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 8942002 : 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 8942002 : if (nrows > MAXIMUM_ROWCOUNT || isnan(nrows))
222 0 : nrows = MAXIMUM_ROWCOUNT;
223 8942002 : else if (nrows <= 1.0)
224 3251690 : nrows = 1.0;
225 : else
226 5690312 : nrows = rint(nrows);
227 :
228 8942002 : 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 1902542 : 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 1902542 : 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 1902542 : 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 45676 : 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 45676 : if (isnan(x))
272 0 : return LONG_MAX;
273 45676 : 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 45120 : 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 426398 : cost_seqscan(Path *path, PlannerInfo *root,
296 : RelOptInfo *baserel, ParamPathInfo *param_info)
297 : {
298 426398 : 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 426398 : if (param_info)
311 840 : path->rows = param_info->ppi_rows;
312 : else
313 425558 : path->rows = baserel->rows;
314 :
315 : /* fetch estimated page cost for tablespace containing table */
316 426398 : get_tablespace_page_costs(baserel->reltablespace,
317 : NULL,
318 : &spc_seq_page_cost);
319 :
320 : /*
321 : * disk costs
322 : */
323 426398 : disk_run_cost = spc_seq_page_cost * baserel->pages;
324 :
325 : /* CPU costs */
326 426398 : get_restriction_qual_cost(root, baserel, param_info, &qpqual_cost);
327 :
328 426398 : startup_cost += qpqual_cost.startup;
329 426398 : cpu_per_tuple = cpu_tuple_cost + qpqual_cost.per_tuple;
330 426398 : cpu_run_cost = cpu_per_tuple * baserel->tuples;
331 : /* tlist eval costs are paid per output row, not per tuple scanned */
332 426398 : startup_cost += path->pathtarget->cost.startup;
333 426398 : cpu_run_cost += path->pathtarget->cost.per_tuple * path->rows;
334 :
335 : /* Adjust costing for parallelism, if used. */
336 426398 : if (path->parallel_workers > 0)
337 : {
338 26118 : double parallel_divisor = get_parallel_divisor(path);
339 :
340 : /* The CPU cost is divided among all the workers. */
341 26118 : 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 26118 : path->rows = clamp_row_est(path->rows / parallel_divisor);
355 : }
356 :
357 426398 : path->disabled_nodes = enable_seqscan ? 0 : 1;
358 426398 : path->startup_cost = startup_cost;
359 426398 : path->total_cost = startup_cost + cpu_run_cost + disk_run_cost;
360 426398 : }
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 19082 : cost_gather(GatherPath *path, PlannerInfo *root,
447 : RelOptInfo *rel, ParamPathInfo *param_info,
448 : double *rows)
449 : {
450 19082 : Cost startup_cost = 0;
451 19082 : Cost run_cost = 0;
452 :
453 : /* Mark the path with the correct row estimate */
454 19082 : if (rows)
455 1752 : path->path.rows = *rows;
456 17330 : else if (param_info)
457 0 : path->path.rows = param_info->ppi_rows;
458 : else
459 17330 : path->path.rows = rel->rows;
460 :
461 19082 : startup_cost = path->subpath->startup_cost;
462 :
463 19082 : run_cost = path->subpath->total_cost - path->subpath->startup_cost;
464 :
465 : /* Parallel setup and communication cost. */
466 19082 : startup_cost += parallel_setup_cost;
467 19082 : run_cost += parallel_tuple_cost * path->path.rows;
468 :
469 19082 : path->path.disabled_nodes = path->subpath->disabled_nodes;
470 19082 : path->path.startup_cost = startup_cost;
471 19082 : path->path.total_cost = (startup_cost + run_cost);
472 19082 : }
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 10134 : 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 10134 : Cost startup_cost = 0;
492 10134 : 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 10134 : if (rows)
499 4612 : path->path.rows = *rows;
500 5522 : else if (param_info)
501 0 : path->path.rows = param_info->ppi_rows;
502 : else
503 5522 : 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 10134 : N = (double) path->num_workers + 1;
512 10134 : logN = LOG2(N);
513 :
514 : /* Assumed cost per tuple comparison */
515 10134 : comparison_cost = 2.0 * cpu_operator_cost;
516 :
517 : /* Heap creation cost */
518 10134 : startup_cost += comparison_cost * N * logN;
519 :
520 : /* Per-tuple heap maintenance cost */
521 10134 : run_cost += path->path.rows * comparison_cost * logN;
522 :
523 : /* small cost for heap management, like cost_merge_append */
524 10134 : 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 10134 : startup_cost += parallel_setup_cost;
533 10134 : run_cost += parallel_tuple_cost * path->path.rows * 1.05;
534 :
535 10134 : path->path.disabled_nodes = input_disabled_nodes
536 10134 : + (enable_gathermerge ? 0 : 1);
537 10134 : path->path.startup_cost = startup_cost + input_startup_cost;
538 10134 : path->path.total_cost = (startup_cost + run_cost + input_total_cost);
539 10134 : }
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 784292 : cost_index(IndexPath *path, PlannerInfo *root, double loop_count,
561 : bool partial_path)
562 : {
563 784292 : IndexOptInfo *index = path->indexinfo;
564 784292 : RelOptInfo *baserel = index->rel;
565 784292 : bool indexonly = (path->path.pathtype == T_IndexOnlyScan);
566 : amcostestimate_function amcostestimate;
567 : List *qpquals;
568 784292 : Cost startup_cost = 0;
569 784292 : Cost run_cost = 0;
570 784292 : 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 784292 : if (path->path.param_info)
601 : {
602 144868 : path->path.rows = path->path.param_info->ppi_rows;
603 : /* qpquals come from the rel's restriction clauses and ppi_clauses */
604 144868 : qpquals = list_concat(extract_nonindex_conditions(path->indexinfo->indrestrictinfo,
605 : path->indexclauses),
606 144868 : extract_nonindex_conditions(path->path.param_info->ppi_clauses,
607 : path->indexclauses));
608 : }
609 : else
610 : {
611 639424 : path->path.rows = baserel->rows;
612 : /* qpquals come from just the rel's restriction clauses */
613 639424 : 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 784292 : 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 784292 : amcostestimate = (amcostestimate_function) index->amcostestimate;
628 784292 : 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 784292 : path->indextotalcost = indexTotalCost;
639 784292 : path->indexselectivity = indexSelectivity;
640 :
641 : /* all costs for touching index itself included here */
642 784292 : startup_cost += indexStartupCost;
643 784292 : run_cost += indexTotalCost - indexStartupCost;
644 :
645 : /* estimate number of main-table tuples fetched */
646 784292 : tuples_fetched = clamp_row_est(indexSelectivity * baserel->tuples);
647 :
648 : /* fetch estimated page costs for tablespace containing table */
649 784292 : 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 784292 : 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 82932 : pages_fetched = index_pages_fetched(tuples_fetched * loop_count,
691 : baserel->pages,
692 82932 : (double) index->pages,
693 : root);
694 :
695 82932 : if (indexonly)
696 9036 : pages_fetched = ceil(pages_fetched * (1.0 - baserel->allvisfrac));
697 :
698 82932 : rand_heap_pages = pages_fetched;
699 :
700 82932 : 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 82932 : pages_fetched = ceil(indexSelectivity * (double) baserel->pages);
713 :
714 82932 : pages_fetched = index_pages_fetched(pages_fetched * loop_count,
715 : baserel->pages,
716 82932 : (double) index->pages,
717 : root);
718 :
719 82932 : if (indexonly)
720 9036 : pages_fetched = ceil(pages_fetched * (1.0 - baserel->allvisfrac));
721 :
722 82932 : 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 701360 : pages_fetched = index_pages_fetched(tuples_fetched,
731 : baserel->pages,
732 701360 : (double) index->pages,
733 : root);
734 :
735 701360 : if (indexonly)
736 64374 : pages_fetched = ceil(pages_fetched * (1.0 - baserel->allvisfrac));
737 :
738 701360 : rand_heap_pages = pages_fetched;
739 :
740 : /* max_IO_cost is for the perfectly uncorrelated case (csquared=0) */
741 701360 : max_IO_cost = pages_fetched * spc_random_page_cost;
742 :
743 : /* min_IO_cost is for the perfectly correlated case (csquared=1) */
744 701360 : pages_fetched = ceil(indexSelectivity * (double) baserel->pages);
745 :
746 701360 : if (indexonly)
747 64374 : pages_fetched = ceil(pages_fetched * (1.0 - baserel->allvisfrac));
748 :
749 701360 : if (pages_fetched > 0)
750 : {
751 631462 : min_IO_cost = spc_random_page_cost;
752 631462 : if (pages_fetched > 1)
753 187834 : min_IO_cost += (pages_fetched - 1) * spc_seq_page_cost;
754 : }
755 : else
756 69898 : min_IO_cost = 0;
757 : }
758 :
759 784292 : 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 272496 : if (indexonly)
767 23214 : 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 272496 : 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 272496 : if (path->path.parallel_workers <= 0)
786 262464 : return;
787 :
788 10032 : 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 521828 : csquared = indexCorrelation * indexCorrelation;
796 :
797 521828 : 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 521828 : cost_qual_eval(&qpqual_cost, qpquals, root);
806 :
807 521828 : startup_cost += qpqual_cost.startup;
808 521828 : cpu_per_tuple = cpu_tuple_cost + qpqual_cost.per_tuple;
809 :
810 521828 : cpu_run_cost += cpu_per_tuple * tuples_fetched;
811 :
812 : /* tlist eval costs are paid per output row, not per tuple scanned */
813 521828 : startup_cost += path->path.pathtarget->cost.startup;
814 521828 : cpu_run_cost += path->path.pathtarget->cost.per_tuple * path->path.rows;
815 :
816 : /* Adjust costing for parallelism, if used. */
817 521828 : if (path->path.parallel_workers > 0)
818 : {
819 10032 : double parallel_divisor = get_parallel_divisor(&path->path);
820 :
821 10032 : path->path.rows = clamp_row_est(path->path.rows / parallel_divisor);
822 :
823 : /* The CPU cost is divided among all the workers. */
824 10032 : cpu_run_cost /= parallel_divisor;
825 : }
826 :
827 521828 : run_cost += cpu_run_cost;
828 :
829 521828 : path->path.startup_cost = startup_cost;
830 521828 : 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 929160 : extract_nonindex_conditions(List *qual_clauses, List *indexclauses)
851 : {
852 929160 : List *result = NIL;
853 : ListCell *lc;
854 :
855 1951374 : foreach(lc, qual_clauses)
856 : {
857 1022214 : RestrictInfo *rinfo = lfirst_node(RestrictInfo, lc);
858 :
859 1022214 : if (rinfo->pseudoconstant)
860 9706 : continue; /* we may drop pseudoconstants here */
861 1012508 : if (is_redundant_with_indexclauses(rinfo, indexclauses))
862 592024 : continue; /* dup or derived from same EquivalenceClass */
863 : /* ... skip the predicate proof attempt createplan.c will try ... */
864 420484 : result = lappend(result, rinfo);
865 : }
866 929160 : 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 1094750 : 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 1094750 : T = (pages > 1) ? (double) pages : 1.0;
918 :
919 : /* Compute number of pages assumed to be competing for cache space */
920 1094750 : total_pages = root->total_table_pages + index_pages;
921 1094750 : total_pages = Max(total_pages, 1.0);
922 : Assert(T <= total_pages);
923 :
924 : /* b is pro-rated share of effective_cache_size */
925 1094750 : b = (double) effective_cache_size * T / total_pages;
926 :
927 : /* force it positive and integral */
928 1094750 : if (b <= 1.0)
929 0 : b = 1.0;
930 : else
931 1094750 : b = ceil(b);
932 :
933 : /* This part is the Mackert and Lohman formula */
934 1094750 : if (T <= b)
935 : {
936 1094750 : pages_fetched =
937 1094750 : (2.0 * T * tuples_fetched) / (2.0 * T + tuples_fetched);
938 1094750 : if (pages_fetched >= T)
939 634444 : pages_fetched = T;
940 : else
941 460306 : 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 1094750 : 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 186226 : get_indexpath_pages(Path *bitmapqual)
974 : {
975 186226 : double result = 0;
976 : ListCell *l;
977 :
978 186226 : if (IsA(bitmapqual, BitmapAndPath))
979 : {
980 23300 : BitmapAndPath *apath = (BitmapAndPath *) bitmapqual;
981 :
982 69900 : foreach(l, apath->bitmapquals)
983 : {
984 46600 : result += get_indexpath_pages((Path *) lfirst(l));
985 : }
986 : }
987 162926 : else if (IsA(bitmapqual, BitmapOrPath))
988 : {
989 74 : BitmapOrPath *opath = (BitmapOrPath *) bitmapqual;
990 :
991 234 : foreach(l, opath->bitmapquals)
992 : {
993 160 : result += get_indexpath_pages((Path *) lfirst(l));
994 : }
995 : }
996 162852 : else if (IsA(bitmapqual, IndexPath))
997 : {
998 162852 : IndexPath *ipath = (IndexPath *) bitmapqual;
999 :
1000 162852 : result = (double) ipath->indexinfo->pages;
1001 : }
1002 : else
1003 0 : elog(ERROR, "unrecognized node type: %d", nodeTag(bitmapqual));
1004 :
1005 186226 : 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 537172 : cost_bitmap_heap_scan(Path *path, PlannerInfo *root, RelOptInfo *baserel,
1024 : ParamPathInfo *param_info,
1025 : Path *bitmapqual, double loop_count)
1026 : {
1027 537172 : Cost startup_cost = 0;
1028 537172 : 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 537172 : if (param_info)
1047 224424 : path->rows = param_info->ppi_rows;
1048 : else
1049 312748 : path->rows = baserel->rows;
1050 :
1051 537172 : pages_fetched = compute_bitmap_pages(root, baserel, bitmapqual,
1052 : loop_count, &indexTotalCost,
1053 : &tuples_fetched);
1054 :
1055 537172 : startup_cost += indexTotalCost;
1056 537172 : T = (baserel->pages > 1) ? (double) baserel->pages : 1.0;
1057 :
1058 : /* Fetch estimated page costs for tablespace containing table. */
1059 537172 : 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 537172 : if (pages_fetched >= 2.0)
1071 108850 : cost_per_page = spc_random_page_cost -
1072 108850 : (spc_random_page_cost - spc_seq_page_cost)
1073 108850 : * sqrt(pages_fetched / T);
1074 : else
1075 428322 : cost_per_page = spc_random_page_cost;
1076 :
1077 537172 : 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 537172 : get_restriction_qual_cost(root, baserel, param_info, &qpqual_cost);
1089 :
1090 537172 : startup_cost += qpqual_cost.startup;
1091 537172 : cpu_per_tuple = cpu_tuple_cost + qpqual_cost.per_tuple;
1092 537172 : cpu_run_cost = cpu_per_tuple * tuples_fetched;
1093 :
1094 : /* Adjust costing for parallelism, if used. */
1095 537172 : if (path->parallel_workers > 0)
1096 : {
1097 4182 : double parallel_divisor = get_parallel_divisor(path);
1098 :
1099 : /* The CPU cost is divided among all the workers. */
1100 4182 : cpu_run_cost /= parallel_divisor;
1101 :
1102 4182 : path->rows = clamp_row_est(path->rows / parallel_divisor);
1103 : }
1104 :
1105 :
1106 537172 : run_cost += cpu_run_cost;
1107 :
1108 : /* tlist eval costs are paid per output row, not per tuple scanned */
1109 537172 : startup_cost += path->pathtarget->cost.startup;
1110 537172 : run_cost += path->pathtarget->cost.per_tuple * path->rows;
1111 :
1112 537172 : path->disabled_nodes = enable_bitmapscan ? 0 : 1;
1113 537172 : path->startup_cost = startup_cost;
1114 537172 : path->total_cost = startup_cost + run_cost;
1115 537172 : }
1116 :
1117 : /*
1118 : * cost_bitmap_tree_node
1119 : * Extract cost and selectivity from a bitmap tree node (index/and/or)
1120 : */
1121 : void
1122 1012496 : cost_bitmap_tree_node(Path *path, Cost *cost, Selectivity *selec)
1123 : {
1124 1012496 : if (IsA(path, IndexPath))
1125 : {
1126 957236 : *cost = ((IndexPath *) path)->indextotalcost;
1127 957236 : *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 957236 : *cost += 0.1 * cpu_operator_cost * path->rows;
1136 : }
1137 55260 : else if (IsA(path, BitmapAndPath))
1138 : {
1139 52024 : *cost = path->total_cost;
1140 52024 : *selec = ((BitmapAndPath *) path)->bitmapselectivity;
1141 : }
1142 3236 : else if (IsA(path, BitmapOrPath))
1143 : {
1144 3236 : *cost = path->total_cost;
1145 3236 : *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 1012496 : }
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 51834 : 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 51834 : totalCost = 0.0;
1181 51834 : selec = 1.0;
1182 155502 : foreach(l, path->bitmapquals)
1183 : {
1184 103668 : Path *subpath = (Path *) lfirst(l);
1185 : Cost subCost;
1186 : Selectivity subselec;
1187 :
1188 103668 : cost_bitmap_tree_node(subpath, &subCost, &subselec);
1189 :
1190 103668 : selec *= subselec;
1191 :
1192 103668 : totalCost += subCost;
1193 103668 : if (l != list_head(path->bitmapquals))
1194 51834 : totalCost += 100.0 * cpu_operator_cost;
1195 : }
1196 51834 : path->bitmapselectivity = selec;
1197 51834 : path->path.rows = 0; /* per above, not used */
1198 51834 : path->path.disabled_nodes = 0;
1199 51834 : path->path.startup_cost = totalCost;
1200 51834 : path->path.total_cost = totalCost;
1201 51834 : }
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 48856 : 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 48856 : if (param_info)
1478 552 : qpquals = list_concat_copy(param_info->ppi_clauses,
1479 552 : baserel->baserestrictinfo);
1480 : else
1481 48304 : qpquals = baserel->baserestrictinfo;
1482 :
1483 48856 : path->path.rows = clamp_row_est(path->subpath->rows *
1484 48856 : 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 48856 : path->path.disabled_nodes = path->subpath->disabled_nodes;
1497 48856 : path->path.startup_cost = path->subpath->startup_cost;
1498 48856 : 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 48856 : if (qpquals == NIL && trivial_pathtarget)
1514 24874 : return;
1515 :
1516 23982 : get_restriction_qual_cost(root, baserel, param_info, &qpqual_cost);
1517 :
1518 23982 : startup_cost = qpqual_cost.startup;
1519 23982 : cpu_per_tuple = cpu_tuple_cost + qpqual_cost.per_tuple;
1520 23982 : run_cost = cpu_per_tuple * path->subpath->rows;
1521 :
1522 : /* tlist eval costs are paid per output row, not per tuple scanned */
1523 23982 : startup_cost += path->path.pathtarget->cost.startup;
1524 23982 : run_cost += path->path.pathtarget->cost.per_tuple * path->path.rows;
1525 :
1526 23982 : path->path.startup_cost += startup_cost;
1527 23982 : 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 55114 : cost_functionscan(Path *path, PlannerInfo *root,
1539 : RelOptInfo *baserel, ParamPathInfo *param_info)
1540 : {
1541 55114 : Cost startup_cost = 0;
1542 55114 : 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 55114 : rte = planner_rt_fetch(baserel->relid, root);
1551 : Assert(rte->rtekind == RTE_FUNCTION);
1552 :
1553 : /* Mark the path with the correct row estimate */
1554 55114 : if (param_info)
1555 8342 : path->rows = param_info->ppi_rows;
1556 : else
1557 46772 : 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 55114 : cost_qual_eval_node(&exprcost, (Node *) rte->functions, root);
1573 :
1574 55114 : startup_cost += exprcost.startup + exprcost.per_tuple;
1575 :
1576 : /* Add scanning CPU costs */
1577 55114 : get_restriction_qual_cost(root, baserel, param_info, &qpqual_cost);
1578 :
1579 55114 : startup_cost += qpqual_cost.startup;
1580 55114 : cpu_per_tuple = cpu_tuple_cost + qpqual_cost.per_tuple;
1581 55114 : run_cost += cpu_per_tuple * baserel->tuples;
1582 :
1583 : /* tlist eval costs are paid per output row, not per tuple scanned */
1584 55114 : startup_cost += path->pathtarget->cost.startup;
1585 55114 : run_cost += path->pathtarget->cost.per_tuple * path->rows;
1586 :
1587 55114 : path->disabled_nodes = 0;
1588 55114 : path->startup_cost = startup_cost;
1589 55114 : path->total_cost = startup_cost + run_cost;
1590 55114 : }
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 8232 : cost_valuesscan(Path *path, PlannerInfo *root,
1658 : RelOptInfo *baserel, ParamPathInfo *param_info)
1659 : {
1660 8232 : Cost startup_cost = 0;
1661 8232 : 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 8232 : if (param_info)
1671 66 : path->rows = param_info->ppi_rows;
1672 : else
1673 8166 : 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 8232 : cpu_per_tuple = cpu_operator_cost;
1680 :
1681 : /* Add scanning CPU costs */
1682 8232 : get_restriction_qual_cost(root, baserel, param_info, &qpqual_cost);
1683 :
1684 8232 : startup_cost += qpqual_cost.startup;
1685 8232 : cpu_per_tuple += cpu_tuple_cost + qpqual_cost.per_tuple;
1686 8232 : run_cost += cpu_per_tuple * baserel->tuples;
1687 :
1688 : /* tlist eval costs are paid per output row, not per tuple scanned */
1689 8232 : startup_cost += path->pathtarget->cost.startup;
1690 8232 : run_cost += path->pathtarget->cost.per_tuple * path->rows;
1691 :
1692 8232 : path->disabled_nodes = 0;
1693 8232 : path->startup_cost = startup_cost;
1694 8232 : path->total_cost = startup_cost + run_cost;
1695 8232 : }
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 5100 : cost_ctescan(Path *path, PlannerInfo *root,
1709 : RelOptInfo *baserel, ParamPathInfo *param_info)
1710 : {
1711 5100 : Cost startup_cost = 0;
1712 5100 : 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 5100 : if (param_info)
1722 0 : path->rows = param_info->ppi_rows;
1723 : else
1724 5100 : path->rows = baserel->rows;
1725 :
1726 : /* Charge one CPU tuple cost per row for tuplestore manipulation */
1727 5100 : cpu_per_tuple = cpu_tuple_cost;
1728 :
1729 : /* Add scanning CPU costs */
1730 5100 : get_restriction_qual_cost(root, baserel, param_info, &qpqual_cost);
1731 :
1732 5100 : startup_cost += qpqual_cost.startup;
1733 5100 : cpu_per_tuple += cpu_tuple_cost + qpqual_cost.per_tuple;
1734 5100 : run_cost += cpu_per_tuple * baserel->tuples;
1735 :
1736 : /* tlist eval costs are paid per output row, not per tuple scanned */
1737 5100 : startup_cost += path->pathtarget->cost.startup;
1738 5100 : run_cost += path->pathtarget->cost.per_tuple * path->rows;
1739 :
1740 5100 : path->disabled_nodes = 0;
1741 5100 : path->startup_cost = startup_cost;
1742 5100 : path->total_cost = startup_cost + run_cost;
1743 5100 : }
1744 :
1745 : /*
1746 : * cost_namedtuplestorescan
1747 : * Determines and returns the cost of scanning a named tuplestore.
1748 : */
1749 : void
1750 462 : cost_namedtuplestorescan(Path *path, PlannerInfo *root,
1751 : RelOptInfo *baserel, ParamPathInfo *param_info)
1752 : {
1753 462 : Cost startup_cost = 0;
1754 462 : 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 462 : if (param_info)
1764 0 : path->rows = param_info->ppi_rows;
1765 : else
1766 462 : path->rows = baserel->rows;
1767 :
1768 : /* Charge one CPU tuple cost per row for tuplestore manipulation */
1769 462 : cpu_per_tuple = cpu_tuple_cost;
1770 :
1771 : /* Add scanning CPU costs */
1772 462 : get_restriction_qual_cost(root, baserel, param_info, &qpqual_cost);
1773 :
1774 462 : startup_cost += qpqual_cost.startup;
1775 462 : cpu_per_tuple += cpu_tuple_cost + qpqual_cost.per_tuple;
1776 462 : run_cost += cpu_per_tuple * baserel->tuples;
1777 :
1778 462 : path->disabled_nodes = 0;
1779 462 : path->startup_cost = startup_cost;
1780 462 : path->total_cost = startup_cost + run_cost;
1781 462 : }
1782 :
1783 : /*
1784 : * cost_resultscan
1785 : * Determines and returns the cost of scanning an RTE_RESULT relation.
1786 : */
1787 : void
1788 4244 : cost_resultscan(Path *path, PlannerInfo *root,
1789 : RelOptInfo *baserel, ParamPathInfo *param_info)
1790 : {
1791 4244 : Cost startup_cost = 0;
1792 4244 : 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 4244 : if (param_info)
1802 150 : path->rows = param_info->ppi_rows;
1803 : else
1804 4094 : path->rows = baserel->rows;
1805 :
1806 : /* We charge qual cost plus cpu_tuple_cost */
1807 4244 : get_restriction_qual_cost(root, baserel, param_info, &qpqual_cost);
1808 :
1809 4244 : startup_cost += qpqual_cost.startup;
1810 4244 : cpu_per_tuple = cpu_tuple_cost + qpqual_cost.per_tuple;
1811 4244 : run_cost += cpu_per_tuple * baserel->tuples;
1812 :
1813 4244 : path->disabled_nodes = 0;
1814 4244 : path->startup_cost = startup_cost;
1815 4244 : path->total_cost = startup_cost + run_cost;
1816 4244 : }
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 1004 : 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 1004 : startup_cost = nrterm->startup_cost;
1834 1004 : total_cost = nrterm->total_cost;
1835 1004 : 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 1004 : total_cost += 10 * rterm->total_cost;
1844 1004 : 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 1004 : total_cost += cpu_tuple_cost * total_rows;
1852 :
1853 1004 : runion->disabled_nodes = nrterm->disabled_nodes + rterm->disabled_nodes;
1854 1004 : runion->startup_cost = startup_cost;
1855 1004 : runion->total_cost = total_cost;
1856 1004 : runion->rows = total_rows;
1857 1004 : runion->pathtarget->width = Max(nrterm->pathtarget->width,
1858 : rterm->pathtarget->width);
1859 1004 : }
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 1712880 : 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 1712880 : double input_bytes = relation_byte_size(tuples, width);
1904 : double output_bytes;
1905 : double output_tuples;
1906 1712880 : 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 1712880 : if (tuples < 2.0)
1913 509520 : tuples = 2.0;
1914 :
1915 : /* Include the default cost-per-comparison */
1916 1712880 : comparison_cost += 2.0 * cpu_operator_cost;
1917 :
1918 : /* Do we have a useful LIMIT? */
1919 1712880 : if (limit_tuples > 0 && limit_tuples < tuples)
1920 : {
1921 1874 : output_tuples = limit_tuples;
1922 1874 : output_bytes = relation_byte_size(output_tuples, width);
1923 : }
1924 : else
1925 : {
1926 1711006 : output_tuples = tuples;
1927 1711006 : output_bytes = input_bytes;
1928 : }
1929 :
1930 1712880 : if (output_bytes > sort_mem_bytes)
1931 : {
1932 : /*
1933 : * We'll have to use a disk-based sort of all the tuples
1934 : */
1935 20356 : double npages = ceil(input_bytes / BLCKSZ);
1936 20356 : double nruns = input_bytes / sort_mem_bytes;
1937 20356 : 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 20356 : *startup_cost = comparison_cost * tuples * LOG2(tuples);
1947 :
1948 : /* Disk costs */
1949 :
1950 : /* Compute logM(r) as log(r) / log(M) */
1951 20356 : if (nruns > mergeorder)
1952 5302 : log_runs = ceil(log(nruns) / log(mergeorder));
1953 : else
1954 15054 : log_runs = 1.0;
1955 20356 : npageaccesses = 2.0 * npages * log_runs;
1956 : /* Assume 3/4ths of accesses are sequential, 1/4th are not */
1957 20356 : *startup_cost += npageaccesses *
1958 20356 : (seq_page_cost * 0.75 + random_page_cost * 0.25);
1959 : }
1960 1692524 : 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 1370 : *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 1691154 : *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 1712880 : *run_cost = cpu_operator_cost * tuples;
1985 1712880 : }
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 11968 : 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 11968 : 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 11968 : List *presortedExprs = NIL;
2016 : ListCell *l;
2017 11968 : 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 11968 : if (input_tuples < 2.0)
2026 7146 : input_tuples = 2.0;
2027 :
2028 : /* Default estimate of number of groups, capped to one group per row. */
2029 11968 : 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 12064 : foreach(l, pathkeys)
2054 : {
2055 12064 : PathKey *key = (PathKey *) lfirst(l);
2056 12064 : EquivalenceMember *member = (EquivalenceMember *)
2057 12064 : 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 12064 : 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 12054 : presortedExprs = lappend(presortedExprs, member->em_expr);
2071 :
2072 12054 : if (foreach_current_index(l) + 1 >= presorted_keys)
2073 11958 : break;
2074 : }
2075 :
2076 : /* Estimate the number of groups with equal presorted keys. */
2077 11968 : if (!unknown_varno)
2078 11958 : input_groups = estimate_num_groups(root, presortedExprs, input_tuples,
2079 : NULL, NULL);
2080 :
2081 11968 : group_tuples = input_tuples / input_groups;
2082 11968 : 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 11968 : 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 11968 : 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 11968 : run_cost = group_run_cost + (group_run_cost + group_startup_cost) *
2106 11968 : (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 11968 : 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 11968 : run_cost += 2.0 * cpu_tuple_cost * input_groups;
2120 :
2121 11968 : path->rows = input_tuples;
2122 :
2123 : /* should not generate these paths when enable_incremental_sort=false */
2124 : Assert(enable_incremental_sort);
2125 11968 : path->disabled_nodes = input_disabled_nodes;
2126 :
2127 11968 : path->startup_cost = startup_cost;
2128 11968 : path->total_cost = startup_cost + run_cost;
2129 11968 : }
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 1700912 : 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 1700912 : cost_tuplesort(&startup_cost, &run_cost,
2155 : tuples, width,
2156 : comparison_cost, sort_mem,
2157 : limit_tuples);
2158 :
2159 1700912 : startup_cost += input_cost;
2160 :
2161 1700912 : path->rows = tuples;
2162 1700912 : path->disabled_nodes = input_disabled_nodes + (enable_sort ? 0 : 1);
2163 1700912 : path->startup_cost = startup_cost;
2164 1700912 : path->total_cost = startup_cost + run_cost;
2165 1700912 : }
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 18312 : 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 18312 : if (numpaths == 0)
2185 14166 : return 0;
2186 :
2187 : /*
2188 : * Array length is number of workers or number of relevant paths,
2189 : * whichever is less.
2190 : */
2191 4146 : arrlen = Min(parallel_workers, numpaths);
2192 4146 : costarr = (Cost *) palloc(sizeof(Cost) * arrlen);
2193 :
2194 : /* The first few paths will each be claimed by a different worker. */
2195 4146 : path_index = 0;
2196 11964 : foreach(cell, subpaths)
2197 : {
2198 8662 : Path *subpath = (Path *) lfirst(cell);
2199 :
2200 8662 : if (path_index == arrlen)
2201 844 : break;
2202 7818 : 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 4146 : 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 4628 : 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 196 : min_index = i;
2231 : }
2232 : }
2233 :
2234 : /* Return the highest cost from the array */
2235 4146 : max_index = 0;
2236 11964 : for (int i = 0; i < arrlen; i++)
2237 : {
2238 7818 : if (costarr[i] > costarr[max_index])
2239 188 : max_index = i;
2240 : }
2241 :
2242 4146 : return costarr[max_index];
2243 : }
2244 :
2245 : /*
2246 : * cost_append
2247 : * Determines and returns the cost of an Append node.
2248 : */
2249 : void
2250 54110 : cost_append(AppendPath *apath)
2251 : {
2252 : ListCell *l;
2253 :
2254 54110 : apath->path.disabled_nodes = 0;
2255 54110 : apath->path.startup_cost = 0;
2256 54110 : apath->path.total_cost = 0;
2257 54110 : apath->path.rows = 0;
2258 :
2259 54110 : if (apath->subpaths == NIL)
2260 1642 : return;
2261 :
2262 52468 : if (!apath->path.parallel_aware)
2263 : {
2264 34156 : List *pathkeys = apath->path.pathkeys;
2265 :
2266 34156 : if (pathkeys == NIL)
2267 : {
2268 32026 : 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 32026 : 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 123948 : foreach(l, apath->subpaths)
2281 : {
2282 91922 : Path *subpath = (Path *) lfirst(l);
2283 :
2284 91922 : apath->path.rows += subpath->rows;
2285 91922 : apath->path.disabled_nodes += subpath->disabled_nodes;
2286 91922 : 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 8318 : foreach(l, apath->subpaths)
2310 : {
2311 6188 : Path *subpath = (Path *) lfirst(l);
2312 : Path sort_path; /* dummy for result of cost_sort */
2313 :
2314 6188 : if (!pathkeys_contained_in(pathkeys, subpath->pathkeys))
2315 : {
2316 : /*
2317 : * We'll need to insert a Sort node, so include costs for
2318 : * that. We can use the parent's LIMIT if any, since we
2319 : * certainly won't pull more than that many tuples from
2320 : * any child.
2321 : */
2322 44 : cost_sort(&sort_path,
2323 : NULL, /* doesn't currently need root */
2324 : pathkeys,
2325 : subpath->disabled_nodes,
2326 : subpath->total_cost,
2327 : subpath->rows,
2328 44 : subpath->pathtarget->width,
2329 : 0.0,
2330 : work_mem,
2331 : apath->limit_tuples);
2332 44 : subpath = &sort_path;
2333 : }
2334 :
2335 6188 : apath->path.rows += subpath->rows;
2336 6188 : apath->path.disabled_nodes += subpath->disabled_nodes;
2337 6188 : apath->path.startup_cost += subpath->startup_cost;
2338 6188 : apath->path.total_cost += subpath->total_cost;
2339 : }
2340 : }
2341 : }
2342 : else /* parallel-aware */
2343 : {
2344 18312 : int i = 0;
2345 18312 : double parallel_divisor = get_parallel_divisor(&apath->path);
2346 :
2347 : /* Parallel-aware Append never produces ordered output. */
2348 : Assert(apath->path.pathkeys == NIL);
2349 :
2350 : /* Calculate startup cost. */
2351 71472 : foreach(l, apath->subpaths)
2352 : {
2353 53160 : Path *subpath = (Path *) lfirst(l);
2354 :
2355 : /*
2356 : * Append will start returning tuples when the child node having
2357 : * lowest startup cost is done setting up. We consider only the
2358 : * first few subplans that immediately get a worker assigned.
2359 : */
2360 53160 : if (i == 0)
2361 18312 : apath->path.startup_cost = subpath->startup_cost;
2362 34848 : else if (i < apath->path.parallel_workers)
2363 17754 : apath->path.startup_cost = Min(apath->path.startup_cost,
2364 : subpath->startup_cost);
2365 :
2366 : /*
2367 : * Apply parallel divisor to subpaths. Scale the number of rows
2368 : * for each partial subpath based on the ratio of the parallel
2369 : * divisor originally used for the subpath to the one we adopted.
2370 : * Also add the cost of partial paths to the total cost, but
2371 : * ignore non-partial paths for now.
2372 : */
2373 53160 : if (i < apath->first_partial_path)
2374 8300 : apath->path.rows += subpath->rows / parallel_divisor;
2375 : else
2376 : {
2377 : double subpath_parallel_divisor;
2378 :
2379 44860 : subpath_parallel_divisor = get_parallel_divisor(subpath);
2380 44860 : apath->path.rows += subpath->rows * (subpath_parallel_divisor /
2381 : parallel_divisor);
2382 44860 : apath->path.total_cost += subpath->total_cost;
2383 : }
2384 :
2385 53160 : apath->path.disabled_nodes += subpath->disabled_nodes;
2386 53160 : apath->path.rows = clamp_row_est(apath->path.rows);
2387 :
2388 53160 : i++;
2389 : }
2390 :
2391 : /* Add cost for non-partial subpaths. */
2392 18312 : apath->path.total_cost +=
2393 18312 : append_nonpartial_cost(apath->subpaths,
2394 : apath->first_partial_path,
2395 : apath->path.parallel_workers);
2396 : }
2397 :
2398 : /*
2399 : * Although Append does not do any selection or projection, it's not free;
2400 : * add a small per-tuple overhead.
2401 : */
2402 52468 : apath->path.total_cost +=
2403 52468 : cpu_tuple_cost * APPEND_CPU_COST_MULTIPLIER * apath->path.rows;
2404 : }
2405 :
2406 : /*
2407 : * cost_merge_append
2408 : * Determines and returns the cost of a MergeAppend node.
2409 : *
2410 : * MergeAppend merges several pre-sorted input streams, using a heap that
2411 : * at any given instant holds the next tuple from each stream. If there
2412 : * are N streams, we need about N*log2(N) tuple comparisons to construct
2413 : * the heap at startup, and then for each output tuple, about log2(N)
2414 : * comparisons to replace the top entry.
2415 : *
2416 : * (The effective value of N will drop once some of the input streams are
2417 : * exhausted, but it seems unlikely to be worth trying to account for that.)
2418 : *
2419 : * The heap is never spilled to disk, since we assume N is not very large.
2420 : * So this is much simpler than cost_sort.
2421 : *
2422 : * As in cost_sort, we charge two operator evals per tuple comparison.
2423 : *
2424 : * 'pathkeys' is a list of sort keys
2425 : * 'n_streams' is the number of input streams
2426 : * 'input_disabled_nodes' is the sum of the input streams' disabled node counts
2427 : * 'input_startup_cost' is the sum of the input streams' startup costs
2428 : * 'input_total_cost' is the sum of the input streams' total costs
2429 : * 'tuples' is the number of tuples in all the streams
2430 : */
2431 : void
2432 4166 : cost_merge_append(Path *path, PlannerInfo *root,
2433 : List *pathkeys, int n_streams,
2434 : int input_disabled_nodes,
2435 : Cost input_startup_cost, Cost input_total_cost,
2436 : double tuples)
2437 : {
2438 4166 : Cost startup_cost = 0;
2439 4166 : Cost run_cost = 0;
2440 : Cost comparison_cost;
2441 : double N;
2442 : double logN;
2443 :
2444 : /*
2445 : * Avoid log(0)...
2446 : */
2447 4166 : N = (n_streams < 2) ? 2.0 : (double) n_streams;
2448 4166 : logN = LOG2(N);
2449 :
2450 : /* Assumed cost per tuple comparison */
2451 4166 : comparison_cost = 2.0 * cpu_operator_cost;
2452 :
2453 : /* Heap creation cost */
2454 4166 : startup_cost += comparison_cost * N * logN;
2455 :
2456 : /* Per-tuple heap maintenance cost */
2457 4166 : run_cost += tuples * comparison_cost * logN;
2458 :
2459 : /*
2460 : * Although MergeAppend does not do any selection or projection, it's not
2461 : * free; add a small per-tuple overhead.
2462 : */
2463 4166 : run_cost += cpu_tuple_cost * APPEND_CPU_COST_MULTIPLIER * tuples;
2464 :
2465 4166 : path->disabled_nodes = input_disabled_nodes;
2466 4166 : path->startup_cost = startup_cost + input_startup_cost;
2467 4166 : path->total_cost = startup_cost + run_cost + input_total_cost;
2468 4166 : }
2469 :
2470 : /*
2471 : * cost_material
2472 : * Determines and returns the cost of materializing a relation, including
2473 : * the cost of reading the input data.
2474 : *
2475 : * If the total volume of data to materialize exceeds work_mem, we will need
2476 : * to write it to disk, so the cost is much higher in that case.
2477 : *
2478 : * Note that here we are estimating the costs for the first scan of the
2479 : * relation, so the materialization is all overhead --- any savings will
2480 : * occur only on rescan, which is estimated in cost_rescan.
2481 : */
2482 : void
2483 530244 : cost_material(Path *path,
2484 : int input_disabled_nodes,
2485 : Cost input_startup_cost, Cost input_total_cost,
2486 : double tuples, int width)
2487 : {
2488 530244 : Cost startup_cost = input_startup_cost;
2489 530244 : Cost run_cost = input_total_cost - input_startup_cost;
2490 530244 : double nbytes = relation_byte_size(tuples, width);
2491 530244 : double work_mem_bytes = work_mem * (Size) 1024;
2492 :
2493 530244 : path->rows = tuples;
2494 :
2495 : /*
2496 : * Whether spilling or not, charge 2x cpu_operator_cost per tuple to
2497 : * reflect bookkeeping overhead. (This rate must be more than what
2498 : * cost_rescan charges for materialize, ie, cpu_operator_cost per tuple;
2499 : * if it is exactly the same then there will be a cost tie between
2500 : * nestloop with A outer, materialized B inner and nestloop with B outer,
2501 : * materialized A inner. The extra cost ensures we'll prefer
2502 : * materializing the smaller rel.) Note that this is normally a good deal
2503 : * less than cpu_tuple_cost; which is OK because a Material plan node
2504 : * doesn't do qual-checking or projection, so it's got less overhead than
2505 : * most plan nodes.
2506 : */
2507 530244 : run_cost += 2 * cpu_operator_cost * tuples;
2508 :
2509 : /*
2510 : * If we will spill to disk, charge at the rate of seq_page_cost per page.
2511 : * This cost is assumed to be evenly spread through the plan run phase,
2512 : * which isn't exactly accurate but our cost model doesn't allow for
2513 : * nonuniform costs within the run phase.
2514 : */
2515 530244 : if (nbytes > work_mem_bytes)
2516 : {
2517 5472 : double npages = ceil(nbytes / BLCKSZ);
2518 :
2519 5472 : run_cost += seq_page_cost * npages;
2520 : }
2521 :
2522 530244 : path->disabled_nodes = input_disabled_nodes + (enable_material ? 0 : 1);
2523 530244 : path->startup_cost = startup_cost;
2524 530244 : path->total_cost = startup_cost + run_cost;
2525 530244 : }
2526 :
2527 : /*
2528 : * cost_memoize_rescan
2529 : * Determines the estimated cost of rescanning a Memoize node.
2530 : *
2531 : * In order to estimate this, we must gain knowledge of how often we expect to
2532 : * be called and how many distinct sets of parameters we are likely to be
2533 : * called with. If we expect a good cache hit ratio, then we can set our
2534 : * costs to account for that hit ratio, plus a little bit of cost for the
2535 : * caching itself. Caching will not work out well if we expect to be called
2536 : * with too many distinct parameter values. The worst-case here is that we
2537 : * never see any parameter value twice, in which case we'd never get a cache
2538 : * hit and caching would be a complete waste of effort.
2539 : */
2540 : static void
2541 290258 : cost_memoize_rescan(PlannerInfo *root, MemoizePath *mpath,
2542 : Cost *rescan_startup_cost, Cost *rescan_total_cost)
2543 : {
2544 : EstimationInfo estinfo;
2545 : ListCell *lc;
2546 290258 : Cost input_startup_cost = mpath->subpath->startup_cost;
2547 290258 : Cost input_total_cost = mpath->subpath->total_cost;
2548 290258 : double tuples = mpath->subpath->rows;
2549 290258 : double calls = mpath->calls;
2550 290258 : int width = mpath->subpath->pathtarget->width;
2551 :
2552 : double hash_mem_bytes;
2553 : double est_entry_bytes;
2554 : double est_cache_entries;
2555 : double ndistinct;
2556 : double evict_ratio;
2557 : double hit_ratio;
2558 : Cost startup_cost;
2559 : Cost total_cost;
2560 :
2561 : /* available cache space */
2562 290258 : hash_mem_bytes = get_hash_memory_limit();
2563 :
2564 : /*
2565 : * Set the number of bytes each cache entry should consume in the cache.
2566 : * To provide us with better estimations on how many cache entries we can
2567 : * store at once, we make a call to the executor here to ask it what
2568 : * memory overheads there are for a single cache entry.
2569 : */
2570 290258 : est_entry_bytes = relation_byte_size(tuples, width) +
2571 290258 : ExecEstimateCacheEntryOverheadBytes(tuples);
2572 :
2573 : /* include the estimated width for the cache keys */
2574 618206 : foreach(lc, mpath->param_exprs)
2575 327948 : est_entry_bytes += get_expr_width(root, (Node *) lfirst(lc));
2576 :
2577 : /* estimate on the upper limit of cache entries we can hold at once */
2578 290258 : est_cache_entries = floor(hash_mem_bytes / est_entry_bytes);
2579 :
2580 : /* estimate on the distinct number of parameter values */
2581 290258 : ndistinct = estimate_num_groups(root, mpath->param_exprs, calls, NULL,
2582 : &estinfo);
2583 :
2584 : /*
2585 : * When the estimation fell back on using a default value, it's a bit too
2586 : * risky to assume that it's ok to use a Memoize node. The use of a
2587 : * default could cause us to use a Memoize node when it's really
2588 : * inappropriate to do so. If we see that this has been done, then we'll
2589 : * assume that every call will have unique parameters, which will almost
2590 : * certainly mean a MemoizePath will never survive add_path().
2591 : */
2592 290258 : if ((estinfo.flags & SELFLAG_USED_DEFAULT) != 0)
2593 15516 : ndistinct = calls;
2594 :
2595 : /*
2596 : * Since we've already estimated the maximum number of entries we can
2597 : * store at once and know the estimated number of distinct values we'll be
2598 : * called with, we'll take this opportunity to set the path's est_entries.
2599 : * This will ultimately determine the hash table size that the executor
2600 : * will use. If we leave this at zero, the executor will just choose the
2601 : * size itself. Really this is not the right place to do this, but it's
2602 : * convenient since everything is already calculated.
2603 : */
2604 290258 : mpath->est_entries = Min(Min(ndistinct, est_cache_entries),
2605 : PG_UINT32_MAX);
2606 :
2607 : /*
2608 : * When the number of distinct parameter values is above the amount we can
2609 : * store in the cache, then we'll have to evict some entries from the
2610 : * cache. This is not free. Here we estimate how often we'll incur the
2611 : * cost of that eviction.
2612 : */
2613 290258 : evict_ratio = 1.0 - Min(est_cache_entries, ndistinct) / ndistinct;
2614 :
2615 : /*
2616 : * In order to estimate how costly a single scan will be, we need to
2617 : * attempt to estimate what the cache hit ratio will be. To do that we
2618 : * must look at how many scans are estimated in total for this node and
2619 : * how many of those scans we expect to get a cache hit.
2620 : */
2621 580516 : hit_ratio = ((calls - ndistinct) / calls) *
2622 290258 : (est_cache_entries / Max(ndistinct, est_cache_entries));
2623 :
2624 : Assert(hit_ratio >= 0 && hit_ratio <= 1.0);
2625 :
2626 : /*
2627 : * Set the total_cost accounting for the expected cache hit ratio. We
2628 : * also add on a cpu_operator_cost to account for a cache lookup. This
2629 : * will happen regardless of whether it's a cache hit or not.
2630 : */
2631 290258 : total_cost = input_total_cost * (1.0 - hit_ratio) + cpu_operator_cost;
2632 :
2633 : /* Now adjust the total cost to account for cache evictions */
2634 :
2635 : /* Charge a cpu_tuple_cost for evicting the actual cache entry */
2636 290258 : total_cost += cpu_tuple_cost * evict_ratio;
2637 :
2638 : /*
2639 : * Charge a 10th of cpu_operator_cost to evict every tuple in that entry.
2640 : * The per-tuple eviction is really just a pfree, so charging a whole
2641 : * cpu_operator_cost seems a little excessive.
2642 : */
2643 290258 : total_cost += cpu_operator_cost / 10.0 * evict_ratio * tuples;
2644 :
2645 : /*
2646 : * Now adjust for storing things in the cache, since that's not free
2647 : * either. Everything must go in the cache. We don't proportion this
2648 : * over any ratio, just apply it once for the scan. We charge a
2649 : * cpu_tuple_cost for the creation of the cache entry and also a
2650 : * cpu_operator_cost for each tuple we expect to cache.
2651 : */
2652 290258 : total_cost += cpu_tuple_cost + cpu_operator_cost * tuples;
2653 :
2654 : /*
2655 : * Getting the first row must be also be proportioned according to the
2656 : * expected cache hit ratio.
2657 : */
2658 290258 : startup_cost = input_startup_cost * (1.0 - hit_ratio);
2659 :
2660 : /*
2661 : * Additionally we charge a cpu_tuple_cost to account for cache lookups,
2662 : * which we'll do regardless of whether it was a cache hit or not.
2663 : */
2664 290258 : startup_cost += cpu_tuple_cost;
2665 :
2666 290258 : *rescan_startup_cost = startup_cost;
2667 290258 : *rescan_total_cost = total_cost;
2668 290258 : }
2669 :
2670 : /*
2671 : * cost_agg
2672 : * Determines and returns the cost of performing an Agg plan node,
2673 : * including the cost of its input.
2674 : *
2675 : * aggcosts can be NULL when there are no actual aggregate functions (i.e.,
2676 : * we are using a hashed Agg node just to do grouping).
2677 : *
2678 : * Note: when aggstrategy == AGG_SORTED, caller must ensure that input costs
2679 : * are for appropriately-sorted input.
2680 : */
2681 : void
2682 69284 : cost_agg(Path *path, PlannerInfo *root,
2683 : AggStrategy aggstrategy, const AggClauseCosts *aggcosts,
2684 : int numGroupCols, double numGroups,
2685 : List *quals,
2686 : int disabled_nodes,
2687 : Cost input_startup_cost, Cost input_total_cost,
2688 : double input_tuples, double input_width)
2689 : {
2690 : double output_tuples;
2691 : Cost startup_cost;
2692 : Cost total_cost;
2693 69284 : const AggClauseCosts dummy_aggcosts = {0};
2694 :
2695 : /* Use all-zero per-aggregate costs if NULL is passed */
2696 69284 : if (aggcosts == NULL)
2697 : {
2698 : Assert(aggstrategy == AGG_HASHED);
2699 12584 : aggcosts = &dummy_aggcosts;
2700 : }
2701 :
2702 : /*
2703 : * The transCost.per_tuple component of aggcosts should be charged once
2704 : * per input tuple, corresponding to the costs of evaluating the aggregate
2705 : * transfns and their input expressions. The finalCost.per_tuple component
2706 : * is charged once per output tuple, corresponding to the costs of
2707 : * evaluating the finalfns. Startup costs are of course charged but once.
2708 : *
2709 : * If we are grouping, we charge an additional cpu_operator_cost per
2710 : * grouping column per input tuple for grouping comparisons.
2711 : *
2712 : * We will produce a single output tuple if not grouping, and a tuple per
2713 : * group otherwise. We charge cpu_tuple_cost for each output tuple.
2714 : *
2715 : * Note: in this cost model, AGG_SORTED and AGG_HASHED have exactly the
2716 : * same total CPU cost, but AGG_SORTED has lower startup cost. If the
2717 : * input path is already sorted appropriately, AGG_SORTED should be
2718 : * preferred (since it has no risk of memory overflow). This will happen
2719 : * as long as the computed total costs are indeed exactly equal --- but if
2720 : * there's roundoff error we might do the wrong thing. So be sure that
2721 : * the computations below form the same intermediate values in the same
2722 : * order.
2723 : */
2724 69284 : if (aggstrategy == AGG_PLAIN)
2725 : {
2726 37096 : startup_cost = input_total_cost;
2727 37096 : startup_cost += aggcosts->transCost.startup;
2728 37096 : startup_cost += aggcosts->transCost.per_tuple * input_tuples;
2729 37096 : startup_cost += aggcosts->finalCost.startup;
2730 37096 : startup_cost += aggcosts->finalCost.per_tuple;
2731 : /* we aren't grouping */
2732 37096 : total_cost = startup_cost + cpu_tuple_cost;
2733 37096 : output_tuples = 1;
2734 : }
2735 32188 : else if (aggstrategy == AGG_SORTED || aggstrategy == AGG_MIXED)
2736 : {
2737 : /* Here we are able to deliver output on-the-fly */
2738 11244 : startup_cost = input_startup_cost;
2739 11244 : total_cost = input_total_cost;
2740 11244 : if (aggstrategy == AGG_MIXED && !enable_hashagg)
2741 456 : ++disabled_nodes;
2742 : /* calcs phrased this way to match HASHED case, see note above */
2743 11244 : total_cost += aggcosts->transCost.startup;
2744 11244 : total_cost += aggcosts->transCost.per_tuple * input_tuples;
2745 11244 : total_cost += (cpu_operator_cost * numGroupCols) * input_tuples;
2746 11244 : total_cost += aggcosts->finalCost.startup;
2747 11244 : total_cost += aggcosts->finalCost.per_tuple * numGroups;
2748 11244 : total_cost += cpu_tuple_cost * numGroups;
2749 11244 : output_tuples = numGroups;
2750 : }
2751 : else
2752 : {
2753 : /* must be AGG_HASHED */
2754 20944 : startup_cost = input_total_cost;
2755 20944 : if (!enable_hashagg)
2756 1578 : ++disabled_nodes;
2757 20944 : startup_cost += aggcosts->transCost.startup;
2758 20944 : startup_cost += aggcosts->transCost.per_tuple * input_tuples;
2759 : /* cost of computing hash value */
2760 20944 : startup_cost += (cpu_operator_cost * numGroupCols) * input_tuples;
2761 20944 : startup_cost += aggcosts->finalCost.startup;
2762 :
2763 20944 : total_cost = startup_cost;
2764 20944 : total_cost += aggcosts->finalCost.per_tuple * numGroups;
2765 : /* cost of retrieving from hash table */
2766 20944 : total_cost += cpu_tuple_cost * numGroups;
2767 20944 : output_tuples = numGroups;
2768 : }
2769 :
2770 : /*
2771 : * Add the disk costs of hash aggregation that spills to disk.
2772 : *
2773 : * Groups that go into the hash table stay in memory until finalized, so
2774 : * spilling and reprocessing tuples doesn't incur additional invocations
2775 : * of transCost or finalCost. Furthermore, the computed hash value is
2776 : * stored with the spilled tuples, so we don't incur extra invocations of
2777 : * the hash function.
2778 : *
2779 : * Hash Agg begins returning tuples after the first batch is complete.
2780 : * Accrue writes (spilled tuples) to startup_cost and to total_cost;
2781 : * accrue reads only to total_cost.
2782 : */
2783 69284 : if (aggstrategy == AGG_HASHED || aggstrategy == AGG_MIXED)
2784 : {
2785 : double pages;
2786 21860 : double pages_written = 0.0;
2787 21860 : double pages_read = 0.0;
2788 : double spill_cost;
2789 : double hashentrysize;
2790 : double nbatches;
2791 : Size mem_limit;
2792 : uint64 ngroups_limit;
2793 : int num_partitions;
2794 : int depth;
2795 :
2796 : /*
2797 : * Estimate number of batches based on the computed limits. If less
2798 : * than or equal to one, all groups are expected to fit in memory;
2799 : * otherwise we expect to spill.
2800 : */
2801 21860 : hashentrysize = hash_agg_entry_size(list_length(root->aggtransinfos),
2802 : input_width,
2803 : aggcosts->transitionSpace);
2804 21860 : hash_agg_set_limits(hashentrysize, numGroups, 0, &mem_limit,
2805 : &ngroups_limit, &num_partitions);
2806 :
2807 21860 : nbatches = Max((numGroups * hashentrysize) / mem_limit,
2808 : numGroups / ngroups_limit);
2809 :
2810 21860 : nbatches = Max(ceil(nbatches), 1.0);
2811 21860 : num_partitions = Max(num_partitions, 2);
2812 :
2813 : /*
2814 : * The number of partitions can change at different levels of
2815 : * recursion; but for the purposes of this calculation assume it stays
2816 : * constant.
2817 : */
2818 21860 : depth = ceil(log(nbatches) / log(num_partitions));
2819 :
2820 : /*
2821 : * Estimate number of pages read and written. For each level of
2822 : * recursion, a tuple must be written and then later read.
2823 : */
2824 21860 : pages = relation_byte_size(input_tuples, input_width) / BLCKSZ;
2825 21860 : pages_written = pages_read = pages * depth;
2826 :
2827 : /*
2828 : * HashAgg has somewhat worse IO behavior than Sort on typical
2829 : * hardware/OS combinations. Account for this with a generic penalty.
2830 : */
2831 21860 : pages_read *= 2.0;
2832 21860 : pages_written *= 2.0;
2833 :
2834 21860 : startup_cost += pages_written * random_page_cost;
2835 21860 : total_cost += pages_written * random_page_cost;
2836 21860 : total_cost += pages_read * seq_page_cost;
2837 :
2838 : /* account for CPU cost of spilling a tuple and reading it back */
2839 21860 : spill_cost = depth * input_tuples * 2.0 * cpu_tuple_cost;
2840 21860 : startup_cost += spill_cost;
2841 21860 : total_cost += spill_cost;
2842 : }
2843 :
2844 : /*
2845 : * If there are quals (HAVING quals), account for their cost and
2846 : * selectivity.
2847 : */
2848 69284 : if (quals)
2849 : {
2850 : QualCost qual_cost;
2851 :
2852 4364 : cost_qual_eval(&qual_cost, quals, root);
2853 4364 : startup_cost += qual_cost.startup;
2854 4364 : total_cost += qual_cost.startup + output_tuples * qual_cost.per_tuple;
2855 :
2856 4364 : output_tuples = clamp_row_est(output_tuples *
2857 4364 : clauselist_selectivity(root,
2858 : quals,
2859 : 0,
2860 : JOIN_INNER,
2861 : NULL));
2862 : }
2863 :
2864 69284 : path->rows = output_tuples;
2865 69284 : path->disabled_nodes = disabled_nodes;
2866 69284 : path->startup_cost = startup_cost;
2867 69284 : path->total_cost = total_cost;
2868 69284 : }
2869 :
2870 : /*
2871 : * get_windowclause_startup_tuples
2872 : * Estimate how many tuples we'll need to fetch from a WindowAgg's
2873 : * subnode before we can output the first WindowAgg tuple.
2874 : *
2875 : * How many tuples need to be read depends on the WindowClause. For example,
2876 : * a WindowClause with no PARTITION BY and no ORDER BY requires that all
2877 : * subnode tuples are read and aggregated before the WindowAgg can output
2878 : * anything. If there's a PARTITION BY, then we only need to look at tuples
2879 : * in the first partition. Here we attempt to estimate just how many
2880 : * 'input_tuples' the WindowAgg will need to read for the given WindowClause
2881 : * before the first tuple can be output.
2882 : */
2883 : static double
2884 2754 : get_windowclause_startup_tuples(PlannerInfo *root, WindowClause *wc,
2885 : double input_tuples)
2886 : {
2887 2754 : int frameOptions = wc->frameOptions;
2888 : double partition_tuples;
2889 : double return_tuples;
2890 : double peer_tuples;
2891 :
2892 : /*
2893 : * First, figure out how many partitions there are likely to be and set
2894 : * partition_tuples according to that estimate.
2895 : */
2896 2754 : if (wc->partitionClause != NIL)
2897 : {
2898 : double num_partitions;
2899 716 : List *partexprs = get_sortgrouplist_exprs(wc->partitionClause,
2900 716 : root->parse->targetList);
2901 :
2902 716 : num_partitions = estimate_num_groups(root, partexprs, input_tuples,
2903 : NULL, NULL);
2904 716 : list_free(partexprs);
2905 :
2906 716 : partition_tuples = input_tuples / num_partitions;
2907 : }
2908 : else
2909 : {
2910 : /* all tuples belong to the same partition */
2911 2038 : partition_tuples = input_tuples;
2912 : }
2913 :
2914 : /* estimate the number of tuples in each peer group */
2915 2754 : if (wc->orderClause != NIL)
2916 : {
2917 : double num_groups;
2918 : List *orderexprs;
2919 :
2920 2274 : orderexprs = get_sortgrouplist_exprs(wc->orderClause,
2921 2274 : root->parse->targetList);
2922 :
2923 : /* estimate out how many peer groups there are in the partition */
2924 2274 : num_groups = estimate_num_groups(root, orderexprs,
2925 : partition_tuples, NULL,
2926 : NULL);
2927 2274 : list_free(orderexprs);
2928 2274 : peer_tuples = partition_tuples / num_groups;
2929 : }
2930 : else
2931 : {
2932 : /* no ORDER BY so only 1 tuple belongs in each peer group */
2933 480 : peer_tuples = 1.0;
2934 : }
2935 :
2936 2754 : if (frameOptions & FRAMEOPTION_END_UNBOUNDED_FOLLOWING)
2937 : {
2938 : /* include all partition rows */
2939 346 : return_tuples = partition_tuples;
2940 : }
2941 2408 : else if (frameOptions & FRAMEOPTION_END_CURRENT_ROW)
2942 : {
2943 1418 : if (frameOptions & FRAMEOPTION_ROWS)
2944 : {
2945 : /* just count the current row */
2946 608 : return_tuples = 1.0;
2947 : }
2948 810 : else if (frameOptions & (FRAMEOPTION_RANGE | FRAMEOPTION_GROUPS))
2949 : {
2950 : /*
2951 : * When in RANGE/GROUPS mode, it's more complex. If there's no
2952 : * ORDER BY, then all rows in the partition are peers, otherwise
2953 : * we'll need to read the first group of peers.
2954 : */
2955 810 : if (wc->orderClause == NIL)
2956 308 : return_tuples = partition_tuples;
2957 : else
2958 502 : return_tuples = peer_tuples;
2959 : }
2960 : else
2961 : {
2962 : /*
2963 : * Something new we don't support yet? This needs attention.
2964 : * We'll just return 1.0 in the meantime.
2965 : */
2966 : Assert(false);
2967 0 : return_tuples = 1.0;
2968 : }
2969 : }
2970 990 : else if (frameOptions & FRAMEOPTION_END_OFFSET_PRECEDING)
2971 : {
2972 : /*
2973 : * BETWEEN ... AND N PRECEDING will only need to read the WindowAgg's
2974 : * subnode after N ROWS/RANGES/GROUPS. N can be 0, but not negative,
2975 : * so we'll just assume only the current row needs to be read to fetch
2976 : * the first WindowAgg row.
2977 : */
2978 108 : return_tuples = 1.0;
2979 : }
2980 882 : else if (frameOptions & FRAMEOPTION_END_OFFSET_FOLLOWING)
2981 : {
2982 882 : Const *endOffset = (Const *) wc->endOffset;
2983 : double end_offset_value;
2984 :
2985 : /* try and figure out the value specified in the endOffset. */
2986 882 : if (IsA(endOffset, Const))
2987 : {
2988 882 : if (endOffset->constisnull)
2989 : {
2990 : /*
2991 : * NULLs are not allowed, but currently, there's no code to
2992 : * error out if there's a NULL Const. We'll only discover
2993 : * this during execution. For now, just pretend everything is
2994 : * fine and assume that just the first row/range/group will be
2995 : * needed.
2996 : */
2997 0 : end_offset_value = 1.0;
2998 : }
2999 : else
3000 : {
3001 882 : switch (endOffset->consttype)
3002 : {
3003 24 : case INT2OID:
3004 24 : end_offset_value =
3005 24 : (double) DatumGetInt16(endOffset->constvalue);
3006 24 : break;
3007 132 : case INT4OID:
3008 132 : end_offset_value =
3009 132 : (double) DatumGetInt32(endOffset->constvalue);
3010 132 : break;
3011 384 : case INT8OID:
3012 384 : end_offset_value =
3013 384 : (double) DatumGetInt64(endOffset->constvalue);
3014 384 : break;
3015 342 : default:
3016 342 : end_offset_value =
3017 342 : partition_tuples / peer_tuples *
3018 : DEFAULT_INEQ_SEL;
3019 342 : break;
3020 : }
3021 : }
3022 : }
3023 : else
3024 : {
3025 : /*
3026 : * When the end bound is not a Const, we'll just need to guess. We
3027 : * just make use of DEFAULT_INEQ_SEL.
3028 : */
3029 0 : end_offset_value =
3030 0 : partition_tuples / peer_tuples * DEFAULT_INEQ_SEL;
3031 : }
3032 :
3033 882 : if (frameOptions & FRAMEOPTION_ROWS)
3034 : {
3035 : /* include the N FOLLOWING and the current row */
3036 222 : return_tuples = end_offset_value + 1.0;
3037 : }
3038 660 : else if (frameOptions & (FRAMEOPTION_RANGE | FRAMEOPTION_GROUPS))
3039 : {
3040 : /* include N FOLLOWING ranges/group and the initial range/group */
3041 660 : return_tuples = peer_tuples * (end_offset_value + 1.0);
3042 : }
3043 : else
3044 : {
3045 : /*
3046 : * Something new we don't support yet? This needs attention.
3047 : * We'll just return 1.0 in the meantime.
3048 : */
3049 : Assert(false);
3050 0 : return_tuples = 1.0;
3051 : }
3052 : }
3053 : else
3054 : {
3055 : /*
3056 : * Something new we don't support yet? This needs attention. We'll
3057 : * just return 1.0 in the meantime.
3058 : */
3059 : Assert(false);
3060 0 : return_tuples = 1.0;
3061 : }
3062 :
3063 2754 : if (wc->partitionClause != NIL || wc->orderClause != NIL)
3064 : {
3065 : /*
3066 : * Cap the return value to the estimated partition tuples and account
3067 : * for the extra tuple WindowAgg will need to read to confirm the next
3068 : * tuple does not belong to the same partition or peer group.
3069 : */
3070 2474 : return_tuples = Min(return_tuples + 1.0, partition_tuples);
3071 : }
3072 : else
3073 : {
3074 : /*
3075 : * Cap the return value so it's never higher than the expected tuples
3076 : * in the partition.
3077 : */
3078 280 : return_tuples = Min(return_tuples, partition_tuples);
3079 : }
3080 :
3081 : /*
3082 : * We needn't worry about any EXCLUDE options as those only exclude rows
3083 : * from being aggregated, not from being read from the WindowAgg's
3084 : * subnode.
3085 : */
3086 :
3087 2754 : return clamp_row_est(return_tuples);
3088 : }
3089 :
3090 : /*
3091 : * cost_windowagg
3092 : * Determines and returns the cost of performing a WindowAgg plan node,
3093 : * including the cost of its input.
3094 : *
3095 : * Input is assumed already properly sorted.
3096 : */
3097 : void
3098 2754 : cost_windowagg(Path *path, PlannerInfo *root,
3099 : List *windowFuncs, WindowClause *winclause,
3100 : int input_disabled_nodes,
3101 : Cost input_startup_cost, Cost input_total_cost,
3102 : double input_tuples)
3103 : {
3104 : Cost startup_cost;
3105 : Cost total_cost;
3106 : double startup_tuples;
3107 : int numPartCols;
3108 : int numOrderCols;
3109 : ListCell *lc;
3110 :
3111 2754 : numPartCols = list_length(winclause->partitionClause);
3112 2754 : numOrderCols = list_length(winclause->orderClause);
3113 :
3114 2754 : startup_cost = input_startup_cost;
3115 2754 : total_cost = input_total_cost;
3116 :
3117 : /*
3118 : * Window functions are assumed to cost their stated execution cost, plus
3119 : * the cost of evaluating their input expressions, per tuple. Since they
3120 : * may in fact evaluate their inputs at multiple rows during each cycle,
3121 : * this could be a drastic underestimate; but without a way to know how
3122 : * many rows the window function will fetch, it's hard to do better. In
3123 : * any case, it's a good estimate for all the built-in window functions,
3124 : * so we'll just do this for now.
3125 : */
3126 6246 : foreach(lc, windowFuncs)
3127 : {
3128 3492 : WindowFunc *wfunc = lfirst_node(WindowFunc, lc);
3129 : Cost wfunccost;
3130 : QualCost argcosts;
3131 :
3132 3492 : argcosts.startup = argcosts.per_tuple = 0;
3133 3492 : add_function_cost(root, wfunc->winfnoid, (Node *) wfunc,
3134 : &argcosts);
3135 3492 : startup_cost += argcosts.startup;
3136 3492 : wfunccost = argcosts.per_tuple;
3137 :
3138 : /* also add the input expressions' cost to per-input-row costs */
3139 3492 : cost_qual_eval_node(&argcosts, (Node *) wfunc->args, root);
3140 3492 : startup_cost += argcosts.startup;
3141 3492 : wfunccost += argcosts.per_tuple;
3142 :
3143 : /*
3144 : * Add the filter's cost to per-input-row costs. XXX We should reduce
3145 : * input expression costs according to filter selectivity.
3146 : */
3147 3492 : cost_qual_eval_node(&argcosts, (Node *) wfunc->aggfilter, root);
3148 3492 : startup_cost += argcosts.startup;
3149 3492 : wfunccost += argcosts.per_tuple;
3150 :
3151 3492 : total_cost += wfunccost * input_tuples;
3152 : }
3153 :
3154 : /*
3155 : * We also charge cpu_operator_cost per grouping column per tuple for
3156 : * grouping comparisons, plus cpu_tuple_cost per tuple for general
3157 : * overhead.
3158 : *
3159 : * XXX this neglects costs of spooling the data to disk when it overflows
3160 : * work_mem. Sooner or later that should get accounted for.
3161 : */
3162 2754 : total_cost += cpu_operator_cost * (numPartCols + numOrderCols) * input_tuples;
3163 2754 : total_cost += cpu_tuple_cost * input_tuples;
3164 :
3165 2754 : path->rows = input_tuples;
3166 2754 : path->disabled_nodes = input_disabled_nodes;
3167 2754 : path->startup_cost = startup_cost;
3168 2754 : path->total_cost = total_cost;
3169 :
3170 : /*
3171 : * Also, take into account how many tuples we need to read from the
3172 : * subnode in order to produce the first tuple from the WindowAgg. To do
3173 : * this we proportion the run cost (total cost not including startup cost)
3174 : * over the estimated startup tuples. We already included the startup
3175 : * cost of the subnode, so we only need to do this when the estimated
3176 : * startup tuples is above 1.0.
3177 : */
3178 2754 : startup_tuples = get_windowclause_startup_tuples(root, winclause,
3179 : input_tuples);
3180 :
3181 2754 : if (startup_tuples > 1.0)
3182 2466 : path->startup_cost += (total_cost - startup_cost) / input_tuples *
3183 2466 : (startup_tuples - 1.0);
3184 2754 : }
3185 :
3186 : /*
3187 : * cost_group
3188 : * Determines and returns the cost of performing a Group plan node,
3189 : * including the cost of its input.
3190 : *
3191 : * Note: caller must ensure that input costs are for appropriately-sorted
3192 : * input.
3193 : */
3194 : void
3195 1214 : cost_group(Path *path, PlannerInfo *root,
3196 : int numGroupCols, double numGroups,
3197 : List *quals,
3198 : int input_disabled_nodes,
3199 : Cost input_startup_cost, Cost input_total_cost,
3200 : double input_tuples)
3201 : {
3202 : double output_tuples;
3203 : Cost startup_cost;
3204 : Cost total_cost;
3205 :
3206 1214 : output_tuples = numGroups;
3207 1214 : startup_cost = input_startup_cost;
3208 1214 : total_cost = input_total_cost;
3209 :
3210 : /*
3211 : * Charge one cpu_operator_cost per comparison per input tuple. We assume
3212 : * all columns get compared at most of the tuples.
3213 : */
3214 1214 : total_cost += cpu_operator_cost * input_tuples * numGroupCols;
3215 :
3216 : /*
3217 : * If there are quals (HAVING quals), account for their cost and
3218 : * selectivity.
3219 : */
3220 1214 : if (quals)
3221 : {
3222 : QualCost qual_cost;
3223 :
3224 0 : cost_qual_eval(&qual_cost, quals, root);
3225 0 : startup_cost += qual_cost.startup;
3226 0 : total_cost += qual_cost.startup + output_tuples * qual_cost.per_tuple;
3227 :
3228 0 : output_tuples = clamp_row_est(output_tuples *
3229 0 : clauselist_selectivity(root,
3230 : quals,
3231 : 0,
3232 : JOIN_INNER,
3233 : NULL));
3234 : }
3235 :
3236 1214 : path->rows = output_tuples;
3237 1214 : path->disabled_nodes = input_disabled_nodes;
3238 1214 : path->startup_cost = startup_cost;
3239 1214 : path->total_cost = total_cost;
3240 1214 : }
3241 :
3242 : /*
3243 : * initial_cost_nestloop
3244 : * Preliminary estimate of the cost of a nestloop join path.
3245 : *
3246 : * This must quickly produce lower-bound estimates of the path's startup and
3247 : * total costs. If we are unable to eliminate the proposed path from
3248 : * consideration using the lower bounds, final_cost_nestloop will be called
3249 : * to obtain the final estimates.
3250 : *
3251 : * The exact division of labor between this function and final_cost_nestloop
3252 : * is private to them, and represents a tradeoff between speed of the initial
3253 : * estimate and getting a tight lower bound. We choose to not examine the
3254 : * join quals here, since that's by far the most expensive part of the
3255 : * calculations. The end result is that CPU-cost considerations must be
3256 : * left for the second phase; and for SEMI/ANTI joins, we must also postpone
3257 : * incorporation of the inner path's run cost.
3258 : *
3259 : * 'workspace' is to be filled with startup_cost, total_cost, and perhaps
3260 : * other data to be used by final_cost_nestloop
3261 : * 'jointype' is the type of join to be performed
3262 : * 'outer_path' is the outer input to the join
3263 : * 'inner_path' is the inner input to the join
3264 : * 'extra' contains miscellaneous information about the join
3265 : */
3266 : void
3267 2877214 : initial_cost_nestloop(PlannerInfo *root, JoinCostWorkspace *workspace,
3268 : JoinType jointype,
3269 : Path *outer_path, Path *inner_path,
3270 : JoinPathExtraData *extra)
3271 : {
3272 : int disabled_nodes;
3273 2877214 : Cost startup_cost = 0;
3274 2877214 : Cost run_cost = 0;
3275 2877214 : double outer_path_rows = outer_path->rows;
3276 : Cost inner_rescan_start_cost;
3277 : Cost inner_rescan_total_cost;
3278 : Cost inner_run_cost;
3279 : Cost inner_rescan_run_cost;
3280 :
3281 : /* Count up disabled nodes. */
3282 2877214 : disabled_nodes = enable_nestloop ? 0 : 1;
3283 2877214 : disabled_nodes += inner_path->disabled_nodes;
3284 2877214 : disabled_nodes += outer_path->disabled_nodes;
3285 :
3286 : /* estimate costs to rescan the inner relation */
3287 2877214 : cost_rescan(root, inner_path,
3288 : &inner_rescan_start_cost,
3289 : &inner_rescan_total_cost);
3290 :
3291 : /* cost of source data */
3292 :
3293 : /*
3294 : * NOTE: clearly, we must pay both outer and inner paths' startup_cost
3295 : * before we can start returning tuples, so the join's startup cost is
3296 : * their sum. We'll also pay the inner path's rescan startup cost
3297 : * multiple times.
3298 : */
3299 2877214 : startup_cost += outer_path->startup_cost + inner_path->startup_cost;
3300 2877214 : run_cost += outer_path->total_cost - outer_path->startup_cost;
3301 2877214 : if (outer_path_rows > 1)
3302 2036094 : run_cost += (outer_path_rows - 1) * inner_rescan_start_cost;
3303 :
3304 2877214 : inner_run_cost = inner_path->total_cost - inner_path->startup_cost;
3305 2877214 : inner_rescan_run_cost = inner_rescan_total_cost - inner_rescan_start_cost;
3306 :
3307 2877214 : if (jointype == JOIN_SEMI || jointype == JOIN_ANTI ||
3308 2818798 : extra->inner_unique)
3309 : {
3310 : /*
3311 : * With a SEMI or ANTI join, or if the innerrel is known unique, the
3312 : * executor will stop after the first match.
3313 : *
3314 : * Getting decent estimates requires inspection of the join quals,
3315 : * which we choose to postpone to final_cost_nestloop.
3316 : */
3317 :
3318 : /* Save private data for final_cost_nestloop */
3319 1326212 : workspace->inner_run_cost = inner_run_cost;
3320 1326212 : workspace->inner_rescan_run_cost = inner_rescan_run_cost;
3321 : }
3322 : else
3323 : {
3324 : /* Normal case; we'll scan whole input rel for each outer row */
3325 1551002 : run_cost += inner_run_cost;
3326 1551002 : if (outer_path_rows > 1)
3327 1107692 : run_cost += (outer_path_rows - 1) * inner_rescan_run_cost;
3328 : }
3329 :
3330 : /* CPU costs left for later */
3331 :
3332 : /* Public result fields */
3333 2877214 : workspace->disabled_nodes = disabled_nodes;
3334 2877214 : workspace->startup_cost = startup_cost;
3335 2877214 : workspace->total_cost = startup_cost + run_cost;
3336 : /* Save private data for final_cost_nestloop */
3337 2877214 : workspace->run_cost = run_cost;
3338 2877214 : }
3339 :
3340 : /*
3341 : * final_cost_nestloop
3342 : * Final estimate of the cost and result size of a nestloop join path.
3343 : *
3344 : * 'path' is already filled in except for the rows and cost fields
3345 : * 'workspace' is the result from initial_cost_nestloop
3346 : * 'extra' contains miscellaneous information about the join
3347 : */
3348 : void
3349 1389978 : final_cost_nestloop(PlannerInfo *root, NestPath *path,
3350 : JoinCostWorkspace *workspace,
3351 : JoinPathExtraData *extra)
3352 : {
3353 1389978 : Path *outer_path = path->jpath.outerjoinpath;
3354 1389978 : Path *inner_path = path->jpath.innerjoinpath;
3355 1389978 : double outer_path_rows = outer_path->rows;
3356 1389978 : double inner_path_rows = inner_path->rows;
3357 1389978 : Cost startup_cost = workspace->startup_cost;
3358 1389978 : Cost run_cost = workspace->run_cost;
3359 : Cost cpu_per_tuple;
3360 : QualCost restrict_qual_cost;
3361 : double ntuples;
3362 :
3363 : /* Set the number of disabled nodes. */
3364 1389978 : path->jpath.path.disabled_nodes = workspace->disabled_nodes;
3365 :
3366 : /* Protect some assumptions below that rowcounts aren't zero */
3367 1389978 : if (outer_path_rows <= 0)
3368 0 : outer_path_rows = 1;
3369 1389978 : if (inner_path_rows <= 0)
3370 678 : inner_path_rows = 1;
3371 : /* Mark the path with the correct row estimate */
3372 1389978 : if (path->jpath.path.param_info)
3373 26740 : path->jpath.path.rows = path->jpath.path.param_info->ppi_rows;
3374 : else
3375 1363238 : path->jpath.path.rows = path->jpath.path.parent->rows;
3376 :
3377 : /* For partial paths, scale row estimate. */
3378 1389978 : if (path->jpath.path.parallel_workers > 0)
3379 : {
3380 12648 : double parallel_divisor = get_parallel_divisor(&path->jpath.path);
3381 :
3382 12648 : path->jpath.path.rows =
3383 12648 : clamp_row_est(path->jpath.path.rows / parallel_divisor);
3384 : }
3385 :
3386 : /* cost of inner-relation source data (we already dealt with outer rel) */
3387 :
3388 1389978 : if (path->jpath.jointype == JOIN_SEMI || path->jpath.jointype == JOIN_ANTI ||
3389 1349660 : extra->inner_unique)
3390 910470 : {
3391 : /*
3392 : * With a SEMI or ANTI join, or if the innerrel is known unique, the
3393 : * executor will stop after the first match.
3394 : */
3395 910470 : Cost inner_run_cost = workspace->inner_run_cost;
3396 910470 : Cost inner_rescan_run_cost = workspace->inner_rescan_run_cost;
3397 : double outer_matched_rows;
3398 : double outer_unmatched_rows;
3399 : Selectivity inner_scan_frac;
3400 :
3401 : /*
3402 : * For an outer-rel row that has at least one match, we can expect the
3403 : * inner scan to stop after a fraction 1/(match_count+1) of the inner
3404 : * rows, if the matches are evenly distributed. Since they probably
3405 : * aren't quite evenly distributed, we apply a fuzz factor of 2.0 to
3406 : * that fraction. (If we used a larger fuzz factor, we'd have to
3407 : * clamp inner_scan_frac to at most 1.0; but since match_count is at
3408 : * least 1, no such clamp is needed now.)
3409 : */
3410 910470 : outer_matched_rows = rint(outer_path_rows * extra->semifactors.outer_match_frac);
3411 910470 : outer_unmatched_rows = outer_path_rows - outer_matched_rows;
3412 910470 : inner_scan_frac = 2.0 / (extra->semifactors.match_count + 1.0);
3413 :
3414 : /*
3415 : * Compute number of tuples processed (not number emitted!). First,
3416 : * account for successfully-matched outer rows.
3417 : */
3418 910470 : ntuples = outer_matched_rows * inner_path_rows * inner_scan_frac;
3419 :
3420 : /*
3421 : * Now we need to estimate the actual costs of scanning the inner
3422 : * relation, which may be quite a bit less than N times inner_run_cost
3423 : * due to early scan stops. We consider two cases. If the inner path
3424 : * is an indexscan using all the joinquals as indexquals, then an
3425 : * unmatched outer row results in an indexscan returning no rows,
3426 : * which is probably quite cheap. Otherwise, the executor will have
3427 : * to scan the whole inner rel for an unmatched row; not so cheap.
3428 : */
3429 910470 : if (has_indexed_join_quals(path))
3430 : {
3431 : /*
3432 : * Successfully-matched outer rows will only require scanning
3433 : * inner_scan_frac of the inner relation. In this case, we don't
3434 : * need to charge the full inner_run_cost even when that's more
3435 : * than inner_rescan_run_cost, because we can assume that none of
3436 : * the inner scans ever scan the whole inner relation. So it's
3437 : * okay to assume that all the inner scan executions can be
3438 : * fractions of the full cost, even if materialization is reducing
3439 : * the rescan cost. At this writing, it's impossible to get here
3440 : * for a materialized inner scan, so inner_run_cost and
3441 : * inner_rescan_run_cost will be the same anyway; but just in
3442 : * case, use inner_run_cost for the first matched tuple and
3443 : * inner_rescan_run_cost for additional ones.
3444 : */
3445 149366 : run_cost += inner_run_cost * inner_scan_frac;
3446 149366 : if (outer_matched_rows > 1)
3447 21166 : run_cost += (outer_matched_rows - 1) * inner_rescan_run_cost * inner_scan_frac;
3448 :
3449 : /*
3450 : * Add the cost of inner-scan executions for unmatched outer rows.
3451 : * We estimate this as the same cost as returning the first tuple
3452 : * of a nonempty scan. We consider that these are all rescans,
3453 : * since we used inner_run_cost once already.
3454 : */
3455 149366 : run_cost += outer_unmatched_rows *
3456 149366 : inner_rescan_run_cost / inner_path_rows;
3457 :
3458 : /*
3459 : * We won't be evaluating any quals at all for unmatched rows, so
3460 : * don't add them to ntuples.
3461 : */
3462 : }
3463 : else
3464 : {
3465 : /*
3466 : * Here, a complicating factor is that rescans may be cheaper than
3467 : * first scans. If we never scan all the way to the end of the
3468 : * inner rel, it might be (depending on the plan type) that we'd
3469 : * never pay the whole inner first-scan run cost. However it is
3470 : * difficult to estimate whether that will happen (and it could
3471 : * not happen if there are any unmatched outer rows!), so be
3472 : * conservative and always charge the whole first-scan cost once.
3473 : * We consider this charge to correspond to the first unmatched
3474 : * outer row, unless there isn't one in our estimate, in which
3475 : * case blame it on the first matched row.
3476 : */
3477 :
3478 : /* First, count all unmatched join tuples as being processed */
3479 761104 : ntuples += outer_unmatched_rows * inner_path_rows;
3480 :
3481 : /* Now add the forced full scan, and decrement appropriate count */
3482 761104 : run_cost += inner_run_cost;
3483 761104 : if (outer_unmatched_rows >= 1)
3484 731382 : outer_unmatched_rows -= 1;
3485 : else
3486 29722 : outer_matched_rows -= 1;
3487 :
3488 : /* Add inner run cost for additional outer tuples having matches */
3489 761104 : if (outer_matched_rows > 0)
3490 266156 : run_cost += outer_matched_rows * inner_rescan_run_cost * inner_scan_frac;
3491 :
3492 : /* Add inner run cost for additional unmatched outer tuples */
3493 761104 : if (outer_unmatched_rows > 0)
3494 504814 : run_cost += outer_unmatched_rows * inner_rescan_run_cost;
3495 : }
3496 : }
3497 : else
3498 : {
3499 : /* Normal-case source costs were included in preliminary estimate */
3500 :
3501 : /* Compute number of tuples processed (not number emitted!) */
3502 479508 : ntuples = outer_path_rows * inner_path_rows;
3503 : }
3504 :
3505 : /* CPU costs */
3506 1389978 : cost_qual_eval(&restrict_qual_cost, path->jpath.joinrestrictinfo, root);
3507 1389978 : startup_cost += restrict_qual_cost.startup;
3508 1389978 : cpu_per_tuple = cpu_tuple_cost + restrict_qual_cost.per_tuple;
3509 1389978 : run_cost += cpu_per_tuple * ntuples;
3510 :
3511 : /* tlist eval costs are paid per output row, not per tuple scanned */
3512 1389978 : startup_cost += path->jpath.path.pathtarget->cost.startup;
3513 1389978 : run_cost += path->jpath.path.pathtarget->cost.per_tuple * path->jpath.path.rows;
3514 :
3515 1389978 : path->jpath.path.startup_cost = startup_cost;
3516 1389978 : path->jpath.path.total_cost = startup_cost + run_cost;
3517 1389978 : }
3518 :
3519 : /*
3520 : * initial_cost_mergejoin
3521 : * Preliminary estimate of the cost of a mergejoin path.
3522 : *
3523 : * This must quickly produce lower-bound estimates of the path's startup and
3524 : * total costs. If we are unable to eliminate the proposed path from
3525 : * consideration using the lower bounds, final_cost_mergejoin will be called
3526 : * to obtain the final estimates.
3527 : *
3528 : * The exact division of labor between this function and final_cost_mergejoin
3529 : * is private to them, and represents a tradeoff between speed of the initial
3530 : * estimate and getting a tight lower bound. We choose to not examine the
3531 : * join quals here, except for obtaining the scan selectivity estimate which
3532 : * is really essential (but fortunately, use of caching keeps the cost of
3533 : * getting that down to something reasonable).
3534 : * We also assume that cost_sort/cost_incremental_sort is cheap enough to use
3535 : * here.
3536 : *
3537 : * 'workspace' is to be filled with startup_cost, total_cost, and perhaps
3538 : * other data to be used by final_cost_mergejoin
3539 : * 'jointype' is the type of join to be performed
3540 : * 'mergeclauses' is the list of joinclauses to be used as merge clauses
3541 : * 'outer_path' is the outer input to the join
3542 : * 'inner_path' is the inner input to the join
3543 : * 'outersortkeys' is the list of sort keys for the outer path
3544 : * 'innersortkeys' is the list of sort keys for the inner path
3545 : * 'extra' contains miscellaneous information about the join
3546 : *
3547 : * Note: outersortkeys and innersortkeys should be NIL if no explicit
3548 : * sort is needed because the respective source path is already ordered.
3549 : */
3550 : void
3551 1270732 : initial_cost_mergejoin(PlannerInfo *root, JoinCostWorkspace *workspace,
3552 : JoinType jointype,
3553 : List *mergeclauses,
3554 : Path *outer_path, Path *inner_path,
3555 : List *outersortkeys, List *innersortkeys,
3556 : JoinPathExtraData *extra)
3557 : {
3558 : int disabled_nodes;
3559 1270732 : Cost startup_cost = 0;
3560 1270732 : Cost run_cost = 0;
3561 1270732 : double outer_path_rows = outer_path->rows;
3562 1270732 : double inner_path_rows = inner_path->rows;
3563 : Cost inner_run_cost;
3564 : double outer_rows,
3565 : inner_rows,
3566 : outer_skip_rows,
3567 : inner_skip_rows;
3568 : Selectivity outerstartsel,
3569 : outerendsel,
3570 : innerstartsel,
3571 : innerendsel;
3572 : Path sort_path; /* dummy for result of
3573 : * cost_sort/cost_incremental_sort */
3574 :
3575 : /* Protect some assumptions below that rowcounts aren't zero */
3576 1270732 : if (outer_path_rows <= 0)
3577 96 : outer_path_rows = 1;
3578 1270732 : if (inner_path_rows <= 0)
3579 126 : inner_path_rows = 1;
3580 :
3581 : /*
3582 : * A merge join will stop as soon as it exhausts either input stream
3583 : * (unless it's an outer join, in which case the outer side has to be
3584 : * scanned all the way anyway). Estimate fraction of the left and right
3585 : * inputs that will actually need to be scanned. Likewise, we can
3586 : * estimate the number of rows that will be skipped before the first join
3587 : * pair is found, which should be factored into startup cost. We use only
3588 : * the first (most significant) merge clause for this purpose. Since
3589 : * mergejoinscansel() is a fairly expensive computation, we cache the
3590 : * results in the merge clause RestrictInfo.
3591 : */
3592 1270732 : if (mergeclauses && jointype != JOIN_FULL)
3593 1264592 : {
3594 1264592 : RestrictInfo *firstclause = (RestrictInfo *) linitial(mergeclauses);
3595 : List *opathkeys;
3596 : List *ipathkeys;
3597 : PathKey *opathkey;
3598 : PathKey *ipathkey;
3599 : MergeScanSelCache *cache;
3600 :
3601 : /* Get the input pathkeys to determine the sort-order details */
3602 1264592 : opathkeys = outersortkeys ? outersortkeys : outer_path->pathkeys;
3603 1264592 : ipathkeys = innersortkeys ? innersortkeys : inner_path->pathkeys;
3604 : Assert(opathkeys);
3605 : Assert(ipathkeys);
3606 1264592 : opathkey = (PathKey *) linitial(opathkeys);
3607 1264592 : ipathkey = (PathKey *) linitial(ipathkeys);
3608 : /* debugging check */
3609 1264592 : if (opathkey->pk_opfamily != ipathkey->pk_opfamily ||
3610 1264592 : opathkey->pk_eclass->ec_collation != ipathkey->pk_eclass->ec_collation ||
3611 1264592 : opathkey->pk_cmptype != ipathkey->pk_cmptype ||
3612 1264592 : opathkey->pk_nulls_first != ipathkey->pk_nulls_first)
3613 0 : elog(ERROR, "left and right pathkeys do not match in mergejoin");
3614 :
3615 : /* Get the selectivity with caching */
3616 1264592 : cache = cached_scansel(root, firstclause, opathkey);
3617 :
3618 1264592 : if (bms_is_subset(firstclause->left_relids,
3619 1264592 : outer_path->parent->relids))
3620 : {
3621 : /* left side of clause is outer */
3622 674802 : outerstartsel = cache->leftstartsel;
3623 674802 : outerendsel = cache->leftendsel;
3624 674802 : innerstartsel = cache->rightstartsel;
3625 674802 : innerendsel = cache->rightendsel;
3626 : }
3627 : else
3628 : {
3629 : /* left side of clause is inner */
3630 589790 : outerstartsel = cache->rightstartsel;
3631 589790 : outerendsel = cache->rightendsel;
3632 589790 : innerstartsel = cache->leftstartsel;
3633 589790 : innerendsel = cache->leftendsel;
3634 : }
3635 1264592 : if (jointype == JOIN_LEFT ||
3636 : jointype == JOIN_ANTI)
3637 : {
3638 209408 : outerstartsel = 0.0;
3639 209408 : outerendsel = 1.0;
3640 : }
3641 1055184 : else if (jointype == JOIN_RIGHT ||
3642 : jointype == JOIN_RIGHT_ANTI)
3643 : {
3644 204200 : innerstartsel = 0.0;
3645 204200 : innerendsel = 1.0;
3646 : }
3647 : }
3648 : else
3649 : {
3650 : /* cope with clauseless or full mergejoin */
3651 6140 : outerstartsel = innerstartsel = 0.0;
3652 6140 : outerendsel = innerendsel = 1.0;
3653 : }
3654 :
3655 : /*
3656 : * Convert selectivities to row counts. We force outer_rows and
3657 : * inner_rows to be at least 1, but the skip_rows estimates can be zero.
3658 : */
3659 1270732 : outer_skip_rows = rint(outer_path_rows * outerstartsel);
3660 1270732 : inner_skip_rows = rint(inner_path_rows * innerstartsel);
3661 1270732 : outer_rows = clamp_row_est(outer_path_rows * outerendsel);
3662 1270732 : inner_rows = clamp_row_est(inner_path_rows * innerendsel);
3663 :
3664 : Assert(outer_skip_rows <= outer_rows);
3665 : Assert(inner_skip_rows <= inner_rows);
3666 :
3667 : /*
3668 : * Readjust scan selectivities to account for above rounding. This is
3669 : * normally an insignificant effect, but when there are only a few rows in
3670 : * the inputs, failing to do this makes for a large percentage error.
3671 : */
3672 1270732 : outerstartsel = outer_skip_rows / outer_path_rows;
3673 1270732 : innerstartsel = inner_skip_rows / inner_path_rows;
3674 1270732 : outerendsel = outer_rows / outer_path_rows;
3675 1270732 : innerendsel = inner_rows / inner_path_rows;
3676 :
3677 : Assert(outerstartsel <= outerendsel);
3678 : Assert(innerstartsel <= innerendsel);
3679 :
3680 1270732 : disabled_nodes = enable_mergejoin ? 0 : 1;
3681 :
3682 : /* cost of source data */
3683 :
3684 1270732 : if (outersortkeys) /* do we need to sort outer? */
3685 : {
3686 599150 : bool use_incremental_sort = false;
3687 : int presorted_keys;
3688 :
3689 : /*
3690 : * We choose to use incremental sort if it is enabled and there are
3691 : * presorted keys; otherwise we use full sort.
3692 : */
3693 599150 : if (enable_incremental_sort)
3694 : {
3695 : bool is_sorted PG_USED_FOR_ASSERTS_ONLY;
3696 :
3697 598220 : is_sorted = pathkeys_count_contained_in(outersortkeys,
3698 : outer_path->pathkeys,
3699 : &presorted_keys);
3700 : Assert(!is_sorted);
3701 :
3702 598220 : if (presorted_keys > 0)
3703 2082 : use_incremental_sort = true;
3704 : }
3705 :
3706 599150 : if (!use_incremental_sort)
3707 : {
3708 597068 : cost_sort(&sort_path,
3709 : root,
3710 : outersortkeys,
3711 : outer_path->disabled_nodes,
3712 : outer_path->total_cost,
3713 : outer_path_rows,
3714 597068 : outer_path->pathtarget->width,
3715 : 0.0,
3716 : work_mem,
3717 : -1.0);
3718 : }
3719 : else
3720 : {
3721 2082 : cost_incremental_sort(&sort_path,
3722 : root,
3723 : outersortkeys,
3724 : presorted_keys,
3725 : outer_path->disabled_nodes,
3726 : outer_path->startup_cost,
3727 : outer_path->total_cost,
3728 : outer_path_rows,
3729 2082 : outer_path->pathtarget->width,
3730 : 0.0,
3731 : work_mem,
3732 : -1.0);
3733 : }
3734 599150 : disabled_nodes += sort_path.disabled_nodes;
3735 599150 : startup_cost += sort_path.startup_cost;
3736 599150 : startup_cost += (sort_path.total_cost - sort_path.startup_cost)
3737 599150 : * outerstartsel;
3738 599150 : run_cost += (sort_path.total_cost - sort_path.startup_cost)
3739 599150 : * (outerendsel - outerstartsel);
3740 : }
3741 : else
3742 : {
3743 671582 : disabled_nodes += outer_path->disabled_nodes;
3744 671582 : startup_cost += outer_path->startup_cost;
3745 671582 : startup_cost += (outer_path->total_cost - outer_path->startup_cost)
3746 671582 : * outerstartsel;
3747 671582 : run_cost += (outer_path->total_cost - outer_path->startup_cost)
3748 671582 : * (outerendsel - outerstartsel);
3749 : }
3750 :
3751 1270732 : if (innersortkeys) /* do we need to sort inner? */
3752 : {
3753 : /*
3754 : * We do not consider incremental sort for inner path, because
3755 : * incremental sort does not support mark/restore.
3756 : */
3757 :
3758 982898 : cost_sort(&sort_path,
3759 : root,
3760 : innersortkeys,
3761 : inner_path->disabled_nodes,
3762 : inner_path->total_cost,
3763 : inner_path_rows,
3764 982898 : inner_path->pathtarget->width,
3765 : 0.0,
3766 : work_mem,
3767 : -1.0);
3768 982898 : disabled_nodes += sort_path.disabled_nodes;
3769 982898 : startup_cost += sort_path.startup_cost;
3770 982898 : startup_cost += (sort_path.total_cost - sort_path.startup_cost)
3771 982898 : * innerstartsel;
3772 982898 : inner_run_cost = (sort_path.total_cost - sort_path.startup_cost)
3773 982898 : * (innerendsel - innerstartsel);
3774 : }
3775 : else
3776 : {
3777 287834 : disabled_nodes += inner_path->disabled_nodes;
3778 287834 : startup_cost += inner_path->startup_cost;
3779 287834 : startup_cost += (inner_path->total_cost - inner_path->startup_cost)
3780 287834 : * innerstartsel;
3781 287834 : inner_run_cost = (inner_path->total_cost - inner_path->startup_cost)
3782 287834 : * (innerendsel - innerstartsel);
3783 : }
3784 :
3785 : /*
3786 : * We can't yet determine whether rescanning occurs, or whether
3787 : * materialization of the inner input should be done. The minimum
3788 : * possible inner input cost, regardless of rescan and materialization
3789 : * considerations, is inner_run_cost. We include that in
3790 : * workspace->total_cost, but not yet in run_cost.
3791 : */
3792 :
3793 : /* CPU costs left for later */
3794 :
3795 : /* Public result fields */
3796 1270732 : workspace->disabled_nodes = disabled_nodes;
3797 1270732 : workspace->startup_cost = startup_cost;
3798 1270732 : workspace->total_cost = startup_cost + run_cost + inner_run_cost;
3799 : /* Save private data for final_cost_mergejoin */
3800 1270732 : workspace->run_cost = run_cost;
3801 1270732 : workspace->inner_run_cost = inner_run_cost;
3802 1270732 : workspace->outer_rows = outer_rows;
3803 1270732 : workspace->inner_rows = inner_rows;
3804 1270732 : workspace->outer_skip_rows = outer_skip_rows;
3805 1270732 : workspace->inner_skip_rows = inner_skip_rows;
3806 1270732 : }
3807 :
3808 : /*
3809 : * final_cost_mergejoin
3810 : * Final estimate of the cost and result size of a mergejoin path.
3811 : *
3812 : * Unlike other costsize functions, this routine makes two actual decisions:
3813 : * whether the executor will need to do mark/restore, and whether we should
3814 : * materialize the inner path. It would be logically cleaner to build
3815 : * separate paths testing these alternatives, but that would require repeating
3816 : * most of the cost calculations, which are not all that cheap. Since the
3817 : * choice will not affect output pathkeys or startup cost, only total cost,
3818 : * there is no possibility of wanting to keep more than one path. So it seems
3819 : * best to make the decisions here and record them in the path's
3820 : * skip_mark_restore and materialize_inner fields.
3821 : *
3822 : * Mark/restore overhead is usually required, but can be skipped if we know
3823 : * that the executor need find only one match per outer tuple, and that the
3824 : * mergeclauses are sufficient to identify a match.
3825 : *
3826 : * We materialize the inner path if we need mark/restore and either the inner
3827 : * path can't support mark/restore, or it's cheaper to use an interposed
3828 : * Material node to handle mark/restore.
3829 : *
3830 : * 'path' is already filled in except for the rows and cost fields and
3831 : * skip_mark_restore and materialize_inner
3832 : * 'workspace' is the result from initial_cost_mergejoin
3833 : * 'extra' contains miscellaneous information about the join
3834 : */
3835 : void
3836 328052 : final_cost_mergejoin(PlannerInfo *root, MergePath *path,
3837 : JoinCostWorkspace *workspace,
3838 : JoinPathExtraData *extra)
3839 : {
3840 328052 : Path *outer_path = path->jpath.outerjoinpath;
3841 328052 : Path *inner_path = path->jpath.innerjoinpath;
3842 328052 : double inner_path_rows = inner_path->rows;
3843 328052 : List *mergeclauses = path->path_mergeclauses;
3844 328052 : List *innersortkeys = path->innersortkeys;
3845 328052 : Cost startup_cost = workspace->startup_cost;
3846 328052 : Cost run_cost = workspace->run_cost;
3847 328052 : Cost inner_run_cost = workspace->inner_run_cost;
3848 328052 : double outer_rows = workspace->outer_rows;
3849 328052 : double inner_rows = workspace->inner_rows;
3850 328052 : double outer_skip_rows = workspace->outer_skip_rows;
3851 328052 : double inner_skip_rows = workspace->inner_skip_rows;
3852 : Cost cpu_per_tuple,
3853 : bare_inner_cost,
3854 : mat_inner_cost;
3855 : QualCost merge_qual_cost;
3856 : QualCost qp_qual_cost;
3857 : double mergejointuples,
3858 : rescannedtuples;
3859 : double rescanratio;
3860 :
3861 : /* Set the number of disabled nodes. */
3862 328052 : path->jpath.path.disabled_nodes = workspace->disabled_nodes;
3863 :
3864 : /* Protect some assumptions below that rowcounts aren't zero */
3865 328052 : if (inner_path_rows <= 0)
3866 90 : inner_path_rows = 1;
3867 :
3868 : /* Mark the path with the correct row estimate */
3869 328052 : if (path->jpath.path.param_info)
3870 776 : path->jpath.path.rows = path->jpath.path.param_info->ppi_rows;
3871 : else
3872 327276 : path->jpath.path.rows = path->jpath.path.parent->rows;
3873 :
3874 : /* For partial paths, scale row estimate. */
3875 328052 : if (path->jpath.path.parallel_workers > 0)
3876 : {
3877 9432 : double parallel_divisor = get_parallel_divisor(&path->jpath.path);
3878 :
3879 9432 : path->jpath.path.rows =
3880 9432 : clamp_row_est(path->jpath.path.rows / parallel_divisor);
3881 : }
3882 :
3883 : /*
3884 : * Compute cost of the mergequals and qpquals (other restriction clauses)
3885 : * separately.
3886 : */
3887 328052 : cost_qual_eval(&merge_qual_cost, mergeclauses, root);
3888 328052 : cost_qual_eval(&qp_qual_cost, path->jpath.joinrestrictinfo, root);
3889 328052 : qp_qual_cost.startup -= merge_qual_cost.startup;
3890 328052 : qp_qual_cost.per_tuple -= merge_qual_cost.per_tuple;
3891 :
3892 : /*
3893 : * With a SEMI or ANTI join, or if the innerrel is known unique, the
3894 : * executor will stop scanning for matches after the first match. When
3895 : * all the joinclauses are merge clauses, this means we don't ever need to
3896 : * back up the merge, and so we can skip mark/restore overhead.
3897 : */
3898 328052 : if ((path->jpath.jointype == JOIN_SEMI ||
3899 321240 : path->jpath.jointype == JOIN_ANTI ||
3900 464940 : extra->inner_unique) &&
3901 151100 : (list_length(path->jpath.joinrestrictinfo) ==
3902 151100 : list_length(path->path_mergeclauses)))
3903 126338 : path->skip_mark_restore = true;
3904 : else
3905 201714 : path->skip_mark_restore = false;
3906 :
3907 : /*
3908 : * Get approx # tuples passing the mergequals. We use approx_tuple_count
3909 : * here because we need an estimate done with JOIN_INNER semantics.
3910 : */
3911 328052 : mergejointuples = approx_tuple_count(root, &path->jpath, mergeclauses);
3912 :
3913 : /*
3914 : * When there are equal merge keys in the outer relation, the mergejoin
3915 : * must rescan any matching tuples in the inner relation. This means
3916 : * re-fetching inner tuples; we have to estimate how often that happens.
3917 : *
3918 : * For regular inner and outer joins, the number of re-fetches can be
3919 : * estimated approximately as size of merge join output minus size of
3920 : * inner relation. Assume that the distinct key values are 1, 2, ..., and
3921 : * denote the number of values of each key in the outer relation as m1,
3922 : * m2, ...; in the inner relation, n1, n2, ... Then we have
3923 : *
3924 : * size of join = m1 * n1 + m2 * n2 + ...
3925 : *
3926 : * number of rescanned tuples = (m1 - 1) * n1 + (m2 - 1) * n2 + ... = m1 *
3927 : * n1 + m2 * n2 + ... - (n1 + n2 + ...) = size of join - size of inner
3928 : * relation
3929 : *
3930 : * This equation works correctly for outer tuples having no inner match
3931 : * (nk = 0), but not for inner tuples having no outer match (mk = 0); we
3932 : * are effectively subtracting those from the number of rescanned tuples,
3933 : * when we should not. Can we do better without expensive selectivity
3934 : * computations?
3935 : *
3936 : * The whole issue is moot if we are working from a unique-ified outer
3937 : * input, or if we know we don't need to mark/restore at all.
3938 : */
3939 328052 : if (IsA(outer_path, UniquePath) || path->skip_mark_restore)
3940 129078 : rescannedtuples = 0;
3941 : else
3942 : {
3943 198974 : rescannedtuples = mergejointuples - inner_path_rows;
3944 : /* Must clamp because of possible underestimate */
3945 198974 : if (rescannedtuples < 0)
3946 79718 : rescannedtuples = 0;
3947 : }
3948 :
3949 : /*
3950 : * We'll inflate various costs this much to account for rescanning. Note
3951 : * that this is to be multiplied by something involving inner_rows, or
3952 : * another number related to the portion of the inner rel we'll scan.
3953 : */
3954 328052 : rescanratio = 1.0 + (rescannedtuples / inner_rows);
3955 :
3956 : /*
3957 : * Decide whether we want to materialize the inner input to shield it from
3958 : * mark/restore and performing re-fetches. Our cost model for regular
3959 : * re-fetches is that a re-fetch costs the same as an original fetch,
3960 : * which is probably an overestimate; but on the other hand we ignore the
3961 : * bookkeeping costs of mark/restore. Not clear if it's worth developing
3962 : * a more refined model. So we just need to inflate the inner run cost by
3963 : * rescanratio.
3964 : */
3965 328052 : bare_inner_cost = inner_run_cost * rescanratio;
3966 :
3967 : /*
3968 : * When we interpose a Material node the re-fetch cost is assumed to be
3969 : * just cpu_operator_cost per tuple, independently of the underlying
3970 : * plan's cost; and we charge an extra cpu_operator_cost per original
3971 : * fetch as well. Note that we're assuming the materialize node will
3972 : * never spill to disk, since it only has to remember tuples back to the
3973 : * last mark. (If there are a huge number of duplicates, our other cost
3974 : * factors will make the path so expensive that it probably won't get
3975 : * chosen anyway.) So we don't use cost_rescan here.
3976 : *
3977 : * Note: keep this estimate in sync with create_mergejoin_plan's labeling
3978 : * of the generated Material node.
3979 : */
3980 328052 : mat_inner_cost = inner_run_cost +
3981 328052 : cpu_operator_cost * inner_rows * rescanratio;
3982 :
3983 : /*
3984 : * If we don't need mark/restore at all, we don't need materialization.
3985 : */
3986 328052 : if (path->skip_mark_restore)
3987 126338 : path->materialize_inner = false;
3988 :
3989 : /*
3990 : * Prefer materializing if it looks cheaper, unless the user has asked to
3991 : * suppress materialization.
3992 : */
3993 201714 : else if (enable_material && mat_inner_cost < bare_inner_cost)
3994 2670 : path->materialize_inner = true;
3995 :
3996 : /*
3997 : * Even if materializing doesn't look cheaper, we *must* do it if the
3998 : * inner path is to be used directly (without sorting) and it doesn't
3999 : * support mark/restore.
4000 : *
4001 : * Since the inner side must be ordered, and only Sorts and IndexScans can
4002 : * create order to begin with, and they both support mark/restore, you
4003 : * might think there's no problem --- but you'd be wrong. Nestloop and
4004 : * merge joins can *preserve* the order of their inputs, so they can be
4005 : * selected as the input of a mergejoin, and they don't support
4006 : * mark/restore at present.
4007 : *
4008 : * We don't test the value of enable_material here, because
4009 : * materialization is required for correctness in this case, and turning
4010 : * it off does not entitle us to deliver an invalid plan.
4011 : */
4012 199044 : else if (innersortkeys == NIL &&
4013 12330 : !ExecSupportsMarkRestore(inner_path))
4014 1534 : path->materialize_inner = true;
4015 :
4016 : /*
4017 : * Also, force materializing if the inner path is to be sorted and the
4018 : * sort is expected to spill to disk. This is because the final merge
4019 : * pass can be done on-the-fly if it doesn't have to support mark/restore.
4020 : * We don't try to adjust the cost estimates for this consideration,
4021 : * though.
4022 : *
4023 : * Since materialization is a performance optimization in this case,
4024 : * rather than necessary for correctness, we skip it if enable_material is
4025 : * off.
4026 : */
4027 197510 : else if (enable_material && innersortkeys != NIL &&
4028 186666 : relation_byte_size(inner_path_rows,
4029 186666 : inner_path->pathtarget->width) >
4030 186666 : work_mem * (Size) 1024)
4031 256 : path->materialize_inner = true;
4032 : else
4033 197254 : path->materialize_inner = false;
4034 :
4035 : /* Charge the right incremental cost for the chosen case */
4036 328052 : if (path->materialize_inner)
4037 4460 : run_cost += mat_inner_cost;
4038 : else
4039 323592 : run_cost += bare_inner_cost;
4040 :
4041 : /* CPU costs */
4042 :
4043 : /*
4044 : * The number of tuple comparisons needed is approximately number of outer
4045 : * rows plus number of inner rows plus number of rescanned tuples (can we
4046 : * refine this?). At each one, we need to evaluate the mergejoin quals.
4047 : */
4048 328052 : startup_cost += merge_qual_cost.startup;
4049 328052 : startup_cost += merge_qual_cost.per_tuple *
4050 328052 : (outer_skip_rows + inner_skip_rows * rescanratio);
4051 328052 : run_cost += merge_qual_cost.per_tuple *
4052 328052 : ((outer_rows - outer_skip_rows) +
4053 328052 : (inner_rows - inner_skip_rows) * rescanratio);
4054 :
4055 : /*
4056 : * For each tuple that gets through the mergejoin proper, we charge
4057 : * cpu_tuple_cost plus the cost of evaluating additional restriction
4058 : * clauses that are to be applied at the join. (This is pessimistic since
4059 : * not all of the quals may get evaluated at each tuple.)
4060 : *
4061 : * Note: we could adjust for SEMI/ANTI joins skipping some qual
4062 : * evaluations here, but it's probably not worth the trouble.
4063 : */
4064 328052 : startup_cost += qp_qual_cost.startup;
4065 328052 : cpu_per_tuple = cpu_tuple_cost + qp_qual_cost.per_tuple;
4066 328052 : run_cost += cpu_per_tuple * mergejointuples;
4067 :
4068 : /* tlist eval costs are paid per output row, not per tuple scanned */
4069 328052 : startup_cost += path->jpath.path.pathtarget->cost.startup;
4070 328052 : run_cost += path->jpath.path.pathtarget->cost.per_tuple * path->jpath.path.rows;
4071 :
4072 328052 : path->jpath.path.startup_cost = startup_cost;
4073 328052 : path->jpath.path.total_cost = startup_cost + run_cost;
4074 328052 : }
4075 :
4076 : /*
4077 : * run mergejoinscansel() with caching
4078 : */
4079 : static MergeScanSelCache *
4080 1264592 : cached_scansel(PlannerInfo *root, RestrictInfo *rinfo, PathKey *pathkey)
4081 : {
4082 : MergeScanSelCache *cache;
4083 : ListCell *lc;
4084 : Selectivity leftstartsel,
4085 : leftendsel,
4086 : rightstartsel,
4087 : rightendsel;
4088 : MemoryContext oldcontext;
4089 :
4090 : /* Do we have this result already? */
4091 1264634 : foreach(lc, rinfo->scansel_cache)
4092 : {
4093 1146406 : cache = (MergeScanSelCache *) lfirst(lc);
4094 1146406 : if (cache->opfamily == pathkey->pk_opfamily &&
4095 1146406 : cache->collation == pathkey->pk_eclass->ec_collation &&
4096 1146406 : cache->cmptype == pathkey->pk_cmptype &&
4097 1146364 : cache->nulls_first == pathkey->pk_nulls_first)
4098 1146364 : return cache;
4099 : }
4100 :
4101 : /* Nope, do the computation */
4102 118228 : mergejoinscansel(root,
4103 118228 : (Node *) rinfo->clause,
4104 : pathkey->pk_opfamily,
4105 : pathkey->pk_cmptype,
4106 118228 : pathkey->pk_nulls_first,
4107 : &leftstartsel,
4108 : &leftendsel,
4109 : &rightstartsel,
4110 : &rightendsel);
4111 :
4112 : /* Cache the result in suitably long-lived workspace */
4113 118228 : oldcontext = MemoryContextSwitchTo(root->planner_cxt);
4114 :
4115 118228 : cache = (MergeScanSelCache *) palloc(sizeof(MergeScanSelCache));
4116 118228 : cache->opfamily = pathkey->pk_opfamily;
4117 118228 : cache->collation = pathkey->pk_eclass->ec_collation;
4118 118228 : cache->cmptype = pathkey->pk_cmptype;
4119 118228 : cache->nulls_first = pathkey->pk_nulls_first;
4120 118228 : cache->leftstartsel = leftstartsel;
4121 118228 : cache->leftendsel = leftendsel;
4122 118228 : cache->rightstartsel = rightstartsel;
4123 118228 : cache->rightendsel = rightendsel;
4124 :
4125 118228 : rinfo->scansel_cache = lappend(rinfo->scansel_cache, cache);
4126 :
4127 118228 : MemoryContextSwitchTo(oldcontext);
4128 :
4129 118228 : return cache;
4130 : }
4131 :
4132 : /*
4133 : * initial_cost_hashjoin
4134 : * Preliminary estimate of the cost of a hashjoin path.
4135 : *
4136 : * This must quickly produce lower-bound estimates of the path's startup and
4137 : * total costs. If we are unable to eliminate the proposed path from
4138 : * consideration using the lower bounds, final_cost_hashjoin will be called
4139 : * to obtain the final estimates.
4140 : *
4141 : * The exact division of labor between this function and final_cost_hashjoin
4142 : * is private to them, and represents a tradeoff between speed of the initial
4143 : * estimate and getting a tight lower bound. We choose to not examine the
4144 : * join quals here (other than by counting the number of hash clauses),
4145 : * so we can't do much with CPU costs. We do assume that
4146 : * ExecChooseHashTableSize is cheap enough to use here.
4147 : *
4148 : * 'workspace' is to be filled with startup_cost, total_cost, and perhaps
4149 : * other data to be used by final_cost_hashjoin
4150 : * 'jointype' is the type of join to be performed
4151 : * 'hashclauses' is the list of joinclauses to be used as hash clauses
4152 : * 'outer_path' is the outer input to the join
4153 : * 'inner_path' is the inner input to the join
4154 : * 'extra' contains miscellaneous information about the join
4155 : * 'parallel_hash' indicates that inner_path is partial and that a shared
4156 : * hash table will be built in parallel
4157 : */
4158 : void
4159 683140 : initial_cost_hashjoin(PlannerInfo *root, JoinCostWorkspace *workspace,
4160 : JoinType jointype,
4161 : List *hashclauses,
4162 : Path *outer_path, Path *inner_path,
4163 : JoinPathExtraData *extra,
4164 : bool parallel_hash)
4165 : {
4166 : int disabled_nodes;
4167 683140 : Cost startup_cost = 0;
4168 683140 : Cost run_cost = 0;
4169 683140 : double outer_path_rows = outer_path->rows;
4170 683140 : double inner_path_rows = inner_path->rows;
4171 683140 : double inner_path_rows_total = inner_path_rows;
4172 683140 : int num_hashclauses = list_length(hashclauses);
4173 : int numbuckets;
4174 : int numbatches;
4175 : int num_skew_mcvs;
4176 : size_t space_allowed; /* unused */
4177 :
4178 : /* Count up disabled nodes. */
4179 683140 : disabled_nodes = enable_hashjoin ? 0 : 1;
4180 683140 : disabled_nodes += inner_path->disabled_nodes;
4181 683140 : disabled_nodes += outer_path->disabled_nodes;
4182 :
4183 : /* cost of source data */
4184 683140 : startup_cost += outer_path->startup_cost;
4185 683140 : run_cost += outer_path->total_cost - outer_path->startup_cost;
4186 683140 : startup_cost += inner_path->total_cost;
4187 :
4188 : /*
4189 : * Cost of computing hash function: must do it once per input tuple. We
4190 : * charge one cpu_operator_cost for each column's hash function. Also,
4191 : * tack on one cpu_tuple_cost per inner row, to model the costs of
4192 : * inserting the row into the hashtable.
4193 : *
4194 : * XXX when a hashclause is more complex than a single operator, we really
4195 : * should charge the extra eval costs of the left or right side, as
4196 : * appropriate, here. This seems more work than it's worth at the moment.
4197 : */
4198 683140 : startup_cost += (cpu_operator_cost * num_hashclauses + cpu_tuple_cost)
4199 683140 : * inner_path_rows;
4200 683140 : run_cost += cpu_operator_cost * num_hashclauses * outer_path_rows;
4201 :
4202 : /*
4203 : * If this is a parallel hash build, then the value we have for
4204 : * inner_rows_total currently refers only to the rows returned by each
4205 : * participant. For shared hash table size estimation, we need the total
4206 : * number, so we need to undo the division.
4207 : */
4208 683140 : if (parallel_hash)
4209 12552 : inner_path_rows_total *= get_parallel_divisor(inner_path);
4210 :
4211 : /*
4212 : * Get hash table size that executor would use for inner relation.
4213 : *
4214 : * XXX for the moment, always assume that skew optimization will be
4215 : * performed. As long as SKEW_HASH_MEM_PERCENT is small, it's not worth
4216 : * trying to determine that for sure.
4217 : *
4218 : * XXX at some point it might be interesting to try to account for skew
4219 : * optimization in the cost estimate, but for now, we don't.
4220 : */
4221 683140 : ExecChooseHashTableSize(inner_path_rows_total,
4222 683140 : inner_path->pathtarget->width,
4223 : true, /* useskew */
4224 : parallel_hash, /* try_combined_hash_mem */
4225 : outer_path->parallel_workers,
4226 : &space_allowed,
4227 : &numbuckets,
4228 : &numbatches,
4229 : &num_skew_mcvs);
4230 :
4231 : /*
4232 : * If inner relation is too big then we will need to "batch" the join,
4233 : * which implies writing and reading most of the tuples to disk an extra
4234 : * time. Charge seq_page_cost per page, since the I/O should be nice and
4235 : * sequential. Writing the inner rel counts as startup cost, all the rest
4236 : * as run cost.
4237 : */
4238 683140 : if (numbatches > 1)
4239 : {
4240 5344 : double outerpages = page_size(outer_path_rows,
4241 5344 : outer_path->pathtarget->width);
4242 5344 : double innerpages = page_size(inner_path_rows,
4243 5344 : inner_path->pathtarget->width);
4244 :
4245 5344 : startup_cost += seq_page_cost * innerpages;
4246 5344 : run_cost += seq_page_cost * (innerpages + 2 * outerpages);
4247 : }
4248 :
4249 : /* CPU costs left for later */
4250 :
4251 : /* Public result fields */
4252 683140 : workspace->disabled_nodes = disabled_nodes;
4253 683140 : workspace->startup_cost = startup_cost;
4254 683140 : workspace->total_cost = startup_cost + run_cost;
4255 : /* Save private data for final_cost_hashjoin */
4256 683140 : workspace->run_cost = run_cost;
4257 683140 : workspace->numbuckets = numbuckets;
4258 683140 : workspace->numbatches = numbatches;
4259 683140 : workspace->inner_rows_total = inner_path_rows_total;
4260 683140 : }
4261 :
4262 : /*
4263 : * final_cost_hashjoin
4264 : * Final estimate of the cost and result size of a hashjoin path.
4265 : *
4266 : * Note: the numbatches estimate is also saved into 'path' for use later
4267 : *
4268 : * 'path' is already filled in except for the rows and cost fields and
4269 : * num_batches
4270 : * 'workspace' is the result from initial_cost_hashjoin
4271 : * 'extra' contains miscellaneous information about the join
4272 : */
4273 : void
4274 294864 : final_cost_hashjoin(PlannerInfo *root, HashPath *path,
4275 : JoinCostWorkspace *workspace,
4276 : JoinPathExtraData *extra)
4277 : {
4278 294864 : Path *outer_path = path->jpath.outerjoinpath;
4279 294864 : Path *inner_path = path->jpath.innerjoinpath;
4280 294864 : double outer_path_rows = outer_path->rows;
4281 294864 : double inner_path_rows = inner_path->rows;
4282 294864 : double inner_path_rows_total = workspace->inner_rows_total;
4283 294864 : List *hashclauses = path->path_hashclauses;
4284 294864 : Cost startup_cost = workspace->startup_cost;
4285 294864 : Cost run_cost = workspace->run_cost;
4286 294864 : int numbuckets = workspace->numbuckets;
4287 294864 : int numbatches = workspace->numbatches;
4288 : Cost cpu_per_tuple;
4289 : QualCost hash_qual_cost;
4290 : QualCost qp_qual_cost;
4291 : double hashjointuples;
4292 : double virtualbuckets;
4293 : Selectivity innerbucketsize;
4294 : Selectivity innermcvfreq;
4295 : ListCell *hcl;
4296 :
4297 : /* Set the number of disabled nodes. */
4298 294864 : path->jpath.path.disabled_nodes = workspace->disabled_nodes;
4299 :
4300 : /* Mark the path with the correct row estimate */
4301 294864 : if (path->jpath.path.param_info)
4302 1422 : path->jpath.path.rows = path->jpath.path.param_info->ppi_rows;
4303 : else
4304 293442 : path->jpath.path.rows = path->jpath.path.parent->rows;
4305 :
4306 : /* For partial paths, scale row estimate. */
4307 294864 : if (path->jpath.path.parallel_workers > 0)
4308 : {
4309 11340 : double parallel_divisor = get_parallel_divisor(&path->jpath.path);
4310 :
4311 11340 : path->jpath.path.rows =
4312 11340 : clamp_row_est(path->jpath.path.rows / parallel_divisor);
4313 : }
4314 :
4315 : /* mark the path with estimated # of batches */
4316 294864 : path->num_batches = numbatches;
4317 :
4318 : /* store the total number of tuples (sum of partial row estimates) */
4319 294864 : path->inner_rows_total = inner_path_rows_total;
4320 :
4321 : /* and compute the number of "virtual" buckets in the whole join */
4322 294864 : virtualbuckets = (double) numbuckets * (double) numbatches;
4323 :
4324 : /*
4325 : * Determine bucketsize fraction and MCV frequency for the inner relation.
4326 : * We use the smallest bucketsize or MCV frequency estimated for any
4327 : * individual hashclause; this is undoubtedly conservative.
4328 : *
4329 : * BUT: if inner relation has been unique-ified, we can assume it's good
4330 : * for hashing. This is important both because it's the right answer, and
4331 : * because we avoid contaminating the cache with a value that's wrong for
4332 : * non-unique-ified paths.
4333 : */
4334 294864 : if (IsA(inner_path, UniquePath))
4335 : {
4336 4716 : innerbucketsize = 1.0 / virtualbuckets;
4337 4716 : innermcvfreq = 0.0;
4338 : }
4339 : else
4340 : {
4341 : List *otherclauses;
4342 :
4343 290148 : innerbucketsize = 1.0;
4344 290148 : innermcvfreq = 1.0;
4345 :
4346 : /* At first, try to estimate bucket size using extended statistics. */
4347 290148 : otherclauses = estimate_multivariate_bucketsize(root,
4348 : inner_path->parent,
4349 : hashclauses,
4350 : &innerbucketsize);
4351 :
4352 : /* Pass through the remaining clauses */
4353 610794 : foreach(hcl, otherclauses)
4354 : {
4355 320646 : RestrictInfo *restrictinfo = lfirst_node(RestrictInfo, hcl);
4356 : Selectivity thisbucketsize;
4357 : Selectivity thismcvfreq;
4358 :
4359 : /*
4360 : * First we have to figure out which side of the hashjoin clause
4361 : * is the inner side.
4362 : *
4363 : * Since we tend to visit the same clauses over and over when
4364 : * planning a large query, we cache the bucket stats estimates in
4365 : * the RestrictInfo node to avoid repeated lookups of statistics.
4366 : */
4367 320646 : if (bms_is_subset(restrictinfo->right_relids,
4368 320646 : inner_path->parent->relids))
4369 : {
4370 : /* righthand side is inner */
4371 167716 : thisbucketsize = restrictinfo->right_bucketsize;
4372 167716 : if (thisbucketsize < 0)
4373 : {
4374 : /* not cached yet */
4375 90438 : estimate_hash_bucket_stats(root,
4376 90438 : get_rightop(restrictinfo->clause),
4377 : virtualbuckets,
4378 : &restrictinfo->right_mcvfreq,
4379 : &restrictinfo->right_bucketsize);
4380 90438 : thisbucketsize = restrictinfo->right_bucketsize;
4381 : }
4382 167716 : thismcvfreq = restrictinfo->right_mcvfreq;
4383 : }
4384 : else
4385 : {
4386 : Assert(bms_is_subset(restrictinfo->left_relids,
4387 : inner_path->parent->relids));
4388 : /* lefthand side is inner */
4389 152930 : thisbucketsize = restrictinfo->left_bucketsize;
4390 152930 : if (thisbucketsize < 0)
4391 : {
4392 : /* not cached yet */
4393 78468 : estimate_hash_bucket_stats(root,
4394 78468 : get_leftop(restrictinfo->clause),
4395 : virtualbuckets,
4396 : &restrictinfo->left_mcvfreq,
4397 : &restrictinfo->left_bucketsize);
4398 78468 : thisbucketsize = restrictinfo->left_bucketsize;
4399 : }
4400 152930 : thismcvfreq = restrictinfo->left_mcvfreq;
4401 : }
4402 :
4403 320646 : if (innerbucketsize > thisbucketsize)
4404 203874 : innerbucketsize = thisbucketsize;
4405 320646 : if (innermcvfreq > thismcvfreq)
4406 290402 : innermcvfreq = thismcvfreq;
4407 : }
4408 : }
4409 :
4410 : /*
4411 : * If the bucket holding the inner MCV would exceed hash_mem, we don't
4412 : * want to hash unless there is really no other alternative, so apply
4413 : * disable_cost. (The executor normally copes with excessive memory usage
4414 : * by splitting batches, but obviously it cannot separate equal values
4415 : * that way, so it will be unable to drive the batch size below hash_mem
4416 : * when this is true.)
4417 : */
4418 294864 : if (relation_byte_size(clamp_row_est(inner_path_rows * innermcvfreq),
4419 589728 : inner_path->pathtarget->width) > get_hash_memory_limit())
4420 6 : startup_cost += disable_cost;
4421 :
4422 : /*
4423 : * Compute cost of the hashquals and qpquals (other restriction clauses)
4424 : * separately.
4425 : */
4426 294864 : cost_qual_eval(&hash_qual_cost, hashclauses, root);
4427 294864 : cost_qual_eval(&qp_qual_cost, path->jpath.joinrestrictinfo, root);
4428 294864 : qp_qual_cost.startup -= hash_qual_cost.startup;
4429 294864 : qp_qual_cost.per_tuple -= hash_qual_cost.per_tuple;
4430 :
4431 : /* CPU costs */
4432 :
4433 294864 : if (path->jpath.jointype == JOIN_SEMI ||
4434 288874 : path->jpath.jointype == JOIN_ANTI ||
4435 283800 : extra->inner_unique)
4436 125584 : {
4437 : double outer_matched_rows;
4438 : Selectivity inner_scan_frac;
4439 :
4440 : /*
4441 : * With a SEMI or ANTI join, or if the innerrel is known unique, the
4442 : * executor will stop after the first match.
4443 : *
4444 : * For an outer-rel row that has at least one match, we can expect the
4445 : * bucket scan to stop after a fraction 1/(match_count+1) of the
4446 : * bucket's rows, if the matches are evenly distributed. Since they
4447 : * probably aren't quite evenly distributed, we apply a fuzz factor of
4448 : * 2.0 to that fraction. (If we used a larger fuzz factor, we'd have
4449 : * to clamp inner_scan_frac to at most 1.0; but since match_count is
4450 : * at least 1, no such clamp is needed now.)
4451 : */
4452 125584 : outer_matched_rows = rint(outer_path_rows * extra->semifactors.outer_match_frac);
4453 125584 : inner_scan_frac = 2.0 / (extra->semifactors.match_count + 1.0);
4454 :
4455 125584 : startup_cost += hash_qual_cost.startup;
4456 251168 : run_cost += hash_qual_cost.per_tuple * outer_matched_rows *
4457 125584 : clamp_row_est(inner_path_rows * innerbucketsize * inner_scan_frac) * 0.5;
4458 :
4459 : /*
4460 : * For unmatched outer-rel rows, the picture is quite a lot different.
4461 : * In the first place, there is no reason to assume that these rows
4462 : * preferentially hit heavily-populated buckets; instead assume they
4463 : * are uncorrelated with the inner distribution and so they see an
4464 : * average bucket size of inner_path_rows / virtualbuckets. In the
4465 : * second place, it seems likely that they will have few if any exact
4466 : * hash-code matches and so very few of the tuples in the bucket will
4467 : * actually require eval of the hash quals. We don't have any good
4468 : * way to estimate how many will, but for the moment assume that the
4469 : * effective cost per bucket entry is one-tenth what it is for
4470 : * matchable tuples.
4471 : */
4472 251168 : run_cost += hash_qual_cost.per_tuple *
4473 251168 : (outer_path_rows - outer_matched_rows) *
4474 125584 : clamp_row_est(inner_path_rows / virtualbuckets) * 0.05;
4475 :
4476 : /* Get # of tuples that will pass the basic join */
4477 125584 : if (path->jpath.jointype == JOIN_ANTI)
4478 5074 : hashjointuples = outer_path_rows - outer_matched_rows;
4479 : else
4480 120510 : hashjointuples = outer_matched_rows;
4481 : }
4482 : else
4483 : {
4484 : /*
4485 : * The number of tuple comparisons needed is the number of outer
4486 : * tuples times the typical number of tuples in a hash bucket, which
4487 : * is the inner relation size times its bucketsize fraction. At each
4488 : * one, we need to evaluate the hashjoin quals. But actually,
4489 : * charging the full qual eval cost at each tuple is pessimistic,
4490 : * since we don't evaluate the quals unless the hash values match
4491 : * exactly. For lack of a better idea, halve the cost estimate to
4492 : * allow for that.
4493 : */
4494 169280 : startup_cost += hash_qual_cost.startup;
4495 338560 : run_cost += hash_qual_cost.per_tuple * outer_path_rows *
4496 169280 : clamp_row_est(inner_path_rows * innerbucketsize) * 0.5;
4497 :
4498 : /*
4499 : * Get approx # tuples passing the hashquals. We use
4500 : * approx_tuple_count here because we need an estimate done with
4501 : * JOIN_INNER semantics.
4502 : */
4503 169280 : hashjointuples = approx_tuple_count(root, &path->jpath, hashclauses);
4504 : }
4505 :
4506 : /*
4507 : * For each tuple that gets through the hashjoin proper, we charge
4508 : * cpu_tuple_cost plus the cost of evaluating additional restriction
4509 : * clauses that are to be applied at the join. (This is pessimistic since
4510 : * not all of the quals may get evaluated at each tuple.)
4511 : */
4512 294864 : startup_cost += qp_qual_cost.startup;
4513 294864 : cpu_per_tuple = cpu_tuple_cost + qp_qual_cost.per_tuple;
4514 294864 : run_cost += cpu_per_tuple * hashjointuples;
4515 :
4516 : /* tlist eval costs are paid per output row, not per tuple scanned */
4517 294864 : startup_cost += path->jpath.path.pathtarget->cost.startup;
4518 294864 : run_cost += path->jpath.path.pathtarget->cost.per_tuple * path->jpath.path.rows;
4519 :
4520 294864 : path->jpath.path.startup_cost = startup_cost;
4521 294864 : path->jpath.path.total_cost = startup_cost + run_cost;
4522 294864 : }
4523 :
4524 :
4525 : /*
4526 : * cost_subplan
4527 : * Figure the costs for a SubPlan (or initplan).
4528 : *
4529 : * Note: we could dig the subplan's Plan out of the root list, but in practice
4530 : * all callers have it handy already, so we make them pass it.
4531 : */
4532 : void
4533 45710 : cost_subplan(PlannerInfo *root, SubPlan *subplan, Plan *plan)
4534 : {
4535 : QualCost sp_cost;
4536 :
4537 : /* Figure any cost for evaluating the testexpr */
4538 45710 : cost_qual_eval(&sp_cost,
4539 45710 : make_ands_implicit((Expr *) subplan->testexpr),
4540 : root);
4541 :
4542 45710 : if (subplan->useHashTable)
4543 : {
4544 : /*
4545 : * If we are using a hash table for the subquery outputs, then the
4546 : * cost of evaluating the query is a one-time cost. We charge one
4547 : * cpu_operator_cost per tuple for the work of loading the hashtable,
4548 : * too.
4549 : */
4550 2174 : sp_cost.startup += plan->total_cost +
4551 2174 : cpu_operator_cost * plan->plan_rows;
4552 :
4553 : /*
4554 : * The per-tuple costs include the cost of evaluating the lefthand
4555 : * expressions, plus the cost of probing the hashtable. We already
4556 : * accounted for the lefthand expressions as part of the testexpr, and
4557 : * will also have counted one cpu_operator_cost for each comparison
4558 : * operator. That is probably too low for the probing cost, but it's
4559 : * hard to make a better estimate, so live with it for now.
4560 : */
4561 : }
4562 : else
4563 : {
4564 : /*
4565 : * Otherwise we will be rescanning the subplan output on each
4566 : * evaluation. We need to estimate how much of the output we will
4567 : * actually need to scan. NOTE: this logic should agree with the
4568 : * tuple_fraction estimates used by make_subplan() in
4569 : * plan/subselect.c.
4570 : */
4571 43536 : Cost plan_run_cost = plan->total_cost - plan->startup_cost;
4572 :
4573 43536 : if (subplan->subLinkType == EXISTS_SUBLINK)
4574 : {
4575 : /* we only need to fetch 1 tuple; clamp to avoid zero divide */
4576 2630 : sp_cost.per_tuple += plan_run_cost / clamp_row_est(plan->plan_rows);
4577 : }
4578 40906 : else if (subplan->subLinkType == ALL_SUBLINK ||
4579 40888 : subplan->subLinkType == ANY_SUBLINK)
4580 : {
4581 : /* assume we need 50% of the tuples */
4582 134 : sp_cost.per_tuple += 0.50 * plan_run_cost;
4583 : /* also charge a cpu_operator_cost per row examined */
4584 134 : sp_cost.per_tuple += 0.50 * plan->plan_rows * cpu_operator_cost;
4585 : }
4586 : else
4587 : {
4588 : /* assume we need all tuples */
4589 40772 : sp_cost.per_tuple += plan_run_cost;
4590 : }
4591 :
4592 : /*
4593 : * Also account for subplan's startup cost. If the subplan is
4594 : * uncorrelated or undirect correlated, AND its topmost node is one
4595 : * that materializes its output, assume that we'll only need to pay
4596 : * its startup cost once; otherwise assume we pay the startup cost
4597 : * every time.
4598 : */
4599 57106 : if (subplan->parParam == NIL &&
4600 13570 : ExecMaterializesOutput(nodeTag(plan)))
4601 690 : sp_cost.startup += plan->startup_cost;
4602 : else
4603 42846 : sp_cost.per_tuple += plan->startup_cost;
4604 : }
4605 :
4606 45710 : subplan->startup_cost = sp_cost.startup;
4607 45710 : subplan->per_call_cost = sp_cost.per_tuple;
4608 45710 : }
4609 :
4610 :
4611 : /*
4612 : * cost_rescan
4613 : * Given a finished Path, estimate the costs of rescanning it after
4614 : * having done so the first time. For some Path types a rescan is
4615 : * cheaper than an original scan (if no parameters change), and this
4616 : * function embodies knowledge about that. The default is to return
4617 : * the same costs stored in the Path. (Note that the cost estimates
4618 : * actually stored in Paths are always for first scans.)
4619 : *
4620 : * This function is not currently intended to model effects such as rescans
4621 : * being cheaper due to disk block caching; what we are concerned with is
4622 : * plan types wherein the executor caches results explicitly, or doesn't
4623 : * redo startup calculations, etc.
4624 : */
4625 : static void
4626 2877214 : cost_rescan(PlannerInfo *root, Path *path,
4627 : Cost *rescan_startup_cost, /* output parameters */
4628 : Cost *rescan_total_cost)
4629 : {
4630 2877214 : switch (path->pathtype)
4631 : {
4632 54446 : case T_FunctionScan:
4633 :
4634 : /*
4635 : * Currently, nodeFunctionscan.c always executes the function to
4636 : * completion before returning any rows, and caches the results in
4637 : * a tuplestore. So the function eval cost is all startup cost
4638 : * and isn't paid over again on rescans. However, all run costs
4639 : * will be paid over again.
4640 : */
4641 54446 : *rescan_startup_cost = 0;
4642 54446 : *rescan_total_cost = path->total_cost - path->startup_cost;
4643 54446 : break;
4644 128374 : case T_HashJoin:
4645 :
4646 : /*
4647 : * If it's a single-batch join, we don't need to rebuild the hash
4648 : * table during a rescan.
4649 : */
4650 128374 : if (((HashPath *) path)->num_batches == 1)
4651 : {
4652 : /* Startup cost is exactly the cost of hash table building */
4653 128374 : *rescan_startup_cost = 0;
4654 128374 : *rescan_total_cost = path->total_cost - path->startup_cost;
4655 : }
4656 : else
4657 : {
4658 : /* Otherwise, no special treatment */
4659 0 : *rescan_startup_cost = path->startup_cost;
4660 0 : *rescan_total_cost = path->total_cost;
4661 : }
4662 128374 : break;
4663 8742 : case T_CteScan:
4664 : case T_WorkTableScan:
4665 : {
4666 : /*
4667 : * These plan types materialize their final result in a
4668 : * tuplestore or tuplesort object. So the rescan cost is only
4669 : * cpu_tuple_cost per tuple, unless the result is large enough
4670 : * to spill to disk.
4671 : */
4672 8742 : Cost run_cost = cpu_tuple_cost * path->rows;
4673 8742 : double nbytes = relation_byte_size(path->rows,
4674 8742 : path->pathtarget->width);
4675 8742 : double work_mem_bytes = work_mem * (Size) 1024;
4676 :
4677 8742 : if (nbytes > work_mem_bytes)
4678 : {
4679 : /* It will spill, so account for re-read cost */
4680 296 : double npages = ceil(nbytes / BLCKSZ);
4681 :
4682 296 : run_cost += seq_page_cost * npages;
4683 : }
4684 8742 : *rescan_startup_cost = 0;
4685 8742 : *rescan_total_cost = run_cost;
4686 : }
4687 8742 : break;
4688 977210 : case T_Material:
4689 : case T_Sort:
4690 : {
4691 : /*
4692 : * These plan types not only materialize their results, but do
4693 : * not implement qual filtering or projection. So they are
4694 : * even cheaper to rescan than the ones above. We charge only
4695 : * cpu_operator_cost per tuple. (Note: keep that in sync with
4696 : * the run_cost charge in cost_sort, and also see comments in
4697 : * cost_material before you change it.)
4698 : */
4699 977210 : Cost run_cost = cpu_operator_cost * path->rows;
4700 977210 : double nbytes = relation_byte_size(path->rows,
4701 977210 : path->pathtarget->width);
4702 977210 : double work_mem_bytes = work_mem * (Size) 1024;
4703 :
4704 977210 : if (nbytes > work_mem_bytes)
4705 : {
4706 : /* It will spill, so account for re-read cost */
4707 11554 : double npages = ceil(nbytes / BLCKSZ);
4708 :
4709 11554 : run_cost += seq_page_cost * npages;
4710 : }
4711 977210 : *rescan_startup_cost = 0;
4712 977210 : *rescan_total_cost = run_cost;
4713 : }
4714 977210 : break;
4715 290258 : case T_Memoize:
4716 : /* All the hard work is done by cost_memoize_rescan */
4717 290258 : cost_memoize_rescan(root, (MemoizePath *) path,
4718 : rescan_startup_cost, rescan_total_cost);
4719 290258 : break;
4720 1418184 : default:
4721 1418184 : *rescan_startup_cost = path->startup_cost;
4722 1418184 : *rescan_total_cost = path->total_cost;
4723 1418184 : break;
4724 : }
4725 2877214 : }
4726 :
4727 :
4728 : /*
4729 : * cost_qual_eval
4730 : * Estimate the CPU costs of evaluating a WHERE clause.
4731 : * The input can be either an implicitly-ANDed list of boolean
4732 : * expressions, or a list of RestrictInfo nodes. (The latter is
4733 : * preferred since it allows caching of the results.)
4734 : * The result includes both a one-time (startup) component,
4735 : * and a per-evaluation component.
4736 : *
4737 : * Note: in some code paths root can be passed as NULL, resulting in
4738 : * slightly worse estimates.
4739 : */
4740 : void
4741 3958882 : cost_qual_eval(QualCost *cost, List *quals, PlannerInfo *root)
4742 : {
4743 : cost_qual_eval_context context;
4744 : ListCell *l;
4745 :
4746 3958882 : context.root = root;
4747 3958882 : context.total.startup = 0;
4748 3958882 : context.total.per_tuple = 0;
4749 :
4750 : /* We don't charge any cost for the implicit ANDing at top level ... */
4751 :
4752 7482174 : foreach(l, quals)
4753 : {
4754 3523292 : Node *qual = (Node *) lfirst(l);
4755 :
4756 3523292 : cost_qual_eval_walker(qual, &context);
4757 : }
4758 :
4759 3958882 : *cost = context.total;
4760 3958882 : }
4761 :
4762 : /*
4763 : * cost_qual_eval_node
4764 : * As above, for a single RestrictInfo or expression.
4765 : */
4766 : void
4767 1820036 : cost_qual_eval_node(QualCost *cost, Node *qual, PlannerInfo *root)
4768 : {
4769 : cost_qual_eval_context context;
4770 :
4771 1820036 : context.root = root;
4772 1820036 : context.total.startup = 0;
4773 1820036 : context.total.per_tuple = 0;
4774 :
4775 1820036 : cost_qual_eval_walker(qual, &context);
4776 :
4777 1820036 : *cost = context.total;
4778 1820036 : }
4779 :
4780 : static bool
4781 8858316 : cost_qual_eval_walker(Node *node, cost_qual_eval_context *context)
4782 : {
4783 8858316 : if (node == NULL)
4784 90594 : return false;
4785 :
4786 : /*
4787 : * RestrictInfo nodes contain an eval_cost field reserved for this
4788 : * routine's use, so that it's not necessary to evaluate the qual clause's
4789 : * cost more than once. If the clause's cost hasn't been computed yet,
4790 : * the field's startup value will contain -1.
4791 : */
4792 8767722 : if (IsA(node, RestrictInfo))
4793 : {
4794 3715342 : RestrictInfo *rinfo = (RestrictInfo *) node;
4795 :
4796 3715342 : if (rinfo->eval_cost.startup < 0)
4797 : {
4798 : cost_qual_eval_context locContext;
4799 :
4800 579094 : locContext.root = context->root;
4801 579094 : locContext.total.startup = 0;
4802 579094 : locContext.total.per_tuple = 0;
4803 :
4804 : /*
4805 : * For an OR clause, recurse into the marked-up tree so that we
4806 : * set the eval_cost for contained RestrictInfos too.
4807 : */
4808 579094 : if (rinfo->orclause)
4809 10386 : cost_qual_eval_walker((Node *) rinfo->orclause, &locContext);
4810 : else
4811 568708 : cost_qual_eval_walker((Node *) rinfo->clause, &locContext);
4812 :
4813 : /*
4814 : * If the RestrictInfo is marked pseudoconstant, it will be tested
4815 : * only once, so treat its cost as all startup cost.
4816 : */
4817 579094 : if (rinfo->pseudoconstant)
4818 : {
4819 : /* count one execution during startup */
4820 9976 : locContext.total.startup += locContext.total.per_tuple;
4821 9976 : locContext.total.per_tuple = 0;
4822 : }
4823 579094 : rinfo->eval_cost = locContext.total;
4824 : }
4825 3715342 : context->total.startup += rinfo->eval_cost.startup;
4826 3715342 : context->total.per_tuple += rinfo->eval_cost.per_tuple;
4827 : /* do NOT recurse into children */
4828 3715342 : return false;
4829 : }
4830 :
4831 : /*
4832 : * For each operator or function node in the given tree, we charge the
4833 : * estimated execution cost given by pg_proc.procost (remember to multiply
4834 : * this by cpu_operator_cost).
4835 : *
4836 : * Vars and Consts are charged zero, and so are boolean operators (AND,
4837 : * OR, NOT). Simplistic, but a lot better than no model at all.
4838 : *
4839 : * Should we try to account for the possibility of short-circuit
4840 : * evaluation of AND/OR? Probably *not*, because that would make the
4841 : * results depend on the clause ordering, and we are not in any position
4842 : * to expect that the current ordering of the clauses is the one that's
4843 : * going to end up being used. The above per-RestrictInfo caching would
4844 : * not mix well with trying to re-order clauses anyway.
4845 : *
4846 : * Another issue that is entirely ignored here is that if a set-returning
4847 : * function is below top level in the tree, the functions/operators above
4848 : * it will need to be evaluated multiple times. In practical use, such
4849 : * cases arise so seldom as to not be worth the added complexity needed;
4850 : * moreover, since our rowcount estimates for functions tend to be pretty
4851 : * phony, the results would also be pretty phony.
4852 : */
4853 5052380 : if (IsA(node, FuncExpr))
4854 : {
4855 351276 : add_function_cost(context->root, ((FuncExpr *) node)->funcid, node,
4856 : &context->total);
4857 : }
4858 4701104 : else if (IsA(node, OpExpr) ||
4859 4057064 : IsA(node, DistinctExpr) ||
4860 4055816 : IsA(node, NullIfExpr))
4861 : {
4862 : /* rely on struct equivalence to treat these all alike */
4863 645412 : set_opfuncid((OpExpr *) node);
4864 645412 : add_function_cost(context->root, ((OpExpr *) node)->opfuncid, node,
4865 : &context->total);
4866 : }
4867 4055692 : else if (IsA(node, ScalarArrayOpExpr))
4868 : {
4869 44480 : ScalarArrayOpExpr *saop = (ScalarArrayOpExpr *) node;
4870 44480 : Node *arraynode = (Node *) lsecond(saop->args);
4871 : QualCost sacosts;
4872 : QualCost hcosts;
4873 44480 : double estarraylen = estimate_array_length(context->root, arraynode);
4874 :
4875 44480 : set_sa_opfuncid(saop);
4876 44480 : sacosts.startup = sacosts.per_tuple = 0;
4877 44480 : add_function_cost(context->root, saop->opfuncid, NULL,
4878 : &sacosts);
4879 :
4880 44480 : if (OidIsValid(saop->hashfuncid))
4881 : {
4882 : /* Handle costs for hashed ScalarArrayOpExpr */
4883 492 : hcosts.startup = hcosts.per_tuple = 0;
4884 :
4885 492 : add_function_cost(context->root, saop->hashfuncid, NULL, &hcosts);
4886 492 : context->total.startup += sacosts.startup + hcosts.startup;
4887 :
4888 : /* Estimate the cost of building the hashtable. */
4889 492 : context->total.startup += estarraylen * hcosts.per_tuple;
4890 :
4891 : /*
4892 : * XXX should we charge a little bit for sacosts.per_tuple when
4893 : * building the table, or is it ok to assume there will be zero
4894 : * hash collision?
4895 : */
4896 :
4897 : /*
4898 : * Charge for hashtable lookups. Charge a single hash and a
4899 : * single comparison.
4900 : */
4901 492 : context->total.per_tuple += hcosts.per_tuple + sacosts.per_tuple;
4902 : }
4903 : else
4904 : {
4905 : /*
4906 : * Estimate that the operator will be applied to about half of the
4907 : * array elements before the answer is determined.
4908 : */
4909 43988 : context->total.startup += sacosts.startup;
4910 87976 : context->total.per_tuple += sacosts.per_tuple *
4911 43988 : estimate_array_length(context->root, arraynode) * 0.5;
4912 : }
4913 : }
4914 4011212 : else if (IsA(node, Aggref) ||
4915 3956090 : IsA(node, WindowFunc))
4916 : {
4917 : /*
4918 : * Aggref and WindowFunc nodes are (and should be) treated like Vars,
4919 : * ie, zero execution cost in the current model, because they behave
4920 : * essentially like Vars at execution. We disregard the costs of
4921 : * their input expressions for the same reason. The actual execution
4922 : * costs of the aggregate/window functions and their arguments have to
4923 : * be factored into plan-node-specific costing of the Agg or WindowAgg
4924 : * plan node.
4925 : */
4926 58646 : return false; /* don't recurse into children */
4927 : }
4928 3952566 : else if (IsA(node, GroupingFunc))
4929 : {
4930 : /* Treat this as having cost 1 */
4931 422 : context->total.per_tuple += cpu_operator_cost;
4932 422 : return false; /* don't recurse into children */
4933 : }
4934 3952144 : else if (IsA(node, CoerceViaIO))
4935 : {
4936 22074 : CoerceViaIO *iocoerce = (CoerceViaIO *) node;
4937 : Oid iofunc;
4938 : Oid typioparam;
4939 : bool typisvarlena;
4940 :
4941 : /* check the result type's input function */
4942 22074 : getTypeInputInfo(iocoerce->resulttype,
4943 : &iofunc, &typioparam);
4944 22074 : add_function_cost(context->root, iofunc, NULL,
4945 : &context->total);
4946 : /* check the input type's output function */
4947 22074 : getTypeOutputInfo(exprType((Node *) iocoerce->arg),
4948 : &iofunc, &typisvarlena);
4949 22074 : add_function_cost(context->root, iofunc, NULL,
4950 : &context->total);
4951 : }
4952 3930070 : else if (IsA(node, ArrayCoerceExpr))
4953 : {
4954 5288 : ArrayCoerceExpr *acoerce = (ArrayCoerceExpr *) node;
4955 : QualCost perelemcost;
4956 :
4957 5288 : cost_qual_eval_node(&perelemcost, (Node *) acoerce->elemexpr,
4958 : context->root);
4959 5288 : context->total.startup += perelemcost.startup;
4960 5288 : if (perelemcost.per_tuple > 0)
4961 66 : context->total.per_tuple += perelemcost.per_tuple *
4962 66 : estimate_array_length(context->root, (Node *) acoerce->arg);
4963 : }
4964 3924782 : else if (IsA(node, RowCompareExpr))
4965 : {
4966 : /* Conservatively assume we will check all the columns */
4967 216 : RowCompareExpr *rcexpr = (RowCompareExpr *) node;
4968 : ListCell *lc;
4969 :
4970 702 : foreach(lc, rcexpr->opnos)
4971 : {
4972 486 : Oid opid = lfirst_oid(lc);
4973 :
4974 486 : add_function_cost(context->root, get_opcode(opid), NULL,
4975 : &context->total);
4976 : }
4977 : }
4978 3924566 : else if (IsA(node, MinMaxExpr) ||
4979 3924306 : IsA(node, SQLValueFunction) ||
4980 3919532 : IsA(node, XmlExpr) ||
4981 3918830 : IsA(node, CoerceToDomain) ||
4982 3909312 : IsA(node, NextValueExpr) ||
4983 3908950 : IsA(node, JsonExpr))
4984 : {
4985 : /* Treat all these as having cost 1 */
4986 18164 : context->total.per_tuple += cpu_operator_cost;
4987 : }
4988 3906402 : else if (IsA(node, SubLink))
4989 : {
4990 : /* This routine should not be applied to un-planned expressions */
4991 0 : elog(ERROR, "cannot handle unplanned sub-select");
4992 : }
4993 3906402 : else if (IsA(node, SubPlan))
4994 : {
4995 : /*
4996 : * A subplan node in an expression typically indicates that the
4997 : * subplan will be executed on each evaluation, so charge accordingly.
4998 : * (Sub-selects that can be executed as InitPlans have already been
4999 : * removed from the expression.)
5000 : */
5001 45400 : SubPlan *subplan = (SubPlan *) node;
5002 :
5003 45400 : context->total.startup += subplan->startup_cost;
5004 45400 : context->total.per_tuple += subplan->per_call_cost;
5005 :
5006 : /*
5007 : * We don't want to recurse into the testexpr, because it was already
5008 : * counted in the SubPlan node's costs. So we're done.
5009 : */
5010 45400 : return false;
5011 : }
5012 3861002 : else if (IsA(node, AlternativeSubPlan))
5013 : {
5014 : /*
5015 : * Arbitrarily use the first alternative plan for costing. (We should
5016 : * certainly only include one alternative, and we don't yet have
5017 : * enough information to know which one the executor is most likely to
5018 : * use.)
5019 : */
5020 1912 : AlternativeSubPlan *asplan = (AlternativeSubPlan *) node;
5021 :
5022 1912 : return cost_qual_eval_walker((Node *) linitial(asplan->subplans),
5023 : context);
5024 : }
5025 3859090 : else if (IsA(node, PlaceHolderVar))
5026 : {
5027 : /*
5028 : * A PlaceHolderVar should be given cost zero when considering general
5029 : * expression evaluation costs. The expense of doing the contained
5030 : * expression is charged as part of the tlist eval costs of the scan
5031 : * or join where the PHV is first computed (see set_rel_width and
5032 : * add_placeholders_to_joinrel). If we charged it again here, we'd be
5033 : * double-counting the cost for each level of plan that the PHV
5034 : * bubbles up through. Hence, return without recursing into the
5035 : * phexpr.
5036 : */
5037 4992 : return false;
5038 : }
5039 :
5040 : /* recurse into children */
5041 4941008 : return expression_tree_walker(node, cost_qual_eval_walker, context);
5042 : }
5043 :
5044 : /*
5045 : * get_restriction_qual_cost
5046 : * Compute evaluation costs of a baserel's restriction quals, plus any
5047 : * movable join quals that have been pushed down to the scan.
5048 : * Results are returned into *qpqual_cost.
5049 : *
5050 : * This is a convenience subroutine that works for seqscans and other cases
5051 : * where all the given quals will be evaluated the hard way. It's not useful
5052 : * for cost_index(), for example, where the index machinery takes care of
5053 : * some of the quals. We assume baserestrictcost was previously set by
5054 : * set_baserel_size_estimates().
5055 : */
5056 : static void
5057 1064428 : get_restriction_qual_cost(PlannerInfo *root, RelOptInfo *baserel,
5058 : ParamPathInfo *param_info,
5059 : QualCost *qpqual_cost)
5060 : {
5061 1064428 : if (param_info)
5062 : {
5063 : /* Include costs of pushed-down clauses */
5064 234468 : cost_qual_eval(qpqual_cost, param_info->ppi_clauses, root);
5065 :
5066 234468 : qpqual_cost->startup += baserel->baserestrictcost.startup;
5067 234468 : qpqual_cost->per_tuple += baserel->baserestrictcost.per_tuple;
5068 : }
5069 : else
5070 829960 : *qpqual_cost = baserel->baserestrictcost;
5071 1064428 : }
5072 :
5073 :
5074 : /*
5075 : * compute_semi_anti_join_factors
5076 : * Estimate how much of the inner input a SEMI, ANTI, or inner_unique join
5077 : * can be expected to scan.
5078 : *
5079 : * In a hash or nestloop SEMI/ANTI join, the executor will stop scanning
5080 : * inner rows as soon as it finds a match to the current outer row.
5081 : * The same happens if we have detected the inner rel is unique.
5082 : * We should therefore adjust some of the cost components for this effect.
5083 : * This function computes some estimates needed for these adjustments.
5084 : * These estimates will be the same regardless of the particular paths used
5085 : * for the outer and inner relation, so we compute these once and then pass
5086 : * them to all the join cost estimation functions.
5087 : *
5088 : * Input parameters:
5089 : * joinrel: join relation under consideration
5090 : * outerrel: outer relation under consideration
5091 : * innerrel: inner relation under consideration
5092 : * jointype: if not JOIN_SEMI or JOIN_ANTI, we assume it's inner_unique
5093 : * sjinfo: SpecialJoinInfo relevant to this join
5094 : * restrictlist: join quals
5095 : * Output parameters:
5096 : * *semifactors is filled in (see pathnodes.h for field definitions)
5097 : */
5098 : void
5099 214156 : compute_semi_anti_join_factors(PlannerInfo *root,
5100 : RelOptInfo *joinrel,
5101 : RelOptInfo *outerrel,
5102 : RelOptInfo *innerrel,
5103 : JoinType jointype,
5104 : SpecialJoinInfo *sjinfo,
5105 : List *restrictlist,
5106 : SemiAntiJoinFactors *semifactors)
5107 : {
5108 : Selectivity jselec;
5109 : Selectivity nselec;
5110 : Selectivity avgmatch;
5111 : SpecialJoinInfo norm_sjinfo;
5112 : List *joinquals;
5113 : ListCell *l;
5114 :
5115 : /*
5116 : * In an ANTI join, we must ignore clauses that are "pushed down", since
5117 : * those won't affect the match logic. In a SEMI join, we do not
5118 : * distinguish joinquals from "pushed down" quals, so just use the whole
5119 : * restrictinfo list. For other outer join types, we should consider only
5120 : * non-pushed-down quals, so that this devolves to an IS_OUTER_JOIN check.
5121 : */
5122 214156 : if (IS_OUTER_JOIN(jointype))
5123 : {
5124 78776 : joinquals = NIL;
5125 173812 : foreach(l, restrictlist)
5126 : {
5127 95036 : RestrictInfo *rinfo = lfirst_node(RestrictInfo, l);
5128 :
5129 95036 : if (!RINFO_IS_PUSHED_DOWN(rinfo, joinrel->relids))
5130 89098 : joinquals = lappend(joinquals, rinfo);
5131 : }
5132 : }
5133 : else
5134 135380 : joinquals = restrictlist;
5135 :
5136 : /*
5137 : * Get the JOIN_SEMI or JOIN_ANTI selectivity of the join clauses.
5138 : */
5139 214156 : jselec = clauselist_selectivity(root,
5140 : joinquals,
5141 : 0,
5142 : (jointype == JOIN_ANTI) ? JOIN_ANTI : JOIN_SEMI,
5143 : sjinfo);
5144 :
5145 : /*
5146 : * Also get the normal inner-join selectivity of the join clauses.
5147 : */
5148 214156 : init_dummy_sjinfo(&norm_sjinfo, outerrel->relids, innerrel->relids);
5149 :
5150 214156 : nselec = clauselist_selectivity(root,
5151 : joinquals,
5152 : 0,
5153 : JOIN_INNER,
5154 : &norm_sjinfo);
5155 :
5156 : /* Avoid leaking a lot of ListCells */
5157 214156 : if (IS_OUTER_JOIN(jointype))
5158 78776 : list_free(joinquals);
5159 :
5160 : /*
5161 : * jselec can be interpreted as the fraction of outer-rel rows that have
5162 : * any matches (this is true for both SEMI and ANTI cases). And nselec is
5163 : * the fraction of the Cartesian product that matches. So, the average
5164 : * number of matches for each outer-rel row that has at least one match is
5165 : * nselec * inner_rows / jselec.
5166 : *
5167 : * Note: it is correct to use the inner rel's "rows" count here, even
5168 : * though we might later be considering a parameterized inner path with
5169 : * fewer rows. This is because we have included all the join clauses in
5170 : * the selectivity estimate.
5171 : */
5172 214156 : if (jselec > 0) /* protect against zero divide */
5173 : {
5174 214124 : avgmatch = nselec * innerrel->rows / jselec;
5175 : /* Clamp to sane range */
5176 214124 : avgmatch = Max(1.0, avgmatch);
5177 : }
5178 : else
5179 32 : avgmatch = 1.0;
5180 :
5181 214156 : semifactors->outer_match_frac = jselec;
5182 214156 : semifactors->match_count = avgmatch;
5183 214156 : }
5184 :
5185 : /*
5186 : * has_indexed_join_quals
5187 : * Check whether all the joinquals of a nestloop join are used as
5188 : * inner index quals.
5189 : *
5190 : * If the inner path of a SEMI/ANTI join is an indexscan (including bitmap
5191 : * indexscan) that uses all the joinquals as indexquals, we can assume that an
5192 : * unmatched outer tuple is cheap to process, whereas otherwise it's probably
5193 : * expensive.
5194 : */
5195 : static bool
5196 910470 : has_indexed_join_quals(NestPath *path)
5197 : {
5198 910470 : JoinPath *joinpath = &path->jpath;
5199 910470 : Relids joinrelids = joinpath->path.parent->relids;
5200 910470 : Path *innerpath = joinpath->innerjoinpath;
5201 : List *indexclauses;
5202 : bool found_one;
5203 : ListCell *lc;
5204 :
5205 : /* If join still has quals to evaluate, it's not fast */
5206 910470 : if (joinpath->joinrestrictinfo != NIL)
5207 649304 : return false;
5208 : /* Nor if the inner path isn't parameterized at all */
5209 261166 : if (innerpath->param_info == NULL)
5210 4800 : return false;
5211 :
5212 : /* Find the indexclauses list for the inner scan */
5213 256366 : switch (innerpath->pathtype)
5214 : {
5215 154704 : case T_IndexScan:
5216 : case T_IndexOnlyScan:
5217 154704 : indexclauses = ((IndexPath *) innerpath)->indexclauses;
5218 154704 : break;
5219 270 : case T_BitmapHeapScan:
5220 : {
5221 : /* Accept only a simple bitmap scan, not AND/OR cases */
5222 270 : Path *bmqual = ((BitmapHeapPath *) innerpath)->bitmapqual;
5223 :
5224 270 : if (IsA(bmqual, IndexPath))
5225 222 : indexclauses = ((IndexPath *) bmqual)->indexclauses;
5226 : else
5227 48 : return false;
5228 222 : break;
5229 : }
5230 101392 : default:
5231 :
5232 : /*
5233 : * If it's not a simple indexscan, it probably doesn't run quickly
5234 : * for zero rows out, even if it's a parameterized path using all
5235 : * the joinquals.
5236 : */
5237 101392 : return false;
5238 : }
5239 :
5240 : /*
5241 : * Examine the inner path's param clauses. Any that are from the outer
5242 : * path must be found in the indexclauses list, either exactly or in an
5243 : * equivalent form generated by equivclass.c. Also, we must find at least
5244 : * one such clause, else it's a clauseless join which isn't fast.
5245 : */
5246 154926 : found_one = false;
5247 308344 : foreach(lc, innerpath->param_info->ppi_clauses)
5248 : {
5249 158474 : RestrictInfo *rinfo = (RestrictInfo *) lfirst(lc);
5250 :
5251 158474 : if (join_clause_is_movable_into(rinfo,
5252 158474 : innerpath->parent->relids,
5253 : joinrelids))
5254 : {
5255 157970 : if (!is_redundant_with_indexclauses(rinfo, indexclauses))
5256 5056 : return false;
5257 152914 : found_one = true;
5258 : }
5259 : }
5260 149870 : return found_one;
5261 : }
5262 :
5263 :
5264 : /*
5265 : * approx_tuple_count
5266 : * Quick-and-dirty estimation of the number of join rows passing
5267 : * a set of qual conditions.
5268 : *
5269 : * The quals can be either an implicitly-ANDed list of boolean expressions,
5270 : * or a list of RestrictInfo nodes (typically the latter).
5271 : *
5272 : * We intentionally compute the selectivity under JOIN_INNER rules, even
5273 : * if it's some type of outer join. This is appropriate because we are
5274 : * trying to figure out how many tuples pass the initial merge or hash
5275 : * join step.
5276 : *
5277 : * This is quick-and-dirty because we bypass clauselist_selectivity, and
5278 : * simply multiply the independent clause selectivities together. Now
5279 : * clauselist_selectivity often can't do any better than that anyhow, but
5280 : * for some situations (such as range constraints) it is smarter. However,
5281 : * we can't effectively cache the results of clauselist_selectivity, whereas
5282 : * the individual clause selectivities can be and are cached.
5283 : *
5284 : * Since we are only using the results to estimate how many potential
5285 : * output tuples are generated and passed through qpqual checking, it
5286 : * seems OK to live with the approximation.
5287 : */
5288 : static double
5289 497332 : approx_tuple_count(PlannerInfo *root, JoinPath *path, List *quals)
5290 : {
5291 : double tuples;
5292 497332 : double outer_tuples = path->outerjoinpath->rows;
5293 497332 : double inner_tuples = path->innerjoinpath->rows;
5294 : SpecialJoinInfo sjinfo;
5295 497332 : Selectivity selec = 1.0;
5296 : ListCell *l;
5297 :
5298 : /*
5299 : * Make up a SpecialJoinInfo for JOIN_INNER semantics.
5300 : */
5301 497332 : init_dummy_sjinfo(&sjinfo, path->outerjoinpath->parent->relids,
5302 497332 : path->innerjoinpath->parent->relids);
5303 :
5304 : /* Get the approximate selectivity */
5305 1067840 : foreach(l, quals)
5306 : {
5307 570508 : Node *qual = (Node *) lfirst(l);
5308 :
5309 : /* Note that clause_selectivity will be able to cache its result */
5310 570508 : selec *= clause_selectivity(root, qual, 0, JOIN_INNER, &sjinfo);
5311 : }
5312 :
5313 : /* Apply it to the input relation sizes */
5314 497332 : tuples = selec * outer_tuples * inner_tuples;
5315 :
5316 497332 : return clamp_row_est(tuples);
5317 : }
5318 :
5319 :
5320 : /*
5321 : * set_baserel_size_estimates
5322 : * Set the size estimates for the given base relation.
5323 : *
5324 : * The rel's targetlist and restrictinfo list must have been constructed
5325 : * already, and rel->tuples must be set.
5326 : *
5327 : * We set the following fields of the rel node:
5328 : * rows: the estimated number of output tuples (after applying
5329 : * restriction clauses).
5330 : * width: the estimated average output tuple width in bytes.
5331 : * baserestrictcost: estimated cost of evaluating baserestrictinfo clauses.
5332 : */
5333 : void
5334 503986 : set_baserel_size_estimates(PlannerInfo *root, RelOptInfo *rel)
5335 : {
5336 : double nrows;
5337 :
5338 : /* Should only be applied to base relations */
5339 : Assert(rel->relid > 0);
5340 :
5341 1007942 : nrows = rel->tuples *
5342 503986 : clauselist_selectivity(root,
5343 : rel->baserestrictinfo,
5344 : 0,
5345 : JOIN_INNER,
5346 : NULL);
5347 :
5348 503956 : rel->rows = clamp_row_est(nrows);
5349 :
5350 503956 : cost_qual_eval(&rel->baserestrictcost, rel->baserestrictinfo, root);
5351 :
5352 503956 : set_rel_width(root, rel);
5353 503956 : }
5354 :
5355 : /*
5356 : * get_parameterized_baserel_size
5357 : * Make a size estimate for a parameterized scan of a base relation.
5358 : *
5359 : * 'param_clauses' lists the additional join clauses to be used.
5360 : *
5361 : * set_baserel_size_estimates must have been applied already.
5362 : */
5363 : double
5364 149918 : get_parameterized_baserel_size(PlannerInfo *root, RelOptInfo *rel,
5365 : List *param_clauses)
5366 : {
5367 : List *allclauses;
5368 : double nrows;
5369 :
5370 : /*
5371 : * Estimate the number of rows returned by the parameterized scan, knowing
5372 : * that it will apply all the extra join clauses as well as the rel's own
5373 : * restriction clauses. Note that we force the clauses to be treated as
5374 : * non-join clauses during selectivity estimation.
5375 : */
5376 149918 : allclauses = list_concat_copy(param_clauses, rel->baserestrictinfo);
5377 299836 : nrows = rel->tuples *
5378 149918 : clauselist_selectivity(root,
5379 : allclauses,
5380 149918 : rel->relid, /* do not use 0! */
5381 : JOIN_INNER,
5382 : NULL);
5383 149918 : nrows = clamp_row_est(nrows);
5384 : /* For safety, make sure result is not more than the base estimate */
5385 149918 : if (nrows > rel->rows)
5386 0 : nrows = rel->rows;
5387 149918 : return nrows;
5388 : }
5389 :
5390 : /*
5391 : * set_joinrel_size_estimates
5392 : * Set the size estimates for the given join relation.
5393 : *
5394 : * The rel's targetlist must have been constructed already, and a
5395 : * restriction clause list that matches the given component rels must
5396 : * be provided.
5397 : *
5398 : * Since there is more than one way to make a joinrel for more than two
5399 : * base relations, the results we get here could depend on which component
5400 : * rel pair is provided. In theory we should get the same answers no matter
5401 : * which pair is provided; in practice, since the selectivity estimation
5402 : * routines don't handle all cases equally well, we might not. But there's
5403 : * not much to be done about it. (Would it make sense to repeat the
5404 : * calculations for each pair of input rels that's encountered, and somehow
5405 : * average the results? Probably way more trouble than it's worth, and
5406 : * anyway we must keep the rowcount estimate the same for all paths for the
5407 : * joinrel.)
5408 : *
5409 : * We set only the rows field here. The reltarget field was already set by
5410 : * build_joinrel_tlist, and baserestrictcost is not used for join rels.
5411 : */
5412 : void
5413 216270 : set_joinrel_size_estimates(PlannerInfo *root, RelOptInfo *rel,
5414 : RelOptInfo *outer_rel,
5415 : RelOptInfo *inner_rel,
5416 : SpecialJoinInfo *sjinfo,
5417 : List *restrictlist)
5418 : {
5419 216270 : rel->rows = calc_joinrel_size_estimate(root,
5420 : rel,
5421 : outer_rel,
5422 : inner_rel,
5423 : outer_rel->rows,
5424 : inner_rel->rows,
5425 : sjinfo,
5426 : restrictlist);
5427 216270 : }
5428 :
5429 : /*
5430 : * get_parameterized_joinrel_size
5431 : * Make a size estimate for a parameterized scan of a join relation.
5432 : *
5433 : * 'rel' is the joinrel under consideration.
5434 : * 'outer_path', 'inner_path' are (probably also parameterized) Paths that
5435 : * produce the relations being joined.
5436 : * 'sjinfo' is any SpecialJoinInfo relevant to this join.
5437 : * 'restrict_clauses' lists the join clauses that need to be applied at the
5438 : * join node (including any movable clauses that were moved down to this join,
5439 : * and not including any movable clauses that were pushed down into the
5440 : * child paths).
5441 : *
5442 : * set_joinrel_size_estimates must have been applied already.
5443 : */
5444 : double
5445 7344 : get_parameterized_joinrel_size(PlannerInfo *root, RelOptInfo *rel,
5446 : Path *outer_path,
5447 : Path *inner_path,
5448 : SpecialJoinInfo *sjinfo,
5449 : List *restrict_clauses)
5450 : {
5451 : double nrows;
5452 :
5453 : /*
5454 : * Estimate the number of rows returned by the parameterized join as the
5455 : * sizes of the input paths times the selectivity of the clauses that have
5456 : * ended up at this join node.
5457 : *
5458 : * As with set_joinrel_size_estimates, the rowcount estimate could depend
5459 : * on the pair of input paths provided, though ideally we'd get the same
5460 : * estimate for any pair with the same parameterization.
5461 : */
5462 7344 : nrows = calc_joinrel_size_estimate(root,
5463 : rel,
5464 : outer_path->parent,
5465 : inner_path->parent,
5466 : outer_path->rows,
5467 : inner_path->rows,
5468 : sjinfo,
5469 : restrict_clauses);
5470 : /* For safety, make sure result is not more than the base estimate */
5471 7344 : if (nrows > rel->rows)
5472 12 : nrows = rel->rows;
5473 7344 : return nrows;
5474 : }
5475 :
5476 : /*
5477 : * calc_joinrel_size_estimate
5478 : * Workhorse for set_joinrel_size_estimates and
5479 : * get_parameterized_joinrel_size.
5480 : *
5481 : * outer_rel/inner_rel are the relations being joined, but they should be
5482 : * assumed to have sizes outer_rows/inner_rows; those numbers might be less
5483 : * than what rel->rows says, when we are considering parameterized paths.
5484 : */
5485 : static double
5486 223614 : calc_joinrel_size_estimate(PlannerInfo *root,
5487 : RelOptInfo *joinrel,
5488 : RelOptInfo *outer_rel,
5489 : RelOptInfo *inner_rel,
5490 : double outer_rows,
5491 : double inner_rows,
5492 : SpecialJoinInfo *sjinfo,
5493 : List *restrictlist)
5494 : {
5495 223614 : JoinType jointype = sjinfo->jointype;
5496 : Selectivity fkselec;
5497 : Selectivity jselec;
5498 : Selectivity pselec;
5499 : double nrows;
5500 :
5501 : /*
5502 : * Compute joinclause selectivity. Note that we are only considering
5503 : * clauses that become restriction clauses at this join level; we are not
5504 : * double-counting them because they were not considered in estimating the
5505 : * sizes of the component rels.
5506 : *
5507 : * First, see whether any of the joinclauses can be matched to known FK
5508 : * constraints. If so, drop those clauses from the restrictlist, and
5509 : * instead estimate their selectivity using FK semantics. (We do this
5510 : * without regard to whether said clauses are local or "pushed down".
5511 : * Probably, an FK-matching clause could never be seen as pushed down at
5512 : * an outer join, since it would be strict and hence would be grounds for
5513 : * join strength reduction.) fkselec gets the net selectivity for
5514 : * FK-matching clauses, or 1.0 if there are none.
5515 : */
5516 223614 : fkselec = get_foreign_key_join_selectivity(root,
5517 : outer_rel->relids,
5518 : inner_rel->relids,
5519 : sjinfo,
5520 : &restrictlist);
5521 :
5522 : /*
5523 : * For an outer join, we have to distinguish the selectivity of the join's
5524 : * own clauses (JOIN/ON conditions) from any clauses that were "pushed
5525 : * down". For inner joins we just count them all as joinclauses.
5526 : */
5527 223614 : if (IS_OUTER_JOIN(jointype))
5528 : {
5529 83816 : List *joinquals = NIL;
5530 83816 : List *pushedquals = NIL;
5531 : ListCell *l;
5532 :
5533 : /* Grovel through the clauses to separate into two lists */
5534 188518 : foreach(l, restrictlist)
5535 : {
5536 104702 : RestrictInfo *rinfo = lfirst_node(RestrictInfo, l);
5537 :
5538 104702 : if (RINFO_IS_PUSHED_DOWN(rinfo, joinrel->relids))
5539 4864 : pushedquals = lappend(pushedquals, rinfo);
5540 : else
5541 99838 : joinquals = lappend(joinquals, rinfo);
5542 : }
5543 :
5544 : /* Get the separate selectivities */
5545 83816 : jselec = clauselist_selectivity(root,
5546 : joinquals,
5547 : 0,
5548 : jointype,
5549 : sjinfo);
5550 83816 : pselec = clauselist_selectivity(root,
5551 : pushedquals,
5552 : 0,
5553 : jointype,
5554 : sjinfo);
5555 :
5556 : /* Avoid leaking a lot of ListCells */
5557 83816 : list_free(joinquals);
5558 83816 : list_free(pushedquals);
5559 : }
5560 : else
5561 : {
5562 139798 : jselec = clauselist_selectivity(root,
5563 : restrictlist,
5564 : 0,
5565 : jointype,
5566 : sjinfo);
5567 139798 : pselec = 0.0; /* not used, keep compiler quiet */
5568 : }
5569 :
5570 : /*
5571 : * Basically, we multiply size of Cartesian product by selectivity.
5572 : *
5573 : * If we are doing an outer join, take that into account: the joinqual
5574 : * selectivity has to be clamped using the knowledge that the output must
5575 : * be at least as large as the non-nullable input. However, any
5576 : * pushed-down quals are applied after the outer join, so their
5577 : * selectivity applies fully.
5578 : *
5579 : * For JOIN_SEMI and JOIN_ANTI, the selectivity is defined as the fraction
5580 : * of LHS rows that have matches, and we apply that straightforwardly.
5581 : */
5582 223614 : switch (jointype)
5583 : {
5584 132002 : case JOIN_INNER:
5585 132002 : nrows = outer_rows * inner_rows * fkselec * jselec;
5586 : /* pselec not used */
5587 132002 : break;
5588 76812 : case JOIN_LEFT:
5589 76812 : nrows = outer_rows * inner_rows * fkselec * jselec;
5590 76812 : if (nrows < outer_rows)
5591 29292 : nrows = outer_rows;
5592 76812 : nrows *= pselec;
5593 76812 : break;
5594 1714 : case JOIN_FULL:
5595 1714 : nrows = outer_rows * inner_rows * fkselec * jselec;
5596 1714 : if (nrows < outer_rows)
5597 1136 : nrows = outer_rows;
5598 1714 : if (nrows < inner_rows)
5599 120 : nrows = inner_rows;
5600 1714 : nrows *= pselec;
5601 1714 : break;
5602 7796 : case JOIN_SEMI:
5603 7796 : nrows = outer_rows * fkselec * jselec;
5604 : /* pselec not used */
5605 7796 : break;
5606 5290 : case JOIN_ANTI:
5607 5290 : nrows = outer_rows * (1.0 - fkselec * jselec);
5608 5290 : nrows *= pselec;
5609 5290 : break;
5610 0 : default:
5611 : /* other values not expected here */
5612 0 : elog(ERROR, "unrecognized join type: %d", (int) jointype);
5613 : nrows = 0; /* keep compiler quiet */
5614 : break;
5615 : }
5616 :
5617 223614 : return clamp_row_est(nrows);
5618 : }
5619 :
5620 : /*
5621 : * get_foreign_key_join_selectivity
5622 : * Estimate join selectivity for foreign-key-related clauses.
5623 : *
5624 : * Remove any clauses that can be matched to FK constraints from *restrictlist,
5625 : * and return a substitute estimate of their selectivity. 1.0 is returned
5626 : * when there are no such clauses.
5627 : *
5628 : * The reason for treating such clauses specially is that we can get better
5629 : * estimates this way than by relying on clauselist_selectivity(), especially
5630 : * for multi-column FKs where that function's assumption that the clauses are
5631 : * independent falls down badly. But even with single-column FKs, we may be
5632 : * able to get a better answer when the pg_statistic stats are missing or out
5633 : * of date.
5634 : */
5635 : static Selectivity
5636 223614 : get_foreign_key_join_selectivity(PlannerInfo *root,
5637 : Relids outer_relids,
5638 : Relids inner_relids,
5639 : SpecialJoinInfo *sjinfo,
5640 : List **restrictlist)
5641 : {
5642 223614 : Selectivity fkselec = 1.0;
5643 223614 : JoinType jointype = sjinfo->jointype;
5644 223614 : List *worklist = *restrictlist;
5645 : ListCell *lc;
5646 :
5647 : /* Consider each FK constraint that is known to match the query */
5648 225636 : foreach(lc, root->fkey_list)
5649 : {
5650 2022 : ForeignKeyOptInfo *fkinfo = (ForeignKeyOptInfo *) lfirst(lc);
5651 : bool ref_is_outer;
5652 : List *removedlist;
5653 : ListCell *cell;
5654 :
5655 : /*
5656 : * This FK is not relevant unless it connects a baserel on one side of
5657 : * this join to a baserel on the other side.
5658 : */
5659 3692 : if (bms_is_member(fkinfo->con_relid, outer_relids) &&
5660 1670 : bms_is_member(fkinfo->ref_relid, inner_relids))
5661 1496 : ref_is_outer = false;
5662 866 : else if (bms_is_member(fkinfo->ref_relid, outer_relids) &&
5663 340 : bms_is_member(fkinfo->con_relid, inner_relids))
5664 130 : ref_is_outer = true;
5665 : else
5666 396 : continue;
5667 :
5668 : /*
5669 : * If we're dealing with a semi/anti join, and the FK's referenced
5670 : * relation is on the outside, then knowledge of the FK doesn't help
5671 : * us figure out what we need to know (which is the fraction of outer
5672 : * rows that have matches). On the other hand, if the referenced rel
5673 : * is on the inside, then all outer rows must have matches in the
5674 : * referenced table (ignoring nulls). But any restriction or join
5675 : * clauses that filter that table will reduce the fraction of matches.
5676 : * We can account for restriction clauses, but it's too hard to guess
5677 : * how many table rows would get through a join that's inside the RHS.
5678 : * Hence, if either case applies, punt and ignore the FK.
5679 : */
5680 1626 : if ((jointype == JOIN_SEMI || jointype == JOIN_ANTI) &&
5681 1104 : (ref_is_outer || bms_membership(inner_relids) != BMS_SINGLETON))
5682 12 : continue;
5683 :
5684 : /*
5685 : * Modify the restrictlist by removing clauses that match the FK (and
5686 : * putting them into removedlist instead). It seems unsafe to modify
5687 : * the originally-passed List structure, so we make a shallow copy the
5688 : * first time through.
5689 : */
5690 1614 : if (worklist == *restrictlist)
5691 1376 : worklist = list_copy(worklist);
5692 :
5693 1614 : removedlist = NIL;
5694 3382 : foreach(cell, worklist)
5695 : {
5696 1768 : RestrictInfo *rinfo = (RestrictInfo *) lfirst(cell);
5697 1768 : bool remove_it = false;
5698 : int i;
5699 :
5700 : /* Drop this clause if it matches any column of the FK */
5701 2246 : for (i = 0; i < fkinfo->nkeys; i++)
5702 : {
5703 2216 : if (rinfo->parent_ec)
5704 : {
5705 : /*
5706 : * EC-derived clauses can only match by EC. It is okay to
5707 : * consider any clause derived from the same EC as
5708 : * matching the FK: even if equivclass.c chose to generate
5709 : * a clause equating some other pair of Vars, it could
5710 : * have generated one equating the FK's Vars. So for
5711 : * purposes of estimation, we can act as though it did so.
5712 : *
5713 : * Note: checking parent_ec is a bit of a cheat because
5714 : * there are EC-derived clauses that don't have parent_ec
5715 : * set; but such clauses must compare expressions that
5716 : * aren't just Vars, so they cannot match the FK anyway.
5717 : */
5718 304 : if (fkinfo->eclass[i] == rinfo->parent_ec)
5719 : {
5720 298 : remove_it = true;
5721 298 : break;
5722 : }
5723 : }
5724 : else
5725 : {
5726 : /*
5727 : * Otherwise, see if rinfo was previously matched to FK as
5728 : * a "loose" clause.
5729 : */
5730 1912 : if (list_member_ptr(fkinfo->rinfos[i], rinfo))
5731 : {
5732 1440 : remove_it = true;
5733 1440 : break;
5734 : }
5735 : }
5736 : }
5737 1768 : if (remove_it)
5738 : {
5739 1738 : worklist = foreach_delete_current(worklist, cell);
5740 1738 : removedlist = lappend(removedlist, rinfo);
5741 : }
5742 : }
5743 :
5744 : /*
5745 : * If we failed to remove all the matching clauses we expected to
5746 : * find, chicken out and ignore this FK; applying its selectivity
5747 : * might result in double-counting. Put any clauses we did manage to
5748 : * remove back into the worklist.
5749 : *
5750 : * Since the matching clauses are known not outerjoin-delayed, they
5751 : * would normally have appeared in the initial joinclause list. If we
5752 : * didn't find them, there are two possibilities:
5753 : *
5754 : * 1. If the FK match is based on an EC that is ec_has_const, it won't
5755 : * have generated any join clauses at all. We discount such ECs while
5756 : * checking to see if we have "all" the clauses. (Below, we'll adjust
5757 : * the selectivity estimate for this case.)
5758 : *
5759 : * 2. The clauses were matched to some other FK in a previous
5760 : * iteration of this loop, and thus removed from worklist. (A likely
5761 : * case is that two FKs are matched to the same EC; there will be only
5762 : * one EC-derived clause in the initial list, so the first FK will
5763 : * consume it.) Applying both FKs' selectivity independently risks
5764 : * underestimating the join size; in particular, this would undo one
5765 : * of the main things that ECs were invented for, namely to avoid
5766 : * double-counting the selectivity of redundant equality conditions.
5767 : * Later we might think of a reasonable way to combine the estimates,
5768 : * but for now, just punt, since this is a fairly uncommon situation.
5769 : */
5770 1614 : if (removedlist == NIL ||
5771 1314 : list_length(removedlist) !=
5772 1314 : (fkinfo->nmatched_ec - fkinfo->nconst_ec + fkinfo->nmatched_ri))
5773 : {
5774 300 : worklist = list_concat(worklist, removedlist);
5775 300 : continue;
5776 : }
5777 :
5778 : /*
5779 : * Finally we get to the payoff: estimate selectivity using the
5780 : * knowledge that each referencing row will match exactly one row in
5781 : * the referenced table.
5782 : *
5783 : * XXX that's not true in the presence of nulls in the referencing
5784 : * column(s), so in principle we should derate the estimate for those.
5785 : * However (1) if there are any strict restriction clauses for the
5786 : * referencing column(s) elsewhere in the query, derating here would
5787 : * be double-counting the null fraction, and (2) it's not very clear
5788 : * how to combine null fractions for multiple referencing columns. So
5789 : * we do nothing for now about correcting for nulls.
5790 : *
5791 : * XXX another point here is that if either side of an FK constraint
5792 : * is an inheritance parent, we estimate as though the constraint
5793 : * covers all its children as well. This is not an unreasonable
5794 : * assumption for a referencing table, ie the user probably applied
5795 : * identical constraints to all child tables (though perhaps we ought
5796 : * to check that). But it's not possible to have done that for a
5797 : * referenced table. Fortunately, precisely because that doesn't
5798 : * work, it is uncommon in practice to have an FK referencing a parent
5799 : * table. So, at least for now, disregard inheritance here.
5800 : */
5801 1314 : if (jointype == JOIN_SEMI || jointype == JOIN_ANTI)
5802 866 : {
5803 : /*
5804 : * For JOIN_SEMI and JOIN_ANTI, we only get here when the FK's
5805 : * referenced table is exactly the inside of the join. The join
5806 : * selectivity is defined as the fraction of LHS rows that have
5807 : * matches. The FK implies that every LHS row has a match *in the
5808 : * referenced table*; but any restriction clauses on it will
5809 : * reduce the number of matches. Hence we take the join
5810 : * selectivity as equal to the selectivity of the table's
5811 : * restriction clauses, which is rows / tuples; but we must guard
5812 : * against tuples == 0.
5813 : */
5814 866 : RelOptInfo *ref_rel = find_base_rel(root, fkinfo->ref_relid);
5815 866 : double ref_tuples = Max(ref_rel->tuples, 1.0);
5816 :
5817 866 : fkselec *= ref_rel->rows / ref_tuples;
5818 : }
5819 : else
5820 : {
5821 : /*
5822 : * Otherwise, selectivity is exactly 1/referenced-table-size; but
5823 : * guard against tuples == 0. Note we should use the raw table
5824 : * tuple count, not any estimate of its filtered or joined size.
5825 : */
5826 448 : RelOptInfo *ref_rel = find_base_rel(root, fkinfo->ref_relid);
5827 448 : double ref_tuples = Max(ref_rel->tuples, 1.0);
5828 :
5829 448 : fkselec *= 1.0 / ref_tuples;
5830 : }
5831 :
5832 : /*
5833 : * If any of the FK columns participated in ec_has_const ECs, then
5834 : * equivclass.c will have generated "var = const" restrictions for
5835 : * each side of the join, thus reducing the sizes of both input
5836 : * relations. Taking the fkselec at face value would amount to
5837 : * double-counting the selectivity of the constant restriction for the
5838 : * referencing Var. Hence, look for the restriction clause(s) that
5839 : * were applied to the referencing Var(s), and divide out their
5840 : * selectivity to correct for this.
5841 : */
5842 1314 : if (fkinfo->nconst_ec > 0)
5843 : {
5844 24 : for (int i = 0; i < fkinfo->nkeys; i++)
5845 : {
5846 18 : EquivalenceClass *ec = fkinfo->eclass[i];
5847 :
5848 18 : if (ec && ec->ec_has_const)
5849 : {
5850 6 : EquivalenceMember *em = fkinfo->fk_eclass_member[i];
5851 6 : RestrictInfo *rinfo = find_derived_clause_for_ec_member(root,
5852 : ec,
5853 : em);
5854 :
5855 6 : if (rinfo)
5856 : {
5857 : Selectivity s0;
5858 :
5859 6 : s0 = clause_selectivity(root,
5860 : (Node *) rinfo,
5861 : 0,
5862 : jointype,
5863 : sjinfo);
5864 6 : if (s0 > 0)
5865 6 : fkselec /= s0;
5866 : }
5867 : }
5868 : }
5869 : }
5870 : }
5871 :
5872 223614 : *restrictlist = worklist;
5873 223614 : CLAMP_PROBABILITY(fkselec);
5874 223614 : return fkselec;
5875 : }
5876 :
5877 : /*
5878 : * set_subquery_size_estimates
5879 : * Set the size estimates for a base relation that is a subquery.
5880 : *
5881 : * The rel's targetlist and restrictinfo list must have been constructed
5882 : * already, and the Paths for the subquery must have been completed.
5883 : * We look at the subquery's PlannerInfo to extract data.
5884 : *
5885 : * We set the same fields as set_baserel_size_estimates.
5886 : */
5887 : void
5888 27936 : set_subquery_size_estimates(PlannerInfo *root, RelOptInfo *rel)
5889 : {
5890 27936 : PlannerInfo *subroot = rel->subroot;
5891 : RelOptInfo *sub_final_rel;
5892 : ListCell *lc;
5893 :
5894 : /* Should only be applied to base relations that are subqueries */
5895 : Assert(rel->relid > 0);
5896 : Assert(planner_rt_fetch(rel->relid, root)->rtekind == RTE_SUBQUERY);
5897 :
5898 : /*
5899 : * Copy raw number of output rows from subquery. All of its paths should
5900 : * have the same output rowcount, so just look at cheapest-total.
5901 : */
5902 27936 : sub_final_rel = fetch_upper_rel(subroot, UPPERREL_FINAL, NULL);
5903 27936 : rel->tuples = sub_final_rel->cheapest_total_path->rows;
5904 :
5905 : /*
5906 : * Compute per-output-column width estimates by examining the subquery's
5907 : * targetlist. For any output that is a plain Var, get the width estimate
5908 : * that was made while planning the subquery. Otherwise, we leave it to
5909 : * set_rel_width to fill in a datatype-based default estimate.
5910 : */
5911 116916 : foreach(lc, subroot->parse->targetList)
5912 : {
5913 88980 : TargetEntry *te = lfirst_node(TargetEntry, lc);
5914 88980 : Node *texpr = (Node *) te->expr;
5915 88980 : int32 item_width = 0;
5916 :
5917 : /* junk columns aren't visible to upper query */
5918 88980 : if (te->resjunk)
5919 1284 : continue;
5920 :
5921 : /*
5922 : * The subquery could be an expansion of a view that's had columns
5923 : * added to it since the current query was parsed, so that there are
5924 : * non-junk tlist columns in it that don't correspond to any column
5925 : * visible at our query level. Ignore such columns.
5926 : */
5927 87696 : if (te->resno < rel->min_attr || te->resno > rel->max_attr)
5928 0 : continue;
5929 :
5930 : /*
5931 : * XXX This currently doesn't work for subqueries containing set
5932 : * operations, because the Vars in their tlists are bogus references
5933 : * to the first leaf subquery, which wouldn't give the right answer
5934 : * even if we could still get to its PlannerInfo.
5935 : *
5936 : * Also, the subquery could be an appendrel for which all branches are
5937 : * known empty due to constraint exclusion, in which case
5938 : * set_append_rel_pathlist will have left the attr_widths set to zero.
5939 : *
5940 : * In either case, we just leave the width estimate zero until
5941 : * set_rel_width fixes it.
5942 : */
5943 87696 : if (IsA(texpr, Var) &&
5944 39856 : subroot->parse->setOperations == NULL)
5945 : {
5946 38082 : Var *var = (Var *) texpr;
5947 38082 : RelOptInfo *subrel = find_base_rel(subroot, var->varno);
5948 :
5949 38082 : item_width = subrel->attr_widths[var->varattno - subrel->min_attr];
5950 : }
5951 87696 : rel->attr_widths[te->resno - rel->min_attr] = item_width;
5952 : }
5953 :
5954 : /* Now estimate number of output rows, etc */
5955 27936 : set_baserel_size_estimates(root, rel);
5956 27936 : }
5957 :
5958 : /*
5959 : * set_function_size_estimates
5960 : * Set the size estimates for a base relation that is a function call.
5961 : *
5962 : * The rel's targetlist and restrictinfo list must have been constructed
5963 : * already.
5964 : *
5965 : * We set the same fields as set_baserel_size_estimates.
5966 : */
5967 : void
5968 55114 : set_function_size_estimates(PlannerInfo *root, RelOptInfo *rel)
5969 : {
5970 : RangeTblEntry *rte;
5971 : ListCell *lc;
5972 :
5973 : /* Should only be applied to base relations that are functions */
5974 : Assert(rel->relid > 0);
5975 55114 : rte = planner_rt_fetch(rel->relid, root);
5976 : Assert(rte->rtekind == RTE_FUNCTION);
5977 :
5978 : /*
5979 : * Estimate number of rows the functions will return. The rowcount of the
5980 : * node is that of the largest function result.
5981 : */
5982 55114 : rel->tuples = 0;
5983 110916 : foreach(lc, rte->functions)
5984 : {
5985 55802 : RangeTblFunction *rtfunc = (RangeTblFunction *) lfirst(lc);
5986 55802 : double ntup = expression_returns_set_rows(root, rtfunc->funcexpr);
5987 :
5988 55802 : if (ntup > rel->tuples)
5989 55138 : rel->tuples = ntup;
5990 : }
5991 :
5992 : /* Now estimate number of output rows, etc */
5993 55114 : set_baserel_size_estimates(root, rel);
5994 55114 : }
5995 :
5996 : /*
5997 : * set_function_size_estimates
5998 : * Set the size estimates for a base relation that is a function call.
5999 : *
6000 : * The rel's targetlist and restrictinfo list must have been constructed
6001 : * already.
6002 : *
6003 : * We set the same fields as set_tablefunc_size_estimates.
6004 : */
6005 : void
6006 626 : set_tablefunc_size_estimates(PlannerInfo *root, RelOptInfo *rel)
6007 : {
6008 : /* Should only be applied to base relations that are functions */
6009 : Assert(rel->relid > 0);
6010 : Assert(planner_rt_fetch(rel->relid, root)->rtekind == RTE_TABLEFUNC);
6011 :
6012 626 : rel->tuples = 100;
6013 :
6014 : /* Now estimate number of output rows, etc */
6015 626 : set_baserel_size_estimates(root, rel);
6016 626 : }
6017 :
6018 : /*
6019 : * set_values_size_estimates
6020 : * Set the size estimates for a base relation that is a values list.
6021 : *
6022 : * The rel's targetlist and restrictinfo list must have been constructed
6023 : * already.
6024 : *
6025 : * We set the same fields as set_baserel_size_estimates.
6026 : */
6027 : void
6028 8232 : set_values_size_estimates(PlannerInfo *root, RelOptInfo *rel)
6029 : {
6030 : RangeTblEntry *rte;
6031 :
6032 : /* Should only be applied to base relations that are values lists */
6033 : Assert(rel->relid > 0);
6034 8232 : rte = planner_rt_fetch(rel->relid, root);
6035 : Assert(rte->rtekind == RTE_VALUES);
6036 :
6037 : /*
6038 : * Estimate number of rows the values list will return. We know this
6039 : * precisely based on the list length (well, barring set-returning
6040 : * functions in list items, but that's a refinement not catered for
6041 : * anywhere else either).
6042 : */
6043 8232 : rel->tuples = list_length(rte->values_lists);
6044 :
6045 : /* Now estimate number of output rows, etc */
6046 8232 : set_baserel_size_estimates(root, rel);
6047 8232 : }
6048 :
6049 : /*
6050 : * set_cte_size_estimates
6051 : * Set the size estimates for a base relation that is a CTE reference.
6052 : *
6053 : * The rel's targetlist and restrictinfo list must have been constructed
6054 : * already, and we need an estimate of the number of rows returned by the CTE
6055 : * (if a regular CTE) or the non-recursive term (if a self-reference).
6056 : *
6057 : * We set the same fields as set_baserel_size_estimates.
6058 : */
6059 : void
6060 5100 : set_cte_size_estimates(PlannerInfo *root, RelOptInfo *rel, double cte_rows)
6061 : {
6062 : RangeTblEntry *rte;
6063 :
6064 : /* Should only be applied to base relations that are CTE references */
6065 : Assert(rel->relid > 0);
6066 5100 : rte = planner_rt_fetch(rel->relid, root);
6067 : Assert(rte->rtekind == RTE_CTE);
6068 :
6069 5100 : if (rte->self_reference)
6070 : {
6071 : /*
6072 : * In a self-reference, we assume the average worktable size is a
6073 : * multiple of the nonrecursive term's size. The best multiplier will
6074 : * vary depending on query "fan-out", so make its value adjustable.
6075 : */
6076 1010 : rel->tuples = clamp_row_est(recursive_worktable_factor * cte_rows);
6077 : }
6078 : else
6079 : {
6080 : /* Otherwise just believe the CTE's rowcount estimate */
6081 4090 : rel->tuples = cte_rows;
6082 : }
6083 :
6084 : /* Now estimate number of output rows, etc */
6085 5100 : set_baserel_size_estimates(root, rel);
6086 5100 : }
6087 :
6088 : /*
6089 : * set_namedtuplestore_size_estimates
6090 : * Set the size estimates for a base relation that is a tuplestore reference.
6091 : *
6092 : * The rel's targetlist and restrictinfo list must have been constructed
6093 : * already.
6094 : *
6095 : * We set the same fields as set_baserel_size_estimates.
6096 : */
6097 : void
6098 462 : set_namedtuplestore_size_estimates(PlannerInfo *root, RelOptInfo *rel)
6099 : {
6100 : RangeTblEntry *rte;
6101 :
6102 : /* Should only be applied to base relations that are tuplestore references */
6103 : Assert(rel->relid > 0);
6104 462 : rte = planner_rt_fetch(rel->relid, root);
6105 : Assert(rte->rtekind == RTE_NAMEDTUPLESTORE);
6106 :
6107 : /*
6108 : * Use the estimate provided by the code which is generating the named
6109 : * tuplestore. In some cases, the actual number might be available; in
6110 : * others the same plan will be re-used, so a "typical" value might be
6111 : * estimated and used.
6112 : */
6113 462 : rel->tuples = rte->enrtuples;
6114 462 : if (rel->tuples < 0)
6115 0 : rel->tuples = 1000;
6116 :
6117 : /* Now estimate number of output rows, etc */
6118 462 : set_baserel_size_estimates(root, rel);
6119 462 : }
6120 :
6121 : /*
6122 : * set_result_size_estimates
6123 : * Set the size estimates for an RTE_RESULT base relation
6124 : *
6125 : * The rel's targetlist and restrictinfo list must have been constructed
6126 : * already.
6127 : *
6128 : * We set the same fields as set_baserel_size_estimates.
6129 : */
6130 : void
6131 4190 : set_result_size_estimates(PlannerInfo *root, RelOptInfo *rel)
6132 : {
6133 : /* Should only be applied to RTE_RESULT base relations */
6134 : Assert(rel->relid > 0);
6135 : Assert(planner_rt_fetch(rel->relid, root)->rtekind == RTE_RESULT);
6136 :
6137 : /* RTE_RESULT always generates a single row, natively */
6138 4190 : rel->tuples = 1;
6139 :
6140 : /* Now estimate number of output rows, etc */
6141 4190 : set_baserel_size_estimates(root, rel);
6142 4190 : }
6143 :
6144 : /*
6145 : * set_foreign_size_estimates
6146 : * Set the size estimates for a base relation that is a foreign table.
6147 : *
6148 : * There is not a whole lot that we can do here; the foreign-data wrapper
6149 : * is responsible for producing useful estimates. We can do a decent job
6150 : * of estimating baserestrictcost, so we set that, and we also set up width
6151 : * using what will be purely datatype-driven estimates from the targetlist.
6152 : * There is no way to do anything sane with the rows value, so we just put
6153 : * a default estimate and hope that the wrapper can improve on it. The
6154 : * wrapper's GetForeignRelSize function will be called momentarily.
6155 : *
6156 : * The rel's targetlist and restrictinfo list must have been constructed
6157 : * already.
6158 : */
6159 : void
6160 2412 : set_foreign_size_estimates(PlannerInfo *root, RelOptInfo *rel)
6161 : {
6162 : /* Should only be applied to base relations */
6163 : Assert(rel->relid > 0);
6164 :
6165 2412 : rel->rows = 1000; /* entirely bogus default estimate */
6166 :
6167 2412 : cost_qual_eval(&rel->baserestrictcost, rel->baserestrictinfo, root);
6168 :
6169 2412 : set_rel_width(root, rel);
6170 2412 : }
6171 :
6172 :
6173 : /*
6174 : * set_rel_width
6175 : * Set the estimated output width of a base relation.
6176 : *
6177 : * The estimated output width is the sum of the per-attribute width estimates
6178 : * for the actually-referenced columns, plus any PHVs or other expressions
6179 : * that have to be calculated at this relation. This is the amount of data
6180 : * we'd need to pass upwards in case of a sort, hash, etc.
6181 : *
6182 : * This function also sets reltarget->cost, so it's a bit misnamed now.
6183 : *
6184 : * NB: this works best on plain relations because it prefers to look at
6185 : * real Vars. For subqueries, set_subquery_size_estimates will already have
6186 : * copied up whatever per-column estimates were made within the subquery,
6187 : * and for other types of rels there isn't much we can do anyway. We fall
6188 : * back on (fairly stupid) datatype-based width estimates if we can't get
6189 : * any better number.
6190 : *
6191 : * The per-attribute width estimates are cached for possible re-use while
6192 : * building join relations or post-scan/join pathtargets.
6193 : */
6194 : static void
6195 506368 : set_rel_width(PlannerInfo *root, RelOptInfo *rel)
6196 : {
6197 506368 : Oid reloid = planner_rt_fetch(rel->relid, root)->relid;
6198 506368 : int64 tuple_width = 0;
6199 506368 : bool have_wholerow_var = false;
6200 : ListCell *lc;
6201 :
6202 : /* Vars are assumed to have cost zero, but other exprs do not */
6203 506368 : rel->reltarget->cost.startup = 0;
6204 506368 : rel->reltarget->cost.per_tuple = 0;
6205 :
6206 1832376 : foreach(lc, rel->reltarget->exprs)
6207 : {
6208 1326008 : Node *node = (Node *) lfirst(lc);
6209 :
6210 : /*
6211 : * Ordinarily, a Var in a rel's targetlist must belong to that rel;
6212 : * but there are corner cases involving LATERAL references where that
6213 : * isn't so. If the Var has the wrong varno, fall through to the
6214 : * generic case (it doesn't seem worth the trouble to be any smarter).
6215 : */
6216 1326008 : if (IsA(node, Var) &&
6217 1301834 : ((Var *) node)->varno == rel->relid)
6218 357698 : {
6219 1301768 : Var *var = (Var *) node;
6220 : int ndx;
6221 : int32 item_width;
6222 :
6223 : Assert(var->varattno >= rel->min_attr);
6224 : Assert(var->varattno <= rel->max_attr);
6225 :
6226 1301768 : ndx = var->varattno - rel->min_attr;
6227 :
6228 : /*
6229 : * If it's a whole-row Var, we'll deal with it below after we have
6230 : * already cached as many attr widths as possible.
6231 : */
6232 1301768 : if (var->varattno == 0)
6233 : {
6234 2956 : have_wholerow_var = true;
6235 2956 : continue;
6236 : }
6237 :
6238 : /*
6239 : * The width may have been cached already (especially if it's a
6240 : * subquery), so don't duplicate effort.
6241 : */
6242 1298812 : if (rel->attr_widths[ndx] > 0)
6243 : {
6244 238584 : tuple_width += rel->attr_widths[ndx];
6245 238584 : continue;
6246 : }
6247 :
6248 : /* Try to get column width from statistics */
6249 1060228 : if (reloid != InvalidOid && var->varattno > 0)
6250 : {
6251 841604 : item_width = get_attavgwidth(reloid, var->varattno);
6252 841604 : if (item_width > 0)
6253 : {
6254 702530 : rel->attr_widths[ndx] = item_width;
6255 702530 : tuple_width += item_width;
6256 702530 : continue;
6257 : }
6258 : }
6259 :
6260 : /*
6261 : * Not a plain relation, or can't find statistics for it. Estimate
6262 : * using just the type info.
6263 : */
6264 357698 : item_width = get_typavgwidth(var->vartype, var->vartypmod);
6265 : Assert(item_width > 0);
6266 357698 : rel->attr_widths[ndx] = item_width;
6267 357698 : tuple_width += item_width;
6268 : }
6269 24240 : else if (IsA(node, PlaceHolderVar))
6270 : {
6271 : /*
6272 : * We will need to evaluate the PHV's contained expression while
6273 : * scanning this rel, so be sure to include it in reltarget->cost.
6274 : */
6275 1960 : PlaceHolderVar *phv = (PlaceHolderVar *) node;
6276 1960 : PlaceHolderInfo *phinfo = find_placeholder_info(root, phv);
6277 : QualCost cost;
6278 :
6279 1960 : tuple_width += phinfo->ph_width;
6280 1960 : cost_qual_eval_node(&cost, (Node *) phv->phexpr, root);
6281 1960 : rel->reltarget->cost.startup += cost.startup;
6282 1960 : rel->reltarget->cost.per_tuple += cost.per_tuple;
6283 : }
6284 : else
6285 : {
6286 : /*
6287 : * We could be looking at an expression pulled up from a subquery,
6288 : * or a ROW() representing a whole-row child Var, etc. Do what we
6289 : * can using the expression type information.
6290 : */
6291 : int32 item_width;
6292 : QualCost cost;
6293 :
6294 22280 : item_width = get_typavgwidth(exprType(node), exprTypmod(node));
6295 : Assert(item_width > 0);
6296 22280 : tuple_width += item_width;
6297 : /* Not entirely clear if we need to account for cost, but do so */
6298 22280 : cost_qual_eval_node(&cost, node, root);
6299 22280 : rel->reltarget->cost.startup += cost.startup;
6300 22280 : rel->reltarget->cost.per_tuple += cost.per_tuple;
6301 : }
6302 : }
6303 :
6304 : /*
6305 : * If we have a whole-row reference, estimate its width as the sum of
6306 : * per-column widths plus heap tuple header overhead.
6307 : */
6308 506368 : if (have_wholerow_var)
6309 : {
6310 2956 : int64 wholerow_width = MAXALIGN(SizeofHeapTupleHeader);
6311 :
6312 2956 : if (reloid != InvalidOid)
6313 : {
6314 : /* Real relation, so estimate true tuple width */
6315 2298 : wholerow_width += get_relation_data_width(reloid,
6316 2298 : rel->attr_widths - rel->min_attr);
6317 : }
6318 : else
6319 : {
6320 : /* Do what we can with info for a phony rel */
6321 : AttrNumber i;
6322 :
6323 1794 : for (i = 1; i <= rel->max_attr; i++)
6324 1136 : wholerow_width += rel->attr_widths[i - rel->min_attr];
6325 : }
6326 :
6327 2956 : rel->attr_widths[0 - rel->min_attr] = clamp_width_est(wholerow_width);
6328 :
6329 : /*
6330 : * Include the whole-row Var as part of the output tuple. Yes, that
6331 : * really is what happens at runtime.
6332 : */
6333 2956 : tuple_width += wholerow_width;
6334 : }
6335 :
6336 506368 : rel->reltarget->width = clamp_width_est(tuple_width);
6337 506368 : }
6338 :
6339 : /*
6340 : * set_pathtarget_cost_width
6341 : * Set the estimated eval cost and output width of a PathTarget tlist.
6342 : *
6343 : * As a notational convenience, returns the same PathTarget pointer passed in.
6344 : *
6345 : * Most, though not quite all, uses of this function occur after we've run
6346 : * set_rel_width() for base relations; so we can usually obtain cached width
6347 : * estimates for Vars. If we can't, fall back on datatype-based width
6348 : * estimates. Present early-planning uses of PathTargets don't need accurate
6349 : * widths badly enough to justify going to the catalogs for better data.
6350 : */
6351 : PathTarget *
6352 608654 : set_pathtarget_cost_width(PlannerInfo *root, PathTarget *target)
6353 : {
6354 608654 : int64 tuple_width = 0;
6355 : ListCell *lc;
6356 :
6357 : /* Vars are assumed to have cost zero, but other exprs do not */
6358 608654 : target->cost.startup = 0;
6359 608654 : target->cost.per_tuple = 0;
6360 :
6361 2120352 : foreach(lc, target->exprs)
6362 : {
6363 1511698 : Node *node = (Node *) lfirst(lc);
6364 :
6365 1511698 : tuple_width += get_expr_width(root, node);
6366 :
6367 : /* For non-Vars, account for evaluation cost */
6368 1511698 : if (!IsA(node, Var))
6369 : {
6370 : QualCost cost;
6371 :
6372 634706 : cost_qual_eval_node(&cost, node, root);
6373 634706 : target->cost.startup += cost.startup;
6374 634706 : target->cost.per_tuple += cost.per_tuple;
6375 : }
6376 : }
6377 :
6378 608654 : target->width = clamp_width_est(tuple_width);
6379 :
6380 608654 : return target;
6381 : }
6382 :
6383 : /*
6384 : * get_expr_width
6385 : * Estimate the width of the given expr attempting to use the width
6386 : * cached in a Var's owning RelOptInfo, else fallback on the type's
6387 : * average width when unable to or when the given Node is not a Var.
6388 : */
6389 : static int32
6390 1839646 : get_expr_width(PlannerInfo *root, const Node *expr)
6391 : {
6392 : int32 width;
6393 :
6394 1839646 : if (IsA(expr, Var))
6395 : {
6396 1192034 : const Var *var = (const Var *) expr;
6397 :
6398 : /* We should not see any upper-level Vars here */
6399 : Assert(var->varlevelsup == 0);
6400 :
6401 : /* Try to get data from RelOptInfo cache */
6402 1192034 : if (!IS_SPECIAL_VARNO(var->varno) &&
6403 1186424 : var->varno < root->simple_rel_array_size)
6404 : {
6405 1186424 : RelOptInfo *rel = root->simple_rel_array[var->varno];
6406 :
6407 1186424 : if (rel != NULL &&
6408 1157496 : var->varattno >= rel->min_attr &&
6409 1157496 : var->varattno <= rel->max_attr)
6410 : {
6411 1157496 : int ndx = var->varattno - rel->min_attr;
6412 :
6413 1157496 : if (rel->attr_widths[ndx] > 0)
6414 1123992 : return rel->attr_widths[ndx];
6415 : }
6416 : }
6417 :
6418 : /*
6419 : * No cached data available, so estimate using just the type info.
6420 : */
6421 68042 : width = get_typavgwidth(var->vartype, var->vartypmod);
6422 : Assert(width > 0);
6423 :
6424 68042 : return width;
6425 : }
6426 :
6427 647612 : width = get_typavgwidth(exprType(expr), exprTypmod(expr));
6428 : Assert(width > 0);
6429 647612 : return width;
6430 : }
6431 :
6432 : /*
6433 : * relation_byte_size
6434 : * Estimate the storage space in bytes for a given number of tuples
6435 : * of a given width (size in bytes).
6436 : */
6437 : static double
6438 4035286 : relation_byte_size(double tuples, int width)
6439 : {
6440 4035286 : return tuples * (MAXALIGN(width) + MAXALIGN(SizeofHeapTupleHeader));
6441 : }
6442 :
6443 : /*
6444 : * page_size
6445 : * Returns an estimate of the number of pages covered by a given
6446 : * number of tuples of a given width (size in bytes).
6447 : */
6448 : static double
6449 10688 : page_size(double tuples, int width)
6450 : {
6451 10688 : return ceil(relation_byte_size(tuples, width) / BLCKSZ);
6452 : }
6453 :
6454 : /*
6455 : * Estimate the fraction of the work that each worker will do given the
6456 : * number of workers budgeted for the path.
6457 : */
6458 : static double
6459 178686 : get_parallel_divisor(Path *path)
6460 : {
6461 178686 : double parallel_divisor = path->parallel_workers;
6462 :
6463 : /*
6464 : * Early experience with parallel query suggests that when there is only
6465 : * one worker, the leader often makes a very substantial contribution to
6466 : * executing the parallel portion of the plan, but as more workers are
6467 : * added, it does less and less, because it's busy reading tuples from the
6468 : * workers and doing whatever non-parallel post-processing is needed. By
6469 : * the time we reach 4 workers, the leader no longer makes a meaningful
6470 : * contribution. Thus, for now, estimate that the leader spends 30% of
6471 : * its time servicing each worker, and the remainder executing the
6472 : * parallel plan.
6473 : */
6474 178686 : if (parallel_leader_participation)
6475 : {
6476 : double leader_contribution;
6477 :
6478 177384 : leader_contribution = 1.0 - (0.3 * path->parallel_workers);
6479 177384 : if (leader_contribution > 0)
6480 175068 : parallel_divisor += leader_contribution;
6481 : }
6482 :
6483 178686 : return parallel_divisor;
6484 : }
6485 :
6486 : /*
6487 : * compute_bitmap_pages
6488 : * Estimate number of pages fetched from heap in a bitmap heap scan.
6489 : *
6490 : * 'baserel' is the relation to be scanned
6491 : * 'bitmapqual' is a tree of IndexPaths, BitmapAndPaths, and BitmapOrPaths
6492 : * 'loop_count' is the number of repetitions of the indexscan to factor into
6493 : * estimates of caching behavior
6494 : *
6495 : * If cost_p isn't NULL, the indexTotalCost estimate is returned in *cost_p.
6496 : * If tuples_p isn't NULL, the tuples_fetched estimate is returned in *tuples_p.
6497 : */
6498 : double
6499 681804 : compute_bitmap_pages(PlannerInfo *root, RelOptInfo *baserel,
6500 : Path *bitmapqual, double loop_count,
6501 : Cost *cost_p, double *tuples_p)
6502 : {
6503 : Cost indexTotalCost;
6504 : Selectivity indexSelectivity;
6505 : double T;
6506 : double pages_fetched;
6507 : double tuples_fetched;
6508 : double heap_pages;
6509 : double maxentries;
6510 :
6511 : /*
6512 : * Fetch total cost of obtaining the bitmap, as well as its total
6513 : * selectivity.
6514 : */
6515 681804 : cost_bitmap_tree_node(bitmapqual, &indexTotalCost, &indexSelectivity);
6516 :
6517 : /*
6518 : * Estimate number of main-table pages fetched.
6519 : */
6520 681804 : tuples_fetched = clamp_row_est(indexSelectivity * baserel->tuples);
6521 :
6522 681804 : T = (baserel->pages > 1) ? (double) baserel->pages : 1.0;
6523 :
6524 : /*
6525 : * For a single scan, the number of heap pages that need to be fetched is
6526 : * the same as the Mackert and Lohman formula for the case T <= b (ie, no
6527 : * re-reads needed).
6528 : */
6529 681804 : pages_fetched = (2.0 * T * tuples_fetched) / (2.0 * T + tuples_fetched);
6530 :
6531 : /*
6532 : * Calculate the number of pages fetched from the heap. Then based on
6533 : * current work_mem estimate get the estimated maxentries in the bitmap.
6534 : * (Note that we always do this calculation based on the number of pages
6535 : * that would be fetched in a single iteration, even if loop_count > 1.
6536 : * That's correct, because only that number of entries will be stored in
6537 : * the bitmap at one time.)
6538 : */
6539 681804 : heap_pages = Min(pages_fetched, baserel->pages);
6540 681804 : maxentries = tbm_calculate_entries(work_mem * (Size) 1024);
6541 :
6542 681804 : if (loop_count > 1)
6543 : {
6544 : /*
6545 : * For repeated bitmap scans, scale up the number of tuples fetched in
6546 : * the Mackert and Lohman formula by the number of scans, so that we
6547 : * estimate the number of pages fetched by all the scans. Then
6548 : * pro-rate for one scan.
6549 : */
6550 139466 : pages_fetched = index_pages_fetched(tuples_fetched * loop_count,
6551 : baserel->pages,
6552 : get_indexpath_pages(bitmapqual),
6553 : root);
6554 139466 : pages_fetched /= loop_count;
6555 : }
6556 :
6557 681804 : if (pages_fetched >= T)
6558 68468 : pages_fetched = T;
6559 : else
6560 613336 : pages_fetched = ceil(pages_fetched);
6561 :
6562 681804 : if (maxentries < heap_pages)
6563 : {
6564 : double exact_pages;
6565 : double lossy_pages;
6566 :
6567 : /*
6568 : * Crude approximation of the number of lossy pages. Because of the
6569 : * way tbm_lossify() is coded, the number of lossy pages increases
6570 : * very sharply as soon as we run short of memory; this formula has
6571 : * that property and seems to perform adequately in testing, but it's
6572 : * possible we could do better somehow.
6573 : */
6574 18 : lossy_pages = Max(0, heap_pages - maxentries / 2);
6575 18 : exact_pages = heap_pages - lossy_pages;
6576 :
6577 : /*
6578 : * If there are lossy pages then recompute the number of tuples
6579 : * processed by the bitmap heap node. We assume here that the chance
6580 : * of a given tuple coming from an exact page is the same as the
6581 : * chance that a given page is exact. This might not be true, but
6582 : * it's not clear how we can do any better.
6583 : */
6584 18 : if (lossy_pages > 0)
6585 : tuples_fetched =
6586 18 : clamp_row_est(indexSelectivity *
6587 18 : (exact_pages / heap_pages) * baserel->tuples +
6588 18 : (lossy_pages / heap_pages) * baserel->tuples);
6589 : }
6590 :
6591 681804 : if (cost_p)
6592 537172 : *cost_p = indexTotalCost;
6593 681804 : if (tuples_p)
6594 537172 : *tuples_p = tuples_fetched;
6595 :
6596 681804 : return pages_fetched;
6597 : }
6598 :
6599 : /*
6600 : * compute_gather_rows
6601 : * Estimate number of rows for gather (merge) nodes.
6602 : *
6603 : * In a parallel plan, each worker's row estimate is determined by dividing the
6604 : * total number of rows by parallel_divisor, which accounts for the leader's
6605 : * contribution in addition to the number of workers. Accordingly, when
6606 : * estimating the number of rows for gather (merge) nodes, we multiply the rows
6607 : * per worker by the same parallel_divisor to undo the division.
6608 : */
6609 : double
6610 29210 : compute_gather_rows(Path *path)
6611 : {
6612 : Assert(path->parallel_workers > 0);
6613 :
6614 29210 : return clamp_row_est(path->rows * get_parallel_divisor(path));
6615 : }
|
