->partitionClause != NIL)
2903 : {
2904 : double num_partitions;
2905 734 : List *partexprs = get_sortgrouplist_exprs(wc->partitionClause,
2906 734 : root->parse->targetList);
2907 :
2908 734 : num_partitions = estimate_num_groups(root, partexprs, input_tuples,
2909 : NULL, NULL);
2910 734 : list_free(partexprs);
2911 :
2912 734 : partition_tuples = input_tuples / num_partitions;
2913 : }
2914 : else
2915 : {
2916 : /* all tuples belong to the same partition */
2917 2230 : partition_tuples = input_tuples;
2918 : }
2919 :
2920 : /* estimate the number of tuples in each peer group */
2921 2964 : if (wc->orderClause != NIL)
2922 : {
2923 : double num_groups;
2924 : List *orderexprs;
2925 :
2926 2358 : orderexprs = get_sortgrouplist_exprs(wc->orderClause,
2927 2358 : root->parse->targetList);
2928 :
2929 : /* estimate out how many peer groups there are in the partition */
2930 2358 : num_groups = estimate_num_groups(root, orderexprs,
2931 : partition_tuples, NULL,
2932 : NULL);
2933 2358 : list_free(orderexprs);
2934 2358 : peer_tuples = partition_tuples / num_groups;
2935 : }
2936 : else
2937 : {
2938 : /* no ORDER BY so only 1 tuple belongs in each peer group */
2939 606 : peer_tuples = 1.0;
2940 : }
2941 :
2942 2964 : if (frameOptions & FRAMEOPTION_END_UNBOUNDED_FOLLOWING)
2943 : {
2944 : /* include all partition rows */
2945 364 : return_tuples = partition_tuples;
2946 : }
2947 2600 : else if (frameOptions & FRAMEOPTION_END_CURRENT_ROW)
2948 : {
2949 1562 : if (frameOptions & FRAMEOPTION_ROWS)
2950 : {
2951 : /* just count the current row */
2952 722 : return_tuples = 1.0;
2953 : }
2954 840 : else if (frameOptions & (FRAMEOPTION_RANGE | FRAMEOPTION_GROUPS))
2955 : {
2956 : /*
2957 : * When in RANGE/GROUPS mode, it's more complex. If there's no
2958 : * ORDER BY, then all rows in the partition are peers, otherwise
2959 : * we'll need to read the first group of peers.
2960 : */
2961 840 : if (wc->orderClause == NIL)
2962 326 : return_tuples = partition_tuples;
2963 : else
2964 514 : return_tuples = peer_tuples;
2965 : }
2966 : else
2967 : {
2968 : /*
2969 : * Something new we don't support yet? This needs attention.
2970 : * We'll just return 1.0 in the meantime.
2971 : */
2972 : Assert(false);
2973 0 : return_tuples = 1.0;
2974 : }
2975 : }
2976 1038 : else if (frameOptions & FRAMEOPTION_END_OFFSET_PRECEDING)
2977 : {
2978 : /*
2979 : * BETWEEN ... AND N PRECEDING will only need to read the WindowAgg's
2980 : * subnode after N ROWS/RANGES/GROUPS. N can be 0, but not negative,
2981 : * so we'll just assume only the current row needs to be read to fetch
2982 : * the first WindowAgg row.
2983 : */
2984 108 : return_tuples = 1.0;
2985 : }
2986 930 : else if (frameOptions & FRAMEOPTION_END_OFFSET_FOLLOWING)
2987 : {
2988 930 : Const *endOffset = (Const *) wc->endOffset;
2989 : double end_offset_value;
2990 :
2991 : /* try and figure out the value specified in the endOffset. */
2992 930 : if (IsA(endOffset, Const))
2993 : {
2994 930 : if (endOffset->constisnull)
2995 : {
2996 : /*
2997 : * NULLs are not allowed, but currently, there's no code to
2998 : * error out if there's a NULL Const. We'll only discover
2999 : * this during execution. For now, just pretend everything is
3000 : * fine and assume that just the first row/range/group will be
3001 : * needed.
3002 : */
3003 0 : end_offset_value = 1.0;
3004 : }
3005 : else
3006 : {
3007 930 : switch (endOffset->consttype)
3008 : {
3009 24 : case INT2OID:
3010 24 : end_offset_value =
3011 24 : (double) DatumGetInt16(endOffset->constvalue);
3012 24 : break;
3013 132 : case INT4OID:
3014 132 : end_offset_value =
3015 132 : (double) DatumGetInt32(endOffset->constvalue);
3016 132 : break;
3017 432 : case INT8OID:
3018 432 : end_offset_value =
3019 432 : (double) DatumGetInt64(endOffset->constvalue);
3020 432 : break;
3021 342 : default:
3022 342 : end_offset_value =
3023 342 : partition_tuples / peer_tuples *
3024 : DEFAULT_INEQ_SEL;
3025 342 : break;
3026 : }
3027 : }
3028 : }
3029 : else
3030 : {
3031 : /*
3032 : * When the end bound is not a Const, we'll just need to guess. We
3033 : * just make use of DEFAULT_INEQ_SEL.
3034 : */
3035 0 : end_offset_value =
3036 0 : partition_tuples / peer_tuples * DEFAULT_INEQ_SEL;
3037 : }
3038 :
3039 930 : if (frameOptions & FRAMEOPTION_ROWS)
3040 : {
3041 : /* include the N FOLLOWING and the current row */
3042 270 : return_tuples = end_offset_value + 1.0;
3043 : }
3044 660 : else if (frameOptions & (FRAMEOPTION_RANGE | FRAMEOPTION_GROUPS))
3045 : {
3046 : /* include N FOLLOWING ranges/group and the initial range/group */
3047 660 : return_tuples = peer_tuples * (end_offset_value + 1.0);
3048 : }
3049 : else
3050 : {
3051 : /*
3052 : * Something new we don't support yet? This needs attention.
3053 : * We'll just return 1.0 in the meantime.
3054 : */
3055 : Assert(false);
3056 0 : return_tuples = 1.0;
3057 : }
3058 : }
3059 : else
3060 : {
3061 : /*
3062 : * Something new we don't support yet? This needs attention. We'll
3063 : * just return 1.0 in the meantime.
3064 : */
3065 : Assert(false);
3066 0 : return_tuples = 1.0;
3067 : }
3068 :
3069 2964 : if (wc->partitionClause != NIL || wc->orderClause != NIL)
3070 : {
3071 : /*
3072 : * Cap the return value to the estimated partition tuples and account
3073 : * for the extra tuple WindowAgg will need to read to confirm the next
3074 : * tuple does not belong to the same partition or peer group.
3075 : */
3076 2570 : return_tuples = Min(return_tuples + 1.0, partition_tuples);
3077 : }
3078 : else
3079 : {
3080 : /*
3081 : * Cap the return value so it's never higher than the expected tuples
3082 : * in the partition.
3083 : */
3084 394 : return_tuples = Min(return_tuples, partition_tuples);
3085 : }
3086 :
3087 : /*
3088 : * We needn't worry about any EXCLUDE options as those only exclude rows
3089 : * from being aggregated, not from being read from the WindowAgg's
3090 : * subnode.
3091 : */
3092 :
3093 2964 : return clamp_row_est(return_tuples);
3094 : }
3095 :
3096 : /*
3097 : * cost_windowagg
3098 : * Determines and returns the cost of performing a WindowAgg plan node,
3099 : * including the cost of its input.
3100 : *
3101 : * Input is assumed already properly sorted.
3102 : */
3103 : void
3104 2964 : cost_windowagg(Path *path, PlannerInfo *root,
3105 : List *windowFuncs, WindowClause *winclause,
3106 : int input_disabled_nodes,
3107 : Cost input_startup_cost, Cost input_total_cost,
3108 : double input_tuples)
3109 : {
3110 : Cost startup_cost;
3111 : Cost total_cost;
3112 : double startup_tuples;
3113 : int numPartCols;
3114 : int numOrderCols;
3115 : ListCell *lc;
3116 :
3117 2964 : numPartCols = list_length(winclause->partitionClause);
3118 2964 : numOrderCols = list_length(winclause->orderClause);
3119 :
3120 2964 : startup_cost = input_startup_cost;
3121 2964 : total_cost = input_total_cost;
3122 :
3123 : /*
3124 : * Window functions are assumed to cost their stated execution cost, plus
3125 : * the cost of evaluating their input expressions, per tuple. Since they
3126 : * may in fact evaluate their inputs at multiple rows during each cycle,
3127 : * this could be a drastic underestimate; but without a way to know how
3128 : * many rows the window function will fetch, it's hard to do better. In
3129 : * any case, it's a good estimate for all the built-in window functions,
3130 : * so we'll just do this for now.
3131 : */
3132 6798 : foreach(lc, windowFuncs)
3133 : {
3134 3834 : WindowFunc *wfunc = lfirst_node(WindowFunc, lc);
3135 : Cost wfunccost;
3136 : QualCost argcosts;
3137 :
3138 3834 : argcosts.startup = argcosts.per_tuple = 0;
3139 3834 : add_function_cost(root, wfunc->winfnoid, (Node *) wfunc,
3140 : &argcosts);
3141 3834 : startup_cost += argcosts.startup;
3142 3834 : wfunccost = argcosts.per_tuple;
3143 :
3144 : /* also add the input expressions' cost to per-input-row costs */
3145 3834 : cost_qual_eval_node(&argcosts, (Node *) wfunc->args, root);
3146 3834 : startup_cost += argcosts.startup;
3147 3834 : wfunccost += argcosts.per_tuple;
3148 :
3149 : /*
3150 : * Add the filter's cost to per-input-row costs. XXX We should reduce
3151 : * input expression costs according to filter selectivity.
3152 : */
3153 3834 : cost_qual_eval_node(&argcosts, (Node *) wfunc->aggfilter, root);
3154 3834 : startup_cost += argcosts.startup;
3155 3834 : wfunccost += argcosts.per_tuple;
3156 :
3157 3834 : total_cost += wfunccost * input_tuples;
3158 : }
3159 :
3160 : /*
3161 : * We also charge cpu_operator_cost per grouping column per tuple for
3162 : * grouping comparisons, plus cpu_tuple_cost per tuple for general
3163 : * overhead.
3164 : *
3165 : * XXX this neglects costs of spooling the data to disk when it overflows
3166 : * work_mem. Sooner or later that should get accounted for.
3167 : */
3168 2964 : total_cost += cpu_operator_cost * (numPartCols + numOrderCols) * input_tuples;
3169 2964 : total_cost += cpu_tuple_cost * input_tuples;
3170 :
3171 2964 : path->rows = input_tuples;
3172 2964 : path->disabled_nodes = input_disabled_nodes;
3173 2964 : path->startup_cost = startup_cost;
3174 2964 : path->total_cost = total_cost;
3175 :
3176 : /*
3177 : * Also, take into account how many tuples we need to read from the
3178 : * subnode in order to produce the first tuple from the WindowAgg. To do
3179 : * this we proportion the run cost (total cost not including startup cost)
3180 : * over the estimated startup tuples. We already included the startup
3181 : * cost of the subnode, so we only need to do this when the estimated
3182 : * startup tuples is above 1.0.
3183 : */
3184 2964 : startup_tuples = get_windowclause_startup_tuples(root, winclause,
3185 : input_tuples);
3186 :
3187 2964 : if (startup_tuples > 1.0)
3188 2556 : path->startup_cost += (total_cost - startup_cost) / input_tuples *
3189 2556 : (startup_tuples - 1.0);
3190 2964 : }
3191 :
3192 : /*
3193 : * cost_group
3194 : * Determines and returns the cost of performing a Group plan node,
3195 : * including the cost of its input.
3196 : *
3197 : * Note: caller must ensure that input costs are for appropriately-sorted
3198 : * input.
3199 : */
3200 : void
3201 1226 : cost_group(Path *path, PlannerInfo *root,
3202 : int numGroupCols, double numGroups,
3203 : List *quals,
3204 : int input_disabled_nodes,
3205 : Cost input_startup_cost, Cost input_total_cost,
3206 : double input_tuples)
3207 : {
3208 : double output_tuples;
3209 : Cost startup_cost;
3210 : Cost total_cost;
3211 :
3212 1226 : output_tuples = numGroups;
3213 1226 : startup_cost = input_startup_cost;
3214 1226 : total_cost = input_total_cost;
3215 :
3216 : /*
3217 : * Charge one cpu_operator_cost per comparison per input tuple. We assume
3218 : * all columns get compared at most of the tuples.
3219 : */
3220 1226 : total_cost += cpu_operator_cost * input_tuples * numGroupCols;
3221 :
3222 : /*
3223 : * If there are quals (HAVING quals), account for their cost and
3224 : * selectivity.
3225 : */
3226 1226 : if (quals)
3227 : {
3228 : QualCost qual_cost;
3229 :
3230 0 : cost_qual_eval(&qual_cost, quals, root);
3231 0 : startup_cost += qual_cost.startup;
3232 0 : total_cost += qual_cost.startup + output_tuples * qual_cost.per_tuple;
3233 :
3234 0 : output_tuples = clamp_row_est(output_tuples *
3235 0 : clauselist_selectivity(root,
3236 : quals,
3237 : 0,
3238 : JOIN_INNER,
3239 : NULL));
3240 : }
3241 :
3242 1226 : path->rows = output_tuples;
3243 1226 : path->disabled_nodes = input_disabled_nodes;
3244 1226 : path->startup_cost = startup_cost;
3245 1226 : path->total_cost = total_cost;
3246 1226 : }
3247 :
3248 : /*
3249 : * initial_cost_nestloop
3250 : * Preliminary estimate of the cost of a nestloop join path.
3251 : *
3252 : * This must quickly produce lower-bound estimates of the path's startup and
3253 : * total costs. If we are unable to eliminate the proposed path from
3254 : * consideration using the lower bounds, final_cost_nestloop will be called
3255 : * to obtain the final estimates.
3256 : *
3257 : * The exact division of labor between this function and final_cost_nestloop
3258 : * is private to them, and represents a tradeoff between speed of the initial
3259 : * estimate and getting a tight lower bound. We choose to not examine the
3260 : * join quals here, since that's by far the most expensive part of the
3261 : * calculations. The end result is that CPU-cost considerations must be
3262 : * left for the second phase; and for SEMI/ANTI joins, we must also postpone
3263 : * incorporation of the inner path's run cost.
3264 : *
3265 : * 'workspace' is to be filled with startup_cost, total_cost, and perhaps
3266 : * other data to be used by final_cost_nestloop
3267 : * 'jointype' is the type of join to be performed
3268 : * 'outer_path' is the outer input to the join
3269 : * 'inner_path' is the inner input to the join
3270 : * 'extra' contains miscellaneous information about the join
3271 : */
3272 : void
3273 3259348 : initial_cost_nestloop(PlannerInfo *root, JoinCostWorkspace *workspace,
3274 : JoinType jointype,
3275 : Path *outer_path, Path *inner_path,
3276 : JoinPathExtraData *extra)
3277 : {
3278 : int disabled_nodes;
3279 3259348 : Cost startup_cost = 0;
3280 3259348 : Cost run_cost = 0;
3281 3259348 : double outer_path_rows = outer_path->rows;
3282 : Cost inner_rescan_start_cost;
3283 : Cost inner_rescan_total_cost;
3284 : Cost inner_run_cost;
3285 : Cost inner_rescan_run_cost;
3286 :
3287 : /* Count up disabled nodes. */
3288 3259348 : disabled_nodes = enable_nestloop ? 0 : 1;
3289 3259348 : disabled_nodes += inner_path->disabled_nodes;
3290 3259348 : disabled_nodes += outer_path->disabled_nodes;
3291 :
3292 : /* estimate costs to rescan the inner relation */
3293 3259348 : cost_rescan(root, inner_path,
3294 : &inner_rescan_start_cost,
3295 : &inner_rescan_total_cost);
3296 :
3297 : /* cost of source data */
3298 :
3299 : /*
3300 : * NOTE: clearly, we must pay both outer and inner paths' startup_cost
3301 : * before we can start returning tuples, so the join's startup cost is
3302 : * their sum. We'll also pay the inner path's rescan startup cost
3303 : * multiple times.
3304 : */
3305 3259348 : startup_cost += outer_path->startup_cost + inner_path->startup_cost;
3306 3259348 : run_cost += outer_path->total_cost - outer_path->startup_cost;
3307 3259348 : if (outer_path_rows > 1)
3308 2383426 : run_cost += (outer_path_rows - 1) * inner_rescan_start_cost;
3309 :
3310 3259348 : inner_run_cost = inner_path->total_cost - inner_path->startup_cost;
3311 3259348 : inner_rescan_run_cost = inner_rescan_total_cost - inner_rescan_start_cost;
3312 :
3313 3259348 : if (jointype == JOIN_SEMI || jointype == JOIN_ANTI ||
3314 3198104 : extra->inner_unique)
3315 : {
3316 : /*
3317 : * With a SEMI or ANTI join, or if the innerrel is known unique, the
3318 : * executor will stop after the first match.
3319 : *
3320 : * Getting decent estimates requires inspection of the join quals,
3321 : * which we choose to postpone to final_cost_nestloop.
3322 : */
3323 :
3324 : /* Save private data for final_cost_nestloop */
3325 1323660 : workspace->inner_run_cost = inner_run_cost;
3326 1323660 : workspace->inner_rescan_run_cost = inner_rescan_run_cost;
3327 : }
3328 : else
3329 : {
3330 : /* Normal case; we'll scan whole input rel for each outer row */
3331 1935688 : run_cost += inner_run_cost;
3332 1935688 : if (outer_path_rows > 1)
3333 1499860 : run_cost += (outer_path_rows - 1) * inner_rescan_run_cost;
3334 : }
3335 :
3336 : /* CPU costs left for later */
3337 :
3338 : /* Public result fields */
3339 3259348 : workspace->disabled_nodes = disabled_nodes;
3340 3259348 : workspace->startup_cost = startup_cost;
3341 3259348 : workspace->total_cost = startup_cost + run_cost;
3342 : /* Save private data for final_cost_nestloop */
3343 3259348 : workspace->run_cost = run_cost;
3344 3259348 : }
3345 :
3346 : /*
3347 : * final_cost_nestloop
3348 : * Final estimate of the cost and result size of a nestloop join path.
3349 : *
3350 : * 'path' is already filled in except for the rows and cost fields
3351 : * 'workspace' is the result from initial_cost_nestloop
3352 : * 'extra' contains miscellaneous information about the join
3353 : */
3354 : void
3355 1460290 : final_cost_nestloop(PlannerInfo *root, NestPath *path,
3356 : JoinCostWorkspace *workspace,
3357 : JoinPathExtraData *extra)
3358 : {
3359 1460290 : Path *outer_path = path->jpath.outerjoinpath;
3360 1460290 : Path *inner_path = path->jpath.innerjoinpath;
3361 1460290 : double outer_path_rows = outer_path->rows;
3362 1460290 : double inner_path_rows = inner_path->rows;
3363 1460290 : Cost startup_cost = workspace->startup_cost;
3364 1460290 : Cost run_cost = workspace->run_cost;
3365 : Cost cpu_per_tuple;
3366 : QualCost restrict_qual_cost;
3367 : double ntuples;
3368 :
3369 : /* Set the number of disabled nodes. */
3370 1460290 : path->jpath.path.disabled_nodes = workspace->disabled_nodes;
3371 :
3372 : /* Protect some assumptions below that rowcounts aren't zero */
3373 1460290 : if (outer_path_rows <= 0)
3374 0 : outer_path_rows = 1;
3375 1460290 : if (inner_path_rows <= 0)
3376 726 : inner_path_rows = 1;
3377 : /* Mark the path with the correct row estimate */
3378 1460290 : if (path->jpath.path.param_info)
3379 32864 : path->jpath.path.rows = path->jpath.path.param_info->ppi_rows;
3380 : else
3381 1427426 : path->jpath.path.rows = path->jpath.path.parent->rows;
3382 :
3383 : /* For partial paths, scale row estimate. */
3384 1460290 : if (path->jpath.path.parallel_workers > 0)
3385 : {
3386 44038 : double parallel_divisor = get_parallel_divisor(&path->jpath.path);
3387 :
3388 44038 : path->jpath.path.rows =
3389 44038 : clamp_row_est(path->jpath.path.rows / parallel_divisor);
3390 : }
3391 :
3392 : /* cost of inner-relation source data (we already dealt with outer rel) */
3393 :
3394 1460290 : if (path->jpath.jointype == JOIN_SEMI || path->jpath.jointype == JOIN_ANTI ||
3395 1417746 : extra->inner_unique)
3396 913154 : {
3397 : /*
3398 : * With a SEMI or ANTI join, or if the innerrel is known unique, the
3399 : * executor will stop after the first match.
3400 : */
3401 913154 : Cost inner_run_cost = workspace->inner_run_cost;
3402 913154 : Cost inner_rescan_run_cost = workspace->inner_rescan_run_cost;
3403 : double outer_matched_rows;
3404 : double outer_unmatched_rows;
3405 : Selectivity inner_scan_frac;
3406 :
3407 : /*
3408 : * For an outer-rel row that has at least one match, we can expect the
3409 : * inner scan to stop after a fraction 1/(match_count+1) of the inner
3410 : * rows, if the matches are evenly distributed. Since they probably
3411 : * aren't quite evenly distributed, we apply a fuzz factor of 2.0 to
3412 : * that fraction. (If we used a larger fuzz factor, we'd have to
3413 : * clamp inner_scan_frac to at most 1.0; but since match_count is at
3414 : * least 1, no such clamp is needed now.)
3415 : */
3416 913154 : outer_matched_rows = rint(outer_path_rows * extra->semifactors.outer_match_frac);
3417 913154 : outer_unmatched_rows = outer_path_rows - outer_matched_rows;
3418 913154 : inner_scan_frac = 2.0 / (extra->semifactors.match_count + 1.0);
3419 :
3420 : /*
3421 : * Compute number of tuples processed (not number emitted!). First,
3422 : * account for successfully-matched outer rows.
3423 : */
3424 913154 : ntuples = outer_matched_rows * inner_path_rows * inner_scan_frac;
3425 :
3426 : /*
3427 : * Now we need to estimate the actual costs of scanning the inner
3428 : * relation, which may be quite a bit less than N times inner_run_cost
3429 : * due to early scan stops. We consider two cases. If the inner path
3430 : * is an indexscan using all the joinquals as indexquals, then an
3431 : * unmatched outer row results in an indexscan returning no rows,
3432 : * which is probably quite cheap. Otherwise, the executor will have
3433 : * to scan the whole inner rel for an unmatched row; not so cheap.
3434 : */
3435 913154 : if (has_indexed_join_quals(path))
3436 : {
3437 : /*
3438 : * Successfully-matched outer rows will only require scanning
3439 : * inner_scan_frac of the inner relation. In this case, we don't
3440 : * need to charge the full inner_run_cost even when that's more
3441 : * than inner_rescan_run_cost, because we can assume that none of
3442 : * the inner scans ever scan the whole inner relation. So it's
3443 : * okay to assume that all the inner scan executions can be
3444 : * fractions of the full cost, even if materialization is reducing
3445 : * the rescan cost. At this writing, it's impossible to get here
3446 : * for a materialized inner scan, so inner_run_cost and
3447 : * inner_rescan_run_cost will be the same anyway; but just in
3448 : * case, use inner_run_cost for the first matched tuple and
3449 : * inner_rescan_run_cost for additional ones.
3450 : */
3451 149722 : run_cost += inner_run_cost * inner_scan_frac;
3452 149722 : if (outer_matched_rows > 1)
3453 22780 : run_cost += (outer_matched_rows - 1) * inner_rescan_run_cost * inner_scan_frac;
3454 :
3455 : /*
3456 : * Add the cost of inner-scan executions for unmatched outer rows.
3457 : * We estimate this as the same cost as returning the first tuple
3458 : * of a nonempty scan. We consider that these are all rescans,
3459 : * since we used inner_run_cost once already.
3460 : */
3461 149722 : run_cost += outer_unmatched_rows *
3462 149722 : inner_rescan_run_cost / inner_path_rows;
3463 :
3464 : /*
3465 : * We won't be evaluating any quals at all for unmatched rows, so
3466 : * don't add them to ntuples.
3467 : */
3468 : }
3469 : else
3470 : {
3471 : /*
3472 : * Here, a complicating factor is that rescans may be cheaper than
3473 : * first scans. If we never scan all the way to the end of the
3474 : * inner rel, it might be (depending on the plan type) that we'd
3475 : * never pay the whole inner first-scan run cost. However it is
3476 : * difficult to estimate whether that will happen (and it could
3477 : * not happen if there are any unmatched outer rows!), so be
3478 : * conservative and always charge the whole first-scan cost once.
3479 : * We consider this charge to correspond to the first unmatched
3480 : * outer row, unless there isn't one in our estimate, in which
3481 : * case blame it on the first matched row.
3482 : */
3483 :
3484 : /* First, count all unmatched join tuples as being processed */
3485 763432 : ntuples += outer_unmatched_rows * inner_path_rows;
3486 :
3487 : /* Now add the forced full scan, and decrement appropriate count */
3488 763432 : run_cost += inner_run_cost;
3489 763432 : if (outer_unmatched_rows >= 1)
3490 726608 : outer_unmatched_rows -= 1;
3491 : else
3492 36824 : outer_matched_rows -= 1;
3493 :
3494 : /* Add inner run cost for additional outer tuples having matches */
3495 763432 : if (outer_matched_rows > 0)
3496 276552 : run_cost += outer_matched_rows * inner_rescan_run_cost * inner_scan_frac;
3497 :
3498 : /* Add inner run cost for additional unmatched outer tuples */
3499 763432 : if (outer_unmatched_rows > 0)
3500 483792 : run_cost += outer_unmatched_rows * inner_rescan_run_cost;
3501 : }
3502 : }
3503 : else
3504 : {
3505 : /* Normal-case source costs were included in preliminary estimate */
3506 :
3507 : /* Compute number of tuples processed (not number emitted!) */
3508 547136 : ntuples = outer_path_rows * inner_path_rows;
3509 : }
3510 :
3511 : /* CPU costs */
3512 1460290 : cost_qual_eval(&restrict_qual_cost, path->jpath.joinrestrictinfo, root);
3513 1460290 : startup_cost += restrict_qual_cost.startup;
3514 1460290 : cpu_per_tuple = cpu_tuple_cost + restrict_qual_cost.per_tuple;
3515 1460290 : run_cost += cpu_per_tuple * ntuples;
3516 :
3517 : /* tlist eval costs are paid per output row, not per tuple scanned */
3518 1460290 : startup_cost += path->jpath.path.pathtarget->cost.startup;
3519 1460290 : run_cost += path->jpath.path.pathtarget->cost.per_tuple * path->jpath.path.rows;
3520 :
3521 1460290 : path->jpath.path.startup_cost = startup_cost;
3522 1460290 : path->jpath.path.total_cost = startup_cost + run_cost;
3523 1460290 : }
3524 :
3525 : /*
3526 : * initial_cost_mergejoin
3527 : * Preliminary estimate of the cost of a mergejoin path.
3528 : *
3529 : * This must quickly produce lower-bound estimates of the path's startup and
3530 : * total costs. If we are unable to eliminate the proposed path from
3531 : * consideration using the lower bounds, final_cost_mergejoin will be called
3532 : * to obtain the final estimates.
3533 : *
3534 : * The exact division of labor between this function and final_cost_mergejoin
3535 : * is private to them, and represents a tradeoff between speed of the initial
3536 : * estimate and getting a tight lower bound. We choose to not examine the
3537 : * join quals here, except for obtaining the scan selectivity estimate which
3538 : * is really essential (but fortunately, use of caching keeps the cost of
3539 : * getting that down to something reasonable).
3540 : * We also assume that cost_sort/cost_incremental_sort is cheap enough to use
3541 : * here.
3542 : *
3543 : * 'workspace' is to be filled with startup_cost, total_cost, and perhaps
3544 : * other data to be used by final_cost_mergejoin
3545 : * 'jointype' is the type of join to be performed
3546 : * 'mergeclauses' is the list of joinclauses to be used as merge clauses
3547 : * 'outer_path' is the outer input to the join
3548 : * 'inner_path' is the inner input to the join
3549 : * 'outersortkeys' is the list of sort keys for the outer path
3550 : * 'innersortkeys' is the list of sort keys for the inner path
3551 : * 'outer_presorted_keys' is the number of presorted keys of the outer path
3552 : * 'extra' contains miscellaneous information about the join
3553 : *
3554 : * Note: outersortkeys and innersortkeys should be NIL if no explicit
3555 : * sort is needed because the respective source path is already ordered.
3556 : */
3557 : void
3558 1476348 : initial_cost_mergejoin(PlannerInfo *root, JoinCostWorkspace *workspace,
3559 : JoinType jointype,
3560 : List *mergeclauses,
3561 : Path *outer_path, Path *inner_path,
3562 : List *outersortkeys, List *innersortkeys,
3563 : int outer_presorted_keys,
3564 : JoinPathExtraData *extra)
3565 : {
3566 : int disabled_nodes;
3567 1476348 : Cost startup_cost = 0;
3568 1476348 : Cost run_cost = 0;
3569 1476348 : double outer_path_rows = outer_path->rows;
3570 1476348 : double inner_path_rows = inner_path->rows;
3571 : Cost inner_run_cost;
3572 : double outer_rows,
3573 : inner_rows,
3574 : outer_skip_rows,
3575 : inner_skip_rows;
3576 : Selectivity outerstartsel,
3577 : outerendsel,
3578 : innerstartsel,
3579 : innerendsel;
3580 : Path sort_path; /* dummy for result of
3581 : * cost_sort/cost_incremental_sort */
3582 :
3583 : /* Protect some assumptions below that rowcounts aren't zero */
3584 1476348 : if (outer_path_rows <= 0)
3585 96 : outer_path_rows = 1;
3586 1476348 : if (inner_path_rows <= 0)
3587 126 : inner_path_rows = 1;
3588 :
3589 : /*
3590 : * A merge join will stop as soon as it exhausts either input stream
3591 : * (unless it's an outer join, in which case the outer side has to be
3592 : * scanned all the way anyway). Estimate fraction of the left and right
3593 : * inputs that will actually need to be scanned. Likewise, we can
3594 : * estimate the number of rows that will be skipped before the first join
3595 : * pair is found, which should be factored into startup cost. We use only
3596 : * the first (most significant) merge clause for this purpose. Since
3597 : * mergejoinscansel() is a fairly expensive computation, we cache the
3598 : * results in the merge clause RestrictInfo.
3599 : */
3600 1476348 : if (mergeclauses && jointype != JOIN_FULL)
3601 1470192 : {
3602 1470192 : RestrictInfo *firstclause = (RestrictInfo *) linitial(mergeclauses);
3603 : List *opathkeys;
3604 : List *ipathkeys;
3605 : PathKey *opathkey;
3606 : PathKey *ipathkey;
3607 : MergeScanSelCache *cache;
3608 :
3609 : /* Get the input pathkeys to determine the sort-order details */
3610 1470192 : opathkeys = outersortkeys ? outersortkeys : outer_path->pathkeys;
3611 1470192 : ipathkeys = innersortkeys ? innersortkeys : inner_path->pathkeys;
3612 : Assert(opathkeys);
3613 : Assert(ipathkeys);
3614 1470192 : opathkey = (PathKey *) linitial(opathkeys);
3615 1470192 : ipathkey = (PathKey *) linitial(ipathkeys);
3616 : /* debugging check */
3617 1470192 : if (opathkey->pk_opfamily != ipathkey->pk_opfamily ||
3618 1470192 : opathkey->pk_eclass->ec_collation != ipathkey->pk_eclass->ec_collation ||
3619 1470192 : opathkey->pk_cmptype != ipathkey->pk_cmptype ||
3620 1470192 : opathkey->pk_nulls_first != ipathkey->pk_nulls_first)
3621 0 : elog(ERROR, "left and right pathkeys do not match in mergejoin");
3622 :
3623 : /* Get the selectivity with caching */
3624 1470192 : cache = cached_scansel(root, firstclause, opathkey);
3625 :
3626 1470192 : if (bms_is_subset(firstclause->left_relids,
3627 1470192 : outer_path->parent->relids))
3628 : {
3629 : /* left side of clause is outer */
3630 766760 : outerstartsel = cache->leftstartsel;
3631 766760 : outerendsel = cache->leftendsel;
3632 766760 : innerstartsel = cache->rightstartsel;
3633 766760 : innerendsel = cache->rightendsel;
3634 : }
3635 : else
3636 : {
3637 : /* left side of clause is inner */
3638 703432 : outerstartsel = cache->rightstartsel;
3639 703432 : outerendsel = cache->rightendsel;
3640 703432 : innerstartsel = cache->leftstartsel;
3641 703432 : innerendsel = cache->leftendsel;
3642 : }
3643 1470192 : if (jointype == JOIN_LEFT ||
3644 : jointype == JOIN_ANTI)
3645 : {
3646 197822 : outerstartsel = 0.0;
3647 197822 : outerendsel = 1.0;
3648 : }
3649 1272370 : else if (jointype == JOIN_RIGHT ||
3650 : jointype == JOIN_RIGHT_ANTI)
3651 : {
3652 197126 : innerstartsel = 0.0;
3653 197126 : innerendsel = 1.0;
3654 : }
3655 : }
3656 : else
3657 : {
3658 : /* cope with clauseless or full mergejoin */
3659 6156 : outerstartsel = innerstartsel = 0.0;
3660 6156 : outerendsel = innerendsel = 1.0;
3661 : }
3662 :
3663 : /*
3664 : * Convert selectivities to row counts. We force outer_rows and
3665 : * inner_rows to be at least 1, but the skip_rows estimates can be zero.
3666 : */
3667 1476348 : outer_skip_rows = rint(outer_path_rows * outerstartsel);
3668 1476348 : inner_skip_rows = rint(inner_path_rows * innerstartsel);
3669 1476348 : outer_rows = clamp_row_est(outer_path_rows * outerendsel);
3670 1476348 : inner_rows = clamp_row_est(inner_path_rows * innerendsel);
3671 :
3672 : Assert(outer_skip_rows <= outer_rows);
3673 : Assert(inner_skip_rows <= inner_rows);
3674 :
3675 : /*
3676 : * Readjust scan selectivities to account for above rounding. This is
3677 : * normally an insignificant effect, but when there are only a few rows in
3678 : * the inputs, failing to do this makes for a large percentage error.
3679 : */
3680 1476348 : outerstartsel = outer_skip_rows / outer_path_rows;
3681 1476348 : innerstartsel = inner_skip_rows / inner_path_rows;
3682 1476348 : outerendsel = outer_rows / outer_path_rows;
3683 1476348 : innerendsel = inner_rows / inner_path_rows;
3684 :
3685 : Assert(outerstartsel <= outerendsel);
3686 : Assert(innerstartsel <= innerendsel);
3687 :
3688 1476348 : disabled_nodes = enable_mergejoin ? 0 : 1;
3689 :
3690 : /* cost of source data */
3691 :
3692 1476348 : if (outersortkeys) /* do we need to sort outer? */
3693 : {
3694 : /*
3695 : * We can assert that the outer path is not already ordered
3696 : * appropriately for the mergejoin; otherwise, outersortkeys would
3697 : * have been set to NIL.
3698 : */
3699 : Assert(!pathkeys_contained_in(outersortkeys, outer_path->pathkeys));
3700 :
3701 : /*
3702 : * We choose to use incremental sort if it is enabled and there are
3703 : * presorted keys; otherwise we use full sort.
3704 : */
3705 754442 : if (enable_incremental_sort && outer_presorted_keys > 0)
3706 : {
3707 1698 : cost_incremental_sort(&sort_path,
3708 : root,
3709 : outersortkeys,
3710 : outer_presorted_keys,
3711 : outer_path->disabled_nodes,
3712 : outer_path->startup_cost,
3713 : outer_path->total_cost,
3714 : outer_path_rows,
3715 1698 : outer_path->pathtarget->width,
3716 : 0.0,
3717 : work_mem,
3718 : -1.0);
3719 : }
3720 : else
3721 : {
3722 752744 : cost_sort(&sort_path,
3723 : root,
3724 : outersortkeys,
3725 : outer_path->disabled_nodes,
3726 : outer_path->total_cost,
3727 : outer_path_rows,
3728 752744 : outer_path->pathtarget->width,
3729 : 0.0,
3730 : work_mem,
3731 : -1.0);
3732 : }
3733 :
3734 754442 : disabled_nodes += sort_path.disabled_nodes;
3735 754442 : startup_cost += sort_path.startup_cost;
3736 754442 : startup_cost += (sort_path.total_cost - sort_path.startup_cost)
3737 754442 : * outerstartsel;
3738 754442 : run_cost += (sort_path.total_cost - sort_path.startup_cost)
3739 754442 : * (outerendsel - outerstartsel);
3740 : }
3741 : else
3742 : {
3743 721906 : disabled_nodes += outer_path->disabled_nodes;
3744 721906 : startup_cost += outer_path->startup_cost;
3745 721906 : startup_cost += (outer_path->total_cost - outer_path->startup_cost)
3746 721906 : * outerstartsel;
3747 721906 : run_cost += (outer_path->total_cost - outer_path->startup_cost)
3748 721906 : * (outerendsel - outerstartsel);
3749 : }
3750 :
3751 1476348 : if (innersortkeys) /* do we need to sort inner? */
3752 : {
3753 : /*
3754 : * We can assert that the inner path is not already ordered
3755 : * appropriately for the mergejoin; otherwise, innersortkeys would
3756 : * have been set to NIL.
3757 : */
3758 : Assert(!pathkeys_contained_in(innersortkeys, inner_path->pathkeys));
3759 :
3760 : /*
3761 : * We do not consider incremental sort for inner path, because
3762 : * incremental sort does not support mark/restore.
3763 : */
3764 :
3765 1188296 : cost_sort(&sort_path,
3766 : root,
3767 : innersortkeys,
3768 : inner_path->disabled_nodes,
3769 : inner_path->total_cost,
3770 : inner_path_rows,
3771 1188296 : inner_path->pathtarget->width,
3772 : 0.0,
3773 : work_mem,
3774 : -1.0);
3775 1188296 : disabled_nodes += sort_path.disabled_nodes;
3776 1188296 : startup_cost += sort_path.startup_cost;
3777 1188296 : startup_cost += (sort_path.total_cost - sort_path.startup_cost)
3778 1188296 : * innerstartsel;
3779 1188296 : inner_run_cost = (sort_path.total_cost - sort_path.startup_cost)
3780 1188296 : * (innerendsel - innerstartsel);
3781 : }
3782 : else
3783 : {
3784 288052 : disabled_nodes += inner_path->disabled_nodes;
3785 288052 : startup_cost += inner_path->startup_cost;
3786 288052 : startup_cost += (inner_path->total_cost - inner_path->startup_cost)
3787 288052 : * innerstartsel;
3788 288052 : inner_run_cost = (inner_path->total_cost - inner_path->startup_cost)
3789 288052 : * (innerendsel - innerstartsel);
3790 : }
3791 :
3792 : /*
3793 : * We can't yet determine whether rescanning occurs, or whether
3794 : * materialization of the inner input should be done. The minimum
3795 : * possible inner input cost, regardless of rescan and materialization
3796 : * considerations, is inner_run_cost. We include that in
3797 : * workspace->total_cost, but not yet in run_cost.
3798 : */
3799 :
3800 : /* CPU costs left for later */
3801 :
3802 : /* Public result fields */
3803 1476348 : workspace->disabled_nodes = disabled_nodes;
3804 1476348 : workspace->startup_cost = startup_cost;
3805 1476348 : workspace->total_cost = startup_cost + run_cost + inner_run_cost;
3806 : /* Save private data for final_cost_mergejoin */
3807 1476348 : workspace->run_cost = run_cost;
3808 1476348 : workspace->inner_run_cost = inner_run_cost;
3809 1476348 : workspace->outer_rows = outer_rows;
3810 1476348 : workspace->inner_rows = inner_rows;
3811 1476348 : workspace->outer_skip_rows = outer_skip_rows;
3812 1476348 : workspace->inner_skip_rows = inner_skip_rows;
3813 1476348 : }
3814 :
3815 : /*
3816 : * final_cost_mergejoin
3817 : * Final estimate of the cost and result size of a mergejoin path.
3818 : *
3819 : * Unlike other costsize functions, this routine makes two actual decisions:
3820 : * whether the executor will need to do mark/restore, and whether we should
3821 : * materialize the inner path. It would be logically cleaner to build
3822 : * separate paths testing these alternatives, but that would require repeating
3823 : * most of the cost calculations, which are not all that cheap. Since the
3824 : * choice will not affect output pathkeys or startup cost, only total cost,
3825 : * there is no possibility of wanting to keep more than one path. So it seems
3826 : * best to make the decisions here and record them in the path's
3827 : * skip_mark_restore and materialize_inner fields.
3828 : *
3829 : * Mark/restore overhead is usually required, but can be skipped if we know
3830 : * that the executor need find only one match per outer tuple, and that the
3831 : * mergeclauses are sufficient to identify a match.
3832 : *
3833 : * We materialize the inner path if we need mark/restore and either the inner
3834 : * path can't support mark/restore, or it's cheaper to use an interposed
3835 : * Material node to handle mark/restore.
3836 : *
3837 : * 'path' is already filled in except for the rows and cost fields and
3838 : * skip_mark_restore and materialize_inner
3839 : * 'workspace' is the result from initial_cost_mergejoin
3840 : * 'extra' contains miscellaneous information about the join
3841 : */
3842 : void
3843 459832 : final_cost_mergejoin(PlannerInfo *root, MergePath *path,
3844 : JoinCostWorkspace *workspace,
3845 : JoinPathExtraData *extra)
3846 : {
3847 459832 : Path *outer_path = path->jpath.outerjoinpath;
3848 459832 : Path *inner_path = path->jpath.innerjoinpath;
3849 459832 : double inner_path_rows = inner_path->rows;
3850 459832 : List *mergeclauses = path->path_mergeclauses;
3851 459832 : List *innersortkeys = path->innersortkeys;
3852 459832 : Cost startup_cost = workspace->startup_cost;
3853 459832 : Cost run_cost = workspace->run_cost;
3854 459832 : Cost inner_run_cost = workspace->inner_run_cost;
3855 459832 : double outer_rows = workspace->outer_rows;
3856 459832 : double inner_rows = workspace->inner_rows;
3857 459832 : double outer_skip_rows = workspace->outer_skip_rows;
3858 459832 : double inner_skip_rows = workspace->inner_skip_rows;
3859 : Cost cpu_per_tuple,
3860 : bare_inner_cost,
3861 : mat_inner_cost;
3862 : QualCost merge_qual_cost;
3863 : QualCost qp_qual_cost;
3864 : double mergejointuples,
3865 : rescannedtuples;
3866 : double rescanratio;
3867 :
3868 : /* Set the number of disabled nodes. */
3869 459832 : path->jpath.path.disabled_nodes = workspace->disabled_nodes;
3870 :
3871 : /* Protect some assumptions below that rowcounts aren't zero */
3872 459832 : if (inner_path_rows <= 0)
3873 90 : inner_path_rows = 1;
3874 :
3875 : /* Mark the path with the correct row estimate */
3876 459832 : if (path->jpath.path.param_info)
3877 1644 : path->jpath.path.rows = path->jpath.path.param_info->ppi_rows;
3878 : else
3879 458188 : path->jpath.path.rows = path->jpath.path.parent->rows;
3880 :
3881 : /* For partial paths, scale row estimate. */
3882 459832 : if (path->jpath.path.parallel_workers > 0)
3883 : {
3884 65832 : double parallel_divisor = get_parallel_divisor(&path->jpath.path);
3885 :
3886 65832 : path->jpath.path.rows =
3887 65832 : clamp_row_est(path->jpath.path.rows / parallel_divisor);
3888 : }
3889 :
3890 : /*
3891 : * Compute cost of the mergequals and qpquals (other restriction clauses)
3892 : * separately.
3893 : */
3894 459832 : cost_qual_eval(&merge_qual_cost, mergeclauses, root);
3895 459832 : cost_qual_eval(&qp_qual_cost, path->jpath.joinrestrictinfo, root);
3896 459832 : qp_qual_cost.startup -= merge_qual_cost.startup;
3897 459832 : qp_qual_cost.per_tuple -= merge_qual_cost.per_tuple;
3898 :
3899 : /*
3900 : * With a SEMI or ANTI join, or if the innerrel is known unique, the
3901 : * executor will stop scanning for matches after the first match. When
3902 : * all the joinclauses are merge clauses, this means we don't ever need to
3903 : * back up the merge, and so we can skip mark/restore overhead.
3904 : */
3905 459832 : if ((path->jpath.jointype == JOIN_SEMI ||
3906 452760 : path->jpath.jointype == JOIN_ANTI ||
3907 604898 : extra->inner_unique) &&
3908 159128 : (list_length(path->jpath.joinrestrictinfo) ==
3909 159128 : list_length(path->path_mergeclauses)))
3910 137356 : path->skip_mark_restore = true;
3911 : else
3912 322476 : path->skip_mark_restore = false;
3913 :
3914 : /*
3915 : * Get approx # tuples passing the mergequals. We use approx_tuple_count
3916 : * here because we need an estimate done with JOIN_INNER semantics.
3917 : */
3918 459832 : mergejointuples = approx_tuple_count(root, &path->jpath, mergeclauses);
3919 :
3920 : /*
3921 : * When there are equal merge keys in the outer relation, the mergejoin
3922 : * must rescan any matching tuples in the inner relation. This means
3923 : * re-fetching inner tuples; we have to estimate how often that happens.
3924 : *
3925 : * For regular inner and outer joins, the number of re-fetches can be
3926 : * estimated approximately as size of merge join output minus size of
3927 : * inner relation. Assume that the distinct key values are 1, 2, ..., and
3928 : * denote the number of values of each key in the outer relation as m1,
3929 : * m2, ...; in the inner relation, n1, n2, ... Then we have
3930 : *
3931 : * size of join = m1 * n1 + m2 * n2 + ...
3932 : *
3933 : * number of rescanned tuples = (m1 - 1) * n1 + (m2 - 1) * n2 + ... = m1 *
3934 : * n1 + m2 * n2 + ... - (n1 + n2 + ...) = size of join - size of inner
3935 : * relation
3936 : *
3937 : * This equation works correctly for outer tuples having no inner match
3938 : * (nk = 0), but not for inner tuples having no outer match (mk = 0); we
3939 : * are effectively subtracting those from the number of rescanned tuples,
3940 : * when we should not. Can we do better without expensive selectivity
3941 : * computations?
3942 : *
3943 : * The whole issue is moot if we know we don't need to mark/restore at
3944 : * all, or if we are working from a unique-ified outer input.
3945 : */
3946 459832 : if (path->skip_mark_restore ||
3947 322476 : RELATION_WAS_MADE_UNIQUE(outer_path->parent, extra->sjinfo,
3948 : path->jpath.jointype))
3949 141670 : rescannedtuples = 0;
3950 : else
3951 : {
3952 318162 : rescannedtuples = mergejointuples - inner_path_rows;
3953 : /* Must clamp because of possible underestimate */
3954 318162 : if (rescannedtuples < 0)
3955 77642 : rescannedtuples = 0;
3956 : }
3957 :
3958 : /*
3959 : * We'll inflate various costs this much to account for rescanning. Note
3960 : * that this is to be multiplied by something involving inner_rows, or
3961 : * another number related to the portion of the inner rel we'll scan.
3962 : */
3963 459832 : rescanratio = 1.0 + (rescannedtuples / inner_rows);
3964 :
3965 : /*
3966 : * Decide whether we want to materialize the inner input to shield it from
3967 : * mark/restore and performing re-fetches. Our cost model for regular
3968 : * re-fetches is that a re-fetch costs the same as an original fetch,
3969 : * which is probably an overestimate; but on the other hand we ignore the
3970 : * bookkeeping costs of mark/restore. Not clear if it's worth developing
3971 : * a more refined model. So we just need to inflate the inner run cost by
3972 : * rescanratio.
3973 : */
3974 459832 : bare_inner_cost = inner_run_cost * rescanratio;
3975 :
3976 : /*
3977 : * When we interpose a Material node the re-fetch cost is assumed to be
3978 : * just cpu_operator_cost per tuple, independently of the underlying
3979 : * plan's cost; and we charge an extra cpu_operator_cost per original
3980 : * fetch as well. Note that we're assuming the materialize node will
3981 : * never spill to disk, since it only has to remember tuples back to the
3982 : * last mark. (If there are a huge number of duplicates, our other cost
3983 : * factors will make the path so expensive that it probably won't get
3984 : * chosen anyway.) So we don't use cost_rescan here.
3985 : *
3986 : * Note: keep this estimate in sync with create_mergejoin_plan's labeling
3987 : * of the generated Material node.
3988 : */
3989 459832 : mat_inner_cost = inner_run_cost +
3990 459832 : cpu_operator_cost * inner_rows * rescanratio;
3991 :
3992 : /*
3993 : * If we don't need mark/restore at all, we don't need materialization.
3994 : */
3995 459832 : if (path->skip_mark_restore)
3996 137356 : path->materialize_inner = false;
3997 :
3998 : /*
3999 : * Prefer materializing if it looks cheaper, unless the user has asked to
4000 : * suppress materialization.
4001 : */
4002 322476 : else if (enable_material && mat_inner_cost < bare_inner_cost)
4003 3542 : path->materialize_inner = true;
4004 :
4005 : /*
4006 : * Even if materializing doesn't look cheaper, we *must* do it if the
4007 : * inner path is to be used directly (without sorting) and it doesn't
4008 : * support mark/restore.
4009 : *
4010 : * Since the inner side must be ordered, and only Sorts and IndexScans can
4011 : * create order to begin with, and they both support mark/restore, you
4012 : * might think there's no problem --- but you'd be wrong. Nestloop and
4013 : * merge joins can *preserve* the order of their inputs, so they can be
4014 : * selected as the input of a mergejoin, and they don't support
4015 : * mark/restore at present.
4016 : *
4017 : * We don't test the value of enable_material here, because
4018 : * materialization is required for correctness in this case, and turning
4019 : * it off does not entitle us to deliver an invalid plan.
4020 : */
4021 318934 : else if (innersortkeys == NIL &&
4022 8624 : !ExecSupportsMarkRestore(inner_path))
4023 1876 : path->materialize_inner = true;
4024 :
4025 : /*
4026 : * Also, force materializing if the inner path is to be sorted and the
4027 : * sort is expected to spill to disk. This is because the final merge
4028 : * pass can be done on-the-fly if it doesn't have to support mark/restore.
4029 : * We don't try to adjust the cost estimates for this consideration,
4030 : * though.
4031 : *
4032 : * Since materialization is a performance optimization in this case,
4033 : * rather than necessary for correctness, we skip it if enable_material is
4034 : * off.
4035 : */
4036 317058 : else if (enable_material && innersortkeys != NIL &&
4037 310262 : relation_byte_size(inner_path_rows,
4038 310262 : inner_path->pathtarget->width) >
4039 310262 : work_mem * (Size) 1024)
4040 284 : path->materialize_inner = true;
4041 : else
4042 316774 : path->materialize_inner = false;
4043 :
4044 : /* Charge the right incremental cost for the chosen case */
4045 459832 : if (path->materialize_inner)
4046 5702 : run_cost += mat_inner_cost;
4047 : else
4048 454130 : run_cost += bare_inner_cost;
4049 :
4050 : /* CPU costs */
4051 :
4052 : /*
4053 : * The number of tuple comparisons needed is approximately number of outer
4054 : * rows plus number of inner rows plus number of rescanned tuples (can we
4055 : * refine this?). At each one, we need to evaluate the mergejoin quals.
4056 : */
4057 459832 : startup_cost += merge_qual_cost.startup;
4058 459832 : startup_cost += merge_qual_cost.per_tuple *
4059 459832 : (outer_skip_rows + inner_skip_rows * rescanratio);
4060 459832 : run_cost += merge_qual_cost.per_tuple *
4061 459832 : ((outer_rows - outer_skip_rows) +
4062 459832 : (inner_rows - inner_skip_rows) * rescanratio);
4063 :
4064 : /*
4065 : * For each tuple that gets through the mergejoin proper, we charge
4066 : * cpu_tuple_cost plus the cost of evaluating additional restriction
4067 : * clauses that are to be applied at the join. (This is pessimistic since
4068 : * not all of the quals may get evaluated at each tuple.)
4069 : *
4070 : * Note: we could adjust for SEMI/ANTI joins skipping some qual
4071 : * evaluations here, but it's probably not worth the trouble.
4072 : */
4073 459832 : startup_cost += qp_qual_cost.startup;
4074 459832 : cpu_per_tuple = cpu_tuple_cost + qp_qual_cost.per_tuple;
4075 459832 : run_cost += cpu_per_tuple * mergejointuples;
4076 :
4077 : /* tlist eval costs are paid per output row, not per tuple scanned */
4078 459832 : startup_cost += path->jpath.path.pathtarget->cost.startup;
4079 459832 : run_cost += path->jpath.path.pathtarget->cost.per_tuple * path->jpath.path.rows;
4080 :
4081 459832 : path->jpath.path.startup_cost = startup_cost;
4082 459832 : path->jpath.path.total_cost = startup_cost + run_cost;
4083 459832 : }
4084 :
4085 : /*
4086 : * run mergejoinscansel() with caching
4087 : */
4088 : static MergeScanSelCache *
4089 1470192 : cached_scansel(PlannerInfo *root, RestrictInfo *rinfo, PathKey *pathkey)
4090 : {
4091 : MergeScanSelCache *cache;
4092 : ListCell *lc;
4093 : Selectivity leftstartsel,
4094 : leftendsel,
4095 : rightstartsel,
4096 : rightendsel;
4097 : MemoryContext oldcontext;
4098 :
4099 : /* Do we have this result already? */
4100 1470270 : foreach(lc, rinfo->scansel_cache)
4101 : {
4102 1328696 : cache = (MergeScanSelCache *) lfirst(lc);
4103 1328696 : if (cache->opfamily == pathkey->pk_opfamily &&
4104 1328696 : cache->collation == pathkey->pk_eclass->ec_collation &&
4105 1328696 : cache->cmptype == pathkey->pk_cmptype &&
4106 1328618 : cache->nulls_first == pathkey->pk_nulls_first)
4107 1328618 : return cache;
4108 : }
4109 :
4110 : /* Nope, do the computation */
4111 141574 : mergejoinscansel(root,
4112 141574 : (Node *) rinfo->clause,
4113 : pathkey->pk_opfamily,
4114 : pathkey->pk_cmptype,
4115 141574 : pathkey->pk_nulls_first,
4116 : &leftstartsel,
4117 : &leftendsel,
4118 : &rightstartsel,
4119 : &rightendsel);
4120 :
4121 : /* Cache the result in suitably long-lived workspace */
4122 141574 : oldcontext = MemoryContextSwitchTo(root->planner_cxt);
4123 :
4124 141574 : cache = (MergeScanSelCache *) palloc(sizeof(MergeScanSelCache));
4125 141574 : cache->opfamily = pathkey->pk_opfamily;
4126 141574 : cache->collation = pathkey->pk_eclass->ec_collation;
4127 141574 : cache->cmptype = pathkey->pk_cmptype;
4128 141574 : cache->nulls_first = pathkey->pk_nulls_first;
4129 141574 : cache->leftstartsel = leftstartsel;
4130 141574 : cache->leftendsel = leftendsel;
4131 141574 : cache->rightstartsel = rightstartsel;
4132 141574 : cache->rightendsel = rightendsel;
4133 :
4134 141574 : rinfo->scansel_cache = lappend(rinfo->scansel_cache, cache);
4135 :
4136 141574 : MemoryContextSwitchTo(oldcontext);
4137 :
4138 141574 : return cache;
4139 : }
4140 :
4141 : /*
4142 : * initial_cost_hashjoin
4143 : * Preliminary estimate of the cost of a hashjoin path.
4144 : *
4145 : * This must quickly produce lower-bound estimates of the path's startup and
4146 : * total costs. If we are unable to eliminate the proposed path from
4147 : * consideration using the lower bounds, final_cost_hashjoin will be called
4148 : * to obtain the final estimates.
4149 : *
4150 : * The exact division of labor between this function and final_cost_hashjoin
4151 : * is private to them, and represents a tradeoff between speed of the initial
4152 : * estimate and getting a tight lower bound. We choose to not examine the
4153 : * join quals here (other than by counting the number of hash clauses),
4154 : * so we can't do much with CPU costs. We do assume that
4155 : * ExecChooseHashTableSize is cheap enough to use here.
4156 : *
4157 : * 'workspace' is to be filled with startup_cost, total_cost, and perhaps
4158 : * other data to be used by final_cost_hashjoin
4159 : * 'jointype' is the type of join to be performed
4160 : * 'hashclauses' is the list of joinclauses to be used as hash clauses
4161 : * 'outer_path' is the outer input to the join
4162 : * 'inner_path' is the inner input to the join
4163 : * 'extra' contains miscellaneous information about the join
4164 : * 'parallel_hash' indicates that inner_path is partial and that a shared
4165 : * hash table will be built in parallel
4166 : */
4167 : void
4168 875454 : initial_cost_hashjoin(PlannerInfo *root, JoinCostWorkspace *workspace,
4169 : JoinType jointype,
4170 : List *hashclauses,
4171 : Path *outer_path, Path *inner_path,
4172 : JoinPathExtraData *extra,
4173 : bool parallel_hash)
4174 : {
4175 : int disabled_nodes;
4176 875454 : Cost startup_cost = 0;
4177 875454 : Cost run_cost = 0;
4178 875454 : double outer_path_rows = outer_path->rows;
4179 875454 : double inner_path_rows = inner_path->rows;
4180 875454 : double inner_path_rows_total = inner_path_rows;
4181 875454 : int num_hashclauses = list_length(hashclauses);
4182 : int numbuckets;
4183 : int numbatches;
4184 : int num_skew_mcvs;
4185 : size_t space_allowed; /* unused */
4186 :
4187 : /* Count up disabled nodes. */
4188 875454 : disabled_nodes = enable_hashjoin ? 0 : 1;
4189 875454 : disabled_nodes += inner_path->disabled_nodes;
4190 875454 : disabled_nodes += outer_path->disabled_nodes;
4191 :
4192 : /* cost of source data */
4193 875454 : startup_cost += outer_path->startup_cost;
4194 875454 : run_cost += outer_path->total_cost - outer_path->startup_cost;
4195 875454 : startup_cost += inner_path->total_cost;
4196 :
4197 : /*
4198 : * Cost of computing hash function: must do it once per input tuple. We
4199 : * charge one cpu_operator_cost for each column's hash function. Also,
4200 : * tack on one cpu_tuple_cost per inner row, to model the costs of
4201 : * inserting the row into the hashtable.
4202 : *
4203 : * XXX when a hashclause is more complex than a single operator, we really
4204 : * should charge the extra eval costs of the left or right side, as
4205 : * appropriate, here. This seems more work than it's worth at the moment.
4206 : */
4207 875454 : startup_cost += (cpu_operator_cost * num_hashclauses + cpu_tuple_cost)
4208 875454 : * inner_path_rows;
4209 875454 : run_cost += cpu_operator_cost * num_hashclauses * outer_path_rows;
4210 :
4211 : /*
4212 : * If this is a parallel hash build, then the value we have for
4213 : * inner_rows_total currently refers only to the rows returned by each
4214 : * participant. For shared hash table size estimation, we need the total
4215 : * number, so we need to undo the division.
4216 : */
4217 875454 : if (parallel_hash)
4218 75156 : inner_path_rows_total *= get_parallel_divisor(inner_path);
4219 :
4220 : /*
4221 : * Get hash table size that executor would use for inner relation.
4222 : *
4223 : * XXX for the moment, always assume that skew optimization will be
4224 : * performed. As long as SKEW_HASH_MEM_PERCENT is small, it's not worth
4225 : * trying to determine that for sure.
4226 : *
4227 : * XXX at some point it might be interesting to try to account for skew
4228 : * optimization in the cost estimate, but for now, we don't.
4229 : */
4230 875454 : ExecChooseHashTableSize(inner_path_rows_total,
4231 875454 : inner_path->pathtarget->width,
4232 : true, /* useskew */
4233 : parallel_hash, /* try_combined_hash_mem */
4234 : outer_path->parallel_workers,
4235 : &space_allowed,
4236 : &numbuckets,
4237 : &numbatches,
4238 : &num_skew_mcvs);
4239 :
4240 : /*
4241 : * If inner relation is too big then we will need to "batch" the join,
4242 : * which implies writing and reading most of the tuples to disk an extra
4243 : * time. Charge seq_page_cost per page, since the I/O should be nice and
4244 : * sequential. Writing the inner rel counts as startup cost, all the rest
4245 : * as run cost.
4246 : */
4247 875454 : if (numbatches > 1)
4248 : {
4249 4664 : double outerpages = page_size(outer_path_rows,
4250 4664 : outer_path->pathtarget->width);
4251 4664 : double innerpages = page_size(inner_path_rows,
4252 4664 : inner_path->pathtarget->width);
4253 :
4254 4664 : startup_cost += seq_page_cost * innerpages;
4255 4664 : run_cost += seq_page_cost * (innerpages + 2 * outerpages);
4256 : }
4257 :
4258 : /* CPU costs left for later */
4259 :
4260 : /* Public result fields */
4261 875454 : workspace->disabled_nodes = disabled_nodes;
4262 875454 : workspace->startup_cost = startup_cost;
4263 875454 : workspace->total_cost = startup_cost + run_cost;
4264 : /* Save private data for final_cost_hashjoin */
4265 875454 : workspace->run_cost = run_cost;
4266 875454 : workspace->numbuckets = numbuckets;
4267 875454 : workspace->numbatches = numbatches;
4268 875454 : workspace->inner_rows_total = inner_path_rows_total;
4269 875454 : }
4270 :
4271 : /*
4272 : * final_cost_hashjoin
4273 : * Final estimate of the cost and result size of a hashjoin path.
4274 : *
4275 : * Note: the numbatches estimate is also saved into 'path' for use later
4276 : *
4277 : * 'path' is already filled in except for the rows and cost fields and
4278 : * num_batches
4279 : * 'workspace' is the result from initial_cost_hashjoin
4280 : * 'extra' contains miscellaneous information about the join
4281 : */
4282 : void
4283 456364 : final_cost_hashjoin(PlannerInfo *root, HashPath *path,
4284 : JoinCostWorkspace *workspace,
4285 : JoinPathExtraData *extra)
4286 : {
4287 456364 : Path *outer_path = path->jpath.outerjoinpath;
4288 456364 : Path *inner_path = path->jpath.innerjoinpath;
4289 456364 : double outer_path_rows = outer_path->rows;
4290 456364 : double inner_path_rows = inner_path->rows;
4291 456364 : double inner_path_rows_total = workspace->inner_rows_total;
4292 456364 : List *hashclauses = path->path_hashclauses;
4293 456364 : Cost startup_cost = workspace->startup_cost;
4294 456364 : Cost run_cost = workspace->run_cost;
4295 456364 : int numbuckets = workspace->numbuckets;
4296 456364 : int numbatches = workspace->numbatches;
4297 : Cost cpu_per_tuple;
4298 : QualCost hash_qual_cost;
4299 : QualCost qp_qual_cost;
4300 : double hashjointuples;
4301 : double virtualbuckets;
4302 : Selectivity innerbucketsize;
4303 : Selectivity innermcvfreq;
4304 : ListCell *hcl;
4305 :
4306 : /* Set the number of disabled nodes. */
4307 456364 : path->jpath.path.disabled_nodes = workspace->disabled_nodes;
4308 :
4309 : /* Mark the path with the correct row estimate */
4310 456364 : if (path->jpath.path.param_info)
4311 3674 : path->jpath.path.rows = path->jpath.path.param_info->ppi_rows;
4312 : else
4313 452690 : path->jpath.path.rows = path->jpath.path.parent->rows;
4314 :
4315 : /* For partial paths, scale row estimate. */
4316 456364 : if (path->jpath.path.parallel_workers > 0)
4317 : {
4318 107100 : double parallel_divisor = get_parallel_divisor(&path->jpath.path);
4319 :
4320 107100 : path->jpath.path.rows =
4321 107100 : clamp_row_est(path->jpath.path.rows / parallel_divisor);
4322 : }
4323 :
4324 : /* mark the path with estimated # of batches */
4325 456364 : path->num_batches = numbatches;
4326 :
4327 : /* store the total number of tuples (sum of partial row estimates) */
4328 456364 : path->inner_rows_total = inner_path_rows_total;
4329 :
4330 : /* and compute the number of "virtual" buckets in the whole join */
4331 456364 : virtualbuckets = (double) numbuckets * (double) numbatches;
4332 :
4333 : /*
4334 : * Determine bucketsize fraction and MCV frequency for the inner relation.
4335 : * We use the smallest bucketsize or MCV frequency estimated for any
4336 : * individual hashclause; this is undoubtedly conservative.
4337 : *
4338 : * BUT: if inner relation has been unique-ified, we can assume it's good
4339 : * for hashing. This is important both because it's the right answer, and
4340 : * because we avoid contaminating the cache with a value that's wrong for
4341 : * non-unique-ified paths.
4342 : */
4343 456364 : if (RELATION_WAS_MADE_UNIQUE(inner_path->parent, extra->sjinfo,
4344 : path->jpath.jointype))
4345 : {
4346 4232 : innerbucketsize = 1.0 / virtualbuckets;
4347 4232 : innermcvfreq = 0.0;
4348 : }
4349 : else
4350 : {
4351 : List *otherclauses;
4352 :
4353 452132 : innerbucketsize = 1.0;
4354 452132 : innermcvfreq = 1.0;
4355 :
4356 : /* At first, try to estimate bucket size using extended statistics. */
4357 452132 : otherclauses = estimate_multivariate_bucketsize(root,
4358 : inner_path->parent,
4359 : hashclauses,
4360 : &innerbucketsize);
4361 :
4362 : /* Pass through the remaining clauses */
4363 941098 : foreach(hcl, otherclauses)
4364 : {
4365 488966 : RestrictInfo *restrictinfo = lfirst_node(RestrictInfo, hcl);
4366 : Selectivity thisbucketsize;
4367 : Selectivity thismcvfreq;
4368 :
4369 : /*
4370 : * First we have to figure out which side of the hashjoin clause
4371 : * is the inner side.
4372 : *
4373 : * Since we tend to visit the same clauses over and over when
4374 : * planning a large query, we cache the bucket stats estimates in
4375 : * the RestrictInfo node to avoid repeated lookups of statistics.
4376 : */
4377 488966 : if (bms_is_subset(restrictinfo->right_relids,
4378 488966 : inner_path->parent->relids))
4379 : {
4380 : /* righthand side is inner */
4381 252898 : thisbucketsize = restrictinfo->right_bucketsize;
4382 252898 : if (thisbucketsize < 0)
4383 : {
4384 : /* not cached yet */
4385 109320 : estimate_hash_bucket_stats(root,
4386 109320 : get_rightop(restrictinfo->clause),
4387 : virtualbuckets,
4388 : &restrictinfo->right_mcvfreq,
4389 : &restrictinfo->right_bucketsize);
4390 109320 : thisbucketsize = restrictinfo->right_bucketsize;
4391 : }
4392 252898 : thismcvfreq = restrictinfo->right_mcvfreq;
4393 : }
4394 : else
4395 : {
4396 : Assert(bms_is_subset(restrictinfo->left_relids,
4397 : inner_path->parent->relids));
4398 : /* lefthand side is inner */
4399 236068 : thisbucketsize = restrictinfo->left_bucketsize;
4400 236068 : if (thisbucketsize < 0)
4401 : {
4402 : /* not cached yet */
4403 95920 : estimate_hash_bucket_stats(root,
4404 95920 : get_leftop(restrictinfo->clause),
4405 : virtualbuckets,
4406 : &restrictinfo->left_mcvfreq,
4407 : &restrictinfo->left_bucketsize);
4408 95920 : thisbucketsize = restrictinfo->left_bucketsize;
4409 : }
4410 236068 : thismcvfreq = restrictinfo->left_mcvfreq;
4411 : }
4412 :
4413 488966 : if (innerbucketsize > thisbucketsize)
4414 370488 : innerbucketsize = thisbucketsize;
4415 488966 : if (innermcvfreq > thismcvfreq)
4416 455146 : innermcvfreq = thismcvfreq;
4417 : }
4418 : }
4419 :
4420 : /*
4421 : * If the bucket holding the inner MCV would exceed hash_mem, we don't
4422 : * want to hash unless there is really no other alternative, so apply
4423 : * disable_cost. (The executor normally copes with excessive memory usage
4424 : * by splitting batches, but obviously it cannot separate equal values
4425 : * that way, so it will be unable to drive the batch size below hash_mem
4426 : * when this is true.)
4427 : */
4428 456364 : if (relation_byte_size(clamp_row_est(inner_path_rows * innermcvfreq),
4429 912728 : inner_path->pathtarget->width) > get_hash_memory_limit())
4430 8 : startup_cost += disable_cost;
4431 :
4432 : /*
4433 : * Compute cost of the hashquals and qpquals (other restriction clauses)
4434 : * separately.
4435 : */
4436 456364 : cost_qual_eval(&hash_qual_cost, hashclauses, root);
4437 456364 : cost_qual_eval(&qp_qual_cost, path->jpath.joinrestrictinfo, root);
4438 456364 : qp_qual_cost.startup -= hash_qual_cost.startup;
4439 456364 : qp_qual_cost.per_tuple -= hash_qual_cost.per_tuple;
4440 :
4441 : /* CPU costs */
4442 :
4443 456364 : if (path->jpath.jointype == JOIN_SEMI ||
4444 450180 : path->jpath.jointype == JOIN_ANTI ||
4445 445636 : extra->inner_unique)
4446 124856 : {
4447 : double outer_matched_rows;
4448 : Selectivity inner_scan_frac;
4449 :
4450 : /*
4451 : * With a SEMI or ANTI join, or if the innerrel is known unique, the
4452 : * executor will stop after the first match.
4453 : *
4454 : * For an outer-rel row that has at least one match, we can expect the
4455 : * bucket scan to stop after a fraction 1/(match_count+1) of the
4456 : * bucket's rows, if the matches are evenly distributed. Since they
4457 : * probably aren't quite evenly distributed, we apply a fuzz factor of
4458 : * 2.0 to that fraction. (If we used a larger fuzz factor, we'd have
4459 : * to clamp inner_scan_frac to at most 1.0; but since match_count is
4460 : * at least 1, no such clamp is needed now.)
4461 : */
4462 124856 : outer_matched_rows = rint(outer_path_rows * extra->semifactors.outer_match_frac);
4463 124856 : inner_scan_frac = 2.0 / (extra->semifactors.match_count + 1.0);
4464 :
4465 124856 : startup_cost += hash_qual_cost.startup;
4466 249712 : run_cost += hash_qual_cost.per_tuple * outer_matched_rows *
4467 124856 : clamp_row_est(inner_path_rows * innerbucketsize * inner_scan_frac) * 0.5;
4468 :
4469 : /*
4470 : * For unmatched outer-rel rows, the picture is quite a lot different.
4471 : * In the first place, there is no reason to assume that these rows
4472 : * preferentially hit heavily-populated buckets; instead assume they
4473 : * are uncorrelated with the inner distribution and so they see an
4474 : * average bucket size of inner_path_rows / virtualbuckets. In the
4475 : * second place, it seems likely that they will have few if any exact
4476 : * hash-code matches and so very few of the tuples in the bucket will
4477 : * actually require eval of the hash quals. We don't have any good
4478 : * way to estimate how many will, but for the moment assume that the
4479 : * effective cost per bucket entry is one-tenth what it is for
4480 : * matchable tuples.
4481 : */
4482 249712 : run_cost += hash_qual_cost.per_tuple *
4483 249712 : (outer_path_rows - outer_matched_rows) *
4484 124856 : clamp_row_est(inner_path_rows / virtualbuckets) * 0.05;
4485 :
4486 : /* Get # of tuples that will pass the basic join */
4487 124856 : if (path->jpath.jointype == JOIN_ANTI)
4488 4544 : hashjointuples = outer_path_rows - outer_matched_rows;
4489 : else
4490 120312 : hashjointuples = outer_matched_rows;
4491 : }
4492 : else
4493 : {
4494 : /*
4495 : * The number of tuple comparisons needed is the number of outer
4496 : * tuples times the typical number of tuples in a hash bucket, which
4497 : * is the inner relation size times its bucketsize fraction. At each
4498 : * one, we need to evaluate the hashjoin quals. But actually,
4499 : * charging the full qual eval cost at each tuple is pessimistic,
4500 : * since we don't evaluate the quals unless the hash values match
4501 : * exactly. For lack of a better idea, halve the cost estimate to
4502 : * allow for that.
4503 : */
4504 331508 : startup_cost += hash_qual_cost.startup;
4505 663016 : run_cost += hash_qual_cost.per_tuple * outer_path_rows *
4506 331508 : clamp_row_est(inner_path_rows * innerbucketsize) * 0.5;
4507 :
4508 : /*
4509 : * Get approx # tuples passing the hashquals. We use
4510 : * approx_tuple_count here because we need an estimate done with
4511 : * JOIN_INNER semantics.
4512 : */
4513 331508 : hashjointuples = approx_tuple_count(root, &path->jpath, hashclauses);
4514 : }
4515 :
4516 : /*
4517 : * For each tuple that gets through the hashjoin proper, we charge
4518 : * cpu_tuple_cost plus the cost of evaluating additional restriction
4519 : * clauses that are to be applied at the join. (This is pessimistic since
4520 : * not all of the quals may get evaluated at each tuple.)
4521 : */
4522 456364 : startup_cost += qp_qual_cost.startup;
4523 456364 : cpu_per_tuple = cpu_tuple_cost + qp_qual_cost.per_tuple;
4524 456364 : run_cost += cpu_per_tuple * hashjointuples;
4525 :
4526 : /* tlist eval costs are paid per output row, not per tuple scanned */
4527 456364 : startup_cost += path->jpath.path.pathtarget->cost.startup;
4528 456364 : run_cost += path->jpath.path.pathtarget->cost.per_tuple * path->jpath.path.rows;
4529 :
4530 456364 : path->jpath.path.startup_cost = startup_cost;
4531 456364 : path->jpath.path.total_cost = startup_cost + run_cost;
4532 456364 : }
4533 :
4534 :
4535 : /*
4536 : * cost_subplan
4537 : * Figure the costs for a SubPlan (or initplan).
4538 : *
4539 : * Note: we could dig the subplan's Plan out of the root list, but in practice
4540 : * all callers have it handy already, so we make them pass it.
4541 : */
4542 : void
4543 43618 : cost_subplan(PlannerInfo *root, SubPlan *subplan, Plan *plan)
4544 : {
4545 : QualCost sp_cost;
4546 :
4547 : /*
4548 : * Figure any cost for evaluating the testexpr.
4549 : *
4550 : * Usually, SubPlan nodes are built very early, before we have constructed
4551 : * any RelOptInfos for the parent query level, which means the parent root
4552 : * does not yet contain enough information to safely consult statistics.
4553 : * Therefore, we pass root as NULL here. cost_qual_eval() is already
4554 : * well-equipped to handle a NULL root.
4555 : *
4556 : * One exception is SubPlan nodes built for the initplans of MIN/MAX
4557 : * aggregates from indexes (cf. SS_make_initplan_from_plan). In this
4558 : * case, having a NULL root is safe because testexpr will be NULL.
4559 : * Besides, an initplan will by definition not consult anything from the
4560 : * parent plan.
4561 : */
4562 43618 : cost_qual_eval(&sp_cost,
4563 43618 : make_ands_implicit((Expr *) subplan->testexpr),
4564 : NULL);
4565 :
4566 43618 : if (subplan->useHashTable)
4567 : {
4568 : /*
4569 : * If we are using a hash table for the subquery outputs, then the
4570 : * cost of evaluating the query is a one-time cost. We charge one
4571 : * cpu_operator_cost per tuple for the work of loading the hashtable,
4572 : * too.
4573 : */
4574 2116 : sp_cost.startup += plan->total_cost +
4575 2116 : cpu_operator_cost * plan->plan_rows;
4576 :
4577 : /*
4578 : * The per-tuple costs include the cost of evaluating the lefthand
4579 : * expressions, plus the cost of probing the hashtable. We already
4580 : * accounted for the lefthand expressions as part of the testexpr, and
4581 : * will also have counted one cpu_operator_cost for each comparison
4582 : * operator. That is probably too low for the probing cost, but it's
4583 : * hard to make a better estimate, so live with it for now.
4584 : */
4585 : }
4586 : else
4587 : {
4588 : /*
4589 : * Otherwise we will be rescanning the subplan output on each
4590 : * evaluation. We need to estimate how much of the output we will
4591 : * actually need to scan. NOTE: this logic should agree with the
4592 : * tuple_fraction estimates used by make_subplan() in
4593 : * plan/subselect.c.
4594 : */
4595 41502 : Cost plan_run_cost = plan->total_cost - plan->startup_cost;
4596 :
4597 41502 : if (subplan->subLinkType == EXISTS_SUBLINK)
4598 : {
4599 : /* we only need to fetch 1 tuple; clamp to avoid zero divide */
4600 2506 : sp_cost.per_tuple += plan_run_cost / clamp_row_est(plan->plan_rows);
4601 : }
4602 38996 : else if (subplan->subLinkType == ALL_SUBLINK ||
4603 38978 : subplan->subLinkType == ANY_SUBLINK)
4604 : {
4605 : /* assume we need 50% of the tuples */
4606 146 : sp_cost.per_tuple += 0.50 * plan_run_cost;
4607 : /* also charge a cpu_operator_cost per row examined */
4608 146 : sp_cost.per_tuple += 0.50 * plan->plan_rows * cpu_operator_cost;
4609 : }
4610 : else
4611 : {
4612 : /* assume we need all tuples */
4613 38850 : sp_cost.per_tuple += plan_run_cost;
4614 : }
4615 :
4616 : /*
4617 : * Also account for subplan's startup cost. If the subplan is
4618 : * uncorrelated or undirect correlated, AND its topmost node is one
4619 : * that materializes its output, assume that we'll only need to pay
4620 : * its startup cost once; otherwise assume we pay the startup cost
4621 : * every time.
4622 : */
4623 54736 : if (subplan->parParam == NIL &&
4624 13234 : ExecMaterializesOutput(nodeTag(plan)))
4625 726 : sp_cost.startup += plan->startup_cost;
4626 : else
4627 40776 : sp_cost.per_tuple += plan->startup_cost;
4628 : }
4629 :
4630 43618 : subplan->startup_cost = sp_cost.startup;
4631 43618 : subplan->per_call_cost = sp_cost.per_tuple;
4632 43618 : }
4633 :
4634 :
4635 : /*
4636 : * cost_rescan
4637 : * Given a finished Path, estimate the costs of rescanning it after
4638 : * having done so the first time. For some Path types a rescan is
4639 : * cheaper than an original scan (if no parameters change), and this
4640 : * function embodies knowledge about that. The default is to return
4641 : * the same costs stored in the Path. (Note that the cost estimates
4642 : * actually stored in Paths are always for first scans.)
4643 : *
4644 : * This function is not currently intended to model effects such as rescans
4645 : * being cheaper due to disk block caching; what we are concerned with is
4646 : * plan types wherein the executor caches results explicitly, or doesn't
4647 : * redo startup calculations, etc.
4648 : */
4649 : static void
4650 3259348 : cost_rescan(PlannerInfo *root, Path *path,
4651 : Cost *rescan_startup_cost, /* output parameters */
4652 : Cost *rescan_total_cost)
4653 : {
4654 3259348 : switch (path->pathtype)
4655 : {
4656 52998 : case T_FunctionScan:
4657 :
4658 : /*
4659 : * Currently, nodeFunctionscan.c always executes the function to
4660 : * completion before returning any rows, and caches the results in
4661 : * a tuplestore. So the function eval cost is all startup cost
4662 : * and isn't paid over again on rescans. However, all run costs
4663 : * will be paid over again.
4664 : */
4665 52998 : *rescan_startup_cost = 0;
4666 52998 : *rescan_total_cost = path->total_cost - path->startup_cost;
4667 52998 : break;
4668 135374 : case T_HashJoin:
4669 :
4670 : /*
4671 : * If it's a single-batch join, we don't need to rebuild the hash
4672 : * table during a rescan.
4673 : */
4674 135374 : if (((HashPath *) path)->num_batches == 1)
4675 : {
4676 : /* Startup cost is exactly the cost of hash table building */
4677 135374 : *rescan_startup_cost = 0;
4678 135374 : *rescan_total_cost = path->total_cost - path->startup_cost;
4679 : }
4680 : else
4681 : {
4682 : /* Otherwise, no special treatment */
4683 0 : *rescan_startup_cost = path->startup_cost;
4684 0 : *rescan_total_cost = path->total_cost;
4685 : }
4686 135374 : break;
4687 8014 : case T_CteScan:
4688 : case T_WorkTableScan:
4689 : {
4690 : /*
4691 : * These plan types materialize their final result in a
4692 : * tuplestore or tuplesort object. So the rescan cost is only
4693 : * cpu_tuple_cost per tuple, unless the result is large enough
4694 : * to spill to disk.
4695 : */
4696 8014 : Cost run_cost = cpu_tuple_cost * path->rows;
4697 8014 : double nbytes = relation_byte_size(path->rows,
4698 8014 : path->pathtarget->width);
4699 8014 : double work_mem_bytes = work_mem * (Size) 1024;
4700 :
4701 8014 : if (nbytes > work_mem_bytes)
4702 : {
4703 : /* It will spill, so account for re-read cost */
4704 352 : double npages = ceil(nbytes / BLCKSZ);
4705 :
4706 352 : run_cost += seq_page_cost * npages;
4707 : }
4708 8014 : *rescan_startup_cost = 0;
4709 8014 : *rescan_total_cost = run_cost;
4710 : }
4711 8014 : break;
4712 1176396 : case T_Material:
4713 : case T_Sort:
4714 : {
4715 : /*
4716 : * These plan types not only materialize their results, but do
4717 : * not implement qual filtering or projection. So they are
4718 : * even cheaper to rescan than the ones above. We charge only
4719 : * cpu_operator_cost per tuple. (Note: keep that in sync with
4720 : * the run_cost charge in cost_sort, and also see comments in
4721 : * cost_material before you change it.)
4722 : */
4723 1176396 : Cost run_cost = cpu_operator_cost * path->rows;
4724 1176396 : double nbytes = relation_byte_size(path->rows,
4725 1176396 : path->pathtarget->width);
4726 1176396 : double work_mem_bytes = work_mem * (Size) 1024;
4727 :
4728 1176396 : if (nbytes > work_mem_bytes)
4729 : {
4730 : /* It will spill, so account for re-read cost */
4731 9942 : double npages = ceil(nbytes / BLCKSZ);
4732 :
4733 9942 : run_cost += seq_page_cost * npages;
4734 : }
4735 1176396 : *rescan_startup_cost = 0;
4736 1176396 : *rescan_total_cost = run_cost;
4737 : }
4738 1176396 : break;
4739 291440 : case T_Memoize:
4740 : /* All the hard work is done by cost_memoize_rescan */
4741 291440 : cost_memoize_rescan(root, (MemoizePath *) path,
4742 : rescan_startup_cost, rescan_total_cost);
4743 291440 : break;
4744 1595126 : default:
4745 1595126 : *rescan_startup_cost = path->startup_cost;
4746 1595126 : *rescan_total_cost = path->total_cost;
4747 1595126 : break;
4748 : }
4749 3259348 : }
4750 :
4751 :
4752 : /*
4753 : * cost_qual_eval
4754 : * Estimate the CPU costs of evaluating a WHERE clause.
4755 : * The input can be either an implicitly-ANDed list of boolean
4756 : * expressions, or a list of RestrictInfo nodes. (The latter is
4757 : * preferred since it allows caching of the results.)
4758 : * The result includes both a one-time (startup) component,
4759 : * and a per-evaluation component.
4760 : *
4761 : * Note: in some code paths root can be passed as NULL, resulting in
4762 : * slightly worse estimates.
4763 : */
4764 : void
4765 4644346 : cost_qual_eval(QualCost *cost, List *quals, PlannerInfo *root)
4766 : {
4767 : cost_qual_eval_context context;
4768 : ListCell *l;
4769 :
4770 4644346 : context.root = root;
4771 4644346 : context.total.startup = 0;
4772 4644346 : context.total.per_tuple = 0;
4773 :
4774 : /* We don't charge any cost for the implicit ANDing at top level ... */
4775 :
4776 8857452 : foreach(l, quals)
4777 : {
4778 4213106 : Node *qual = (Node *) lfirst(l);
4779 :
4780 4213106 : cost_qual_eval_walker(qual, &context);
4781 : }
4782 :
4783 4644346 : *cost = context.total;
4784 4644346 : }
4785 :
4786 : /*
4787 : * cost_qual_eval_node
4788 : * As above, for a single RestrictInfo or expression.
4789 : */
4790 : void
4791 1827182 : cost_qual_eval_node(QualCost *cost, Node *qual, PlannerInfo *root)
4792 : {
4793 : cost_qual_eval_context context;
4794 :
4795 1827182 : context.root = root;
4796 1827182 : context.total.startup = 0;
4797 1827182 : context.total.per_tuple = 0;
4798 :
4799 1827182 : cost_qual_eval_walker(qual, &context);
4800 :
4801 1827182 : *cost = context.total;
4802 1827182 : }
4803 :
4804 : static bool
4805 9481420 : cost_qual_eval_walker(Node *node, cost_qual_eval_context *context)
4806 : {
4807 9481420 : if (node == NULL)
4808 87926 : return false;
4809 :
4810 : /*
4811 : * RestrictInfo nodes contain an eval_cost field reserved for this
4812 : * routine's use, so that it's not necessary to evaluate the qual clause's
4813 : * cost more than once. If the clause's cost hasn't been computed yet,
4814 : * the field's startup value will contain -1.
4815 : */
4816 9393494 : if (IsA(node, RestrictInfo))
4817 : {
4818 4400302 : RestrictInfo *rinfo = (RestrictInfo *) node;
4819 :
4820 4400302 : if (rinfo->eval_cost.startup < 0)
4821 : {
4822 : cost_qual_eval_context locContext;
4823 :
4824 597028 : locContext.root = context->root;
4825 597028 : locContext.total.startup = 0;
4826 597028 : locContext.total.per_tuple = 0;
4827 :
4828 : /*
4829 : * For an OR clause, recurse into the marked-up tree so that we
4830 : * set the eval_cost for contained RestrictInfos too.
4831 : */
4832 597028 : if (rinfo->orclause)
4833 9530 : cost_qual_eval_walker((Node *) rinfo->orclause, &locContext);
4834 : else
4835 587498 : cost_qual_eval_walker((Node *) rinfo->clause, &locContext);
4836 :
4837 : /*
4838 : * If the RestrictInfo is marked pseudoconstant, it will be tested
4839 : * only once, so treat its cost as all startup cost.
4840 : */
4841 597028 : if (rinfo->pseudoconstant)
4842 : {
4843 : /* count one execution during startup */
4844 10064 : locContext.total.startup += locContext.total.per_tuple;
4845 10064 : locContext.total.per_tuple = 0;
4846 : }
4847 597028 : rinfo->eval_cost = locContext.total;
4848 : }
4849 4400302 : context->total.startup += rinfo->eval_cost.startup;
4850 4400302 : context->total.per_tuple += rinfo->eval_cost.per_tuple;
4851 : /* do NOT recurse into children */
4852 4400302 : return false;
4853 : }
4854 :
4855 : /*
4856 : * For each operator or function node in the given tree, we charge the
4857 : * estimated execution cost given by pg_proc.procost (remember to multiply
4858 : * this by cpu_operator_cost).
4859 : *
4860 : * Vars and Consts are charged zero, and so are boolean operators (AND,
4861 : * OR, NOT). Simplistic, but a lot better than no model at all.
4862 : *
4863 : * Should we try to account for the possibility of short-circuit
4864 : * evaluation of AND/OR? Probably *not*, because that would make the
4865 : * results depend on the clause ordering, and we are not in any position
4866 : * to expect that the current ordering of the clauses is the one that's
4867 : * going to end up being used. The above per-RestrictInfo caching would
4868 : * not mix well with trying to re-order clauses anyway.
4869 : *
4870 : * Another issue that is entirely ignored here is that if a set-returning
4871 : * function is below top level in the tree, the functions/operators above
4872 : * it will need to be evaluated multiple times. In practical use, such
4873 : * cases arise so seldom as to not be worth the added complexity needed;
4874 : * moreover, since our rowcount estimates for functions tend to be pretty
4875 : * phony, the results would also be pretty phony.
4876 : */
4877 4993192 : if (IsA(node, FuncExpr))
4878 : {
4879 342150 : add_function_cost(context->root, ((FuncExpr *) node)->funcid, node,
4880 : &context->total);
4881 : }
4882 4651042 : else if (IsA(node, OpExpr) ||
4883 3998546 : IsA(node, DistinctExpr) ||
4884 3997472 : IsA(node, NullIfExpr))
4885 : {
4886 : /* rely on struct equivalence to treat these all alike */
4887 653694 : set_opfuncid((OpExpr *) node);
4888 653694 : add_function_cost(context->root, ((OpExpr *) node)->opfuncid, node,
4889 : &context->total);
4890 : }
4891 3997348 : else if (IsA(node, ScalarArrayOpExpr))
4892 : {
4893 44388 : ScalarArrayOpExpr *saop = (ScalarArrayOpExpr *) node;
4894 44388 : Node *arraynode = (Node *) lsecond(saop->args);
4895 : QualCost sacosts;
4896 : QualCost hcosts;
4897 44388 : double estarraylen = estimate_array_length(context->root, arraynode);
4898 :
4899 44388 : set_sa_opfuncid(saop);
4900 44388 : sacosts.startup = sacosts.per_tuple = 0;
4901 44388 : add_function_cost(context->root, saop->opfuncid, NULL,
4902 : &sacosts);
4903 :
4904 44388 : if (OidIsValid(saop->hashfuncid))
4905 : {
4906 : /* Handle costs for hashed ScalarArrayOpExpr */
4907 430 : hcosts.startup = hcosts.per_tuple = 0;
4908 :
4909 430 : add_function_cost(context->root, saop->hashfuncid, NULL, &hcosts);
4910 430 : context->total.startup += sacosts.startup + hcosts.startup;
4911 :
4912 : /* Estimate the cost of building the hashtable. */
4913 430 : context->total.startup += estarraylen * hcosts.per_tuple;
4914 :
4915 : /*
4916 : * XXX should we charge a little bit for sacosts.per_tuple when
4917 : * building the table, or is it ok to assume there will be zero
4918 : * hash collision?
4919 : */
4920 :
4921 : /*
4922 : * Charge for hashtable lookups. Charge a single hash and a
4923 : * single comparison.
4924 : */
4925 430 : context->total.per_tuple += hcosts.per_tuple + sacosts.per_tuple;
4926 : }
4927 : else
4928 : {
4929 : /*
4930 : * Estimate that the operator will be applied to about half of the
4931 : * array elements before the answer is determined.
4932 : */
4933 43958 : context->total.startup += sacosts.startup;
4934 87916 : context->total.per_tuple += sacosts.per_tuple *
4935 43958 : estimate_array_length(context->root, arraynode) * 0.5;
4936 : }
4937 : }
4938 3952960 : else if (IsA(node, Aggref) ||
4939 3887684 : IsA(node, WindowFunc))
4940 : {
4941 : /*
4942 : * Aggref and WindowFunc nodes are (and should be) treated like Vars,
4943 : * ie, zero execution cost in the current model, because they behave
4944 : * essentially like Vars at execution. We disregard the costs of
4945 : * their input expressions for the same reason. The actual execution
4946 : * costs of the aggregate/window functions and their arguments have to
4947 : * be factored into plan-node-specific costing of the Agg or WindowAgg
4948 : * plan node.
4949 : */
4950 69142 : return false; /* don't recurse into children */
4951 : }
4952 3883818 : else if (IsA(node, GroupingFunc))
4953 : {
4954 : /* Treat this as having cost 1 */
4955 422 : context->total.per_tuple += cpu_operator_cost;
4956 422 : return false; /* don't recurse into children */
4957 : }
4958 3883396 : else if (IsA(node, CoerceViaIO))
4959 : {
4960 22242 : CoerceViaIO *iocoerce = (CoerceViaIO *) node;
4961 : Oid iofunc;
4962 : Oid typioparam;
4963 : bool typisvarlena;
4964 :
4965 : /* check the result type's input function */
4966 22242 : getTypeInputInfo(iocoerce->resulttype,
4967 : &iofunc, &typioparam);
4968 22242 : add_function_cost(context->root, iofunc, NULL,
4969 : &context->total);
4970 : /* check the input type's output function */
4971 22242 : getTypeOutputInfo(exprType((Node *) iocoerce->arg),
4972 : &iofunc, &typisvarlena);
4973 22242 : add_function_cost(context->root, iofunc, NULL,
4974 : &context->total);
4975 : }
4976 3861154 : else if (IsA(node, ArrayCoerceExpr))
4977 : {
4978 5102 : ArrayCoerceExpr *acoerce = (ArrayCoerceExpr *) node;
4979 : QualCost perelemcost;
4980 :
4981 5102 : cost_qual_eval_node(&perelemcost, (Node *) acoerce->elemexpr,
4982 : context->root);
4983 5102 : context->total.startup += perelemcost.startup;
4984 5102 : if (perelemcost.per_tuple > 0)
4985 66 : context->total.per_tuple += perelemcost.per_tuple *
4986 66 : estimate_array_length(context->root, (Node *) acoerce->arg);
4987 : }
4988 3856052 : else if (IsA(node, RowCompareExpr))
4989 : {
4990 : /* Conservatively assume we will check all the columns */
4991 252 : RowCompareExpr *rcexpr = (RowCompareExpr *) node;
4992 : ListCell *lc;
4993 :
4994 810 : foreach(lc, rcexpr->opnos)
4995 : {
4996 558 : Oid opid = lfirst_oid(lc);
4997 :
4998 558 : add_function_cost(context->root, get_opcode(opid), NULL,
4999 : &context->total);
5000 : }
5001 : }
5002 3855800 : else if (IsA(node, MinMaxExpr) ||
5003 3855528 : IsA(node, SQLValueFunction) ||
5004 3850762 : IsA(node, XmlExpr) ||
5005 3850060 : IsA(node, CoerceToDomain) ||
5006 3840342 : IsA(node, NextValueExpr) ||
5007 3839982 : IsA(node, JsonExpr))
5008 : {
5009 : /* Treat all these as having cost 1 */
5010 18390 : context->total.per_tuple += cpu_operator_cost;
5011 : }
5012 3837410 : else if (IsA(node, SubLink))
5013 : {
5014 : /* This routine should not be applied to un-planned expressions */
5015 0 : elog(ERROR, "cannot handle unplanned sub-select");
5016 : }
5017 3837410 : else if (IsA(node, SubPlan))
5018 : {
5019 : /*
5020 : * A subplan node in an expression typically indicates that the
5021 : * subplan will be executed on each evaluation, so charge accordingly.
5022 : * (Sub-selects that can be executed as InitPlans have already been
5023 : * removed from the expression.)
5024 : */
5025 42974 : SubPlan *subplan = (SubPlan *) node;
5026 :
5027 42974 : context->total.startup += subplan->startup_cost;
5028 42974 : context->total.per_tuple += subplan->per_call_cost;
5029 :
5030 : /*
5031 : * We don't want to recurse into the testexpr, because it was already
5032 : * counted in the SubPlan node's costs. So we're done.
5033 : */
5034 42974 : return false;
5035 : }
5036 3794436 : else if (IsA(node, AlternativeSubPlan))
5037 : {
5038 : /*
5039 : * Arbitrarily use the first alternative plan for costing. (We should
5040 : * certainly only include one alternative, and we don't yet have
5041 : * enough information to know which one the executor is most likely to
5042 : * use.)
5043 : */
5044 1848 : AlternativeSubPlan *asplan = (AlternativeSubPlan *) node;
5045 :
5046 1848 : return cost_qual_eval_walker((Node *) linitial(asplan->subplans),
5047 : context);
5048 : }
5049 3792588 : else if (IsA(node, PlaceHolderVar))
5050 : {
5051 : /*
5052 : * A PlaceHolderVar should be given cost zero when considering general
5053 : * expression evaluation costs. The expense of doing the contained
5054 : * expression is charged as part of the tlist eval costs of the scan
5055 : * or join where the PHV is first computed (see set_rel_width and
5056 : * add_placeholders_to_joinrel). If we charged it again here, we'd be
5057 : * double-counting the cost for each level of plan that the PHV
5058 : * bubbles up through. Hence, return without recursing into the
5059 : * phexpr.
5060 : */
5061 5196 : return false;
5062 : }
5063 :
5064 : /* recurse into children */
5065 4873610 : return expression_tree_walker(node, cost_qual_eval_walker, context);
5066 : }
5067 :
5068 : /*
5069 : * get_restriction_qual_cost
5070 : * Compute evaluation costs of a baserel's restriction quals, plus any
5071 : * movable join quals that have been pushed down to the scan.
5072 : * Results are returned into *qpqual_cost.
5073 : *
5074 : * This is a convenience subroutine that works for seqscans and other cases
5075 : * where all the given quals will be evaluated the hard way. It's not useful
5076 : * for cost_index(), for example, where the index machinery takes care of
5077 : * some of the quals. We assume baserestrictcost was previously set by
5078 : * set_baserel_size_estimates().
5079 : */
5080 : static void
5081 1078884 : get_restriction_qual_cost(PlannerInfo *root, RelOptInfo *baserel,
5082 : ParamPathInfo *param_info,
5083 : QualCost *qpqual_cost)
5084 : {
5085 1078884 : if (param_info)
5086 : {
5087 : /* Include costs of pushed-down clauses */
5088 242702 : cost_qual_eval(qpqual_cost, param_info->ppi_clauses, root);
5089 :
5090 242702 : qpqual_cost->startup += baserel->baserestrictcost.startup;
5091 242702 : qpqual_cost->per_tuple += baserel->baserestrictcost.per_tuple;
5092 : }
5093 : else
5094 836182 : *qpqual_cost = baserel->baserestrictcost;
5095 1078884 : }
5096 :
5097 :
5098 : /*
5099 : * compute_semi_anti_join_factors
5100 : * Estimate how much of the inner input a SEMI, ANTI, or inner_unique join
5101 : * can be expected to scan.
5102 : *
5103 : * In a hash or nestloop SEMI/ANTI join, the executor will stop scanning
5104 : * inner rows as soon as it finds a match to the current outer row.
5105 : * The same happens if we have detected the inner rel is unique.
5106 : * We should therefore adjust some of the cost components for this effect.
5107 : * This function computes some estimates needed for these adjustments.
5108 : * These estimates will be the same regardless of the particular paths used
5109 : * for the outer and inner relation, so we compute these once and then pass
5110 : * them to all the join cost estimation functions.
5111 : *
5112 : * Input parameters:
5113 : * joinrel: join relation under consideration
5114 : * outerrel: outer relation under consideration
5115 : * innerrel: inner relation under consideration
5116 : * jointype: if not JOIN_SEMI or JOIN_ANTI, we assume it's inner_unique
5117 : * sjinfo: SpecialJoinInfo relevant to this join
5118 : * restrictlist: join quals
5119 : * Output parameters:
5120 : * *semifactors is filled in (see pathnodes.h for field definitions)
5121 : */
5122 : void
5123 212866 : compute_semi_anti_join_factors(PlannerInfo *root,
5124 : RelOptInfo *joinrel,
5125 : RelOptInfo *outerrel,
5126 : RelOptInfo *innerrel,
5127 : JoinType jointype,
5128 : SpecialJoinInfo *sjinfo,
5129 : List *restrictlist,
5130 : SemiAntiJoinFactors *semifactors)
5131 : {
5132 : Selectivity jselec;
5133 : Selectivity nselec;
5134 : Selectivity avgmatch;
5135 : SpecialJoinInfo norm_sjinfo;
5136 : List *joinquals;
5137 : ListCell *l;
5138 :
5139 : /*
5140 : * In an ANTI join, we must ignore clauses that are "pushed down", since
5141 : * those won't affect the match logic. In a SEMI join, we do not
5142 : * distinguish joinquals from "pushed down" quals, so just use the whole
5143 : * restrictinfo list. For other outer join types, we should consider only
5144 : * non-pushed-down quals, so that this devolves to an IS_OUTER_JOIN check.
5145 : */
5146 212866 : if (IS_OUTER_JOIN(jointype))
5147 : {
5148 74608 : joinquals = NIL;
5149 163336 : foreach(l, restrictlist)
5150 : {
5151 88728 : RestrictInfo *rinfo = lfirst_node(RestrictInfo, l);
5152 :
5153 88728 : if (!RINFO_IS_PUSHED_DOWN(rinfo, joinrel->relids))
5154 83906 : joinquals = lappend(joinquals, rinfo);
5155 : }
5156 : }
5157 : else
5158 138258 : joinquals = restrictlist;
5159 :
5160 : /*
5161 : * Get the JOIN_SEMI or JOIN_ANTI selectivity of the join clauses.
5162 : */
5163 212866 : jselec = clauselist_selectivity(root,
5164 : joinquals,
5165 : 0,
5166 : (jointype == JOIN_ANTI) ? JOIN_ANTI : JOIN_SEMI,
5167 : sjinfo);
5168 :
5169 : /*
5170 : * Also get the normal inner-join selectivity of the join clauses.
5171 : */
5172 212866 : init_dummy_sjinfo(&norm_sjinfo, outerrel->relids, innerrel->relids);
5173 :
5174 212866 : nselec = clauselist_selectivity(root,
5175 : joinquals,
5176 : 0,
5177 : JOIN_INNER,
5178 : &norm_sjinfo);
5179 :
5180 : /* Avoid leaking a lot of ListCells */
5181 212866 : if (IS_OUTER_JOIN(jointype))
5182 74608 : list_free(joinquals);
5183 :
5184 : /*
5185 : * jselec can be interpreted as the fraction of outer-rel rows that have
5186 : * any matches (this is true for both SEMI and ANTI cases). And nselec is
5187 : * the fraction of the Cartesian product that matches. So, the average
5188 : * number of matches for each outer-rel row that has at least one match is
5189 : * nselec * inner_rows / jselec.
5190 : *
5191 : * Note: it is correct to use the inner rel's "rows" count here, even
5192 : * though we might later be considering a parameterized inner path with
5193 : * fewer rows. This is because we have included all the join clauses in
5194 : * the selectivity estimate.
5195 : */
5196 212866 : if (jselec > 0) /* protect against zero divide */
5197 : {
5198 212474 : avgmatch = nselec * innerrel->rows / jselec;
5199 : /* Clamp to sane range */
5200 212474 : avgmatch = Max(1.0, avgmatch);
5201 : }
5202 : else
5203 392 : avgmatch = 1.0;
5204 :
5205 212866 : semifactors->outer_match_frac = jselec;
5206 212866 : semifactors->match_count = avgmatch;
5207 212866 : }
5208 :
5209 : /*
5210 : * has_indexed_join_quals
5211 : * Check whether all the joinquals of a nestloop join are used as
5212 : * inner index quals.
5213 : *
5214 : * If the inner path of a SEMI/ANTI join is an indexscan (including bitmap
5215 : * indexscan) that uses all the joinquals as indexquals, we can assume that an
5216 : * unmatched outer tuple is cheap to process, whereas otherwise it's probably
5217 : * expensive.
5218 : */
5219 : static bool
5220 913154 : has_indexed_join_quals(NestPath *path)
5221 : {
5222 913154 : JoinPath *joinpath = &path->jpath;
5223 913154 : Relids joinrelids = joinpath->path.parent->relids;
5224 913154 : Path *innerpath = joinpath->innerjoinpath;
5225 : List *indexclauses;
5226 : bool found_one;
5227 : ListCell *lc;
5228 :
5229 : /* If join still has quals to evaluate, it's not fast */
5230 913154 : if (joinpath->joinrestrictinfo != NIL)
5231 648312 : return false;
5232 : /* Nor if the inner path isn't parameterized at all */
5233 264842 : if (innerpath->param_info == NULL)
5234 3300 : return false;
5235 :
5236 : /* Find the indexclauses list for the inner scan */
5237 261542 : switch (innerpath->pathtype)
5238 : {
5239 158728 : case T_IndexScan:
5240 : case T_IndexOnlyScan:
5241 158728 : indexclauses = ((IndexPath *) innerpath)->indexclauses;
5242 158728 : break;
5243 270 : case T_BitmapHeapScan:
5244 : {
5245 : /* Accept only a simple bitmap scan, not AND/OR cases */
5246 270 : Path *bmqual = ((BitmapHeapPath *) innerpath)->bitmapqual;
5247 :
5248 270 : if (IsA(bmqual, IndexPath))
5249 222 : indexclauses = ((IndexPath *) bmqual)->indexclauses;
5250 : else
5251 48 : return false;
5252 222 : break;
5253 : }
5254 102544 : default:
5255 :
5256 : /*
5257 : * If it's not a simple indexscan, it probably doesn't run quickly
5258 : * for zero rows out, even if it's a parameterized path using all
5259 : * the joinquals.
5260 : */
5261 102544 : return false;
5262 : }
5263 :
5264 : /*
5265 : * Examine the inner path's param clauses. Any that are from the outer
5266 : * path must be found in the indexclauses list, either exactly or in an
5267 : * equivalent form generated by equivclass.c. Also, we must find at least
5268 : * one such clause, else it's a clauseless join which isn't fast.
5269 : */
5270 158950 : found_one = false;
5271 313792 : foreach(lc, innerpath->param_info->ppi_clauses)
5272 : {
5273 163518 : RestrictInfo *rinfo = (RestrictInfo *) lfirst(lc);
5274 :
5275 163518 : if (join_clause_is_movable_into(rinfo,
5276 163518 : innerpath->parent->relids,
5277 : joinrelids))
5278 : {
5279 162966 : if (!is_redundant_with_indexclauses(rinfo, indexclauses))
5280 8676 : return false;
5281 154290 : found_one = true;
5282 : }
5283 : }
5284 150274 : return found_one;
5285 : }
5286 :
5287 :
5288 : /*
5289 : * approx_tuple_count
5290 : * Quick-and-dirty estimation of the number of join rows passing
5291 : * a set of qual conditions.
5292 : *
5293 : * The quals can be either an implicitly-ANDed list of boolean expressions,
5294 : * or a list of RestrictInfo nodes (typically the latter).
5295 : *
5296 : * We intentionally compute the selectivity under JOIN_INNER rules, even
5297 : * if it's some type of outer join. This is appropriate because we are
5298 : * trying to figure out how many tuples pass the initial merge or hash
5299 : * join step.
5300 : *
5301 : * This is quick-and-dirty because we bypass clauselist_selectivity, and
5302 : * simply multiply the independent clause selectivities together. Now
5303 : * clauselist_selectivity often can't do any better than that anyhow, but
5304 : * for some situations (such as range constraints) it is smarter. However,
5305 : * we can't effectively cache the results of clauselist_selectivity, whereas
5306 : * the individual clause selectivities can be and are cached.
5307 : *
5308 : * Since we are only using the results to estimate how many potential
5309 : * output tuples are generated and passed through qpqual checking, it
5310 : * seems OK to live with the approximation.
5311 : */
5312 : static double
5313 791340 : approx_tuple_count(PlannerInfo *root, JoinPath *path, List *quals)
5314 : {
5315 : double tuples;
5316 791340 : double outer_tuples = path->outerjoinpath->rows;
5317 791340 : double inner_tuples = path->innerjoinpath->rows;
5318 : SpecialJoinInfo sjinfo;
5319 791340 : Selectivity selec = 1.0;
5320 : ListCell *l;
5321 :
5322 : /*
5323 : * Make up a SpecialJoinInfo for JOIN_INNER semantics.
5324 : */
5325 791340 : init_dummy_sjinfo(&sjinfo, path->outerjoinpath->parent->relids,
5326 791340 : path->innerjoinpath->parent->relids);
5327 :
5328 : /* Get the approximate selectivity */
5329 1673196 : foreach(l, quals)
5330 : {
5331 881856 : Node *qual = (Node *) lfirst(l);
5332 :
5333 : /* Note that clause_selectivity will be able to cache its result */
5334 881856 : selec *= clause_selectivity(root, qual, 0, JOIN_INNER, &sjinfo);
5335 : }
5336 :
5337 : /* Apply it to the input relation sizes */
5338 791340 : tuples = selec * outer_tuples * inner_tuples;
5339 :
5340 791340 : return clamp_row_est(tuples);
5341 : }
5342 :
5343 :
5344 : /*
5345 : * set_baserel_size_estimates
5346 : * Set the size estimates for the given base relation.
5347 : *
5348 : * The rel's targetlist and restrictinfo list must have been constructed
5349 : * already, and rel->tuples must be set.
5350 : *
5351 : * We set the following fields of the rel node:
5352 : * rows: the estimated number of output tuples (after applying
5353 : * restriction clauses).
5354 : * width: the estimated average output tuple width in bytes.
5355 : * baserestrictcost: estimated cost of evaluating baserestrictinfo clauses.
5356 : */
5357 : void
5358 512176 : set_baserel_size_estimates(PlannerInfo *root, RelOptInfo *rel)
5359 : {
5360 : double nrows;
5361 :
5362 : /* Should only be applied to base relations */
5363 : Assert(rel->relid > 0);
5364 :
5365 1024322 : nrows = rel->tuples *
5366 512176 : clauselist_selectivity(root,
5367 : rel->baserestrictinfo,
5368 : 0,
5369 : JOIN_INNER,
5370 : NULL);
5371 :
5372 512146 : rel->rows = clamp_row_est(nrows);
5373 :
5374 512146 : cost_qual_eval(&rel->baserestrictcost, rel->baserestrictinfo, root);
5375 :
5376 512146 : set_rel_width(root, rel);
5377 512146 : }
5378 :
5379 : /*
5380 : * get_parameterized_baserel_size
5381 : * Make a size estimate for a parameterized scan of a base relation.
5382 : *
5383 : * 'param_clauses' lists the additional join clauses to be used.
5384 : *
5385 : * set_baserel_size_estimates must have been applied already.
5386 : */
5387 : double
5388 159668 : get_parameterized_baserel_size(PlannerInfo *root, RelOptInfo *rel,
5389 : List *param_clauses)
5390 : {
5391 : List *allclauses;
5392 : double nrows;
5393 :
5394 : /*
5395 : * Estimate the number of rows returned by the parameterized scan, knowing
5396 : * that it will apply all the extra join clauses as well as the rel's own
5397 : * restriction clauses. Note that we force the clauses to be treated as
5398 : * non-join clauses during selectivity estimation.
5399 : */
5400 159668 : allclauses = list_concat_copy(param_clauses, rel->baserestrictinfo);
5401 319336 : nrows = rel->tuples *
5402 159668 : clauselist_selectivity(root,
5403 : allclauses,
5404 159668 : rel->relid, /* do not use 0! */
5405 : JOIN_INNER,
5406 : NULL);
5407 159668 : nrows = clamp_row_est(nrows);
5408 : /* For safety, make sure result is not more than the base estimate */
5409 159668 : if (nrows > rel->rows)
5410 0 : nrows = rel->rows;
5411 159668 : return nrows;
5412 : }
5413 :
5414 : /*
5415 : * set_joinrel_size_estimates
5416 : * Set the size estimates for the given join relation.
5417 : *
5418 : * The rel's targetlist must have been constructed already, and a
5419 : * restriction clause list that matches the given component rels must
5420 : * be provided.
5421 : *
5422 : * Since there is more than one way to make a joinrel for more than two
5423 : * base relations, the results we get here could depend on which component
5424 : * rel pair is provided. In theory we should get the same answers no matter
5425 : * which pair is provided; in practice, since the selectivity estimation
5426 : * routines don't handle all cases equally well, we might not. But there's
5427 : * not much to be done about it. (Would it make sense to repeat the
5428 : * calculations for each pair of input rels that's encountered, and somehow
5429 : * average the results? Probably way more trouble than it's worth, and
5430 : * anyway we must keep the rowcount estimate the same for all paths for the
5431 : * joinrel.)
5432 : *
5433 : * We set only the rows field here. The reltarget field was already set by
5434 : * build_joinrel_tlist, and baserestrictcost is not used for join rels.
5435 : */
5436 : void
5437 253956 : set_joinrel_size_estimates(PlannerInfo *root, RelOptInfo *rel,
5438 : RelOptInfo *outer_rel,
5439 : RelOptInfo *inner_rel,
5440 : SpecialJoinInfo *sjinfo,
5441 : List *restrictlist)
5442 : {
5443 253956 : rel->rows = calc_joinrel_size_estimate(root,
5444 : rel,
5445 : outer_rel,
5446 : inner_rel,
5447 : outer_rel->rows,
5448 : inner_rel->rows,
5449 : sjinfo,
5450 : restrictlist);
5451 253956 : }
5452 :
5453 : /*
5454 : * get_parameterized_joinrel_size
5455 : * Make a size estimate for a parameterized scan of a join relation.
5456 : *
5457 : * 'rel' is the joinrel under consideration.
5458 : * 'outer_path', 'inner_path' are (probably also parameterized) Paths that
5459 : * produce the relations being joined.
5460 : * 'sjinfo' is any SpecialJoinInfo relevant to this join.
5461 : * 'restrict_clauses' lists the join clauses that need to be applied at the
5462 : * join node (including any movable clauses that were moved down to this join,
5463 : * and not including any movable clauses that were pushed down into the
5464 : * child paths).
5465 : *
5466 : * set_joinrel_size_estimates must have been applied already.
5467 : */
5468 : double
5469 10048 : get_parameterized_joinrel_size(PlannerInfo *root, RelOptInfo *rel,
5470 : Path *outer_path,
5471 : Path *inner_path,
5472 : SpecialJoinInfo *sjinfo,
5473 : List *restrict_clauses)
5474 : {
5475 : double nrows;
5476 :
5477 : /*
5478 : * Estimate the number of rows returned by the parameterized join as the
5479 : * sizes of the input paths times the selectivity of the clauses that have
5480 : * ended up at this join node.
5481 : *
5482 : * As with set_joinrel_size_estimates, the rowcount estimate could depend
5483 : * on the pair of input paths provided, though ideally we'd get the same
5484 : * estimate for any pair with the same parameterization.
5485 : */
5486 10048 : nrows = calc_joinrel_size_estimate(root,
5487 : rel,
5488 : outer_path->parent,
5489 : inner_path->parent,
5490 : outer_path->rows,
5491 : inner_path->rows,
5492 : sjinfo,
5493 : restrict_clauses);
5494 : /* For safety, make sure result is not more than the base estimate */
5495 10048 : if (nrows > rel->rows)
5496 12 : nrows = rel->rows;
5497 10048 : return nrows;
5498 : }
5499 :
5500 : /*
5501 : * calc_joinrel_size_estimate
5502 : * Workhorse for set_joinrel_size_estimates and
5503 : * get_parameterized_joinrel_size.
5504 : *
5505 : * outer_rel/inner_rel are the relations being joined, but they should be
5506 : * assumed to have sizes outer_rows/inner_rows; those numbers might be less
5507 : * than what rel->rows says, when we are considering parameterized paths.
5508 : */
5509 : static double
5510 264004 : calc_joinrel_size_estimate(PlannerInfo *root,
5511 : RelOptInfo *joinrel,
5512 : RelOptInfo *outer_rel,
5513 : RelOptInfo *inner_rel,
5514 : double outer_rows,
5515 : double inner_rows,
5516 : SpecialJoinInfo *sjinfo,
5517 : List *restrictlist)
5518 : {
5519 264004 : JoinType jointype = sjinfo->jointype;
5520 : Selectivity fkselec;
5521 : Selectivity jselec;
5522 : Selectivity pselec;
5523 : double nrows;
5524 :
5525 : /*
5526 : * Compute joinclause selectivity. Note that we are only considering
5527 : * clauses that become restriction clauses at this join level; we are not
5528 : * double-counting them because they were not considered in estimating the
5529 : * sizes of the component rels.
5530 : *
5531 : * First, see whether any of the joinclauses can be matched to known FK
5532 : * constraints. If so, drop those clauses from the restrictlist, and
5533 : * instead estimate their selectivity using FK semantics. (We do this
5534 : * without regard to whether said clauses are local or "pushed down".
5535 : * Probably, an FK-matching clause could never be seen as pushed down at
5536 : * an outer join, since it would be strict and hence would be grounds for
5537 : * join strength reduction.) fkselec gets the net selectivity for
5538 : * FK-matching clauses, or 1.0 if there are none.
5539 : */
5540 264004 : fkselec = get_foreign_key_join_selectivity(root,
5541 : outer_rel->relids,
5542 : inner_rel->relids,
5543 : sjinfo,
5544 : &restrictlist);
5545 :
5546 : /*
5547 : * For an outer join, we have to distinguish the selectivity of the join's
5548 : * own clauses (JOIN/ON conditions) from any clauses that were "pushed
5549 : * down". For inner joins we just count them all as joinclauses.
5550 : */
5551 264004 : if (IS_OUTER_JOIN(jointype))
5552 : {
5553 80016 : List *joinquals = NIL;
5554 80016 : List *pushedquals = NIL;
5555 : ListCell *l;
5556 :
5557 : /* Grovel through the clauses to separate into two lists */
5558 180308 : foreach(l, restrictlist)
5559 : {
5560 100292 : RestrictInfo *rinfo = lfirst_node(RestrictInfo, l);
5561 :
5562 100292 : if (RINFO_IS_PUSHED_DOWN(rinfo, joinrel->relids))
5563 4270 : pushedquals = lappend(pushedquals, rinfo);
5564 : else
5565 96022 : joinquals = lappend(joinquals, rinfo);
5566 : }
5567 :
5568 : /* Get the separate selectivities */
5569 80016 : jselec = clauselist_selectivity(root,
5570 : joinquals,
5571 : 0,
5572 : jointype,
5573 : sjinfo);
5574 80016 : pselec = clauselist_selectivity(root,
5575 : pushedquals,
5576 : 0,
5577 : jointype,
5578 : sjinfo);
5579 :
5580 : /* Avoid leaking a lot of ListCells */
5581 80016 : list_free(joinquals);
5582 80016 : list_free(pushedquals);
5583 : }
5584 : else
5585 : {
5586 183988 : jselec = clauselist_selectivity(root,
5587 : restrictlist,
5588 : 0,
5589 : jointype,
5590 : sjinfo);
5591 183988 : pselec = 0.0; /* not used, keep compiler quiet */
5592 : }
5593 :
5594 : /*
5595 : * Basically, we multiply size of Cartesian product by selectivity.
5596 : *
5597 : * If we are doing an outer join, take that into account: the joinqual
5598 : * selectivity has to be clamped using the knowledge that the output must
5599 : * be at least as large as the non-nullable input. However, any
5600 : * pushed-down quals are applied after the outer join, so their
5601 : * selectivity applies fully.
5602 : *
5603 : * For JOIN_SEMI and JOIN_ANTI, the selectivity is defined as the fraction
5604 : * of LHS rows that have matches, and we apply that straightforwardly.
5605 : */
5606 264004 : switch (jointype)
5607 : {
5608 175896 : case JOIN_INNER:
5609 175896 : nrows = outer_rows * inner_rows * fkselec * jselec;
5610 : /* pselec not used */
5611 175896 : break;
5612 73298 : case JOIN_LEFT:
5613 73298 : nrows = outer_rows * inner_rows * fkselec * jselec;
5614 73298 : if (nrows < outer_rows)
5615 28664 : nrows = outer_rows;
5616 73298 : nrows *= pselec;
5617 73298 : break;
5618 1720 : case JOIN_FULL:
5619 1720 : nrows = outer_rows * inner_rows * fkselec * jselec;
5620 1720 : if (nrows < outer_rows)
5621 1142 : nrows = outer_rows;
5622 1720 : if (nrows < inner_rows)
5623 120 : nrows = inner_rows;
5624 1720 : nrows *= pselec;
5625 1720 : break;
5626 8092 : case JOIN_SEMI:
5627 8092 : nrows = outer_rows * fkselec * jselec;
5628 : /* pselec not used */
5629 8092 : break;
5630 4998 : case JOIN_ANTI:
5631 4998 : nrows = outer_rows * (1.0 - fkselec * jselec);
5632 4998 : nrows *= pselec;
5633 4998 : break;
5634 0 : default:
5635 : /* other values not expected here */
5636 0 : elog(ERROR, "unrecognized join type: %d", (int) jointype);
5637 : nrows = 0; /* keep compiler quiet */
5638 : break;
5639 : }
5640 :
5641 264004 : return clamp_row_est(nrows);
5642 : }
5643 :
5644 : /*
5645 : * get_foreign_key_join_selectivity
5646 : * Estimate join selectivity for foreign-key-related clauses.
5647 : *
5648 : * Remove any clauses that can be matched to FK constraints from *restrictlist,
5649 : * and return a substitute estimate of their selectivity. 1.0 is returned
5650 : * when there are no such clauses.
5651 : *
5652 : * The reason for treating such clauses specially is that we can get better
5653 : * estimates this way than by relying on clauselist_selectivity(), especially
5654 : * for multi-column FKs where that function's assumption that the clauses are
5655 : * independent falls down badly. But even with single-column FKs, we may be
5656 : * able to get a better answer when the pg_statistic stats are missing or out
5657 : * of date.
5658 : */
5659 : static Selectivity
5660 264004 : get_foreign_key_join_selectivity(PlannerInfo *root,
5661 : Relids outer_relids,
5662 : Relids inner_relids,
5663 : SpecialJoinInfo *sjinfo,
5664 : List **restrictlist)
5665 : {
5666 264004 : Selectivity fkselec = 1.0;
5667 264004 : JoinType jointype = sjinfo->jointype;
5668 264004 : List *worklist = *restrictlist;
5669 : ListCell *lc;
5670 :
5671 : /* Consider each FK constraint that is known to match the query */
5672 265970 : foreach(lc, root->fkey_list)
5673 : {
5674 1966 : ForeignKeyOptInfo *fkinfo = (ForeignKeyOptInfo *) lfirst(lc);
5675 : bool ref_is_outer;
5676 : List *removedlist;
5677 : ListCell *cell;
5678 :
5679 : /*
5680 : * This FK is not relevant unless it connects a baserel on one side of
5681 : * this join to a baserel on the other side.
5682 : */
5683 3580 : if (bms_is_member(fkinfo->con_relid, outer_relids) &&
5684 1614 : bms_is_member(fkinfo->ref_relid, inner_relids))
5685 1440 : ref_is_outer = false;
5686 866 : else if (bms_is_member(fkinfo->ref_relid, outer_relids) &&
5687 340 : bms_is_member(fkinfo->con_relid, inner_relids))
5688 130 : ref_is_outer = true;
5689 : else
5690 396 : continue;
5691 :
5692 : /*
5693 : * If we're dealing with a semi/anti join, and the FK's referenced
5694 : * relation is on the outside, then knowledge of the FK doesn't help
5695 : * us figure out what we need to know (which is the fraction of outer
5696 : * rows that have matches). On the other hand, if the referenced rel
5697 : * is on the inside, then all outer rows must have matches in the
5698 : * referenced table (ignoring nulls). But any restriction or join
5699 : * clauses that filter that table will reduce the fraction of matches.
5700 : * We can account for restriction clauses, but it's too hard to guess
5701 : * how many table rows would get through a join that's inside the RHS.
5702 : * Hence, if either case applies, punt and ignore the FK.
5703 : */
5704 1570 : if ((jointype == JOIN_SEMI || jointype == JOIN_ANTI) &&
5705 1048 : (ref_is_outer || bms_membership(inner_relids) != BMS_SINGLETON))
5706 12 : continue;
5707 :
5708 : /*
5709 : * Modify the restrictlist by removing clauses that match the FK (and
5710 : * putting them into removedlist instead). It seems unsafe to modify
5711 : * the originally-passed List structure, so we make a shallow copy the
5712 : * first time through.
5713 : */
5714 1558 : if (worklist == *restrictlist)
5715 1334 : worklist = list_copy(worklist);
5716 :
5717 1558 : removedlist = NIL;
5718 3252 : foreach(cell, worklist)
5719 : {
5720 1694 : RestrictInfo *rinfo = (RestrictInfo *) lfirst(cell);
5721 1694 : bool remove_it = false;
5722 : int i;
5723 :
5724 : /* Drop this clause if it matches any column of the FK */
5725 2140 : for (i = 0; i < fkinfo->nkeys; i++)
5726 : {
5727 2110 : if (rinfo->parent_ec)
5728 : {
5729 : /*
5730 : * EC-derived clauses can only match by EC. It is okay to
5731 : * consider any clause derived from the same EC as
5732 : * matching the FK: even if equivclass.c chose to generate
5733 : * a clause equating some other pair of Vars, it could
5734 : * have generated one equating the FK's Vars. So for
5735 : * purposes of estimation, we can act as though it did so.
5736 : *
5737 : * Note: checking parent_ec is a bit of a cheat because
5738 : * there are EC-derived clauses that don't have parent_ec
5739 : * set; but such clauses must compare expressions that
5740 : * aren't just Vars, so they cannot match the FK anyway.
5741 : */
5742 304 : if (fkinfo->eclass[i] == rinfo->parent_ec)
5743 : {
5744 298 : remove_it = true;
5745 298 : break;
5746 : }
5747 : }
5748 : else
5749 : {
5750 : /*
5751 : * Otherwise, see if rinfo was previously matched to FK as
5752 : * a "loose" clause.
5753 : */
5754 1806 : if (list_member_ptr(fkinfo->rinfos[i], rinfo))
5755 : {
5756 1366 : remove_it = true;
5757 1366 : break;
5758 : }
5759 : }
5760 : }
5761 1694 : if (remove_it)
5762 : {
5763 1664 : worklist = foreach_delete_current(worklist, cell);
5764 1664 : removedlist = lappend(removedlist, rinfo);
5765 : }
5766 : }
5767 :
5768 : /*
5769 : * If we failed to remove all the matching clauses we expected to
5770 : * find, chicken out and ignore this FK; applying its selectivity
5771 : * might result in double-counting. Put any clauses we did manage to
5772 : * remove back into the worklist.
5773 : *
5774 : * Since the matching clauses are known not outerjoin-delayed, they
5775 : * would normally have appeared in the initial joinclause list. If we
5776 : * didn't find them, there are two possibilities:
5777 : *
5778 : * 1. If the FK match is based on an EC that is ec_has_const, it won't
5779 : * have generated any join clauses at all. We discount such ECs while
5780 : * checking to see if we have "all" the clauses. (Below, we'll adjust
5781 : * the selectivity estimate for this case.)
5782 : *
5783 : * 2. The clauses were matched to some other FK in a previous
5784 : * iteration of this loop, and thus removed from worklist. (A likely
5785 : * case is that two FKs are matched to the same EC; there will be only
5786 : * one EC-derived clause in the initial list, so the first FK will
5787 : * consume it.) Applying both FKs' selectivity independently risks
5788 : * underestimating the join size; in particular, this would undo one
5789 : * of the main things that ECs were invented for, namely to avoid
5790 : * double-counting the selectivity of redundant equality conditions.
5791 : * Later we might think of a reasonable way to combine the estimates,
5792 : * but for now, just punt, since this is a fairly uncommon situation.
5793 : */
5794 1558 : if (removedlist == NIL ||
5795 1272 : list_length(removedlist) !=
5796 1272 : (fkinfo->nmatched_ec - fkinfo->nconst_ec + fkinfo->nmatched_ri))
5797 : {
5798 286 : worklist = list_concat(worklist, removedlist);
5799 286 : continue;
5800 : }
5801 :
5802 : /*
5803 : * Finally we get to the payoff: estimate selectivity using the
5804 : * knowledge that each referencing row will match exactly one row in
5805 : * the referenced table.
5806 : *
5807 : * XXX that's not true in the presence of nulls in the referencing
5808 : * column(s), so in principle we should derate the estimate for those.
5809 : * However (1) if there are any strict restriction clauses for the
5810 : * referencing column(s) elsewhere in the query, derating here would
5811 : * be double-counting the null fraction, and (2) it's not very clear
5812 : * how to combine null fractions for multiple referencing columns. So
5813 : * we do nothing for now about correcting for nulls.
5814 : *
5815 : * XXX another point here is that if either side of an FK constraint
5816 : * is an inheritance parent, we estimate as though the constraint
5817 : * covers all its children as well. This is not an unreasonable
5818 : * assumption for a referencing table, ie the user probably applied
5819 : * identical constraints to all child tables (though perhaps we ought
5820 : * to check that). But it's not possible to have done that for a
5821 : * referenced table. Fortunately, precisely because that doesn't
5822 : * work, it is uncommon in practice to have an FK referencing a parent
5823 : * table. So, at least for now, disregard inheritance here.
5824 : */
5825 1272 : if (jointype == JOIN_SEMI || jointype == JOIN_ANTI)
5826 824 : {
5827 : /*
5828 : * For JOIN_SEMI and JOIN_ANTI, we only get here when the FK's
5829 : * referenced table is exactly the inside of the join. The join
5830 : * selectivity is defined as the fraction of LHS rows that have
5831 : * matches. The FK implies that every LHS row has a match *in the
5832 : * referenced table*; but any restriction clauses on it will
5833 : * reduce the number of matches. Hence we take the join
5834 : * selectivity as equal to the selectivity of the table's
5835 : * restriction clauses, which is rows / tuples; but we must guard
5836 : * against tuples == 0.
5837 : */
5838 824 : RelOptInfo *ref_rel = find_base_rel(root, fkinfo->ref_relid);
5839 824 : double ref_tuples = Max(ref_rel->tuples, 1.0);
5840 :
5841 824 : fkselec *= ref_rel->rows / ref_tuples;
5842 : }
5843 : else
5844 : {
5845 : /*
5846 : * Otherwise, selectivity is exactly 1/referenced-table-size; but
5847 : * guard against tuples == 0. Note we should use the raw table
5848 : * tuple count, not any estimate of its filtered or joined size.
5849 : */
5850 448 : RelOptInfo *ref_rel = find_base_rel(root, fkinfo->ref_relid);
5851 448 : double ref_tuples = Max(ref_rel->tuples, 1.0);
5852 :
5853 448 : fkselec *= 1.0 / ref_tuples;
5854 : }
5855 :
5856 : /*
5857 : * If any of the FK columns participated in ec_has_const ECs, then
5858 : * equivclass.c will have generated "var = const" restrictions for
5859 : * each side of the join, thus reducing the sizes of both input
5860 : * relations. Taking the fkselec at face value would amount to
5861 : * double-counting the selectivity of the constant restriction for the
5862 : * referencing Var. Hence, look for the restriction clause(s) that
5863 : * were applied to the referencing Var(s), and divide out their
5864 : * selectivity to correct for this.
5865 : */
5866 1272 : if (fkinfo->nconst_ec > 0)
5867 : {
5868 24 : for (int i = 0; i < fkinfo->nkeys; i++)
5869 : {
5870 18 : EquivalenceClass *ec = fkinfo->eclass[i];
5871 :
5872 18 : if (ec && ec->ec_has_const)
5873 : {
5874 6 : EquivalenceMember *em = fkinfo->fk_eclass_member[i];
5875 6 : RestrictInfo *rinfo = find_derived_clause_for_ec_member(root,
5876 : ec,
5877 : em);
5878 :
5879 6 : if (rinfo)
5880 : {
5881 : Selectivity s0;
5882 :
5883 6 : s0 = clause_selectivity(root,
5884 : (Node *) rinfo,
5885 : 0,
5886 : jointype,
5887 : sjinfo);
5888 6 : if (s0 > 0)
5889 6 : fkselec /= s0;
5890 : }
5891 : }
5892 : }
5893 : }
5894 : }
5895 :
5896 264004 : *restrictlist = worklist;
5897 264004 : CLAMP_PROBABILITY(fkselec);
5898 264004 : return fkselec;
5899 : }
5900 :
5901 : /*
5902 : * set_subquery_size_estimates
5903 : * Set the size estimates for a base relation that is a subquery.
5904 : *
5905 : * The rel's targetlist and restrictinfo list must have been constructed
5906 : * already, and the Paths for the subquery must have been completed.
5907 : * We look at the subquery's PlannerInfo to extract data.
5908 : *
5909 : * We set the same fields as set_baserel_size_estimates.
5910 : */
5911 : void
5912 33498 : set_subquery_size_estimates(PlannerInfo *root, RelOptInfo *rel)
5913 : {
5914 33498 : PlannerInfo *subroot = rel->subroot;
5915 : RelOptInfo *sub_final_rel;
5916 : ListCell *lc;
5917 :
5918 : /* Should only be applied to base relations that are subqueries */
5919 : Assert(rel->relid > 0);
5920 : Assert(planner_rt_fetch(rel->relid, root)->rtekind == RTE_SUBQUERY);
5921 :
5922 : /*
5923 : * Copy raw number of output rows from subquery. All of its paths should
5924 : * have the same output rowcount, so just look at cheapest-total.
5925 : */
5926 33498 : sub_final_rel = fetch_upper_rel(subroot, UPPERREL_FINAL, NULL);
5927 33498 : rel->tuples = sub_final_rel->cheapest_total_path->rows;
5928 :
5929 : /*
5930 : * Compute per-output-column width estimates by examining the subquery's
5931 : * targetlist. For any output that is a plain Var, get the width estimate
5932 : * that was made while planning the subquery. Otherwise, we leave it to
5933 : * set_rel_width to fill in a datatype-based default estimate.
5934 : */
5935 162566 : foreach(lc, subroot->parse->targetList)
5936 : {
5937 129068 : TargetEntry *te = lfirst_node(TargetEntry, lc);
5938 129068 : Node *texpr = (Node *) te->expr;
5939 129068 : int32 item_width = 0;
5940 :
5941 : /* junk columns aren't visible to upper query */
5942 129068 : if (te->resjunk)
5943 1368 : continue;
5944 :
5945 : /*
5946 : * The subquery could be an expansion of a view that's had columns
5947 : * added to it since the current query was parsed, so that there are
5948 : * non-junk tlist columns in it that don't correspond to any column
5949 : * visible at our query level. Ignore such columns.
5950 : */
5951 127700 : if (te->resno < rel->min_attr || te->resno > rel->max_attr)
5952 0 : continue;
5953 :
5954 : /*
5955 : * XXX This currently doesn't work for subqueries containing set
5956 : * operations, because the Vars in their tlists are bogus references
5957 : * to the first leaf subquery, which wouldn't give the right answer
5958 : * even if we could still get to its PlannerInfo.
5959 : *
5960 : * Also, the subquery could be an appendrel for which all branches are
5961 : * known empty due to constraint exclusion, in which case
5962 : * set_append_rel_pathlist will have left the attr_widths set to zero.
5963 : *
5964 : * In either case, we just leave the width estimate zero until
5965 : * set_rel_width fixes it.
5966 : */
5967 127700 : if (IsA(texpr, Var) &&
5968 62084 : subroot->parse->setOperations == NULL)
5969 : {
5970 60236 : Var *var = (Var *) texpr;
5971 60236 : RelOptInfo *subrel = find_base_rel(subroot, var->varno);
5972 :
5973 60236 : item_width = subrel->attr_widths[var->varattno - subrel->min_attr];
5974 : }
5975 127700 : rel->attr_widths[te->resno - rel->min_attr] = item_width;
5976 : }
5977 :
5978 : /* Now estimate number of output rows, etc */
5979 33498 : set_baserel_size_estimates(root, rel);
5980 33498 : }
5981 :
5982 : /*
5983 : * set_function_size_estimates
5984 : * Set the size estimates for a base relation that is a function call.
5985 : *
5986 : * The rel's targetlist and restrictinfo list must have been constructed
5987 : * already.
5988 : *
5989 : * We set the same fields as set_baserel_size_estimates.
5990 : */
5991 : void
5992 52258 : set_function_size_estimates(PlannerInfo *root, RelOptInfo *rel)
5993 : {
5994 : RangeTblEntry *rte;
5995 : ListCell *lc;
5996 :
5997 : /* Should only be applied to base relations that are functions */
5998 : Assert(rel->relid > 0);
5999 52258 : rte = planner_rt_fetch(rel->relid, root);
6000 : Assert(rte->rtekind == RTE_FUNCTION);
6001 :
6002 : /*
6003 : * Estimate number of rows the functions will return. The rowcount of the
6004 : * node is that of the largest function result.
6005 : */
6006 52258 : rel->tuples = 0;
6007 105016 : foreach(lc, rte->functions)
6008 : {
6009 52758 : RangeTblFunction *rtfunc = (RangeTblFunction *) lfirst(lc);
6010 52758 : double ntup = expression_returns_set_rows(root, rtfunc->funcexpr);
6011 :
6012 52758 : if (ntup > rel->tuples)
6013 52282 : rel->tuples = ntup;
6014 : }
6015 :
6016 : /* Now estimate number of output rows, etc */
6017 52258 : set_baserel_size_estimates(root, rel);
6018 52258 : }
6019 :
6020 : /*
6021 : * set_function_size_estimates
6022 : * Set the size estimates for a base relation that is a function call.
6023 : *
6024 : * The rel's targetlist and restrictinfo list must have been constructed
6025 : * already.
6026 : *
6027 : * We set the same fields as set_tablefunc_size_estimates.
6028 : */
6029 : void
6030 626 : set_tablefunc_size_estimates(PlannerInfo *root, RelOptInfo *rel)
6031 : {
6032 : /* Should only be applied to base relations that are functions */
6033 : Assert(rel->relid > 0);
6034 : Assert(planner_rt_fetch(rel->relid, root)->rtekind == RTE_TABLEFUNC);
6035 :
6036 626 : rel->tuples = 100;
6037 :
6038 : /* Now estimate number of output rows, etc */
6039 626 : set_baserel_size_estimates(root, rel);
6040 626 : }
6041 :
6042 : /*
6043 : * set_values_size_estimates
6044 : * Set the size estimates for a base relation that is a values list.
6045 : *
6046 : * The rel's targetlist and restrictinfo list must have been constructed
6047 : * already.
6048 : *
6049 : * We set the same fields as set_baserel_size_estimates.
6050 : */
6051 : void
6052 8286 : set_values_size_estimates(PlannerInfo *root, RelOptInfo *rel)
6053 : {
6054 : RangeTblEntry *rte;
6055 :
6056 : /* Should only be applied to base relations that are values lists */
6057 : Assert(rel->relid > 0);
6058 8286 : rte = planner_rt_fetch(rel->relid, root);
6059 : Assert(rte->rtekind == RTE_VALUES);
6060 :
6061 : /*
6062 : * Estimate number of rows the values list will return. We know this
6063 : * precisely based on the list length (well, barring set-returning
6064 : * functions in list items, but that's a refinement not catered for
6065 : * anywhere else either).
6066 : */
6067 8286 : rel->tuples = list_length(rte->values_lists);
6068 :
6069 : /* Now estimate number of output rows, etc */
6070 8286 : set_baserel_size_estimates(root, rel);
6071 8286 : }
6072 :
6073 : /*
6074 : * set_cte_size_estimates
6075 : * Set the size estimates for a base relation that is a CTE reference.
6076 : *
6077 : * The rel's targetlist and restrictinfo list must have been constructed
6078 : * already, and we need an estimate of the number of rows returned by the CTE
6079 : * (if a regular CTE) or the non-recursive term (if a self-reference).
6080 : *
6081 : * We set the same fields as set_baserel_size_estimates.
6082 : */
6083 : void
6084 5194 : set_cte_size_estimates(PlannerInfo *root, RelOptInfo *rel, double cte_rows)
6085 : {
6086 : RangeTblEntry *rte;
6087 :
6088 : /* Should only be applied to base relations that are CTE references */
6089 : Assert(rel->relid > 0);
6090 5194 : rte = planner_rt_fetch(rel->relid, root);
6091 : Assert(rte->rtekind == RTE_CTE);
6092 :
6093 5194 : if (rte->self_reference)
6094 : {
6095 : /*
6096 : * In a self-reference, we assume the average worktable size is a
6097 : * multiple of the nonrecursive term's size. The best multiplier will
6098 : * vary depending on query "fan-out", so make its value adjustable.
6099 : */
6100 934 : rel->tuples = clamp_row_est(recursive_worktable_factor * cte_rows);
6101 : }
6102 : else
6103 : {
6104 : /* Otherwise just believe the CTE's rowcount estimate */
6105 4260 : rel->tuples = cte_rows;
6106 : }
6107 :
6108 : /* Now estimate number of output rows, etc */
6109 5194 : set_baserel_size_estimates(root, rel);
6110 5194 : }
6111 :
6112 : /*
6113 : * set_namedtuplestore_size_estimates
6114 : * Set the size estimates for a base relation that is a tuplestore reference.
6115 : *
6116 : * The rel's targetlist and restrictinfo list must have been constructed
6117 : * already.
6118 : *
6119 : * We set the same fields as set_baserel_size_estimates.
6120 : */
6121 : void
6122 474 : set_namedtuplestore_size_estimates(PlannerInfo *root, RelOptInfo *rel)
6123 : {
6124 : RangeTblEntry *rte;
6125 :
6126 : /* Should only be applied to base relations that are tuplestore references */
6127 : Assert(rel->relid > 0);
6128 474 : rte = planner_rt_fetch(rel->relid, root);
6129 : Assert(rte->rtekind == RTE_NAMEDTUPLESTORE);
6130 :
6131 : /*
6132 : * Use the estimate provided by the code which is generating the named
6133 : * tuplestore. In some cases, the actual number might be available; in
6134 : * others the same plan will be re-used, so a "typical" value might be
6135 : * estimated and used.
6136 : */
6137 474 : rel->tuples = rte->enrtuples;
6138 474 : if (rel->tuples < 0)
6139 0 : rel->tuples = 1000;
6140 :
6141 : /* Now estimate number of output rows, etc */
6142 474 : set_baserel_size_estimates(root, rel);
6143 474 : }
6144 :
6145 : /*
6146 : * set_result_size_estimates
6147 : * Set the size estimates for an RTE_RESULT base relation
6148 : *
6149 : * The rel's targetlist and restrictinfo list must have been constructed
6150 : * already.
6151 : *
6152 : * We set the same fields as set_baserel_size_estimates.
6153 : */
6154 : void
6155 4208 : set_result_size_estimates(PlannerInfo *root, RelOptInfo *rel)
6156 : {
6157 : /* Should only be applied to RTE_RESULT base relations */
6158 : Assert(rel->relid > 0);
6159 : Assert(planner_rt_fetch(rel->relid, root)->rtekind == RTE_RESULT);
6160 :
6161 : /* RTE_RESULT always generates a single row, natively */
6162 4208 : rel->tuples = 1;
6163 :
6164 : /* Now estimate number of output rows, etc */
6165 4208 : set_baserel_size_estimates(root, rel);
6166 4208 : }
6167 :
6168 : /*
6169 : * set_foreign_size_estimates
6170 : * Set the size estimates for a base relation that is a foreign table.
6171 : *
6172 : * There is not a whole lot that we can do here; the foreign-data wrapper
6173 : * is responsible for producing useful estimates. We can do a decent job
6174 : * of estimating baserestrictcost, so we set that, and we also set up width
6175 : * using what will be purely datatype-driven estimates from the targetlist.
6176 : * There is no way to do anything sane with the rows value, so we just put
6177 : * a default estimate and hope that the wrapper can improve on it. The
6178 : * wrapper's GetForeignRelSize function will be called momentarily.
6179 : *
6180 : * The rel's targetlist and restrictinfo list must have been constructed
6181 : * already.
6182 : */
6183 : void
6184 2464 : set_foreign_size_estimates(PlannerInfo *root, RelOptInfo *rel)
6185 : {
6186 : /* Should only be applied to base relations */
6187 : Assert(rel->relid > 0);
6188 :
6189 2464 : rel->rows = 1000; /* entirely bogus default estimate */
6190 :
6191 2464 : cost_qual_eval(&rel->baserestrictcost, rel->baserestrictinfo, root);
6192 :
6193 2464 : set_rel_width(root, rel);
6194 2464 : }
6195 :
6196 :
6197 : /*
6198 : * set_rel_width
6199 : * Set the estimated output width of a base relation.
6200 : *
6201 : * The estimated output width is the sum of the per-attribute width estimates
6202 : * for the actually-referenced columns, plus any PHVs or other expressions
6203 : * that have to be calculated at this relation. This is the amount of data
6204 : * we'd need to pass upwards in case of a sort, hash, etc.
6205 : *
6206 : * This function also sets reltarget->cost, so it's a bit misnamed now.
6207 : *
6208 : * NB: this works best on plain relations because it prefers to look at
6209 : * real Vars. For subqueries, set_subquery_size_estimates will already have
6210 : * copied up whatever per-column estimates were made within the subquery,
6211 : * and for other types of rels there isn't much we can do anyway. We fall
6212 : * back on (fairly stupid) datatype-based width estimates if we can't get
6213 : * any better number.
6214 : *
6215 : * The per-attribute width estimates are cached for possible re-use while
6216 : * building join relations or post-scan/join pathtargets.
6217 : */
6218 : static void
6219 514610 : set_rel_width(PlannerInfo *root, RelOptInfo *rel)
6220 : {
6221 514610 : Oid reloid = planner_rt_fetch(rel->relid, root)->relid;
6222 514610 : int64 tuple_width = 0;
6223 514610 : bool have_wholerow_var = false;
6224 : ListCell *lc;
6225 :
6226 : /* Vars are assumed to have cost zero, but other exprs do not */
6227 514610 : rel->reltarget->cost.startup = 0;
6228 514610 : rel->reltarget->cost.per_tuple = 0;
6229 :
6230 1853024 : foreach(lc, rel->reltarget->exprs)
6231 : {
6232 1338414 : Node *node = (Node *) lfirst(lc);
6233 :
6234 : /*
6235 : * Ordinarily, a Var in a rel's targetlist must belong to that rel;
6236 : * but there are corner cases involving LATERAL references where that
6237 : * isn't so. If the Var has the wrong varno, fall through to the
6238 : * generic case (it doesn't seem worth the trouble to be any smarter).
6239 : */
6240 1338414 : if (IsA(node, Var) &&
6241 1314462 : ((Var *) node)->varno == rel->relid)
6242 360872 : {
6243 1314396 : Var *var = (Var *) node;
6244 : int ndx;
6245 : int32 item_width;
6246 :
6247 : Assert(var->varattno >= rel->min_attr);
6248 : Assert(var->varattno <= rel->max_attr);
6249 :
6250 1314396 : ndx = var->varattno - rel->min_attr;
6251 :
6252 : /*
6253 : * If it's a whole-row Var, we'll deal with it below after we have
6254 : * already cached as many attr widths as possible.
6255 : */
6256 1314396 : if (var->varattno == 0)
6257 : {
6258 3050 : have_wholerow_var = true;
6259 3050 : continue;
6260 : }
6261 :
6262 : /*
6263 : * The width may have been cached already (especially if it's a
6264 : * subquery), so don't duplicate effort.
6265 : */
6266 1311346 : if (rel->attr_widths[ndx] > 0)
6267 : {
6268 260058 : tuple_width += rel->attr_widths[ndx];
6269 260058 : continue;
6270 : }
6271 :
6272 : /* Try to get column width from statistics */
6273 1051288 : if (reloid != InvalidOid && var->varattno > 0)
6274 : {
6275 829102 : item_width = get_attavgwidth(reloid, var->varattno);
6276 829102 : if (item_width > 0)
6277 : {
6278 690416 : rel->attr_widths[ndx] = item_width;
6279 690416 : tuple_width += item_width;
6280 690416 : continue;
6281 : }
6282 : }
6283 :
6284 : /*
6285 : * Not a plain relation, or can't find statistics for it. Estimate
6286 : * using just the type info.
6287 : */
6288 360872 : item_width = get_typavgwidth(var->vartype, var->vartypmod);
6289 : Assert(item_width > 0);
6290 360872 : rel->attr_widths[ndx] = item_width;
6291 360872 : tuple_width += item_width;
6292 : }
6293 24018 : else if (IsA(node, PlaceHolderVar))
6294 : {
6295 : /*
6296 : * We will need to evaluate the PHV's contained expression while
6297 : * scanning this rel, so be sure to include it in reltarget->cost.
6298 : */
6299 2008 : PlaceHolderVar *phv = (PlaceHolderVar *) node;
6300 2008 : PlaceHolderInfo *phinfo = find_placeholder_info(root, phv);
6301 : QualCost cost;
6302 :
6303 2008 : tuple_width += phinfo->ph_width;
6304 2008 : cost_qual_eval_node(&cost, (Node *) phv->phexpr, root);
6305 2008 : rel->reltarget->cost.startup += cost.startup;
6306 2008 : rel->reltarget->cost.per_tuple += cost.per_tuple;
6307 : }
6308 : else
6309 : {
6310 : /*
6311 : * We could be looking at an expression pulled up from a subquery,
6312 : * or a ROW() representing a whole-row child Var, etc. Do what we
6313 : * can using the expression type information.
6314 : */
6315 : int32 item_width;
6316 : QualCost cost;
6317 :
6318 22010 : item_width = get_typavgwidth(exprType(node), exprTypmod(node));
6319 : Assert(item_width > 0);
6320 22010 : tuple_width += item_width;
6321 : /* Not entirely clear if we need to account for cost, but do so */
6322 22010 : cost_qual_eval_node(&cost, node, root);
6323 22010 : rel->reltarget->cost.startup += cost.startup;
6324 22010 : rel->reltarget->cost.per_tuple += cost.per_tuple;
6325 : }
6326 : }
6327 :
6328 : /*
6329 : * If we have a whole-row reference, estimate its width as the sum of
6330 : * per-column widths plus heap tuple header overhead.
6331 : */
6332 514610 : if (have_wholerow_var)
6333 : {
6334 3050 : int64 wholerow_width = MAXALIGN(SizeofHeapTupleHeader);
6335 :
6336 3050 : if (reloid != InvalidOid)
6337 : {
6338 : /* Real relation, so estimate true tuple width */
6339 2392 : wholerow_width += get_relation_data_width(reloid,
6340 2392 : rel->attr_widths - rel->min_attr);
6341 : }
6342 : else
6343 : {
6344 : /* Do what we can with info for a phony rel */
6345 : AttrNumber i;
6346 :
6347 1794 : for (i = 1; i <= rel->max_attr; i++)
6348 1136 : wholerow_width += rel->attr_widths[i - rel->min_attr];
6349 : }
6350 :
6351 3050 : rel->attr_widths[0 - rel->min_attr] = clamp_width_est(wholerow_width);
6352 :
6353 : /*
6354 : * Include the whole-row Var as part of the output tuple. Yes, that
6355 : * really is what happens at runtime.
6356 : */
6357 3050 : tuple_width += wholerow_width;
6358 : }
6359 :
6360 514610 : rel->reltarget->width = clamp_width_est(tuple_width);
6361 514610 : }
6362 :
6363 : /*
6364 : * set_pathtarget_cost_width
6365 : * Set the estimated eval cost and output width of a PathTarget tlist.
6366 : *
6367 : * As a notational convenience, returns the same PathTarget pointer passed in.
6368 : *
6369 : * Most, though not quite all, uses of this function occur after we've run
6370 : * set_rel_width() for base relations; so we can usually obtain cached width
6371 : * estimates for Vars. If we can't, fall back on datatype-based width
6372 : * estimates. Present early-planning uses of PathTargets don't need accurate
6373 : * widths badly enough to justify going to the catalogs for better data.
6374 : */
6375 : PathTarget *
6376 614642 : set_pathtarget_cost_width(PlannerInfo *root, PathTarget *target)
6377 : {
6378 614642 : int64 tuple_width = 0;
6379 : ListCell *lc;
6380 :
6381 : /* Vars are assumed to have cost zero, but other exprs do not */
6382 614642 : target->cost.startup = 0;
6383 614642 : target->cost.per_tuple = 0;
6384 :
6385 2138030 : foreach(lc, target->exprs)
6386 : {
6387 1523388 : Node *node = (Node *) lfirst(lc);
6388 :
6389 1523388 : tuple_width += get_expr_width(root, node);
6390 :
6391 : /* For non-Vars, account for evaluation cost */
6392 1523388 : if (!IsA(node, Var))
6393 : {
6394 : QualCost cost;
6395 :
6396 651394 : cost_qual_eval_node(&cost, node, root);
6397 651394 : target->cost.startup += cost.startup;
6398 651394 : target->cost.per_tuple += cost.per_tuple;
6399 : }
6400 : }
6401 :
6402 614642 : target->width = clamp_width_est(tuple_width);
6403 :
6404 614642 : return target;
6405 : }
6406 :
6407 : /*
6408 : * get_expr_width
6409 : * Estimate the width of the given expr attempting to use the width
6410 : * cached in a Var's owning RelOptInfo, else fallback on the type's
6411 : * average width when unable to or when the given Node is not a Var.
6412 : */
6413 : static int32
6414 1853186 : get_expr_width(PlannerInfo *root, const Node *expr)
6415 : {
6416 : int32 width;
6417 :
6418 1853186 : if (IsA(expr, Var))
6419 : {
6420 1189060 : const Var *var = (const Var *) expr;
6421 :
6422 : /* We should not see any upper-level Vars here */
6423 : Assert(var->varlevelsup == 0);
6424 :
6425 : /* Try to get data from RelOptInfo cache */
6426 1189060 : if (!IS_SPECIAL_VARNO(var->varno) &&
6427 1183222 : var->varno < root->simple_rel_array_size)
6428 : {
6429 1183222 : RelOptInfo *rel = root->simple_rel_array[var->varno];
6430 :
6431 1183222 : if (rel != NULL &&
6432 1165538 : var->varattno >= rel->min_attr &&
6433 1165538 : var->varattno <= rel->max_attr)
6434 : {
6435 1165538 : int ndx = var->varattno - rel->min_attr;
6436 :
6437 1165538 : if (rel->attr_widths[ndx] > 0)
6438 1133584 : return rel->attr_widths[ndx];
6439 : }
6440 : }
6441 :
6442 : /*
6443 : * No cached data available, so estimate using just the type info.
6444 : */
6445 55476 : width = get_typavgwidth(var->vartype, var->vartypmod);
6446 : Assert(width > 0);
6447 :
6448 55476 : return width;
6449 : }
6450 :
6451 664126 : width = get_typavgwidth(exprType(expr), exprTypmod(expr));
6452 : Assert(width > 0);
6453 664126 : return width;
6454 : }
6455 :
6456 : /*
6457 : * relation_byte_size
6458 : * Estimate the storage space in bytes for a given number of tuples
6459 : * of a given width (size in bytes).
6460 : */
6461 : static double
6462 5040228 : relation_byte_size(double tuples, int width)
6463 : {
6464 5040228 : return tuples * (MAXALIGN(width) + MAXALIGN(SizeofHeapTupleHeader));
6465 : }
6466 :
6467 : /*
6468 : * page_size
6469 : * Returns an estimate of the number of pages covered by a given
6470 : * number of tuples of a given width (size in bytes).
6471 : */
6472 : static double
6473 9328 : page_size(double tuples, int width)
6474 : {
6475 9328 : return ceil(relation_byte_size(tuples, width) / BLCKSZ);
6476 : }
6477 :
6478 : /*
6479 : * Estimate the fraction of the work that each worker will do given the
6480 : * number of workers budgeted for the path.
6481 : */
6482 : static double
6483 463068 : get_parallel_divisor(Path *path)
6484 : {
6485 463068 : double parallel_divisor = path->parallel_workers;
6486 :
6487 : /*
6488 : * Early experience with parallel query suggests that when there is only
6489 : * one worker, the leader often makes a very substantial contribution to
6490 : * executing the parallel portion of the plan, but as more workers are
6491 : * added, it does less and less, because it's busy reading tuples from the
6492 : * workers and doing whatever non-parallel post-processing is needed. By
6493 : * the time we reach 4 workers, the leader no longer makes a meaningful
6494 : * contribution. Thus, for now, estimate that the leader spends 30% of
6495 : * its time servicing each worker, and the remainder executing the
6496 : * parallel plan.
6497 : */
6498 463068 : if (parallel_leader_participation)
6499 : {
6500 : double leader_contribution;
6501 :
6502 461766 : leader_contribution = 1.0 - (0.3 * path->parallel_workers);
6503 461766 : if (leader_contribution > 0)
6504 459396 : parallel_divisor += leader_contribution;
6505 : }
6506 :
6507 463068 : return parallel_divisor;
6508 : }
6509 :
6510 : /*
6511 : * compute_bitmap_pages
6512 : * Estimate number of pages fetched from heap in a bitmap heap scan.
6513 : *
6514 : * 'baserel' is the relation to be scanned
6515 : * 'bitmapqual' is a tree of IndexPaths, BitmapAndPaths, and BitmapOrPaths
6516 : * 'loop_count' is the number of repetitions of the indexscan to factor into
6517 : * estimates of caching behavior
6518 : *
6519 : * If cost_p isn't NULL, the indexTotalCost estimate is returned in *cost_p.
6520 : * If tuples_p isn't NULL, the tuples_fetched estimate is returned in *tuples_p.
6521 : */
6522 : double
6523 687152 : compute_bitmap_pages(PlannerInfo *root, RelOptInfo *baserel,
6524 : Path *bitmapqual, double loop_count,
6525 : Cost *cost_p, double *tuples_p)
6526 : {
6527 : Cost indexTotalCost;
6528 : Selectivity indexSelectivity;
6529 : double T;
6530 : double pages_fetched;
6531 : double tuples_fetched;
6532 : double heap_pages;
6533 : double maxentries;
6534 :
6535 : /*
6536 : * Fetch total cost of obtaining the bitmap, as well as its total
6537 : * selectivity.
6538 : */
6539 687152 : cost_bitmap_tree_node(bitmapqual, &indexTotalCost, &indexSelectivity);
6540 :
6541 : /*
6542 : * Estimate number of main-table pages fetched.
6543 : */
6544 687152 : tuples_fetched = clamp_row_est(indexSelectivity * baserel->tuples);
6545 :
6546 687152 : T = (baserel->pages > 1) ? (double) baserel->pages : 1.0;
6547 :
6548 : /*
6549 : * For a single scan, the number of heap pages that need to be fetched is
6550 : * the same as the Mackert and Lohman formula for the case T <= b (ie, no
6551 : * re-reads needed).
6552 : */
6553 687152 : pages_fetched = (2.0 * T * tuples_fetched) / (2.0 * T + tuples_fetched);
6554 :
6555 : /*
6556 : * Calculate the number of pages fetched from the heap. Then based on
6557 : * current work_mem estimate get the estimated maxentries in the bitmap.
6558 : * (Note that we always do this calculation based on the number of pages
6559 : * that would be fetched in a single iteration, even if loop_count > 1.
6560 : * That's correct, because only that number of entries will be stored in
6561 : * the bitmap at one time.)
6562 : */
6563 687152 : heap_pages = Min(pages_fetched, baserel->pages);
6564 687152 : maxentries = tbm_calculate_entries(work_mem * (Size) 1024);
6565 :
6566 687152 : if (loop_count > 1)
6567 : {
6568 : /*
6569 : * For repeated bitmap scans, scale up the number of tuples fetched in
6570 : * the Mackert and Lohman formula by the number of scans, so that we
6571 : * estimate the number of pages fetched by all the scans. Then
6572 : * pro-rate for one scan.
6573 : */
6574 145024 : pages_fetched = index_pages_fetched(tuples_fetched * loop_count,
6575 : baserel->pages,
6576 : get_indexpath_pages(bitmapqual),
6577 : root);
6578 145024 : pages_fetched /= loop_count;
6579 : }
6580 :
6581 687152 : if (pages_fetched >= T)
6582 65818 : pages_fetched = T;
6583 : else
6584 621334 : pages_fetched = ceil(pages_fetched);
6585 :
6586 687152 : if (maxentries < heap_pages)
6587 : {
6588 : double exact_pages;
6589 : double lossy_pages;
6590 :
6591 : /*
6592 : * Crude approximation of the number of lossy pages. Because of the
6593 : * way tbm_lossify() is coded, the number of lossy pages increases
6594 : * very sharply as soon as we run short of memory; this formula has
6595 : * that property and seems to perform adequately in testing, but it's
6596 : * possible we could do better somehow.
6597 : */
6598 18 : lossy_pages = Max(0, heap_pages - maxentries / 2);
6599 18 : exact_pages = heap_pages - lossy_pages;
6600 :
6601 : /*
6602 : * If there are lossy pages then recompute the number of tuples
6603 : * processed by the bitmap heap node. We assume here that the chance
6604 : * of a given tuple coming from an exact page is the same as the
6605 : * chance that a given page is exact. This might not be true, but
6606 : * it's not clear how we can do any better.
6607 : */
6608 18 : if (lossy_pages > 0)
6609 : tuples_fetched =
6610 18 : clamp_row_est(indexSelectivity *
6611 18 : (exact_pages / heap_pages) * baserel->tuples +
6612 18 : (lossy_pages / heap_pages) * baserel->tuples);
6613 : }
6614 :
6615 687152 : if (cost_p)
6616 542954 : *cost_p = indexTotalCost;
6617 687152 : if (tuples_p)
6618 542954 : *tuples_p = tuples_fetched;
6619 :
6620 687152 : return pages_fetched;
6621 : }
6622 :
6623 : /*
6624 : * compute_gather_rows
6625 : * Estimate number of rows for gather (merge) nodes.
6626 : *
6627 : * In a parallel plan, each worker's row estimate is determined by dividing the
6628 : * total number of rows by parallel_divisor, which accounts for the leader's
6629 : * contribution in addition to the number of workers. Accordingly, when
6630 : * estimating the number of rows for gather (merge) nodes, we multiply the rows
6631 : * per worker by the same parallel_divisor to undo the division.
6632 : */
6633 : double
6634 42324 : compute_gather_rows(Path *path)
6635 : {
6636 : Assert(path->parallel_workers > 0);
6637 :
6638 42324 : return clamp_row_est(path->rows * get_parallel_divisor(path));
6639 : }
|
