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 9008926 : 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 9008926 : if (nrows > MAXIMUM_ROWCOUNT || isnan(nrows))
222 0 : nrows = MAXIMUM_ROWCOUNT;
223 9008926 : else if (nrows <= 1.0)
224 3213630 : nrows = 1.0;
225 : else
226 5795296 : nrows = rint(nrows);
227 :
228 9008926 : 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 1895846 : 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 1895846 : 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 1895846 : return (int32) tuple_width;
258 : }
259 :
260 : /*
261 : * clamp_cardinality_to_long
262 : * Cast a Cardinality value to a sane long value.
263 : */
264 : long
265 45860 : clamp_cardinality_to_long(Cardinality x)
266 : {
267 : /*
268 : * Just for paranoia's sake, ensure we do something sane with negative or
269 : * NaN values.
270 : */
271 45860 : if (isnan(x))
272 0 : return LONG_MAX;
273 45860 : if (x <= 0)
274 556 : return 0;
275 :
276 : /*
277 : * If "long" is 64 bits, then LONG_MAX cannot be represented exactly as a
278 : * double. Casting it to double and back may well result in overflow due
279 : * to rounding, so avoid doing that. We trust that any double value that
280 : * compares strictly less than "(double) LONG_MAX" will cast to a
281 : * representable "long" value.
282 : */
283 45304 : return (x < (double) LONG_MAX) ? (long) x : LONG_MAX;
284 : }
285 :
286 :
287 : /*
288 : * cost_seqscan
289 : * Determines and returns the cost of scanning a relation sequentially.
290 : *
291 : * 'baserel' is the relation to be scanned
292 : * 'param_info' is the ParamPathInfo if this is a parameterized path, else NULL
293 : */
294 : void
295 427500 : cost_seqscan(Path *path, PlannerInfo *root,
296 : RelOptInfo *baserel, ParamPathInfo *param_info)
297 : {
298 427500 : Cost startup_cost = 0;
299 : Cost cpu_run_cost;
300 : Cost disk_run_cost;
301 : double spc_seq_page_cost;
302 : QualCost qpqual_cost;
303 : Cost cpu_per_tuple;
304 :
305 : /* Should only be applied to base relations */
306 : Assert(baserel->relid > 0);
307 : Assert(baserel->rtekind == RTE_RELATION);
308 :
309 : /* Mark the path with the correct row estimate */
310 427500 : if (param_info)
311 840 : path->rows = param_info->ppi_rows;
312 : else
313 426660 : path->rows = baserel->rows;
314 :
315 : /* fetch estimated page cost for tablespace containing table */
316 427500 : get_tablespace_page_costs(baserel->reltablespace,
317 : NULL,
318 : &spc_seq_page_cost);
319 :
320 : /*
321 : * disk costs
322 : */
323 427500 : disk_run_cost = spc_seq_page_cost * baserel->pages;
324 :
325 : /* CPU costs */
326 427500 : get_restriction_qual_cost(root, baserel, param_info, &qpqual_cost);
327 :
328 427500 : startup_cost += qpqual_cost.startup;
329 427500 : cpu_per_tuple = cpu_tuple_cost + qpqual_cost.per_tuple;
330 427500 : cpu_run_cost = cpu_per_tuple * baserel->tuples;
331 : /* tlist eval costs are paid per output row, not per tuple scanned */
332 427500 : startup_cost += path->pathtarget->cost.startup;
333 427500 : cpu_run_cost += path->pathtarget->cost.per_tuple * path->rows;
334 :
335 : /* Adjust costing for parallelism, if used. */
336 427500 : if (path->parallel_workers > 0)
337 : {
338 26156 : double parallel_divisor = get_parallel_divisor(path);
339 :
340 : /* The CPU cost is divided among all the workers. */
341 26156 : cpu_run_cost /= parallel_divisor;
342 :
343 : /*
344 : * It may be possible to amortize some of the I/O cost, but probably
345 : * not very much, because most operating systems already do aggressive
346 : * prefetching. For now, we assume that the disk run cost can't be
347 : * amortized at all.
348 : */
349 :
350 : /*
351 : * In the case of a parallel plan, the row count needs to represent
352 : * the number of tuples processed per worker.
353 : */
354 26156 : path->rows = clamp_row_est(path->rows / parallel_divisor);
355 : }
356 :
357 427500 : path->disabled_nodes = enable_seqscan ? 0 : 1;
358 427500 : path->startup_cost = startup_cost;
359 427500 : path->total_cost = startup_cost + cpu_run_cost + disk_run_cost;
360 427500 : }
361 :
362 : /*
363 : * cost_samplescan
364 : * Determines and returns the cost of scanning a relation using sampling.
365 : *
366 : * 'baserel' is the relation to be scanned
367 : * 'param_info' is the ParamPathInfo if this is a parameterized path, else NULL
368 : */
369 : void
370 306 : cost_samplescan(Path *path, PlannerInfo *root,
371 : RelOptInfo *baserel, ParamPathInfo *param_info)
372 : {
373 306 : Cost startup_cost = 0;
374 306 : Cost run_cost = 0;
375 : RangeTblEntry *rte;
376 : TableSampleClause *tsc;
377 : TsmRoutine *tsm;
378 : double spc_seq_page_cost,
379 : spc_random_page_cost,
380 : spc_page_cost;
381 : QualCost qpqual_cost;
382 : Cost cpu_per_tuple;
383 :
384 : /* Should only be applied to base relations with tablesample clauses */
385 : Assert(baserel->relid > 0);
386 306 : rte = planner_rt_fetch(baserel->relid, root);
387 : Assert(rte->rtekind == RTE_RELATION);
388 306 : tsc = rte->tablesample;
389 : Assert(tsc != NULL);
390 306 : tsm = GetTsmRoutine(tsc->tsmhandler);
391 :
392 : /* Mark the path with the correct row estimate */
393 306 : if (param_info)
394 72 : path->rows = param_info->ppi_rows;
395 : else
396 234 : path->rows = baserel->rows;
397 :
398 : /* fetch estimated page cost for tablespace containing table */
399 306 : get_tablespace_page_costs(baserel->reltablespace,
400 : &spc_random_page_cost,
401 : &spc_seq_page_cost);
402 :
403 : /* if NextSampleBlock is used, assume random access, else sequential */
404 612 : spc_page_cost = (tsm->NextSampleBlock != NULL) ?
405 306 : spc_random_page_cost : spc_seq_page_cost;
406 :
407 : /*
408 : * disk costs (recall that baserel->pages has already been set to the
409 : * number of pages the sampling method will visit)
410 : */
411 306 : run_cost += spc_page_cost * baserel->pages;
412 :
413 : /*
414 : * CPU costs (recall that baserel->tuples has already been set to the
415 : * number of tuples the sampling method will select). Note that we ignore
416 : * execution cost of the TABLESAMPLE parameter expressions; they will be
417 : * evaluated only once per scan, and in most usages they'll likely be
418 : * simple constants anyway. We also don't charge anything for the
419 : * calculations the sampling method might do internally.
420 : */
421 306 : get_restriction_qual_cost(root, baserel, param_info, &qpqual_cost);
422 :
423 306 : startup_cost += qpqual_cost.startup;
424 306 : cpu_per_tuple = cpu_tuple_cost + qpqual_cost.per_tuple;
425 306 : run_cost += cpu_per_tuple * baserel->tuples;
426 : /* tlist eval costs are paid per output row, not per tuple scanned */
427 306 : startup_cost += path->pathtarget->cost.startup;
428 306 : run_cost += path->pathtarget->cost.per_tuple * path->rows;
429 :
430 306 : path->disabled_nodes = 0;
431 306 : path->startup_cost = startup_cost;
432 306 : path->total_cost = startup_cost + run_cost;
433 306 : }
434 :
435 : /*
436 : * cost_gather
437 : * Determines and returns the cost of gather path.
438 : *
439 : * 'rel' is the relation to be operated upon
440 : * 'param_info' is the ParamPathInfo if this is a parameterized path, else NULL
441 : * 'rows' may be used to point to a row estimate; if non-NULL, it overrides
442 : * both 'rel' and 'param_info'. This is useful when the path doesn't exactly
443 : * correspond to any particular RelOptInfo.
444 : */
445 : void
446 19134 : cost_gather(GatherPath *path, PlannerInfo *root,
447 : RelOptInfo *rel, ParamPathInfo *param_info,
448 : double *rows)
449 : {
450 19134 : Cost startup_cost = 0;
451 19134 : Cost run_cost = 0;
452 :
453 : /* Mark the path with the correct row estimate */
454 19134 : if (rows)
455 1752 : path->path.rows = *rows;
456 17382 : else if (param_info)
457 0 : path->path.rows = param_info->ppi_rows;
458 : else
459 17382 : path->path.rows = rel->rows;
460 :
461 19134 : startup_cost = path->subpath->startup_cost;
462 :
463 19134 : run_cost = path->subpath->total_cost - path->subpath->startup_cost;
464 :
465 : /* Parallel setup and communication cost. */
466 19134 : startup_cost += parallel_setup_cost;
467 19134 : run_cost += parallel_tuple_cost * path->path.rows;
468 :
469 19134 : path->path.disabled_nodes = path->subpath->disabled_nodes;
470 19134 : path->path.startup_cost = startup_cost;
471 19134 : path->path.total_cost = (startup_cost + run_cost);
472 19134 : }
473 :
474 : /*
475 : * cost_gather_merge
476 : * Determines and returns the cost of gather merge path.
477 : *
478 : * GatherMerge merges several pre-sorted input streams, using a heap that at
479 : * any given instant holds the next tuple from each stream. If there are N
480 : * streams, we need about N*log2(N) tuple comparisons to construct the heap at
481 : * startup, and then for each output tuple, about log2(N) comparisons to
482 : * replace the top heap entry with the next tuple from the same stream.
483 : */
484 : void
485 10190 : cost_gather_merge(GatherMergePath *path, PlannerInfo *root,
486 : RelOptInfo *rel, ParamPathInfo *param_info,
487 : int input_disabled_nodes,
488 : Cost input_startup_cost, Cost input_total_cost,
489 : double *rows)
490 : {
491 10190 : Cost startup_cost = 0;
492 10190 : Cost run_cost = 0;
493 : Cost comparison_cost;
494 : double N;
495 : double logN;
496 :
497 : /* Mark the path with the correct row estimate */
498 10190 : if (rows)
499 4624 : path->path.rows = *rows;
500 5566 : else if (param_info)
501 0 : path->path.rows = param_info->ppi_rows;
502 : else
503 5566 : path->path.rows = rel->rows;
504 :
505 : /*
506 : * Add one to the number of workers to account for the leader. This might
507 : * be overgenerous since the leader will do less work than other workers
508 : * in typical cases, but we'll go with it for now.
509 : */
510 : Assert(path->num_workers > 0);
511 10190 : N = (double) path->num_workers + 1;
512 10190 : logN = LOG2(N);
513 :
514 : /* Assumed cost per tuple comparison */
515 10190 : comparison_cost = 2.0 * cpu_operator_cost;
516 :
517 : /* Heap creation cost */
518 10190 : startup_cost += comparison_cost * N * logN;
519 :
520 : /* Per-tuple heap maintenance cost */
521 10190 : run_cost += path->path.rows * comparison_cost * logN;
522 :
523 : /* small cost for heap management, like cost_merge_append */
524 10190 : run_cost += cpu_operator_cost * path->path.rows;
525 :
526 : /*
527 : * Parallel setup and communication cost. Since Gather Merge, unlike
528 : * Gather, requires us to block until a tuple is available from every
529 : * worker, we bump the IPC cost up a little bit as compared with Gather.
530 : * For lack of a better idea, charge an extra 5%.
531 : */
532 10190 : startup_cost += parallel_setup_cost;
533 10190 : run_cost += parallel_tuple_cost * path->path.rows * 1.05;
534 :
535 10190 : path->path.disabled_nodes = input_disabled_nodes
536 10190 : + (enable_gathermerge ? 0 : 1);
537 10190 : path->path.startup_cost = startup_cost + input_startup_cost;
538 10190 : path->path.total_cost = (startup_cost + run_cost + input_total_cost);
539 10190 : }
540 :
541 : /*
542 : * cost_index
543 : * Determines and returns the cost of scanning a relation using an index.
544 : *
545 : * 'path' describes the indexscan under consideration, and is complete
546 : * except for the fields to be set by this routine
547 : * 'loop_count' is the number of repetitions of the indexscan to factor into
548 : * estimates of caching behavior
549 : *
550 : * In addition to rows, startup_cost and total_cost, cost_index() sets the
551 : * path's indextotalcost and indexselectivity fields. These values will be
552 : * needed if the IndexPath is used in a BitmapIndexScan.
553 : *
554 : * NOTE: path->indexquals must contain only clauses usable as index
555 : * restrictions. Any additional quals evaluated as qpquals may reduce the
556 : * number of returned tuples, but they won't reduce the number of tuples
557 : * we have to fetch from the table, so they don't reduce the scan cost.
558 : */
559 : void
560 783406 : cost_index(IndexPath *path, PlannerInfo *root, double loop_count,
561 : bool partial_path)
562 : {
563 783406 : IndexOptInfo *index = path->indexinfo;
564 783406 : RelOptInfo *baserel = index->rel;
565 783406 : bool indexonly = (path->path.pathtype == T_IndexOnlyScan);
566 : amcostestimate_function amcostestimate;
567 : List *qpquals;
568 783406 : Cost startup_cost = 0;
569 783406 : Cost run_cost = 0;
570 783406 : Cost cpu_run_cost = 0;
571 : Cost indexStartupCost;
572 : Cost indexTotalCost;
573 : Selectivity indexSelectivity;
574 : double indexCorrelation,
575 : csquared;
576 : double spc_seq_page_cost,
577 : spc_random_page_cost;
578 : Cost min_IO_cost,
579 : max_IO_cost;
580 : QualCost qpqual_cost;
581 : Cost cpu_per_tuple;
582 : double tuples_fetched;
583 : double pages_fetched;
584 : double rand_heap_pages;
585 : double index_pages;
586 :
587 : /* Should only be applied to base relations */
588 : Assert(IsA(baserel, RelOptInfo) &&
589 : IsA(index, IndexOptInfo));
590 : Assert(baserel->relid > 0);
591 : Assert(baserel->rtekind == RTE_RELATION);
592 :
593 : /*
594 : * Mark the path with the correct row estimate, and identify which quals
595 : * will need to be enforced as qpquals. We need not check any quals that
596 : * are implied by the index's predicate, so we can use indrestrictinfo not
597 : * baserestrictinfo as the list of relevant restriction clauses for the
598 : * rel.
599 : */
600 783406 : if (path->path.param_info)
601 : {
602 143600 : path->path.rows = path->path.param_info->ppi_rows;
603 : /* qpquals come from the rel's restriction clauses and ppi_clauses */
604 143600 : qpquals = list_concat(extract_nonindex_conditions(path->indexinfo->indrestrictinfo,
605 : path->indexclauses),
606 143600 : extract_nonindex_conditions(path->path.param_info->ppi_clauses,
607 : path->indexclauses));
608 : }
609 : else
610 : {
611 639806 : path->path.rows = baserel->rows;
612 : /* qpquals come from just the rel's restriction clauses */
613 639806 : qpquals = extract_nonindex_conditions(path->indexinfo->indrestrictinfo,
614 : path->indexclauses);
615 : }
616 :
617 : /* we don't need to check enable_indexonlyscan; indxpath.c does that */
618 783406 : path->path.disabled_nodes = enable_indexscan ? 0 : 1;
619 :
620 : /*
621 : * Call index-access-method-specific code to estimate the processing cost
622 : * for scanning the index, as well as the selectivity of the index (ie,
623 : * the fraction of main-table tuples we will have to retrieve) and its
624 : * correlation to the main-table tuple order. We need a cast here because
625 : * pathnodes.h uses a weak function type to avoid including amapi.h.
626 : */
627 783406 : amcostestimate = (amcostestimate_function) index->amcostestimate;
628 783406 : amcostestimate(root, path, loop_count,
629 : &indexStartupCost, &indexTotalCost,
630 : &indexSelectivity, &indexCorrelation,
631 : &index_pages);
632 :
633 : /*
634 : * Save amcostestimate's results for possible use in bitmap scan planning.
635 : * We don't bother to save indexStartupCost or indexCorrelation, because a
636 : * bitmap scan doesn't care about either.
637 : */
638 783406 : path->indextotalcost = indexTotalCost;
639 783406 : path->indexselectivity = indexSelectivity;
640 :
641 : /* all costs for touching index itself included here */
642 783406 : startup_cost += indexStartupCost;
643 783406 : run_cost += indexTotalCost - indexStartupCost;
644 :
645 : /* estimate number of main-table tuples fetched */
646 783406 : tuples_fetched = clamp_row_est(indexSelectivity * baserel->tuples);
647 :
648 : /* fetch estimated page costs for tablespace containing table */
649 783406 : get_tablespace_page_costs(baserel->reltablespace,
650 : &spc_random_page_cost,
651 : &spc_seq_page_cost);
652 :
653 : /*----------
654 : * Estimate number of main-table pages fetched, and compute I/O cost.
655 : *
656 : * When the index ordering is uncorrelated with the table ordering,
657 : * we use an approximation proposed by Mackert and Lohman (see
658 : * index_pages_fetched() for details) to compute the number of pages
659 : * fetched, and then charge spc_random_page_cost per page fetched.
660 : *
661 : * When the index ordering is exactly correlated with the table ordering
662 : * (just after a CLUSTER, for example), the number of pages fetched should
663 : * be exactly selectivity * table_size. What's more, all but the first
664 : * will be sequential fetches, not the random fetches that occur in the
665 : * uncorrelated case. So if the number of pages is more than 1, we
666 : * ought to charge
667 : * spc_random_page_cost + (pages_fetched - 1) * spc_seq_page_cost
668 : * For partially-correlated indexes, we ought to charge somewhere between
669 : * these two estimates. We currently interpolate linearly between the
670 : * estimates based on the correlation squared (XXX is that appropriate?).
671 : *
672 : * If it's an index-only scan, then we will not need to fetch any heap
673 : * pages for which the visibility map shows all tuples are visible.
674 : * Hence, reduce the estimated number of heap fetches accordingly.
675 : * We use the measured fraction of the entire heap that is all-visible,
676 : * which might not be particularly relevant to the subset of the heap
677 : * that this query will fetch; but it's not clear how to do better.
678 : *----------
679 : */
680 783406 : if (loop_count > 1)
681 : {
682 : /*
683 : * For repeated indexscans, the appropriate estimate for the
684 : * uncorrelated case is to scale up the number of tuples fetched in
685 : * the Mackert and Lohman formula by the number of scans, so that we
686 : * estimate the number of pages fetched by all the scans; then
687 : * pro-rate the costs for one scan. In this case we assume all the
688 : * fetches are random accesses.
689 : */
690 82992 : pages_fetched = index_pages_fetched(tuples_fetched * loop_count,
691 : baserel->pages,
692 82992 : (double) index->pages,
693 : root);
694 :
695 82992 : if (indexonly)
696 9360 : pages_fetched = ceil(pages_fetched * (1.0 - baserel->allvisfrac));
697 :
698 82992 : rand_heap_pages = pages_fetched;
699 :
700 82992 : max_IO_cost = (pages_fetched * spc_random_page_cost) / loop_count;
701 :
702 : /*
703 : * In the perfectly correlated case, the number of pages touched by
704 : * each scan is selectivity * table_size, and we can use the Mackert
705 : * and Lohman formula at the page level to estimate how much work is
706 : * saved by caching across scans. We still assume all the fetches are
707 : * random, though, which is an overestimate that's hard to correct for
708 : * without double-counting the cache effects. (But in most cases
709 : * where such a plan is actually interesting, only one page would get
710 : * fetched per scan anyway, so it shouldn't matter much.)
711 : */
712 82992 : pages_fetched = ceil(indexSelectivity * (double) baserel->pages);
713 :
714 82992 : pages_fetched = index_pages_fetched(pages_fetched * loop_count,
715 : baserel->pages,
716 82992 : (double) index->pages,
717 : root);
718 :
719 82992 : if (indexonly)
720 9360 : pages_fetched = ceil(pages_fetched * (1.0 - baserel->allvisfrac));
721 :
722 82992 : min_IO_cost = (pages_fetched * spc_random_page_cost) / loop_count;
723 : }
724 : else
725 : {
726 : /*
727 : * Normal case: apply the Mackert and Lohman formula, and then
728 : * interpolate between that and the correlation-derived result.
729 : */
730 700414 : pages_fetched = index_pages_fetched(tuples_fetched,
731 : baserel->pages,
732 700414 : (double) index->pages,
733 : root);
734 :
735 700414 : if (indexonly)
736 64364 : pages_fetched = ceil(pages_fetched * (1.0 - baserel->allvisfrac));
737 :
738 700414 : rand_heap_pages = pages_fetched;
739 :
740 : /* max_IO_cost is for the perfectly uncorrelated case (csquared=0) */
741 700414 : max_IO_cost = pages_fetched * spc_random_page_cost;
742 :
743 : /* min_IO_cost is for the perfectly correlated case (csquared=1) */
744 700414 : pages_fetched = ceil(indexSelectivity * (double) baserel->pages);
745 :
746 700414 : if (indexonly)
747 64364 : pages_fetched = ceil(pages_fetched * (1.0 - baserel->allvisfrac));
748 :
749 700414 : if (pages_fetched > 0)
750 : {
751 635782 : min_IO_cost = spc_random_page_cost;
752 635782 : if (pages_fetched > 1)
753 187546 : min_IO_cost += (pages_fetched - 1) * spc_seq_page_cost;
754 : }
755 : else
756 64632 : min_IO_cost = 0;
757 : }
758 :
759 783406 : if (partial_path)
760 : {
761 : /*
762 : * For index only scans compute workers based on number of index pages
763 : * fetched; the number of heap pages we fetch might be so small as to
764 : * effectively rule out parallelism, which we don't want to do.
765 : */
766 272286 : if (indexonly)
767 23216 : rand_heap_pages = -1;
768 :
769 : /*
770 : * Estimate the number of parallel workers required to scan index. Use
771 : * the number of heap pages computed considering heap fetches won't be
772 : * sequential as for parallel scans the pages are accessed in random
773 : * order.
774 : */
775 272286 : path->path.parallel_workers = compute_parallel_worker(baserel,
776 : rand_heap_pages,
777 : index_pages,
778 : max_parallel_workers_per_gather);
779 :
780 : /*
781 : * Fall out if workers can't be assigned for parallel scan, because in
782 : * such a case this path will be rejected. So there is no benefit in
783 : * doing extra computation.
784 : */
785 272286 : if (path->path.parallel_workers <= 0)
786 262222 : return;
787 :
788 10064 : path->path.parallel_aware = true;
789 : }
790 :
791 : /*
792 : * Now interpolate based on estimated index order correlation to get total
793 : * disk I/O cost for main table accesses.
794 : */
795 521184 : csquared = indexCorrelation * indexCorrelation;
796 :
797 521184 : run_cost += max_IO_cost + csquared * (min_IO_cost - max_IO_cost);
798 :
799 : /*
800 : * Estimate CPU costs per tuple.
801 : *
802 : * What we want here is cpu_tuple_cost plus the evaluation costs of any
803 : * qual clauses that we have to evaluate as qpquals.
804 : */
805 521184 : cost_qual_eval(&qpqual_cost, qpquals, root);
806 :
807 521184 : startup_cost += qpqual_cost.startup;
808 521184 : cpu_per_tuple = cpu_tuple_cost + qpqual_cost.per_tuple;
809 :
810 521184 : cpu_run_cost += cpu_per_tuple * tuples_fetched;
811 :
812 : /* tlist eval costs are paid per output row, not per tuple scanned */
813 521184 : startup_cost += path->path.pathtarget->cost.startup;
814 521184 : cpu_run_cost += path->path.pathtarget->cost.per_tuple * path->path.rows;
815 :
816 : /* Adjust costing for parallelism, if used. */
817 521184 : if (path->path.parallel_workers > 0)
818 : {
819 10064 : double parallel_divisor = get_parallel_divisor(&path->path);
820 :
821 10064 : path->path.rows = clamp_row_est(path->path.rows / parallel_divisor);
822 :
823 : /* The CPU cost is divided among all the workers. */
824 10064 : cpu_run_cost /= parallel_divisor;
825 : }
826 :
827 521184 : run_cost += cpu_run_cost;
828 :
829 521184 : path->path.startup_cost = startup_cost;
830 521184 : path->path.total_cost = startup_cost + run_cost;
831 : }
832 :
833 : /*
834 : * extract_nonindex_conditions
835 : *
836 : * Given a list of quals to be enforced in an indexscan, extract the ones that
837 : * will have to be applied as qpquals (ie, the index machinery won't handle
838 : * them). Here we detect only whether a qual clause is directly redundant
839 : * with some indexclause. If the index path is chosen for use, createplan.c
840 : * will try a bit harder to get rid of redundant qual conditions; specifically
841 : * it will see if quals can be proven to be implied by the indexquals. But
842 : * it does not seem worth the cycles to try to factor that in at this stage,
843 : * since we're only trying to estimate qual eval costs. Otherwise this must
844 : * match the logic in create_indexscan_plan().
845 : *
846 : * qual_clauses, and the result, are lists of RestrictInfos.
847 : * indexclauses is a list of IndexClauses.
848 : */
849 : static List *
850 927006 : extract_nonindex_conditions(List *qual_clauses, List *indexclauses)
851 : {
852 927006 : List *result = NIL;
853 : ListCell *lc;
854 :
855 1945178 : foreach(lc, qual_clauses)
856 : {
857 1018172 : RestrictInfo *rinfo = lfirst_node(RestrictInfo, lc);
858 :
859 1018172 : if (rinfo->pseudoconstant)
860 9694 : continue; /* we may drop pseudoconstants here */
861 1008478 : if (is_redundant_with_indexclauses(rinfo, indexclauses))
862 591930 : continue; /* dup or derived from same EquivalenceClass */
863 : /* ... skip the predicate proof attempt createplan.c will try ... */
864 416548 : result = lappend(result, rinfo);
865 : }
866 927006 : return result;
867 : }
868 :
869 : /*
870 : * index_pages_fetched
871 : * Estimate the number of pages actually fetched after accounting for
872 : * cache effects.
873 : *
874 : * We use an approximation proposed by Mackert and Lohman, "Index Scans
875 : * Using a Finite LRU Buffer: A Validated I/O Model", ACM Transactions
876 : * on Database Systems, Vol. 14, No. 3, September 1989, Pages 401-424.
877 : * The Mackert and Lohman approximation is that the number of pages
878 : * fetched is
879 : * PF =
880 : * min(2TNs/(2T+Ns), T) when T <= b
881 : * 2TNs/(2T+Ns) when T > b and Ns <= 2Tb/(2T-b)
882 : * b + (Ns - 2Tb/(2T-b))*(T-b)/T when T > b and Ns > 2Tb/(2T-b)
883 : * where
884 : * T = # pages in table
885 : * N = # tuples in table
886 : * s = selectivity = fraction of table to be scanned
887 : * b = # buffer pages available (we include kernel space here)
888 : *
889 : * We assume that effective_cache_size is the total number of buffer pages
890 : * available for the whole query, and pro-rate that space across all the
891 : * tables in the query and the index currently under consideration. (This
892 : * ignores space needed for other indexes used by the query, but since we
893 : * don't know which indexes will get used, we can't estimate that very well;
894 : * and in any case counting all the tables may well be an overestimate, since
895 : * depending on the join plan not all the tables may be scanned concurrently.)
896 : *
897 : * The product Ns is the number of tuples fetched; we pass in that
898 : * product rather than calculating it here. "pages" is the number of pages
899 : * in the object under consideration (either an index or a table).
900 : * "index_pages" is the amount to add to the total table space, which was
901 : * computed for us by make_one_rel.
902 : *
903 : * Caller is expected to have ensured that tuples_fetched is greater than zero
904 : * and rounded to integer (see clamp_row_est). The result will likewise be
905 : * greater than zero and integral.
906 : */
907 : double
908 1093930 : index_pages_fetched(double tuples_fetched, BlockNumber pages,
909 : double index_pages, PlannerInfo *root)
910 : {
911 : double pages_fetched;
912 : double total_pages;
913 : double T,
914 : b;
915 :
916 : /* T is # pages in table, but don't allow it to be zero */
917 1093930 : T = (pages > 1) ? (double) pages : 1.0;
918 :
919 : /* Compute number of pages assumed to be competing for cache space */
920 1093930 : total_pages = root->total_table_pages + index_pages;
921 1093930 : total_pages = Max(total_pages, 1.0);
922 : Assert(T <= total_pages);
923 :
924 : /* b is pro-rated share of effective_cache_size */
925 1093930 : b = (double) effective_cache_size * T / total_pages;
926 :
927 : /* force it positive and integral */
928 1093930 : if (b <= 1.0)
929 0 : b = 1.0;
930 : else
931 1093930 : b = ceil(b);
932 :
933 : /* This part is the Mackert and Lohman formula */
934 1093930 : if (T <= b)
935 : {
936 1093930 : pages_fetched =
937 1093930 : (2.0 * T * tuples_fetched) / (2.0 * T + tuples_fetched);
938 1093930 : if (pages_fetched >= T)
939 634516 : pages_fetched = T;
940 : else
941 459414 : pages_fetched = ceil(pages_fetched);
942 : }
943 : else
944 : {
945 : double lim;
946 :
947 0 : lim = (2.0 * T * b) / (2.0 * T - b);
948 0 : if (tuples_fetched <= lim)
949 : {
950 0 : pages_fetched =
951 0 : (2.0 * T * tuples_fetched) / (2.0 * T + tuples_fetched);
952 : }
953 : else
954 : {
955 0 : pages_fetched =
956 0 : b + (tuples_fetched - lim) * (T - b) / T;
957 : }
958 0 : pages_fetched = ceil(pages_fetched);
959 : }
960 1093930 : return pages_fetched;
961 : }
962 :
963 : /*
964 : * get_indexpath_pages
965 : * Determine the total size of the indexes used in a bitmap index path.
966 : *
967 : * Note: if the same index is used more than once in a bitmap tree, we will
968 : * count it multiple times, which perhaps is the wrong thing ... but it's
969 : * not completely clear, and detecting duplicates is difficult, so ignore it
970 : * for now.
971 : */
972 : static double
973 186588 : get_indexpath_pages(Path *bitmapqual)
974 : {
975 186588 : double result = 0;
976 : ListCell *l;
977 :
978 186588 : if (IsA(bitmapqual, BitmapAndPath))
979 : {
980 23352 : BitmapAndPath *apath = (BitmapAndPath *) bitmapqual;
981 :
982 70056 : foreach(l, apath->bitmapquals)
983 : {
984 46704 : result += get_indexpath_pages((Path *) lfirst(l));
985 : }
986 : }
987 163236 : else if (IsA(bitmapqual, BitmapOrPath))
988 : {
989 70 : BitmapOrPath *opath = (BitmapOrPath *) bitmapqual;
990 :
991 222 : foreach(l, opath->bitmapquals)
992 : {
993 152 : result += get_indexpath_pages((Path *) lfirst(l));
994 : }
995 : }
996 163166 : else if (IsA(bitmapqual, IndexPath))
997 : {
998 163166 : IndexPath *ipath = (IndexPath *) bitmapqual;
999 :
1000 163166 : result = (double) ipath->indexinfo->pages;
1001 : }
1002 : else
1003 0 : elog(ERROR, "unrecognized node type: %d", nodeTag(bitmapqual));
1004 :
1005 186588 : return result;
1006 : }
1007 :
1008 : /*
1009 : * cost_bitmap_heap_scan
1010 : * Determines and returns the cost of scanning a relation using a bitmap
1011 : * index-then-heap plan.
1012 : *
1013 : * 'baserel' is the relation to be scanned
1014 : * 'param_info' is the ParamPathInfo if this is a parameterized path, else NULL
1015 : * 'bitmapqual' is a tree of IndexPaths, BitmapAndPaths, and BitmapOrPaths
1016 : * 'loop_count' is the number of repetitions of the indexscan to factor into
1017 : * estimates of caching behavior
1018 : *
1019 : * Note: the component IndexPaths in bitmapqual should have been costed
1020 : * using the same loop_count.
1021 : */
1022 : void
1023 535218 : cost_bitmap_heap_scan(Path *path, PlannerInfo *root, RelOptInfo *baserel,
1024 : ParamPathInfo *param_info,
1025 : Path *bitmapqual, double loop_count)
1026 : {
1027 535218 : Cost startup_cost = 0;
1028 535218 : Cost run_cost = 0;
1029 : Cost indexTotalCost;
1030 : QualCost qpqual_cost;
1031 : Cost cpu_per_tuple;
1032 : Cost cost_per_page;
1033 : Cost cpu_run_cost;
1034 : double tuples_fetched;
1035 : double pages_fetched;
1036 : double spc_seq_page_cost,
1037 : spc_random_page_cost;
1038 : double T;
1039 :
1040 : /* Should only be applied to base relations */
1041 : Assert(IsA(baserel, RelOptInfo));
1042 : Assert(baserel->relid > 0);
1043 : Assert(baserel->rtekind == RTE_RELATION);
1044 :
1045 : /* Mark the path with the correct row estimate */
1046 535218 : if (param_info)
1047 222466 : path->rows = param_info->ppi_rows;
1048 : else
1049 312752 : path->rows = baserel->rows;
1050 :
1051 535218 : pages_fetched = compute_bitmap_pages(root, baserel, bitmapqual,
1052 : loop_count, &indexTotalCost,
1053 : &tuples_fetched);
1054 :
1055 535218 : startup_cost += indexTotalCost;
1056 535218 : T = (baserel->pages > 1) ? (double) baserel->pages : 1.0;
1057 :
1058 : /* Fetch estimated page costs for tablespace containing table. */
1059 535218 : get_tablespace_page_costs(baserel->reltablespace,
1060 : &spc_random_page_cost,
1061 : &spc_seq_page_cost);
1062 :
1063 : /*
1064 : * For small numbers of pages we should charge spc_random_page_cost
1065 : * apiece, while if nearly all the table's pages are being read, it's more
1066 : * appropriate to charge spc_seq_page_cost apiece. The effect is
1067 : * nonlinear, too. For lack of a better idea, interpolate like this to
1068 : * determine the cost per page.
1069 : */
1070 535218 : if (pages_fetched >= 2.0)
1071 113116 : cost_per_page = spc_random_page_cost -
1072 113116 : (spc_random_page_cost - spc_seq_page_cost)
1073 113116 : * sqrt(pages_fetched / T);
1074 : else
1075 422102 : cost_per_page = spc_random_page_cost;
1076 :
1077 535218 : run_cost += pages_fetched * cost_per_page;
1078 :
1079 : /*
1080 : * Estimate CPU costs per tuple.
1081 : *
1082 : * Often the indexquals don't need to be rechecked at each tuple ... but
1083 : * not always, especially not if there are enough tuples involved that the
1084 : * bitmaps become lossy. For the moment, just assume they will be
1085 : * rechecked always. This means we charge the full freight for all the
1086 : * scan clauses.
1087 : */
1088 535218 : get_restriction_qual_cost(root, baserel, param_info, &qpqual_cost);
1089 :
1090 535218 : startup_cost += qpqual_cost.startup;
1091 535218 : cpu_per_tuple = cpu_tuple_cost + qpqual_cost.per_tuple;
1092 535218 : cpu_run_cost = cpu_per_tuple * tuples_fetched;
1093 :
1094 : /* Adjust costing for parallelism, if used. */
1095 535218 : if (path->parallel_workers > 0)
1096 : {
1097 4190 : double parallel_divisor = get_parallel_divisor(path);
1098 :
1099 : /* The CPU cost is divided among all the workers. */
1100 4190 : cpu_run_cost /= parallel_divisor;
1101 :
1102 4190 : path->rows = clamp_row_est(path->rows / parallel_divisor);
1103 : }
1104 :
1105 :
1106 535218 : run_cost += cpu_run_cost;
1107 :
1108 : /* tlist eval costs are paid per output row, not per tuple scanned */
1109 535218 : startup_cost += path->pathtarget->cost.startup;
1110 535218 : run_cost += path->pathtarget->cost.per_tuple * path->rows;
1111 :
1112 535218 : path->disabled_nodes = enable_bitmapscan ? 0 : 1;
1113 535218 : path->startup_cost = startup_cost;
1114 535218 : path->total_cost = startup_cost + run_cost;
1115 535218 : }
1116 :
1117 : /*
1118 : * cost_bitmap_tree_node
1119 : * Extract cost and selectivity from a bitmap tree node (index/and/or)
1120 : */
1121 : void
1122 1002220 : cost_bitmap_tree_node(Path *path, Cost *cost, Selectivity *selec)
1123 : {
1124 1002220 : if (IsA(path, IndexPath))
1125 : {
1126 947418 : *cost = ((IndexPath *) path)->indextotalcost;
1127 947418 : *selec = ((IndexPath *) path)->indexselectivity;
1128 :
1129 : /*
1130 : * Charge a small amount per retrieved tuple to reflect the costs of
1131 : * manipulating the bitmap. This is mostly to make sure that a bitmap
1132 : * scan doesn't look to be the same cost as an indexscan to retrieve a
1133 : * single tuple.
1134 : */
1135 947418 : *cost += 0.1 * cpu_operator_cost * path->rows;
1136 : }
1137 54802 : else if (IsA(path, BitmapAndPath))
1138 : {
1139 51570 : *cost = path->total_cost;
1140 51570 : *selec = ((BitmapAndPath *) path)->bitmapselectivity;
1141 : }
1142 3232 : else if (IsA(path, BitmapOrPath))
1143 : {
1144 3232 : *cost = path->total_cost;
1145 3232 : *selec = ((BitmapOrPath *) path)->bitmapselectivity;
1146 : }
1147 : else
1148 : {
1149 0 : elog(ERROR, "unrecognized node type: %d", nodeTag(path));
1150 : *cost = *selec = 0; /* keep compiler quiet */
1151 : }
1152 1002220 : }
1153 :
1154 : /*
1155 : * cost_bitmap_and_node
1156 : * Estimate the cost of a BitmapAnd node
1157 : *
1158 : * Note that this considers only the costs of index scanning and bitmap
1159 : * creation, not the eventual heap access. In that sense the object isn't
1160 : * truly a Path, but it has enough path-like properties (costs in particular)
1161 : * to warrant treating it as one. We don't bother to set the path rows field,
1162 : * however.
1163 : */
1164 : void
1165 51384 : cost_bitmap_and_node(BitmapAndPath *path, PlannerInfo *root)
1166 : {
1167 : Cost totalCost;
1168 : Selectivity selec;
1169 : ListCell *l;
1170 :
1171 : /*
1172 : * We estimate AND selectivity on the assumption that the inputs are
1173 : * independent. This is probably often wrong, but we don't have the info
1174 : * to do better.
1175 : *
1176 : * The runtime cost of the BitmapAnd itself is estimated at 100x
1177 : * cpu_operator_cost for each tbm_intersect needed. Probably too small,
1178 : * definitely too simplistic?
1179 : */
1180 51384 : totalCost = 0.0;
1181 51384 : selec = 1.0;
1182 154152 : foreach(l, path->bitmapquals)
1183 : {
1184 102768 : Path *subpath = (Path *) lfirst(l);
1185 : Cost subCost;
1186 : Selectivity subselec;
1187 :
1188 102768 : cost_bitmap_tree_node(subpath, &subCost, &subselec);
1189 :
1190 102768 : selec *= subselec;
1191 :
1192 102768 : totalCost += subCost;
1193 102768 : if (l != list_head(path->bitmapquals))
1194 51384 : totalCost += 100.0 * cpu_operator_cost;
1195 : }
1196 51384 : path->bitmapselectivity = selec;
1197 51384 : path->path.rows = 0; /* per above, not used */
1198 51384 : path->path.disabled_nodes = 0;
1199 51384 : path->path.startup_cost = totalCost;
1200 51384 : path->path.total_cost = totalCost;
1201 51384 : }
1202 :
1203 : /*
1204 : * cost_bitmap_or_node
1205 : * Estimate the cost of a BitmapOr node
1206 : *
1207 : * See comments for cost_bitmap_and_node.
1208 : */
1209 : void
1210 976 : cost_bitmap_or_node(BitmapOrPath *path, PlannerInfo *root)
1211 : {
1212 : Cost totalCost;
1213 : Selectivity selec;
1214 : ListCell *l;
1215 :
1216 : /*
1217 : * We estimate OR selectivity on the assumption that the inputs are
1218 : * non-overlapping, since that's often the case in "x IN (list)" type
1219 : * situations. Of course, we clamp to 1.0 at the end.
1220 : *
1221 : * The runtime cost of the BitmapOr itself is estimated at 100x
1222 : * cpu_operator_cost for each tbm_union needed. Probably too small,
1223 : * definitely too simplistic? We are aware that the tbm_unions are
1224 : * optimized out when the inputs are BitmapIndexScans.
1225 : */
1226 976 : totalCost = 0.0;
1227 976 : selec = 0.0;
1228 2736 : foreach(l, path->bitmapquals)
1229 : {
1230 1760 : Path *subpath = (Path *) lfirst(l);
1231 : Cost subCost;
1232 : Selectivity subselec;
1233 :
1234 1760 : cost_bitmap_tree_node(subpath, &subCost, &subselec);
1235 :
1236 1760 : selec += subselec;
1237 :
1238 1760 : totalCost += subCost;
1239 1760 : if (l != list_head(path->bitmapquals) &&
1240 784 : !IsA(subpath, IndexPath))
1241 6 : totalCost += 100.0 * cpu_operator_cost;
1242 : }
1243 976 : path->bitmapselectivity = Min(selec, 1.0);
1244 976 : path->path.rows = 0; /* per above, not used */
1245 976 : path->path.startup_cost = totalCost;
1246 976 : path->path.total_cost = totalCost;
1247 976 : }
1248 :
1249 : /*
1250 : * cost_tidscan
1251 : * Determines and returns the cost of scanning a relation using TIDs.
1252 : *
1253 : * 'baserel' is the relation to be scanned
1254 : * 'tidquals' is the list of TID-checkable quals
1255 : * 'param_info' is the ParamPathInfo if this is a parameterized path, else NULL
1256 : */
1257 : void
1258 852 : cost_tidscan(Path *path, PlannerInfo *root,
1259 : RelOptInfo *baserel, List *tidquals, ParamPathInfo *param_info)
1260 : {
1261 852 : Cost startup_cost = 0;
1262 852 : Cost run_cost = 0;
1263 : QualCost qpqual_cost;
1264 : Cost cpu_per_tuple;
1265 : QualCost tid_qual_cost;
1266 : double ntuples;
1267 : ListCell *l;
1268 : double spc_random_page_cost;
1269 :
1270 : /* Should only be applied to base relations */
1271 : Assert(baserel->relid > 0);
1272 : Assert(baserel->rtekind == RTE_RELATION);
1273 : Assert(tidquals != NIL);
1274 :
1275 : /* Mark the path with the correct row estimate */
1276 852 : if (param_info)
1277 144 : path->rows = param_info->ppi_rows;
1278 : else
1279 708 : path->rows = baserel->rows;
1280 :
1281 : /* Count how many tuples we expect to retrieve */
1282 852 : ntuples = 0;
1283 1728 : foreach(l, tidquals)
1284 : {
1285 876 : RestrictInfo *rinfo = lfirst_node(RestrictInfo, l);
1286 876 : Expr *qual = rinfo->clause;
1287 :
1288 : /*
1289 : * We must use a TID scan for CurrentOfExpr; in any other case, we
1290 : * should be generating a TID scan only if enable_tidscan=true. Also,
1291 : * if CurrentOfExpr is the qual, there should be only one.
1292 : */
1293 : Assert(enable_tidscan || IsA(qual, CurrentOfExpr));
1294 : Assert(list_length(tidquals) == 1 || !IsA(qual, CurrentOfExpr));
1295 :
1296 876 : if (IsA(qual, ScalarArrayOpExpr))
1297 : {
1298 : /* Each element of the array yields 1 tuple */
1299 50 : ScalarArrayOpExpr *saop = (ScalarArrayOpExpr *) qual;
1300 50 : Node *arraynode = (Node *) lsecond(saop->args);
1301 :
1302 50 : ntuples += estimate_array_length(root, arraynode);
1303 : }
1304 826 : else if (IsA(qual, CurrentOfExpr))
1305 : {
1306 : /* CURRENT OF yields 1 tuple */
1307 404 : ntuples++;
1308 : }
1309 : else
1310 : {
1311 : /* It's just CTID = something, count 1 tuple */
1312 422 : ntuples++;
1313 : }
1314 : }
1315 :
1316 : /*
1317 : * The TID qual expressions will be computed once, any other baserestrict
1318 : * quals once per retrieved tuple.
1319 : */
1320 852 : cost_qual_eval(&tid_qual_cost, tidquals, root);
1321 :
1322 : /* fetch estimated page cost for tablespace containing table */
1323 852 : get_tablespace_page_costs(baserel->reltablespace,
1324 : &spc_random_page_cost,
1325 : NULL);
1326 :
1327 : /* disk costs --- assume each tuple on a different page */
1328 852 : run_cost += spc_random_page_cost * ntuples;
1329 :
1330 : /* Add scanning CPU costs */
1331 852 : get_restriction_qual_cost(root, baserel, param_info, &qpqual_cost);
1332 :
1333 : /* XXX currently we assume TID quals are a subset of qpquals */
1334 852 : startup_cost += qpqual_cost.startup + tid_qual_cost.per_tuple;
1335 852 : cpu_per_tuple = cpu_tuple_cost + qpqual_cost.per_tuple -
1336 852 : tid_qual_cost.per_tuple;
1337 852 : run_cost += cpu_per_tuple * ntuples;
1338 :
1339 : /* tlist eval costs are paid per output row, not per tuple scanned */
1340 852 : startup_cost += path->pathtarget->cost.startup;
1341 852 : run_cost += path->pathtarget->cost.per_tuple * path->rows;
1342 :
1343 : /*
1344 : * There are assertions above verifying that we only reach this function
1345 : * either when enable_tidscan=true or when the TID scan is the only legal
1346 : * path, so it's safe to set disabled_nodes to zero here.
1347 : */
1348 852 : path->disabled_nodes = 0;
1349 852 : path->startup_cost = startup_cost;
1350 852 : path->total_cost = startup_cost + run_cost;
1351 852 : }
1352 :
1353 : /*
1354 : * cost_tidrangescan
1355 : * Determines and sets the costs of scanning a relation using a range of
1356 : * TIDs for 'path'
1357 : *
1358 : * 'baserel' is the relation to be scanned
1359 : * 'tidrangequals' is the list of TID-checkable range quals
1360 : * 'param_info' is the ParamPathInfo if this is a parameterized path, else NULL
1361 : */
1362 : void
1363 1940 : cost_tidrangescan(Path *path, PlannerInfo *root,
1364 : RelOptInfo *baserel, List *tidrangequals,
1365 : ParamPathInfo *param_info)
1366 : {
1367 : Selectivity selectivity;
1368 : double pages;
1369 1940 : Cost startup_cost = 0;
1370 1940 : Cost run_cost = 0;
1371 : QualCost qpqual_cost;
1372 : Cost cpu_per_tuple;
1373 : QualCost tid_qual_cost;
1374 : double ntuples;
1375 : double nseqpages;
1376 : double spc_random_page_cost;
1377 : double spc_seq_page_cost;
1378 :
1379 : /* Should only be applied to base relations */
1380 : Assert(baserel->relid > 0);
1381 : Assert(baserel->rtekind == RTE_RELATION);
1382 :
1383 : /* Mark the path with the correct row estimate */
1384 1940 : if (param_info)
1385 0 : path->rows = param_info->ppi_rows;
1386 : else
1387 1940 : path->rows = baserel->rows;
1388 :
1389 : /* Count how many tuples and pages we expect to scan */
1390 1940 : selectivity = clauselist_selectivity(root, tidrangequals, baserel->relid,
1391 : JOIN_INNER, NULL);
1392 1940 : pages = ceil(selectivity * baserel->pages);
1393 :
1394 1940 : if (pages <= 0.0)
1395 42 : pages = 1.0;
1396 :
1397 : /*
1398 : * The first page in a range requires a random seek, but each subsequent
1399 : * page is just a normal sequential page read. NOTE: it's desirable for
1400 : * TID Range Scans to cost more than the equivalent Sequential Scans,
1401 : * because Seq Scans have some performance advantages such as scan
1402 : * synchronization and parallelizability, and we'd prefer one of them to
1403 : * be picked unless a TID Range Scan really is better.
1404 : */
1405 1940 : ntuples = selectivity * baserel->tuples;
1406 1940 : nseqpages = pages - 1.0;
1407 :
1408 : /*
1409 : * The TID qual expressions will be computed once, any other baserestrict
1410 : * quals once per retrieved tuple.
1411 : */
1412 1940 : cost_qual_eval(&tid_qual_cost, tidrangequals, root);
1413 :
1414 : /* fetch estimated page cost for tablespace containing table */
1415 1940 : get_tablespace_page_costs(baserel->reltablespace,
1416 : &spc_random_page_cost,
1417 : &spc_seq_page_cost);
1418 :
1419 : /* disk costs; 1 random page and the remainder as seq pages */
1420 1940 : run_cost += spc_random_page_cost + spc_seq_page_cost * nseqpages;
1421 :
1422 : /* Add scanning CPU costs */
1423 1940 : get_restriction_qual_cost(root, baserel, param_info, &qpqual_cost);
1424 :
1425 : /*
1426 : * XXX currently we assume TID quals are a subset of qpquals at this
1427 : * point; they will be removed (if possible) when we create the plan, so
1428 : * we subtract their cost from the total qpqual cost. (If the TID quals
1429 : * can't be removed, this is a mistake and we're going to underestimate
1430 : * the CPU cost a bit.)
1431 : */
1432 1940 : startup_cost += qpqual_cost.startup + tid_qual_cost.per_tuple;
1433 1940 : cpu_per_tuple = cpu_tuple_cost + qpqual_cost.per_tuple -
1434 1940 : tid_qual_cost.per_tuple;
1435 1940 : run_cost += cpu_per_tuple * ntuples;
1436 :
1437 : /* tlist eval costs are paid per output row, not per tuple scanned */
1438 1940 : startup_cost += path->pathtarget->cost.startup;
1439 1940 : run_cost += path->pathtarget->cost.per_tuple * path->rows;
1440 :
1441 : /* we should not generate this path type when enable_tidscan=false */
1442 : Assert(enable_tidscan);
1443 1940 : path->disabled_nodes = 0;
1444 1940 : path->startup_cost = startup_cost;
1445 1940 : path->total_cost = startup_cost + run_cost;
1446 1940 : }
1447 :
1448 : /*
1449 : * cost_subqueryscan
1450 : * Determines and returns the cost of scanning a subquery RTE.
1451 : *
1452 : * 'baserel' is the relation to be scanned
1453 : * 'param_info' is the ParamPathInfo if this is a parameterized path, else NULL
1454 : * 'trivial_pathtarget' is true if the pathtarget is believed to be trivial.
1455 : */
1456 : void
1457 48672 : cost_subqueryscan(SubqueryScanPath *path, PlannerInfo *root,
1458 : RelOptInfo *baserel, ParamPathInfo *param_info,
1459 : bool trivial_pathtarget)
1460 : {
1461 : Cost startup_cost;
1462 : Cost run_cost;
1463 : List *qpquals;
1464 : QualCost qpqual_cost;
1465 : Cost cpu_per_tuple;
1466 :
1467 : /* Should only be applied to base relations that are subqueries */
1468 : Assert(baserel->relid > 0);
1469 : Assert(baserel->rtekind == RTE_SUBQUERY);
1470 :
1471 : /*
1472 : * We compute the rowcount estimate as the subplan's estimate times the
1473 : * selectivity of relevant restriction clauses. In simple cases this will
1474 : * come out the same as baserel->rows; but when dealing with parallelized
1475 : * paths we must do it like this to get the right answer.
1476 : */
1477 48672 : if (param_info)
1478 606 : qpquals = list_concat_copy(param_info->ppi_clauses,
1479 606 : baserel->baserestrictinfo);
1480 : else
1481 48066 : qpquals = baserel->baserestrictinfo;
1482 :
1483 48672 : path->path.rows = clamp_row_est(path->subpath->rows *
1484 48672 : clauselist_selectivity(root,
1485 : qpquals,
1486 : 0,
1487 : JOIN_INNER,
1488 : NULL));
1489 :
1490 : /*
1491 : * Cost of path is cost of evaluating the subplan, plus cost of evaluating
1492 : * any restriction clauses and tlist that will be attached to the
1493 : * SubqueryScan node, plus cpu_tuple_cost to account for selection and
1494 : * projection overhead.
1495 : */
1496 48672 : path->path.disabled_nodes = path->subpath->disabled_nodes;
1497 48672 : path->path.startup_cost = path->subpath->startup_cost;
1498 48672 : path->path.total_cost = path->subpath->total_cost;
1499 :
1500 : /*
1501 : * However, if there are no relevant restriction clauses and the
1502 : * pathtarget is trivial, then we expect that setrefs.c will optimize away
1503 : * the SubqueryScan plan node altogether, so we should just make its cost
1504 : * and rowcount equal to the input path's.
1505 : *
1506 : * Note: there are some edge cases where createplan.c will apply a
1507 : * different targetlist to the SubqueryScan node, thus falsifying our
1508 : * current estimate of whether the target is trivial, and making the cost
1509 : * estimate (though not the rowcount) wrong. It does not seem worth the
1510 : * extra complication to try to account for that exactly, especially since
1511 : * that behavior falsifies other cost estimates as well.
1512 : */
1513 48672 : if (qpquals == NIL && trivial_pathtarget)
1514 24682 : return;
1515 :
1516 23990 : get_restriction_qual_cost(root, baserel, param_info, &qpqual_cost);
1517 :
1518 23990 : startup_cost = qpqual_cost.startup;
1519 23990 : cpu_per_tuple = cpu_tuple_cost + qpqual_cost.per_tuple;
1520 23990 : run_cost = cpu_per_tuple * path->subpath->rows;
1521 :
1522 : /* tlist eval costs are paid per output row, not per tuple scanned */
1523 23990 : startup_cost += path->path.pathtarget->cost.startup;
1524 23990 : run_cost += path->path.pathtarget->cost.per_tuple * path->path.rows;
1525 :
1526 23990 : path->path.startup_cost += startup_cost;
1527 23990 : path->path.total_cost += startup_cost + run_cost;
1528 : }
1529 :
1530 : /*
1531 : * cost_functionscan
1532 : * Determines and returns the cost of scanning a function RTE.
1533 : *
1534 : * 'baserel' is the relation to be scanned
1535 : * 'param_info' is the ParamPathInfo if this is a parameterized path, else NULL
1536 : */
1537 : void
1538 51610 : cost_functionscan(Path *path, PlannerInfo *root,
1539 : RelOptInfo *baserel, ParamPathInfo *param_info)
1540 : {
1541 51610 : Cost startup_cost = 0;
1542 51610 : Cost run_cost = 0;
1543 : QualCost qpqual_cost;
1544 : Cost cpu_per_tuple;
1545 : RangeTblEntry *rte;
1546 : QualCost exprcost;
1547 :
1548 : /* Should only be applied to base relations that are functions */
1549 : Assert(baserel->relid > 0);
1550 51610 : rte = planner_rt_fetch(baserel->relid, root);
1551 : Assert(rte->rtekind == RTE_FUNCTION);
1552 :
1553 : /* Mark the path with the correct row estimate */
1554 51610 : if (param_info)
1555 8294 : path->rows = param_info->ppi_rows;
1556 : else
1557 43316 : path->rows = baserel->rows;
1558 :
1559 : /*
1560 : * Estimate costs of executing the function expression(s).
1561 : *
1562 : * Currently, nodeFunctionscan.c always executes the functions to
1563 : * completion before returning any rows, and caches the results in a
1564 : * tuplestore. So the function eval cost is all startup cost, and per-row
1565 : * costs are minimal.
1566 : *
1567 : * XXX in principle we ought to charge tuplestore spill costs if the
1568 : * number of rows is large. However, given how phony our rowcount
1569 : * estimates for functions tend to be, there's not a lot of point in that
1570 : * refinement right now.
1571 : */
1572 51610 : cost_qual_eval_node(&exprcost, (Node *) rte->functions, root);
1573 :
1574 51610 : startup_cost += exprcost.startup + exprcost.per_tuple;
1575 :
1576 : /* Add scanning CPU costs */
1577 51610 : get_restriction_qual_cost(root, baserel, param_info, &qpqual_cost);
1578 :
1579 51610 : startup_cost += qpqual_cost.startup;
1580 51610 : cpu_per_tuple = cpu_tuple_cost + qpqual_cost.per_tuple;
1581 51610 : run_cost += cpu_per_tuple * baserel->tuples;
1582 :
1583 : /* tlist eval costs are paid per output row, not per tuple scanned */
1584 51610 : startup_cost += path->pathtarget->cost.startup;
1585 51610 : run_cost += path->pathtarget->cost.per_tuple * path->rows;
1586 :
1587 51610 : path->disabled_nodes = 0;
1588 51610 : path->startup_cost = startup_cost;
1589 51610 : path->total_cost = startup_cost + run_cost;
1590 51610 : }
1591 :
1592 : /*
1593 : * cost_tablefuncscan
1594 : * Determines and returns the cost of scanning a table function.
1595 : *
1596 : * 'baserel' is the relation to be scanned
1597 : * 'param_info' is the ParamPathInfo if this is a parameterized path, else NULL
1598 : */
1599 : void
1600 626 : cost_tablefuncscan(Path *path, PlannerInfo *root,
1601 : RelOptInfo *baserel, ParamPathInfo *param_info)
1602 : {
1603 626 : Cost startup_cost = 0;
1604 626 : Cost run_cost = 0;
1605 : QualCost qpqual_cost;
1606 : Cost cpu_per_tuple;
1607 : RangeTblEntry *rte;
1608 : QualCost exprcost;
1609 :
1610 : /* Should only be applied to base relations that are functions */
1611 : Assert(baserel->relid > 0);
1612 626 : rte = planner_rt_fetch(baserel->relid, root);
1613 : Assert(rte->rtekind == RTE_TABLEFUNC);
1614 :
1615 : /* Mark the path with the correct row estimate */
1616 626 : if (param_info)
1617 234 : path->rows = param_info->ppi_rows;
1618 : else
1619 392 : path->rows = baserel->rows;
1620 :
1621 : /*
1622 : * Estimate costs of executing the table func expression(s).
1623 : *
1624 : * XXX in principle we ought to charge tuplestore spill costs if the
1625 : * number of rows is large. However, given how phony our rowcount
1626 : * estimates for tablefuncs tend to be, there's not a lot of point in that
1627 : * refinement right now.
1628 : */
1629 626 : cost_qual_eval_node(&exprcost, (Node *) rte->tablefunc, root);
1630 :
1631 626 : startup_cost += exprcost.startup + exprcost.per_tuple;
1632 :
1633 : /* Add scanning CPU costs */
1634 626 : get_restriction_qual_cost(root, baserel, param_info, &qpqual_cost);
1635 :
1636 626 : startup_cost += qpqual_cost.startup;
1637 626 : cpu_per_tuple = cpu_tuple_cost + qpqual_cost.per_tuple;
1638 626 : run_cost += cpu_per_tuple * baserel->tuples;
1639 :
1640 : /* tlist eval costs are paid per output row, not per tuple scanned */
1641 626 : startup_cost += path->pathtarget->cost.startup;
1642 626 : run_cost += path->pathtarget->cost.per_tuple * path->rows;
1643 :
1644 626 : path->disabled_nodes = 0;
1645 626 : path->startup_cost = startup_cost;
1646 626 : path->total_cost = startup_cost + run_cost;
1647 626 : }
1648 :
1649 : /*
1650 : * cost_valuesscan
1651 : * Determines and returns the cost of scanning a VALUES RTE.
1652 : *
1653 : * 'baserel' is the relation to be scanned
1654 : * 'param_info' is the ParamPathInfo if this is a parameterized path, else NULL
1655 : */
1656 : void
1657 8246 : cost_valuesscan(Path *path, PlannerInfo *root,
1658 : RelOptInfo *baserel, ParamPathInfo *param_info)
1659 : {
1660 8246 : Cost startup_cost = 0;
1661 8246 : Cost run_cost = 0;
1662 : QualCost qpqual_cost;
1663 : Cost cpu_per_tuple;
1664 :
1665 : /* Should only be applied to base relations that are values lists */
1666 : Assert(baserel->relid > 0);
1667 : Assert(baserel->rtekind == RTE_VALUES);
1668 :
1669 : /* Mark the path with the correct row estimate */
1670 8246 : if (param_info)
1671 66 : path->rows = param_info->ppi_rows;
1672 : else
1673 8180 : path->rows = baserel->rows;
1674 :
1675 : /*
1676 : * For now, estimate list evaluation cost at one operator eval per list
1677 : * (probably pretty bogus, but is it worth being smarter?)
1678 : */
1679 8246 : cpu_per_tuple = cpu_operator_cost;
1680 :
1681 : /* Add scanning CPU costs */
1682 8246 : get_restriction_qual_cost(root, baserel, param_info, &qpqual_cost);
1683 :
1684 8246 : startup_cost += qpqual_cost.startup;
1685 8246 : cpu_per_tuple += cpu_tuple_cost + qpqual_cost.per_tuple;
1686 8246 : run_cost += cpu_per_tuple * baserel->tuples;
1687 :
1688 : /* tlist eval costs are paid per output row, not per tuple scanned */
1689 8246 : startup_cost += path->pathtarget->cost.startup;
1690 8246 : run_cost += path->pathtarget->cost.per_tuple * path->rows;
1691 :
1692 8246 : path->disabled_nodes = 0;
1693 8246 : path->startup_cost = startup_cost;
1694 8246 : path->total_cost = startup_cost + run_cost;
1695 8246 : }
1696 :
1697 : /*
1698 : * cost_ctescan
1699 : * Determines and returns the cost of scanning a CTE RTE.
1700 : *
1701 : * Note: this is used for both self-reference and regular CTEs; the
1702 : * possible cost differences are below the threshold of what we could
1703 : * estimate accurately anyway. Note that the costs of evaluating the
1704 : * referenced CTE query are added into the final plan as initplan costs,
1705 : * and should NOT be counted here.
1706 : */
1707 : void
1708 5094 : cost_ctescan(Path *path, PlannerInfo *root,
1709 : RelOptInfo *baserel, ParamPathInfo *param_info)
1710 : {
1711 5094 : Cost startup_cost = 0;
1712 5094 : Cost run_cost = 0;
1713 : QualCost qpqual_cost;
1714 : Cost cpu_per_tuple;
1715 :
1716 : /* Should only be applied to base relations that are CTEs */
1717 : Assert(baserel->relid > 0);
1718 : Assert(baserel->rtekind == RTE_CTE);
1719 :
1720 : /* Mark the path with the correct row estimate */
1721 5094 : if (param_info)
1722 0 : path->rows = param_info->ppi_rows;
1723 : else
1724 5094 : path->rows = baserel->rows;
1725 :
1726 : /* Charge one CPU tuple cost per row for tuplestore manipulation */
1727 5094 : cpu_per_tuple = cpu_tuple_cost;
1728 :
1729 : /* Add scanning CPU costs */
1730 5094 : get_restriction_qual_cost(root, baserel, param_info, &qpqual_cost);
1731 :
1732 5094 : startup_cost += qpqual_cost.startup;
1733 5094 : cpu_per_tuple += cpu_tuple_cost + qpqual_cost.per_tuple;
1734 5094 : run_cost += cpu_per_tuple * baserel->tuples;
1735 :
1736 : /* tlist eval costs are paid per output row, not per tuple scanned */
1737 5094 : startup_cost += path->pathtarget->cost.startup;
1738 5094 : run_cost += path->pathtarget->cost.per_tuple * path->rows;
1739 :
1740 5094 : path->disabled_nodes = 0;
1741 5094 : path->startup_cost = startup_cost;
1742 5094 : path->total_cost = startup_cost + run_cost;
1743 5094 : }
1744 :
1745 : /*
1746 : * cost_namedtuplestorescan
1747 : * Determines and returns the cost of scanning a named tuplestore.
1748 : */
1749 : void
1750 466 : cost_namedtuplestorescan(Path *path, PlannerInfo *root,
1751 : RelOptInfo *baserel, ParamPathInfo *param_info)
1752 : {
1753 466 : Cost startup_cost = 0;
1754 466 : Cost run_cost = 0;
1755 : QualCost qpqual_cost;
1756 : Cost cpu_per_tuple;
1757 :
1758 : /* Should only be applied to base relations that are Tuplestores */
1759 : Assert(baserel->relid > 0);
1760 : Assert(baserel->rtekind == RTE_NAMEDTUPLESTORE);
1761 :
1762 : /* Mark the path with the correct row estimate */
1763 466 : if (param_info)
1764 0 : path->rows = param_info->ppi_rows;
1765 : else
1766 466 : path->rows = baserel->rows;
1767 :
1768 : /* Charge one CPU tuple cost per row for tuplestore manipulation */
1769 466 : cpu_per_tuple = cpu_tuple_cost;
1770 :
1771 : /* Add scanning CPU costs */
1772 466 : get_restriction_qual_cost(root, baserel, param_info, &qpqual_cost);
1773 :
1774 466 : startup_cost += qpqual_cost.startup;
1775 466 : cpu_per_tuple += cpu_tuple_cost + qpqual_cost.per_tuple;
1776 466 : run_cost += cpu_per_tuple * baserel->tuples;
1777 :
1778 466 : path->disabled_nodes = 0;
1779 466 : path->startup_cost = startup_cost;
1780 466 : path->total_cost = startup_cost + run_cost;
1781 466 : }
1782 :
1783 : /*
1784 : * cost_resultscan
1785 : * Determines and returns the cost of scanning an RTE_RESULT relation.
1786 : */
1787 : void
1788 4280 : cost_resultscan(Path *path, PlannerInfo *root,
1789 : RelOptInfo *baserel, ParamPathInfo *param_info)
1790 : {
1791 4280 : Cost startup_cost = 0;
1792 4280 : Cost run_cost = 0;
1793 : QualCost qpqual_cost;
1794 : Cost cpu_per_tuple;
1795 :
1796 : /* Should only be applied to RTE_RESULT base relations */
1797 : Assert(baserel->relid > 0);
1798 : Assert(baserel->rtekind == RTE_RESULT);
1799 :
1800 : /* Mark the path with the correct row estimate */
1801 4280 : if (param_info)
1802 156 : path->rows = param_info->ppi_rows;
1803 : else
1804 4124 : path->rows = baserel->rows;
1805 :
1806 : /* We charge qual cost plus cpu_tuple_cost */
1807 4280 : get_restriction_qual_cost(root, baserel, param_info, &qpqual_cost);
1808 :
1809 4280 : startup_cost += qpqual_cost.startup;
1810 4280 : cpu_per_tuple = cpu_tuple_cost + qpqual_cost.per_tuple;
1811 4280 : run_cost += cpu_per_tuple * baserel->tuples;
1812 :
1813 4280 : path->disabled_nodes = 0;
1814 4280 : path->startup_cost = startup_cost;
1815 4280 : path->total_cost = startup_cost + run_cost;
1816 4280 : }
1817 :
1818 : /*
1819 : * cost_recursive_union
1820 : * Determines and returns the cost of performing a recursive union,
1821 : * and also the estimated output size.
1822 : *
1823 : * We are given Paths for the nonrecursive and recursive terms.
1824 : */
1825 : void
1826 1000 : cost_recursive_union(Path *runion, Path *nrterm, Path *rterm)
1827 : {
1828 : Cost startup_cost;
1829 : Cost total_cost;
1830 : double total_rows;
1831 :
1832 : /* We probably have decent estimates for the non-recursive term */
1833 1000 : startup_cost = nrterm->startup_cost;
1834 1000 : total_cost = nrterm->total_cost;
1835 1000 : total_rows = nrterm->rows;
1836 :
1837 : /*
1838 : * We arbitrarily assume that about 10 recursive iterations will be
1839 : * needed, and that we've managed to get a good fix on the cost and output
1840 : * size of each one of them. These are mighty shaky assumptions but it's
1841 : * hard to see how to do better.
1842 : */
1843 1000 : total_cost += 10 * rterm->total_cost;
1844 1000 : total_rows += 10 * rterm->rows;
1845 :
1846 : /*
1847 : * Also charge cpu_tuple_cost per row to account for the costs of
1848 : * manipulating the tuplestores. (We don't worry about possible
1849 : * spill-to-disk costs.)
1850 : */
1851 1000 : total_cost += cpu_tuple_cost * total_rows;
1852 :
1853 1000 : runion->disabled_nodes = nrterm->disabled_nodes + rterm->disabled_nodes;
1854 1000 : runion->startup_cost = startup_cost;
1855 1000 : runion->total_cost = total_cost;
1856 1000 : runion->rows = total_rows;
1857 1000 : runion->pathtarget->width = Max(nrterm->pathtarget->width,
1858 : rterm->pathtarget->width);
1859 1000 : }
1860 :
1861 : /*
1862 : * cost_tuplesort
1863 : * Determines and returns the cost of sorting a relation using tuplesort,
1864 : * not including the cost of reading the input data.
1865 : *
1866 : * If the total volume of data to sort is less than sort_mem, we will do
1867 : * an in-memory sort, which requires no I/O and about t*log2(t) tuple
1868 : * comparisons for t tuples.
1869 : *
1870 : * If the total volume exceeds sort_mem, we switch to a tape-style merge
1871 : * algorithm. There will still be about t*log2(t) tuple comparisons in
1872 : * total, but we will also need to write and read each tuple once per
1873 : * merge pass. We expect about ceil(logM(r)) merge passes where r is the
1874 : * number of initial runs formed and M is the merge order used by tuplesort.c.
1875 : * Since the average initial run should be about sort_mem, we have
1876 : * disk traffic = 2 * relsize * ceil(logM(p / sort_mem))
1877 : * cpu = comparison_cost * t * log2(t)
1878 : *
1879 : * If the sort is bounded (i.e., only the first k result tuples are needed)
1880 : * and k tuples can fit into sort_mem, we use a heap method that keeps only
1881 : * k tuples in the heap; this will require about t*log2(k) tuple comparisons.
1882 : *
1883 : * The disk traffic is assumed to be 3/4ths sequential and 1/4th random
1884 : * accesses (XXX can't we refine that guess?)
1885 : *
1886 : * By default, we charge two operator evals per tuple comparison, which should
1887 : * be in the right ballpark in most cases. The caller can tweak this by
1888 : * specifying nonzero comparison_cost; typically that's used for any extra
1889 : * work that has to be done to prepare the inputs to the comparison operators.
1890 : *
1891 : * 'tuples' is the number of tuples in the relation
1892 : * 'width' is the average tuple width in bytes
1893 : * 'comparison_cost' is the extra cost per comparison, if any
1894 : * 'sort_mem' is the number of kilobytes of work memory allowed for the sort
1895 : * 'limit_tuples' is the bound on the number of output tuples; -1 if no bound
1896 : */
1897 : static void
1898 1753248 : cost_tuplesort(Cost *startup_cost, Cost *run_cost,
1899 : double tuples, int width,
1900 : Cost comparison_cost, int sort_mem,
1901 : double limit_tuples)
1902 : {
1903 1753248 : double input_bytes = relation_byte_size(tuples, width);
1904 : double output_bytes;
1905 : double output_tuples;
1906 1753248 : int64 sort_mem_bytes = sort_mem * (int64) 1024;
1907 :
1908 : /*
1909 : * We want to be sure the cost of a sort is never estimated as zero, even
1910 : * if passed-in tuple count is zero. Besides, mustn't do log(0)...
1911 : */
1912 1753248 : if (tuples < 2.0)
1913 516000 : tuples = 2.0;
1914 :
1915 : /* Include the default cost-per-comparison */
1916 1753248 : comparison_cost += 2.0 * cpu_operator_cost;
1917 :
1918 : /* Do we have a useful LIMIT? */
1919 1753248 : if (limit_tuples > 0 && limit_tuples < tuples)
1920 : {
1921 1846 : output_tuples = limit_tuples;
1922 1846 : output_bytes = relation_byte_size(output_tuples, width);
1923 : }
1924 : else
1925 : {
1926 1751402 : output_tuples = tuples;
1927 1751402 : output_bytes = input_bytes;
1928 : }
1929 :
1930 1753248 : if (output_bytes > sort_mem_bytes)
1931 : {
1932 : /*
1933 : * We'll have to use a disk-based sort of all the tuples
1934 : */
1935 20224 : double npages = ceil(input_bytes / BLCKSZ);
1936 20224 : double nruns = input_bytes / sort_mem_bytes;
1937 20224 : double mergeorder = tuplesort_merge_order(sort_mem_bytes);
1938 : double log_runs;
1939 : double npageaccesses;
1940 :
1941 : /*
1942 : * CPU costs
1943 : *
1944 : * Assume about N log2 N comparisons
1945 : */
1946 20224 : *startup_cost = comparison_cost * tuples * LOG2(tuples);
1947 :
1948 : /* Disk costs */
1949 :
1950 : /* Compute logM(r) as log(r) / log(M) */
1951 20224 : if (nruns > mergeorder)
1952 4856 : log_runs = ceil(log(nruns) / log(mergeorder));
1953 : else
1954 15368 : log_runs = 1.0;
1955 20224 : npageaccesses = 2.0 * npages * log_runs;
1956 : /* Assume 3/4ths of accesses are sequential, 1/4th are not */
1957 20224 : *startup_cost += npageaccesses *
1958 20224 : (seq_page_cost * 0.75 + random_page_cost * 0.25);
1959 : }
1960 1733024 : else if (tuples > 2 * output_tuples || input_bytes > sort_mem_bytes)
1961 : {
1962 : /*
1963 : * We'll use a bounded heap-sort keeping just K tuples in memory, for
1964 : * a total number of tuple comparisons of N log2 K; but the constant
1965 : * factor is a bit higher than for quicksort. Tweak it so that the
1966 : * cost curve is continuous at the crossover point.
1967 : */
1968 1376 : *startup_cost = comparison_cost * tuples * LOG2(2.0 * output_tuples);
1969 : }
1970 : else
1971 : {
1972 : /* We'll use plain quicksort on all the input tuples */
1973 1731648 : *startup_cost = comparison_cost * tuples * LOG2(tuples);
1974 : }
1975 :
1976 : /*
1977 : * Also charge a small amount (arbitrarily set equal to operator cost) per
1978 : * extracted tuple. We don't charge cpu_tuple_cost because a Sort node
1979 : * doesn't do qual-checking or projection, so it has less overhead than
1980 : * most plan nodes. Note it's correct to use tuples not output_tuples
1981 : * here --- the upper LIMIT will pro-rate the run cost so we'd be double
1982 : * counting the LIMIT otherwise.
1983 : */
1984 1753248 : *run_cost = cpu_operator_cost * tuples;
1985 1753248 : }
1986 :
1987 : /*
1988 : * cost_incremental_sort
1989 : * Determines and returns the cost of sorting a relation incrementally, when
1990 : * the input path is presorted by a prefix of the pathkeys.
1991 : *
1992 : * 'presorted_keys' is the number of leading pathkeys by which the input path
1993 : * is sorted.
1994 : *
1995 : * We estimate the number of groups into which the relation is divided by the
1996 : * leading pathkeys, and then calculate the cost of sorting a single group
1997 : * with tuplesort using cost_tuplesort().
1998 : */
1999 : void
2000 11734 : cost_incremental_sort(Path *path,
2001 : PlannerInfo *root, List *pathkeys, int presorted_keys,
2002 : int input_disabled_nodes,
2003 : Cost input_startup_cost, Cost input_total_cost,
2004 : double input_tuples, int width, Cost comparison_cost, int sort_mem,
2005 : double limit_tuples)
2006 : {
2007 : Cost startup_cost,
2008 : run_cost,
2009 11734 : input_run_cost = input_total_cost - input_startup_cost;
2010 : double group_tuples,
2011 : input_groups;
2012 : Cost group_startup_cost,
2013 : group_run_cost,
2014 : group_input_run_cost;
2015 11734 : List *presortedExprs = NIL;
2016 : ListCell *l;
2017 11734 : bool unknown_varno = false;
2018 :
2019 : Assert(presorted_keys > 0 && presorted_keys < list_length(pathkeys));
2020 :
2021 : /*
2022 : * We want to be sure the cost of a sort is never estimated as zero, even
2023 : * if passed-in tuple count is zero. Besides, mustn't do log(0)...
2024 : */
2025 11734 : if (input_tuples < 2.0)
2026 6760 : input_tuples = 2.0;
2027 :
2028 : /* Default estimate of number of groups, capped to one group per row. */
2029 11734 : input_groups = Min(input_tuples, DEFAULT_NUM_DISTINCT);
2030 :
2031 : /*
2032 : * Extract presorted keys as list of expressions.
2033 : *
2034 : * We need to be careful about Vars containing "varno 0" which might have
2035 : * been introduced by generate_append_tlist, which would confuse
2036 : * estimate_num_groups (in fact it'd fail for such expressions). See
2037 : * recurse_set_operations which has to deal with the same issue.
2038 : *
2039 : * Unlike recurse_set_operations we can't access the original target list
2040 : * here, and even if we could it's not very clear how useful would that be
2041 : * for a set operation combining multiple tables. So we simply detect if
2042 : * there are any expressions with "varno 0" and use the default
2043 : * DEFAULT_NUM_DISTINCT in that case.
2044 : *
2045 : * We might also use either 1.0 (a single group) or input_tuples (each row
2046 : * being a separate group), pretty much the worst and best case for
2047 : * incremental sort. But those are extreme cases and using something in
2048 : * between seems reasonable. Furthermore, generate_append_tlist is used
2049 : * for set operations, which are likely to produce mostly unique output
2050 : * anyway - from that standpoint the DEFAULT_NUM_DISTINCT is defensive
2051 : * while maintaining lower startup cost.
2052 : */
2053 11830 : foreach(l, pathkeys)
2054 : {
2055 11830 : PathKey *key = (PathKey *) lfirst(l);
2056 11830 : EquivalenceMember *member = (EquivalenceMember *)
2057 11830 : linitial(key->pk_eclass->ec_members);
2058 :
2059 : /*
2060 : * Check if the expression contains Var with "varno 0" so that we
2061 : * don't call estimate_num_groups in that case.
2062 : */
2063 11830 : if (bms_is_member(0, pull_varnos(root, (Node *) member->em_expr)))
2064 : {
2065 10 : unknown_varno = true;
2066 10 : break;
2067 : }
2068 :
2069 : /* expression not containing any Vars with "varno 0" */
2070 11820 : presortedExprs = lappend(presortedExprs, member->em_expr);
2071 :
2072 11820 : if (foreach_current_index(l) + 1 >= presorted_keys)
2073 11724 : break;
2074 : }
2075 :
2076 : /* Estimate the number of groups with equal presorted keys. */
2077 11734 : if (!unknown_varno)
2078 11724 : input_groups = estimate_num_groups(root, presortedExprs, input_tuples,
2079 : NULL, NULL);
2080 :
2081 11734 : group_tuples = input_tuples / input_groups;
2082 11734 : group_input_run_cost = input_run_cost / input_groups;
2083 :
2084 : /*
2085 : * Estimate the average cost of sorting of one group where presorted keys
2086 : * are equal.
2087 : */
2088 11734 : cost_tuplesort(&group_startup_cost, &group_run_cost,
2089 : group_tuples, width, comparison_cost, sort_mem,
2090 : limit_tuples);
2091 :
2092 : /*
2093 : * Startup cost of incremental sort is the startup cost of its first group
2094 : * plus the cost of its input.
2095 : */
2096 11734 : startup_cost = group_startup_cost + input_startup_cost +
2097 : group_input_run_cost;
2098 :
2099 : /*
2100 : * After we started producing tuples from the first group, the cost of
2101 : * producing all the tuples is given by the cost to finish processing this
2102 : * group, plus the total cost to process the remaining groups, plus the
2103 : * remaining cost of input.
2104 : */
2105 11734 : run_cost = group_run_cost + (group_run_cost + group_startup_cost) *
2106 11734 : (input_groups - 1) + group_input_run_cost * (input_groups - 1);
2107 :
2108 : /*
2109 : * Incremental sort adds some overhead by itself. Firstly, it has to
2110 : * detect the sort groups. This is roughly equal to one extra copy and
2111 : * comparison per tuple.
2112 : */
2113 11734 : run_cost += (cpu_tuple_cost + comparison_cost) * input_tuples;
2114 :
2115 : /*
2116 : * Additionally, we charge double cpu_tuple_cost for each input group to
2117 : * account for the tuplesort_reset that's performed after each group.
2118 : */
2119 11734 : run_cost += 2.0 * cpu_tuple_cost * input_groups;
2120 :
2121 11734 : path->rows = input_tuples;
2122 :
2123 : /* should not generate these paths when enable_incremental_sort=false */
2124 : Assert(enable_incremental_sort);
2125 11734 : path->disabled_nodes = input_disabled_nodes;
2126 :
2127 11734 : path->startup_cost = startup_cost;
2128 11734 : path->total_cost = startup_cost + run_cost;
2129 11734 : }
2130 :
2131 : /*
2132 : * cost_sort
2133 : * Determines and returns the cost of sorting a relation, including
2134 : * the cost of reading the input data.
2135 : *
2136 : * NOTE: some callers currently pass NIL for pathkeys because they
2137 : * can't conveniently supply the sort keys. Since this routine doesn't
2138 : * currently do anything with pathkeys anyway, that doesn't matter...
2139 : * but if it ever does, it should react gracefully to lack of key data.
2140 : * (Actually, the thing we'd most likely be interested in is just the number
2141 : * of sort keys, which all callers *could* supply.)
2142 : */
2143 : void
2144 1741514 : cost_sort(Path *path, PlannerInfo *root,
2145 : List *pathkeys, int input_disabled_nodes,
2146 : Cost input_cost, double tuples, int width,
2147 : Cost comparison_cost, int sort_mem,
2148 : double limit_tuples)
2149 :
2150 : {
2151 : Cost startup_cost;
2152 : Cost run_cost;
2153 :
2154 1741514 : cost_tuplesort(&startup_cost, &run_cost,
2155 : tuples, width,
2156 : comparison_cost, sort_mem,
2157 : limit_tuples);
2158 :
2159 1741514 : startup_cost += input_cost;
2160 :
2161 1741514 : path->rows = tuples;
2162 1741514 : path->disabled_nodes = input_disabled_nodes + (enable_sort ? 0 : 1);
2163 1741514 : path->startup_cost = startup_cost;
2164 1741514 : path->total_cost = startup_cost + run_cost;
2165 1741514 : }
2166 :
2167 : /*
2168 : * append_nonpartial_cost
2169 : * Estimate the cost of the non-partial paths in a Parallel Append.
2170 : * The non-partial paths are assumed to be the first "numpaths" paths
2171 : * from the subpaths list, and to be in order of decreasing cost.
2172 : */
2173 : static Cost
2174 18330 : append_nonpartial_cost(List *subpaths, int numpaths, int parallel_workers)
2175 : {
2176 : Cost *costarr;
2177 : int arrlen;
2178 : ListCell *l;
2179 : ListCell *cell;
2180 : int path_index;
2181 : int min_index;
2182 : int max_index;
2183 :
2184 18330 : if (numpaths == 0)
2185 14190 : return 0;
2186 :
2187 : /*
2188 : * Array length is number of workers or number of relevant paths,
2189 : * whichever is less.
2190 : */
2191 4140 : arrlen = Min(parallel_workers, numpaths);
2192 4140 : costarr = (Cost *) palloc(sizeof(Cost) * arrlen);
2193 :
2194 : /* The first few paths will each be claimed by a different worker. */
2195 4140 : path_index = 0;
2196 11946 : foreach(cell, subpaths)
2197 : {
2198 8650 : Path *subpath = (Path *) lfirst(cell);
2199 :
2200 8650 : if (path_index == arrlen)
2201 844 : break;
2202 7806 : costarr[path_index++] = subpath->total_cost;
2203 : }
2204 :
2205 : /*
2206 : * Since subpaths are sorted by decreasing cost, the last one will have
2207 : * the minimum cost.
2208 : */
2209 4140 : min_index = arrlen - 1;
2210 :
2211 : /*
2212 : * For each of the remaining subpaths, add its cost to the array element
2213 : * with minimum cost.
2214 : */
2215 4622 : for_each_cell(l, subpaths, cell)
2216 : {
2217 1028 : Path *subpath = (Path *) lfirst(l);
2218 :
2219 : /* Consider only the non-partial paths */
2220 1028 : if (path_index++ == numpaths)
2221 546 : break;
2222 :
2223 482 : costarr[min_index] += subpath->total_cost;
2224 :
2225 : /* Update the new min cost array index */
2226 482 : min_index = 0;
2227 1482 : for (int i = 0; i < arrlen; i++)
2228 : {
2229 1000 : if (costarr[i] < costarr[min_index])
2230 202 : min_index = i;
2231 : }
2232 : }
2233 :
2234 : /* Return the highest cost from the array */
2235 4140 : max_index = 0;
2236 11946 : for (int i = 0; i < arrlen; i++)
2237 : {
2238 7806 : if (costarr[i] > costarr[max_index])
2239 182 : max_index = i;
2240 : }
2241 :
2242 4140 : return costarr[max_index];
2243 : }
2244 :
2245 : /*
2246 : * cost_append
2247 : * Determines and returns the cost of an Append node.
2248 : */
2249 : void
2250 54580 : cost_append(AppendPath *apath, PlannerInfo *root)
2251 : {
2252 : ListCell *l;
2253 :
2254 54580 : apath->path.disabled_nodes = 0;
2255 54580 : apath->path.startup_cost = 0;
2256 54580 : apath->path.total_cost = 0;
2257 54580 : apath->path.rows = 0;
2258 :
2259 54580 : if (apath->subpaths == NIL)
2260 1738 : return;
2261 :
2262 52842 : if (!apath->path.parallel_aware)
2263 : {
2264 34512 : List *pathkeys = apath->path.pathkeys;
2265 :
2266 34512 : if (pathkeys == NIL)
2267 : {
2268 32298 : Path *firstsubpath = (Path *) linitial(apath->subpaths);
2269 :
2270 : /*
2271 : * For an unordered, non-parallel-aware Append we take the startup
2272 : * cost as the startup cost of the first subpath.
2273 : */
2274 32298 : apath->path.startup_cost = firstsubpath->startup_cost;
2275 :
2276 : /*
2277 : * Compute rows, number of disabled nodes, and total cost as sums
2278 : * of underlying subplan values.
2279 : */
2280 124900 : foreach(l, apath->subpaths)
2281 : {
2282 92602 : Path *subpath = (Path *) lfirst(l);
2283 :
2284 92602 : apath->path.rows += subpath->rows;
2285 92602 : apath->path.disabled_nodes += subpath->disabled_nodes;
2286 92602 : apath->path.total_cost += subpath->total_cost;
2287 : }
2288 : }
2289 : else
2290 : {
2291 : /*
2292 : * For an ordered, non-parallel-aware Append we take the startup
2293 : * cost as the sum of the subpath startup costs. This ensures
2294 : * that we don't underestimate the startup cost when a query's
2295 : * LIMIT is such that several of the children have to be run to
2296 : * satisfy it. This might be overkill --- another plausible hack
2297 : * would be to take the Append's startup cost as the maximum of
2298 : * the child startup costs. But we don't want to risk believing
2299 : * that an ORDER BY LIMIT query can be satisfied at small cost
2300 : * when the first child has small startup cost but later ones
2301 : * don't. (If we had the ability to deal with nonlinear cost
2302 : * interpolation for partial retrievals, we would not need to be
2303 : * so conservative about this.)
2304 : *
2305 : * This case is also different from the above in that we have to
2306 : * account for possibly injecting sorts into subpaths that aren't
2307 : * natively ordered.
2308 : */
2309 8588 : foreach(l, apath->subpaths)
2310 : {
2311 6374 : Path *subpath = (Path *) lfirst(l);
2312 : int presorted_keys;
2313 : Path sort_path; /* dummy for result of
2314 : * cost_sort/cost_incremental_sort */
2315 :
2316 6374 : if (!pathkeys_count_contained_in(pathkeys, subpath->pathkeys,
2317 : &presorted_keys))
2318 : {
2319 : /*
2320 : * We'll need to insert a Sort node, so include costs for
2321 : * that. We choose to use incremental sort if it is
2322 : * enabled and there are presorted keys; otherwise we use
2323 : * full sort.
2324 : *
2325 : * We can use the parent's LIMIT if any, since we
2326 : * certainly won't pull more than that many tuples from
2327 : * any child.
2328 : */
2329 56 : if (enable_incremental_sort && presorted_keys > 0)
2330 : {
2331 12 : cost_incremental_sort(&sort_path,
2332 : root,
2333 : pathkeys,
2334 : presorted_keys,
2335 : subpath->disabled_nodes,
2336 : subpath->startup_cost,
2337 : subpath->total_cost,
2338 : subpath->rows,
2339 12 : subpath->pathtarget->width,
2340 : 0.0,
2341 : work_mem,
2342 : apath->limit_tuples);
2343 : }
2344 : else
2345 : {
2346 44 : cost_sort(&sort_path,
2347 : root,
2348 : pathkeys,
2349 : subpath->disabled_nodes,
2350 : subpath->total_cost,
2351 : subpath->rows,
2352 44 : subpath->pathtarget->width,
2353 : 0.0,
2354 : work_mem,
2355 : apath->limit_tuples);
2356 : }
2357 :
2358 56 : subpath = &sort_path;
2359 : }
2360 :
2361 6374 : apath->path.rows += subpath->rows;
2362 6374 : apath->path.disabled_nodes += subpath->disabled_nodes;
2363 6374 : apath->path.startup_cost += subpath->startup_cost;
2364 6374 : apath->path.total_cost += subpath->total_cost;
2365 : }
2366 : }
2367 : }
2368 : else /* parallel-aware */
2369 : {
2370 18330 : int i = 0;
2371 18330 : double parallel_divisor = get_parallel_divisor(&apath->path);
2372 :
2373 : /* Parallel-aware Append never produces ordered output. */
2374 : Assert(apath->path.pathkeys == NIL);
2375 :
2376 : /* Calculate startup cost. */
2377 71526 : foreach(l, apath->subpaths)
2378 : {
2379 53196 : Path *subpath = (Path *) lfirst(l);
2380 :
2381 : /*
2382 : * Append will start returning tuples when the child node having
2383 : * lowest startup cost is done setting up. We consider only the
2384 : * first few subplans that immediately get a worker assigned.
2385 : */
2386 53196 : if (i == 0)
2387 18330 : apath->path.startup_cost = subpath->startup_cost;
2388 34866 : else if (i < apath->path.parallel_workers)
2389 17772 : apath->path.startup_cost = Min(apath->path.startup_cost,
2390 : subpath->startup_cost);
2391 :
2392 : /*
2393 : * Apply parallel divisor to subpaths. Scale the number of rows
2394 : * for each partial subpath based on the ratio of the parallel
2395 : * divisor originally used for the subpath to the one we adopted.
2396 : * Also add the cost of partial paths to the total cost, but
2397 : * ignore non-partial paths for now.
2398 : */
2399 53196 : if (i < apath->first_partial_path)
2400 8288 : apath->path.rows += subpath->rows / parallel_divisor;
2401 : else
2402 : {
2403 : double subpath_parallel_divisor;
2404 :
2405 44908 : subpath_parallel_divisor = get_parallel_divisor(subpath);
2406 44908 : apath->path.rows += subpath->rows * (subpath_parallel_divisor /
2407 : parallel_divisor);
2408 44908 : apath->path.total_cost += subpath->total_cost;
2409 : }
2410 :
2411 53196 : apath->path.disabled_nodes += subpath->disabled_nodes;
2412 53196 : apath->path.rows = clamp_row_est(apath->path.rows);
2413 :
2414 53196 : i++;
2415 : }
2416 :
2417 : /* Add cost for non-partial subpaths. */
2418 18330 : apath->path.total_cost +=
2419 18330 : append_nonpartial_cost(apath->subpaths,
2420 : apath->first_partial_path,
2421 : apath->path.parallel_workers);
2422 : }
2423 :
2424 : /*
2425 : * Although Append does not do any selection or projection, it's not free;
2426 : * add a small per-tuple overhead.
2427 : */
2428 52842 : apath->path.total_cost +=
2429 52842 : cpu_tuple_cost * APPEND_CPU_COST_MULTIPLIER * apath->path.rows;
2430 : }
2431 :
2432 : /*
2433 : * cost_merge_append
2434 : * Determines and returns the cost of a MergeAppend node.
2435 : *
2436 : * MergeAppend merges several pre-sorted input streams, using a heap that
2437 : * at any given instant holds the next tuple from each stream. If there
2438 : * are N streams, we need about N*log2(N) tuple comparisons to construct
2439 : * the heap at startup, and then for each output tuple, about log2(N)
2440 : * comparisons to replace the top entry.
2441 : *
2442 : * (The effective value of N will drop once some of the input streams are
2443 : * exhausted, but it seems unlikely to be worth trying to account for that.)
2444 : *
2445 : * The heap is never spilled to disk, since we assume N is not very large.
2446 : * So this is much simpler than cost_sort.
2447 : *
2448 : * As in cost_sort, we charge two operator evals per tuple comparison.
2449 : *
2450 : * 'pathkeys' is a list of sort keys
2451 : * 'n_streams' is the number of input streams
2452 : * 'input_disabled_nodes' is the sum of the input streams' disabled node counts
2453 : * 'input_startup_cost' is the sum of the input streams' startup costs
2454 : * 'input_total_cost' is the sum of the input streams' total costs
2455 : * 'tuples' is the number of tuples in all the streams
2456 : */
2457 : void
2458 4220 : cost_merge_append(Path *path, PlannerInfo *root,
2459 : List *pathkeys, int n_streams,
2460 : int input_disabled_nodes,
2461 : Cost input_startup_cost, Cost input_total_cost,
2462 : double tuples)
2463 : {
2464 4220 : Cost startup_cost = 0;
2465 4220 : Cost run_cost = 0;
2466 : Cost comparison_cost;
2467 : double N;
2468 : double logN;
2469 :
2470 : /*
2471 : * Avoid log(0)...
2472 : */
2473 4220 : N = (n_streams < 2) ? 2.0 : (double) n_streams;
2474 4220 : logN = LOG2(N);
2475 :
2476 : /* Assumed cost per tuple comparison */
2477 4220 : comparison_cost = 2.0 * cpu_operator_cost;
2478 :
2479 : /* Heap creation cost */
2480 4220 : startup_cost += comparison_cost * N * logN;
2481 :
2482 : /* Per-tuple heap maintenance cost */
2483 4220 : run_cost += tuples * comparison_cost * logN;
2484 :
2485 : /*
2486 : * Although MergeAppend does not do any selection or projection, it's not
2487 : * free; add a small per-tuple overhead.
2488 : */
2489 4220 : run_cost += cpu_tuple_cost * APPEND_CPU_COST_MULTIPLIER * tuples;
2490 :
2491 4220 : path->disabled_nodes = input_disabled_nodes;
2492 4220 : path->startup_cost = startup_cost + input_startup_cost;
2493 4220 : path->total_cost = startup_cost + run_cost + input_total_cost;
2494 4220 : }
2495 :
2496 : /*
2497 : * cost_material
2498 : * Determines and returns the cost of materializing a relation, including
2499 : * the cost of reading the input data.
2500 : *
2501 : * If the total volume of data to materialize exceeds work_mem, we will need
2502 : * to write it to disk, so the cost is much higher in that case.
2503 : *
2504 : * Note that here we are estimating the costs for the first scan of the
2505 : * relation, so the materialization is all overhead --- any savings will
2506 : * occur only on rescan, which is estimated in cost_rescan.
2507 : */
2508 : void
2509 532066 : cost_material(Path *path,
2510 : int input_disabled_nodes,
2511 : Cost input_startup_cost, Cost input_total_cost,
2512 : double tuples, int width)
2513 : {
2514 532066 : Cost startup_cost = input_startup_cost;
2515 532066 : Cost run_cost = input_total_cost - input_startup_cost;
2516 532066 : double nbytes = relation_byte_size(tuples, width);
2517 532066 : double work_mem_bytes = work_mem * (Size) 1024;
2518 :
2519 532066 : path->rows = tuples;
2520 :
2521 : /*
2522 : * Whether spilling or not, charge 2x cpu_operator_cost per tuple to
2523 : * reflect bookkeeping overhead. (This rate must be more than what
2524 : * cost_rescan charges for materialize, ie, cpu_operator_cost per tuple;
2525 : * if it is exactly the same then there will be a cost tie between
2526 : * nestloop with A outer, materialized B inner and nestloop with B outer,
2527 : * materialized A inner. The extra cost ensures we'll prefer
2528 : * materializing the smaller rel.) Note that this is normally a good deal
2529 : * less than cpu_tuple_cost; which is OK because a Material plan node
2530 : * doesn't do qual-checking or projection, so it's got less overhead than
2531 : * most plan nodes.
2532 : */
2533 532066 : run_cost += 2 * cpu_operator_cost * tuples;
2534 :
2535 : /*
2536 : * If we will spill to disk, charge at the rate of seq_page_cost per page.
2537 : * This cost is assumed to be evenly spread through the plan run phase,
2538 : * which isn't exactly accurate but our cost model doesn't allow for
2539 : * nonuniform costs within the run phase.
2540 : */
2541 532066 : if (nbytes > work_mem_bytes)
2542 : {
2543 5464 : double npages = ceil(nbytes / BLCKSZ);
2544 :
2545 5464 : run_cost += seq_page_cost * npages;
2546 : }
2547 :
2548 532066 : path->disabled_nodes = input_disabled_nodes + (enable_material ? 0 : 1);
2549 532066 : path->startup_cost = startup_cost;
2550 532066 : path->total_cost = startup_cost + run_cost;
2551 532066 : }
2552 :
2553 : /*
2554 : * cost_memoize_rescan
2555 : * Determines the estimated cost of rescanning a Memoize node.
2556 : *
2557 : * In order to estimate this, we must gain knowledge of how often we expect to
2558 : * be called and how many distinct sets of parameters we are likely to be
2559 : * called with. If we expect a good cache hit ratio, then we can set our
2560 : * costs to account for that hit ratio, plus a little bit of cost for the
2561 : * caching itself. Caching will not work out well if we expect to be called
2562 : * with too many distinct parameter values. The worst-case here is that we
2563 : * never see any parameter value twice, in which case we'd never get a cache
2564 : * hit and caching would be a complete waste of effort.
2565 : */
2566 : static void
2567 299506 : cost_memoize_rescan(PlannerInfo *root, MemoizePath *mpath,
2568 : Cost *rescan_startup_cost, Cost *rescan_total_cost)
2569 : {
2570 : EstimationInfo estinfo;
2571 : ListCell *lc;
2572 299506 : Cost input_startup_cost = mpath->subpath->startup_cost;
2573 299506 : Cost input_total_cost = mpath->subpath->total_cost;
2574 299506 : double tuples = mpath->subpath->rows;
2575 299506 : double calls = mpath->calls;
2576 299506 : int width = mpath->subpath->pathtarget->width;
2577 :
2578 : double hash_mem_bytes;
2579 : double est_entry_bytes;
2580 : double est_cache_entries;
2581 : double ndistinct;
2582 : double evict_ratio;
2583 : double hit_ratio;
2584 : Cost startup_cost;
2585 : Cost total_cost;
2586 :
2587 : /* available cache space */
2588 299506 : hash_mem_bytes = get_hash_memory_limit();
2589 :
2590 : /*
2591 : * Set the number of bytes each cache entry should consume in the cache.
2592 : * To provide us with better estimations on how many cache entries we can
2593 : * store at once, we make a call to the executor here to ask it what
2594 : * memory overheads there are for a single cache entry.
2595 : */
2596 299506 : est_entry_bytes = relation_byte_size(tuples, width) +
2597 299506 : ExecEstimateCacheEntryOverheadBytes(tuples);
2598 :
2599 : /* include the estimated width for the cache keys */
2600 643636 : foreach(lc, mpath->param_exprs)
2601 344130 : est_entry_bytes += get_expr_width(root, (Node *) lfirst(lc));
2602 :
2603 : /* estimate on the upper limit of cache entries we can hold at once */
2604 299506 : est_cache_entries = floor(hash_mem_bytes / est_entry_bytes);
2605 :
2606 : /* estimate on the distinct number of parameter values */
2607 299506 : ndistinct = estimate_num_groups(root, mpath->param_exprs, calls, NULL,
2608 : &estinfo);
2609 :
2610 : /*
2611 : * When the estimation fell back on using a default value, it's a bit too
2612 : * risky to assume that it's ok to use a Memoize node. The use of a
2613 : * default could cause us to use a Memoize node when it's really
2614 : * inappropriate to do so. If we see that this has been done, then we'll
2615 : * assume that every call will have unique parameters, which will almost
2616 : * certainly mean a MemoizePath will never survive add_path().
2617 : */
2618 299506 : if ((estinfo.flags & SELFLAG_USED_DEFAULT) != 0)
2619 16610 : ndistinct = calls;
2620 :
2621 : /*
2622 : * Since we've already estimated the maximum number of entries we can
2623 : * store at once and know the estimated number of distinct values we'll be
2624 : * called with, we'll take this opportunity to set the path's est_entries.
2625 : * This will ultimately determine the hash table size that the executor
2626 : * will use. If we leave this at zero, the executor will just choose the
2627 : * size itself. Really this is not the right place to do this, but it's
2628 : * convenient since everything is already calculated.
2629 : */
2630 299506 : mpath->est_entries = Min(Min(ndistinct, est_cache_entries),
2631 : PG_UINT32_MAX);
2632 :
2633 : /*
2634 : * When the number of distinct parameter values is above the amount we can
2635 : * store in the cache, then we'll have to evict some entries from the
2636 : * cache. This is not free. Here we estimate how often we'll incur the
2637 : * cost of that eviction.
2638 : */
2639 299506 : evict_ratio = 1.0 - Min(est_cache_entries, ndistinct) / ndistinct;
2640 :
2641 : /*
2642 : * In order to estimate how costly a single scan will be, we need to
2643 : * attempt to estimate what the cache hit ratio will be. To do that we
2644 : * must look at how many scans are estimated in total for this node and
2645 : * how many of those scans we expect to get a cache hit.
2646 : */
2647 599012 : hit_ratio = ((calls - ndistinct) / calls) *
2648 299506 : (est_cache_entries / Max(ndistinct, est_cache_entries));
2649 :
2650 : Assert(hit_ratio >= 0 && hit_ratio <= 1.0);
2651 :
2652 : /*
2653 : * Set the total_cost accounting for the expected cache hit ratio. We
2654 : * also add on a cpu_operator_cost to account for a cache lookup. This
2655 : * will happen regardless of whether it's a cache hit or not.
2656 : */
2657 299506 : total_cost = input_total_cost * (1.0 - hit_ratio) + cpu_operator_cost;
2658 :
2659 : /* Now adjust the total cost to account for cache evictions */
2660 :
2661 : /* Charge a cpu_tuple_cost for evicting the actual cache entry */
2662 299506 : total_cost += cpu_tuple_cost * evict_ratio;
2663 :
2664 : /*
2665 : * Charge a 10th of cpu_operator_cost to evict every tuple in that entry.
2666 : * The per-tuple eviction is really just a pfree, so charging a whole
2667 : * cpu_operator_cost seems a little excessive.
2668 : */
2669 299506 : total_cost += cpu_operator_cost / 10.0 * evict_ratio * tuples;
2670 :
2671 : /*
2672 : * Now adjust for storing things in the cache, since that's not free
2673 : * either. Everything must go in the cache. We don't proportion this
2674 : * over any ratio, just apply it once for the scan. We charge a
2675 : * cpu_tuple_cost for the creation of the cache entry and also a
2676 : * cpu_operator_cost for each tuple we expect to cache.
2677 : */
2678 299506 : total_cost += cpu_tuple_cost + cpu_operator_cost * tuples;
2679 :
2680 : /*
2681 : * Getting the first row must be also be proportioned according to the
2682 : * expected cache hit ratio.
2683 : */
2684 299506 : startup_cost = input_startup_cost * (1.0 - hit_ratio);
2685 :
2686 : /*
2687 : * Additionally we charge a cpu_tuple_cost to account for cache lookups,
2688 : * which we'll do regardless of whether it was a cache hit or not.
2689 : */
2690 299506 : startup_cost += cpu_tuple_cost;
2691 :
2692 299506 : *rescan_startup_cost = startup_cost;
2693 299506 : *rescan_total_cost = total_cost;
2694 299506 : }
2695 :
2696 : /*
2697 : * cost_agg
2698 : * Determines and returns the cost of performing an Agg plan node,
2699 : * including the cost of its input.
2700 : *
2701 : * aggcosts can be NULL when there are no actual aggregate functions (i.e.,
2702 : * we are using a hashed Agg node just to do grouping).
2703 : *
2704 : * Note: when aggstrategy == AGG_SORTED, caller must ensure that input costs
2705 : * are for appropriately-sorted input.
2706 : */
2707 : void
2708 69476 : cost_agg(Path *path, PlannerInfo *root,
2709 : AggStrategy aggstrategy, const AggClauseCosts *aggcosts,
2710 : int numGroupCols, double numGroups,
2711 : List *quals,
2712 : int disabled_nodes,
2713 : Cost input_startup_cost, Cost input_total_cost,
2714 : double input_tuples, double input_width)
2715 : {
2716 : double output_tuples;
2717 : Cost startup_cost;
2718 : Cost total_cost;
2719 69476 : const AggClauseCosts dummy_aggcosts = {0};
2720 :
2721 : /* Use all-zero per-aggregate costs if NULL is passed */
2722 69476 : if (aggcosts == NULL)
2723 : {
2724 : Assert(aggstrategy == AGG_HASHED);
2725 12588 : aggcosts = &dummy_aggcosts;
2726 : }
2727 :
2728 : /*
2729 : * The transCost.per_tuple component of aggcosts should be charged once
2730 : * per input tuple, corresponding to the costs of evaluating the aggregate
2731 : * transfns and their input expressions. The finalCost.per_tuple component
2732 : * is charged once per output tuple, corresponding to the costs of
2733 : * evaluating the finalfns. Startup costs are of course charged but once.
2734 : *
2735 : * If we are grouping, we charge an additional cpu_operator_cost per
2736 : * grouping column per input tuple for grouping comparisons.
2737 : *
2738 : * We will produce a single output tuple if not grouping, and a tuple per
2739 : * group otherwise. We charge cpu_tuple_cost for each output tuple.
2740 : *
2741 : * Note: in this cost model, AGG_SORTED and AGG_HASHED have exactly the
2742 : * same total CPU cost, but AGG_SORTED has lower startup cost. If the
2743 : * input path is already sorted appropriately, AGG_SORTED should be
2744 : * preferred (since it has no risk of memory overflow). This will happen
2745 : * as long as the computed total costs are indeed exactly equal --- but if
2746 : * there's roundoff error we might do the wrong thing. So be sure that
2747 : * the computations below form the same intermediate values in the same
2748 : * order.
2749 : */
2750 69476 : if (aggstrategy == AGG_PLAIN)
2751 : {
2752 37230 : startup_cost = input_total_cost;
2753 37230 : startup_cost += aggcosts->transCost.startup;
2754 37230 : startup_cost += aggcosts->transCost.per_tuple * input_tuples;
2755 37230 : startup_cost += aggcosts->finalCost.startup;
2756 37230 : startup_cost += aggcosts->finalCost.per_tuple;
2757 : /* we aren't grouping */
2758 37230 : total_cost = startup_cost + cpu_tuple_cost;
2759 37230 : output_tuples = 1;
2760 : }
2761 32246 : else if (aggstrategy == AGG_SORTED || aggstrategy == AGG_MIXED)
2762 : {
2763 : /* Here we are able to deliver output on-the-fly */
2764 11266 : startup_cost = input_startup_cost;
2765 11266 : total_cost = input_total_cost;
2766 11266 : if (aggstrategy == AGG_MIXED && !enable_hashagg)
2767 456 : ++disabled_nodes;
2768 : /* calcs phrased this way to match HASHED case, see note above */
2769 11266 : total_cost += aggcosts->transCost.startup;
2770 11266 : total_cost += aggcosts->transCost.per_tuple * input_tuples;
2771 11266 : total_cost += (cpu_operator_cost * numGroupCols) * input_tuples;
2772 11266 : total_cost += aggcosts->finalCost.startup;
2773 11266 : total_cost += aggcosts->finalCost.per_tuple * numGroups;
2774 11266 : total_cost += cpu_tuple_cost * numGroups;
2775 11266 : output_tuples = numGroups;
2776 : }
2777 : else
2778 : {
2779 : /* must be AGG_HASHED */
2780 20980 : startup_cost = input_total_cost;
2781 20980 : if (!enable_hashagg)
2782 1578 : ++disabled_nodes;
2783 20980 : startup_cost += aggcosts->transCost.startup;
2784 20980 : startup_cost += aggcosts->transCost.per_tuple * input_tuples;
2785 : /* cost of computing hash value */
2786 20980 : startup_cost += (cpu_operator_cost * numGroupCols) * input_tuples;
2787 20980 : startup_cost += aggcosts->finalCost.startup;
2788 :
2789 20980 : total_cost = startup_cost;
2790 20980 : total_cost += aggcosts->finalCost.per_tuple * numGroups;
2791 : /* cost of retrieving from hash table */
2792 20980 : total_cost += cpu_tuple_cost * numGroups;
2793 20980 : output_tuples = numGroups;
2794 : }
2795 :
2796 : /*
2797 : * Add the disk costs of hash aggregation that spills to disk.
2798 : *
2799 : * Groups that go into the hash table stay in memory until finalized, so
2800 : * spilling and reprocessing tuples doesn't incur additional invocations
2801 : * of transCost or finalCost. Furthermore, the computed hash value is
2802 : * stored with the spilled tuples, so we don't incur extra invocations of
2803 : * the hash function.
2804 : *
2805 : * Hash Agg begins returning tuples after the first batch is complete.
2806 : * Accrue writes (spilled tuples) to startup_cost and to total_cost;
2807 : * accrue reads only to total_cost.
2808 : */
2809 69476 : if (aggstrategy == AGG_HASHED || aggstrategy == AGG_MIXED)
2810 : {
2811 : double pages;
2812 21896 : double pages_written = 0.0;
2813 21896 : double pages_read = 0.0;
2814 : double spill_cost;
2815 : double hashentrysize;
2816 : double nbatches;
2817 : Size mem_limit;
2818 : uint64 ngroups_limit;
2819 : int num_partitions;
2820 : int depth;
2821 :
2822 : /*
2823 : * Estimate number of batches based on the computed limits. If less
2824 : * than or equal to one, all groups are expected to fit in memory;
2825 : * otherwise we expect to spill.
2826 : */
2827 21896 : hashentrysize = hash_agg_entry_size(list_length(root->aggtransinfos),
2828 : input_width,
2829 21896 : aggcosts->transitionSpace);
2830 21896 : hash_agg_set_limits(hashentrysize, numGroups, 0, &mem_limit,
2831 : &ngroups_limit, &num_partitions);
2832 :
2833 21896 : nbatches = Max((numGroups * hashentrysize) / mem_limit,
2834 : numGroups / ngroups_limit);
2835 :
2836 21896 : nbatches = Max(ceil(nbatches), 1.0);
2837 21896 : num_partitions = Max(num_partitions, 2);
2838 :
2839 : /*
2840 : * The number of partitions can change at different levels of
2841 : * recursion; but for the purposes of this calculation assume it stays
2842 : * constant.
2843 : */
2844 21896 : depth = ceil(log(nbatches) / log(num_partitions));
2845 :
2846 : /*
2847 : * Estimate number of pages read and written. For each level of
2848 : * recursion, a tuple must be written and then later read.
2849 : */
2850 21896 : pages = relation_byte_size(input_tuples, input_width) / BLCKSZ;
2851 21896 : pages_written = pages_read = pages * depth;
2852 :
2853 : /*
2854 : * HashAgg has somewhat worse IO behavior than Sort on typical
2855 : * hardware/OS combinations. Account for this with a generic penalty.
2856 : */
2857 21896 : pages_read *= 2.0;
2858 21896 : pages_written *= 2.0;
2859 :
2860 21896 : startup_cost += pages_written * random_page_cost;
2861 21896 : total_cost += pages_written * random_page_cost;
2862 21896 : total_cost += pages_read * seq_page_cost;
2863 :
2864 : /* account for CPU cost of spilling a tuple and reading it back */
2865 21896 : spill_cost = depth * input_tuples * 2.0 * cpu_tuple_cost;
2866 21896 : startup_cost += spill_cost;
2867 21896 : total_cost += spill_cost;
2868 : }
2869 :
2870 : /*
2871 : * If there are quals (HAVING quals), account for their cost and
2872 : * selectivity.
2873 : */
2874 69476 : if (quals)
2875 : {
2876 : QualCost qual_cost;
2877 :
2878 4400 : cost_qual_eval(&qual_cost, quals, root);
2879 4400 : startup_cost += qual_cost.startup;
2880 4400 : total_cost += qual_cost.startup + output_tuples * qual_cost.per_tuple;
2881 :
2882 4400 : output_tuples = clamp_row_est(output_tuples *
2883 4400 : clauselist_selectivity(root,
2884 : quals,
2885 : 0,
2886 : JOIN_INNER,
2887 : NULL));
2888 : }
2889 :
2890 69476 : path->rows = output_tuples;
2891 69476 : path->disabled_nodes = disabled_nodes;
2892 69476 : path->startup_cost = startup_cost;
2893 69476 : path->total_cost = total_cost;
2894 69476 : }
2895 :
2896 : /*
2897 : * get_windowclause_startup_tuples
2898 : * Estimate how many tuples we'll need to fetch from a WindowAgg's
2899 : * subnode before we can output the first WindowAgg tuple.
2900 : *
2901 : * How many tuples need to be read depends on the WindowClause. For example,
2902 : * a WindowClause with no PARTITION BY and no ORDER BY requires that all
2903 : * subnode tuples are read and aggregated before the WindowAgg can output
2904 : * anything. If there's a PARTITION BY, then we only need to look at tuples
2905 : * in the first partition. Here we attempt to estimate just how many
2906 : * 'input_tuples' the WindowAgg will need to read for the given WindowClause
2907 : * before the first tuple can be output.
2908 : */
2909 : static double
2910 2754 : get_windowclause_startup_tuples(PlannerInfo *root, WindowClause *wc,
2911 : double input_tuples)
2912 : {
2913 2754 : int frameOptions = wc->frameOptions;
2914 : double partition_tuples;
2915 : double return_tuples;
2916 : double peer_tuples;
2917 :
2918 : /*
2919 : * First, figure out how many partitions there are likely to be and set
2920 : * partition_tuples according to that estimate.
2921 : */
2922 2754 : if (wc->partitionClause != NIL)
2923 : {
2924 : double num_partitions;
2925 716 : List *partexprs = get_sortgrouplist_exprs(wc->partitionClause,
2926 716 : root->parse->targetList);
2927 :
2928 716 : num_partitions = estimate_num_groups(root, partexprs, input_tuples,
2929 : NULL, NULL);
2930 716 : list_free(partexprs);
2931 :
2932 716 : partition_tuples = input_tuples / num_partitions;
2933 : }
2934 : else
2935 : {
2936 : /* all tuples belong to the same partition */
2937 2038 : partition_tuples = input_tuples;
2938 : }
2939 :
2940 : /* estimate the number of tuples in each peer group */
2941 2754 : if (wc->orderClause != NIL)
2942 : {
2943 : double num_groups;
2944 : List *orderexprs;
2945 :
2946 2274 : orderexprs = get_sortgrouplist_exprs(wc->orderClause,
2947 2274 : root->parse->targetList);
2948 :
2949 : /* estimate out how many peer groups there are in the partition */
2950 2274 : num_groups = estimate_num_groups(root, orderexprs,
2951 : partition_tuples, NULL,
2952 : NULL);
2953 2274 : list_free(orderexprs);
2954 2274 : peer_tuples = partition_tuples / num_groups;
2955 : }
2956 : else
2957 : {
2958 : /* no ORDER BY so only 1 tuple belongs in each peer group */
2959 480 : peer_tuples = 1.0;
2960 : }
2961 :
2962 2754 : if (frameOptions & FRAMEOPTION_END_UNBOUNDED_FOLLOWING)
2963 : {
2964 : /* include all partition rows */
2965 346 : return_tuples = partition_tuples;
2966 : }
2967 2408 : else if (frameOptions & FRAMEOPTION_END_CURRENT_ROW)
2968 : {
2969 1418 : if (frameOptions & FRAMEOPTION_ROWS)
2970 : {
2971 : /* just count the current row */
2972 608 : return_tuples = 1.0;
2973 : }
2974 810 : else if (frameOptions & (FRAMEOPTION_RANGE | FRAMEOPTION_GROUPS))
2975 : {
2976 : /*
2977 : * When in RANGE/GROUPS mode, it's more complex. If there's no
2978 : * ORDER BY, then all rows in the partition are peers, otherwise
2979 : * we'll need to read the first group of peers.
2980 : */
2981 810 : if (wc->orderClause == NIL)
2982 308 : return_tuples = partition_tuples;
2983 : else
2984 502 : return_tuples = peer_tuples;
2985 : }
2986 : else
2987 : {
2988 : /*
2989 : * Something new we don't support yet? This needs attention.
2990 : * We'll just return 1.0 in the meantime.
2991 : */
2992 : Assert(false);
2993 0 : return_tuples = 1.0;
2994 : }
2995 : }
2996 990 : else if (frameOptions & FRAMEOPTION_END_OFFSET_PRECEDING)
2997 : {
2998 : /*
2999 : * BETWEEN ... AND N PRECEDING will only need to read the WindowAgg's
3000 : * subnode after N ROWS/RANGES/GROUPS. N can be 0, but not negative,
3001 : * so we'll just assume only the current row needs to be read to fetch
3002 : * the first WindowAgg row.
3003 : */
3004 108 : return_tuples = 1.0;
3005 : }
3006 882 : else if (frameOptions & FRAMEOPTION_END_OFFSET_FOLLOWING)
3007 : {
3008 882 : Const *endOffset = (Const *) wc->endOffset;
3009 : double end_offset_value;
3010 :
3011 : /* try and figure out the value specified in the endOffset. */
3012 882 : if (IsA(endOffset, Const))
3013 : {
3014 882 : if (endOffset->constisnull)
3015 : {
3016 : /*
3017 : * NULLs are not allowed, but currently, there's no code to
3018 : * error out if there's a NULL Const. We'll only discover
3019 : * this during execution. For now, just pretend everything is
3020 : * fine and assume that just the first row/range/group will be
3021 : * needed.
3022 : */
3023 0 : end_offset_value = 1.0;
3024 : }
3025 : else
3026 : {
3027 882 : switch (endOffset->consttype)
3028 : {
3029 24 : case INT2OID:
3030 24 : end_offset_value =
3031 24 : (double) DatumGetInt16(endOffset->constvalue);
3032 24 : break;
3033 132 : case INT4OID:
3034 132 : end_offset_value =
3035 132 : (double) DatumGetInt32(endOffset->constvalue);
3036 132 : break;
3037 384 : case INT8OID:
3038 384 : end_offset_value =
3039 384 : (double) DatumGetInt64(endOffset->constvalue);
3040 384 : break;
3041 342 : default:
3042 342 : end_offset_value =
3043 342 : partition_tuples / peer_tuples *
3044 : DEFAULT_INEQ_SEL;
3045 342 : break;
3046 : }
3047 : }
3048 : }
3049 : else
3050 : {
3051 : /*
3052 : * When the end bound is not a Const, we'll just need to guess. We
3053 : * just make use of DEFAULT_INEQ_SEL.
3054 : */
3055 0 : end_offset_value =
3056 0 : partition_tuples / peer_tuples * DEFAULT_INEQ_SEL;
3057 : }
3058 :
3059 882 : if (frameOptions & FRAMEOPTION_ROWS)
3060 : {
3061 : /* include the N FOLLOWING and the current row */
3062 222 : return_tuples = end_offset_value + 1.0;
3063 : }
3064 660 : else if (frameOptions & (FRAMEOPTION_RANGE | FRAMEOPTION_GROUPS))
3065 : {
3066 : /* include N FOLLOWING ranges/group and the initial range/group */
3067 660 : return_tuples = peer_tuples * (end_offset_value + 1.0);
3068 : }
3069 : else
3070 : {
3071 : /*
3072 : * Something new we don't support yet? This needs attention.
3073 : * We'll just return 1.0 in the meantime.
3074 : */
3075 : Assert(false);
3076 0 : return_tuples = 1.0;
3077 : }
3078 : }
3079 : else
3080 : {
3081 : /*
3082 : * Something new we don't support yet? This needs attention. We'll
3083 : * just return 1.0 in the meantime.
3084 : */
3085 : Assert(false);
3086 0 : return_tuples = 1.0;
3087 : }
3088 :
3089 2754 : if (wc->partitionClause != NIL || wc->orderClause != NIL)
3090 : {
3091 : /*
3092 : * Cap the return value to the estimated partition tuples and account
3093 : * for the extra tuple WindowAgg will need to read to confirm the next
3094 : * tuple does not belong to the same partition or peer group.
3095 : */
3096 2474 : return_tuples = Min(return_tuples + 1.0, partition_tuples);
3097 : }
3098 : else
3099 : {
3100 : /*
3101 : * Cap the return value so it's never higher than the expected tuples
3102 : * in the partition.
3103 : */
3104 280 : return_tuples = Min(return_tuples, partition_tuples);
3105 : }
3106 :
3107 : /*
3108 : * We needn't worry about any EXCLUDE options as those only exclude rows
3109 : * from being aggregated, not from being read from the WindowAgg's
3110 : * subnode.
3111 : */
3112 :
3113 2754 : return clamp_row_est(return_tuples);
3114 : }
3115 :
3116 : /*
3117 : * cost_windowagg
3118 : * Determines and returns the cost of performing a WindowAgg plan node,
3119 : * including the cost of its input.
3120 : *
3121 : * Input is assumed already properly sorted.
3122 : */
3123 : void
3124 2754 : cost_windowagg(Path *path, PlannerInfo *root,
3125 : List *windowFuncs, WindowClause *winclause,
3126 : int input_disabled_nodes,
3127 : Cost input_startup_cost, Cost input_total_cost,
3128 : double input_tuples)
3129 : {
3130 : Cost startup_cost;
3131 : Cost total_cost;
3132 : double startup_tuples;
3133 : int numPartCols;
3134 : int numOrderCols;
3135 : ListCell *lc;
3136 :
3137 2754 : numPartCols = list_length(winclause->partitionClause);
3138 2754 : numOrderCols = list_length(winclause->orderClause);
3139 :
3140 2754 : startup_cost = input_startup_cost;
3141 2754 : total_cost = input_total_cost;
3142 :
3143 : /*
3144 : * Window functions are assumed to cost their stated execution cost, plus
3145 : * the cost of evaluating their input expressions, per tuple. Since they
3146 : * may in fact evaluate their inputs at multiple rows during each cycle,
3147 : * this could be a drastic underestimate; but without a way to know how
3148 : * many rows the window function will fetch, it's hard to do better. In
3149 : * any case, it's a good estimate for all the built-in window functions,
3150 : * so we'll just do this for now.
3151 : */
3152 6246 : foreach(lc, windowFuncs)
3153 : {
3154 3492 : WindowFunc *wfunc = lfirst_node(WindowFunc, lc);
3155 : Cost wfunccost;
3156 : QualCost argcosts;
3157 :
3158 3492 : argcosts.startup = argcosts.per_tuple = 0;
3159 3492 : add_function_cost(root, wfunc->winfnoid, (Node *) wfunc,
3160 : &argcosts);
3161 3492 : startup_cost += argcosts.startup;
3162 3492 : wfunccost = argcosts.per_tuple;
3163 :
3164 : /* also add the input expressions' cost to per-input-row costs */
3165 3492 : cost_qual_eval_node(&argcosts, (Node *) wfunc->args, root);
3166 3492 : startup_cost += argcosts.startup;
3167 3492 : wfunccost += argcosts.per_tuple;
3168 :
3169 : /*
3170 : * Add the filter's cost to per-input-row costs. XXX We should reduce
3171 : * input expression costs according to filter selectivity.
3172 : */
3173 3492 : cost_qual_eval_node(&argcosts, (Node *) wfunc->aggfilter, root);
3174 3492 : startup_cost += argcosts.startup;
3175 3492 : wfunccost += argcosts.per_tuple;
3176 :
3177 3492 : total_cost += wfunccost * input_tuples;
3178 : }
3179 :
3180 : /*
3181 : * We also charge cpu_operator_cost per grouping column per tuple for
3182 : * grouping comparisons, plus cpu_tuple_cost per tuple for general
3183 : * overhead.
3184 : *
3185 : * XXX this neglects costs of spooling the data to disk when it overflows
3186 : * work_mem. Sooner or later that should get accounted for.
3187 : */
3188 2754 : total_cost += cpu_operator_cost * (numPartCols + numOrderCols) * input_tuples;
3189 2754 : total_cost += cpu_tuple_cost * input_tuples;
3190 :
3191 2754 : path->rows = input_tuples;
3192 2754 : path->disabled_nodes = input_disabled_nodes;
3193 2754 : path->startup_cost = startup_cost;
3194 2754 : path->total_cost = total_cost;
3195 :
3196 : /*
3197 : * Also, take into account how many tuples we need to read from the
3198 : * subnode in order to produce the first tuple from the WindowAgg. To do
3199 : * this we proportion the run cost (total cost not including startup cost)
3200 : * over the estimated startup tuples. We already included the startup
3201 : * cost of the subnode, so we only need to do this when the estimated
3202 : * startup tuples is above 1.0.
3203 : */
3204 2754 : startup_tuples = get_windowclause_startup_tuples(root, winclause,
3205 : input_tuples);
3206 :
3207 2754 : if (startup_tuples > 1.0)
3208 2466 : path->startup_cost += (total_cost - startup_cost) / input_tuples *
3209 2466 : (startup_tuples - 1.0);
3210 2754 : }
3211 :
3212 : /*
3213 : * cost_group
3214 : * Determines and returns the cost of performing a Group plan node,
3215 : * including the cost of its input.
3216 : *
3217 : * Note: caller must ensure that input costs are for appropriately-sorted
3218 : * input.
3219 : */
3220 : void
3221 1214 : cost_group(Path *path, PlannerInfo *root,
3222 : int numGroupCols, double numGroups,
3223 : List *quals,
3224 : int input_disabled_nodes,
3225 : Cost input_startup_cost, Cost input_total_cost,
3226 : double input_tuples)
3227 : {
3228 : double output_tuples;
3229 : Cost startup_cost;
3230 : Cost total_cost;
3231 :
3232 1214 : output_tuples = numGroups;
3233 1214 : startup_cost = input_startup_cost;
3234 1214 : total_cost = input_total_cost;
3235 :
3236 : /*
3237 : * Charge one cpu_operator_cost per comparison per input tuple. We assume
3238 : * all columns get compared at most of the tuples.
3239 : */
3240 1214 : total_cost += cpu_operator_cost * input_tuples * numGroupCols;
3241 :
3242 : /*
3243 : * If there are quals (HAVING quals), account for their cost and
3244 : * selectivity.
3245 : */
3246 1214 : if (quals)
3247 : {
3248 : QualCost qual_cost;
3249 :
3250 0 : cost_qual_eval(&qual_cost, quals, root);
3251 0 : startup_cost += qual_cost.startup;
3252 0 : total_cost += qual_cost.startup + output_tuples * qual_cost.per_tuple;
3253 :
3254 0 : output_tuples = clamp_row_est(output_tuples *
3255 0 : clauselist_selectivity(root,
3256 : quals,
3257 : 0,
3258 : JOIN_INNER,
3259 : NULL));
3260 : }
3261 :
3262 1214 : path->rows = output_tuples;
3263 1214 : path->disabled_nodes = input_disabled_nodes;
3264 1214 : path->startup_cost = startup_cost;
3265 1214 : path->total_cost = total_cost;
3266 1214 : }
3267 :
3268 : /*
3269 : * initial_cost_nestloop
3270 : * Preliminary estimate of the cost of a nestloop join path.
3271 : *
3272 : * This must quickly produce lower-bound estimates of the path's startup and
3273 : * total costs. If we are unable to eliminate the proposed path from
3274 : * consideration using the lower bounds, final_cost_nestloop will be called
3275 : * to obtain the final estimates.
3276 : *
3277 : * The exact division of labor between this function and final_cost_nestloop
3278 : * is private to them, and represents a tradeoff between speed of the initial
3279 : * estimate and getting a tight lower bound. We choose to not examine the
3280 : * join quals here, since that's by far the most expensive part of the
3281 : * calculations. The end result is that CPU-cost considerations must be
3282 : * left for the second phase; and for SEMI/ANTI joins, we must also postpone
3283 : * incorporation of the inner path's run cost.
3284 : *
3285 : * 'workspace' is to be filled with startup_cost, total_cost, and perhaps
3286 : * other data to be used by final_cost_nestloop
3287 : * 'jointype' is the type of join to be performed
3288 : * 'outer_path' is the outer input to the join
3289 : * 'inner_path' is the inner input to the join
3290 : * 'extra' contains miscellaneous information about the join
3291 : */
3292 : void
3293 2915176 : initial_cost_nestloop(PlannerInfo *root, JoinCostWorkspace *workspace,
3294 : JoinType jointype,
3295 : Path *outer_path, Path *inner_path,
3296 : JoinPathExtraData *extra)
3297 : {
3298 : int disabled_nodes;
3299 2915176 : Cost startup_cost = 0;
3300 2915176 : Cost run_cost = 0;
3301 2915176 : double outer_path_rows = outer_path->rows;
3302 : Cost inner_rescan_start_cost;
3303 : Cost inner_rescan_total_cost;
3304 : Cost inner_run_cost;
3305 : Cost inner_rescan_run_cost;
3306 :
3307 : /* Count up disabled nodes. */
3308 2915176 : disabled_nodes = enable_nestloop ? 0 : 1;
3309 2915176 : disabled_nodes += inner_path->disabled_nodes;
3310 2915176 : disabled_nodes += outer_path->disabled_nodes;
3311 :
3312 : /* estimate costs to rescan the inner relation */
3313 2915176 : cost_rescan(root, inner_path,
3314 : &inner_rescan_start_cost,
3315 : &inner_rescan_total_cost);
3316 :
3317 : /* cost of source data */
3318 :
3319 : /*
3320 : * NOTE: clearly, we must pay both outer and inner paths' startup_cost
3321 : * before we can start returning tuples, so the join's startup cost is
3322 : * their sum. We'll also pay the inner path's rescan startup cost
3323 : * multiple times.
3324 : */
3325 2915176 : startup_cost += outer_path->startup_cost + inner_path->startup_cost;
3326 2915176 : run_cost += outer_path->total_cost - outer_path->startup_cost;
3327 2915176 : if (outer_path_rows > 1)
3328 2072332 : run_cost += (outer_path_rows - 1) * inner_rescan_start_cost;
3329 :
3330 2915176 : inner_run_cost = inner_path->total_cost - inner_path->startup_cost;
3331 2915176 : inner_rescan_run_cost = inner_rescan_total_cost - inner_rescan_start_cost;
3332 :
3333 2915176 : if (jointype == JOIN_SEMI || jointype == JOIN_ANTI ||
3334 2852572 : extra->inner_unique)
3335 : {
3336 : /*
3337 : * With a SEMI or ANTI join, or if the innerrel is known unique, the
3338 : * executor will stop after the first match.
3339 : *
3340 : * Getting decent estimates requires inspection of the join quals,
3341 : * which we choose to postpone to final_cost_nestloop.
3342 : */
3343 :
3344 : /* Save private data for final_cost_nestloop */
3345 1333264 : workspace->inner_run_cost = inner_run_cost;
3346 1333264 : workspace->inner_rescan_run_cost = inner_rescan_run_cost;
3347 : }
3348 : else
3349 : {
3350 : /* Normal case; we'll scan whole input rel for each outer row */
3351 1581912 : run_cost += inner_run_cost;
3352 1581912 : if (outer_path_rows > 1)
3353 1138416 : run_cost += (outer_path_rows - 1) * inner_rescan_run_cost;
3354 : }
3355 :
3356 : /* CPU costs left for later */
3357 :
3358 : /* Public result fields */
3359 2915176 : workspace->disabled_nodes = disabled_nodes;
3360 2915176 : workspace->startup_cost = startup_cost;
3361 2915176 : workspace->total_cost = startup_cost + run_cost;
3362 : /* Save private data for final_cost_nestloop */
3363 2915176 : workspace->run_cost = run_cost;
3364 2915176 : }
3365 :
3366 : /*
3367 : * final_cost_nestloop
3368 : * Final estimate of the cost and result size of a nestloop join path.
3369 : *
3370 : * 'path' is already filled in except for the rows and cost fields
3371 : * 'workspace' is the result from initial_cost_nestloop
3372 : * 'extra' contains miscellaneous information about the join
3373 : */
3374 : void
3375 1405576 : final_cost_nestloop(PlannerInfo *root, NestPath *path,
3376 : JoinCostWorkspace *workspace,
3377 : JoinPathExtraData *extra)
3378 : {
3379 1405576 : Path *outer_path = path->jpath.outerjoinpath;
3380 1405576 : Path *inner_path = path->jpath.innerjoinpath;
3381 1405576 : double outer_path_rows = outer_path->rows;
3382 1405576 : double inner_path_rows = inner_path->rows;
3383 1405576 : Cost startup_cost = workspace->startup_cost;
3384 1405576 : Cost run_cost = workspace->run_cost;
3385 : Cost cpu_per_tuple;
3386 : QualCost restrict_qual_cost;
3387 : double ntuples;
3388 :
3389 : /* Set the number of disabled nodes. */
3390 1405576 : path->jpath.path.disabled_nodes = workspace->disabled_nodes;
3391 :
3392 : /* Protect some assumptions below that rowcounts aren't zero */
3393 1405576 : if (outer_path_rows <= 0)
3394 0 : outer_path_rows = 1;
3395 1405576 : if (inner_path_rows <= 0)
3396 702 : inner_path_rows = 1;
3397 : /* Mark the path with the correct row estimate */
3398 1405576 : if (path->jpath.path.param_info)
3399 29138 : path->jpath.path.rows = path->jpath.path.param_info->ppi_rows;
3400 : else
3401 1376438 : path->jpath.path.rows = path->jpath.path.parent->rows;
3402 :
3403 : /* For partial paths, scale row estimate. */
3404 1405576 : if (path->jpath.path.parallel_workers > 0)
3405 : {
3406 12758 : double parallel_divisor = get_parallel_divisor(&path->jpath.path);
3407 :
3408 12758 : path->jpath.path.rows =
3409 12758 : clamp_row_est(path->jpath.path.rows / parallel_divisor);
3410 : }
3411 :
3412 : /* cost of inner-relation source data (we already dealt with outer rel) */
3413 :
3414 1405576 : if (path->jpath.jointype == JOIN_SEMI || path->jpath.jointype == JOIN_ANTI ||
3415 1362990 : extra->inner_unique)
3416 910432 : {
3417 : /*
3418 : * With a SEMI or ANTI join, or if the innerrel is known unique, the
3419 : * executor will stop after the first match.
3420 : */
3421 910432 : Cost inner_run_cost = workspace->inner_run_cost;
3422 910432 : Cost inner_rescan_run_cost = workspace->inner_rescan_run_cost;
3423 : double outer_matched_rows;
3424 : double outer_unmatched_rows;
3425 : Selectivity inner_scan_frac;
3426 :
3427 : /*
3428 : * For an outer-rel row that has at least one match, we can expect the
3429 : * inner scan to stop after a fraction 1/(match_count+1) of the inner
3430 : * rows, if the matches are evenly distributed. Since they probably
3431 : * aren't quite evenly distributed, we apply a fuzz factor of 2.0 to
3432 : * that fraction. (If we used a larger fuzz factor, we'd have to
3433 : * clamp inner_scan_frac to at most 1.0; but since match_count is at
3434 : * least 1, no such clamp is needed now.)
3435 : */
3436 910432 : outer_matched_rows = rint(outer_path_rows * extra->semifactors.outer_match_frac);
3437 910432 : outer_unmatched_rows = outer_path_rows - outer_matched_rows;
3438 910432 : inner_scan_frac = 2.0 / (extra->semifactors.match_count + 1.0);
3439 :
3440 : /*
3441 : * Compute number of tuples processed (not number emitted!). First,
3442 : * account for successfully-matched outer rows.
3443 : */
3444 910432 : ntuples = outer_matched_rows * inner_path_rows * inner_scan_frac;
3445 :
3446 : /*
3447 : * Now we need to estimate the actual costs of scanning the inner
3448 : * relation, which may be quite a bit less than N times inner_run_cost
3449 : * due to early scan stops. We consider two cases. If the inner path
3450 : * is an indexscan using all the joinquals as indexquals, then an
3451 : * unmatched outer row results in an indexscan returning no rows,
3452 : * which is probably quite cheap. Otherwise, the executor will have
3453 : * to scan the whole inner rel for an unmatched row; not so cheap.
3454 : */
3455 910432 : if (has_indexed_join_quals(path))
3456 : {
3457 : /*
3458 : * Successfully-matched outer rows will only require scanning
3459 : * inner_scan_frac of the inner relation. In this case, we don't
3460 : * need to charge the full inner_run_cost even when that's more
3461 : * than inner_rescan_run_cost, because we can assume that none of
3462 : * the inner scans ever scan the whole inner relation. So it's
3463 : * okay to assume that all the inner scan executions can be
3464 : * fractions of the full cost, even if materialization is reducing
3465 : * the rescan cost. At this writing, it's impossible to get here
3466 : * for a materialized inner scan, so inner_run_cost and
3467 : * inner_rescan_run_cost will be the same anyway; but just in
3468 : * case, use inner_run_cost for the first matched tuple and
3469 : * inner_rescan_run_cost for additional ones.
3470 : */
3471 149778 : run_cost += inner_run_cost * inner_scan_frac;
3472 149778 : if (outer_matched_rows > 1)
3473 20642 : run_cost += (outer_matched_rows - 1) * inner_rescan_run_cost * inner_scan_frac;
3474 :
3475 : /*
3476 : * Add the cost of inner-scan executions for unmatched outer rows.
3477 : * We estimate this as the same cost as returning the first tuple
3478 : * of a nonempty scan. We consider that these are all rescans,
3479 : * since we used inner_run_cost once already.
3480 : */
3481 149778 : run_cost += outer_unmatched_rows *
3482 149778 : inner_rescan_run_cost / inner_path_rows;
3483 :
3484 : /*
3485 : * We won't be evaluating any quals at all for unmatched rows, so
3486 : * don't add them to ntuples.
3487 : */
3488 : }
3489 : else
3490 : {
3491 : /*
3492 : * Here, a complicating factor is that rescans may be cheaper than
3493 : * first scans. If we never scan all the way to the end of the
3494 : * inner rel, it might be (depending on the plan type) that we'd
3495 : * never pay the whole inner first-scan run cost. However it is
3496 : * difficult to estimate whether that will happen (and it could
3497 : * not happen if there are any unmatched outer rows!), so be
3498 : * conservative and always charge the whole first-scan cost once.
3499 : * We consider this charge to correspond to the first unmatched
3500 : * outer row, unless there isn't one in our estimate, in which
3501 : * case blame it on the first matched row.
3502 : */
3503 :
3504 : /* First, count all unmatched join tuples as being processed */
3505 760654 : ntuples += outer_unmatched_rows * inner_path_rows;
3506 :
3507 : /* Now add the forced full scan, and decrement appropriate count */
3508 760654 : run_cost += inner_run_cost;
3509 760654 : if (outer_unmatched_rows >= 1)
3510 728896 : outer_unmatched_rows -= 1;
3511 : else
3512 31758 : outer_matched_rows -= 1;
3513 :
3514 : /* Add inner run cost for additional outer tuples having matches */
3515 760654 : if (outer_matched_rows > 0)
3516 268356 : run_cost += outer_matched_rows * inner_rescan_run_cost * inner_scan_frac;
3517 :
3518 : /* Add inner run cost for additional unmatched outer tuples */
3519 760654 : if (outer_unmatched_rows > 0)
3520 507888 : run_cost += outer_unmatched_rows * inner_rescan_run_cost;
3521 : }
3522 : }
3523 : else
3524 : {
3525 : /* Normal-case source costs were included in preliminary estimate */
3526 :
3527 : /* Compute number of tuples processed (not number emitted!) */
3528 495144 : ntuples = outer_path_rows * inner_path_rows;
3529 : }
3530 :
3531 : /* CPU costs */
3532 1405576 : cost_qual_eval(&restrict_qual_cost, path->jpath.joinrestrictinfo, root);
3533 1405576 : startup_cost += restrict_qual_cost.startup;
3534 1405576 : cpu_per_tuple = cpu_tuple_cost + restrict_qual_cost.per_tuple;
3535 1405576 : run_cost += cpu_per_tuple * ntuples;
3536 :
3537 : /* tlist eval costs are paid per output row, not per tuple scanned */
3538 1405576 : startup_cost += path->jpath.path.pathtarget->cost.startup;
3539 1405576 : run_cost += path->jpath.path.pathtarget->cost.per_tuple * path->jpath.path.rows;
3540 :
3541 1405576 : path->jpath.path.startup_cost = startup_cost;
3542 1405576 : path->jpath.path.total_cost = startup_cost + run_cost;
3543 1405576 : }
3544 :
3545 : /*
3546 : * initial_cost_mergejoin
3547 : * Preliminary estimate of the cost of a mergejoin path.
3548 : *
3549 : * This must quickly produce lower-bound estimates of the path's startup and
3550 : * total costs. If we are unable to eliminate the proposed path from
3551 : * consideration using the lower bounds, final_cost_mergejoin will be called
3552 : * to obtain the final estimates.
3553 : *
3554 : * The exact division of labor between this function and final_cost_mergejoin
3555 : * is private to them, and represents a tradeoff between speed of the initial
3556 : * estimate and getting a tight lower bound. We choose to not examine the
3557 : * join quals here, except for obtaining the scan selectivity estimate which
3558 : * is really essential (but fortunately, use of caching keeps the cost of
3559 : * getting that down to something reasonable).
3560 : * We also assume that cost_sort/cost_incremental_sort is cheap enough to use
3561 : * here.
3562 : *
3563 : * 'workspace' is to be filled with startup_cost, total_cost, and perhaps
3564 : * other data to be used by final_cost_mergejoin
3565 : * 'jointype' is the type of join to be performed
3566 : * 'mergeclauses' is the list of joinclauses to be used as merge clauses
3567 : * 'outer_path' is the outer input to the join
3568 : * 'inner_path' is the inner input to the join
3569 : * 'outersortkeys' is the list of sort keys for the outer path
3570 : * 'innersortkeys' is the list of sort keys for the inner path
3571 : * 'outer_presorted_keys' is the number of presorted keys of the outer path
3572 : * 'extra' contains miscellaneous information about the join
3573 : *
3574 : * Note: outersortkeys and innersortkeys should be NIL if no explicit
3575 : * sort is needed because the respective source path is already ordered.
3576 : */
3577 : void
3578 1284824 : initial_cost_mergejoin(PlannerInfo *root, JoinCostWorkspace *workspace,
3579 : JoinType jointype,
3580 : List *mergeclauses,
3581 : Path *outer_path, Path *inner_path,
3582 : List *outersortkeys, List *innersortkeys,
3583 : int outer_presorted_keys,
3584 : JoinPathExtraData *extra)
3585 : {
3586 : int disabled_nodes;
3587 1284824 : Cost startup_cost = 0;
3588 1284824 : Cost run_cost = 0;
3589 1284824 : double outer_path_rows = outer_path->rows;
3590 1284824 : double inner_path_rows = inner_path->rows;
3591 : Cost inner_run_cost;
3592 : double outer_rows,
3593 : inner_rows,
3594 : outer_skip_rows,
3595 : inner_skip_rows;
3596 : Selectivity outerstartsel,
3597 : outerendsel,
3598 : innerstartsel,
3599 : innerendsel;
3600 : Path sort_path; /* dummy for result of
3601 : * cost_sort/cost_incremental_sort */
3602 :
3603 : /* Protect some assumptions below that rowcounts aren't zero */
3604 1284824 : if (outer_path_rows <= 0)
3605 96 : outer_path_rows = 1;
3606 1284824 : if (inner_path_rows <= 0)
3607 126 : inner_path_rows = 1;
3608 :
3609 : /*
3610 : * A merge join will stop as soon as it exhausts either input stream
3611 : * (unless it's an outer join, in which case the outer side has to be
3612 : * scanned all the way anyway). Estimate fraction of the left and right
3613 : * inputs that will actually need to be scanned. Likewise, we can
3614 : * estimate the number of rows that will be skipped before the first join
3615 : * pair is found, which should be factored into startup cost. We use only
3616 : * the first (most significant) merge clause for this purpose. Since
3617 : * mergejoinscansel() is a fairly expensive computation, we cache the
3618 : * results in the merge clause RestrictInfo.
3619 : */
3620 1284824 : if (mergeclauses && jointype != JOIN_FULL)
3621 1278684 : {
3622 1278684 : RestrictInfo *firstclause = (RestrictInfo *) linitial(mergeclauses);
3623 : List *opathkeys;
3624 : List *ipathkeys;
3625 : PathKey *opathkey;
3626 : PathKey *ipathkey;
3627 : MergeScanSelCache *cache;
3628 :
3629 : /* Get the input pathkeys to determine the sort-order details */
3630 1278684 : opathkeys = outersortkeys ? outersortkeys : outer_path->pathkeys;
3631 1278684 : ipathkeys = innersortkeys ? innersortkeys : inner_path->pathkeys;
3632 : Assert(opathkeys);
3633 : Assert(ipathkeys);
3634 1278684 : opathkey = (PathKey *) linitial(opathkeys);
3635 1278684 : ipathkey = (PathKey *) linitial(ipathkeys);
3636 : /* debugging check */
3637 1278684 : if (opathkey->pk_opfamily != ipathkey->pk_opfamily ||
3638 1278684 : opathkey->pk_eclass->ec_collation != ipathkey->pk_eclass->ec_collation ||
3639 1278684 : opathkey->pk_cmptype != ipathkey->pk_cmptype ||
3640 1278684 : opathkey->pk_nulls_first != ipathkey->pk_nulls_first)
3641 0 : elog(ERROR, "left and right pathkeys do not match in mergejoin");
3642 :
3643 : /* Get the selectivity with caching */
3644 1278684 : cache = cached_scansel(root, firstclause, opathkey);
3645 :
3646 1278684 : if (bms_is_subset(firstclause->left_relids,
3647 1278684 : outer_path->parent->relids))
3648 : {
3649 : /* left side of clause is outer */
3650 687210 : outerstartsel = cache->leftstartsel;
3651 687210 : outerendsel = cache->leftendsel;
3652 687210 : innerstartsel = cache->rightstartsel;
3653 687210 : innerendsel = cache->rightendsel;
3654 : }
3655 : else
3656 : {
3657 : /* left side of clause is inner */
3658 591474 : outerstartsel = cache->rightstartsel;
3659 591474 : outerendsel = cache->rightendsel;
3660 591474 : innerstartsel = cache->leftstartsel;
3661 591474 : innerendsel = cache->leftendsel;
3662 : }
3663 1278684 : if (jointype == JOIN_LEFT ||
3664 : jointype == JOIN_ANTI)
3665 : {
3666 222462 : outerstartsel = 0.0;
3667 222462 : outerendsel = 1.0;
3668 : }
3669 1056222 : else if (jointype == JOIN_RIGHT ||
3670 : jointype == JOIN_RIGHT_ANTI)
3671 : {
3672 215144 : innerstartsel = 0.0;
3673 215144 : innerendsel = 1.0;
3674 : }
3675 : }
3676 : else
3677 : {
3678 : /* cope with clauseless or full mergejoin */
3679 6140 : outerstartsel = innerstartsel = 0.0;
3680 6140 : outerendsel = innerendsel = 1.0;
3681 : }
3682 :
3683 : /*
3684 : * Convert selectivities to row counts. We force outer_rows and
3685 : * inner_rows to be at least 1, but the skip_rows estimates can be zero.
3686 : */
3687 1284824 : outer_skip_rows = rint(outer_path_rows * outerstartsel);
3688 1284824 : inner_skip_rows = rint(inner_path_rows * innerstartsel);
3689 1284824 : outer_rows = clamp_row_est(outer_path_rows * outerendsel);
3690 1284824 : inner_rows = clamp_row_est(inner_path_rows * innerendsel);
3691 :
3692 : Assert(outer_skip_rows <= outer_rows);
3693 : Assert(inner_skip_rows <= inner_rows);
3694 :
3695 : /*
3696 : * Readjust scan selectivities to account for above rounding. This is
3697 : * normally an insignificant effect, but when there are only a few rows in
3698 : * the inputs, failing to do this makes for a large percentage error.
3699 : */
3700 1284824 : outerstartsel = outer_skip_rows / outer_path_rows;
3701 1284824 : innerstartsel = inner_skip_rows / inner_path_rows;
3702 1284824 : outerendsel = outer_rows / outer_path_rows;
3703 1284824 : innerendsel = inner_rows / inner_path_rows;
3704 :
3705 : Assert(outerstartsel <= outerendsel);
3706 : Assert(innerstartsel <= innerendsel);
3707 :
3708 1284824 : disabled_nodes = enable_mergejoin ? 0 : 1;
3709 :
3710 : /* cost of source data */
3711 :
3712 1284824 : if (outersortkeys) /* do we need to sort outer? */
3713 : {
3714 : /*
3715 : * We can assert that the outer path is not already ordered
3716 : * appropriately for the mergejoin; otherwise, outersortkeys would
3717 : * have been set to NIL.
3718 : */
3719 : Assert(!pathkeys_contained_in(outersortkeys, outer_path->pathkeys));
3720 :
3721 : /*
3722 : * We choose to use incremental sort if it is enabled and there are
3723 : * presorted keys; otherwise we use full sort.
3724 : */
3725 616130 : if (enable_incremental_sort && outer_presorted_keys > 0)
3726 : {
3727 1978 : cost_incremental_sort(&sort_path,
3728 : root,
3729 : outersortkeys,
3730 : outer_presorted_keys,
3731 : outer_path->disabled_nodes,
3732 : outer_path->startup_cost,
3733 : outer_path->total_cost,
3734 : outer_path_rows,
3735 1978 : outer_path->pathtarget->width,
3736 : 0.0,
3737 : work_mem,
3738 : -1.0);
3739 : }
3740 : else
3741 : {
3742 614152 : cost_sort(&sort_path,
3743 : root,
3744 : outersortkeys,
3745 : outer_path->disabled_nodes,
3746 : outer_path->total_cost,
3747 : outer_path_rows,
3748 614152 : outer_path->pathtarget->width,
3749 : 0.0,
3750 : work_mem,
3751 : -1.0);
3752 : }
3753 :
3754 616130 : disabled_nodes += sort_path.disabled_nodes;
3755 616130 : startup_cost += sort_path.startup_cost;
3756 616130 : startup_cost += (sort_path.total_cost - sort_path.startup_cost)
3757 616130 : * outerstartsel;
3758 616130 : run_cost += (sort_path.total_cost - sort_path.startup_cost)
3759 616130 : * (outerendsel - outerstartsel);
3760 : }
3761 : else
3762 : {
3763 668694 : disabled_nodes += outer_path->disabled_nodes;
3764 668694 : startup_cost += outer_path->startup_cost;
3765 668694 : startup_cost += (outer_path->total_cost - outer_path->startup_cost)
3766 668694 : * outerstartsel;
3767 668694 : run_cost += (outer_path->total_cost - outer_path->startup_cost)
3768 668694 : * (outerendsel - outerstartsel);
3769 : }
3770 :
3771 1284824 : if (innersortkeys) /* do we need to sort inner? */
3772 : {
3773 : /*
3774 : * We can assert that the inner path is not already ordered
3775 : * appropriately for the mergejoin; otherwise, innersortkeys would
3776 : * have been set to NIL.
3777 : */
3778 : Assert(!pathkeys_contained_in(innersortkeys, inner_path->pathkeys));
3779 :
3780 : /*
3781 : * We do not consider incremental sort for inner path, because
3782 : * incremental sort does not support mark/restore.
3783 : */
3784 :
3785 1006564 : cost_sort(&sort_path,
3786 : root,
3787 : innersortkeys,
3788 : inner_path->disabled_nodes,
3789 : inner_path->total_cost,
3790 : inner_path_rows,
3791 1006564 : inner_path->pathtarget->width,
3792 : 0.0,
3793 : work_mem,
3794 : -1.0);
3795 1006564 : disabled_nodes += sort_path.disabled_nodes;
3796 1006564 : startup_cost += sort_path.startup_cost;
3797 1006564 : startup_cost += (sort_path.total_cost - sort_path.startup_cost)
3798 1006564 : * innerstartsel;
3799 1006564 : inner_run_cost = (sort_path.total_cost - sort_path.startup_cost)
3800 1006564 : * (innerendsel - innerstartsel);
3801 : }
3802 : else
3803 : {
3804 278260 : disabled_nodes += inner_path->disabled_nodes;
3805 278260 : startup_cost += inner_path->startup_cost;
3806 278260 : startup_cost += (inner_path->total_cost - inner_path->startup_cost)
3807 278260 : * innerstartsel;
3808 278260 : inner_run_cost = (inner_path->total_cost - inner_path->startup_cost)
3809 278260 : * (innerendsel - innerstartsel);
3810 : }
3811 :
3812 : /*
3813 : * We can't yet determine whether rescanning occurs, or whether
3814 : * materialization of the inner input should be done. The minimum
3815 : * possible inner input cost, regardless of rescan and materialization
3816 : * considerations, is inner_run_cost. We include that in
3817 : * workspace->total_cost, but not yet in run_cost.
3818 : */
3819 :
3820 : /* CPU costs left for later */
3821 :
3822 : /* Public result fields */
3823 1284824 : workspace->disabled_nodes = disabled_nodes;
3824 1284824 : workspace->startup_cost = startup_cost;
3825 1284824 : workspace->total_cost = startup_cost + run_cost + inner_run_cost;
3826 : /* Save private data for final_cost_mergejoin */
3827 1284824 : workspace->run_cost = run_cost;
3828 1284824 : workspace->inner_run_cost = inner_run_cost;
3829 1284824 : workspace->outer_rows = outer_rows;
3830 1284824 : workspace->inner_rows = inner_rows;
3831 1284824 : workspace->outer_skip_rows = outer_skip_rows;
3832 1284824 : workspace->inner_skip_rows = inner_skip_rows;
3833 1284824 : }
3834 :
3835 : /*
3836 : * final_cost_mergejoin
3837 : * Final estimate of the cost and result size of a mergejoin path.
3838 : *
3839 : * Unlike other costsize functions, this routine makes two actual decisions:
3840 : * whether the executor will need to do mark/restore, and whether we should
3841 : * materialize the inner path. It would be logically cleaner to build
3842 : * separate paths testing these alternatives, but that would require repeating
3843 : * most of the cost calculations, which are not all that cheap. Since the
3844 : * choice will not affect output pathkeys or startup cost, only total cost,
3845 : * there is no possibility of wanting to keep more than one path. So it seems
3846 : * best to make the decisions here and record them in the path's
3847 : * skip_mark_restore and materialize_inner fields.
3848 : *
3849 : * Mark/restore overhead is usually required, but can be skipped if we know
3850 : * that the executor need find only one match per outer tuple, and that the
3851 : * mergeclauses are sufficient to identify a match.
3852 : *
3853 : * We materialize the inner path if we need mark/restore and either the inner
3854 : * path can't support mark/restore, or it's cheaper to use an interposed
3855 : * Material node to handle mark/restore.
3856 : *
3857 : * 'path' is already filled in except for the rows and cost fields and
3858 : * skip_mark_restore and materialize_inner
3859 : * 'workspace' is the result from initial_cost_mergejoin
3860 : * 'extra' contains miscellaneous information about the join
3861 : */
3862 : void
3863 327262 : final_cost_mergejoin(PlannerInfo *root, MergePath *path,
3864 : JoinCostWorkspace *workspace,
3865 : JoinPathExtraData *extra)
3866 : {
3867 327262 : Path *outer_path = path->jpath.outerjoinpath;
3868 327262 : Path *inner_path = path->jpath.innerjoinpath;
3869 327262 : double inner_path_rows = inner_path->rows;
3870 327262 : List *mergeclauses = path->path_mergeclauses;
3871 327262 : List *innersortkeys = path->innersortkeys;
3872 327262 : Cost startup_cost = workspace->startup_cost;
3873 327262 : Cost run_cost = workspace->run_cost;
3874 327262 : Cost inner_run_cost = workspace->inner_run_cost;
3875 327262 : double outer_rows = workspace->outer_rows;
3876 327262 : double inner_rows = workspace->inner_rows;
3877 327262 : double outer_skip_rows = workspace->outer_skip_rows;
3878 327262 : double inner_skip_rows = workspace->inner_skip_rows;
3879 : Cost cpu_per_tuple,
3880 : bare_inner_cost,
3881 : mat_inner_cost;
3882 : QualCost merge_qual_cost;
3883 : QualCost qp_qual_cost;
3884 : double mergejointuples,
3885 : rescannedtuples;
3886 : double rescanratio;
3887 :
3888 : /* Set the number of disabled nodes. */
3889 327262 : path->jpath.path.disabled_nodes = workspace->disabled_nodes;
3890 :
3891 : /* Protect some assumptions below that rowcounts aren't zero */
3892 327262 : if (inner_path_rows <= 0)
3893 90 : inner_path_rows = 1;
3894 :
3895 : /* Mark the path with the correct row estimate */
3896 327262 : if (path->jpath.path.param_info)
3897 764 : path->jpath.path.rows = path->jpath.path.param_info->ppi_rows;
3898 : else
3899 326498 : path->jpath.path.rows = path->jpath.path.parent->rows;
3900 :
3901 : /* For partial paths, scale row estimate. */
3902 327262 : if (path->jpath.path.parallel_workers > 0)
3903 : {
3904 9474 : double parallel_divisor = get_parallel_divisor(&path->jpath.path);
3905 :
3906 9474 : path->jpath.path.rows =
3907 9474 : clamp_row_est(path->jpath.path.rows / parallel_divisor);
3908 : }
3909 :
3910 : /*
3911 : * Compute cost of the mergequals and qpquals (other restriction clauses)
3912 : * separately.
3913 : */
3914 327262 : cost_qual_eval(&merge_qual_cost, mergeclauses, root);
3915 327262 : cost_qual_eval(&qp_qual_cost, path->jpath.joinrestrictinfo, root);
3916 327262 : qp_qual_cost.startup -= merge_qual_cost.startup;
3917 327262 : qp_qual_cost.per_tuple -= merge_qual_cost.per_tuple;
3918 :
3919 : /*
3920 : * With a SEMI or ANTI join, or if the innerrel is known unique, the
3921 : * executor will stop scanning for matches after the first match. When
3922 : * all the joinclauses are merge clauses, this means we don't ever need to
3923 : * back up the merge, and so we can skip mark/restore overhead.
3924 : */
3925 327262 : if ((path->jpath.jointype == JOIN_SEMI ||
3926 320450 : path->jpath.jointype == JOIN_ANTI ||
3927 460500 : extra->inner_unique) &&
3928 147510 : (list_length(path->jpath.joinrestrictinfo) ==
3929 147510 : list_length(path->path_mergeclauses)))
3930 123870 : path->skip_mark_restore = true;
3931 : else
3932 203392 : path->skip_mark_restore = false;
3933 :
3934 : /*
3935 : * Get approx # tuples passing the mergequals. We use approx_tuple_count
3936 : * here because we need an estimate done with JOIN_INNER semantics.
3937 : */
3938 327262 : mergejointuples = approx_tuple_count(root, &path->jpath, mergeclauses);
3939 :
3940 : /*
3941 : * When there are equal merge keys in the outer relation, the mergejoin
3942 : * must rescan any matching tuples in the inner relation. This means
3943 : * re-fetching inner tuples; we have to estimate how often that happens.
3944 : *
3945 : * For regular inner and outer joins, the number of re-fetches can be
3946 : * estimated approximately as size of merge join output minus size of
3947 : * inner relation. Assume that the distinct key values are 1, 2, ..., and
3948 : * denote the number of values of each key in the outer relation as m1,
3949 : * m2, ...; in the inner relation, n1, n2, ... Then we have
3950 : *
3951 : * size of join = m1 * n1 + m2 * n2 + ...
3952 : *
3953 : * number of rescanned tuples = (m1 - 1) * n1 + (m2 - 1) * n2 + ... = m1 *
3954 : * n1 + m2 * n2 + ... - (n1 + n2 + ...) = size of join - size of inner
3955 : * relation
3956 : *
3957 : * This equation works correctly for outer tuples having no inner match
3958 : * (nk = 0), but not for inner tuples having no outer match (mk = 0); we
3959 : * are effectively subtracting those from the number of rescanned tuples,
3960 : * when we should not. Can we do better without expensive selectivity
3961 : * computations?
3962 : *
3963 : * The whole issue is moot if we are working from a unique-ified outer
3964 : * input, or if we know we don't need to mark/restore at all.
3965 : */
3966 327262 : if (IsA(outer_path, UniquePath) || path->skip_mark_restore)
3967 126420 : rescannedtuples = 0;
3968 : else
3969 : {
3970 200842 : rescannedtuples = mergejointuples - inner_path_rows;
3971 : /* Must clamp because of possible underestimate */
3972 200842 : if (rescannedtuples < 0)
3973 84362 : rescannedtuples = 0;
3974 : }
3975 :
3976 : /*
3977 : * We'll inflate various costs this much to account for rescanning. Note
3978 : * that this is to be multiplied by something involving inner_rows, or
3979 : * another number related to the portion of the inner rel we'll scan.
3980 : */
3981 327262 : rescanratio = 1.0 + (rescannedtuples / inner_rows);
3982 :
3983 : /*
3984 : * Decide whether we want to materialize the inner input to shield it from
3985 : * mark/restore and performing re-fetches. Our cost model for regular
3986 : * re-fetches is that a re-fetch costs the same as an original fetch,
3987 : * which is probably an overestimate; but on the other hand we ignore the
3988 : * bookkeeping costs of mark/restore. Not clear if it's worth developing
3989 : * a more refined model. So we just need to inflate the inner run cost by
3990 : * rescanratio.
3991 : */
3992 327262 : bare_inner_cost = inner_run_cost * rescanratio;
3993 :
3994 : /*
3995 : * When we interpose a Material node the re-fetch cost is assumed to be
3996 : * just cpu_operator_cost per tuple, independently of the underlying
3997 : * plan's cost; and we charge an extra cpu_operator_cost per original
3998 : * fetch as well. Note that we're assuming the materialize node will
3999 : * never spill to disk, since it only has to remember tuples back to the
4000 : * last mark. (If there are a huge number of duplicates, our other cost
4001 : * factors will make the path so expensive that it probably won't get
4002 : * chosen anyway.) So we don't use cost_rescan here.
4003 : *
4004 : * Note: keep this estimate in sync with create_mergejoin_plan's labeling
4005 : * of the generated Material node.
4006 : */
4007 327262 : mat_inner_cost = inner_run_cost +
4008 327262 : cpu_operator_cost * inner_rows * rescanratio;
4009 :
4010 : /*
4011 : * If we don't need mark/restore at all, we don't need materialization.
4012 : */
4013 327262 : if (path->skip_mark_restore)
4014 123870 : path->materialize_inner = false;
4015 :
4016 : /*
4017 : * Prefer materializing if it looks cheaper, unless the user has asked to
4018 : * suppress materialization.
4019 : */
4020 203392 : else if (enable_material && mat_inner_cost < bare_inner_cost)
4021 2650 : path->materialize_inner = true;
4022 :
4023 : /*
4024 : * Even if materializing doesn't look cheaper, we *must* do it if the
4025 : * inner path is to be used directly (without sorting) and it doesn't
4026 : * support mark/restore.
4027 : *
4028 : * Since the inner side must be ordered, and only Sorts and IndexScans can
4029 : * create order to begin with, and they both support mark/restore, you
4030 : * might think there's no problem --- but you'd be wrong. Nestloop and
4031 : * merge joins can *preserve* the order of their inputs, so they can be
4032 : * selected as the input of a mergejoin, and they don't support
4033 : * mark/restore at present.
4034 : *
4035 : * We don't test the value of enable_material here, because
4036 : * materialization is required for correctness in this case, and turning
4037 : * it off does not entitle us to deliver an invalid plan.
4038 : */
4039 200742 : else if (innersortkeys == NIL &&
4040 8274 : !ExecSupportsMarkRestore(inner_path))
4041 1540 : path->materialize_inner = true;
4042 :
4043 : /*
4044 : * Also, force materializing if the inner path is to be sorted and the
4045 : * sort is expected to spill to disk. This is because the final merge
4046 : * pass can be done on-the-fly if it doesn't have to support mark/restore.
4047 : * We don't try to adjust the cost estimates for this consideration,
4048 : * though.
4049 : *
4050 : * Since materialization is a performance optimization in this case,
4051 : * rather than necessary for correctness, we skip it if enable_material is
4052 : * off.
4053 : */
4054 199202 : else if (enable_material && innersortkeys != NIL &&
4055 192420 : relation_byte_size(inner_path_rows,
4056 192420 : inner_path->pathtarget->width) >
4057 192420 : work_mem * (Size) 1024)
4058 256 : path->materialize_inner = true;
4059 : else
4060 198946 : path->materialize_inner = false;
4061 :
4062 : /* Charge the right incremental cost for the chosen case */
4063 327262 : if (path->materialize_inner)
4064 4446 : run_cost += mat_inner_cost;
4065 : else
4066 322816 : run_cost += bare_inner_cost;
4067 :
4068 : /* CPU costs */
4069 :
4070 : /*
4071 : * The number of tuple comparisons needed is approximately number of outer
4072 : * rows plus number of inner rows plus number of rescanned tuples (can we
4073 : * refine this?). At each one, we need to evaluate the mergejoin quals.
4074 : */
4075 327262 : startup_cost += merge_qual_cost.startup;
4076 327262 : startup_cost += merge_qual_cost.per_tuple *
4077 327262 : (outer_skip_rows + inner_skip_rows * rescanratio);
4078 327262 : run_cost += merge_qual_cost.per_tuple *
4079 327262 : ((outer_rows - outer_skip_rows) +
4080 327262 : (inner_rows - inner_skip_rows) * rescanratio);
4081 :
4082 : /*
4083 : * For each tuple that gets through the mergejoin proper, we charge
4084 : * cpu_tuple_cost plus the cost of evaluating additional restriction
4085 : * clauses that are to be applied at the join. (This is pessimistic since
4086 : * not all of the quals may get evaluated at each tuple.)
4087 : *
4088 : * Note: we could adjust for SEMI/ANTI joins skipping some qual
4089 : * evaluations here, but it's probably not worth the trouble.
4090 : */
4091 327262 : startup_cost += qp_qual_cost.startup;
4092 327262 : cpu_per_tuple = cpu_tuple_cost + qp_qual_cost.per_tuple;
4093 327262 : run_cost += cpu_per_tuple * mergejointuples;
4094 :
4095 : /* tlist eval costs are paid per output row, not per tuple scanned */
4096 327262 : startup_cost += path->jpath.path.pathtarget->cost.startup;
4097 327262 : run_cost += path->jpath.path.pathtarget->cost.per_tuple * path->jpath.path.rows;
4098 :
4099 327262 : path->jpath.path.startup_cost = startup_cost;
4100 327262 : path->jpath.path.total_cost = startup_cost + run_cost;
4101 327262 : }
4102 :
4103 : /*
4104 : * run mergejoinscansel() with caching
4105 : */
4106 : static MergeScanSelCache *
4107 1278684 : cached_scansel(PlannerInfo *root, RestrictInfo *rinfo, PathKey *pathkey)
4108 : {
4109 : MergeScanSelCache *cache;
4110 : ListCell *lc;
4111 : Selectivity leftstartsel,
4112 : leftendsel,
4113 : rightstartsel,
4114 : rightendsel;
4115 : MemoryContext oldcontext;
4116 :
4117 : /* Do we have this result already? */
4118 1278762 : foreach(lc, rinfo->scansel_cache)
4119 : {
4120 1160130 : cache = (MergeScanSelCache *) lfirst(lc);
4121 1160130 : if (cache->opfamily == pathkey->pk_opfamily &&
4122 1160130 : cache->collation == pathkey->pk_eclass->ec_collation &&
4123 1160130 : cache->cmptype == pathkey->pk_cmptype &&
4124 1160052 : cache->nulls_first == pathkey->pk_nulls_first)
4125 1160052 : return cache;
4126 : }
4127 :
4128 : /* Nope, do the computation */
4129 118632 : mergejoinscansel(root,
4130 118632 : (Node *) rinfo->clause,
4131 : pathkey->pk_opfamily,
4132 : pathkey->pk_cmptype,
4133 118632 : pathkey->pk_nulls_first,
4134 : &leftstartsel,
4135 : &leftendsel,
4136 : &rightstartsel,
4137 : &rightendsel);
4138 :
4139 : /* Cache the result in suitably long-lived workspace */
4140 118632 : oldcontext = MemoryContextSwitchTo(root->planner_cxt);
4141 :
4142 118632 : cache = (MergeScanSelCache *) palloc(sizeof(MergeScanSelCache));
4143 118632 : cache->opfamily = pathkey->pk_opfamily;
4144 118632 : cache->collation = pathkey->pk_eclass->ec_collation;
4145 118632 : cache->cmptype = pathkey->pk_cmptype;
4146 118632 : cache->nulls_first = pathkey->pk_nulls_first;
4147 118632 : cache->leftstartsel = leftstartsel;
4148 118632 : cache->leftendsel = leftendsel;
4149 118632 : cache->rightstartsel = rightstartsel;
4150 118632 : cache->rightendsel = rightendsel;
4151 :
4152 118632 : rinfo->scansel_cache = lappend(rinfo->scansel_cache, cache);
4153 :
4154 118632 : MemoryContextSwitchTo(oldcontext);
4155 :
4156 118632 : return cache;
4157 : }
4158 :
4159 : /*
4160 : * initial_cost_hashjoin
4161 : * Preliminary estimate of the cost of a hashjoin path.
4162 : *
4163 : * This must quickly produce lower-bound estimates of the path's startup and
4164 : * total costs. If we are unable to eliminate the proposed path from
4165 : * consideration using the lower bounds, final_cost_hashjoin will be called
4166 : * to obtain the final estimates.
4167 : *
4168 : * The exact division of labor between this function and final_cost_hashjoin
4169 : * is private to them, and represents a tradeoff between speed of the initial
4170 : * estimate and getting a tight lower bound. We choose to not examine the
4171 : * join quals here (other than by counting the number of hash clauses),
4172 : * so we can't do much with CPU costs. We do assume that
4173 : * ExecChooseHashTableSize is cheap enough to use here.
4174 : *
4175 : * 'workspace' is to be filled with startup_cost, total_cost, and perhaps
4176 : * other data to be used by final_cost_hashjoin
4177 : * 'jointype' is the type of join to be performed
4178 : * 'hashclauses' is the list of joinclauses to be used as hash clauses
4179 : * 'outer_path' is the outer input to the join
4180 : * 'inner_path' is the inner input to the join
4181 : * 'extra' contains miscellaneous information about the join
4182 : * 'parallel_hash' indicates that inner_path is partial and that a shared
4183 : * hash table will be built in parallel
4184 : */
4185 : void
4186 690778 : initial_cost_hashjoin(PlannerInfo *root, JoinCostWorkspace *workspace,
4187 : JoinType jointype,
4188 : List *hashclauses,
4189 : Path *outer_path, Path *inner_path,
4190 : JoinPathExtraData *extra,
4191 : bool parallel_hash)
4192 : {
4193 : int disabled_nodes;
4194 690778 : Cost startup_cost = 0;
4195 690778 : Cost run_cost = 0;
4196 690778 : double outer_path_rows = outer_path->rows;
4197 690778 : double inner_path_rows = inner_path->rows;
4198 690778 : double inner_path_rows_total = inner_path_rows;
4199 690778 : int num_hashclauses = list_length(hashclauses);
4200 : int numbuckets;
4201 : int numbatches;
4202 : int num_skew_mcvs;
4203 : size_t space_allowed; /* unused */
4204 :
4205 : /* Count up disabled nodes. */
4206 690778 : disabled_nodes = enable_hashjoin ? 0 : 1;
4207 690778 : disabled_nodes += inner_path->disabled_nodes;
4208 690778 : disabled_nodes += outer_path->disabled_nodes;
4209 :
4210 : /* cost of source data */
4211 690778 : startup_cost += outer_path->startup_cost;
4212 690778 : run_cost += outer_path->total_cost - outer_path->startup_cost;
4213 690778 : startup_cost += inner_path->total_cost;
4214 :
4215 : /*
4216 : * Cost of computing hash function: must do it once per input tuple. We
4217 : * charge one cpu_operator_cost for each column's hash function. Also,
4218 : * tack on one cpu_tuple_cost per inner row, to model the costs of
4219 : * inserting the row into the hashtable.
4220 : *
4221 : * XXX when a hashclause is more complex than a single operator, we really
4222 : * should charge the extra eval costs of the left or right side, as
4223 : * appropriate, here. This seems more work than it's worth at the moment.
4224 : */
4225 690778 : startup_cost += (cpu_operator_cost * num_hashclauses + cpu_tuple_cost)
4226 690778 : * inner_path_rows;
4227 690778 : run_cost += cpu_operator_cost * num_hashclauses * outer_path_rows;
4228 :
4229 : /*
4230 : * If this is a parallel hash build, then the value we have for
4231 : * inner_rows_total currently refers only to the rows returned by each
4232 : * participant. For shared hash table size estimation, we need the total
4233 : * number, so we need to undo the division.
4234 : */
4235 690778 : if (parallel_hash)
4236 12552 : inner_path_rows_total *= get_parallel_divisor(inner_path);
4237 :
4238 : /*
4239 : * Get hash table size that executor would use for inner relation.
4240 : *
4241 : * XXX for the moment, always assume that skew optimization will be
4242 : * performed. As long as SKEW_HASH_MEM_PERCENT is small, it's not worth
4243 : * trying to determine that for sure.
4244 : *
4245 : * XXX at some point it might be interesting to try to account for skew
4246 : * optimization in the cost estimate, but for now, we don't.
4247 : */
4248 690778 : ExecChooseHashTableSize(inner_path_rows_total,
4249 690778 : inner_path->pathtarget->width,
4250 : true, /* useskew */
4251 : parallel_hash, /* try_combined_hash_mem */
4252 : outer_path->parallel_workers,
4253 : &space_allowed,
4254 : &numbuckets,
4255 : &numbatches,
4256 : &num_skew_mcvs);
4257 :
4258 : /*
4259 : * If inner relation is too big then we will need to "batch" the join,
4260 : * which implies writing and reading most of the tuples to disk an extra
4261 : * time. Charge seq_page_cost per page, since the I/O should be nice and
4262 : * sequential. Writing the inner rel counts as startup cost, all the rest
4263 : * as run cost.
4264 : */
4265 690778 : if (numbatches > 1)
4266 : {
4267 5376 : double outerpages = page_size(outer_path_rows,
4268 5376 : outer_path->pathtarget->width);
4269 5376 : double innerpages = page_size(inner_path_rows,
4270 5376 : inner_path->pathtarget->width);
4271 :
4272 5376 : startup_cost += seq_page_cost * innerpages;
4273 5376 : run_cost += seq_page_cost * (innerpages + 2 * outerpages);
4274 : }
4275 :
4276 : /* CPU costs left for later */
4277 :
4278 : /* Public result fields */
4279 690778 : workspace->disabled_nodes = disabled_nodes;
4280 690778 : workspace->startup_cost = startup_cost;
4281 690778 : workspace->total_cost = startup_cost + run_cost;
4282 : /* Save private data for final_cost_hashjoin */
4283 690778 : workspace->run_cost = run_cost;
4284 690778 : workspace->numbuckets = numbuckets;
4285 690778 : workspace->numbatches = numbatches;
4286 690778 : workspace->inner_rows_total = inner_path_rows_total;
4287 690778 : }
4288 :
4289 : /*
4290 : * final_cost_hashjoin
4291 : * Final estimate of the cost and result size of a hashjoin path.
4292 : *
4293 : * Note: the numbatches estimate is also saved into 'path' for use later
4294 : *
4295 : * 'path' is already filled in except for the rows and cost fields and
4296 : * num_batches
4297 : * 'workspace' is the result from initial_cost_hashjoin
4298 : * 'extra' contains miscellaneous information about the join
4299 : */
4300 : void
4301 295816 : final_cost_hashjoin(PlannerInfo *root, HashPath *path,
4302 : JoinCostWorkspace *workspace,
4303 : JoinPathExtraData *extra)
4304 : {
4305 295816 : Path *outer_path = path->jpath.outerjoinpath;
4306 295816 : Path *inner_path = path->jpath.innerjoinpath;
4307 295816 : double outer_path_rows = outer_path->rows;
4308 295816 : double inner_path_rows = inner_path->rows;
4309 295816 : double inner_path_rows_total = workspace->inner_rows_total;
4310 295816 : List *hashclauses = path->path_hashclauses;
4311 295816 : Cost startup_cost = workspace->startup_cost;
4312 295816 : Cost run_cost = workspace->run_cost;
4313 295816 : int numbuckets = workspace->numbuckets;
4314 295816 : int numbatches = workspace->numbatches;
4315 : Cost cpu_per_tuple;
4316 : QualCost hash_qual_cost;
4317 : QualCost qp_qual_cost;
4318 : double hashjointuples;
4319 : double virtualbuckets;
4320 : Selectivity innerbucketsize;
4321 : Selectivity innermcvfreq;
4322 : ListCell *hcl;
4323 :
4324 : /* Set the number of disabled nodes. */
4325 295816 : path->jpath.path.disabled_nodes = workspace->disabled_nodes;
4326 :
4327 : /* Mark the path with the correct row estimate */
4328 295816 : if (path->jpath.path.param_info)
4329 1482 : path->jpath.path.rows = path->jpath.path.param_info->ppi_rows;
4330 : else
4331 294334 : path->jpath.path.rows = path->jpath.path.parent->rows;
4332 :
4333 : /* For partial paths, scale row estimate. */
4334 295816 : if (path->jpath.path.parallel_workers > 0)
4335 : {
4336 11332 : double parallel_divisor = get_parallel_divisor(&path->jpath.path);
4337 :
4338 11332 : path->jpath.path.rows =
4339 11332 : clamp_row_est(path->jpath.path.rows / parallel_divisor);
4340 : }
4341 :
4342 : /* mark the path with estimated # of batches */
4343 295816 : path->num_batches = numbatches;
4344 :
4345 : /* store the total number of tuples (sum of partial row estimates) */
4346 295816 : path->inner_rows_total = inner_path_rows_total;
4347 :
4348 : /* and compute the number of "virtual" buckets in the whole join */
4349 295816 : virtualbuckets = (double) numbuckets * (double) numbatches;
4350 :
4351 : /*
4352 : * Determine bucketsize fraction and MCV frequency for the inner relation.
4353 : * We use the smallest bucketsize or MCV frequency estimated for any
4354 : * individual hashclause; this is undoubtedly conservative.
4355 : *
4356 : * BUT: if inner relation has been unique-ified, we can assume it's good
4357 : * for hashing. This is important both because it's the right answer, and
4358 : * because we avoid contaminating the cache with a value that's wrong for
4359 : * non-unique-ified paths.
4360 : */
4361 295816 : if (IsA(inner_path, UniquePath))
4362 : {
4363 4656 : innerbucketsize = 1.0 / virtualbuckets;
4364 4656 : innermcvfreq = 0.0;
4365 : }
4366 : else
4367 : {
4368 : List *otherclauses;
4369 :
4370 291160 : innerbucketsize = 1.0;
4371 291160 : innermcvfreq = 1.0;
4372 :
4373 : /* At first, try to estimate bucket size using extended statistics. */
4374 291160 : otherclauses = estimate_multivariate_bucketsize(root,
4375 : inner_path->parent,
4376 : hashclauses,
4377 : &innerbucketsize);
4378 :
4379 : /* Pass through the remaining clauses */
4380 617810 : foreach(hcl, otherclauses)
4381 : {
4382 326650 : RestrictInfo *restrictinfo = lfirst_node(RestrictInfo, hcl);
4383 : Selectivity thisbucketsize;
4384 : Selectivity thismcvfreq;
4385 :
4386 : /*
4387 : * First we have to figure out which side of the hashjoin clause
4388 : * is the inner side.
4389 : *
4390 : * Since we tend to visit the same clauses over and over when
4391 : * planning a large query, we cache the bucket stats estimates in
4392 : * the RestrictInfo node to avoid repeated lookups of statistics.
4393 : */
4394 326650 : if (bms_is_subset(restrictinfo->right_relids,
4395 326650 : inner_path->parent->relids))
4396 : {
4397 : /* righthand side is inner */
4398 170980 : thisbucketsize = restrictinfo->right_bucketsize;
4399 170980 : if (thisbucketsize < 0)
4400 : {
4401 : /* not cached yet */
4402 90780 : estimate_hash_bucket_stats(root,
4403 90780 : get_rightop(restrictinfo->clause),
4404 : virtualbuckets,
4405 : &restrictinfo->right_mcvfreq,
4406 : &restrictinfo->right_bucketsize);
4407 90780 : thisbucketsize = restrictinfo->right_bucketsize;
4408 : }
4409 170980 : thismcvfreq = restrictinfo->right_mcvfreq;
4410 : }
4411 : else
4412 : {
4413 : Assert(bms_is_subset(restrictinfo->left_relids,
4414 : inner_path->parent->relids));
4415 : /* lefthand side is inner */
4416 155670 : thisbucketsize = restrictinfo->left_bucketsize;
4417 155670 : if (thisbucketsize < 0)
4418 : {
4419 : /* not cached yet */
4420 78184 : estimate_hash_bucket_stats(root,
4421 78184 : get_leftop(restrictinfo->clause),
4422 : virtualbuckets,
4423 : &restrictinfo->left_mcvfreq,
4424 : &restrictinfo->left_bucketsize);
4425 78184 : thisbucketsize = restrictinfo->left_bucketsize;
4426 : }
4427 155670 : thismcvfreq = restrictinfo->left_mcvfreq;
4428 : }
4429 :
4430 326650 : if (innerbucketsize > thisbucketsize)
4431 207802 : innerbucketsize = thisbucketsize;
4432 326650 : if (innermcvfreq > thismcvfreq)
4433 295052 : innermcvfreq = thismcvfreq;
4434 : }
4435 : }
4436 :
4437 : /*
4438 : * If the bucket holding the inner MCV would exceed hash_mem, we don't
4439 : * want to hash unless there is really no other alternative, so apply
4440 : * disable_cost. (The executor normally copes with excessive memory usage
4441 : * by splitting batches, but obviously it cannot separate equal values
4442 : * that way, so it will be unable to drive the batch size below hash_mem
4443 : * when this is true.)
4444 : */
4445 295816 : if (relation_byte_size(clamp_row_est(inner_path_rows * innermcvfreq),
4446 591632 : inner_path->pathtarget->width) > get_hash_memory_limit())
4447 6 : startup_cost += disable_cost;
4448 :
4449 : /*
4450 : * Compute cost of the hashquals and qpquals (other restriction clauses)
4451 : * separately.
4452 : */
4453 295816 : cost_qual_eval(&hash_qual_cost, hashclauses, root);
4454 295816 : cost_qual_eval(&qp_qual_cost, path->jpath.joinrestrictinfo, root);
4455 295816 : qp_qual_cost.startup -= hash_qual_cost.startup;
4456 295816 : qp_qual_cost.per_tuple -= hash_qual_cost.per_tuple;
4457 :
4458 : /* CPU costs */
4459 :
4460 295816 : if (path->jpath.jointype == JOIN_SEMI ||
4461 289848 : path->jpath.jointype == JOIN_ANTI ||
4462 284954 : extra->inner_unique)
4463 123714 : {
4464 : double outer_matched_rows;
4465 : Selectivity inner_scan_frac;
4466 :
4467 : /*
4468 : * With a SEMI or ANTI join, or if the innerrel is known unique, the
4469 : * executor will stop after the first match.
4470 : *
4471 : * For an outer-rel row that has at least one match, we can expect the
4472 : * bucket scan to stop after a fraction 1/(match_count+1) of the
4473 : * bucket's rows, if the matches are evenly distributed. Since they
4474 : * probably aren't quite evenly distributed, we apply a fuzz factor of
4475 : * 2.0 to that fraction. (If we used a larger fuzz factor, we'd have
4476 : * to clamp inner_scan_frac to at most 1.0; but since match_count is
4477 : * at least 1, no such clamp is needed now.)
4478 : */
4479 123714 : outer_matched_rows = rint(outer_path_rows * extra->semifactors.outer_match_frac);
4480 123714 : inner_scan_frac = 2.0 / (extra->semifactors.match_count + 1.0);
4481 :
4482 123714 : startup_cost += hash_qual_cost.startup;
4483 247428 : run_cost += hash_qual_cost.per_tuple * outer_matched_rows *
4484 123714 : clamp_row_est(inner_path_rows * innerbucketsize * inner_scan_frac) * 0.5;
4485 :
4486 : /*
4487 : * For unmatched outer-rel rows, the picture is quite a lot different.
4488 : * In the first place, there is no reason to assume that these rows
4489 : * preferentially hit heavily-populated buckets; instead assume they
4490 : * are uncorrelated with the inner distribution and so they see an
4491 : * average bucket size of inner_path_rows / virtualbuckets. In the
4492 : * second place, it seems likely that they will have few if any exact
4493 : * hash-code matches and so very few of the tuples in the bucket will
4494 : * actually require eval of the hash quals. We don't have any good
4495 : * way to estimate how many will, but for the moment assume that the
4496 : * effective cost per bucket entry is one-tenth what it is for
4497 : * matchable tuples.
4498 : */
4499 247428 : run_cost += hash_qual_cost.per_tuple *
4500 247428 : (outer_path_rows - outer_matched_rows) *
4501 123714 : clamp_row_est(inner_path_rows / virtualbuckets) * 0.05;
4502 :
4503 : /* Get # of tuples that will pass the basic join */
4504 123714 : if (path->jpath.jointype == JOIN_ANTI)
4505 4894 : hashjointuples = outer_path_rows - outer_matched_rows;
4506 : else
4507 118820 : hashjointuples = outer_matched_rows;
4508 : }
4509 : else
4510 : {
4511 : /*
4512 : * The number of tuple comparisons needed is the number of outer
4513 : * tuples times the typical number of tuples in a hash bucket, which
4514 : * is the inner relation size times its bucketsize fraction. At each
4515 : * one, we need to evaluate the hashjoin quals. But actually,
4516 : * charging the full qual eval cost at each tuple is pessimistic,
4517 : * since we don't evaluate the quals unless the hash values match
4518 : * exactly. For lack of a better idea, halve the cost estimate to
4519 : * allow for that.
4520 : */
4521 172102 : startup_cost += hash_qual_cost.startup;
4522 344204 : run_cost += hash_qual_cost.per_tuple * outer_path_rows *
4523 172102 : clamp_row_est(inner_path_rows * innerbucketsize) * 0.5;
4524 :
4525 : /*
4526 : * Get approx # tuples passing the hashquals. We use
4527 : * approx_tuple_count here because we need an estimate done with
4528 : * JOIN_INNER semantics.
4529 : */
4530 172102 : hashjointuples = approx_tuple_count(root, &path->jpath, hashclauses);
4531 : }
4532 :
4533 : /*
4534 : * For each tuple that gets through the hashjoin proper, we charge
4535 : * cpu_tuple_cost plus the cost of evaluating additional restriction
4536 : * clauses that are to be applied at the join. (This is pessimistic since
4537 : * not all of the quals may get evaluated at each tuple.)
4538 : */
4539 295816 : startup_cost += qp_qual_cost.startup;
4540 295816 : cpu_per_tuple = cpu_tuple_cost + qp_qual_cost.per_tuple;
4541 295816 : run_cost += cpu_per_tuple * hashjointuples;
4542 :
4543 : /* tlist eval costs are paid per output row, not per tuple scanned */
4544 295816 : startup_cost += path->jpath.path.pathtarget->cost.startup;
4545 295816 : run_cost += path->jpath.path.pathtarget->cost.per_tuple * path->jpath.path.rows;
4546 :
4547 295816 : path->jpath.path.startup_cost = startup_cost;
4548 295816 : path->jpath.path.total_cost = startup_cost + run_cost;
4549 295816 : }
4550 :
4551 :
4552 : /*
4553 : * cost_subplan
4554 : * Figure the costs for a SubPlan (or initplan).
4555 : *
4556 : * Note: we could dig the subplan's Plan out of the root list, but in practice
4557 : * all callers have it handy already, so we make them pass it.
4558 : */
4559 : void
4560 45832 : cost_subplan(PlannerInfo *root, SubPlan *subplan, Plan *plan)
4561 : {
4562 : QualCost sp_cost;
4563 :
4564 : /* Figure any cost for evaluating the testexpr */
4565 45832 : cost_qual_eval(&sp_cost,
4566 45832 : make_ands_implicit((Expr *) subplan->testexpr),
4567 : root);
4568 :
4569 45832 : if (subplan->useHashTable)
4570 : {
4571 : /*
4572 : * If we are using a hash table for the subquery outputs, then the
4573 : * cost of evaluating the query is a one-time cost. We charge one
4574 : * cpu_operator_cost per tuple for the work of loading the hashtable,
4575 : * too.
4576 : */
4577 2172 : sp_cost.startup += plan->total_cost +
4578 2172 : cpu_operator_cost * plan->plan_rows;
4579 :
4580 : /*
4581 : * The per-tuple costs include the cost of evaluating the lefthand
4582 : * expressions, plus the cost of probing the hashtable. We already
4583 : * accounted for the lefthand expressions as part of the testexpr, and
4584 : * will also have counted one cpu_operator_cost for each comparison
4585 : * operator. That is probably too low for the probing cost, but it's
4586 : * hard to make a better estimate, so live with it for now.
4587 : */
4588 : }
4589 : else
4590 : {
4591 : /*
4592 : * Otherwise we will be rescanning the subplan output on each
4593 : * evaluation. We need to estimate how much of the output we will
4594 : * actually need to scan. NOTE: this logic should agree with the
4595 : * tuple_fraction estimates used by make_subplan() in
4596 : * plan/subselect.c.
4597 : */
4598 43660 : Cost plan_run_cost = plan->total_cost - plan->startup_cost;
4599 :
4600 43660 : if (subplan->subLinkType == EXISTS_SUBLINK)
4601 : {
4602 : /* we only need to fetch 1 tuple; clamp to avoid zero divide */
4603 2644 : sp_cost.per_tuple += plan_run_cost / clamp_row_est(plan->plan_rows);
4604 : }
4605 41016 : else if (subplan->subLinkType == ALL_SUBLINK ||
4606 40998 : subplan->subLinkType == ANY_SUBLINK)
4607 : {
4608 : /* assume we need 50% of the tuples */
4609 134 : sp_cost.per_tuple += 0.50 * plan_run_cost;
4610 : /* also charge a cpu_operator_cost per row examined */
4611 134 : sp_cost.per_tuple += 0.50 * plan->plan_rows * cpu_operator_cost;
4612 : }
4613 : else
4614 : {
4615 : /* assume we need all tuples */
4616 40882 : sp_cost.per_tuple += plan_run_cost;
4617 : }
4618 :
4619 : /*
4620 : * Also account for subplan's startup cost. If the subplan is
4621 : * uncorrelated or undirect correlated, AND its topmost node is one
4622 : * that materializes its output, assume that we'll only need to pay
4623 : * its startup cost once; otherwise assume we pay the startup cost
4624 : * every time.
4625 : */
4626 57394 : if (subplan->parParam == NIL &&
4627 13734 : ExecMaterializesOutput(nodeTag(plan)))
4628 694 : sp_cost.startup += plan->startup_cost;
4629 : else
4630 42966 : sp_cost.per_tuple += plan->startup_cost;
4631 : }
4632 :
4633 45832 : subplan->startup_cost = sp_cost.startup;
4634 45832 : subplan->per_call_cost = sp_cost.per_tuple;
4635 45832 : }
4636 :
4637 :
4638 : /*
4639 : * cost_rescan
4640 : * Given a finished Path, estimate the costs of rescanning it after
4641 : * having done so the first time. For some Path types a rescan is
4642 : * cheaper than an original scan (if no parameters change), and this
4643 : * function embodies knowledge about that. The default is to return
4644 : * the same costs stored in the Path. (Note that the cost estimates
4645 : * actually stored in Paths are always for first scans.)
4646 : *
4647 : * This function is not currently intended to model effects such as rescans
4648 : * being cheaper due to disk block caching; what we are concerned with is
4649 : * plan types wherein the executor caches results explicitly, or doesn't
4650 : * redo startup calculations, etc.
4651 : */
4652 : static void
4653 2915176 : cost_rescan(PlannerInfo *root, Path *path,
4654 : Cost *rescan_startup_cost, /* output parameters */
4655 : Cost *rescan_total_cost)
4656 : {
4657 2915176 : switch (path->pathtype)
4658 : {
4659 56198 : case T_FunctionScan:
4660 :
4661 : /*
4662 : * Currently, nodeFunctionscan.c always executes the function to
4663 : * completion before returning any rows, and caches the results in
4664 : * a tuplestore. So the function eval cost is all startup cost
4665 : * and isn't paid over again on rescans. However, all run costs
4666 : * will be paid over again.
4667 : */
4668 56198 : *rescan_startup_cost = 0;
4669 56198 : *rescan_total_cost = path->total_cost - path->startup_cost;
4670 56198 : break;
4671 132126 : case T_HashJoin:
4672 :
4673 : /*
4674 : * If it's a single-batch join, we don't need to rebuild the hash
4675 : * table during a rescan.
4676 : */
4677 132126 : if (((HashPath *) path)->num_batches == 1)
4678 : {
4679 : /* Startup cost is exactly the cost of hash table building */
4680 132126 : *rescan_startup_cost = 0;
4681 132126 : *rescan_total_cost = path->total_cost - path->startup_cost;
4682 : }
4683 : else
4684 : {
4685 : /* Otherwise, no special treatment */
4686 0 : *rescan_startup_cost = path->startup_cost;
4687 0 : *rescan_total_cost = path->total_cost;
4688 : }
4689 132126 : break;
4690 8716 : case T_CteScan:
4691 : case T_WorkTableScan:
4692 : {
4693 : /*
4694 : * These plan types materialize their final result in a
4695 : * tuplestore or tuplesort object. So the rescan cost is only
4696 : * cpu_tuple_cost per tuple, unless the result is large enough
4697 : * to spill to disk.
4698 : */
4699 8716 : Cost run_cost = cpu_tuple_cost * path->rows;
4700 8716 : double nbytes = relation_byte_size(path->rows,
4701 8716 : path->pathtarget->width);
4702 8716 : double work_mem_bytes = work_mem * (Size) 1024;
4703 :
4704 8716 : if (nbytes > work_mem_bytes)
4705 : {
4706 : /* It will spill, so account for re-read cost */
4707 296 : double npages = ceil(nbytes / BLCKSZ);
4708 :
4709 296 : run_cost += seq_page_cost * npages;
4710 : }
4711 8716 : *rescan_startup_cost = 0;
4712 8716 : *rescan_total_cost = run_cost;
4713 : }
4714 8716 : break;
4715 985730 : case T_Material:
4716 : case T_Sort:
4717 : {
4718 : /*
4719 : * These plan types not only materialize their results, but do
4720 : * not implement qual filtering or projection. So they are
4721 : * even cheaper to rescan than the ones above. We charge only
4722 : * cpu_operator_cost per tuple. (Note: keep that in sync with
4723 : * the run_cost charge in cost_sort, and also see comments in
4724 : * cost_material before you change it.)
4725 : */
4726 985730 : Cost run_cost = cpu_operator_cost * path->rows;
4727 985730 : double nbytes = relation_byte_size(path->rows,
4728 985730 : path->pathtarget->width);
4729 985730 : double work_mem_bytes = work_mem * (Size) 1024;
4730 :
4731 985730 : if (nbytes > work_mem_bytes)
4732 : {
4733 : /* It will spill, so account for re-read cost */
4734 11514 : double npages = ceil(nbytes / BLCKSZ);
4735 :
4736 11514 : run_cost += seq_page_cost * npages;
4737 : }
4738 985730 : *rescan_startup_cost = 0;
4739 985730 : *rescan_total_cost = run_cost;
4740 : }
4741 985730 : break;
4742 299506 : case T_Memoize:
4743 : /* All the hard work is done by cost_memoize_rescan */
4744 299506 : cost_memoize_rescan(root, (MemoizePath *) path,
4745 : rescan_startup_cost, rescan_total_cost);
4746 299506 : break;
4747 1432900 : default:
4748 1432900 : *rescan_startup_cost = path->startup_cost;
4749 1432900 : *rescan_total_cost = path->total_cost;
4750 1432900 : break;
4751 : }
4752 2915176 : }
4753 :
4754 :
4755 : /*
4756 : * cost_qual_eval
4757 : * Estimate the CPU costs of evaluating a WHERE clause.
4758 : * The input can be either an implicitly-ANDed list of boolean
4759 : * expressions, or a list of RestrictInfo nodes. (The latter is
4760 : * preferred since it allows caching of the results.)
4761 : * The result includes both a one-time (startup) component,
4762 : * and a per-evaluation component.
4763 : *
4764 : * Note: in some code paths root can be passed as NULL, resulting in
4765 : * slightly worse estimates.
4766 : */
4767 : void
4768 3969926 : cost_qual_eval(QualCost *cost, List *quals, PlannerInfo *root)
4769 : {
4770 : cost_qual_eval_context context;
4771 : ListCell *l;
4772 :
4773 3969926 : context.root = root;
4774 3969926 : context.total.startup = 0;
4775 3969926 : context.total.per_tuple = 0;
4776 :
4777 : /* We don't charge any cost for the implicit ANDing at top level ... */
4778 :
4779 7498918 : foreach(l, quals)
4780 : {
4781 3528992 : Node *qual = (Node *) lfirst(l);
4782 :
4783 3528992 : cost_qual_eval_walker(qual, &context);
4784 : }
4785 :
4786 3969926 : *cost = context.total;
4787 3969926 : }
4788 :
4789 : /*
4790 : * cost_qual_eval_node
4791 : * As above, for a single RestrictInfo or expression.
4792 : */
4793 : void
4794 1794360 : cost_qual_eval_node(QualCost *cost, Node *qual, PlannerInfo *root)
4795 : {
4796 : cost_qual_eval_context context;
4797 :
4798 1794360 : context.root = root;
4799 1794360 : context.total.startup = 0;
4800 1794360 : context.total.per_tuple = 0;
4801 :
4802 1794360 : cost_qual_eval_walker(qual, &context);
4803 :
4804 1794360 : *cost = context.total;
4805 1794360 : }
4806 :
4807 : static bool
4808 8732252 : cost_qual_eval_walker(Node *node, cost_qual_eval_context *context)
4809 : {
4810 8732252 : if (node == NULL)
4811 87530 : return false;
4812 :
4813 : /*
4814 : * RestrictInfo nodes contain an eval_cost field reserved for this
4815 : * routine's use, so that it's not necessary to evaluate the qual clause's
4816 : * cost more than once. If the clause's cost hasn't been computed yet,
4817 : * the field's startup value will contain -1.
4818 : */
4819 8644722 : if (IsA(node, RestrictInfo))
4820 : {
4821 3712906 : RestrictInfo *rinfo = (RestrictInfo *) node;
4822 :
4823 3712906 : if (rinfo->eval_cost.startup < 0)
4824 : {
4825 : cost_qual_eval_context locContext;
4826 :
4827 579698 : locContext.root = context->root;
4828 579698 : locContext.total.startup = 0;
4829 579698 : locContext.total.per_tuple = 0;
4830 :
4831 : /*
4832 : * For an OR clause, recurse into the marked-up tree so that we
4833 : * set the eval_cost for contained RestrictInfos too.
4834 : */
4835 579698 : if (rinfo->orclause)
4836 10180 : cost_qual_eval_walker((Node *) rinfo->orclause, &locContext);
4837 : else
4838 569518 : cost_qual_eval_walker((Node *) rinfo->clause, &locContext);
4839 :
4840 : /*
4841 : * If the RestrictInfo is marked pseudoconstant, it will be tested
4842 : * only once, so treat its cost as all startup cost.
4843 : */
4844 579698 : if (rinfo->pseudoconstant)
4845 : {
4846 : /* count one execution during startup */
4847 9880 : locContext.total.startup += locContext.total.per_tuple;
4848 9880 : locContext.total.per_tuple = 0;
4849 : }
4850 579698 : rinfo->eval_cost = locContext.total;
4851 : }
4852 3712906 : context->total.startup += rinfo->eval_cost.startup;
4853 3712906 : context->total.per_tuple += rinfo->eval_cost.per_tuple;
4854 : /* do NOT recurse into children */
4855 3712906 : return false;
4856 : }
4857 :
4858 : /*
4859 : * For each operator or function node in the given tree, we charge the
4860 : * estimated execution cost given by pg_proc.procost (remember to multiply
4861 : * this by cpu_operator_cost).
4862 : *
4863 : * Vars and Consts are charged zero, and so are boolean operators (AND,
4864 : * OR, NOT). Simplistic, but a lot better than no model at all.
4865 : *
4866 : * Should we try to account for the possibility of short-circuit
4867 : * evaluation of AND/OR? Probably *not*, because that would make the
4868 : * results depend on the clause ordering, and we are not in any position
4869 : * to expect that the current ordering of the clauses is the one that's
4870 : * going to end up being used. The above per-RestrictInfo caching would
4871 : * not mix well with trying to re-order clauses anyway.
4872 : *
4873 : * Another issue that is entirely ignored here is that if a set-returning
4874 : * function is below top level in the tree, the functions/operators above
4875 : * it will need to be evaluated multiple times. In practical use, such
4876 : * cases arise so seldom as to not be worth the added complexity needed;
4877 : * moreover, since our rowcount estimates for functions tend to be pretty
4878 : * phony, the results would also be pretty phony.
4879 : */
4880 4931816 : if (IsA(node, FuncExpr))
4881 : {
4882 347878 : add_function_cost(context->root, ((FuncExpr *) node)->funcid, node,
4883 : &context->total);
4884 : }
4885 4583938 : else if (IsA(node, OpExpr) ||
4886 3948680 : IsA(node, DistinctExpr) ||
4887 3947422 : IsA(node, NullIfExpr))
4888 : {
4889 : /* rely on struct equivalence to treat these all alike */
4890 636640 : set_opfuncid((OpExpr *) node);
4891 636640 : add_function_cost(context->root, ((OpExpr *) node)->opfuncid, node,
4892 : &context->total);
4893 : }
4894 3947298 : else if (IsA(node, ScalarArrayOpExpr))
4895 : {
4896 44180 : ScalarArrayOpExpr *saop = (ScalarArrayOpExpr *) node;
4897 44180 : Node *arraynode = (Node *) lsecond(saop->args);
4898 : QualCost sacosts;
4899 : QualCost hcosts;
4900 44180 : double estarraylen = estimate_array_length(context->root, arraynode);
4901 :
4902 44180 : set_sa_opfuncid(saop);
4903 44180 : sacosts.startup = sacosts.per_tuple = 0;
4904 44180 : add_function_cost(context->root, saop->opfuncid, NULL,
4905 : &sacosts);
4906 :
4907 44180 : if (OidIsValid(saop->hashfuncid))
4908 : {
4909 : /* Handle costs for hashed ScalarArrayOpExpr */
4910 438 : hcosts.startup = hcosts.per_tuple = 0;
4911 :
4912 438 : add_function_cost(context->root, saop->hashfuncid, NULL, &hcosts);
4913 438 : context->total.startup += sacosts.startup + hcosts.startup;
4914 :
4915 : /* Estimate the cost of building the hashtable. */
4916 438 : context->total.startup += estarraylen * hcosts.per_tuple;
4917 :
4918 : /*
4919 : * XXX should we charge a little bit for sacosts.per_tuple when
4920 : * building the table, or is it ok to assume there will be zero
4921 : * hash collision?
4922 : */
4923 :
4924 : /*
4925 : * Charge for hashtable lookups. Charge a single hash and a
4926 : * single comparison.
4927 : */
4928 438 : context->total.per_tuple += hcosts.per_tuple + sacosts.per_tuple;
4929 : }
4930 : else
4931 : {
4932 : /*
4933 : * Estimate that the operator will be applied to about half of the
4934 : * array elements before the answer is determined.
4935 : */
4936 43742 : context->total.startup += sacosts.startup;
4937 87484 : context->total.per_tuple += sacosts.per_tuple *
4938 43742 : estimate_array_length(context->root, arraynode) * 0.5;
4939 : }
4940 : }
4941 3903118 : else if (IsA(node, Aggref) ||
4942 3847844 : IsA(node, WindowFunc))
4943 : {
4944 : /*
4945 : * Aggref and WindowFunc nodes are (and should be) treated like Vars,
4946 : * ie, zero execution cost in the current model, because they behave
4947 : * essentially like Vars at execution. We disregard the costs of
4948 : * their input expressions for the same reason. The actual execution
4949 : * costs of the aggregate/window functions and their arguments have to
4950 : * be factored into plan-node-specific costing of the Agg or WindowAgg
4951 : * plan node.
4952 : */
4953 58798 : return false; /* don't recurse into children */
4954 : }
4955 3844320 : else if (IsA(node, GroupingFunc))
4956 : {
4957 : /* Treat this as having cost 1 */
4958 422 : context->total.per_tuple += cpu_operator_cost;
4959 422 : return false; /* don't recurse into children */
4960 : }
4961 3843898 : else if (IsA(node, CoerceViaIO))
4962 : {
4963 22202 : CoerceViaIO *iocoerce = (CoerceViaIO *) node;
4964 : Oid iofunc;
4965 : Oid typioparam;
4966 : bool typisvarlena;
4967 :
4968 : /* check the result type's input function */
4969 22202 : getTypeInputInfo(iocoerce->resulttype,
4970 : &iofunc, &typioparam);
4971 22202 : add_function_cost(context->root, iofunc, NULL,
4972 : &context->total);
4973 : /* check the input type's output function */
4974 22202 : getTypeOutputInfo(exprType((Node *) iocoerce->arg),
4975 : &iofunc, &typisvarlena);
4976 22202 : add_function_cost(context->root, iofunc, NULL,
4977 : &context->total);
4978 : }
4979 3821696 : else if (IsA(node, ArrayCoerceExpr))
4980 : {
4981 5288 : ArrayCoerceExpr *acoerce = (ArrayCoerceExpr *) node;
4982 : QualCost perelemcost;
4983 :
4984 5288 : cost_qual_eval_node(&perelemcost, (Node *) acoerce->elemexpr,
4985 : context->root);
4986 5288 : context->total.startup += perelemcost.startup;
4987 5288 : if (perelemcost.per_tuple > 0)
4988 66 : context->total.per_tuple += perelemcost.per_tuple *
4989 66 : estimate_array_length(context->root, (Node *) acoerce->arg);
4990 : }
4991 3816408 : else if (IsA(node, RowCompareExpr))
4992 : {
4993 : /* Conservatively assume we will check all the columns */
4994 252 : RowCompareExpr *rcexpr = (RowCompareExpr *) node;
4995 : ListCell *lc;
4996 :
4997 810 : foreach(lc, rcexpr->opnos)
4998 : {
4999 558 : Oid opid = lfirst_oid(lc);
5000 :
5001 558 : add_function_cost(context->root, get_opcode(opid), NULL,
5002 : &context->total);
5003 : }
5004 : }
5005 3816156 : else if (IsA(node, MinMaxExpr) ||
5006 3815896 : IsA(node, SQLValueFunction) ||
5007 3811126 : IsA(node, XmlExpr) ||
5008 3810424 : IsA(node, CoerceToDomain) ||
5009 3800566 : IsA(node, NextValueExpr) ||
5010 3800204 : IsA(node, JsonExpr))
5011 : {
5012 : /* Treat all these as having cost 1 */
5013 18500 : context->total.per_tuple += cpu_operator_cost;
5014 : }
5015 3797656 : else if (IsA(node, SubLink))
5016 : {
5017 : /* This routine should not be applied to un-planned expressions */
5018 0 : elog(ERROR, "cannot handle unplanned sub-select");
5019 : }
5020 3797656 : else if (IsA(node, SubPlan))
5021 : {
5022 : /*
5023 : * A subplan node in an expression typically indicates that the
5024 : * subplan will be executed on each evaluation, so charge accordingly.
5025 : * (Sub-selects that can be executed as InitPlans have already been
5026 : * removed from the expression.)
5027 : */
5028 45338 : SubPlan *subplan = (SubPlan *) node;
5029 :
5030 45338 : context->total.startup += subplan->startup_cost;
5031 45338 : context->total.per_tuple += subplan->per_call_cost;
5032 :
5033 : /*
5034 : * We don't want to recurse into the testexpr, because it was already
5035 : * counted in the SubPlan node's costs. So we're done.
5036 : */
5037 45338 : return false;
5038 : }
5039 3752318 : else if (IsA(node, AlternativeSubPlan))
5040 : {
5041 : /*
5042 : * Arbitrarily use the first alternative plan for costing. (We should
5043 : * certainly only include one alternative, and we don't yet have
5044 : * enough information to know which one the executor is most likely to
5045 : * use.)
5046 : */
5047 1910 : AlternativeSubPlan *asplan = (AlternativeSubPlan *) node;
5048 :
5049 1910 : return cost_qual_eval_walker((Node *) linitial(asplan->subplans),
5050 : context);
5051 : }
5052 3750408 : else if (IsA(node, PlaceHolderVar))
5053 : {
5054 : /*
5055 : * A PlaceHolderVar should be given cost zero when considering general
5056 : * expression evaluation costs. The expense of doing the contained
5057 : * expression is charged as part of the tlist eval costs of the scan
5058 : * or join where the PHV is first computed (see set_rel_width and
5059 : * add_placeholders_to_joinrel). If we charged it again here, we'd be
5060 : * double-counting the cost for each level of plan that the PHV
5061 : * bubbles up through. Hence, return without recursing into the
5062 : * phexpr.
5063 : */
5064 5154 : return false;
5065 : }
5066 :
5067 : /* recurse into children */
5068 4820194 : return expression_tree_walker(node, cost_qual_eval_walker, context);
5069 : }
5070 :
5071 : /*
5072 : * get_restriction_qual_cost
5073 : * Compute evaluation costs of a baserel's restriction quals, plus any
5074 : * movable join quals that have been pushed down to the scan.
5075 : * Results are returned into *qpqual_cost.
5076 : *
5077 : * This is a convenience subroutine that works for seqscans and other cases
5078 : * where all the given quals will be evaluated the hard way. It's not useful
5079 : * for cost_index(), for example, where the index machinery takes care of
5080 : * some of the quals. We assume baserestrictcost was previously set by
5081 : * set_baserel_size_estimates().
5082 : */
5083 : static void
5084 1060128 : get_restriction_qual_cost(PlannerInfo *root, RelOptInfo *baserel,
5085 : ParamPathInfo *param_info,
5086 : QualCost *qpqual_cost)
5087 : {
5088 1060128 : if (param_info)
5089 : {
5090 : /* Include costs of pushed-down clauses */
5091 232504 : cost_qual_eval(qpqual_cost, param_info->ppi_clauses, root);
5092 :
5093 232504 : qpqual_cost->startup += baserel->baserestrictcost.startup;
5094 232504 : qpqual_cost->per_tuple += baserel->baserestrictcost.per_tuple;
5095 : }
5096 : else
5097 827624 : *qpqual_cost = baserel->baserestrictcost;
5098 1060128 : }
5099 :
5100 :
5101 : /*
5102 : * compute_semi_anti_join_factors
5103 : * Estimate how much of the inner input a SEMI, ANTI, or inner_unique join
5104 : * can be expected to scan.
5105 : *
5106 : * In a hash or nestloop SEMI/ANTI join, the executor will stop scanning
5107 : * inner rows as soon as it finds a match to the current outer row.
5108 : * The same happens if we have detected the inner rel is unique.
5109 : * We should therefore adjust some of the cost components for this effect.
5110 : * This function computes some estimates needed for these adjustments.
5111 : * These estimates will be the same regardless of the particular paths used
5112 : * for the outer and inner relation, so we compute these once and then pass
5113 : * them to all the join cost estimation functions.
5114 : *
5115 : * Input parameters:
5116 : * joinrel: join relation under consideration
5117 : * outerrel: outer relation under consideration
5118 : * innerrel: inner relation under consideration
5119 : * jointype: if not JOIN_SEMI or JOIN_ANTI, we assume it's inner_unique
5120 : * sjinfo: SpecialJoinInfo relevant to this join
5121 : * restrictlist: join quals
5122 : * Output parameters:
5123 : * *semifactors is filled in (see pathnodes.h for field definitions)
5124 : */
5125 : void
5126 213488 : compute_semi_anti_join_factors(PlannerInfo *root,
5127 : RelOptInfo *joinrel,
5128 : RelOptInfo *outerrel,
5129 : RelOptInfo *innerrel,
5130 : JoinType jointype,
5131 : SpecialJoinInfo *sjinfo,
5132 : List *restrictlist,
5133 : SemiAntiJoinFactors *semifactors)
5134 : {
5135 : Selectivity jselec;
5136 : Selectivity nselec;
5137 : Selectivity avgmatch;
5138 : SpecialJoinInfo norm_sjinfo;
5139 : List *joinquals;
5140 : ListCell *l;
5141 :
5142 : /*
5143 : * In an ANTI join, we must ignore clauses that are "pushed down", since
5144 : * those won't affect the match logic. In a SEMI join, we do not
5145 : * distinguish joinquals from "pushed down" quals, so just use the whole
5146 : * restrictinfo list. For other outer join types, we should consider only
5147 : * non-pushed-down quals, so that this devolves to an IS_OUTER_JOIN check.
5148 : */
5149 213488 : if (IS_OUTER_JOIN(jointype))
5150 : {
5151 79452 : joinquals = NIL;
5152 175456 : foreach(l, restrictlist)
5153 : {
5154 96004 : RestrictInfo *rinfo = lfirst_node(RestrictInfo, l);
5155 :
5156 96004 : if (!RINFO_IS_PUSHED_DOWN(rinfo, joinrel->relids))
5157 90246 : joinquals = lappend(joinquals, rinfo);
5158 : }
5159 : }
5160 : else
5161 134036 : joinquals = restrictlist;
5162 :
5163 : /*
5164 : * Get the JOIN_SEMI or JOIN_ANTI selectivity of the join clauses.
5165 : */
5166 213488 : jselec = clauselist_selectivity(root,
5167 : joinquals,
5168 : 0,
5169 : (jointype == JOIN_ANTI) ? JOIN_ANTI : JOIN_SEMI,
5170 : sjinfo);
5171 :
5172 : /*
5173 : * Also get the normal inner-join selectivity of the join clauses.
5174 : */
5175 213488 : init_dummy_sjinfo(&norm_sjinfo, outerrel->relids, innerrel->relids);
5176 :
5177 213488 : nselec = clauselist_selectivity(root,
5178 : joinquals,
5179 : 0,
5180 : JOIN_INNER,
5181 : &norm_sjinfo);
5182 :
5183 : /* Avoid leaking a lot of ListCells */
5184 213488 : if (IS_OUTER_JOIN(jointype))
5185 79452 : list_free(joinquals);
5186 :
5187 : /*
5188 : * jselec can be interpreted as the fraction of outer-rel rows that have
5189 : * any matches (this is true for both SEMI and ANTI cases). And nselec is
5190 : * the fraction of the Cartesian product that matches. So, the average
5191 : * number of matches for each outer-rel row that has at least one match is
5192 : * nselec * inner_rows / jselec.
5193 : *
5194 : * Note: it is correct to use the inner rel's "rows" count here, even
5195 : * though we might later be considering a parameterized inner path with
5196 : * fewer rows. This is because we have included all the join clauses in
5197 : * the selectivity estimate.
5198 : */
5199 213488 : if (jselec > 0) /* protect against zero divide */
5200 : {
5201 212984 : avgmatch = nselec * innerrel->rows / jselec;
5202 : /* Clamp to sane range */
5203 212984 : avgmatch = Max(1.0, avgmatch);
5204 : }
5205 : else
5206 504 : avgmatch = 1.0;
5207 :
5208 213488 : semifactors->outer_match_frac = jselec;
5209 213488 : semifactors->match_count = avgmatch;
5210 213488 : }
5211 :
5212 : /*
5213 : * has_indexed_join_quals
5214 : * Check whether all the joinquals of a nestloop join are used as
5215 : * inner index quals.
5216 : *
5217 : * If the inner path of a SEMI/ANTI join is an indexscan (including bitmap
5218 : * indexscan) that uses all the joinquals as indexquals, we can assume that an
5219 : * unmatched outer tuple is cheap to process, whereas otherwise it's probably
5220 : * expensive.
5221 : */
5222 : static bool
5223 910432 : has_indexed_join_quals(NestPath *path)
5224 : {
5225 910432 : JoinPath *joinpath = &path->jpath;
5226 910432 : Relids joinrelids = joinpath->path.parent->relids;
5227 910432 : Path *innerpath = joinpath->innerjoinpath;
5228 : List *indexclauses;
5229 : bool found_one;
5230 : ListCell *lc;
5231 :
5232 : /* If join still has quals to evaluate, it's not fast */
5233 910432 : if (joinpath->joinrestrictinfo != NIL)
5234 645950 : return false;
5235 : /* Nor if the inner path isn't parameterized at all */
5236 264482 : if (innerpath->param_info == NULL)
5237 4800 : return false;
5238 :
5239 : /* Find the indexclauses list for the inner scan */
5240 259682 : switch (innerpath->pathtype)
5241 : {
5242 154940 : case T_IndexScan:
5243 : case T_IndexOnlyScan:
5244 154940 : indexclauses = ((IndexPath *) innerpath)->indexclauses;
5245 154940 : break;
5246 270 : case T_BitmapHeapScan:
5247 : {
5248 : /* Accept only a simple bitmap scan, not AND/OR cases */
5249 270 : Path *bmqual = ((BitmapHeapPath *) innerpath)->bitmapqual;
5250 :
5251 270 : if (IsA(bmqual, IndexPath))
5252 222 : indexclauses = ((IndexPath *) bmqual)->indexclauses;
5253 : else
5254 48 : return false;
5255 222 : break;
5256 : }
5257 104472 : default:
5258 :
5259 : /*
5260 : * If it's not a simple indexscan, it probably doesn't run quickly
5261 : * for zero rows out, even if it's a parameterized path using all
5262 : * the joinquals.
5263 : */
5264 104472 : return false;
5265 : }
5266 :
5267 : /*
5268 : * Examine the inner path's param clauses. Any that are from the outer
5269 : * path must be found in the indexclauses list, either exactly or in an
5270 : * equivalent form generated by equivclass.c. Also, we must find at least
5271 : * one such clause, else it's a clauseless join which isn't fast.
5272 : */
5273 155162 : found_one = false;
5274 308648 : foreach(lc, innerpath->param_info->ppi_clauses)
5275 : {
5276 158366 : RestrictInfo *rinfo = (RestrictInfo *) lfirst(lc);
5277 :
5278 158366 : if (join_clause_is_movable_into(rinfo,
5279 158366 : innerpath->parent->relids,
5280 : joinrelids))
5281 : {
5282 157862 : if (!is_redundant_with_indexclauses(rinfo, indexclauses))
5283 4880 : return false;
5284 152982 : found_one = true;
5285 : }
5286 : }
5287 150282 : return found_one;
5288 : }
5289 :
5290 :
5291 : /*
5292 : * approx_tuple_count
5293 : * Quick-and-dirty estimation of the number of join rows passing
5294 : * a set of qual conditions.
5295 : *
5296 : * The quals can be either an implicitly-ANDed list of boolean expressions,
5297 : * or a list of RestrictInfo nodes (typically the latter).
5298 : *
5299 : * We intentionally compute the selectivity under JOIN_INNER rules, even
5300 : * if it's some type of outer join. This is appropriate because we are
5301 : * trying to figure out how many tuples pass the initial merge or hash
5302 : * join step.
5303 : *
5304 : * This is quick-and-dirty because we bypass clauselist_selectivity, and
5305 : * simply multiply the independent clause selectivities together. Now
5306 : * clauselist_selectivity often can't do any better than that anyhow, but
5307 : * for some situations (such as range constraints) it is smarter. However,
5308 : * we can't effectively cache the results of clauselist_selectivity, whereas
5309 : * the individual clause selectivities can be and are cached.
5310 : *
5311 : * Since we are only using the results to estimate how many potential
5312 : * output tuples are generated and passed through qpqual checking, it
5313 : * seems OK to live with the approximation.
5314 : */
5315 : static double
5316 499364 : approx_tuple_count(PlannerInfo *root, JoinPath *path, List *quals)
5317 : {
5318 : double tuples;
5319 499364 : double outer_tuples = path->outerjoinpath->rows;
5320 499364 : double inner_tuples = path->innerjoinpath->rows;
5321 : SpecialJoinInfo sjinfo;
5322 499364 : Selectivity selec = 1.0;
5323 : ListCell *l;
5324 :
5325 : /*
5326 : * Make up a SpecialJoinInfo for JOIN_INNER semantics.
5327 : */
5328 499364 : init_dummy_sjinfo(&sjinfo, path->outerjoinpath->parent->relids,
5329 499364 : path->innerjoinpath->parent->relids);
5330 :
5331 : /* Get the approximate selectivity */
5332 1080070 : foreach(l, quals)
5333 : {
5334 580706 : Node *qual = (Node *) lfirst(l);
5335 :
5336 : /* Note that clause_selectivity will be able to cache its result */
5337 580706 : selec *= clause_selectivity(root, qual, 0, JOIN_INNER, &sjinfo);
5338 : }
5339 :
5340 : /* Apply it to the input relation sizes */
5341 499364 : tuples = selec * outer_tuples * inner_tuples;
5342 :
5343 499364 : return clamp_row_est(tuples);
5344 : }
5345 :
5346 :
5347 : /*
5348 : * set_baserel_size_estimates
5349 : * Set the size estimates for the given base relation.
5350 : *
5351 : * The rel's targetlist and restrictinfo list must have been constructed
5352 : * already, and rel->tuples must be set.
5353 : *
5354 : * We set the following fields of the rel node:
5355 : * rows: the estimated number of output tuples (after applying
5356 : * restriction clauses).
5357 : * width: the estimated average output tuple width in bytes.
5358 : * baserestrictcost: estimated cost of evaluating baserestrictinfo clauses.
5359 : */
5360 : void
5361 501452 : set_baserel_size_estimates(PlannerInfo *root, RelOptInfo *rel)
5362 : {
5363 : double nrows;
5364 :
5365 : /* Should only be applied to base relations */
5366 : Assert(rel->relid > 0);
5367 :
5368 1002874 : nrows = rel->tuples *
5369 501452 : clauselist_selectivity(root,
5370 : rel->baserestrictinfo,
5371 : 0,
5372 : JOIN_INNER,
5373 : NULL);
5374 :
5375 501422 : rel->rows = clamp_row_est(nrows);
5376 :
5377 501422 : cost_qual_eval(&rel->baserestrictcost, rel->baserestrictinfo, root);
5378 :
5379 501422 : set_rel_width(root, rel);
5380 501422 : }
5381 :
5382 : /*
5383 : * get_parameterized_baserel_size
5384 : * Make a size estimate for a parameterized scan of a base relation.
5385 : *
5386 : * 'param_clauses' lists the additional join clauses to be used.
5387 : *
5388 : * set_baserel_size_estimates must have been applied already.
5389 : */
5390 : double
5391 149102 : get_parameterized_baserel_size(PlannerInfo *root, RelOptInfo *rel,
5392 : List *param_clauses)
5393 : {
5394 : List *allclauses;
5395 : double nrows;
5396 :
5397 : /*
5398 : * Estimate the number of rows returned by the parameterized scan, knowing
5399 : * that it will apply all the extra join clauses as well as the rel's own
5400 : * restriction clauses. Note that we force the clauses to be treated as
5401 : * non-join clauses during selectivity estimation.
5402 : */
5403 149102 : allclauses = list_concat_copy(param_clauses, rel->baserestrictinfo);
5404 298204 : nrows = rel->tuples *
5405 149102 : clauselist_selectivity(root,
5406 : allclauses,
5407 149102 : rel->relid, /* do not use 0! */
5408 : JOIN_INNER,
5409 : NULL);
5410 149102 : nrows = clamp_row_est(nrows);
5411 : /* For safety, make sure result is not more than the base estimate */
5412 149102 : if (nrows > rel->rows)
5413 0 : nrows = rel->rows;
5414 149102 : return nrows;
5415 : }
5416 :
5417 : /*
5418 : * set_joinrel_size_estimates
5419 : * Set the size estimates for the given join relation.
5420 : *
5421 : * The rel's targetlist must have been constructed already, and a
5422 : * restriction clause list that matches the given component rels must
5423 : * be provided.
5424 : *
5425 : * Since there is more than one way to make a joinrel for more than two
5426 : * base relations, the results we get here could depend on which component
5427 : * rel pair is provided. In theory we should get the same answers no matter
5428 : * which pair is provided; in practice, since the selectivity estimation
5429 : * routines don't handle all cases equally well, we might not. But there's
5430 : * not much to be done about it. (Would it make sense to repeat the
5431 : * calculations for each pair of input rels that's encountered, and somehow
5432 : * average the results? Probably way more trouble than it's worth, and
5433 : * anyway we must keep the rowcount estimate the same for all paths for the
5434 : * joinrel.)
5435 : *
5436 : * We set only the rows field here. The reltarget field was already set by
5437 : * build_joinrel_tlist, and baserestrictcost is not used for join rels.
5438 : */
5439 : void
5440 216968 : set_joinrel_size_estimates(PlannerInfo *root, RelOptInfo *rel,
5441 : RelOptInfo *outer_rel,
5442 : RelOptInfo *inner_rel,
5443 : SpecialJoinInfo *sjinfo,
5444 : List *restrictlist)
5445 : {
5446 216968 : rel->rows = calc_joinrel_size_estimate(root,
5447 : rel,
5448 : outer_rel,
5449 : inner_rel,
5450 : outer_rel->rows,
5451 : inner_rel->rows,
5452 : sjinfo,
5453 : restrictlist);
5454 216968 : }
5455 :
5456 : /*
5457 : * get_parameterized_joinrel_size
5458 : * Make a size estimate for a parameterized scan of a join relation.
5459 : *
5460 : * 'rel' is the joinrel under consideration.
5461 : * 'outer_path', 'inner_path' are (probably also parameterized) Paths that
5462 : * produce the relations being joined.
5463 : * 'sjinfo' is any SpecialJoinInfo relevant to this join.
5464 : * 'restrict_clauses' lists the join clauses that need to be applied at the
5465 : * join node (including any movable clauses that were moved down to this join,
5466 : * and not including any movable clauses that were pushed down into the
5467 : * child paths).
5468 : *
5469 : * set_joinrel_size_estimates must have been applied already.
5470 : */
5471 : double
5472 8170 : get_parameterized_joinrel_size(PlannerInfo *root, RelOptInfo *rel,
5473 : Path *outer_path,
5474 : Path *inner_path,
5475 : SpecialJoinInfo *sjinfo,
5476 : List *restrict_clauses)
5477 : {
5478 : double nrows;
5479 :
5480 : /*
5481 : * Estimate the number of rows returned by the parameterized join as the
5482 : * sizes of the input paths times the selectivity of the clauses that have
5483 : * ended up at this join node.
5484 : *
5485 : * As with set_joinrel_size_estimates, the rowcount estimate could depend
5486 : * on the pair of input paths provided, though ideally we'd get the same
5487 : * estimate for any pair with the same parameterization.
5488 : */
5489 8170 : nrows = calc_joinrel_size_estimate(root,
5490 : rel,
5491 : outer_path->parent,
5492 : inner_path->parent,
5493 : outer_path->rows,
5494 : inner_path->rows,
5495 : sjinfo,
5496 : restrict_clauses);
5497 : /* For safety, make sure result is not more than the base estimate */
5498 8170 : if (nrows > rel->rows)
5499 12 : nrows = rel->rows;
5500 8170 : return nrows;
5501 : }
5502 :
5503 : /*
5504 : * calc_joinrel_size_estimate
5505 : * Workhorse for set_joinrel_size_estimates and
5506 : * get_parameterized_joinrel_size.
5507 : *
5508 : * outer_rel/inner_rel are the relations being joined, but they should be
5509 : * assumed to have sizes outer_rows/inner_rows; those numbers might be less
5510 : * than what rel->rows says, when we are considering parameterized paths.
5511 : */
5512 : static double
5513 225138 : calc_joinrel_size_estimate(PlannerInfo *root,
5514 : RelOptInfo *joinrel,
5515 : RelOptInfo *outer_rel,
5516 : RelOptInfo *inner_rel,
5517 : double outer_rows,
5518 : double inner_rows,
5519 : SpecialJoinInfo *sjinfo,
5520 : List *restrictlist)
5521 : {
5522 225138 : JoinType jointype = sjinfo->jointype;
5523 : Selectivity fkselec;
5524 : Selectivity jselec;
5525 : Selectivity pselec;
5526 : double nrows;
5527 :
5528 : /*
5529 : * Compute joinclause selectivity. Note that we are only considering
5530 : * clauses that become restriction clauses at this join level; we are not
5531 : * double-counting them because they were not considered in estimating the
5532 : * sizes of the component rels.
5533 : *
5534 : * First, see whether any of the joinclauses can be matched to known FK
5535 : * constraints. If so, drop those clauses from the restrictlist, and
5536 : * instead estimate their selectivity using FK semantics. (We do this
5537 : * without regard to whether said clauses are local or "pushed down".
5538 : * Probably, an FK-matching clause could never be seen as pushed down at
5539 : * an outer join, since it would be strict and hence would be grounds for
5540 : * join strength reduction.) fkselec gets the net selectivity for
5541 : * FK-matching clauses, or 1.0 if there are none.
5542 : */
5543 225138 : fkselec = get_foreign_key_join_selectivity(root,
5544 : outer_rel->relids,
5545 : inner_rel->relids,
5546 : sjinfo,
5547 : &restrictlist);
5548 :
5549 : /*
5550 : * For an outer join, we have to distinguish the selectivity of the join's
5551 : * own clauses (JOIN/ON conditions) from any clauses that were "pushed
5552 : * down". For inner joins we just count them all as joinclauses.
5553 : */
5554 225138 : if (IS_OUTER_JOIN(jointype))
5555 : {
5556 85886 : List *joinquals = NIL;
5557 85886 : List *pushedquals = NIL;
5558 : ListCell *l;
5559 :
5560 : /* Grovel through the clauses to separate into two lists */
5561 195264 : foreach(l, restrictlist)
5562 : {
5563 109378 : RestrictInfo *rinfo = lfirst_node(RestrictInfo, l);
5564 :
5565 109378 : if (RINFO_IS_PUSHED_DOWN(rinfo, joinrel->relids))
5566 4764 : pushedquals = lappend(pushedquals, rinfo);
5567 : else
5568 104614 : joinquals = lappend(joinquals, rinfo);
5569 : }
5570 :
5571 : /* Get the separate selectivities */
5572 85886 : jselec = clauselist_selectivity(root,
5573 : joinquals,
5574 : 0,
5575 : jointype,
5576 : sjinfo);
5577 85886 : pselec = clauselist_selectivity(root,
5578 : pushedquals,
5579 : 0,
5580 : jointype,
5581 : sjinfo);
5582 :
5583 : /* Avoid leaking a lot of ListCells */
5584 85886 : list_free(joinquals);
5585 85886 : list_free(pushedquals);
5586 : }
5587 : else
5588 : {
5589 139252 : jselec = clauselist_selectivity(root,
5590 : restrictlist,
5591 : 0,
5592 : jointype,
5593 : sjinfo);
5594 139252 : pselec = 0.0; /* not used, keep compiler quiet */
5595 : }
5596 :
5597 : /*
5598 : * Basically, we multiply size of Cartesian product by selectivity.
5599 : *
5600 : * If we are doing an outer join, take that into account: the joinqual
5601 : * selectivity has to be clamped using the knowledge that the output must
5602 : * be at least as large as the non-nullable input. However, any
5603 : * pushed-down quals are applied after the outer join, so their
5604 : * selectivity applies fully.
5605 : *
5606 : * For JOIN_SEMI and JOIN_ANTI, the selectivity is defined as the fraction
5607 : * of LHS rows that have matches, and we apply that straightforwardly.
5608 : */
5609 225138 : switch (jointype)
5610 : {
5611 131444 : case JOIN_INNER:
5612 131444 : nrows = outer_rows * inner_rows * fkselec * jselec;
5613 : /* pselec not used */
5614 131444 : break;
5615 78838 : case JOIN_LEFT:
5616 78838 : nrows = outer_rows * inner_rows * fkselec * jselec;
5617 78838 : if (nrows < outer_rows)
5618 32184 : nrows = outer_rows;
5619 78838 : nrows *= pselec;
5620 78838 : break;
5621 1714 : case JOIN_FULL:
5622 1714 : nrows = outer_rows * inner_rows * fkselec * jselec;
5623 1714 : if (nrows < outer_rows)
5624 1136 : nrows = outer_rows;
5625 1714 : if (nrows < inner_rows)
5626 120 : nrows = inner_rows;
5627 1714 : nrows *= pselec;
5628 1714 : break;
5629 7808 : case JOIN_SEMI:
5630 7808 : nrows = outer_rows * fkselec * jselec;
5631 : /* pselec not used */
5632 7808 : break;
5633 5334 : case JOIN_ANTI:
5634 5334 : nrows = outer_rows * (1.0 - fkselec * jselec);
5635 5334 : nrows *= pselec;
5636 5334 : break;
5637 0 : default:
5638 : /* other values not expected here */
5639 0 : elog(ERROR, "unrecognized join type: %d", (int) jointype);
5640 : nrows = 0; /* keep compiler quiet */
5641 : break;
5642 : }
5643 :
5644 225138 : return clamp_row_est(nrows);
5645 : }
5646 :
5647 : /*
5648 : * get_foreign_key_join_selectivity
5649 : * Estimate join selectivity for foreign-key-related clauses.
5650 : *
5651 : * Remove any clauses that can be matched to FK constraints from *restrictlist,
5652 : * and return a substitute estimate of their selectivity. 1.0 is returned
5653 : * when there are no such clauses.
5654 : *
5655 : * The reason for treating such clauses specially is that we can get better
5656 : * estimates this way than by relying on clauselist_selectivity(), especially
5657 : * for multi-column FKs where that function's assumption that the clauses are
5658 : * independent falls down badly. But even with single-column FKs, we may be
5659 : * able to get a better answer when the pg_statistic stats are missing or out
5660 : * of date.
5661 : */
5662 : static Selectivity
5663 225138 : get_foreign_key_join_selectivity(PlannerInfo *root,
5664 : Relids outer_relids,
5665 : Relids inner_relids,
5666 : SpecialJoinInfo *sjinfo,
5667 : List **restrictlist)
5668 : {
5669 225138 : Selectivity fkselec = 1.0;
5670 225138 : JoinType jointype = sjinfo->jointype;
5671 225138 : List *worklist = *restrictlist;
5672 : ListCell *lc;
5673 :
5674 : /* Consider each FK constraint that is known to match the query */
5675 227154 : foreach(lc, root->fkey_list)
5676 : {
5677 2016 : ForeignKeyOptInfo *fkinfo = (ForeignKeyOptInfo *) lfirst(lc);
5678 : bool ref_is_outer;
5679 : List *removedlist;
5680 : ListCell *cell;
5681 :
5682 : /*
5683 : * This FK is not relevant unless it connects a baserel on one side of
5684 : * this join to a baserel on the other side.
5685 : */
5686 3680 : if (bms_is_member(fkinfo->con_relid, outer_relids) &&
5687 1664 : bms_is_member(fkinfo->ref_relid, inner_relids))
5688 1490 : ref_is_outer = false;
5689 866 : else if (bms_is_member(fkinfo->ref_relid, outer_relids) &&
5690 340 : bms_is_member(fkinfo->con_relid, inner_relids))
5691 130 : ref_is_outer = true;
5692 : else
5693 396 : continue;
5694 :
5695 : /*
5696 : * If we're dealing with a semi/anti join, and the FK's referenced
5697 : * relation is on the outside, then knowledge of the FK doesn't help
5698 : * us figure out what we need to know (which is the fraction of outer
5699 : * rows that have matches). On the other hand, if the referenced rel
5700 : * is on the inside, then all outer rows must have matches in the
5701 : * referenced table (ignoring nulls). But any restriction or join
5702 : * clauses that filter that table will reduce the fraction of matches.
5703 : * We can account for restriction clauses, but it's too hard to guess
5704 : * how many table rows would get through a join that's inside the RHS.
5705 : * Hence, if either case applies, punt and ignore the FK.
5706 : */
5707 1620 : if ((jointype == JOIN_SEMI || jointype == JOIN_ANTI) &&
5708 1098 : (ref_is_outer || bms_membership(inner_relids) != BMS_SINGLETON))
5709 12 : continue;
5710 :
5711 : /*
5712 : * Modify the restrictlist by removing clauses that match the FK (and
5713 : * putting them into removedlist instead). It seems unsafe to modify
5714 : * the originally-passed List structure, so we make a shallow copy the
5715 : * first time through.
5716 : */
5717 1608 : if (worklist == *restrictlist)
5718 1370 : worklist = list_copy(worklist);
5719 :
5720 1608 : removedlist = NIL;
5721 3364 : foreach(cell, worklist)
5722 : {
5723 1756 : RestrictInfo *rinfo = (RestrictInfo *) lfirst(cell);
5724 1756 : bool remove_it = false;
5725 : int i;
5726 :
5727 : /* Drop this clause if it matches any column of the FK */
5728 2228 : for (i = 0; i < fkinfo->nkeys; i++)
5729 : {
5730 2198 : if (rinfo->parent_ec)
5731 : {
5732 : /*
5733 : * EC-derived clauses can only match by EC. It is okay to
5734 : * consider any clause derived from the same EC as
5735 : * matching the FK: even if equivclass.c chose to generate
5736 : * a clause equating some other pair of Vars, it could
5737 : * have generated one equating the FK's Vars. So for
5738 : * purposes of estimation, we can act as though it did so.
5739 : *
5740 : * Note: checking parent_ec is a bit of a cheat because
5741 : * there are EC-derived clauses that don't have parent_ec
5742 : * set; but such clauses must compare expressions that
5743 : * aren't just Vars, so they cannot match the FK anyway.
5744 : */
5745 304 : if (fkinfo->eclass[i] == rinfo->parent_ec)
5746 : {
5747 298 : remove_it = true;
5748 298 : break;
5749 : }
5750 : }
5751 : else
5752 : {
5753 : /*
5754 : * Otherwise, see if rinfo was previously matched to FK as
5755 : * a "loose" clause.
5756 : */
5757 1894 : if (list_member_ptr(fkinfo->rinfos[i], rinfo))
5758 : {
5759 1428 : remove_it = true;
5760 1428 : break;
5761 : }
5762 : }
5763 : }
5764 1756 : if (remove_it)
5765 : {
5766 1726 : worklist = foreach_delete_current(worklist, cell);
5767 1726 : removedlist = lappend(removedlist, rinfo);
5768 : }
5769 : }
5770 :
5771 : /*
5772 : * If we failed to remove all the matching clauses we expected to
5773 : * find, chicken out and ignore this FK; applying its selectivity
5774 : * might result in double-counting. Put any clauses we did manage to
5775 : * remove back into the worklist.
5776 : *
5777 : * Since the matching clauses are known not outerjoin-delayed, they
5778 : * would normally have appeared in the initial joinclause list. If we
5779 : * didn't find them, there are two possibilities:
5780 : *
5781 : * 1. If the FK match is based on an EC that is ec_has_const, it won't
5782 : * have generated any join clauses at all. We discount such ECs while
5783 : * checking to see if we have "all" the clauses. (Below, we'll adjust
5784 : * the selectivity estimate for this case.)
5785 : *
5786 : * 2. The clauses were matched to some other FK in a previous
5787 : * iteration of this loop, and thus removed from worklist. (A likely
5788 : * case is that two FKs are matched to the same EC; there will be only
5789 : * one EC-derived clause in the initial list, so the first FK will
5790 : * consume it.) Applying both FKs' selectivity independently risks
5791 : * underestimating the join size; in particular, this would undo one
5792 : * of the main things that ECs were invented for, namely to avoid
5793 : * double-counting the selectivity of redundant equality conditions.
5794 : * Later we might think of a reasonable way to combine the estimates,
5795 : * but for now, just punt, since this is a fairly uncommon situation.
5796 : */
5797 1608 : if (removedlist == NIL ||
5798 1308 : list_length(removedlist) !=
5799 1308 : (fkinfo->nmatched_ec - fkinfo->nconst_ec + fkinfo->nmatched_ri))
5800 : {
5801 300 : worklist = list_concat(worklist, removedlist);
5802 300 : continue;
5803 : }
5804 :
5805 : /*
5806 : * Finally we get to the payoff: estimate selectivity using the
5807 : * knowledge that each referencing row will match exactly one row in
5808 : * the referenced table.
5809 : *
5810 : * XXX that's not true in the presence of nulls in the referencing
5811 : * column(s), so in principle we should derate the estimate for those.
5812 : * However (1) if there are any strict restriction clauses for the
5813 : * referencing column(s) elsewhere in the query, derating here would
5814 : * be double-counting the null fraction, and (2) it's not very clear
5815 : * how to combine null fractions for multiple referencing columns. So
5816 : * we do nothing for now about correcting for nulls.
5817 : *
5818 : * XXX another point here is that if either side of an FK constraint
5819 : * is an inheritance parent, we estimate as though the constraint
5820 : * covers all its children as well. This is not an unreasonable
5821 : * assumption for a referencing table, ie the user probably applied
5822 : * identical constraints to all child tables (though perhaps we ought
5823 : * to check that). But it's not possible to have done that for a
5824 : * referenced table. Fortunately, precisely because that doesn't
5825 : * work, it is uncommon in practice to have an FK referencing a parent
5826 : * table. So, at least for now, disregard inheritance here.
5827 : */
5828 1308 : if (jointype == JOIN_SEMI || jointype == JOIN_ANTI)
5829 860 : {
5830 : /*
5831 : * For JOIN_SEMI and JOIN_ANTI, we only get here when the FK's
5832 : * referenced table is exactly the inside of the join. The join
5833 : * selectivity is defined as the fraction of LHS rows that have
5834 : * matches. The FK implies that every LHS row has a match *in the
5835 : * referenced table*; but any restriction clauses on it will
5836 : * reduce the number of matches. Hence we take the join
5837 : * selectivity as equal to the selectivity of the table's
5838 : * restriction clauses, which is rows / tuples; but we must guard
5839 : * against tuples == 0.
5840 : */
5841 860 : RelOptInfo *ref_rel = find_base_rel(root, fkinfo->ref_relid);
5842 860 : double ref_tuples = Max(ref_rel->tuples, 1.0);
5843 :
5844 860 : fkselec *= ref_rel->rows / ref_tuples;
5845 : }
5846 : else
5847 : {
5848 : /*
5849 : * Otherwise, selectivity is exactly 1/referenced-table-size; but
5850 : * guard against tuples == 0. Note we should use the raw table
5851 : * tuple count, not any estimate of its filtered or joined size.
5852 : */
5853 448 : RelOptInfo *ref_rel = find_base_rel(root, fkinfo->ref_relid);
5854 448 : double ref_tuples = Max(ref_rel->tuples, 1.0);
5855 :
5856 448 : fkselec *= 1.0 / ref_tuples;
5857 : }
5858 :
5859 : /*
5860 : * If any of the FK columns participated in ec_has_const ECs, then
5861 : * equivclass.c will have generated "var = const" restrictions for
5862 : * each side of the join, thus reducing the sizes of both input
5863 : * relations. Taking the fkselec at face value would amount to
5864 : * double-counting the selectivity of the constant restriction for the
5865 : * referencing Var. Hence, look for the restriction clause(s) that
5866 : * were applied to the referencing Var(s), and divide out their
5867 : * selectivity to correct for this.
5868 : */
5869 1308 : if (fkinfo->nconst_ec > 0)
5870 : {
5871 24 : for (int i = 0; i < fkinfo->nkeys; i++)
5872 : {
5873 18 : EquivalenceClass *ec = fkinfo->eclass[i];
5874 :
5875 18 : if (ec && ec->ec_has_const)
5876 : {
5877 6 : EquivalenceMember *em = fkinfo->fk_eclass_member[i];
5878 6 : RestrictInfo *rinfo = find_derived_clause_for_ec_member(root,
5879 : ec,
5880 : em);
5881 :
5882 6 : if (rinfo)
5883 : {
5884 : Selectivity s0;
5885 :
5886 6 : s0 = clause_selectivity(root,
5887 : (Node *) rinfo,
5888 : 0,
5889 : jointype,
5890 : sjinfo);
5891 6 : if (s0 > 0)
5892 6 : fkselec /= s0;
5893 : }
5894 : }
5895 : }
5896 : }
5897 : }
5898 :
5899 225138 : *restrictlist = worklist;
5900 225138 : CLAMP_PROBABILITY(fkselec);
5901 225138 : return fkselec;
5902 : }
5903 :
5904 : /*
5905 : * set_subquery_size_estimates
5906 : * Set the size estimates for a base relation that is a subquery.
5907 : *
5908 : * The rel's targetlist and restrictinfo list must have been constructed
5909 : * already, and the Paths for the subquery must have been completed.
5910 : * We look at the subquery's PlannerInfo to extract data.
5911 : *
5912 : * We set the same fields as set_baserel_size_estimates.
5913 : */
5914 : void
5915 27788 : set_subquery_size_estimates(PlannerInfo *root, RelOptInfo *rel)
5916 : {
5917 27788 : PlannerInfo *subroot = rel->subroot;
5918 : RelOptInfo *sub_final_rel;
5919 : ListCell *lc;
5920 :
5921 : /* Should only be applied to base relations that are subqueries */
5922 : Assert(rel->relid > 0);
5923 : Assert(planner_rt_fetch(rel->relid, root)->rtekind == RTE_SUBQUERY);
5924 :
5925 : /*
5926 : * Copy raw number of output rows from subquery. All of its paths should
5927 : * have the same output rowcount, so just look at cheapest-total.
5928 : */
5929 27788 : sub_final_rel = fetch_upper_rel(subroot, UPPERREL_FINAL, NULL);
5930 27788 : rel->tuples = sub_final_rel->cheapest_total_path->rows;
5931 :
5932 : /*
5933 : * Compute per-output-column width estimates by examining the subquery's
5934 : * targetlist. For any output that is a plain Var, get the width estimate
5935 : * that was made while planning the subquery. Otherwise, we leave it to
5936 : * set_rel_width to fill in a datatype-based default estimate.
5937 : */
5938 113540 : foreach(lc, subroot->parse->targetList)
5939 : {
5940 85752 : TargetEntry *te = lfirst_node(TargetEntry, lc);
5941 85752 : Node *texpr = (Node *) te->expr;
5942 85752 : int32 item_width = 0;
5943 :
5944 : /* junk columns aren't visible to upper query */
5945 85752 : if (te->resjunk)
5946 1292 : continue;
5947 :
5948 : /*
5949 : * The subquery could be an expansion of a view that's had columns
5950 : * added to it since the current query was parsed, so that there are
5951 : * non-junk tlist columns in it that don't correspond to any column
5952 : * visible at our query level. Ignore such columns.
5953 : */
5954 84460 : if (te->resno < rel->min_attr || te->resno > rel->max_attr)
5955 0 : continue;
5956 :
5957 : /*
5958 : * XXX This currently doesn't work for subqueries containing set
5959 : * operations, because the Vars in their tlists are bogus references
5960 : * to the first leaf subquery, which wouldn't give the right answer
5961 : * even if we could still get to its PlannerInfo.
5962 : *
5963 : * Also, the subquery could be an appendrel for which all branches are
5964 : * known empty due to constraint exclusion, in which case
5965 : * set_append_rel_pathlist will have left the attr_widths set to zero.
5966 : *
5967 : * In either case, we just leave the width estimate zero until
5968 : * set_rel_width fixes it.
5969 : */
5970 84460 : if (IsA(texpr, Var) &&
5971 38446 : subroot->parse->setOperations == NULL)
5972 : {
5973 36668 : Var *var = (Var *) texpr;
5974 36668 : RelOptInfo *subrel = find_base_rel(subroot, var->varno);
5975 :
5976 36668 : item_width = subrel->attr_widths[var->varattno - subrel->min_attr];
5977 : }
5978 84460 : rel->attr_widths[te->resno - rel->min_attr] = item_width;
5979 : }
5980 :
5981 : /* Now estimate number of output rows, etc */
5982 27788 : set_baserel_size_estimates(root, rel);
5983 27788 : }
5984 :
5985 : /*
5986 : * set_function_size_estimates
5987 : * Set the size estimates for a base relation that is a function call.
5988 : *
5989 : * The rel's targetlist and restrictinfo list must have been constructed
5990 : * already.
5991 : *
5992 : * We set the same fields as set_baserel_size_estimates.
5993 : */
5994 : void
5995 51610 : set_function_size_estimates(PlannerInfo *root, RelOptInfo *rel)
5996 : {
5997 : RangeTblEntry *rte;
5998 : ListCell *lc;
5999 :
6000 : /* Should only be applied to base relations that are functions */
6001 : Assert(rel->relid > 0);
6002 51610 : rte = planner_rt_fetch(rel->relid, root);
6003 : Assert(rte->rtekind == RTE_FUNCTION);
6004 :
6005 : /*
6006 : * Estimate number of rows the functions will return. The rowcount of the
6007 : * node is that of the largest function result.
6008 : */
6009 51610 : rel->tuples = 0;
6010 103712 : foreach(lc, rte->functions)
6011 : {
6012 52102 : RangeTblFunction *rtfunc = (RangeTblFunction *) lfirst(lc);
6013 52102 : double ntup = expression_returns_set_rows(root, rtfunc->funcexpr);
6014 :
6015 52102 : if (ntup > rel->tuples)
6016 51634 : rel->tuples = ntup;
6017 : }
6018 :
6019 : /* Now estimate number of output rows, etc */
6020 51610 : set_baserel_size_estimates(root, rel);
6021 51610 : }
6022 :
6023 : /*
6024 : * set_function_size_estimates
6025 : * Set the size estimates for a base relation that is a function call.
6026 : *
6027 : * The rel's targetlist and restrictinfo list must have been constructed
6028 : * already.
6029 : *
6030 : * We set the same fields as set_tablefunc_size_estimates.
6031 : */
6032 : void
6033 626 : set_tablefunc_size_estimates(PlannerInfo *root, RelOptInfo *rel)
6034 : {
6035 : /* Should only be applied to base relations that are functions */
6036 : Assert(rel->relid > 0);
6037 : Assert(planner_rt_fetch(rel->relid, root)->rtekind == RTE_TABLEFUNC);
6038 :
6039 626 : rel->tuples = 100;
6040 :
6041 : /* Now estimate number of output rows, etc */
6042 626 : set_baserel_size_estimates(root, rel);
6043 626 : }
6044 :
6045 : /*
6046 : * set_values_size_estimates
6047 : * Set the size estimates for a base relation that is a values list.
6048 : *
6049 : * The rel's targetlist and restrictinfo list must have been constructed
6050 : * already.
6051 : *
6052 : * We set the same fields as set_baserel_size_estimates.
6053 : */
6054 : void
6055 8246 : set_values_size_estimates(PlannerInfo *root, RelOptInfo *rel)
6056 : {
6057 : RangeTblEntry *rte;
6058 :
6059 : /* Should only be applied to base relations that are values lists */
6060 : Assert(rel->relid > 0);
6061 8246 : rte = planner_rt_fetch(rel->relid, root);
6062 : Assert(rte->rtekind == RTE_VALUES);
6063 :
6064 : /*
6065 : * Estimate number of rows the values list will return. We know this
6066 : * precisely based on the list length (well, barring set-returning
6067 : * functions in list items, but that's a refinement not catered for
6068 : * anywhere else either).
6069 : */
6070 8246 : rel->tuples = list_length(rte->values_lists);
6071 :
6072 : /* Now estimate number of output rows, etc */
6073 8246 : set_baserel_size_estimates(root, rel);
6074 8246 : }
6075 :
6076 : /*
6077 : * set_cte_size_estimates
6078 : * Set the size estimates for a base relation that is a CTE reference.
6079 : *
6080 : * The rel's targetlist and restrictinfo list must have been constructed
6081 : * already, and we need an estimate of the number of rows returned by the CTE
6082 : * (if a regular CTE) or the non-recursive term (if a self-reference).
6083 : *
6084 : * We set the same fields as set_baserel_size_estimates.
6085 : */
6086 : void
6087 5094 : set_cte_size_estimates(PlannerInfo *root, RelOptInfo *rel, double cte_rows)
6088 : {
6089 : RangeTblEntry *rte;
6090 :
6091 : /* Should only be applied to base relations that are CTE references */
6092 : Assert(rel->relid > 0);
6093 5094 : rte = planner_rt_fetch(rel->relid, root);
6094 : Assert(rte->rtekind == RTE_CTE);
6095 :
6096 5094 : if (rte->self_reference)
6097 : {
6098 : /*
6099 : * In a self-reference, we assume the average worktable size is a
6100 : * multiple of the nonrecursive term's size. The best multiplier will
6101 : * vary depending on query "fan-out", so make its value adjustable.
6102 : */
6103 1006 : rel->tuples = clamp_row_est(recursive_worktable_factor * cte_rows);
6104 : }
6105 : else
6106 : {
6107 : /* Otherwise just believe the CTE's rowcount estimate */
6108 4088 : rel->tuples = cte_rows;
6109 : }
6110 :
6111 : /* Now estimate number of output rows, etc */
6112 5094 : set_baserel_size_estimates(root, rel);
6113 5094 : }
6114 :
6115 : /*
6116 : * set_namedtuplestore_size_estimates
6117 : * Set the size estimates for a base relation that is a tuplestore reference.
6118 : *
6119 : * The rel's targetlist and restrictinfo list must have been constructed
6120 : * already.
6121 : *
6122 : * We set the same fields as set_baserel_size_estimates.
6123 : */
6124 : void
6125 466 : set_namedtuplestore_size_estimates(PlannerInfo *root, RelOptInfo *rel)
6126 : {
6127 : RangeTblEntry *rte;
6128 :
6129 : /* Should only be applied to base relations that are tuplestore references */
6130 : Assert(rel->relid > 0);
6131 466 : rte = planner_rt_fetch(rel->relid, root);
6132 : Assert(rte->rtekind == RTE_NAMEDTUPLESTORE);
6133 :
6134 : /*
6135 : * Use the estimate provided by the code which is generating the named
6136 : * tuplestore. In some cases, the actual number might be available; in
6137 : * others the same plan will be re-used, so a "typical" value might be
6138 : * estimated and used.
6139 : */
6140 466 : rel->tuples = rte->enrtuples;
6141 466 : if (rel->tuples < 0)
6142 0 : rel->tuples = 1000;
6143 :
6144 : /* Now estimate number of output rows, etc */
6145 466 : set_baserel_size_estimates(root, rel);
6146 466 : }
6147 :
6148 : /*
6149 : * set_result_size_estimates
6150 : * Set the size estimates for an RTE_RESULT base relation
6151 : *
6152 : * The rel's targetlist and restrictinfo list must have been constructed
6153 : * already.
6154 : *
6155 : * We set the same fields as set_baserel_size_estimates.
6156 : */
6157 : void
6158 4220 : set_result_size_estimates(PlannerInfo *root, RelOptInfo *rel)
6159 : {
6160 : /* Should only be applied to RTE_RESULT base relations */
6161 : Assert(rel->relid > 0);
6162 : Assert(planner_rt_fetch(rel->relid, root)->rtekind == RTE_RESULT);
6163 :
6164 : /* RTE_RESULT always generates a single row, natively */
6165 4220 : rel->tuples = 1;
6166 :
6167 : /* Now estimate number of output rows, etc */
6168 4220 : set_baserel_size_estimates(root, rel);
6169 4220 : }
6170 :
6171 : /*
6172 : * set_foreign_size_estimates
6173 : * Set the size estimates for a base relation that is a foreign table.
6174 : *
6175 : * There is not a whole lot that we can do here; the foreign-data wrapper
6176 : * is responsible for producing useful estimates. We can do a decent job
6177 : * of estimating baserestrictcost, so we set that, and we also set up width
6178 : * using what will be purely datatype-driven estimates from the targetlist.
6179 : * There is no way to do anything sane with the rows value, so we just put
6180 : * a default estimate and hope that the wrapper can improve on it. The
6181 : * wrapper's GetForeignRelSize function will be called momentarily.
6182 : *
6183 : * The rel's targetlist and restrictinfo list must have been constructed
6184 : * already.
6185 : */
6186 : void
6187 2424 : set_foreign_size_estimates(PlannerInfo *root, RelOptInfo *rel)
6188 : {
6189 : /* Should only be applied to base relations */
6190 : Assert(rel->relid > 0);
6191 :
6192 2424 : rel->rows = 1000; /* entirely bogus default estimate */
6193 :
6194 2424 : cost_qual_eval(&rel->baserestrictcost, rel->baserestrictinfo, root);
6195 :
6196 2424 : set_rel_width(root, rel);
6197 2424 : }
6198 :
6199 :
6200 : /*
6201 : * set_rel_width
6202 : * Set the estimated output width of a base relation.
6203 : *
6204 : * The estimated output width is the sum of the per-attribute width estimates
6205 : * for the actually-referenced columns, plus any PHVs or other expressions
6206 : * that have to be calculated at this relation. This is the amount of data
6207 : * we'd need to pass upwards in case of a sort, hash, etc.
6208 : *
6209 : * This function also sets reltarget->cost, so it's a bit misnamed now.
6210 : *
6211 : * NB: this works best on plain relations because it prefers to look at
6212 : * real Vars. For subqueries, set_subquery_size_estimates will already have
6213 : * copied up whatever per-column estimates were made within the subquery,
6214 : * and for other types of rels there isn't much we can do anyway. We fall
6215 : * back on (fairly stupid) datatype-based width estimates if we can't get
6216 : * any better number.
6217 : *
6218 : * The per-attribute width estimates are cached for possible re-use while
6219 : * building join relations or post-scan/join pathtargets.
6220 : */
6221 : static void
6222 503846 : set_rel_width(PlannerInfo *root, RelOptInfo *rel)
6223 : {
6224 503846 : Oid reloid = planner_rt_fetch(rel->relid, root)->relid;
6225 503846 : int64 tuple_width = 0;
6226 503846 : bool have_wholerow_var = false;
6227 : ListCell *lc;
6228 :
6229 : /* Vars are assumed to have cost zero, but other exprs do not */
6230 503846 : rel->reltarget->cost.startup = 0;
6231 503846 : rel->reltarget->cost.per_tuple = 0;
6232 :
6233 1820816 : foreach(lc, rel->reltarget->exprs)
6234 : {
6235 1316970 : Node *node = (Node *) lfirst(lc);
6236 :
6237 : /*
6238 : * Ordinarily, a Var in a rel's targetlist must belong to that rel;
6239 : * but there are corner cases involving LATERAL references where that
6240 : * isn't so. If the Var has the wrong varno, fall through to the
6241 : * generic case (it doesn't seem worth the trouble to be any smarter).
6242 : */
6243 1316970 : if (IsA(node, Var) &&
6244 1292802 : ((Var *) node)->varno == rel->relid)
6245 342538 : {
6246 1292736 : Var *var = (Var *) node;
6247 : int ndx;
6248 : int32 item_width;
6249 :
6250 : Assert(var->varattno >= rel->min_attr);
6251 : Assert(var->varattno <= rel->max_attr);
6252 :
6253 1292736 : ndx = var->varattno - rel->min_attr;
6254 :
6255 : /*
6256 : * If it's a whole-row Var, we'll deal with it below after we have
6257 : * already cached as many attr widths as possible.
6258 : */
6259 1292736 : if (var->varattno == 0)
6260 : {
6261 2990 : have_wholerow_var = true;
6262 2990 : continue;
6263 : }
6264 :
6265 : /*
6266 : * The width may have been cached already (especially if it's a
6267 : * subquery), so don't duplicate effort.
6268 : */
6269 1289746 : if (rel->attr_widths[ndx] > 0)
6270 : {
6271 237042 : tuple_width += rel->attr_widths[ndx];
6272 237042 : continue;
6273 : }
6274 :
6275 : /* Try to get column width from statistics */
6276 1052704 : if (reloid != InvalidOid && var->varattno > 0)
6277 : {
6278 837462 : item_width = get_attavgwidth(reloid, var->varattno);
6279 837462 : if (item_width > 0)
6280 : {
6281 710166 : rel->attr_widths[ndx] = item_width;
6282 710166 : tuple_width += item_width;
6283 710166 : continue;
6284 : }
6285 : }
6286 :
6287 : /*
6288 : * Not a plain relation, or can't find statistics for it. Estimate
6289 : * using just the type info.
6290 : */
6291 342538 : item_width = get_typavgwidth(var->vartype, var->vartypmod);
6292 : Assert(item_width > 0);
6293 342538 : rel->attr_widths[ndx] = item_width;
6294 342538 : tuple_width += item_width;
6295 : }
6296 24234 : else if (IsA(node, PlaceHolderVar))
6297 : {
6298 : /*
6299 : * We will need to evaluate the PHV's contained expression while
6300 : * scanning this rel, so be sure to include it in reltarget->cost.
6301 : */
6302 1978 : PlaceHolderVar *phv = (PlaceHolderVar *) node;
6303 1978 : PlaceHolderInfo *phinfo = find_placeholder_info(root, phv);
6304 : QualCost cost;
6305 :
6306 1978 : tuple_width += phinfo->ph_width;
6307 1978 : cost_qual_eval_node(&cost, (Node *) phv->phexpr, root);
6308 1978 : rel->reltarget->cost.startup += cost.startup;
6309 1978 : rel->reltarget->cost.per_tuple += cost.per_tuple;
6310 : }
6311 : else
6312 : {
6313 : /*
6314 : * We could be looking at an expression pulled up from a subquery,
6315 : * or a ROW() representing a whole-row child Var, etc. Do what we
6316 : * can using the expression type information.
6317 : */
6318 : int32 item_width;
6319 : QualCost cost;
6320 :
6321 22256 : item_width = get_typavgwidth(exprType(node), exprTypmod(node));
6322 : Assert(item_width > 0);
6323 22256 : tuple_width += item_width;
6324 : /* Not entirely clear if we need to account for cost, but do so */
6325 22256 : cost_qual_eval_node(&cost, node, root);
6326 22256 : rel->reltarget->cost.startup += cost.startup;
6327 22256 : rel->reltarget->cost.per_tuple += cost.per_tuple;
6328 : }
6329 : }
6330 :
6331 : /*
6332 : * If we have a whole-row reference, estimate its width as the sum of
6333 : * per-column widths plus heap tuple header overhead.
6334 : */
6335 503846 : if (have_wholerow_var)
6336 : {
6337 2990 : int64 wholerow_width = MAXALIGN(SizeofHeapTupleHeader);
6338 :
6339 2990 : if (reloid != InvalidOid)
6340 : {
6341 : /* Real relation, so estimate true tuple width */
6342 2332 : wholerow_width += get_relation_data_width(reloid,
6343 2332 : rel->attr_widths - rel->min_attr);
6344 : }
6345 : else
6346 : {
6347 : /* Do what we can with info for a phony rel */
6348 : AttrNumber i;
6349 :
6350 1794 : for (i = 1; i <= rel->max_attr; i++)
6351 1136 : wholerow_width += rel->attr_widths[i - rel->min_attr];
6352 : }
6353 :
6354 2990 : rel->attr_widths[0 - rel->min_attr] = clamp_width_est(wholerow_width);
6355 :
6356 : /*
6357 : * Include the whole-row Var as part of the output tuple. Yes, that
6358 : * really is what happens at runtime.
6359 : */
6360 2990 : tuple_width += wholerow_width;
6361 : }
6362 :
6363 503846 : rel->reltarget->width = clamp_width_est(tuple_width);
6364 503846 : }
6365 :
6366 : /*
6367 : * set_pathtarget_cost_width
6368 : * Set the estimated eval cost and output width of a PathTarget tlist.
6369 : *
6370 : * As a notational convenience, returns the same PathTarget pointer passed in.
6371 : *
6372 : * Most, though not quite all, uses of this function occur after we've run
6373 : * set_rel_width() for base relations; so we can usually obtain cached width
6374 : * estimates for Vars. If we can't, fall back on datatype-based width
6375 : * estimates. Present early-planning uses of PathTargets don't need accurate
6376 : * widths badly enough to justify going to the catalogs for better data.
6377 : */
6378 : PathTarget *
6379 600952 : set_pathtarget_cost_width(PlannerInfo *root, PathTarget *target)
6380 : {
6381 600952 : int64 tuple_width = 0;
6382 : ListCell *lc;
6383 :
6384 : /* Vars are assumed to have cost zero, but other exprs do not */
6385 600952 : target->cost.startup = 0;
6386 600952 : target->cost.per_tuple = 0;
6387 :
6388 2094896 : foreach(lc, target->exprs)
6389 : {
6390 1493944 : Node *node = (Node *) lfirst(lc);
6391 :
6392 1493944 : tuple_width += get_expr_width(root, node);
6393 :
6394 : /* For non-Vars, account for evaluation cost */
6395 1493944 : if (!IsA(node, Var))
6396 : {
6397 : QualCost cost;
6398 :
6399 622716 : cost_qual_eval_node(&cost, node, root);
6400 622716 : target->cost.startup += cost.startup;
6401 622716 : target->cost.per_tuple += cost.per_tuple;
6402 : }
6403 : }
6404 :
6405 600952 : target->width = clamp_width_est(tuple_width);
6406 :
6407 600952 : return target;
6408 : }
6409 :
6410 : /*
6411 : * get_expr_width
6412 : * Estimate the width of the given expr attempting to use the width
6413 : * cached in a Var's owning RelOptInfo, else fallback on the type's
6414 : * average width when unable to or when the given Node is not a Var.
6415 : */
6416 : static int32
6417 1838074 : get_expr_width(PlannerInfo *root, const Node *expr)
6418 : {
6419 : int32 width;
6420 :
6421 1838074 : if (IsA(expr, Var))
6422 : {
6423 1202094 : const Var *var = (const Var *) expr;
6424 :
6425 : /* We should not see any upper-level Vars here */
6426 : Assert(var->varlevelsup == 0);
6427 :
6428 : /* Try to get data from RelOptInfo cache */
6429 1202094 : if (!IS_SPECIAL_VARNO(var->varno) &&
6430 1196388 : var->varno < root->simple_rel_array_size)
6431 : {
6432 1196388 : RelOptInfo *rel = root->simple_rel_array[var->varno];
6433 :
6434 1196388 : if (rel != NULL &&
6435 1167506 : var->varattno >= rel->min_attr &&
6436 1167506 : var->varattno <= rel->max_attr)
6437 : {
6438 1167506 : int ndx = var->varattno - rel->min_attr;
6439 :
6440 1167506 : if (rel->attr_widths[ndx] > 0)
6441 1134070 : return rel->attr_widths[ndx];
6442 : }
6443 : }
6444 :
6445 : /*
6446 : * No cached data available, so estimate using just the type info.
6447 : */
6448 68024 : width = get_typavgwidth(var->vartype, var->vartypmod);
6449 : Assert(width > 0);
6450 :
6451 68024 : return width;
6452 : }
6453 :
6454 635980 : width = get_typavgwidth(exprType(expr), exprTypmod(expr));
6455 : Assert(width > 0);
6456 635980 : return width;
6457 : }
6458 :
6459 : /*
6460 : * relation_byte_size
6461 : * Estimate the storage space in bytes for a given number of tuples
6462 : * of a given width (size in bytes).
6463 : */
6464 : static double
6465 4101996 : relation_byte_size(double tuples, int width)
6466 : {
6467 4101996 : return tuples * (MAXALIGN(width) + MAXALIGN(SizeofHeapTupleHeader));
6468 : }
6469 :
6470 : /*
6471 : * page_size
6472 : * Returns an estimate of the number of pages covered by a given
6473 : * number of tuples of a given width (size in bytes).
6474 : */
6475 : static double
6476 10752 : page_size(double tuples, int width)
6477 : {
6478 10752 : return ceil(relation_byte_size(tuples, width) / BLCKSZ);
6479 : }
6480 :
6481 : /*
6482 : * Estimate the fraction of the work that each worker will do given the
6483 : * number of workers budgeted for the path.
6484 : */
6485 : static double
6486 179082 : get_parallel_divisor(Path *path)
6487 : {
6488 179082 : double parallel_divisor = path->parallel_workers;
6489 :
6490 : /*
6491 : * Early experience with parallel query suggests that when there is only
6492 : * one worker, the leader often makes a very substantial contribution to
6493 : * executing the parallel portion of the plan, but as more workers are
6494 : * added, it does less and less, because it's busy reading tuples from the
6495 : * workers and doing whatever non-parallel post-processing is needed. By
6496 : * the time we reach 4 workers, the leader no longer makes a meaningful
6497 : * contribution. Thus, for now, estimate that the leader spends 30% of
6498 : * its time servicing each worker, and the remainder executing the
6499 : * parallel plan.
6500 : */
6501 179082 : if (parallel_leader_participation)
6502 : {
6503 : double leader_contribution;
6504 :
6505 177780 : leader_contribution = 1.0 - (0.3 * path->parallel_workers);
6506 177780 : if (leader_contribution > 0)
6507 175464 : parallel_divisor += leader_contribution;
6508 : }
6509 :
6510 179082 : return parallel_divisor;
6511 : }
6512 :
6513 : /*
6514 : * compute_bitmap_pages
6515 : * Estimate number of pages fetched from heap in a bitmap heap scan.
6516 : *
6517 : * 'baserel' is the relation to be scanned
6518 : * 'bitmapqual' is a tree of IndexPaths, BitmapAndPaths, and BitmapOrPaths
6519 : * 'loop_count' is the number of repetitions of the indexscan to factor into
6520 : * estimates of caching behavior
6521 : *
6522 : * If cost_p isn't NULL, the indexTotalCost estimate is returned in *cost_p.
6523 : * If tuples_p isn't NULL, the tuples_fetched estimate is returned in *tuples_p.
6524 : */
6525 : double
6526 680148 : compute_bitmap_pages(PlannerInfo *root, RelOptInfo *baserel,
6527 : Path *bitmapqual, double loop_count,
6528 : Cost *cost_p, double *tuples_p)
6529 : {
6530 : Cost indexTotalCost;
6531 : Selectivity indexSelectivity;
6532 : double T;
6533 : double pages_fetched;
6534 : double tuples_fetched;
6535 : double heap_pages;
6536 : double maxentries;
6537 :
6538 : /*
6539 : * Fetch total cost of obtaining the bitmap, as well as its total
6540 : * selectivity.
6541 : */
6542 680148 : cost_bitmap_tree_node(bitmapqual, &indexTotalCost, &indexSelectivity);
6543 :
6544 : /*
6545 : * Estimate number of main-table pages fetched.
6546 : */
6547 680148 : tuples_fetched = clamp_row_est(indexSelectivity * baserel->tuples);
6548 :
6549 680148 : T = (baserel->pages > 1) ? (double) baserel->pages : 1.0;
6550 :
6551 : /*
6552 : * For a single scan, the number of heap pages that need to be fetched is
6553 : * the same as the Mackert and Lohman formula for the case T <= b (ie, no
6554 : * re-reads needed).
6555 : */
6556 680148 : pages_fetched = (2.0 * T * tuples_fetched) / (2.0 * T + tuples_fetched);
6557 :
6558 : /*
6559 : * Calculate the number of pages fetched from the heap. Then based on
6560 : * current work_mem estimate get the estimated maxentries in the bitmap.
6561 : * (Note that we always do this calculation based on the number of pages
6562 : * that would be fetched in a single iteration, even if loop_count > 1.
6563 : * That's correct, because only that number of entries will be stored in
6564 : * the bitmap at one time.)
6565 : */
6566 680148 : heap_pages = Min(pages_fetched, baserel->pages);
6567 680148 : maxentries = tbm_calculate_entries(work_mem * (Size) 1024);
6568 :
6569 680148 : if (loop_count > 1)
6570 : {
6571 : /*
6572 : * For repeated bitmap scans, scale up the number of tuples fetched in
6573 : * the Mackert and Lohman formula by the number of scans, so that we
6574 : * estimate the number of pages fetched by all the scans. Then
6575 : * pro-rate for one scan.
6576 : */
6577 139732 : pages_fetched = index_pages_fetched(tuples_fetched * loop_count,
6578 : baserel->pages,
6579 : get_indexpath_pages(bitmapqual),
6580 : root);
6581 139732 : pages_fetched /= loop_count;
6582 : }
6583 :
6584 680148 : if (pages_fetched >= T)
6585 69016 : pages_fetched = T;
6586 : else
6587 611132 : pages_fetched = ceil(pages_fetched);
6588 :
6589 680148 : if (maxentries < heap_pages)
6590 : {
6591 : double exact_pages;
6592 : double lossy_pages;
6593 :
6594 : /*
6595 : * Crude approximation of the number of lossy pages. Because of the
6596 : * way tbm_lossify() is coded, the number of lossy pages increases
6597 : * very sharply as soon as we run short of memory; this formula has
6598 : * that property and seems to perform adequately in testing, but it's
6599 : * possible we could do better somehow.
6600 : */
6601 18 : lossy_pages = Max(0, heap_pages - maxentries / 2);
6602 18 : exact_pages = heap_pages - lossy_pages;
6603 :
6604 : /*
6605 : * If there are lossy pages then recompute the number of tuples
6606 : * processed by the bitmap heap node. We assume here that the chance
6607 : * of a given tuple coming from an exact page is the same as the
6608 : * chance that a given page is exact. This might not be true, but
6609 : * it's not clear how we can do any better.
6610 : */
6611 18 : if (lossy_pages > 0)
6612 : tuples_fetched =
6613 18 : clamp_row_est(indexSelectivity *
6614 18 : (exact_pages / heap_pages) * baserel->tuples +
6615 18 : (lossy_pages / heap_pages) * baserel->tuples);
6616 : }
6617 :
6618 680148 : if (cost_p)
6619 535218 : *cost_p = indexTotalCost;
6620 680148 : if (tuples_p)
6621 535218 : *tuples_p = tuples_fetched;
6622 :
6623 680148 : return pages_fetched;
6624 : }
6625 :
6626 : /*
6627 : * compute_gather_rows
6628 : * Estimate number of rows for gather (merge) nodes.
6629 : *
6630 : * In a parallel plan, each worker's row estimate is determined by dividing the
6631 : * total number of rows by parallel_divisor, which accounts for the leader's
6632 : * contribution in addition to the number of workers. Accordingly, when
6633 : * estimating the number of rows for gather (merge) nodes, we multiply the rows
6634 : * per worker by the same parallel_divisor to undo the division.
6635 : */
6636 : double
6637 29318 : compute_gather_rows(Path *path)
6638 : {
6639 : Assert(path->parallel_workers > 0);
6640 :
6641 29318 : return clamp_row_est(path->rows * get_parallel_divisor(path));
6642 : }
|
