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 7261926 : 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 7261926 : if (nrows > MAXIMUM_ROWCOUNT || isnan(nrows))
222 0 : nrows = MAXIMUM_ROWCOUNT;
223 7261926 : else if (nrows <= 1.0)
224 2784928 : nrows = 1.0;
225 : else
226 4476998 : nrows = rint(nrows);
227 :
228 7261926 : 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 1707638 : 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 1707638 : 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 1707638 : 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 43222 : 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 43222 : if (isnan(x))
272 0 : return LONG_MAX;
273 43222 : if (x <= 0)
274 488 : 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 42734 : 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 386594 : cost_seqscan(Path *path, PlannerInfo *root,
296 : RelOptInfo *baserel, ParamPathInfo *param_info)
297 : {
298 386594 : 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 386594 : if (param_info)
311 834 : path->rows = param_info->ppi_rows;
312 : else
313 385760 : path->rows = baserel->rows;
314 :
315 : /* fetch estimated page cost for tablespace containing table */
316 386594 : get_tablespace_page_costs(baserel->reltablespace,
317 : NULL,
318 : &spc_seq_page_cost);
319 :
320 : /*
321 : * disk costs
322 : */
323 386594 : disk_run_cost = spc_seq_page_cost * baserel->pages;
324 :
325 : /* CPU costs */
326 386594 : get_restriction_qual_cost(root, baserel, param_info, &qpqual_cost);
327 :
328 386594 : startup_cost += qpqual_cost.startup;
329 386594 : cpu_per_tuple = cpu_tuple_cost + qpqual_cost.per_tuple;
330 386594 : cpu_run_cost = cpu_per_tuple * baserel->tuples;
331 : /* tlist eval costs are paid per output row, not per tuple scanned */
332 386594 : startup_cost += path->pathtarget->cost.startup;
333 386594 : cpu_run_cost += path->pathtarget->cost.per_tuple * path->rows;
334 :
335 : /* Adjust costing for parallelism, if used. */
336 386594 : if (path->parallel_workers > 0)
337 : {
338 25718 : double parallel_divisor = get_parallel_divisor(path);
339 :
340 : /* The CPU cost is divided among all the workers. */
341 25718 : 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 25718 : path->rows = clamp_row_est(path->rows / parallel_divisor);
355 : }
356 :
357 386594 : path->disabled_nodes = enable_seqscan ? 0 : 1;
358 386594 : path->startup_cost = startup_cost;
359 386594 : path->total_cost = startup_cost + cpu_run_cost + disk_run_cost;
360 386594 : }
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
370 300 : cost_samplescan(Path *path, PlannerInfo *root,
371 : RelOptInfo *baserel, ParamPathInfo *param_info)
372 : {
373 300 : Cost startup_cost = 0;
374 300 : 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);
386 300 : rte = planner_rt_fetch(baserel->relid, root);
387 : Assert(rte->rtekind == RTE_RELATION);
388 300 : tsc = rte->tablesample;
389 : Assert(tsc != NULL);
390 300 : tsm = GetTsmRoutine(tsc->tsmhandler);
391 :
392 : /* Mark the path with the correct row estimate */
393 300 : if (param_info)
394 66 : path->rows = param_info->ppi_rows;
395 : else
396 234 : path->rows = baserel->rows;
397 :
398 : /* fetch estimated page cost for tablespace containing table */
399 300 : 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 */
404 600 : spc_page_cost = (tsm->NextSampleBlock != NULL) ?
405 300 : 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 : */
411 300 : 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 : */
421 300 : get_restriction_qual_cost(root, baserel, param_info, &qpqual_cost);
422 :
423 300 : startup_cost += qpqual_cost.startup;
424 300 : cpu_per_tuple = cpu_tuple_cost + qpqual_cost.per_tuple;
425 300 : run_cost += cpu_per_tuple * baserel->tuples;
426 : /* tlist eval costs are paid per output row, not per tuple scanned */
427 300 : startup_cost += path->pathtarget->cost.startup;
428 300 : run_cost += path->pathtarget->cost.per_tuple * path->rows;
429 :
430 300 : path->disabled_nodes = 0;
431 300 : path->startup_cost = startup_cost;
432 300 : path->total_cost = startup_cost + run_cost;
433 300 : }
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 16286 : cost_gather(GatherPath *path, PlannerInfo *root,
447 : RelOptInfo *rel, ParamPathInfo *param_info,
448 : double *rows)
449 : {
450 16286 : Cost startup_cost = 0;
451 16286 : Cost run_cost = 0;
452 :
453 : /* Mark the path with the correct row estimate */
454 16286 : if (rows)
455 1738 : path->path.rows = *rows;
456 14548 : else if (param_info)
457 0 : path->path.rows = param_info->ppi_rows;
458 : else
459 14548 : path->path.rows = rel->rows;
460 :
461 16286 : startup_cost = path->subpath->startup_cost;
462 :
463 16286 : run_cost = path->subpath->total_cost - path->subpath->startup_cost;
464 :
465 : /* Parallel setup and communication cost. */
466 16286 : startup_cost += parallel_setup_cost;
467 16286 : run_cost += parallel_tuple_cost * path->path.rows;
468 :
469 16286 : path->path.disabled_nodes = path->subpath->disabled_nodes;
470 16286 : path->path.startup_cost = startup_cost;
471 16286 : path->path.total_cost = (startup_cost + run_cost);
472 16286 : }
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 9972 : 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 9972 : Cost startup_cost = 0;
492 9972 : 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 9972 : if (rows)
499 4552 : path->path.rows = *rows;
500 5420 : else if (param_info)
501 0 : path->path.rows = param_info->ppi_rows;
502 : else
503 5420 : 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 9972 : N = (double) path->num_workers + 1;
512 9972 : logN = LOG2(N);
513 :
514 : /* Assumed cost per tuple comparison */
515 9972 : comparison_cost = 2.0 * cpu_operator_cost;
516 :
517 : /* Heap creation cost */
518 9972 : startup_cost += comparison_cost * N * logN;
519 :
520 : /* Per-tuple heap maintenance cost */
521 9972 : run_cost += path->path.rows * comparison_cost * logN;
522 :
523 : /* small cost for heap management, like cost_merge_append */
524 9972 : 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 9972 : startup_cost += parallel_setup_cost;
533 9972 : run_cost += parallel_tuple_cost * path->path.rows * 1.05;
534 :
535 9972 : path->path.disabled_nodes = input_disabled_nodes
536 9972 : + (enable_gathermerge ? 0 : 1);
537 9972 : path->path.startup_cost = startup_cost + input_startup_cost;
538 9972 : path->path.total_cost = (startup_cost + run_cost + input_total_cost);
539 9972 : }
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 680252 : cost_index(IndexPath *path, PlannerInfo *root, double loop_count,
561 : bool partial_path)
562 : {
563 680252 : IndexOptInfo *index = path->indexinfo;
564 680252 : RelOptInfo *baserel = index->rel;
565 680252 : bool indexonly = (path->path.pathtype == T_IndexOnlyScan);
566 : amcostestimate_function amcostestimate;
567 : List *qpquals;
568 680252 : Cost startup_cost = 0;
569 680252 : Cost run_cost = 0;
570 680252 : 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 680252 : if (path->path.param_info)
601 : {
602 125040 : path->path.rows = path->path.param_info->ppi_rows;
603 : /* qpquals come from the rel's restriction clauses and ppi_clauses */
604 125040 : qpquals = list_concat(extract_nonindex_conditions(path->indexinfo->indrestrictinfo,
605 : path->indexclauses),
606 125040 : extract_nonindex_conditions(path->path.param_info->ppi_clauses,
607 : path->indexclauses));
608 : }
609 : else
610 : {
611 555212 : path->path.rows = baserel->rows;
612 : /* qpquals come from just the rel's restriction clauses */
613 555212 : 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 680252 : 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 680252 : amcostestimate = (amcostestimate_function) index->amcostestimate;
628 680252 : 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 680252 : path->indextotalcost = indexTotalCost;
639 680252 : path->indexselectivity = indexSelectivity;
640 :
641 : /* all costs for touching index itself included here */
642 680252 : startup_cost += indexStartupCost;
643 680252 : run_cost += indexTotalCost - indexStartupCost;
644 :
645 : /* estimate number of main-table tuples fetched */
646 680252 : tuples_fetched = clamp_row_est(indexSelectivity * baserel->tuples);
647 :
648 : /* fetch estimated page costs for tablespace containing table */
649 680252 : 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 680252 : 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 69728 : pages_fetched = index_pages_fetched(tuples_fetched * loop_count,
691 : baserel->pages,
692 69728 : (double) index->pages,
693 : root);
694 :
695 69728 : if (indexonly)
696 8348 : pages_fetched = ceil(pages_fetched * (1.0 - baserel->allvisfrac));
697 :
698 69728 : rand_heap_pages = pages_fetched;
699 :
700 69728 : 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 69728 : pages_fetched = ceil(indexSelectivity * (double) baserel->pages);
713 :
714 69728 : pages_fetched = index_pages_fetched(pages_fetched * loop_count,
715 : baserel->pages,
716 69728 : (double) index->pages,
717 : root);
718 :
719 69728 : if (indexonly)
720 8348 : pages_fetched = ceil(pages_fetched * (1.0 - baserel->allvisfrac));
721 :
722 69728 : 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 610524 : pages_fetched = index_pages_fetched(tuples_fetched,
731 : baserel->pages,
732 610524 : (double) index->pages,
733 : root);
734 :
735 610524 : if (indexonly)
736 59744 : pages_fetched = ceil(pages_fetched * (1.0 - baserel->allvisfrac));
737 :
738 610524 : rand_heap_pages = pages_fetched;
739 :
740 : /* max_IO_cost is for the perfectly uncorrelated case (csquared=0) */
741 610524 : max_IO_cost = pages_fetched * spc_random_page_cost;
742 :
743 : /* min_IO_cost is for the perfectly correlated case (csquared=1) */
744 610524 : pages_fetched = ceil(indexSelectivity * (double) baserel->pages);
745 :
746 610524 : if (indexonly)
747 59744 : pages_fetched = ceil(pages_fetched * (1.0 - baserel->allvisfrac));
748 :
749 610524 : if (pages_fetched > 0)
750 : {
751 550452 : min_IO_cost = spc_random_page_cost;
752 550452 : if (pages_fetched > 1)
753 144790 : min_IO_cost += (pages_fetched - 1) * spc_seq_page_cost;
754 : }
755 : else
756 60072 : min_IO_cost = 0;
757 : }
758 :
759 680252 : 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 233180 : if (indexonly)
767 21550 : 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 233180 : 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 233180 : if (path->path.parallel_workers <= 0)
786 223268 : return;
787 :
788 9912 : 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 456984 : csquared = indexCorrelation * indexCorrelation;
796 :
797 456984 : 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 456984 : cost_qual_eval(&qpqual_cost, qpquals, root);
806 :
807 456984 : startup_cost += qpqual_cost.startup;
808 456984 : cpu_per_tuple = cpu_tuple_cost + qpqual_cost.per_tuple;
809 :
810 456984 : cpu_run_cost += cpu_per_tuple * tuples_fetched;
811 :
812 : /* tlist eval costs are paid per output row, not per tuple scanned */
813 456984 : startup_cost += path->path.pathtarget->cost.startup;
814 456984 : cpu_run_cost += path->path.pathtarget->cost.per_tuple * path->path.rows;
815 :
816 : /* Adjust costing for parallelism, if used. */
817 456984 : if (path->path.parallel_workers > 0)
818 : {
819 9912 : double parallel_divisor = get_parallel_divisor(&path->path);
820 :
821 9912 : path->path.rows = clamp_row_est(path->path.rows / parallel_divisor);
822 :
823 : /* The CPU cost is divided among all the workers. */
824 9912 : cpu_run_cost /= parallel_divisor;
825 : }
826 :
827 456984 : run_cost += cpu_run_cost;
828 :
829 456984 : path->path.startup_cost = startup_cost;
830 456984 : 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 805292 : extract_nonindex_conditions(List *qual_clauses, List *indexclauses)
851 : {
852 805292 : List *result = NIL;
853 : ListCell *lc;
854 :
855 1669864 : foreach(lc, qual_clauses)
856 : {
857 864572 : RestrictInfo *rinfo = lfirst_node(RestrictInfo, lc);
858 :
859 864572 : if (rinfo->pseudoconstant)
860 8962 : continue; /* we may drop pseudoconstants here */
861 855610 : if (is_redundant_with_indexclauses(rinfo, indexclauses))
862 522300 : continue; /* dup or derived from same EquivalenceClass */
863 : /* ... skip the predicate proof attempt createplan.c will try ... */
864 333310 : result = lappend(result, rinfo);
865 : }
866 805292 : 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 941980 : 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 941980 : T = (pages > 1) ? (double) pages : 1.0;
918 :
919 : /* Compute number of pages assumed to be competing for cache space */
920 941980 : total_pages = root->total_table_pages + index_pages;
921 941980 : total_pages = Max(total_pages, 1.0);
922 : Assert(T <= total_pages);
923 :
924 : /* b is pro-rated share of effective_cache_size */
925 941980 : b = (double) effective_cache_size * T / total_pages;
926 :
927 : /* force it positive and integral */
928 941980 : if (b <= 1.0)
929 0 : b = 1.0;
930 : else
931 941980 : b = ceil(b);
932 :
933 : /* This part is the Mackert and Lohman formula */
934 941980 : if (T <= b)
935 : {
936 941980 : pages_fetched =
937 941980 : (2.0 * T * tuples_fetched) / (2.0 * T + tuples_fetched);
938 941980 : if (pages_fetched >= T)
939 526090 : pages_fetched = T;
940 : else
941 415890 : 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 941980 : 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 158802 : get_indexpath_pages(Path *bitmapqual)
974 : {
975 158802 : double result = 0;
976 : ListCell *l;
977 :
978 158802 : if (IsA(bitmapqual, BitmapAndPath))
979 : {
980 20148 : BitmapAndPath *apath = (BitmapAndPath *) bitmapqual;
981 :
982 60444 : foreach(l, apath->bitmapquals)
983 : {
984 40296 : result += get_indexpath_pages((Path *) lfirst(l));
985 : }
986 : }
987 138654 : else if (IsA(bitmapqual, BitmapOrPath))
988 : {
989 62 : BitmapOrPath *opath = (BitmapOrPath *) bitmapqual;
990 :
991 186 : foreach(l, opath->bitmapquals)
992 : {
993 124 : result += get_indexpath_pages((Path *) lfirst(l));
994 : }
995 : }
996 138592 : else if (IsA(bitmapqual, IndexPath))
997 : {
998 138592 : IndexPath *ipath = (IndexPath *) bitmapqual;
999 :
1000 138592 : result = (double) ipath->indexinfo->pages;
1001 : }
1002 : else
1003 0 : elog(ERROR, "unrecognized node type: %d", nodeTag(bitmapqual));
1004 :
1005 158802 : 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 470930 : cost_bitmap_heap_scan(Path *path, PlannerInfo *root, RelOptInfo *baserel,
1024 : ParamPathInfo *param_info,
1025 : Path *bitmapqual, double loop_count)
1026 : {
1027 470930 : Cost startup_cost = 0;
1028 470930 : 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 470930 : if (param_info)
1047 194308 : path->rows = param_info->ppi_rows;
1048 : else
1049 276622 : path->rows = baserel->rows;
1050 :
1051 470930 : pages_fetched = compute_bitmap_pages(root, baserel, bitmapqual,
1052 : loop_count, &indexTotalCost,
1053 : &tuples_fetched);
1054 :
1055 470930 : startup_cost += indexTotalCost;
1056 470930 : T = (baserel->pages > 1) ? (double) baserel->pages : 1.0;
1057 :
1058 : /* Fetch estimated page costs for tablespace containing table. */
1059 470930 : 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 470930 : if (pages_fetched >= 2.0)
1071 88982 : cost_per_page = spc_random_page_cost -
1072 88982 : (spc_random_page_cost - spc_seq_page_cost)
1073 88982 : * sqrt(pages_fetched / T);
1074 : else
1075 381948 : cost_per_page = spc_random_page_cost;
1076 :
1077 470930 : 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 470930 : get_restriction_qual_cost(root, baserel, param_info, &qpqual_cost);
1089 :
1090 470930 : startup_cost += qpqual_cost.startup;
1091 470930 : cpu_per_tuple = cpu_tuple_cost + qpqual_cost.per_tuple;
1092 470930 : cpu_run_cost = cpu_per_tuple * tuples_fetched;
1093 :
1094 : /* Adjust costing for parallelism, if used. */
1095 470930 : if (path->parallel_workers > 0)
1096 : {
1097 4176 : double parallel_divisor = get_parallel_divisor(path);
1098 :
1099 : /* The CPU cost is divided among all the workers. */
1100 4176 : cpu_run_cost /= parallel_divisor;
1101 :
1102 4176 : path->rows = clamp_row_est(path->rows / parallel_divisor);
1103 : }
1104 :
1105 :
1106 470930 : run_cost += cpu_run_cost;
1107 :
1108 : /* tlist eval costs are paid per output row, not per tuple scanned */
1109 470930 : startup_cost += path->pathtarget->cost.startup;
1110 470930 : run_cost += path->pathtarget->cost.per_tuple * path->rows;
1111 :
1112 470930 : path->disabled_nodes = enable_bitmapscan ? 0 : 1;
1113 470930 : path->startup_cost = startup_cost;
1114 470930 : path->total_cost = startup_cost + run_cost;
1115 470930 : }
1116 :
1117 : /*
1118 : * cost_bitmap_tree_node
1119 : * Extract cost and selectivity from a bitmap tree node (index/and/or)
1120 : */
1121 : void
1122 886482 : cost_bitmap_tree_node(Path *path, Cost *cost, Selectivity *selec)
1123 : {
1124 886482 : if (IsA(path, IndexPath))
1125 : {
1126 838902 : *cost = ((IndexPath *) path)->indextotalcost;
1127 838902 : *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 838902 : *cost += 0.1 * cpu_operator_cost * path->rows;
1136 : }
1137 47580 : else if (IsA(path, BitmapAndPath))
1138 : {
1139 44156 : *cost = path->total_cost;
1140 44156 : *selec = ((BitmapAndPath *) path)->bitmapselectivity;
1141 : }
1142 3424 : else if (IsA(path, BitmapOrPath))
1143 : {
1144 3424 : *cost = path->total_cost;
1145 3424 : *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 886482 : }
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 44040 : 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 44040 : totalCost = 0.0;
1181 44040 : selec = 1.0;
1182 132120 : foreach(l, path->bitmapquals)
1183 : {
1184 88080 : Path *subpath = (Path *) lfirst(l);
1185 : Cost subCost;
1186 : Selectivity subselec;
1187 :
1188 88080 : cost_bitmap_tree_node(subpath, &subCost, &subselec);
1189 :
1190 88080 : selec *= subselec;
1191 :
1192 88080 : totalCost += subCost;
1193 88080 : if (l != list_head(path->bitmapquals))
1194 44040 : totalCost += 100.0 * cpu_operator_cost;
1195 : }
1196 44040 : path->bitmapselectivity = selec;
1197 44040 : path->path.rows = 0; /* per above, not used */
1198 44040 : path->path.disabled_nodes = 0;
1199 44040 : path->path.startup_cost = totalCost;
1200 44040 : path->path.total_cost = totalCost;
1201 44040 : }
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 1194 : 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 1194 : totalCost = 0.0;
1227 1194 : selec = 0.0;
1228 3396 : foreach(l, path->bitmapquals)
1229 : {
1230 2202 : Path *subpath = (Path *) lfirst(l);
1231 : Cost subCost;
1232 : Selectivity subselec;
1233 :
1234 2202 : cost_bitmap_tree_node(subpath, &subCost, &subselec);
1235 :
1236 2202 : selec += subselec;
1237 :
1238 2202 : totalCost += subCost;
1239 2202 : if (l != list_head(path->bitmapquals) &&
1240 1008 : !IsA(subpath, IndexPath))
1241 6 : totalCost += 100.0 * cpu_operator_cost;
1242 : }
1243 1194 : path->bitmapselectivity = Min(selec, 1.0);
1244 1194 : path->path.rows = 0; /* per above, not used */
1245 1194 : path->path.startup_cost = totalCost;
1246 1194 : path->path.total_cost = totalCost;
1247 1194 : }
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 820 : cost_tidscan(Path *path, PlannerInfo *root,
1259 : RelOptInfo *baserel, List *tidquals, ParamPathInfo *param_info)
1260 : {
1261 820 : Cost startup_cost = 0;
1262 820 : 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 820 : if (param_info)
1277 144 : path->rows = param_info->ppi_rows;
1278 : else
1279 676 : path->rows = baserel->rows;
1280 :
1281 : /* Count how many tuples we expect to retrieve */
1282 820 : ntuples = 0;
1283 1664 : foreach(l, tidquals)
1284 : {
1285 844 : RestrictInfo *rinfo = lfirst_node(RestrictInfo, l);
1286 844 : 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 844 : 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 794 : 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 390 : 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 820 : cost_qual_eval(&tid_qual_cost, tidquals, root);
1321 :
1322 : /* fetch estimated page cost for tablespace containing table */
1323 820 : 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 820 : run_cost += spc_random_page_cost * ntuples;
1329 :
1330 : /* Add scanning CPU costs */
1331 820 : get_restriction_qual_cost(root, baserel, param_info, &qpqual_cost);
1332 :
1333 : /* XXX currently we assume TID quals are a subset of qpquals */
1334 820 : startup_cost += qpqual_cost.startup + tid_qual_cost.per_tuple;
1335 820 : cpu_per_tuple = cpu_tuple_cost + qpqual_cost.per_tuple -
1336 820 : tid_qual_cost.per_tuple;
1337 820 : run_cost += cpu_per_tuple * ntuples;
1338 :
1339 : /* tlist eval costs are paid per output row, not per tuple scanned */
1340 820 : startup_cost += path->pathtarget->cost.startup;
1341 820 : 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 820 : path->disabled_nodes = 0;
1349 820 : path->startup_cost = startup_cost;
1350 820 : path->total_cost = startup_cost + run_cost;
1351 820 : }
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 202 : cost_tidrangescan(Path *path, PlannerInfo *root,
1364 : RelOptInfo *baserel, List *tidrangequals,
1365 : ParamPathInfo *param_info)
1366 : {
1367 : Selectivity selectivity;
1368 : double pages;
1369 202 : Cost startup_cost = 0;
1370 202 : 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 202 : if (param_info)
1385 0 : path->rows = param_info->ppi_rows;
1386 : else
1387 202 : path->rows = baserel->rows;
1388 :
1389 : /* Count how many tuples and pages we expect to scan */
1390 202 : selectivity = clauselist_selectivity(root, tidrangequals, baserel->relid,
1391 : JOIN_INNER, NULL);
1392 202 : pages = ceil(selectivity * baserel->pages);
1393 :
1394 202 : 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 202 : ntuples = selectivity * baserel->tuples;
1406 202 : 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 202 : cost_qual_eval(&tid_qual_cost, tidrangequals, root);
1413 :
1414 : /* fetch estimated page cost for tablespace containing table */
1415 202 : 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 202 : run_cost += spc_random_page_cost + spc_seq_page_cost * nseqpages;
1421 :
1422 : /* Add scanning CPU costs */
1423 202 : 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 202 : startup_cost += qpqual_cost.startup + tid_qual_cost.per_tuple;
1433 202 : cpu_per_tuple = cpu_tuple_cost + qpqual_cost.per_tuple -
1434 202 : tid_qual_cost.per_tuple;
1435 202 : run_cost += cpu_per_tuple * ntuples;
1436 :
1437 : /* tlist eval costs are paid per output row, not per tuple scanned */
1438 202 : startup_cost += path->pathtarget->cost.startup;
1439 202 : 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 202 : path->disabled_nodes = 0;
1444 202 : path->startup_cost = startup_cost;
1445 202 : path->total_cost = startup_cost + run_cost;
1446 202 : }
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 42076 : 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 42076 : if (param_info)
1478 534 : qpquals = list_concat_copy(param_info->ppi_clauses,
1479 534 : baserel->baserestrictinfo);
1480 : else
1481 41542 : qpquals = baserel->baserestrictinfo;
1482 :
1483 42076 : path->path.rows = clamp_row_est(path->subpath->rows *
1484 42076 : 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 42076 : path->path.disabled_nodes = path->subpath->disabled_nodes;
1497 42076 : path->path.startup_cost = path->subpath->startup_cost;
1498 42076 : 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 42076 : if (qpquals == NIL && trivial_pathtarget)
1514 20022 : return;
1515 :
1516 22054 : get_restriction_qual_cost(root, baserel, param_info, &qpqual_cost);
1517 :
1518 22054 : startup_cost = qpqual_cost.startup;
1519 22054 : cpu_per_tuple = cpu_tuple_cost + qpqual_cost.per_tuple;
1520 22054 : run_cost = cpu_per_tuple * path->subpath->rows;
1521 :
1522 : /* tlist eval costs are paid per output row, not per tuple scanned */
1523 22054 : startup_cost += path->path.pathtarget->cost.startup;
1524 22054 : run_cost += path->path.pathtarget->cost.per_tuple * path->path.rows;
1525 :
1526 22054 : path->path.startup_cost += startup_cost;
1527 22054 : 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 44176 : cost_functionscan(Path *path, PlannerInfo *root,
1539 : RelOptInfo *baserel, ParamPathInfo *param_info)
1540 : {
1541 44176 : Cost startup_cost = 0;
1542 44176 : 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 44176 : rte = planner_rt_fetch(baserel->relid, root);
1551 : Assert(rte->rtekind == RTE_FUNCTION);
1552 :
1553 : /* Mark the path with the correct row estimate */
1554 44176 : if (param_info)
1555 8498 : path->rows = param_info->ppi_rows;
1556 : else
1557 35678 : 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 44176 : cost_qual_eval_node(&exprcost, (Node *) rte->functions, root);
1573 :
1574 44176 : startup_cost += exprcost.startup + exprcost.per_tuple;
1575 :
1576 : /* Add scanning CPU costs */
1577 44176 : get_restriction_qual_cost(root, baserel, param_info, &qpqual_cost);
1578 :
1579 44176 : startup_cost += qpqual_cost.startup;
1580 44176 : cpu_per_tuple = cpu_tuple_cost + qpqual_cost.per_tuple;
1581 44176 : run_cost += cpu_per_tuple * baserel->tuples;
1582 :
1583 : /* tlist eval costs are paid per output row, not per tuple scanned */
1584 44176 : startup_cost += path->pathtarget->cost.startup;
1585 44176 : run_cost += path->pathtarget->cost.per_tuple * path->rows;
1586 :
1587 44176 : path->disabled_nodes = 0;
1588 44176 : path->startup_cost = startup_cost;
1589 44176 : path->total_cost = startup_cost + run_cost;
1590 44176 : }
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 7980 : cost_valuesscan(Path *path, PlannerInfo *root,
1658 : RelOptInfo *baserel, ParamPathInfo *param_info)
1659 : {
1660 7980 : Cost startup_cost = 0;
1661 7980 : 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 7980 : if (param_info)
1671 48 : path->rows = param_info->ppi_rows;
1672 : else
1673 7932 : 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 7980 : cpu_per_tuple = cpu_operator_cost;
1680 :
1681 : /* Add scanning CPU costs */
1682 7980 : get_restriction_qual_cost(root, baserel, param_info, &qpqual_cost);
1683 :
1684 7980 : startup_cost += qpqual_cost.startup;
1685 7980 : cpu_per_tuple += cpu_tuple_cost + qpqual_cost.per_tuple;
1686 7980 : run_cost += cpu_per_tuple * baserel->tuples;
1687 :
1688 : /* tlist eval costs are paid per output row, not per tuple scanned */
1689 7980 : startup_cost += path->pathtarget->cost.startup;
1690 7980 : run_cost += path->pathtarget->cost.per_tuple * path->rows;
1691 :
1692 7980 : path->disabled_nodes = 0;
1693 7980 : path->startup_cost = startup_cost;
1694 7980 : path->total_cost = startup_cost + run_cost;
1695 7980 : }
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 4016 : cost_ctescan(Path *path, PlannerInfo *root,
1709 : RelOptInfo *baserel, ParamPathInfo *param_info)
1710 : {
1711 4016 : Cost startup_cost = 0;
1712 4016 : 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 4016 : if (param_info)
1722 0 : path->rows = param_info->ppi_rows;
1723 : else
1724 4016 : path->rows = baserel->rows;
1725 :
1726 : /* Charge one CPU tuple cost per row for tuplestore manipulation */
1727 4016 : cpu_per_tuple = cpu_tuple_cost;
1728 :
1729 : /* Add scanning CPU costs */
1730 4016 : get_restriction_qual_cost(root, baserel, param_info, &qpqual_cost);
1731 :
1732 4016 : startup_cost += qpqual_cost.startup;
1733 4016 : cpu_per_tuple += cpu_tuple_cost + qpqual_cost.per_tuple;
1734 4016 : run_cost += cpu_per_tuple * baserel->tuples;
1735 :
1736 : /* tlist eval costs are paid per output row, not per tuple scanned */
1737 4016 : startup_cost += path->pathtarget->cost.startup;
1738 4016 : run_cost += path->pathtarget->cost.per_tuple * path->rows;
1739 :
1740 4016 : path->disabled_nodes = 0;
1741 4016 : path->startup_cost = startup_cost;
1742 4016 : path->total_cost = startup_cost + run_cost;
1743 4016 : }
1744 :
1745 : /*
1746 : * cost_namedtuplestorescan
1747 : * Determines and returns the cost of scanning a named tuplestore.
1748 : */
1749 : void
1750 438 : cost_namedtuplestorescan(Path *path, PlannerInfo *root,
1751 : RelOptInfo *baserel, ParamPathInfo *param_info)
1752 : {
1753 438 : Cost startup_cost = 0;
1754 438 : 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 438 : if (param_info)
1764 0 : path->rows = param_info->ppi_rows;
1765 : else
1766 438 : path->rows = baserel->rows;
1767 :
1768 : /* Charge one CPU tuple cost per row for tuplestore manipulation */
1769 438 : cpu_per_tuple = cpu_tuple_cost;
1770 :
1771 : /* Add scanning CPU costs */
1772 438 : get_restriction_qual_cost(root, baserel, param_info, &qpqual_cost);
1773 :
1774 438 : startup_cost += qpqual_cost.startup;
1775 438 : cpu_per_tuple += cpu_tuple_cost + qpqual_cost.per_tuple;
1776 438 : run_cost += cpu_per_tuple * baserel->tuples;
1777 :
1778 438 : path->disabled_nodes = 0;
1779 438 : path->startup_cost = startup_cost;
1780 438 : path->total_cost = startup_cost + run_cost;
1781 438 : }
1782 :
1783 : /*
1784 : * cost_resultscan
1785 : * Determines and returns the cost of scanning an RTE_RESULT relation.
1786 : */
1787 : void
1788 1646 : cost_resultscan(Path *path, PlannerInfo *root,
1789 : RelOptInfo *baserel, ParamPathInfo *param_info)
1790 : {
1791 1646 : Cost startup_cost = 0;
1792 1646 : 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 1646 : if (param_info)
1802 144 : path->rows = param_info->ppi_rows;
1803 : else
1804 1502 : path->rows = baserel->rows;
1805 :
1806 : /* We charge qual cost plus cpu_tuple_cost */
1807 1646 : get_restriction_qual_cost(root, baserel, param_info, &qpqual_cost);
1808 :
1809 1646 : startup_cost += qpqual_cost.startup;
1810 1646 : cpu_per_tuple = cpu_tuple_cost + qpqual_cost.per_tuple;
1811 1646 : run_cost += cpu_per_tuple * baserel->tuples;
1812 :
1813 1646 : path->disabled_nodes = 0;
1814 1646 : path->startup_cost = startup_cost;
1815 1646 : path->total_cost = startup_cost + run_cost;
1816 1646 : }
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 816 : 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 816 : startup_cost = nrterm->startup_cost;
1834 816 : total_cost = nrterm->total_cost;
1835 816 : 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 816 : total_cost += 10 * rterm->total_cost;
1844 816 : 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 816 : total_cost += cpu_tuple_cost * total_rows;
1852 :
1853 816 : runion->disabled_nodes = nrterm->disabled_nodes + rterm->disabled_nodes;
1854 816 : runion->startup_cost = startup_cost;
1855 816 : runion->total_cost = total_cost;
1856 816 : runion->rows = total_rows;
1857 816 : runion->pathtarget->width = Max(nrterm->pathtarget->width,
1858 : rterm->pathtarget->width);
1859 816 : }
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 1378310 : 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 1378310 : double input_bytes = relation_byte_size(tuples, width);
1904 : double output_bytes;
1905 : double output_tuples;
1906 1378310 : long sort_mem_bytes = sort_mem * 1024L;
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 1378310 : if (tuples < 2.0)
1913 425154 : tuples = 2.0;
1914 :
1915 : /* Include the default cost-per-comparison */
1916 1378310 : comparison_cost += 2.0 * cpu_operator_cost;
1917 :
1918 : /* Do we have a useful LIMIT? */
1919 1378310 : if (limit_tuples > 0 && limit_tuples < tuples)
1920 : {
1921 1818 : output_tuples = limit_tuples;
1922 1818 : output_bytes = relation_byte_size(output_tuples, width);
1923 : }
1924 : else
1925 : {
1926 1376492 : output_tuples = tuples;
1927 1376492 : output_bytes = input_bytes;
1928 : }
1929 :
1930 1378310 : if (output_bytes > sort_mem_bytes)
1931 : {
1932 : /*
1933 : * We'll have to use a disk-based sort of all the tuples
1934 : */
1935 16842 : double npages = ceil(input_bytes / BLCKSZ);
1936 16842 : double nruns = input_bytes / sort_mem_bytes;
1937 16842 : 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 16842 : *startup_cost = comparison_cost * tuples * LOG2(tuples);
1947 :
1948 : /* Disk costs */
1949 :
1950 : /* Compute logM(r) as log(r) / log(M) */
1951 16842 : if (nruns > mergeorder)
1952 4674 : log_runs = ceil(log(nruns) / log(mergeorder));
1953 : else
1954 12168 : log_runs = 1.0;
1955 16842 : npageaccesses = 2.0 * npages * log_runs;
1956 : /* Assume 3/4ths of accesses are sequential, 1/4th are not */
1957 16842 : *startup_cost += npageaccesses *
1958 16842 : (seq_page_cost * 0.75 + random_page_cost * 0.25);
1959 : }
1960 1361468 : 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 1342 : *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 1360126 : *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 1378310 : *run_cost = cpu_operator_cost * tuples;
1985 1378310 : }
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 9964 : 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 9964 : 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 9964 : List *presortedExprs = NIL;
2016 : ListCell *l;
2017 9964 : 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 9964 : if (input_tuples < 2.0)
2026 5878 : input_tuples = 2.0;
2027 :
2028 : /* Default estimate of number of groups, capped to one group per row. */
2029 9964 : 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 10060 : foreach(l, pathkeys)
2054 : {
2055 10060 : PathKey *key = (PathKey *) lfirst(l);
2056 10060 : EquivalenceMember *member = (EquivalenceMember *)
2057 10060 : 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 10060 : 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 10050 : presortedExprs = lappend(presortedExprs, member->em_expr);
2071 :
2072 10050 : if (foreach_current_index(l) + 1 >= presorted_keys)
2073 9954 : break;
2074 : }
2075 :
2076 : /* Estimate the number of groups with equal presorted keys. */
2077 9964 : if (!unknown_varno)
2078 9954 : input_groups = estimate_num_groups(root, presortedExprs, input_tuples,
2079 : NULL, NULL);
2080 :
2081 9964 : group_tuples = input_tuples / input_groups;
2082 9964 : 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 9964 : 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 9964 : 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 9964 : run_cost = group_run_cost + (group_run_cost + group_startup_cost) *
2106 9964 : (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 9964 : 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 9964 : run_cost += 2.0 * cpu_tuple_cost * input_groups;
2120 :
2121 9964 : path->rows = input_tuples;
2122 :
2123 : /* should not generate these paths when enable_incremental_sort=false */
2124 : Assert(enable_incremental_sort);
2125 9964 : path->disabled_nodes = input_disabled_nodes;
2126 :
2127 9964 : path->startup_cost = startup_cost;
2128 9964 : path->total_cost = startup_cost + run_cost;
2129 9964 : }
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 1368346 : 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 1368346 : cost_tuplesort(&startup_cost, &run_cost,
2155 : tuples, width,
2156 : comparison_cost, sort_mem,
2157 : limit_tuples);
2158 :
2159 1368346 : startup_cost += input_cost;
2160 :
2161 1368346 : path->rows = tuples;
2162 1368346 : path->disabled_nodes = input_disabled_nodes + (enable_sort ? 0 : 1);
2163 1368346 : path->startup_cost = startup_cost;
2164 1368346 : path->total_cost = startup_cost + run_cost;
2165 1368346 : }
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 15490 : 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 15490 : if (numpaths == 0)
2185 13906 : return 0;
2186 :
2187 : /*
2188 : * Array length is number of workers or number of relevant paths,
2189 : * whichever is less.
2190 : */
2191 1584 : arrlen = Min(parallel_workers, numpaths);
2192 1584 : costarr = (Cost *) palloc(sizeof(Cost) * arrlen);
2193 :
2194 : /* The first few paths will each be claimed by a different worker. */
2195 1584 : path_index = 0;
2196 4284 : foreach(cell, subpaths)
2197 : {
2198 3538 : Path *subpath = (Path *) lfirst(cell);
2199 :
2200 3538 : if (path_index == arrlen)
2201 838 : break;
2202 2700 : 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 1584 : 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 2066 : for_each_cell(l, subpaths, cell)
2216 : {
2217 1022 : Path *subpath = (Path *) lfirst(l);
2218 :
2219 : /* Consider only the non-partial paths */
2220 1022 : if (path_index++ == numpaths)
2221 540 : break;
2222 :
2223 482 : costarr[min_index] += subpath->total_cost;
2224 :
2225 : /* Update the new min cost array index */
2226 482 : min_index = 0;
2227 1482 : for (int i = 0; i < arrlen; i++)
2228 : {
2229 1000 : if (costarr[i] < costarr[min_index])
2230 202 : min_index = i;
2231 : }
2232 : }
2233 :
2234 : /* Return the highest cost from the array */
2235 1584 : max_index = 0;
2236 4284 : for (int i = 0; i < arrlen; i++)
2237 : {
2238 2700 : if (costarr[i] > costarr[max_index])
2239 182 : max_index = i;
2240 : }
2241 :
2242 1584 : return costarr[max_index];
2243 : }
2244 :
2245 : /*
2246 : * cost_append
2247 : * Determines and returns the cost of an Append node.
2248 : */
2249 : void
2250 47168 : cost_append(AppendPath *apath)
2251 : {
2252 : ListCell *l;
2253 :
2254 47168 : apath->path.disabled_nodes = 0;
2255 47168 : apath->path.startup_cost = 0;
2256 47168 : apath->path.total_cost = 0;
2257 47168 : apath->path.rows = 0;
2258 :
2259 47168 : if (apath->subpaths == NIL)
2260 1584 : return;
2261 :
2262 45584 : if (!apath->path.parallel_aware)
2263 : {
2264 30094 : List *pathkeys = apath->path.pathkeys;
2265 :
2266 30094 : if (pathkeys == NIL)
2267 : {
2268 28058 : 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 28058 : 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 110674 : foreach(l, apath->subpaths)
2281 : {
2282 82616 : Path *subpath = (Path *) lfirst(l);
2283 :
2284 82616 : apath->path.rows += subpath->rows;
2285 82616 : apath->path.disabled_nodes += subpath->disabled_nodes;
2286 82616 : 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 7944 : foreach(l, apath->subpaths)
2310 : {
2311 5908 : Path *subpath = (Path *) lfirst(l);
2312 : Path sort_path; /* dummy for result of cost_sort */
2313 :
2314 5908 : 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 5908 : apath->path.rows += subpath->rows;
2336 5908 : apath->path.disabled_nodes += subpath->disabled_nodes;
2337 5908 : apath->path.startup_cost += subpath->startup_cost;
2338 5908 : apath->path.total_cost += subpath->total_cost;
2339 : }
2340 : }
2341 : }
2342 : else /* parallel-aware */
2343 : {
2344 15490 : int i = 0;
2345 15490 : 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 62786 : foreach(l, apath->subpaths)
2352 : {
2353 47296 : 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 47296 : if (i == 0)
2361 15490 : apath->path.startup_cost = subpath->startup_cost;
2362 31806 : else if (i < apath->path.parallel_workers)
2363 14956 : 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 47296 : if (i < apath->first_partial_path)
2374 3182 : apath->path.rows += subpath->rows / parallel_divisor;
2375 : else
2376 : {
2377 : double subpath_parallel_divisor;
2378 :
2379 44114 : subpath_parallel_divisor = get_parallel_divisor(subpath);
2380 44114 : apath->path.rows += subpath->rows * (subpath_parallel_divisor /
2381 : parallel_divisor);
2382 44114 : apath->path.total_cost += subpath->total_cost;
2383 : }
2384 :
2385 47296 : apath->path.disabled_nodes += subpath->disabled_nodes;
2386 47296 : apath->path.rows = clamp_row_est(apath->path.rows);
2387 :
2388 47296 : i++;
2389 : }
2390 :
2391 : /* Add cost for non-partial subpaths. */
2392 15490 : apath->path.total_cost +=
2393 15490 : 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 45584 : apath->path.total_cost +=
2403 45584 : 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 4010 : 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 4010 : Cost startup_cost = 0;
2439 4010 : Cost run_cost = 0;
2440 : Cost comparison_cost;
2441 : double N;
2442 : double logN;
2443 :
2444 : /*
2445 : * Avoid log(0)...
2446 : */
2447 4010 : N = (n_streams < 2) ? 2.0 : (double) n_streams;
2448 4010 : logN = LOG2(N);
2449 :
2450 : /* Assumed cost per tuple comparison */
2451 4010 : comparison_cost = 2.0 * cpu_operator_cost;
2452 :
2453 : /* Heap creation cost */
2454 4010 : startup_cost += comparison_cost * N * logN;
2455 :
2456 : /* Per-tuple heap maintenance cost */
2457 4010 : 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 4010 : run_cost += cpu_tuple_cost * APPEND_CPU_COST_MULTIPLIER * tuples;
2464 :
2465 4010 : path->disabled_nodes = input_disabled_nodes;
2466 4010 : path->startup_cost = startup_cost + input_startup_cost;
2467 4010 : path->total_cost = startup_cost + run_cost + input_total_cost;
2468 4010 : }
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 446226 : 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 446226 : Cost startup_cost = input_startup_cost;
2489 446226 : Cost run_cost = input_total_cost - input_startup_cost;
2490 446226 : double nbytes = relation_byte_size(tuples, width);
2491 446226 : long work_mem_bytes = work_mem * 1024L;
2492 :
2493 446226 : 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 446226 : 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 446226 : if (nbytes > work_mem_bytes)
2516 : {
2517 4602 : double npages = ceil(nbytes / BLCKSZ);
2518 :
2519 4602 : run_cost += seq_page_cost * npages;
2520 : }
2521 :
2522 446226 : path->disabled_nodes = input_disabled_nodes + (enable_material ? 0 : 1);
2523 446226 : path->startup_cost = startup_cost;
2524 446226 : path->total_cost = startup_cost + run_cost;
2525 446226 : }
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 234280 : cost_memoize_rescan(PlannerInfo *root, MemoizePath *mpath,
2542 : Cost *rescan_startup_cost, Cost *rescan_total_cost)
2543 : {
2544 : EstimationInfo estinfo;
2545 : ListCell *lc;
2546 234280 : Cost input_startup_cost = mpath->subpath->startup_cost;
2547 234280 : Cost input_total_cost = mpath->subpath->total_cost;
2548 234280 : double tuples = mpath->subpath->rows;
2549 234280 : double calls = mpath->calls;
2550 234280 : 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 234280 : 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 234280 : est_entry_bytes = relation_byte_size(tuples, width) +
2571 234280 : ExecEstimateCacheEntryOverheadBytes(tuples);
2572 :
2573 : /* include the estimated width for the cache keys */
2574 493432 : foreach(lc, mpath->param_exprs)
2575 259152 : 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 234280 : est_cache_entries = floor(hash_mem_bytes / est_entry_bytes);
2579 :
2580 : /* estimate on the distinct number of parameter values */
2581 234280 : 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 234280 : if ((estinfo.flags & SELFLAG_USED_DEFAULT) != 0)
2593 13158 : 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 234280 : 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 234280 : 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 468560 : hit_ratio = ((calls - ndistinct) / calls) *
2622 234280 : (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 234280 : 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 234280 : 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 234280 : 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 234280 : 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 234280 : 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 234280 : startup_cost += cpu_tuple_cost;
2665 :
2666 234280 : *rescan_startup_cost = startup_cost;
2667 234280 : *rescan_total_cost = total_cost;
2668 234280 : }
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 63800 : 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 : AggClauseCosts dummy_aggcosts;
2694 :
2695 : /* Use all-zero per-aggregate costs if NULL is passed */
2696 63800 : if (aggcosts == NULL)
2697 : {
2698 : Assert(aggstrategy == AGG_HASHED);
2699 55392 : MemSet(&dummy_aggcosts, 0, sizeof(AggClauseCosts));
2700 9232 : aggcosts = &dummy_aggcosts;
2701 : }
2702 :
2703 : /*
2704 : * The transCost.per_tuple component of aggcosts should be charged once
2705 : * per input tuple, corresponding to the costs of evaluating the aggregate
2706 : * transfns and their input expressions. The finalCost.per_tuple component
2707 : * is charged once per output tuple, corresponding to the costs of
2708 : * evaluating the finalfns. Startup costs are of course charged but once.
2709 : *
2710 : * If we are grouping, we charge an additional cpu_operator_cost per
2711 : * grouping column per input tuple for grouping comparisons.
2712 : *
2713 : * We will produce a single output tuple if not grouping, and a tuple per
2714 : * group otherwise. We charge cpu_tuple_cost for each output tuple.
2715 : *
2716 : * Note: in this cost model, AGG_SORTED and AGG_HASHED have exactly the
2717 : * same total CPU cost, but AGG_SORTED has lower startup cost. If the
2718 : * input path is already sorted appropriately, AGG_SORTED should be
2719 : * preferred (since it has no risk of memory overflow). This will happen
2720 : * as long as the computed total costs are indeed exactly equal --- but if
2721 : * there's roundoff error we might do the wrong thing. So be sure that
2722 : * the computations below form the same intermediate values in the same
2723 : * order.
2724 : */
2725 63800 : if (aggstrategy == AGG_PLAIN)
2726 : {
2727 35244 : startup_cost = input_total_cost;
2728 35244 : startup_cost += aggcosts->transCost.startup;
2729 35244 : startup_cost += aggcosts->transCost.per_tuple * input_tuples;
2730 35244 : startup_cost += aggcosts->finalCost.startup;
2731 35244 : startup_cost += aggcosts->finalCost.per_tuple;
2732 : /* we aren't grouping */
2733 35244 : total_cost = startup_cost + cpu_tuple_cost;
2734 35244 : output_tuples = 1;
2735 : }
2736 28556 : else if (aggstrategy == AGG_SORTED || aggstrategy == AGG_MIXED)
2737 : {
2738 : /* Here we are able to deliver output on-the-fly */
2739 11100 : startup_cost = input_startup_cost;
2740 11100 : total_cost = input_total_cost;
2741 11100 : if (aggstrategy == AGG_MIXED && !enable_hashagg)
2742 444 : ++disabled_nodes;
2743 : /* calcs phrased this way to match HASHED case, see note above */
2744 11100 : total_cost += aggcosts->transCost.startup;
2745 11100 : total_cost += aggcosts->transCost.per_tuple * input_tuples;
2746 11100 : total_cost += (cpu_operator_cost * numGroupCols) * input_tuples;
2747 11100 : total_cost += aggcosts->finalCost.startup;
2748 11100 : total_cost += aggcosts->finalCost.per_tuple * numGroups;
2749 11100 : total_cost += cpu_tuple_cost * numGroups;
2750 11100 : output_tuples = numGroups;
2751 : }
2752 : else
2753 : {
2754 : /* must be AGG_HASHED */
2755 17456 : startup_cost = input_total_cost;
2756 17456 : if (!enable_hashagg)
2757 1554 : ++disabled_nodes;
2758 17456 : startup_cost += aggcosts->transCost.startup;
2759 17456 : startup_cost += aggcosts->transCost.per_tuple * input_tuples;
2760 : /* cost of computing hash value */
2761 17456 : startup_cost += (cpu_operator_cost * numGroupCols) * input_tuples;
2762 17456 : startup_cost += aggcosts->finalCost.startup;
2763 :
2764 17456 : total_cost = startup_cost;
2765 17456 : total_cost += aggcosts->finalCost.per_tuple * numGroups;
2766 : /* cost of retrieving from hash table */
2767 17456 : total_cost += cpu_tuple_cost * numGroups;
2768 17456 : output_tuples = numGroups;
2769 : }
2770 :
2771 : /*
2772 : * Add the disk costs of hash aggregation that spills to disk.
2773 : *
2774 : * Groups that go into the hash table stay in memory until finalized, so
2775 : * spilling and reprocessing tuples doesn't incur additional invocations
2776 : * of transCost or finalCost. Furthermore, the computed hash value is
2777 : * stored with the spilled tuples, so we don't incur extra invocations of
2778 : * the hash function.
2779 : *
2780 : * Hash Agg begins returning tuples after the first batch is complete.
2781 : * Accrue writes (spilled tuples) to startup_cost and to total_cost;
2782 : * accrue reads only to total_cost.
2783 : */
2784 63800 : if (aggstrategy == AGG_HASHED || aggstrategy == AGG_MIXED)
2785 : {
2786 : double pages;
2787 18348 : double pages_written = 0.0;
2788 18348 : double pages_read = 0.0;
2789 : double spill_cost;
2790 : double hashentrysize;
2791 : double nbatches;
2792 : Size mem_limit;
2793 : uint64 ngroups_limit;
2794 : int num_partitions;
2795 : int depth;
2796 :
2797 : /*
2798 : * Estimate number of batches based on the computed limits. If less
2799 : * than or equal to one, all groups are expected to fit in memory;
2800 : * otherwise we expect to spill.
2801 : */
2802 18348 : hashentrysize = hash_agg_entry_size(list_length(root->aggtransinfos),
2803 : input_width,
2804 : aggcosts->transitionSpace);
2805 18348 : hash_agg_set_limits(hashentrysize, numGroups, 0, &mem_limit,
2806 : &ngroups_limit, &num_partitions);
2807 :
2808 18348 : nbatches = Max((numGroups * hashentrysize) / mem_limit,
2809 : numGroups / ngroups_limit);
2810 :
2811 18348 : nbatches = Max(ceil(nbatches), 1.0);
2812 18348 : num_partitions = Max(num_partitions, 2);
2813 :
2814 : /*
2815 : * The number of partitions can change at different levels of
2816 : * recursion; but for the purposes of this calculation assume it stays
2817 : * constant.
2818 : */
2819 18348 : depth = ceil(log(nbatches) / log(num_partitions));
2820 :
2821 : /*
2822 : * Estimate number of pages read and written. For each level of
2823 : * recursion, a tuple must be written and then later read.
2824 : */
2825 18348 : pages = relation_byte_size(input_tuples, input_width) / BLCKSZ;
2826 18348 : pages_written = pages_read = pages * depth;
2827 :
2828 : /*
2829 : * HashAgg has somewhat worse IO behavior than Sort on typical
2830 : * hardware/OS combinations. Account for this with a generic penalty.
2831 : */
2832 18348 : pages_read *= 2.0;
2833 18348 : pages_written *= 2.0;
2834 :
2835 18348 : startup_cost += pages_written * random_page_cost;
2836 18348 : total_cost += pages_written * random_page_cost;
2837 18348 : total_cost += pages_read * seq_page_cost;
2838 :
2839 : /* account for CPU cost of spilling a tuple and reading it back */
2840 18348 : spill_cost = depth * input_tuples * 2.0 * cpu_tuple_cost;
2841 18348 : startup_cost += spill_cost;
2842 18348 : total_cost += spill_cost;
2843 : }
2844 :
2845 : /*
2846 : * If there are quals (HAVING quals), account for their cost and
2847 : * selectivity.
2848 : */
2849 63800 : if (quals)
2850 : {
2851 : QualCost qual_cost;
2852 :
2853 4488 : cost_qual_eval(&qual_cost, quals, root);
2854 4488 : startup_cost += qual_cost.startup;
2855 4488 : total_cost += qual_cost.startup + output_tuples * qual_cost.per_tuple;
2856 :
2857 4488 : output_tuples = clamp_row_est(output_tuples *
2858 4488 : clauselist_selectivity(root,
2859 : quals,
2860 : 0,
2861 : JOIN_INNER,
2862 : NULL));
2863 : }
2864 :
2865 63800 : path->rows = output_tuples;
2866 63800 : path->disabled_nodes = disabled_nodes;
2867 63800 : path->startup_cost = startup_cost;
2868 63800 : path->total_cost = total_cost;
2869 63800 : }
2870 :
2871 : /*
2872 : * get_windowclause_startup_tuples
2873 : * Estimate how many tuples we'll need to fetch from a WindowAgg's
2874 : * subnode before we can output the first WindowAgg tuple.
2875 : *
2876 : * How many tuples need to be read depends on the WindowClause. For example,
2877 : * a WindowClause with no PARTITION BY and no ORDER BY requires that all
2878 : * subnode tuples are read and aggregated before the WindowAgg can output
2879 : * anything. If there's a PARTITION BY, then we only need to look at tuples
2880 : * in the first partition. Here we attempt to estimate just how many
2881 : * 'input_tuples' the WindowAgg will need to read for the given WindowClause
2882 : * before the first tuple can be output.
2883 : */
2884 : static double
2885 2684 : get_windowclause_startup_tuples(PlannerInfo *root, WindowClause *wc,
2886 : double input_tuples)
2887 : {
2888 2684 : int frameOptions = wc->frameOptions;
2889 : double partition_tuples;
2890 : double return_tuples;
2891 : double peer_tuples;
2892 :
2893 : /*
2894 : * First, figure out how many partitions there are likely to be and set
2895 : * partition_tuples according to that estimate.
2896 : */
2897 2684 : if (wc->partitionClause != NIL)
2898 : {
2899 : double num_partitions;
2900 656 : List *partexprs = get_sortgrouplist_exprs(wc->partitionClause,
2901 656 : root->parse->targetList);
2902 :
2903 656 : num_partitions = estimate_num_groups(root, partexprs, input_tuples,
2904 : NULL, NULL);
2905 656 : list_free(partexprs);
2906 :
2907 656 : partition_tuples = input_tuples / num_partitions;
2908 : }
2909 : else
2910 : {
2911 : /* all tuples belong to the same partition */
2912 2028 : partition_tuples = input_tuples;
2913 : }
2914 :
2915 : /* estimate the number of tuples in each peer group */
2916 2684 : if (wc->orderClause != NIL)
2917 : {
2918 : double num_groups;
2919 : List *orderexprs;
2920 :
2921 2234 : orderexprs = get_sortgrouplist_exprs(wc->orderClause,
2922 2234 : root->parse->targetList);
2923 :
2924 : /* estimate out how many peer groups there are in the partition */
2925 2234 : num_groups = estimate_num_groups(root, orderexprs,
2926 : partition_tuples, NULL,
2927 : NULL);
2928 2234 : list_free(orderexprs);
2929 2234 : peer_tuples = partition_tuples / num_groups;
2930 : }
2931 : else
2932 : {
2933 : /* no ORDER BY so only 1 tuple belongs in each peer group */
2934 450 : peer_tuples = 1.0;
2935 : }
2936 :
2937 2684 : if (frameOptions & FRAMEOPTION_END_UNBOUNDED_FOLLOWING)
2938 : {
2939 : /* include all partition rows */
2940 346 : return_tuples = partition_tuples;
2941 : }
2942 2338 : else if (frameOptions & FRAMEOPTION_END_CURRENT_ROW)
2943 : {
2944 1348 : if (frameOptions & FRAMEOPTION_ROWS)
2945 : {
2946 : /* just count the current row */
2947 606 : return_tuples = 1.0;
2948 : }
2949 742 : else if (frameOptions & (FRAMEOPTION_RANGE | FRAMEOPTION_GROUPS))
2950 : {
2951 : /*
2952 : * When in RANGE/GROUPS mode, it's more complex. If there's no
2953 : * ORDER BY, then all rows in the partition are peers, otherwise
2954 : * we'll need to read the first group of peers.
2955 : */
2956 742 : if (wc->orderClause == NIL)
2957 282 : return_tuples = partition_tuples;
2958 : else
2959 460 : return_tuples = peer_tuples;
2960 : }
2961 : else
2962 : {
2963 : /*
2964 : * Something new we don't support yet? This needs attention.
2965 : * We'll just return 1.0 in the meantime.
2966 : */
2967 : Assert(false);
2968 0 : return_tuples = 1.0;
2969 : }
2970 : }
2971 990 : else if (frameOptions & FRAMEOPTION_END_OFFSET_PRECEDING)
2972 : {
2973 : /*
2974 : * BETWEEN ... AND N PRECEDING will only need to read the WindowAgg's
2975 : * subnode after N ROWS/RANGES/GROUPS. N can be 0, but not negative,
2976 : * so we'll just assume only the current row needs to be read to fetch
2977 : * the first WindowAgg row.
2978 : */
2979 108 : return_tuples = 1.0;
2980 : }
2981 882 : else if (frameOptions & FRAMEOPTION_END_OFFSET_FOLLOWING)
2982 : {
2983 882 : Const *endOffset = (Const *) wc->endOffset;
2984 : double end_offset_value;
2985 :
2986 : /* try and figure out the value specified in the endOffset. */
2987 882 : if (IsA(endOffset, Const))
2988 : {
2989 870 : if (endOffset->constisnull)
2990 : {
2991 : /*
2992 : * NULLs are not allowed, but currently, there's no code to
2993 : * error out if there's a NULL Const. We'll only discover
2994 : * this during execution. For now, just pretend everything is
2995 : * fine and assume that just the first row/range/group will be
2996 : * needed.
2997 : */
2998 0 : end_offset_value = 1.0;
2999 : }
3000 : else
3001 : {
3002 870 : switch (endOffset->consttype)
3003 : {
3004 24 : case INT2OID:
3005 24 : end_offset_value =
3006 24 : (double) DatumGetInt16(endOffset->constvalue);
3007 24 : break;
3008 132 : case INT4OID:
3009 132 : end_offset_value =
3010 132 : (double) DatumGetInt32(endOffset->constvalue);
3011 132 : break;
3012 372 : case INT8OID:
3013 372 : end_offset_value =
3014 372 : (double) DatumGetInt64(endOffset->constvalue);
3015 372 : break;
3016 342 : default:
3017 342 : end_offset_value =
3018 342 : partition_tuples / peer_tuples *
3019 : DEFAULT_INEQ_SEL;
3020 342 : break;
3021 : }
3022 : }
3023 : }
3024 : else
3025 : {
3026 : /*
3027 : * When the end bound is not a Const, we'll just need to guess. We
3028 : * just make use of DEFAULT_INEQ_SEL.
3029 : */
3030 12 : end_offset_value =
3031 12 : partition_tuples / peer_tuples * DEFAULT_INEQ_SEL;
3032 : }
3033 :
3034 882 : if (frameOptions & FRAMEOPTION_ROWS)
3035 : {
3036 : /* include the N FOLLOWING and the current row */
3037 222 : return_tuples = end_offset_value + 1.0;
3038 : }
3039 660 : else if (frameOptions & (FRAMEOPTION_RANGE | FRAMEOPTION_GROUPS))
3040 : {
3041 : /* include N FOLLOWING ranges/group and the initial range/group */
3042 660 : return_tuples = peer_tuples * (end_offset_value + 1.0);
3043 : }
3044 : else
3045 : {
3046 : /*
3047 : * Something new we don't support yet? This needs attention.
3048 : * We'll just return 1.0 in the meantime.
3049 : */
3050 : Assert(false);
3051 0 : return_tuples = 1.0;
3052 : }
3053 : }
3054 : else
3055 : {
3056 : /*
3057 : * Something new we don't support yet? This needs attention. We'll
3058 : * just return 1.0 in the meantime.
3059 : */
3060 : Assert(false);
3061 0 : return_tuples = 1.0;
3062 : }
3063 :
3064 2684 : if (wc->partitionClause != NIL || wc->orderClause != NIL)
3065 : {
3066 : /*
3067 : * Cap the return value to the estimated partition tuples and account
3068 : * for the extra tuple WindowAgg will need to read to confirm the next
3069 : * tuple does not belong to the same partition or peer group.
3070 : */
3071 2408 : return_tuples = Min(return_tuples + 1.0, partition_tuples);
3072 : }
3073 : else
3074 : {
3075 : /*
3076 : * Cap the return value so it's never higher than the expected tuples
3077 : * in the partition.
3078 : */
3079 276 : return_tuples = Min(return_tuples, partition_tuples);
3080 : }
3081 :
3082 : /*
3083 : * We needn't worry about any EXCLUDE options as those only exclude rows
3084 : * from being aggregated, not from being read from the WindowAgg's
3085 : * subnode.
3086 : */
3087 :
3088 2684 : return clamp_row_est(return_tuples);
3089 : }
3090 :
3091 : /*
3092 : * cost_windowagg
3093 : * Determines and returns the cost of performing a WindowAgg plan node,
3094 : * including the cost of its input.
3095 : *
3096 : * Input is assumed already properly sorted.
3097 : */
3098 : void
3099 2684 : cost_windowagg(Path *path, PlannerInfo *root,
3100 : List *windowFuncs, WindowClause *winclause,
3101 : int input_disabled_nodes,
3102 : Cost input_startup_cost, Cost input_total_cost,
3103 : double input_tuples)
3104 : {
3105 : Cost startup_cost;
3106 : Cost total_cost;
3107 : double startup_tuples;
3108 : int numPartCols;
3109 : int numOrderCols;
3110 : ListCell *lc;
3111 :
3112 2684 : numPartCols = list_length(winclause->partitionClause);
3113 2684 : numOrderCols = list_length(winclause->orderClause);
3114 :
3115 2684 : startup_cost = input_startup_cost;
3116 2684 : total_cost = input_total_cost;
3117 :
3118 : /*
3119 : * Window functions are assumed to cost their stated execution cost, plus
3120 : * the cost of evaluating their input expressions, per tuple. Since they
3121 : * may in fact evaluate their inputs at multiple rows during each cycle,
3122 : * this could be a drastic underestimate; but without a way to know how
3123 : * many rows the window function will fetch, it's hard to do better. In
3124 : * any case, it's a good estimate for all the built-in window functions,
3125 : * so we'll just do this for now.
3126 : */
3127 6076 : foreach(lc, windowFuncs)
3128 : {
3129 3392 : WindowFunc *wfunc = lfirst_node(WindowFunc, lc);
3130 : Cost wfunccost;
3131 : QualCost argcosts;
3132 :
3133 3392 : argcosts.startup = argcosts.per_tuple = 0;
3134 3392 : add_function_cost(root, wfunc->winfnoid, (Node *) wfunc,
3135 : &argcosts);
3136 3392 : startup_cost += argcosts.startup;
3137 3392 : wfunccost = argcosts.per_tuple;
3138 :
3139 : /* also add the input expressions' cost to per-input-row costs */
3140 3392 : cost_qual_eval_node(&argcosts, (Node *) wfunc->args, root);
3141 3392 : startup_cost += argcosts.startup;
3142 3392 : wfunccost += argcosts.per_tuple;
3143 :
3144 : /*
3145 : * Add the filter's cost to per-input-row costs. XXX We should reduce
3146 : * input expression costs according to filter selectivity.
3147 : */
3148 3392 : cost_qual_eval_node(&argcosts, (Node *) wfunc->aggfilter, root);
3149 3392 : startup_cost += argcosts.startup;
3150 3392 : wfunccost += argcosts.per_tuple;
3151 :
3152 3392 : total_cost += wfunccost * input_tuples;
3153 : }
3154 :
3155 : /*
3156 : * We also charge cpu_operator_cost per grouping column per tuple for
3157 : * grouping comparisons, plus cpu_tuple_cost per tuple for general
3158 : * overhead.
3159 : *
3160 : * XXX this neglects costs of spooling the data to disk when it overflows
3161 : * work_mem. Sooner or later that should get accounted for.
3162 : */
3163 2684 : total_cost += cpu_operator_cost * (numPartCols + numOrderCols) * input_tuples;
3164 2684 : total_cost += cpu_tuple_cost * input_tuples;
3165 :
3166 2684 : path->rows = input_tuples;
3167 2684 : path->disabled_nodes = input_disabled_nodes;
3168 2684 : path->startup_cost = startup_cost;
3169 2684 : path->total_cost = total_cost;
3170 :
3171 : /*
3172 : * Also, take into account how many tuples we need to read from the
3173 : * subnode in order to produce the first tuple from the WindowAgg. To do
3174 : * this we proportion the run cost (total cost not including startup cost)
3175 : * over the estimated startup tuples. We already included the startup
3176 : * cost of the subnode, so we only need to do this when the estimated
3177 : * startup tuples is above 1.0.
3178 : */
3179 2684 : startup_tuples = get_windowclause_startup_tuples(root, winclause,
3180 : input_tuples);
3181 :
3182 2684 : if (startup_tuples > 1.0)
3183 2406 : path->startup_cost += (total_cost - startup_cost) / input_tuples *
3184 2406 : (startup_tuples - 1.0);
3185 2684 : }
3186 :
3187 : /*
3188 : * cost_group
3189 : * Determines and returns the cost of performing a Group plan node,
3190 : * including the cost of its input.
3191 : *
3192 : * Note: caller must ensure that input costs are for appropriately-sorted
3193 : * input.
3194 : */
3195 : void
3196 1184 : cost_group(Path *path, PlannerInfo *root,
3197 : int numGroupCols, double numGroups,
3198 : List *quals,
3199 : int input_disabled_nodes,
3200 : Cost input_startup_cost, Cost input_total_cost,
3201 : double input_tuples)
3202 : {
3203 : double output_tuples;
3204 : Cost startup_cost;
3205 : Cost total_cost;
3206 :
3207 1184 : output_tuples = numGroups;
3208 1184 : startup_cost = input_startup_cost;
3209 1184 : total_cost = input_total_cost;
3210 :
3211 : /*
3212 : * Charge one cpu_operator_cost per comparison per input tuple. We assume
3213 : * all columns get compared at most of the tuples.
3214 : */
3215 1184 : total_cost += cpu_operator_cost * input_tuples * numGroupCols;
3216 :
3217 : /*
3218 : * If there are quals (HAVING quals), account for their cost and
3219 : * selectivity.
3220 : */
3221 1184 : if (quals)
3222 : {
3223 : QualCost qual_cost;
3224 :
3225 0 : cost_qual_eval(&qual_cost, quals, root);
3226 0 : startup_cost += qual_cost.startup;
3227 0 : total_cost += qual_cost.startup + output_tuples * qual_cost.per_tuple;
3228 :
3229 0 : output_tuples = clamp_row_est(output_tuples *
3230 0 : clauselist_selectivity(root,
3231 : quals,
3232 : 0,
3233 : JOIN_INNER,
3234 : NULL));
3235 : }
3236 :
3237 1184 : path->rows = output_tuples;
3238 1184 : path->disabled_nodes = input_disabled_nodes;
3239 1184 : path->startup_cost = startup_cost;
3240 1184 : path->total_cost = total_cost;
3241 1184 : }
3242 :
3243 : /*
3244 : * initial_cost_nestloop
3245 : * Preliminary estimate of the cost of a nestloop join path.
3246 : *
3247 : * This must quickly produce lower-bound estimates of the path's startup and
3248 : * total costs. If we are unable to eliminate the proposed path from
3249 : * consideration using the lower bounds, final_cost_nestloop will be called
3250 : * to obtain the final estimates.
3251 : *
3252 : * The exact division of labor between this function and final_cost_nestloop
3253 : * is private to them, and represents a tradeoff between speed of the initial
3254 : * estimate and getting a tight lower bound. We choose to not examine the
3255 : * join quals here, since that's by far the most expensive part of the
3256 : * calculations. The end result is that CPU-cost considerations must be
3257 : * left for the second phase; and for SEMI/ANTI joins, we must also postpone
3258 : * incorporation of the inner path's run cost.
3259 : *
3260 : * 'workspace' is to be filled with startup_cost, total_cost, and perhaps
3261 : * other data to be used by final_cost_nestloop
3262 : * 'jointype' is the type of join to be performed
3263 : * 'outer_path' is the outer input to the join
3264 : * 'inner_path' is the inner input to the join
3265 : * 'extra' contains miscellaneous information about the join
3266 : */
3267 : void
3268 2328140 : initial_cost_nestloop(PlannerInfo *root, JoinCostWorkspace *workspace,
3269 : JoinType jointype,
3270 : Path *outer_path, Path *inner_path,
3271 : JoinPathExtraData *extra)
3272 : {
3273 : int disabled_nodes;
3274 2328140 : Cost startup_cost = 0;
3275 2328140 : Cost run_cost = 0;
3276 2328140 : double outer_path_rows = outer_path->rows;
3277 : Cost inner_rescan_start_cost;
3278 : Cost inner_rescan_total_cost;
3279 : Cost inner_run_cost;
3280 : Cost inner_rescan_run_cost;
3281 :
3282 : /* Count up disabled nodes. */
3283 2328140 : disabled_nodes = enable_nestloop ? 0 : 1;
3284 2328140 : disabled_nodes += inner_path->disabled_nodes;
3285 2328140 : disabled_nodes += outer_path->disabled_nodes;
3286 :
3287 : /* estimate costs to rescan the inner relation */
3288 2328140 : cost_rescan(root, inner_path,
3289 : &inner_rescan_start_cost,
3290 : &inner_rescan_total_cost);
3291 :
3292 : /* cost of source data */
3293 :
3294 : /*
3295 : * NOTE: clearly, we must pay both outer and inner paths' startup_cost
3296 : * before we can start returning tuples, so the join's startup cost is
3297 : * their sum. We'll also pay the inner path's rescan startup cost
3298 : * multiple times.
3299 : */
3300 2328140 : startup_cost += outer_path->startup_cost + inner_path->startup_cost;
3301 2328140 : run_cost += outer_path->total_cost - outer_path->startup_cost;
3302 2328140 : if (outer_path_rows > 1)
3303 1597016 : run_cost += (outer_path_rows - 1) * inner_rescan_start_cost;
3304 :
3305 2328140 : inner_run_cost = inner_path->total_cost - inner_path->startup_cost;
3306 2328140 : inner_rescan_run_cost = inner_rescan_total_cost - inner_rescan_start_cost;
3307 :
3308 2328140 : if (jointype == JOIN_SEMI || jointype == JOIN_ANTI ||
3309 2287342 : extra->inner_unique)
3310 : {
3311 : /*
3312 : * With a SEMI or ANTI join, or if the innerrel is known unique, the
3313 : * executor will stop after the first match.
3314 : *
3315 : * Getting decent estimates requires inspection of the join quals,
3316 : * which we choose to postpone to final_cost_nestloop.
3317 : */
3318 :
3319 : /* Save private data for final_cost_nestloop */
3320 1064920 : workspace->inner_run_cost = inner_run_cost;
3321 1064920 : workspace->inner_rescan_run_cost = inner_rescan_run_cost;
3322 : }
3323 : else
3324 : {
3325 : /* Normal case; we'll scan whole input rel for each outer row */
3326 1263220 : run_cost += inner_run_cost;
3327 1263220 : if (outer_path_rows > 1)
3328 876658 : run_cost += (outer_path_rows - 1) * inner_rescan_run_cost;
3329 : }
3330 :
3331 : /* CPU costs left for later */
3332 :
3333 : /* Public result fields */
3334 2328140 : workspace->disabled_nodes = disabled_nodes;
3335 2328140 : workspace->startup_cost = startup_cost;
3336 2328140 : workspace->total_cost = startup_cost + run_cost;
3337 : /* Save private data for final_cost_nestloop */
3338 2328140 : workspace->run_cost = run_cost;
3339 2328140 : }
3340 :
3341 : /*
3342 : * final_cost_nestloop
3343 : * Final estimate of the cost and result size of a nestloop join path.
3344 : *
3345 : * 'path' is already filled in except for the rows and cost fields
3346 : * 'workspace' is the result from initial_cost_nestloop
3347 : * 'extra' contains miscellaneous information about the join
3348 : */
3349 : void
3350 1150978 : final_cost_nestloop(PlannerInfo *root, NestPath *path,
3351 : JoinCostWorkspace *workspace,
3352 : JoinPathExtraData *extra)
3353 : {
3354 1150978 : Path *outer_path = path->jpath.outerjoinpath;
3355 1150978 : Path *inner_path = path->jpath.innerjoinpath;
3356 1150978 : double outer_path_rows = outer_path->rows;
3357 1150978 : double inner_path_rows = inner_path->rows;
3358 1150978 : Cost startup_cost = workspace->startup_cost;
3359 1150978 : Cost run_cost = workspace->run_cost;
3360 : Cost cpu_per_tuple;
3361 : QualCost restrict_qual_cost;
3362 : double ntuples;
3363 :
3364 : /* Set the number of disabled nodes. */
3365 1150978 : path->jpath.path.disabled_nodes = workspace->disabled_nodes;
3366 :
3367 : /* Protect some assumptions below that rowcounts aren't zero */
3368 1150978 : if (outer_path_rows <= 0)
3369 0 : outer_path_rows = 1;
3370 1150978 : if (inner_path_rows <= 0)
3371 690 : inner_path_rows = 1;
3372 : /* Mark the path with the correct row estimate */
3373 1150978 : if (path->jpath.path.param_info)
3374 24894 : path->jpath.path.rows = path->jpath.path.param_info->ppi_rows;
3375 : else
3376 1126084 : path->jpath.path.rows = path->jpath.path.parent->rows;
3377 :
3378 : /* For partial paths, scale row estimate. */
3379 1150978 : if (path->jpath.path.parallel_workers > 0)
3380 : {
3381 12246 : double parallel_divisor = get_parallel_divisor(&path->jpath.path);
3382 :
3383 12246 : path->jpath.path.rows =
3384 12246 : clamp_row_est(path->jpath.path.rows / parallel_divisor);
3385 : }
3386 :
3387 : /* cost of inner-relation source data (we already dealt with outer rel) */
3388 :
3389 1150978 : if (path->jpath.jointype == JOIN_SEMI || path->jpath.jointype == JOIN_ANTI ||
3390 1124338 : extra->inner_unique)
3391 732470 : {
3392 : /*
3393 : * With a SEMI or ANTI join, or if the innerrel is known unique, the
3394 : * executor will stop after the first match.
3395 : */
3396 732470 : Cost inner_run_cost = workspace->inner_run_cost;
3397 732470 : Cost inner_rescan_run_cost = workspace->inner_rescan_run_cost;
3398 : double outer_matched_rows;
3399 : double outer_unmatched_rows;
3400 : Selectivity inner_scan_frac;
3401 :
3402 : /*
3403 : * For an outer-rel row that has at least one match, we can expect the
3404 : * inner scan to stop after a fraction 1/(match_count+1) of the inner
3405 : * rows, if the matches are evenly distributed. Since they probably
3406 : * aren't quite evenly distributed, we apply a fuzz factor of 2.0 to
3407 : * that fraction. (If we used a larger fuzz factor, we'd have to
3408 : * clamp inner_scan_frac to at most 1.0; but since match_count is at
3409 : * least 1, no such clamp is needed now.)
3410 : */
3411 732470 : outer_matched_rows = rint(outer_path_rows * extra->semifactors.outer_match_frac);
3412 732470 : outer_unmatched_rows = outer_path_rows - outer_matched_rows;
3413 732470 : inner_scan_frac = 2.0 / (extra->semifactors.match_count + 1.0);
3414 :
3415 : /*
3416 : * Compute number of tuples processed (not number emitted!). First,
3417 : * account for successfully-matched outer rows.
3418 : */
3419 732470 : ntuples = outer_matched_rows * inner_path_rows * inner_scan_frac;
3420 :
3421 : /*
3422 : * Now we need to estimate the actual costs of scanning the inner
3423 : * relation, which may be quite a bit less than N times inner_run_cost
3424 : * due to early scan stops. We consider two cases. If the inner path
3425 : * is an indexscan using all the joinquals as indexquals, then an
3426 : * unmatched outer row results in an indexscan returning no rows,
3427 : * which is probably quite cheap. Otherwise, the executor will have
3428 : * to scan the whole inner rel for an unmatched row; not so cheap.
3429 : */
3430 732470 : if (has_indexed_join_quals(path))
3431 : {
3432 : /*
3433 : * Successfully-matched outer rows will only require scanning
3434 : * inner_scan_frac of the inner relation. In this case, we don't
3435 : * need to charge the full inner_run_cost even when that's more
3436 : * than inner_rescan_run_cost, because we can assume that none of
3437 : * the inner scans ever scan the whole inner relation. So it's
3438 : * okay to assume that all the inner scan executions can be
3439 : * fractions of the full cost, even if materialization is reducing
3440 : * the rescan cost. At this writing, it's impossible to get here
3441 : * for a materialized inner scan, so inner_run_cost and
3442 : * inner_rescan_run_cost will be the same anyway; but just in
3443 : * case, use inner_run_cost for the first matched tuple and
3444 : * inner_rescan_run_cost for additional ones.
3445 : */
3446 128512 : run_cost += inner_run_cost * inner_scan_frac;
3447 128512 : if (outer_matched_rows > 1)
3448 16566 : run_cost += (outer_matched_rows - 1) * inner_rescan_run_cost * inner_scan_frac;
3449 :
3450 : /*
3451 : * Add the cost of inner-scan executions for unmatched outer rows.
3452 : * We estimate this as the same cost as returning the first tuple
3453 : * of a nonempty scan. We consider that these are all rescans,
3454 : * since we used inner_run_cost once already.
3455 : */
3456 128512 : run_cost += outer_unmatched_rows *
3457 128512 : inner_rescan_run_cost / inner_path_rows;
3458 :
3459 : /*
3460 : * We won't be evaluating any quals at all for unmatched rows, so
3461 : * don't add them to ntuples.
3462 : */
3463 : }
3464 : else
3465 : {
3466 : /*
3467 : * Here, a complicating factor is that rescans may be cheaper than
3468 : * first scans. If we never scan all the way to the end of the
3469 : * inner rel, it might be (depending on the plan type) that we'd
3470 : * never pay the whole inner first-scan run cost. However it is
3471 : * difficult to estimate whether that will happen (and it could
3472 : * not happen if there are any unmatched outer rows!), so be
3473 : * conservative and always charge the whole first-scan cost once.
3474 : * We consider this charge to correspond to the first unmatched
3475 : * outer row, unless there isn't one in our estimate, in which
3476 : * case blame it on the first matched row.
3477 : */
3478 :
3479 : /* First, count all unmatched join tuples as being processed */
3480 603958 : ntuples += outer_unmatched_rows * inner_path_rows;
3481 :
3482 : /* Now add the forced full scan, and decrement appropriate count */
3483 603958 : run_cost += inner_run_cost;
3484 603958 : if (outer_unmatched_rows >= 1)
3485 588416 : outer_unmatched_rows -= 1;
3486 : else
3487 15542 : outer_matched_rows -= 1;
3488 :
3489 : /* Add inner run cost for additional outer tuples having matches */
3490 603958 : if (outer_matched_rows > 0)
3491 208768 : run_cost += outer_matched_rows * inner_rescan_run_cost * inner_scan_frac;
3492 :
3493 : /* Add inner run cost for additional unmatched outer tuples */
3494 603958 : if (outer_unmatched_rows > 0)
3495 386984 : run_cost += outer_unmatched_rows * inner_rescan_run_cost;
3496 : }
3497 : }
3498 : else
3499 : {
3500 : /* Normal-case source costs were included in preliminary estimate */
3501 :
3502 : /* Compute number of tuples processed (not number emitted!) */
3503 418508 : ntuples = outer_path_rows * inner_path_rows;
3504 : }
3505 :
3506 : /* CPU costs */
3507 1150978 : cost_qual_eval(&restrict_qual_cost, path->jpath.joinrestrictinfo, root);
3508 1150978 : startup_cost += restrict_qual_cost.startup;
3509 1150978 : cpu_per_tuple = cpu_tuple_cost + restrict_qual_cost.per_tuple;
3510 1150978 : run_cost += cpu_per_tuple * ntuples;
3511 :
3512 : /* tlist eval costs are paid per output row, not per tuple scanned */
3513 1150978 : startup_cost += path->jpath.path.pathtarget->cost.startup;
3514 1150978 : run_cost += path->jpath.path.pathtarget->cost.per_tuple * path->jpath.path.rows;
3515 :
3516 1150978 : path->jpath.path.startup_cost = startup_cost;
3517 1150978 : path->jpath.path.total_cost = startup_cost + run_cost;
3518 1150978 : }
3519 :
3520 : /*
3521 : * initial_cost_mergejoin
3522 : * Preliminary estimate of the cost of a mergejoin path.
3523 : *
3524 : * This must quickly produce lower-bound estimates of the path's startup and
3525 : * total costs. If we are unable to eliminate the proposed path from
3526 : * consideration using the lower bounds, final_cost_mergejoin will be called
3527 : * to obtain the final estimates.
3528 : *
3529 : * The exact division of labor between this function and final_cost_mergejoin
3530 : * is private to them, and represents a tradeoff between speed of the initial
3531 : * estimate and getting a tight lower bound. We choose to not examine the
3532 : * join quals here, except for obtaining the scan selectivity estimate which
3533 : * is really essential (but fortunately, use of caching keeps the cost of
3534 : * getting that down to something reasonable).
3535 : * We also assume that cost_sort/cost_incremental_sort is cheap enough to use
3536 : * here.
3537 : *
3538 : * 'workspace' is to be filled with startup_cost, total_cost, and perhaps
3539 : * other data to be used by final_cost_mergejoin
3540 : * 'jointype' is the type of join to be performed
3541 : * 'mergeclauses' is the list of joinclauses to be used as merge clauses
3542 : * 'outer_path' is the outer input to the join
3543 : * 'inner_path' is the inner input to the join
3544 : * 'outersortkeys' is the list of sort keys for the outer path
3545 : * 'innersortkeys' is the list of sort keys for the inner path
3546 : * 'extra' contains miscellaneous information about the join
3547 : *
3548 : * Note: outersortkeys and innersortkeys should be NIL if no explicit
3549 : * sort is needed because the respective source path is already ordered.
3550 : */
3551 : void
3552 1007884 : initial_cost_mergejoin(PlannerInfo *root, JoinCostWorkspace *workspace,
3553 : JoinType jointype,
3554 : List *mergeclauses,
3555 : Path *outer_path, Path *inner_path,
3556 : List *outersortkeys, List *innersortkeys,
3557 : JoinPathExtraData *extra)
3558 : {
3559 : int disabled_nodes;
3560 1007884 : Cost startup_cost = 0;
3561 1007884 : Cost run_cost = 0;
3562 1007884 : double outer_path_rows = outer_path->rows;
3563 1007884 : double inner_path_rows = inner_path->rows;
3564 : Cost inner_run_cost;
3565 : double outer_rows,
3566 : inner_rows,
3567 : outer_skip_rows,
3568 : inner_skip_rows;
3569 : Selectivity outerstartsel,
3570 : outerendsel,
3571 : innerstartsel,
3572 : innerendsel;
3573 : Path sort_path; /* dummy for result of
3574 : * cost_sort/cost_incremental_sort */
3575 :
3576 : /* Protect some assumptions below that rowcounts aren't zero */
3577 1007884 : if (outer_path_rows <= 0)
3578 96 : outer_path_rows = 1;
3579 1007884 : if (inner_path_rows <= 0)
3580 126 : inner_path_rows = 1;
3581 :
3582 : /*
3583 : * A merge join will stop as soon as it exhausts either input stream
3584 : * (unless it's an outer join, in which case the outer side has to be
3585 : * scanned all the way anyway). Estimate fraction of the left and right
3586 : * inputs that will actually need to be scanned. Likewise, we can
3587 : * estimate the number of rows that will be skipped before the first join
3588 : * pair is found, which should be factored into startup cost. We use only
3589 : * the first (most significant) merge clause for this purpose. Since
3590 : * mergejoinscansel() is a fairly expensive computation, we cache the
3591 : * results in the merge clause RestrictInfo.
3592 : */
3593 1007884 : if (mergeclauses && jointype != JOIN_FULL)
3594 1001768 : {
3595 1001768 : RestrictInfo *firstclause = (RestrictInfo *) linitial(mergeclauses);
3596 : List *opathkeys;
3597 : List *ipathkeys;
3598 : PathKey *opathkey;
3599 : PathKey *ipathkey;
3600 : MergeScanSelCache *cache;
3601 :
3602 : /* Get the input pathkeys to determine the sort-order details */
3603 1001768 : opathkeys = outersortkeys ? outersortkeys : outer_path->pathkeys;
3604 1001768 : ipathkeys = innersortkeys ? innersortkeys : inner_path->pathkeys;
3605 : Assert(opathkeys);
3606 : Assert(ipathkeys);
3607 1001768 : opathkey = (PathKey *) linitial(opathkeys);
3608 1001768 : ipathkey = (PathKey *) linitial(ipathkeys);
3609 : /* debugging check */
3610 1001768 : if (opathkey->pk_opfamily != ipathkey->pk_opfamily ||
3611 1001768 : opathkey->pk_eclass->ec_collation != ipathkey->pk_eclass->ec_collation ||
3612 1001768 : opathkey->pk_strategy != ipathkey->pk_strategy ||
3613 1001768 : opathkey->pk_nulls_first != ipathkey->pk_nulls_first)
3614 0 : elog(ERROR, "left and right pathkeys do not match in mergejoin");
3615 :
3616 : /* Get the selectivity with caching */
3617 1001768 : cache = cached_scansel(root, firstclause, opathkey);
3618 :
3619 1001768 : if (bms_is_subset(firstclause->left_relids,
3620 1001768 : outer_path->parent->relids))
3621 : {
3622 : /* left side of clause is outer */
3623 534302 : outerstartsel = cache->leftstartsel;
3624 534302 : outerendsel = cache->leftendsel;
3625 534302 : innerstartsel = cache->rightstartsel;
3626 534302 : innerendsel = cache->rightendsel;
3627 : }
3628 : else
3629 : {
3630 : /* left side of clause is inner */
3631 467466 : outerstartsel = cache->rightstartsel;
3632 467466 : outerendsel = cache->rightendsel;
3633 467466 : innerstartsel = cache->leftstartsel;
3634 467466 : innerendsel = cache->leftendsel;
3635 : }
3636 1001768 : if (jointype == JOIN_LEFT ||
3637 : jointype == JOIN_ANTI)
3638 : {
3639 163058 : outerstartsel = 0.0;
3640 163058 : outerendsel = 1.0;
3641 : }
3642 838710 : else if (jointype == JOIN_RIGHT ||
3643 : jointype == JOIN_RIGHT_ANTI)
3644 : {
3645 166686 : innerstartsel = 0.0;
3646 166686 : innerendsel = 1.0;
3647 : }
3648 : }
3649 : else
3650 : {
3651 : /* cope with clauseless or full mergejoin */
3652 6116 : outerstartsel = innerstartsel = 0.0;
3653 6116 : outerendsel = innerendsel = 1.0;
3654 : }
3655 :
3656 : /*
3657 : * Convert selectivities to row counts. We force outer_rows and
3658 : * inner_rows to be at least 1, but the skip_rows estimates can be zero.
3659 : */
3660 1007884 : outer_skip_rows = rint(outer_path_rows * outerstartsel);
3661 1007884 : inner_skip_rows = rint(inner_path_rows * innerstartsel);
3662 1007884 : outer_rows = clamp_row_est(outer_path_rows * outerendsel);
3663 1007884 : inner_rows = clamp_row_est(inner_path_rows * innerendsel);
3664 :
3665 : Assert(outer_skip_rows <= outer_rows);
3666 : Assert(inner_skip_rows <= inner_rows);
3667 :
3668 : /*
3669 : * Readjust scan selectivities to account for above rounding. This is
3670 : * normally an insignificant effect, but when there are only a few rows in
3671 : * the inputs, failing to do this makes for a large percentage error.
3672 : */
3673 1007884 : outerstartsel = outer_skip_rows / outer_path_rows;
3674 1007884 : innerstartsel = inner_skip_rows / inner_path_rows;
3675 1007884 : outerendsel = outer_rows / outer_path_rows;
3676 1007884 : innerendsel = inner_rows / inner_path_rows;
3677 :
3678 : Assert(outerstartsel <= outerendsel);
3679 : Assert(innerstartsel <= innerendsel);
3680 :
3681 1007884 : disabled_nodes = enable_mergejoin ? 0 : 1;
3682 :
3683 : /* cost of source data */
3684 :
3685 1007884 : if (outersortkeys) /* do we need to sort outer? */
3686 : {
3687 480668 : bool use_incremental_sort = false;
3688 : int presorted_keys;
3689 :
3690 : /*
3691 : * We choose to use incremental sort if it is enabled and there are
3692 : * presorted keys; otherwise we use full sort.
3693 : */
3694 480668 : if (enable_incremental_sort)
3695 : {
3696 : bool is_sorted PG_USED_FOR_ASSERTS_ONLY;
3697 :
3698 479738 : is_sorted = pathkeys_count_contained_in(outersortkeys,
3699 : outer_path->pathkeys,
3700 : &presorted_keys);
3701 : Assert(!is_sorted);
3702 :
3703 479738 : if (presorted_keys > 0)
3704 1442 : use_incremental_sort = true;
3705 : }
3706 :
3707 480668 : if (!use_incremental_sort)
3708 : {
3709 479226 : cost_sort(&sort_path,
3710 : root,
3711 : outersortkeys,
3712 : outer_path->disabled_nodes,
3713 : outer_path->total_cost,
3714 : outer_path_rows,
3715 479226 : outer_path->pathtarget->width,
3716 : 0.0,
3717 : work_mem,
3718 : -1.0);
3719 : }
3720 : else
3721 : {
3722 1442 : cost_incremental_sort(&sort_path,
3723 : root,
3724 : outersortkeys,
3725 : presorted_keys,
3726 : outer_path->disabled_nodes,
3727 : outer_path->startup_cost,
3728 : outer_path->total_cost,
3729 : outer_path_rows,
3730 1442 : outer_path->pathtarget->width,
3731 : 0.0,
3732 : work_mem,
3733 : -1.0);
3734 : }
3735 480668 : disabled_nodes += sort_path.disabled_nodes;
3736 480668 : startup_cost += sort_path.startup_cost;
3737 480668 : startup_cost += (sort_path.total_cost - sort_path.startup_cost)
3738 480668 : * outerstartsel;
3739 480668 : run_cost += (sort_path.total_cost - sort_path.startup_cost)
3740 480668 : * (outerendsel - outerstartsel);
3741 : }
3742 : else
3743 : {
3744 527216 : disabled_nodes += outer_path->disabled_nodes;
3745 527216 : startup_cost += outer_path->startup_cost;
3746 527216 : startup_cost += (outer_path->total_cost - outer_path->startup_cost)
3747 527216 : * outerstartsel;
3748 527216 : run_cost += (outer_path->total_cost - outer_path->startup_cost)
3749 527216 : * (outerendsel - outerstartsel);
3750 : }
3751 :
3752 1007884 : if (innersortkeys) /* do we need to sort inner? */
3753 : {
3754 : /*
3755 : * We do not consider incremental sort for inner path, because
3756 : * incremental sort does not support mark/restore.
3757 : */
3758 :
3759 785714 : cost_sort(&sort_path,
3760 : root,
3761 : innersortkeys,
3762 : inner_path->disabled_nodes,
3763 : inner_path->total_cost,
3764 : inner_path_rows,
3765 785714 : inner_path->pathtarget->width,
3766 : 0.0,
3767 : work_mem,
3768 : -1.0);
3769 785714 : disabled_nodes += sort_path.disabled_nodes;
3770 785714 : startup_cost += sort_path.startup_cost;
3771 785714 : startup_cost += (sort_path.total_cost - sort_path.startup_cost)
3772 785714 : * innerstartsel;
3773 785714 : inner_run_cost = (sort_path.total_cost - sort_path.startup_cost)
3774 785714 : * (innerendsel - innerstartsel);
3775 : }
3776 : else
3777 : {
3778 222170 : disabled_nodes += inner_path->disabled_nodes;
3779 222170 : startup_cost += inner_path->startup_cost;
3780 222170 : startup_cost += (inner_path->total_cost - inner_path->startup_cost)
3781 222170 : * innerstartsel;
3782 222170 : inner_run_cost = (inner_path->total_cost - inner_path->startup_cost)
3783 222170 : * (innerendsel - innerstartsel);
3784 : }
3785 :
3786 : /*
3787 : * We can't yet determine whether rescanning occurs, or whether
3788 : * materialization of the inner input should be done. The minimum
3789 : * possible inner input cost, regardless of rescan and materialization
3790 : * considerations, is inner_run_cost. We include that in
3791 : * workspace->total_cost, but not yet in run_cost.
3792 : */
3793 :
3794 : /* CPU costs left for later */
3795 :
3796 : /* Public result fields */
3797 1007884 : workspace->disabled_nodes = disabled_nodes;
3798 1007884 : workspace->startup_cost = startup_cost;
3799 1007884 : workspace->total_cost = startup_cost + run_cost + inner_run_cost;
3800 : /* Save private data for final_cost_mergejoin */
3801 1007884 : workspace->run_cost = run_cost;
3802 1007884 : workspace->inner_run_cost = inner_run_cost;
3803 1007884 : workspace->outer_rows = outer_rows;
3804 1007884 : workspace->inner_rows = inner_rows;
3805 1007884 : workspace->outer_skip_rows = outer_skip_rows;
3806 1007884 : workspace->inner_skip_rows = inner_skip_rows;
3807 1007884 : }
3808 :
3809 : /*
3810 : * final_cost_mergejoin
3811 : * Final estimate of the cost and result size of a mergejoin path.
3812 : *
3813 : * Unlike other costsize functions, this routine makes two actual decisions:
3814 : * whether the executor will need to do mark/restore, and whether we should
3815 : * materialize the inner path. It would be logically cleaner to build
3816 : * separate paths testing these alternatives, but that would require repeating
3817 : * most of the cost calculations, which are not all that cheap. Since the
3818 : * choice will not affect output pathkeys or startup cost, only total cost,
3819 : * there is no possibility of wanting to keep more than one path. So it seems
3820 : * best to make the decisions here and record them in the path's
3821 : * skip_mark_restore and materialize_inner fields.
3822 : *
3823 : * Mark/restore overhead is usually required, but can be skipped if we know
3824 : * that the executor need find only one match per outer tuple, and that the
3825 : * mergeclauses are sufficient to identify a match.
3826 : *
3827 : * We materialize the inner path if we need mark/restore and either the inner
3828 : * path can't support mark/restore, or it's cheaper to use an interposed
3829 : * Material node to handle mark/restore.
3830 : *
3831 : * 'path' is already filled in except for the rows and cost fields and
3832 : * skip_mark_restore and materialize_inner
3833 : * 'workspace' is the result from initial_cost_mergejoin
3834 : * 'extra' contains miscellaneous information about the join
3835 : */
3836 : void
3837 268538 : final_cost_mergejoin(PlannerInfo *root, MergePath *path,
3838 : JoinCostWorkspace *workspace,
3839 : JoinPathExtraData *extra)
3840 : {
3841 268538 : Path *outer_path = path->jpath.outerjoinpath;
3842 268538 : Path *inner_path = path->jpath.innerjoinpath;
3843 268538 : double inner_path_rows = inner_path->rows;
3844 268538 : List *mergeclauses = path->path_mergeclauses;
3845 268538 : List *innersortkeys = path->innersortkeys;
3846 268538 : Cost startup_cost = workspace->startup_cost;
3847 268538 : Cost run_cost = workspace->run_cost;
3848 268538 : Cost inner_run_cost = workspace->inner_run_cost;
3849 268538 : double outer_rows = workspace->outer_rows;
3850 268538 : double inner_rows = workspace->inner_rows;
3851 268538 : double outer_skip_rows = workspace->outer_skip_rows;
3852 268538 : double inner_skip_rows = workspace->inner_skip_rows;
3853 : Cost cpu_per_tuple,
3854 : bare_inner_cost,
3855 : mat_inner_cost;
3856 : QualCost merge_qual_cost;
3857 : QualCost qp_qual_cost;
3858 : double mergejointuples,
3859 : rescannedtuples;
3860 : double rescanratio;
3861 :
3862 : /* Set the number of disabled nodes. */
3863 268538 : path->jpath.path.disabled_nodes = workspace->disabled_nodes;
3864 :
3865 : /* Protect some assumptions below that rowcounts aren't zero */
3866 268538 : if (inner_path_rows <= 0)
3867 90 : inner_path_rows = 1;
3868 :
3869 : /* Mark the path with the correct row estimate */
3870 268538 : if (path->jpath.path.param_info)
3871 654 : path->jpath.path.rows = path->jpath.path.param_info->ppi_rows;
3872 : else
3873 267884 : path->jpath.path.rows = path->jpath.path.parent->rows;
3874 :
3875 : /* For partial paths, scale row estimate. */
3876 268538 : if (path->jpath.path.parallel_workers > 0)
3877 : {
3878 9180 : double parallel_divisor = get_parallel_divisor(&path->jpath.path);
3879 :
3880 9180 : path->jpath.path.rows =
3881 9180 : clamp_row_est(path->jpath.path.rows / parallel_divisor);
3882 : }
3883 :
3884 : /*
3885 : * Compute cost of the mergequals and qpquals (other restriction clauses)
3886 : * separately.
3887 : */
3888 268538 : cost_qual_eval(&merge_qual_cost, mergeclauses, root);
3889 268538 : cost_qual_eval(&qp_qual_cost, path->jpath.joinrestrictinfo, root);
3890 268538 : qp_qual_cost.startup -= merge_qual_cost.startup;
3891 268538 : qp_qual_cost.per_tuple -= merge_qual_cost.per_tuple;
3892 :
3893 : /*
3894 : * With a SEMI or ANTI join, or if the innerrel is known unique, the
3895 : * executor will stop scanning for matches after the first match. When
3896 : * all the joinclauses are merge clauses, this means we don't ever need to
3897 : * back up the merge, and so we can skip mark/restore overhead.
3898 : */
3899 268538 : if ((path->jpath.jointype == JOIN_SEMI ||
3900 264760 : path->jpath.jointype == JOIN_ANTI ||
3901 383528 : extra->inner_unique) &&
3902 125020 : (list_length(path->jpath.joinrestrictinfo) ==
3903 125020 : list_length(path->path_mergeclauses)))
3904 107000 : path->skip_mark_restore = true;
3905 : else
3906 161538 : path->skip_mark_restore = false;
3907 :
3908 : /*
3909 : * Get approx # tuples passing the mergequals. We use approx_tuple_count
3910 : * here because we need an estimate done with JOIN_INNER semantics.
3911 : */
3912 268538 : mergejointuples = approx_tuple_count(root, &path->jpath, mergeclauses);
3913 :
3914 : /*
3915 : * When there are equal merge keys in the outer relation, the mergejoin
3916 : * must rescan any matching tuples in the inner relation. This means
3917 : * re-fetching inner tuples; we have to estimate how often that happens.
3918 : *
3919 : * For regular inner and outer joins, the number of re-fetches can be
3920 : * estimated approximately as size of merge join output minus size of
3921 : * inner relation. Assume that the distinct key values are 1, 2, ..., and
3922 : * denote the number of values of each key in the outer relation as m1,
3923 : * m2, ...; in the inner relation, n1, n2, ... Then we have
3924 : *
3925 : * size of join = m1 * n1 + m2 * n2 + ...
3926 : *
3927 : * number of rescanned tuples = (m1 - 1) * n1 + (m2 - 1) * n2 + ... = m1 *
3928 : * n1 + m2 * n2 + ... - (n1 + n2 + ...) = size of join - size of inner
3929 : * relation
3930 : *
3931 : * This equation works correctly for outer tuples having no inner match
3932 : * (nk = 0), but not for inner tuples having no outer match (mk = 0); we
3933 : * are effectively subtracting those from the number of rescanned tuples,
3934 : * when we should not. Can we do better without expensive selectivity
3935 : * computations?
3936 : *
3937 : * The whole issue is moot if we are working from a unique-ified outer
3938 : * input, or if we know we don't need to mark/restore at all.
3939 : */
3940 268538 : if (IsA(outer_path, UniquePath) || path->skip_mark_restore)
3941 108006 : rescannedtuples = 0;
3942 : else
3943 : {
3944 160532 : rescannedtuples = mergejointuples - inner_path_rows;
3945 : /* Must clamp because of possible underestimate */
3946 160532 : if (rescannedtuples < 0)
3947 59884 : rescannedtuples = 0;
3948 : }
3949 :
3950 : /*
3951 : * We'll inflate various costs this much to account for rescanning. Note
3952 : * that this is to be multiplied by something involving inner_rows, or
3953 : * another number related to the portion of the inner rel we'll scan.
3954 : */
3955 268538 : rescanratio = 1.0 + (rescannedtuples / inner_rows);
3956 :
3957 : /*
3958 : * Decide whether we want to materialize the inner input to shield it from
3959 : * mark/restore and performing re-fetches. Our cost model for regular
3960 : * re-fetches is that a re-fetch costs the same as an original fetch,
3961 : * which is probably an overestimate; but on the other hand we ignore the
3962 : * bookkeeping costs of mark/restore. Not clear if it's worth developing
3963 : * a more refined model. So we just need to inflate the inner run cost by
3964 : * rescanratio.
3965 : */
3966 268538 : bare_inner_cost = inner_run_cost * rescanratio;
3967 :
3968 : /*
3969 : * When we interpose a Material node the re-fetch cost is assumed to be
3970 : * just cpu_operator_cost per tuple, independently of the underlying
3971 : * plan's cost; and we charge an extra cpu_operator_cost per original
3972 : * fetch as well. Note that we're assuming the materialize node will
3973 : * never spill to disk, since it only has to remember tuples back to the
3974 : * last mark. (If there are a huge number of duplicates, our other cost
3975 : * factors will make the path so expensive that it probably won't get
3976 : * chosen anyway.) So we don't use cost_rescan here.
3977 : *
3978 : * Note: keep this estimate in sync with create_mergejoin_plan's labeling
3979 : * of the generated Material node.
3980 : */
3981 268538 : mat_inner_cost = inner_run_cost +
3982 268538 : cpu_operator_cost * inner_rows * rescanratio;
3983 :
3984 : /*
3985 : * If we don't need mark/restore at all, we don't need materialization.
3986 : */
3987 268538 : if (path->skip_mark_restore)
3988 107000 : path->materialize_inner = false;
3989 :
3990 : /*
3991 : * Prefer materializing if it looks cheaper, unless the user has asked to
3992 : * suppress materialization.
3993 : */
3994 161538 : else if (enable_material && mat_inner_cost < bare_inner_cost)
3995 2168 : path->materialize_inner = true;
3996 :
3997 : /*
3998 : * Even if materializing doesn't look cheaper, we *must* do it if the
3999 : * inner path is to be used directly (without sorting) and it doesn't
4000 : * support mark/restore.
4001 : *
4002 : * Since the inner side must be ordered, and only Sorts and IndexScans can
4003 : * create order to begin with, and they both support mark/restore, you
4004 : * might think there's no problem --- but you'd be wrong. Nestloop and
4005 : * merge joins can *preserve* the order of their inputs, so they can be
4006 : * selected as the input of a mergejoin, and they don't support
4007 : * mark/restore at present.
4008 : *
4009 : * We don't test the value of enable_material here, because
4010 : * materialization is required for correctness in this case, and turning
4011 : * it off does not entitle us to deliver an invalid plan.
4012 : */
4013 159370 : else if (innersortkeys == NIL &&
4014 10810 : !ExecSupportsMarkRestore(inner_path))
4015 1402 : path->materialize_inner = true;
4016 :
4017 : /*
4018 : * Also, force materializing if the inner path is to be sorted and the
4019 : * sort is expected to spill to disk. This is because the final merge
4020 : * pass can be done on-the-fly if it doesn't have to support mark/restore.
4021 : * We don't try to adjust the cost estimates for this consideration,
4022 : * though.
4023 : *
4024 : * Since materialization is a performance optimization in this case,
4025 : * rather than necessary for correctness, we skip it if enable_material is
4026 : * off.
4027 : */
4028 157968 : else if (enable_material && innersortkeys != NIL &&
4029 148512 : relation_byte_size(inner_path_rows,
4030 148512 : inner_path->pathtarget->width) >
4031 148512 : (work_mem * 1024L))
4032 172 : path->materialize_inner = true;
4033 : else
4034 157796 : path->materialize_inner = false;
4035 :
4036 : /* Charge the right incremental cost for the chosen case */
4037 268538 : if (path->materialize_inner)
4038 3742 : run_cost += mat_inner_cost;
4039 : else
4040 264796 : run_cost += bare_inner_cost;
4041 :
4042 : /* CPU costs */
4043 :
4044 : /*
4045 : * The number of tuple comparisons needed is approximately number of outer
4046 : * rows plus number of inner rows plus number of rescanned tuples (can we
4047 : * refine this?). At each one, we need to evaluate the mergejoin quals.
4048 : */
4049 268538 : startup_cost += merge_qual_cost.startup;
4050 268538 : startup_cost += merge_qual_cost.per_tuple *
4051 268538 : (outer_skip_rows + inner_skip_rows * rescanratio);
4052 268538 : run_cost += merge_qual_cost.per_tuple *
4053 268538 : ((outer_rows - outer_skip_rows) +
4054 268538 : (inner_rows - inner_skip_rows) * rescanratio);
4055 :
4056 : /*
4057 : * For each tuple that gets through the mergejoin proper, we charge
4058 : * cpu_tuple_cost plus the cost of evaluating additional restriction
4059 : * clauses that are to be applied at the join. (This is pessimistic since
4060 : * not all of the quals may get evaluated at each tuple.)
4061 : *
4062 : * Note: we could adjust for SEMI/ANTI joins skipping some qual
4063 : * evaluations here, but it's probably not worth the trouble.
4064 : */
4065 268538 : startup_cost += qp_qual_cost.startup;
4066 268538 : cpu_per_tuple = cpu_tuple_cost + qp_qual_cost.per_tuple;
4067 268538 : run_cost += cpu_per_tuple * mergejointuples;
4068 :
4069 : /* tlist eval costs are paid per output row, not per tuple scanned */
4070 268538 : startup_cost += path->jpath.path.pathtarget->cost.startup;
4071 268538 : run_cost += path->jpath.path.pathtarget->cost.per_tuple * path->jpath.path.rows;
4072 :
4073 268538 : path->jpath.path.startup_cost = startup_cost;
4074 268538 : path->jpath.path.total_cost = startup_cost + run_cost;
4075 268538 : }
4076 :
4077 : /*
4078 : * run mergejoinscansel() with caching
4079 : */
4080 : static MergeScanSelCache *
4081 1001768 : cached_scansel(PlannerInfo *root, RestrictInfo *rinfo, PathKey *pathkey)
4082 : {
4083 : MergeScanSelCache *cache;
4084 : ListCell *lc;
4085 : Selectivity leftstartsel,
4086 : leftendsel,
4087 : rightstartsel,
4088 : rightendsel;
4089 : MemoryContext oldcontext;
4090 :
4091 : /* Do we have this result already? */
4092 1001810 : foreach(lc, rinfo->scansel_cache)
4093 : {
4094 901494 : cache = (MergeScanSelCache *) lfirst(lc);
4095 901494 : if (cache->opfamily == pathkey->pk_opfamily &&
4096 901494 : cache->collation == pathkey->pk_eclass->ec_collation &&
4097 901494 : cache->strategy == pathkey->pk_strategy &&
4098 901452 : cache->nulls_first == pathkey->pk_nulls_first)
4099 901452 : return cache;
4100 : }
4101 :
4102 : /* Nope, do the computation */
4103 100316 : mergejoinscansel(root,
4104 100316 : (Node *) rinfo->clause,
4105 : pathkey->pk_opfamily,
4106 : pathkey->pk_strategy,
4107 100316 : pathkey->pk_nulls_first,
4108 : &leftstartsel,
4109 : &leftendsel,
4110 : &rightstartsel,
4111 : &rightendsel);
4112 :
4113 : /* Cache the result in suitably long-lived workspace */
4114 100316 : oldcontext = MemoryContextSwitchTo(root->planner_cxt);
4115 :
4116 100316 : cache = (MergeScanSelCache *) palloc(sizeof(MergeScanSelCache));
4117 100316 : cache->opfamily = pathkey->pk_opfamily;
4118 100316 : cache->collation = pathkey->pk_eclass->ec_collation;
4119 100316 : cache->strategy = pathkey->pk_strategy;
4120 100316 : cache->nulls_first = pathkey->pk_nulls_first;
4121 100316 : cache->leftstartsel = leftstartsel;
4122 100316 : cache->leftendsel = leftendsel;
4123 100316 : cache->rightstartsel = rightstartsel;
4124 100316 : cache->rightendsel = rightendsel;
4125 :
4126 100316 : rinfo->scansel_cache = lappend(rinfo->scansel_cache, cache);
4127 :
4128 100316 : MemoryContextSwitchTo(oldcontext);
4129 :
4130 100316 : return cache;
4131 : }
4132 :
4133 : /*
4134 : * initial_cost_hashjoin
4135 : * Preliminary estimate of the cost of a hashjoin path.
4136 : *
4137 : * This must quickly produce lower-bound estimates of the path's startup and
4138 : * total costs. If we are unable to eliminate the proposed path from
4139 : * consideration using the lower bounds, final_cost_hashjoin will be called
4140 : * to obtain the final estimates.
4141 : *
4142 : * The exact division of labor between this function and final_cost_hashjoin
4143 : * is private to them, and represents a tradeoff between speed of the initial
4144 : * estimate and getting a tight lower bound. We choose to not examine the
4145 : * join quals here (other than by counting the number of hash clauses),
4146 : * so we can't do much with CPU costs. We do assume that
4147 : * ExecChooseHashTableSize is cheap enough to use here.
4148 : *
4149 : * 'workspace' is to be filled with startup_cost, total_cost, and perhaps
4150 : * other data to be used by final_cost_hashjoin
4151 : * 'jointype' is the type of join to be performed
4152 : * 'hashclauses' is the list of joinclauses to be used as hash clauses
4153 : * 'outer_path' is the outer input to the join
4154 : * 'inner_path' is the inner input to the join
4155 : * 'extra' contains miscellaneous information about the join
4156 : * 'parallel_hash' indicates that inner_path is partial and that a shared
4157 : * hash table will be built in parallel
4158 : */
4159 : void
4160 553082 : initial_cost_hashjoin(PlannerInfo *root, JoinCostWorkspace *workspace,
4161 : JoinType jointype,
4162 : List *hashclauses,
4163 : Path *outer_path, Path *inner_path,
4164 : JoinPathExtraData *extra,
4165 : bool parallel_hash)
4166 : {
4167 : int disabled_nodes;
4168 553082 : Cost startup_cost = 0;
4169 553082 : Cost run_cost = 0;
4170 553082 : double outer_path_rows = outer_path->rows;
4171 553082 : double inner_path_rows = inner_path->rows;
4172 553082 : double inner_path_rows_total = inner_path_rows;
4173 553082 : int num_hashclauses = list_length(hashclauses);
4174 : int numbuckets;
4175 : int numbatches;
4176 : int num_skew_mcvs;
4177 : size_t space_allowed; /* unused */
4178 :
4179 : /* Count up disabled nodes. */
4180 553082 : disabled_nodes = enable_hashjoin ? 0 : 1;
4181 553082 : disabled_nodes += inner_path->disabled_nodes;
4182 553082 : disabled_nodes += outer_path->disabled_nodes;
4183 :
4184 : /* cost of source data */
4185 553082 : startup_cost += outer_path->startup_cost;
4186 553082 : run_cost += outer_path->total_cost - outer_path->startup_cost;
4187 553082 : startup_cost += inner_path->total_cost;
4188 :
4189 : /*
4190 : * Cost of computing hash function: must do it once per input tuple. We
4191 : * charge one cpu_operator_cost for each column's hash function. Also,
4192 : * tack on one cpu_tuple_cost per inner row, to model the costs of
4193 : * inserting the row into the hashtable.
4194 : *
4195 : * XXX when a hashclause is more complex than a single operator, we really
4196 : * should charge the extra eval costs of the left or right side, as
4197 : * appropriate, here. This seems more work than it's worth at the moment.
4198 : */
4199 553082 : startup_cost += (cpu_operator_cost * num_hashclauses + cpu_tuple_cost)
4200 553082 : * inner_path_rows;
4201 553082 : run_cost += cpu_operator_cost * num_hashclauses * outer_path_rows;
4202 :
4203 : /*
4204 : * If this is a parallel hash build, then the value we have for
4205 : * inner_rows_total currently refers only to the rows returned by each
4206 : * participant. For shared hash table size estimation, we need the total
4207 : * number, so we need to undo the division.
4208 : */
4209 553082 : if (parallel_hash)
4210 12264 : inner_path_rows_total *= get_parallel_divisor(inner_path);
4211 :
4212 : /*
4213 : * Get hash table size that executor would use for inner relation.
4214 : *
4215 : * XXX for the moment, always assume that skew optimization will be
4216 : * performed. As long as SKEW_HASH_MEM_PERCENT is small, it's not worth
4217 : * trying to determine that for sure.
4218 : *
4219 : * XXX at some point it might be interesting to try to account for skew
4220 : * optimization in the cost estimate, but for now, we don't.
4221 : */
4222 553082 : ExecChooseHashTableSize(inner_path_rows_total,
4223 553082 : inner_path->pathtarget->width,
4224 : true, /* useskew */
4225 : parallel_hash, /* try_combined_hash_mem */
4226 : outer_path->parallel_workers,
4227 : &space_allowed,
4228 : &numbuckets,
4229 : &numbatches,
4230 : &num_skew_mcvs);
4231 :
4232 : /*
4233 : * If inner relation is too big then we will need to "batch" the join,
4234 : * which implies writing and reading most of the tuples to disk an extra
4235 : * time. Charge seq_page_cost per page, since the I/O should be nice and
4236 : * sequential. Writing the inner rel counts as startup cost, all the rest
4237 : * as run cost.
4238 : */
4239 553082 : if (numbatches > 1)
4240 : {
4241 4454 : double outerpages = page_size(outer_path_rows,
4242 4454 : outer_path->pathtarget->width);
4243 4454 : double innerpages = page_size(inner_path_rows,
4244 4454 : inner_path->pathtarget->width);
4245 :
4246 4454 : startup_cost += seq_page_cost * innerpages;
4247 4454 : run_cost += seq_page_cost * (innerpages + 2 * outerpages);
4248 : }
4249 :
4250 : /* CPU costs left for later */
4251 :
4252 : /* Public result fields */
4253 553082 : workspace->disabled_nodes = disabled_nodes;
4254 553082 : workspace->startup_cost = startup_cost;
4255 553082 : workspace->total_cost = startup_cost + run_cost;
4256 : /* Save private data for final_cost_hashjoin */
4257 553082 : workspace->run_cost = run_cost;
4258 553082 : workspace->numbuckets = numbuckets;
4259 553082 : workspace->numbatches = numbatches;
4260 553082 : workspace->inner_rows_total = inner_path_rows_total;
4261 553082 : }
4262 :
4263 : /*
4264 : * final_cost_hashjoin
4265 : * Final estimate of the cost and result size of a hashjoin path.
4266 : *
4267 : * Note: the numbatches estimate is also saved into 'path' for use later
4268 : *
4269 : * 'path' is already filled in except for the rows and cost fields and
4270 : * num_batches
4271 : * 'workspace' is the result from initial_cost_hashjoin
4272 : * 'extra' contains miscellaneous information about the join
4273 : */
4274 : void
4275 237510 : final_cost_hashjoin(PlannerInfo *root, HashPath *path,
4276 : JoinCostWorkspace *workspace,
4277 : JoinPathExtraData *extra)
4278 : {
4279 237510 : Path *outer_path = path->jpath.outerjoinpath;
4280 237510 : Path *inner_path = path->jpath.innerjoinpath;
4281 237510 : double outer_path_rows = outer_path->rows;
4282 237510 : double inner_path_rows = inner_path->rows;
4283 237510 : double inner_path_rows_total = workspace->inner_rows_total;
4284 237510 : List *hashclauses = path->path_hashclauses;
4285 237510 : Cost startup_cost = workspace->startup_cost;
4286 237510 : Cost run_cost = workspace->run_cost;
4287 237510 : int numbuckets = workspace->numbuckets;
4288 237510 : int numbatches = workspace->numbatches;
4289 : Cost cpu_per_tuple;
4290 : QualCost hash_qual_cost;
4291 : QualCost qp_qual_cost;
4292 : double hashjointuples;
4293 : double virtualbuckets;
4294 : Selectivity innerbucketsize;
4295 : Selectivity innermcvfreq;
4296 : ListCell *hcl;
4297 :
4298 : /* Set the number of disabled nodes. */
4299 237510 : path->jpath.path.disabled_nodes = workspace->disabled_nodes;
4300 :
4301 : /* Mark the path with the correct row estimate */
4302 237510 : if (path->jpath.path.param_info)
4303 1354 : path->jpath.path.rows = path->jpath.path.param_info->ppi_rows;
4304 : else
4305 236156 : path->jpath.path.rows = path->jpath.path.parent->rows;
4306 :
4307 : /* For partial paths, scale row estimate. */
4308 237510 : if (path->jpath.path.parallel_workers > 0)
4309 : {
4310 10950 : double parallel_divisor = get_parallel_divisor(&path->jpath.path);
4311 :
4312 10950 : path->jpath.path.rows =
4313 10950 : clamp_row_est(path->jpath.path.rows / parallel_divisor);
4314 : }
4315 :
4316 : /* mark the path with estimated # of batches */
4317 237510 : path->num_batches = numbatches;
4318 :
4319 : /* store the total number of tuples (sum of partial row estimates) */
4320 237510 : path->inner_rows_total = inner_path_rows_total;
4321 :
4322 : /* and compute the number of "virtual" buckets in the whole join */
4323 237510 : virtualbuckets = (double) numbuckets * (double) numbatches;
4324 :
4325 : /*
4326 : * Determine bucketsize fraction and MCV frequency for the inner relation.
4327 : * We use the smallest bucketsize or MCV frequency estimated for any
4328 : * individual hashclause; this is undoubtedly conservative.
4329 : *
4330 : * BUT: if inner relation has been unique-ified, we can assume it's good
4331 : * for hashing. This is important both because it's the right answer, and
4332 : * because we avoid contaminating the cache with a value that's wrong for
4333 : * non-unique-ified paths.
4334 : */
4335 237510 : if (IsA(inner_path, UniquePath))
4336 : {
4337 2086 : innerbucketsize = 1.0 / virtualbuckets;
4338 2086 : innermcvfreq = 0.0;
4339 : }
4340 : else
4341 : {
4342 235424 : innerbucketsize = 1.0;
4343 235424 : innermcvfreq = 1.0;
4344 495240 : foreach(hcl, hashclauses)
4345 : {
4346 259816 : RestrictInfo *restrictinfo = lfirst_node(RestrictInfo, hcl);
4347 : Selectivity thisbucketsize;
4348 : Selectivity thismcvfreq;
4349 :
4350 : /*
4351 : * First we have to figure out which side of the hashjoin clause
4352 : * is the inner side.
4353 : *
4354 : * Since we tend to visit the same clauses over and over when
4355 : * planning a large query, we cache the bucket stats estimates in
4356 : * the RestrictInfo node to avoid repeated lookups of statistics.
4357 : */
4358 259816 : if (bms_is_subset(restrictinfo->right_relids,
4359 259816 : inner_path->parent->relids))
4360 : {
4361 : /* righthand side is inner */
4362 137938 : thisbucketsize = restrictinfo->right_bucketsize;
4363 137938 : if (thisbucketsize < 0)
4364 : {
4365 : /* not cached yet */
4366 75578 : estimate_hash_bucket_stats(root,
4367 75578 : get_rightop(restrictinfo->clause),
4368 : virtualbuckets,
4369 : &restrictinfo->right_mcvfreq,
4370 : &restrictinfo->right_bucketsize);
4371 75578 : thisbucketsize = restrictinfo->right_bucketsize;
4372 : }
4373 137938 : thismcvfreq = restrictinfo->right_mcvfreq;
4374 : }
4375 : else
4376 : {
4377 : Assert(bms_is_subset(restrictinfo->left_relids,
4378 : inner_path->parent->relids));
4379 : /* lefthand side is inner */
4380 121878 : thisbucketsize = restrictinfo->left_bucketsize;
4381 121878 : if (thisbucketsize < 0)
4382 : {
4383 : /* not cached yet */
4384 64986 : estimate_hash_bucket_stats(root,
4385 64986 : get_leftop(restrictinfo->clause),
4386 : virtualbuckets,
4387 : &restrictinfo->left_mcvfreq,
4388 : &restrictinfo->left_bucketsize);
4389 64986 : thisbucketsize = restrictinfo->left_bucketsize;
4390 : }
4391 121878 : thismcvfreq = restrictinfo->left_mcvfreq;
4392 : }
4393 :
4394 259816 : if (innerbucketsize > thisbucketsize)
4395 163306 : innerbucketsize = thisbucketsize;
4396 259816 : if (innermcvfreq > thismcvfreq)
4397 238602 : innermcvfreq = thismcvfreq;
4398 : }
4399 : }
4400 :
4401 : /*
4402 : * If the bucket holding the inner MCV would exceed hash_mem, we don't
4403 : * want to hash unless there is really no other alternative, so apply
4404 : * disable_cost. (The executor normally copes with excessive memory usage
4405 : * by splitting batches, but obviously it cannot separate equal values
4406 : * that way, so it will be unable to drive the batch size below hash_mem
4407 : * when this is true.)
4408 : */
4409 237510 : if (relation_byte_size(clamp_row_est(inner_path_rows * innermcvfreq),
4410 475020 : inner_path->pathtarget->width) > get_hash_memory_limit())
4411 0 : startup_cost += disable_cost;
4412 :
4413 : /*
4414 : * Compute cost of the hashquals and qpquals (other restriction clauses)
4415 : * separately.
4416 : */
4417 237510 : cost_qual_eval(&hash_qual_cost, hashclauses, root);
4418 237510 : cost_qual_eval(&qp_qual_cost, path->jpath.joinrestrictinfo, root);
4419 237510 : qp_qual_cost.startup -= hash_qual_cost.startup;
4420 237510 : qp_qual_cost.per_tuple -= hash_qual_cost.per_tuple;
4421 :
4422 : /* CPU costs */
4423 :
4424 237510 : if (path->jpath.jointype == JOIN_SEMI ||
4425 234560 : path->jpath.jointype == JOIN_ANTI ||
4426 230300 : extra->inner_unique)
4427 100816 : {
4428 : double outer_matched_rows;
4429 : Selectivity inner_scan_frac;
4430 :
4431 : /*
4432 : * With a SEMI or ANTI join, or if the innerrel is known unique, the
4433 : * executor will stop after the first match.
4434 : *
4435 : * For an outer-rel row that has at least one match, we can expect the
4436 : * bucket scan to stop after a fraction 1/(match_count+1) of the
4437 : * bucket's rows, if the matches are evenly distributed. Since they
4438 : * probably aren't quite evenly distributed, we apply a fuzz factor of
4439 : * 2.0 to that fraction. (If we used a larger fuzz factor, we'd have
4440 : * to clamp inner_scan_frac to at most 1.0; but since match_count is
4441 : * at least 1, no such clamp is needed now.)
4442 : */
4443 100816 : outer_matched_rows = rint(outer_path_rows * extra->semifactors.outer_match_frac);
4444 100816 : inner_scan_frac = 2.0 / (extra->semifactors.match_count + 1.0);
4445 :
4446 100816 : startup_cost += hash_qual_cost.startup;
4447 201632 : run_cost += hash_qual_cost.per_tuple * outer_matched_rows *
4448 100816 : clamp_row_est(inner_path_rows * innerbucketsize * inner_scan_frac) * 0.5;
4449 :
4450 : /*
4451 : * For unmatched outer-rel rows, the picture is quite a lot different.
4452 : * In the first place, there is no reason to assume that these rows
4453 : * preferentially hit heavily-populated buckets; instead assume they
4454 : * are uncorrelated with the inner distribution and so they see an
4455 : * average bucket size of inner_path_rows / virtualbuckets. In the
4456 : * second place, it seems likely that they will have few if any exact
4457 : * hash-code matches and so very few of the tuples in the bucket will
4458 : * actually require eval of the hash quals. We don't have any good
4459 : * way to estimate how many will, but for the moment assume that the
4460 : * effective cost per bucket entry is one-tenth what it is for
4461 : * matchable tuples.
4462 : */
4463 201632 : run_cost += hash_qual_cost.per_tuple *
4464 201632 : (outer_path_rows - outer_matched_rows) *
4465 100816 : clamp_row_est(inner_path_rows / virtualbuckets) * 0.05;
4466 :
4467 : /* Get # of tuples that will pass the basic join */
4468 100816 : if (path->jpath.jointype == JOIN_ANTI)
4469 4260 : hashjointuples = outer_path_rows - outer_matched_rows;
4470 : else
4471 96556 : hashjointuples = outer_matched_rows;
4472 : }
4473 : else
4474 : {
4475 : /*
4476 : * The number of tuple comparisons needed is the number of outer
4477 : * tuples times the typical number of tuples in a hash bucket, which
4478 : * is the inner relation size times its bucketsize fraction. At each
4479 : * one, we need to evaluate the hashjoin quals. But actually,
4480 : * charging the full qual eval cost at each tuple is pessimistic,
4481 : * since we don't evaluate the quals unless the hash values match
4482 : * exactly. For lack of a better idea, halve the cost estimate to
4483 : * allow for that.
4484 : */
4485 136694 : startup_cost += hash_qual_cost.startup;
4486 273388 : run_cost += hash_qual_cost.per_tuple * outer_path_rows *
4487 136694 : clamp_row_est(inner_path_rows * innerbucketsize) * 0.5;
4488 :
4489 : /*
4490 : * Get approx # tuples passing the hashquals. We use
4491 : * approx_tuple_count here because we need an estimate done with
4492 : * JOIN_INNER semantics.
4493 : */
4494 136694 : hashjointuples = approx_tuple_count(root, &path->jpath, hashclauses);
4495 : }
4496 :
4497 : /*
4498 : * For each tuple that gets through the hashjoin proper, we charge
4499 : * cpu_tuple_cost plus the cost of evaluating additional restriction
4500 : * clauses that are to be applied at the join. (This is pessimistic since
4501 : * not all of the quals may get evaluated at each tuple.)
4502 : */
4503 237510 : startup_cost += qp_qual_cost.startup;
4504 237510 : cpu_per_tuple = cpu_tuple_cost + qp_qual_cost.per_tuple;
4505 237510 : run_cost += cpu_per_tuple * hashjointuples;
4506 :
4507 : /* tlist eval costs are paid per output row, not per tuple scanned */
4508 237510 : startup_cost += path->jpath.path.pathtarget->cost.startup;
4509 237510 : run_cost += path->jpath.path.pathtarget->cost.per_tuple * path->jpath.path.rows;
4510 :
4511 237510 : path->jpath.path.startup_cost = startup_cost;
4512 237510 : path->jpath.path.total_cost = startup_cost + run_cost;
4513 237510 : }
4514 :
4515 :
4516 : /*
4517 : * cost_subplan
4518 : * Figure the costs for a SubPlan (or initplan).
4519 : *
4520 : * Note: we could dig the subplan's Plan out of the root list, but in practice
4521 : * all callers have it handy already, so we make them pass it.
4522 : */
4523 : void
4524 38882 : cost_subplan(PlannerInfo *root, SubPlan *subplan, Plan *plan)
4525 : {
4526 : QualCost sp_cost;
4527 :
4528 : /* Figure any cost for evaluating the testexpr */
4529 38882 : cost_qual_eval(&sp_cost,
4530 38882 : make_ands_implicit((Expr *) subplan->testexpr),
4531 : root);
4532 :
4533 38882 : if (subplan->useHashTable)
4534 : {
4535 : /*
4536 : * If we are using a hash table for the subquery outputs, then the
4537 : * cost of evaluating the query is a one-time cost. We charge one
4538 : * cpu_operator_cost per tuple for the work of loading the hashtable,
4539 : * too.
4540 : */
4541 1902 : sp_cost.startup += plan->total_cost +
4542 1902 : cpu_operator_cost * plan->plan_rows;
4543 :
4544 : /*
4545 : * The per-tuple costs include the cost of evaluating the lefthand
4546 : * expressions, plus the cost of probing the hashtable. We already
4547 : * accounted for the lefthand expressions as part of the testexpr, and
4548 : * will also have counted one cpu_operator_cost for each comparison
4549 : * operator. That is probably too low for the probing cost, but it's
4550 : * hard to make a better estimate, so live with it for now.
4551 : */
4552 : }
4553 : else
4554 : {
4555 : /*
4556 : * Otherwise we will be rescanning the subplan output on each
4557 : * evaluation. We need to estimate how much of the output we will
4558 : * actually need to scan. NOTE: this logic should agree with the
4559 : * tuple_fraction estimates used by make_subplan() in
4560 : * plan/subselect.c.
4561 : */
4562 36980 : Cost plan_run_cost = plan->total_cost - plan->startup_cost;
4563 :
4564 36980 : if (subplan->subLinkType == EXISTS_SUBLINK)
4565 : {
4566 : /* we only need to fetch 1 tuple; clamp to avoid zero divide */
4567 2086 : sp_cost.per_tuple += plan_run_cost / clamp_row_est(plan->plan_rows);
4568 : }
4569 34894 : else if (subplan->subLinkType == ALL_SUBLINK ||
4570 34876 : subplan->subLinkType == ANY_SUBLINK)
4571 : {
4572 : /* assume we need 50% of the tuples */
4573 112 : sp_cost.per_tuple += 0.50 * plan_run_cost;
4574 : /* also charge a cpu_operator_cost per row examined */
4575 112 : sp_cost.per_tuple += 0.50 * plan->plan_rows * cpu_operator_cost;
4576 : }
4577 : else
4578 : {
4579 : /* assume we need all tuples */
4580 34782 : sp_cost.per_tuple += plan_run_cost;
4581 : }
4582 :
4583 : /*
4584 : * Also account for subplan's startup cost. If the subplan is
4585 : * uncorrelated or undirect correlated, AND its topmost node is one
4586 : * that materializes its output, assume that we'll only need to pay
4587 : * its startup cost once; otherwise assume we pay the startup cost
4588 : * every time.
4589 : */
4590 48452 : if (subplan->parParam == NIL &&
4591 11472 : ExecMaterializesOutput(nodeTag(plan)))
4592 584 : sp_cost.startup += plan->startup_cost;
4593 : else
4594 36396 : sp_cost.per_tuple += plan->startup_cost;
4595 : }
4596 :
4597 38882 : subplan->startup_cost = sp_cost.startup;
4598 38882 : subplan->per_call_cost = sp_cost.per_tuple;
4599 38882 : }
4600 :
4601 :
4602 : /*
4603 : * cost_rescan
4604 : * Given a finished Path, estimate the costs of rescanning it after
4605 : * having done so the first time. For some Path types a rescan is
4606 : * cheaper than an original scan (if no parameters change), and this
4607 : * function embodies knowledge about that. The default is to return
4608 : * the same costs stored in the Path. (Note that the cost estimates
4609 : * actually stored in Paths are always for first scans.)
4610 : *
4611 : * This function is not currently intended to model effects such as rescans
4612 : * being cheaper due to disk block caching; what we are concerned with is
4613 : * plan types wherein the executor caches results explicitly, or doesn't
4614 : * redo startup calculations, etc.
4615 : */
4616 : static void
4617 2328140 : cost_rescan(PlannerInfo *root, Path *path,
4618 : Cost *rescan_startup_cost, /* output parameters */
4619 : Cost *rescan_total_cost)
4620 : {
4621 2328140 : switch (path->pathtype)
4622 : {
4623 41082 : case T_FunctionScan:
4624 :
4625 : /*
4626 : * Currently, nodeFunctionscan.c always executes the function to
4627 : * completion before returning any rows, and caches the results in
4628 : * a tuplestore. So the function eval cost is all startup cost
4629 : * and isn't paid over again on rescans. However, all run costs
4630 : * will be paid over again.
4631 : */
4632 41082 : *rescan_startup_cost = 0;
4633 41082 : *rescan_total_cost = path->total_cost - path->startup_cost;
4634 41082 : break;
4635 102070 : case T_HashJoin:
4636 :
4637 : /*
4638 : * If it's a single-batch join, we don't need to rebuild the hash
4639 : * table during a rescan.
4640 : */
4641 102070 : if (((HashPath *) path)->num_batches == 1)
4642 : {
4643 : /* Startup cost is exactly the cost of hash table building */
4644 102070 : *rescan_startup_cost = 0;
4645 102070 : *rescan_total_cost = path->total_cost - path->startup_cost;
4646 : }
4647 : else
4648 : {
4649 : /* Otherwise, no special treatment */
4650 0 : *rescan_startup_cost = path->startup_cost;
4651 0 : *rescan_total_cost = path->total_cost;
4652 : }
4653 102070 : break;
4654 6390 : case T_CteScan:
4655 : case T_WorkTableScan:
4656 : {
4657 : /*
4658 : * These plan types materialize their final result in a
4659 : * tuplestore or tuplesort object. So the rescan cost is only
4660 : * cpu_tuple_cost per tuple, unless the result is large enough
4661 : * to spill to disk.
4662 : */
4663 6390 : Cost run_cost = cpu_tuple_cost * path->rows;
4664 6390 : double nbytes = relation_byte_size(path->rows,
4665 6390 : path->pathtarget->width);
4666 6390 : long work_mem_bytes = work_mem * 1024L;
4667 :
4668 6390 : if (nbytes > work_mem_bytes)
4669 : {
4670 : /* It will spill, so account for re-read cost */
4671 160 : double npages = ceil(nbytes / BLCKSZ);
4672 :
4673 160 : run_cost += seq_page_cost * npages;
4674 : }
4675 6390 : *rescan_startup_cost = 0;
4676 6390 : *rescan_total_cost = run_cost;
4677 : }
4678 6390 : break;
4679 798526 : case T_Material:
4680 : case T_Sort:
4681 : {
4682 : /*
4683 : * These plan types not only materialize their results, but do
4684 : * not implement qual filtering or projection. So they are
4685 : * even cheaper to rescan than the ones above. We charge only
4686 : * cpu_operator_cost per tuple. (Note: keep that in sync with
4687 : * the run_cost charge in cost_sort, and also see comments in
4688 : * cost_material before you change it.)
4689 : */
4690 798526 : Cost run_cost = cpu_operator_cost * path->rows;
4691 798526 : double nbytes = relation_byte_size(path->rows,
4692 798526 : path->pathtarget->width);
4693 798526 : long work_mem_bytes = work_mem * 1024L;
4694 :
4695 798526 : if (nbytes > work_mem_bytes)
4696 : {
4697 : /* It will spill, so account for re-read cost */
4698 9322 : double npages = ceil(nbytes / BLCKSZ);
4699 :
4700 9322 : run_cost += seq_page_cost * npages;
4701 : }
4702 798526 : *rescan_startup_cost = 0;
4703 798526 : *rescan_total_cost = run_cost;
4704 : }
4705 798526 : break;
4706 234280 : case T_Memoize:
4707 : /* All the hard work is done by cost_memoize_rescan */
4708 234280 : cost_memoize_rescan(root, (MemoizePath *) path,
4709 : rescan_startup_cost, rescan_total_cost);
4710 234280 : break;
4711 1145792 : default:
4712 1145792 : *rescan_startup_cost = path->startup_cost;
4713 1145792 : *rescan_total_cost = path->total_cost;
4714 1145792 : break;
4715 : }
4716 2328140 : }
4717 :
4718 :
4719 : /*
4720 : * cost_qual_eval
4721 : * Estimate the CPU costs of evaluating a WHERE clause.
4722 : * The input can be either an implicitly-ANDed list of boolean
4723 : * expressions, or a list of RestrictInfo nodes. (The latter is
4724 : * preferred since it allows caching of the results.)
4725 : * The result includes both a one-time (startup) component,
4726 : * and a per-evaluation component.
4727 : *
4728 : * Note: in some code paths root can be passed as NULL, resulting in
4729 : * slightly worse estimates.
4730 : */
4731 : void
4732 3339026 : cost_qual_eval(QualCost *cost, List *quals, PlannerInfo *root)
4733 : {
4734 : cost_qual_eval_context context;
4735 : ListCell *l;
4736 :
4737 3339026 : context.root = root;
4738 3339026 : context.total.startup = 0;
4739 3339026 : context.total.per_tuple = 0;
4740 :
4741 : /* We don't charge any cost for the implicit ANDing at top level ... */
4742 :
4743 6229806 : foreach(l, quals)
4744 : {
4745 2890780 : Node *qual = (Node *) lfirst(l);
4746 :
4747 2890780 : cost_qual_eval_walker(qual, &context);
4748 : }
4749 :
4750 3339026 : *cost = context.total;
4751 3339026 : }
4752 :
4753 : /*
4754 : * cost_qual_eval_node
4755 : * As above, for a single RestrictInfo or expression.
4756 : */
4757 : void
4758 1609704 : cost_qual_eval_node(QualCost *cost, Node *qual, PlannerInfo *root)
4759 : {
4760 : cost_qual_eval_context context;
4761 :
4762 1609704 : context.root = root;
4763 1609704 : context.total.startup = 0;
4764 1609704 : context.total.per_tuple = 0;
4765 :
4766 1609704 : cost_qual_eval_walker(qual, &context);
4767 :
4768 1609704 : *cost = context.total;
4769 1609704 : }
4770 :
4771 : static bool
4772 7659660 : cost_qual_eval_walker(Node *node, cost_qual_eval_context *context)
4773 : {
4774 7659660 : if (node == NULL)
4775 95796 : return false;
4776 :
4777 : /*
4778 : * RestrictInfo nodes contain an eval_cost field reserved for this
4779 : * routine's use, so that it's not necessary to evaluate the qual clause's
4780 : * cost more than once. If the clause's cost hasn't been computed yet,
4781 : * the field's startup value will contain -1.
4782 : */
4783 7563864 : if (IsA(node, RestrictInfo))
4784 : {
4785 3034596 : RestrictInfo *rinfo = (RestrictInfo *) node;
4786 :
4787 3034596 : if (rinfo->eval_cost.startup < 0)
4788 : {
4789 : cost_qual_eval_context locContext;
4790 :
4791 512220 : locContext.root = context->root;
4792 512220 : locContext.total.startup = 0;
4793 512220 : locContext.total.per_tuple = 0;
4794 :
4795 : /*
4796 : * For an OR clause, recurse into the marked-up tree so that we
4797 : * set the eval_cost for contained RestrictInfos too.
4798 : */
4799 512220 : if (rinfo->orclause)
4800 8248 : cost_qual_eval_walker((Node *) rinfo->orclause, &locContext);
4801 : else
4802 503972 : cost_qual_eval_walker((Node *) rinfo->clause, &locContext);
4803 :
4804 : /*
4805 : * If the RestrictInfo is marked pseudoconstant, it will be tested
4806 : * only once, so treat its cost as all startup cost.
4807 : */
4808 512220 : if (rinfo->pseudoconstant)
4809 : {
4810 : /* count one execution during startup */
4811 9034 : locContext.total.startup += locContext.total.per_tuple;
4812 9034 : locContext.total.per_tuple = 0;
4813 : }
4814 512220 : rinfo->eval_cost = locContext.total;
4815 : }
4816 3034596 : context->total.startup += rinfo->eval_cost.startup;
4817 3034596 : context->total.per_tuple += rinfo->eval_cost.per_tuple;
4818 : /* do NOT recurse into children */
4819 3034596 : return false;
4820 : }
4821 :
4822 : /*
4823 : * For each operator or function node in the given tree, we charge the
4824 : * estimated execution cost given by pg_proc.procost (remember to multiply
4825 : * this by cpu_operator_cost).
4826 : *
4827 : * Vars and Consts are charged zero, and so are boolean operators (AND,
4828 : * OR, NOT). Simplistic, but a lot better than no model at all.
4829 : *
4830 : * Should we try to account for the possibility of short-circuit
4831 : * evaluation of AND/OR? Probably *not*, because that would make the
4832 : * results depend on the clause ordering, and we are not in any position
4833 : * to expect that the current ordering of the clauses is the one that's
4834 : * going to end up being used. The above per-RestrictInfo caching would
4835 : * not mix well with trying to re-order clauses anyway.
4836 : *
4837 : * Another issue that is entirely ignored here is that if a set-returning
4838 : * function is below top level in the tree, the functions/operators above
4839 : * it will need to be evaluated multiple times. In practical use, such
4840 : * cases arise so seldom as to not be worth the added complexity needed;
4841 : * moreover, since our rowcount estimates for functions tend to be pretty
4842 : * phony, the results would also be pretty phony.
4843 : */
4844 4529268 : if (IsA(node, FuncExpr))
4845 : {
4846 312736 : add_function_cost(context->root, ((FuncExpr *) node)->funcid, node,
4847 : &context->total);
4848 : }
4849 4216532 : else if (IsA(node, OpExpr) ||
4850 3611106 : IsA(node, DistinctExpr) ||
4851 3610236 : IsA(node, NullIfExpr))
4852 : {
4853 : /* rely on struct equivalence to treat these all alike */
4854 606408 : set_opfuncid((OpExpr *) node);
4855 606408 : add_function_cost(context->root, ((OpExpr *) node)->opfuncid, node,
4856 : &context->total);
4857 : }
4858 3610124 : else if (IsA(node, ScalarArrayOpExpr))
4859 : {
4860 38770 : ScalarArrayOpExpr *saop = (ScalarArrayOpExpr *) node;
4861 38770 : Node *arraynode = (Node *) lsecond(saop->args);
4862 : QualCost sacosts;
4863 : QualCost hcosts;
4864 38770 : double estarraylen = estimate_array_length(context->root, arraynode);
4865 :
4866 38770 : set_sa_opfuncid(saop);
4867 38770 : sacosts.startup = sacosts.per_tuple = 0;
4868 38770 : add_function_cost(context->root, saop->opfuncid, NULL,
4869 : &sacosts);
4870 :
4871 38770 : if (OidIsValid(saop->hashfuncid))
4872 : {
4873 : /* Handle costs for hashed ScalarArrayOpExpr */
4874 266 : hcosts.startup = hcosts.per_tuple = 0;
4875 :
4876 266 : add_function_cost(context->root, saop->hashfuncid, NULL, &hcosts);
4877 266 : context->total.startup += sacosts.startup + hcosts.startup;
4878 :
4879 : /* Estimate the cost of building the hashtable. */
4880 266 : context->total.startup += estarraylen * hcosts.per_tuple;
4881 :
4882 : /*
4883 : * XXX should we charge a little bit for sacosts.per_tuple when
4884 : * building the table, or is it ok to assume there will be zero
4885 : * hash collision?
4886 : */
4887 :
4888 : /*
4889 : * Charge for hashtable lookups. Charge a single hash and a
4890 : * single comparison.
4891 : */
4892 266 : context->total.per_tuple += hcosts.per_tuple + sacosts.per_tuple;
4893 : }
4894 : else
4895 : {
4896 : /*
4897 : * Estimate that the operator will be applied to about half of the
4898 : * array elements before the answer is determined.
4899 : */
4900 38504 : context->total.startup += sacosts.startup;
4901 77008 : context->total.per_tuple += sacosts.per_tuple *
4902 38504 : estimate_array_length(context->root, arraynode) * 0.5;
4903 : }
4904 : }
4905 3571354 : else if (IsA(node, Aggref) ||
4906 3518264 : IsA(node, WindowFunc))
4907 : {
4908 : /*
4909 : * Aggref and WindowFunc nodes are (and should be) treated like Vars,
4910 : * ie, zero execution cost in the current model, because they behave
4911 : * essentially like Vars at execution. We disregard the costs of
4912 : * their input expressions for the same reason. The actual execution
4913 : * costs of the aggregate/window functions and their arguments have to
4914 : * be factored into plan-node-specific costing of the Agg or WindowAgg
4915 : * plan node.
4916 : */
4917 56488 : return false; /* don't recurse into children */
4918 : }
4919 3514866 : else if (IsA(node, GroupingFunc))
4920 : {
4921 : /* Treat this as having cost 1 */
4922 422 : context->total.per_tuple += cpu_operator_cost;
4923 422 : return false; /* don't recurse into children */
4924 : }
4925 3514444 : else if (IsA(node, CoerceViaIO))
4926 : {
4927 21104 : CoerceViaIO *iocoerce = (CoerceViaIO *) node;
4928 : Oid iofunc;
4929 : Oid typioparam;
4930 : bool typisvarlena;
4931 :
4932 : /* check the result type's input function */
4933 21104 : getTypeInputInfo(iocoerce->resulttype,
4934 : &iofunc, &typioparam);
4935 21104 : add_function_cost(context->root, iofunc, NULL,
4936 : &context->total);
4937 : /* check the input type's output function */
4938 21104 : getTypeOutputInfo(exprType((Node *) iocoerce->arg),
4939 : &iofunc, &typisvarlena);
4940 21104 : add_function_cost(context->root, iofunc, NULL,
4941 : &context->total);
4942 : }
4943 3493340 : else if (IsA(node, ArrayCoerceExpr))
4944 : {
4945 4710 : ArrayCoerceExpr *acoerce = (ArrayCoerceExpr *) node;
4946 : QualCost perelemcost;
4947 :
4948 4710 : cost_qual_eval_node(&perelemcost, (Node *) acoerce->elemexpr,
4949 : context->root);
4950 4710 : context->total.startup += perelemcost.startup;
4951 4710 : if (perelemcost.per_tuple > 0)
4952 58 : context->total.per_tuple += perelemcost.per_tuple *
4953 58 : estimate_array_length(context->root, (Node *) acoerce->arg);
4954 : }
4955 3488630 : else if (IsA(node, RowCompareExpr))
4956 : {
4957 : /* Conservatively assume we will check all the columns */
4958 216 : RowCompareExpr *rcexpr = (RowCompareExpr *) node;
4959 : ListCell *lc;
4960 :
4961 702 : foreach(lc, rcexpr->opnos)
4962 : {
4963 486 : Oid opid = lfirst_oid(lc);
4964 :
4965 486 : add_function_cost(context->root, get_opcode(opid), NULL,
4966 : &context->total);
4967 : }
4968 : }
4969 3488414 : else if (IsA(node, MinMaxExpr) ||
4970 3488230 : IsA(node, SQLValueFunction) ||
4971 3483732 : IsA(node, XmlExpr) ||
4972 3483030 : IsA(node, CoerceToDomain) ||
4973 3473686 : IsA(node, NextValueExpr) ||
4974 3473324 : IsA(node, JsonExpr))
4975 : {
4976 : /* Treat all these as having cost 1 */
4977 17638 : context->total.per_tuple += cpu_operator_cost;
4978 : }
4979 3470776 : else if (IsA(node, SubLink))
4980 : {
4981 : /* This routine should not be applied to un-planned expressions */
4982 0 : elog(ERROR, "cannot handle unplanned sub-select");
4983 : }
4984 3470776 : else if (IsA(node, SubPlan))
4985 : {
4986 : /*
4987 : * A subplan node in an expression typically indicates that the
4988 : * subplan will be executed on each evaluation, so charge accordingly.
4989 : * (Sub-selects that can be executed as InitPlans have already been
4990 : * removed from the expression.)
4991 : */
4992 38720 : SubPlan *subplan = (SubPlan *) node;
4993 :
4994 38720 : context->total.startup += subplan->startup_cost;
4995 38720 : context->total.per_tuple += subplan->per_call_cost;
4996 :
4997 : /*
4998 : * We don't want to recurse into the testexpr, because it was already
4999 : * counted in the SubPlan node's costs. So we're done.
5000 : */
5001 38720 : return false;
5002 : }
5003 3432056 : else if (IsA(node, AlternativeSubPlan))
5004 : {
5005 : /*
5006 : * Arbitrarily use the first alternative plan for costing. (We should
5007 : * certainly only include one alternative, and we don't yet have
5008 : * enough information to know which one the executor is most likely to
5009 : * use.)
5010 : */
5011 1654 : AlternativeSubPlan *asplan = (AlternativeSubPlan *) node;
5012 :
5013 1654 : return cost_qual_eval_walker((Node *) linitial(asplan->subplans),
5014 : context);
5015 : }
5016 3430402 : else if (IsA(node, PlaceHolderVar))
5017 : {
5018 : /*
5019 : * A PlaceHolderVar should be given cost zero when considering general
5020 : * expression evaluation costs. The expense of doing the contained
5021 : * expression is charged as part of the tlist eval costs of the scan
5022 : * or join where the PHV is first computed (see set_rel_width and
5023 : * add_placeholders_to_joinrel). If we charged it again here, we'd be
5024 : * double-counting the cost for each level of plan that the PHV
5025 : * bubbles up through. Hence, return without recursing into the
5026 : * phexpr.
5027 : */
5028 2952 : return false;
5029 : }
5030 :
5031 : /* recurse into children */
5032 4429032 : return expression_tree_walker(node, cost_qual_eval_walker, context);
5033 : }
5034 :
5035 : /*
5036 : * get_restriction_qual_cost
5037 : * Compute evaluation costs of a baserel's restriction quals, plus any
5038 : * movable join quals that have been pushed down to the scan.
5039 : * Results are returned into *qpqual_cost.
5040 : *
5041 : * This is a convenience subroutine that works for seqscans and other cases
5042 : * where all the given quals will be evaluated the hard way. It's not useful
5043 : * for cost_index(), for example, where the index machinery takes care of
5044 : * some of the quals. We assume baserestrictcost was previously set by
5045 : * set_baserel_size_estimates().
5046 : */
5047 : static void
5048 939782 : get_restriction_qual_cost(PlannerInfo *root, RelOptInfo *baserel,
5049 : ParamPathInfo *param_info,
5050 : QualCost *qpqual_cost)
5051 : {
5052 939782 : if (param_info)
5053 : {
5054 : /* Include costs of pushed-down clauses */
5055 204472 : cost_qual_eval(qpqual_cost, param_info->ppi_clauses, root);
5056 :
5057 204472 : qpqual_cost->startup += baserel->baserestrictcost.startup;
5058 204472 : qpqual_cost->per_tuple += baserel->baserestrictcost.per_tuple;
5059 : }
5060 : else
5061 735310 : *qpqual_cost = baserel->baserestrictcost;
5062 939782 : }
5063 :
5064 :
5065 : /*
5066 : * compute_semi_anti_join_factors
5067 : * Estimate how much of the inner input a SEMI, ANTI, or inner_unique join
5068 : * can be expected to scan.
5069 : *
5070 : * In a hash or nestloop SEMI/ANTI join, the executor will stop scanning
5071 : * inner rows as soon as it finds a match to the current outer row.
5072 : * The same happens if we have detected the inner rel is unique.
5073 : * We should therefore adjust some of the cost components for this effect.
5074 : * This function computes some estimates needed for these adjustments.
5075 : * These estimates will be the same regardless of the particular paths used
5076 : * for the outer and inner relation, so we compute these once and then pass
5077 : * them to all the join cost estimation functions.
5078 : *
5079 : * Input parameters:
5080 : * joinrel: join relation under consideration
5081 : * outerrel: outer relation under consideration
5082 : * innerrel: inner relation under consideration
5083 : * jointype: if not JOIN_SEMI or JOIN_ANTI, we assume it's inner_unique
5084 : * sjinfo: SpecialJoinInfo relevant to this join
5085 : * restrictlist: join quals
5086 : * Output parameters:
5087 : * *semifactors is filled in (see pathnodes.h for field definitions)
5088 : */
5089 : void
5090 176318 : compute_semi_anti_join_factors(PlannerInfo *root,
5091 : RelOptInfo *joinrel,
5092 : RelOptInfo *outerrel,
5093 : RelOptInfo *innerrel,
5094 : JoinType jointype,
5095 : SpecialJoinInfo *sjinfo,
5096 : List *restrictlist,
5097 : SemiAntiJoinFactors *semifactors)
5098 : {
5099 : Selectivity jselec;
5100 : Selectivity nselec;
5101 : Selectivity avgmatch;
5102 : SpecialJoinInfo norm_sjinfo;
5103 : List *joinquals;
5104 : ListCell *l;
5105 :
5106 : /*
5107 : * In an ANTI join, we must ignore clauses that are "pushed down", since
5108 : * those won't affect the match logic. In a SEMI join, we do not
5109 : * distinguish joinquals from "pushed down" quals, so just use the whole
5110 : * restrictinfo list. For other outer join types, we should consider only
5111 : * non-pushed-down quals, so that this devolves to an IS_OUTER_JOIN check.
5112 : */
5113 176318 : if (IS_OUTER_JOIN(jointype))
5114 : {
5115 66130 : joinquals = NIL;
5116 143806 : foreach(l, restrictlist)
5117 : {
5118 77676 : RestrictInfo *rinfo = lfirst_node(RestrictInfo, l);
5119 :
5120 77676 : if (!RINFO_IS_PUSHED_DOWN(rinfo, joinrel->relids))
5121 73798 : joinquals = lappend(joinquals, rinfo);
5122 : }
5123 : }
5124 : else
5125 110188 : joinquals = restrictlist;
5126 :
5127 : /*
5128 : * Get the JOIN_SEMI or JOIN_ANTI selectivity of the join clauses.
5129 : */
5130 176318 : jselec = clauselist_selectivity(root,
5131 : joinquals,
5132 : 0,
5133 : (jointype == JOIN_ANTI) ? JOIN_ANTI : JOIN_SEMI,
5134 : sjinfo);
5135 :
5136 : /*
5137 : * Also get the normal inner-join selectivity of the join clauses.
5138 : */
5139 176318 : init_dummy_sjinfo(&norm_sjinfo, outerrel->relids, innerrel->relids);
5140 :
5141 176318 : nselec = clauselist_selectivity(root,
5142 : joinquals,
5143 : 0,
5144 : JOIN_INNER,
5145 : &norm_sjinfo);
5146 :
5147 : /* Avoid leaking a lot of ListCells */
5148 176318 : if (IS_OUTER_JOIN(jointype))
5149 66130 : list_free(joinquals);
5150 :
5151 : /*
5152 : * jselec can be interpreted as the fraction of outer-rel rows that have
5153 : * any matches (this is true for both SEMI and ANTI cases). And nselec is
5154 : * the fraction of the Cartesian product that matches. So, the average
5155 : * number of matches for each outer-rel row that has at least one match is
5156 : * nselec * inner_rows / jselec.
5157 : *
5158 : * Note: it is correct to use the inner rel's "rows" count here, even
5159 : * though we might later be considering a parameterized inner path with
5160 : * fewer rows. This is because we have included all the join clauses in
5161 : * the selectivity estimate.
5162 : */
5163 176318 : if (jselec > 0) /* protect against zero divide */
5164 : {
5165 175978 : avgmatch = nselec * innerrel->rows / jselec;
5166 : /* Clamp to sane range */
5167 175978 : avgmatch = Max(1.0, avgmatch);
5168 : }
5169 : else
5170 340 : avgmatch = 1.0;
5171 :
5172 176318 : semifactors->outer_match_frac = jselec;
5173 176318 : semifactors->match_count = avgmatch;
5174 176318 : }
5175 :
5176 : /*
5177 : * has_indexed_join_quals
5178 : * Check whether all the joinquals of a nestloop join are used as
5179 : * inner index quals.
5180 : *
5181 : * If the inner path of a SEMI/ANTI join is an indexscan (including bitmap
5182 : * indexscan) that uses all the joinquals as indexquals, we can assume that an
5183 : * unmatched outer tuple is cheap to process, whereas otherwise it's probably
5184 : * expensive.
5185 : */
5186 : static bool
5187 732470 : has_indexed_join_quals(NestPath *path)
5188 : {
5189 732470 : JoinPath *joinpath = &path->jpath;
5190 732470 : Relids joinrelids = joinpath->path.parent->relids;
5191 732470 : Path *innerpath = joinpath->innerjoinpath;
5192 : List *indexclauses;
5193 : bool found_one;
5194 : ListCell *lc;
5195 :
5196 : /* If join still has quals to evaluate, it's not fast */
5197 732470 : if (joinpath->joinrestrictinfo != NIL)
5198 514618 : return false;
5199 : /* Nor if the inner path isn't parameterized at all */
5200 217852 : if (innerpath->param_info == NULL)
5201 4800 : return false;
5202 :
5203 : /* Find the indexclauses list for the inner scan */
5204 213052 : switch (innerpath->pathtype)
5205 : {
5206 132260 : case T_IndexScan:
5207 : case T_IndexOnlyScan:
5208 132260 : indexclauses = ((IndexPath *) innerpath)->indexclauses;
5209 132260 : break;
5210 270 : case T_BitmapHeapScan:
5211 : {
5212 : /* Accept only a simple bitmap scan, not AND/OR cases */
5213 270 : Path *bmqual = ((BitmapHeapPath *) innerpath)->bitmapqual;
5214 :
5215 270 : if (IsA(bmqual, IndexPath))
5216 222 : indexclauses = ((IndexPath *) bmqual)->indexclauses;
5217 : else
5218 48 : return false;
5219 222 : break;
5220 : }
5221 80522 : default:
5222 :
5223 : /*
5224 : * If it's not a simple indexscan, it probably doesn't run quickly
5225 : * for zero rows out, even if it's a parameterized path using all
5226 : * the joinquals.
5227 : */
5228 80522 : return false;
5229 : }
5230 :
5231 : /*
5232 : * Examine the inner path's param clauses. Any that are from the outer
5233 : * path must be found in the indexclauses list, either exactly or in an
5234 : * equivalent form generated by equivclass.c. Also, we must find at least
5235 : * one such clause, else it's a clauseless join which isn't fast.
5236 : */
5237 132482 : found_one = false;
5238 264066 : foreach(lc, innerpath->param_info->ppi_clauses)
5239 : {
5240 135086 : RestrictInfo *rinfo = (RestrictInfo *) lfirst(lc);
5241 :
5242 135086 : if (join_clause_is_movable_into(rinfo,
5243 135086 : innerpath->parent->relids,
5244 : joinrelids))
5245 : {
5246 134618 : if (!is_redundant_with_indexclauses(rinfo, indexclauses))
5247 3502 : return false;
5248 131116 : found_one = true;
5249 : }
5250 : }
5251 128980 : return found_one;
5252 : }
5253 :
5254 :
5255 : /*
5256 : * approx_tuple_count
5257 : * Quick-and-dirty estimation of the number of join rows passing
5258 : * a set of qual conditions.
5259 : *
5260 : * The quals can be either an implicitly-ANDed list of boolean expressions,
5261 : * or a list of RestrictInfo nodes (typically the latter).
5262 : *
5263 : * We intentionally compute the selectivity under JOIN_INNER rules, even
5264 : * if it's some type of outer join. This is appropriate because we are
5265 : * trying to figure out how many tuples pass the initial merge or hash
5266 : * join step.
5267 : *
5268 : * This is quick-and-dirty because we bypass clauselist_selectivity, and
5269 : * simply multiply the independent clause selectivities together. Now
5270 : * clauselist_selectivity often can't do any better than that anyhow, but
5271 : * for some situations (such as range constraints) it is smarter. However,
5272 : * we can't effectively cache the results of clauselist_selectivity, whereas
5273 : * the individual clause selectivities can be and are cached.
5274 : *
5275 : * Since we are only using the results to estimate how many potential
5276 : * output tuples are generated and passed through qpqual checking, it
5277 : * seems OK to live with the approximation.
5278 : */
5279 : static double
5280 405232 : approx_tuple_count(PlannerInfo *root, JoinPath *path, List *quals)
5281 : {
5282 : double tuples;
5283 405232 : double outer_tuples = path->outerjoinpath->rows;
5284 405232 : double inner_tuples = path->innerjoinpath->rows;
5285 : SpecialJoinInfo sjinfo;
5286 405232 : Selectivity selec = 1.0;
5287 : ListCell *l;
5288 :
5289 : /*
5290 : * Make up a SpecialJoinInfo for JOIN_INNER semantics.
5291 : */
5292 405232 : init_dummy_sjinfo(&sjinfo, path->outerjoinpath->parent->relids,
5293 405232 : path->innerjoinpath->parent->relids);
5294 :
5295 : /* Get the approximate selectivity */
5296 867110 : foreach(l, quals)
5297 : {
5298 461878 : Node *qual = (Node *) lfirst(l);
5299 :
5300 : /* Note that clause_selectivity will be able to cache its result */
5301 461878 : selec *= clause_selectivity(root, qual, 0, JOIN_INNER, &sjinfo);
5302 : }
5303 :
5304 : /* Apply it to the input relation sizes */
5305 405232 : tuples = selec * outer_tuples * inner_tuples;
5306 :
5307 405232 : return clamp_row_est(tuples);
5308 : }
5309 :
5310 :
5311 : /*
5312 : * set_baserel_size_estimates
5313 : * Set the size estimates for the given base relation.
5314 : *
5315 : * The rel's targetlist and restrictinfo list must have been constructed
5316 : * already, and rel->tuples must be set.
5317 : *
5318 : * We set the following fields of the rel node:
5319 : * rows: the estimated number of output tuples (after applying
5320 : * restriction clauses).
5321 : * width: the estimated average output tuple width in bytes.
5322 : * baserestrictcost: estimated cost of evaluating baserestrictinfo clauses.
5323 : */
5324 : void
5325 444384 : set_baserel_size_estimates(PlannerInfo *root, RelOptInfo *rel)
5326 : {
5327 : double nrows;
5328 :
5329 : /* Should only be applied to base relations */
5330 : Assert(rel->relid > 0);
5331 :
5332 888744 : nrows = rel->tuples *
5333 444384 : clauselist_selectivity(root,
5334 : rel->baserestrictinfo,
5335 : 0,
5336 : JOIN_INNER,
5337 : NULL);
5338 :
5339 444360 : rel->rows = clamp_row_est(nrows);
5340 :
5341 444360 : cost_qual_eval(&rel->baserestrictcost, rel->baserestrictinfo, root);
5342 :
5343 444360 : set_rel_width(root, rel);
5344 444360 : }
5345 :
5346 : /*
5347 : * get_parameterized_baserel_size
5348 : * Make a size estimate for a parameterized scan of a base relation.
5349 : *
5350 : * 'param_clauses' lists the additional join clauses to be used.
5351 : *
5352 : * set_baserel_size_estimates must have been applied already.
5353 : */
5354 : double
5355 131634 : get_parameterized_baserel_size(PlannerInfo *root, RelOptInfo *rel,
5356 : List *param_clauses)
5357 : {
5358 : List *allclauses;
5359 : double nrows;
5360 :
5361 : /*
5362 : * Estimate the number of rows returned by the parameterized scan, knowing
5363 : * that it will apply all the extra join clauses as well as the rel's own
5364 : * restriction clauses. Note that we force the clauses to be treated as
5365 : * non-join clauses during selectivity estimation.
5366 : */
5367 131634 : allclauses = list_concat_copy(param_clauses, rel->baserestrictinfo);
5368 263268 : nrows = rel->tuples *
5369 131634 : clauselist_selectivity(root,
5370 : allclauses,
5371 131634 : rel->relid, /* do not use 0! */
5372 : JOIN_INNER,
5373 : NULL);
5374 131634 : nrows = clamp_row_est(nrows);
5375 : /* For safety, make sure result is not more than the base estimate */
5376 131634 : if (nrows > rel->rows)
5377 0 : nrows = rel->rows;
5378 131634 : return nrows;
5379 : }
5380 :
5381 : /*
5382 : * set_joinrel_size_estimates
5383 : * Set the size estimates for the given join relation.
5384 : *
5385 : * The rel's targetlist must have been constructed already, and a
5386 : * restriction clause list that matches the given component rels must
5387 : * be provided.
5388 : *
5389 : * Since there is more than one way to make a joinrel for more than two
5390 : * base relations, the results we get here could depend on which component
5391 : * rel pair is provided. In theory we should get the same answers no matter
5392 : * which pair is provided; in practice, since the selectivity estimation
5393 : * routines don't handle all cases equally well, we might not. But there's
5394 : * not much to be done about it. (Would it make sense to repeat the
5395 : * calculations for each pair of input rels that's encountered, and somehow
5396 : * average the results? Probably way more trouble than it's worth, and
5397 : * anyway we must keep the rowcount estimate the same for all paths for the
5398 : * joinrel.)
5399 : *
5400 : * We set only the rows field here. The reltarget field was already set by
5401 : * build_joinrel_tlist, and baserestrictcost is not used for join rels.
5402 : */
5403 : void
5404 185790 : set_joinrel_size_estimates(PlannerInfo *root, RelOptInfo *rel,
5405 : RelOptInfo *outer_rel,
5406 : RelOptInfo *inner_rel,
5407 : SpecialJoinInfo *sjinfo,
5408 : List *restrictlist)
5409 : {
5410 185790 : rel->rows = calc_joinrel_size_estimate(root,
5411 : rel,
5412 : outer_rel,
5413 : inner_rel,
5414 : outer_rel->rows,
5415 : inner_rel->rows,
5416 : sjinfo,
5417 : restrictlist);
5418 185790 : }
5419 :
5420 : /*
5421 : * get_parameterized_joinrel_size
5422 : * Make a size estimate for a parameterized scan of a join relation.
5423 : *
5424 : * 'rel' is the joinrel under consideration.
5425 : * 'outer_path', 'inner_path' are (probably also parameterized) Paths that
5426 : * produce the relations being joined.
5427 : * 'sjinfo' is any SpecialJoinInfo relevant to this join.
5428 : * 'restrict_clauses' lists the join clauses that need to be applied at the
5429 : * join node (including any movable clauses that were moved down to this join,
5430 : * and not including any movable clauses that were pushed down into the
5431 : * child paths).
5432 : *
5433 : * set_joinrel_size_estimates must have been applied already.
5434 : */
5435 : double
5436 6920 : get_parameterized_joinrel_size(PlannerInfo *root, RelOptInfo *rel,
5437 : Path *outer_path,
5438 : Path *inner_path,
5439 : SpecialJoinInfo *sjinfo,
5440 : List *restrict_clauses)
5441 : {
5442 : double nrows;
5443 :
5444 : /*
5445 : * Estimate the number of rows returned by the parameterized join as the
5446 : * sizes of the input paths times the selectivity of the clauses that have
5447 : * ended up at this join node.
5448 : *
5449 : * As with set_joinrel_size_estimates, the rowcount estimate could depend
5450 : * on the pair of input paths provided, though ideally we'd get the same
5451 : * estimate for any pair with the same parameterization.
5452 : */
5453 6920 : nrows = calc_joinrel_size_estimate(root,
5454 : rel,
5455 : outer_path->parent,
5456 : inner_path->parent,
5457 : outer_path->rows,
5458 : inner_path->rows,
5459 : sjinfo,
5460 : restrict_clauses);
5461 : /* For safety, make sure result is not more than the base estimate */
5462 6920 : if (nrows > rel->rows)
5463 12 : nrows = rel->rows;
5464 6920 : return nrows;
5465 : }
5466 :
5467 : /*
5468 : * calc_joinrel_size_estimate
5469 : * Workhorse for set_joinrel_size_estimates and
5470 : * get_parameterized_joinrel_size.
5471 : *
5472 : * outer_rel/inner_rel are the relations being joined, but they should be
5473 : * assumed to have sizes outer_rows/inner_rows; those numbers might be less
5474 : * than what rel->rows says, when we are considering parameterized paths.
5475 : */
5476 : static double
5477 192710 : calc_joinrel_size_estimate(PlannerInfo *root,
5478 : RelOptInfo *joinrel,
5479 : RelOptInfo *outer_rel,
5480 : RelOptInfo *inner_rel,
5481 : double outer_rows,
5482 : double inner_rows,
5483 : SpecialJoinInfo *sjinfo,
5484 : List *restrictlist)
5485 : {
5486 192710 : JoinType jointype = sjinfo->jointype;
5487 : Selectivity fkselec;
5488 : Selectivity jselec;
5489 : Selectivity pselec;
5490 : double nrows;
5491 :
5492 : /*
5493 : * Compute joinclause selectivity. Note that we are only considering
5494 : * clauses that become restriction clauses at this join level; we are not
5495 : * double-counting them because they were not considered in estimating the
5496 : * sizes of the component rels.
5497 : *
5498 : * First, see whether any of the joinclauses can be matched to known FK
5499 : * constraints. If so, drop those clauses from the restrictlist, and
5500 : * instead estimate their selectivity using FK semantics. (We do this
5501 : * without regard to whether said clauses are local or "pushed down".
5502 : * Probably, an FK-matching clause could never be seen as pushed down at
5503 : * an outer join, since it would be strict and hence would be grounds for
5504 : * join strength reduction.) fkselec gets the net selectivity for
5505 : * FK-matching clauses, or 1.0 if there are none.
5506 : */
5507 192710 : fkselec = get_foreign_key_join_selectivity(root,
5508 : outer_rel->relids,
5509 : inner_rel->relids,
5510 : sjinfo,
5511 : &restrictlist);
5512 :
5513 : /*
5514 : * For an outer join, we have to distinguish the selectivity of the join's
5515 : * own clauses (JOIN/ON conditions) from any clauses that were "pushed
5516 : * down". For inner joins we just count them all as joinclauses.
5517 : */
5518 192710 : if (IS_OUTER_JOIN(jointype))
5519 : {
5520 70412 : List *joinquals = NIL;
5521 70412 : List *pushedquals = NIL;
5522 : ListCell *l;
5523 :
5524 : /* Grovel through the clauses to separate into two lists */
5525 156814 : foreach(l, restrictlist)
5526 : {
5527 86402 : RestrictInfo *rinfo = lfirst_node(RestrictInfo, l);
5528 :
5529 86402 : if (RINFO_IS_PUSHED_DOWN(rinfo, joinrel->relids))
5530 3744 : pushedquals = lappend(pushedquals, rinfo);
5531 : else
5532 82658 : joinquals = lappend(joinquals, rinfo);
5533 : }
5534 :
5535 : /* Get the separate selectivities */
5536 70412 : jselec = clauselist_selectivity(root,
5537 : joinquals,
5538 : 0,
5539 : jointype,
5540 : sjinfo);
5541 70412 : pselec = clauselist_selectivity(root,
5542 : pushedquals,
5543 : 0,
5544 : jointype,
5545 : sjinfo);
5546 :
5547 : /* Avoid leaking a lot of ListCells */
5548 70412 : list_free(joinquals);
5549 70412 : list_free(pushedquals);
5550 : }
5551 : else
5552 : {
5553 122298 : jselec = clauselist_selectivity(root,
5554 : restrictlist,
5555 : 0,
5556 : jointype,
5557 : sjinfo);
5558 122298 : pselec = 0.0; /* not used, keep compiler quiet */
5559 : }
5560 :
5561 : /*
5562 : * Basically, we multiply size of Cartesian product by selectivity.
5563 : *
5564 : * If we are doing an outer join, take that into account: the joinqual
5565 : * selectivity has to be clamped using the knowledge that the output must
5566 : * be at least as large as the non-nullable input. However, any
5567 : * pushed-down quals are applied after the outer join, so their
5568 : * selectivity applies fully.
5569 : *
5570 : * For JOIN_SEMI and JOIN_ANTI, the selectivity is defined as the fraction
5571 : * of LHS rows that have matches, and we apply that straightforwardly.
5572 : */
5573 192710 : switch (jointype)
5574 : {
5575 117676 : case JOIN_INNER:
5576 117676 : nrows = outer_rows * inner_rows * fkselec * jselec;
5577 : /* pselec not used */
5578 117676 : break;
5579 64220 : case JOIN_LEFT:
5580 64220 : nrows = outer_rows * inner_rows * fkselec * jselec;
5581 64220 : if (nrows < outer_rows)
5582 21410 : nrows = outer_rows;
5583 64220 : nrows *= pselec;
5584 64220 : break;
5585 1690 : case JOIN_FULL:
5586 1690 : nrows = outer_rows * inner_rows * fkselec * jselec;
5587 1690 : if (nrows < outer_rows)
5588 1080 : nrows = outer_rows;
5589 1690 : if (nrows < inner_rows)
5590 120 : nrows = inner_rows;
5591 1690 : nrows *= pselec;
5592 1690 : break;
5593 4622 : case JOIN_SEMI:
5594 4622 : nrows = outer_rows * fkselec * jselec;
5595 : /* pselec not used */
5596 4622 : break;
5597 4502 : case JOIN_ANTI:
5598 4502 : nrows = outer_rows * (1.0 - fkselec * jselec);
5599 4502 : nrows *= pselec;
5600 4502 : break;
5601 0 : default:
5602 : /* other values not expected here */
5603 0 : elog(ERROR, "unrecognized join type: %d", (int) jointype);
5604 : nrows = 0; /* keep compiler quiet */
5605 : break;
5606 : }
5607 :
5608 192710 : return clamp_row_est(nrows);
5609 : }
5610 :
5611 : /*
5612 : * get_foreign_key_join_selectivity
5613 : * Estimate join selectivity for foreign-key-related clauses.
5614 : *
5615 : * Remove any clauses that can be matched to FK constraints from *restrictlist,
5616 : * and return a substitute estimate of their selectivity. 1.0 is returned
5617 : * when there are no such clauses.
5618 : *
5619 : * The reason for treating such clauses specially is that we can get better
5620 : * estimates this way than by relying on clauselist_selectivity(), especially
5621 : * for multi-column FKs where that function's assumption that the clauses are
5622 : * independent falls down badly. But even with single-column FKs, we may be
5623 : * able to get a better answer when the pg_statistic stats are missing or out
5624 : * of date.
5625 : */
5626 : static Selectivity
5627 192710 : get_foreign_key_join_selectivity(PlannerInfo *root,
5628 : Relids outer_relids,
5629 : Relids inner_relids,
5630 : SpecialJoinInfo *sjinfo,
5631 : List **restrictlist)
5632 : {
5633 192710 : Selectivity fkselec = 1.0;
5634 192710 : JoinType jointype = sjinfo->jointype;
5635 192710 : List *worklist = *restrictlist;
5636 : ListCell *lc;
5637 :
5638 : /* Consider each FK constraint that is known to match the query */
5639 194622 : foreach(lc, root->fkey_list)
5640 : {
5641 1912 : ForeignKeyOptInfo *fkinfo = (ForeignKeyOptInfo *) lfirst(lc);
5642 : bool ref_is_outer;
5643 : List *removedlist;
5644 : ListCell *cell;
5645 :
5646 : /*
5647 : * This FK is not relevant unless it connects a baserel on one side of
5648 : * this join to a baserel on the other side.
5649 : */
5650 3472 : if (bms_is_member(fkinfo->con_relid, outer_relids) &&
5651 1560 : bms_is_member(fkinfo->ref_relid, inner_relids))
5652 1368 : ref_is_outer = false;
5653 896 : else if (bms_is_member(fkinfo->ref_relid, outer_relids) &&
5654 352 : bms_is_member(fkinfo->con_relid, inner_relids))
5655 130 : ref_is_outer = true;
5656 : else
5657 414 : continue;
5658 :
5659 : /*
5660 : * If we're dealing with a semi/anti join, and the FK's referenced
5661 : * relation is on the outside, then knowledge of the FK doesn't help
5662 : * us figure out what we need to know (which is the fraction of outer
5663 : * rows that have matches). On the other hand, if the referenced rel
5664 : * is on the inside, then all outer rows must have matches in the
5665 : * referenced table (ignoring nulls). But any restriction or join
5666 : * clauses that filter that table will reduce the fraction of matches.
5667 : * We can account for restriction clauses, but it's too hard to guess
5668 : * how many table rows would get through a join that's inside the RHS.
5669 : * Hence, if either case applies, punt and ignore the FK.
5670 : */
5671 1498 : if ((jointype == JOIN_SEMI || jointype == JOIN_ANTI) &&
5672 976 : (ref_is_outer || bms_membership(inner_relids) != BMS_SINGLETON))
5673 12 : continue;
5674 :
5675 : /*
5676 : * Modify the restrictlist by removing clauses that match the FK (and
5677 : * putting them into removedlist instead). It seems unsafe to modify
5678 : * the originally-passed List structure, so we make a shallow copy the
5679 : * first time through.
5680 : */
5681 1486 : if (worklist == *restrictlist)
5682 1262 : worklist = list_copy(worklist);
5683 :
5684 1486 : removedlist = NIL;
5685 3048 : foreach(cell, worklist)
5686 : {
5687 1562 : RestrictInfo *rinfo = (RestrictInfo *) lfirst(cell);
5688 1562 : bool remove_it = false;
5689 : int i;
5690 :
5691 : /* Drop this clause if it matches any column of the FK */
5692 1948 : for (i = 0; i < fkinfo->nkeys; i++)
5693 : {
5694 1918 : if (rinfo->parent_ec)
5695 : {
5696 : /*
5697 : * EC-derived clauses can only match by EC. It is okay to
5698 : * consider any clause derived from the same EC as
5699 : * matching the FK: even if equivclass.c chose to generate
5700 : * a clause equating some other pair of Vars, it could
5701 : * have generated one equating the FK's Vars. So for
5702 : * purposes of estimation, we can act as though it did so.
5703 : *
5704 : * Note: checking parent_ec is a bit of a cheat because
5705 : * there are EC-derived clauses that don't have parent_ec
5706 : * set; but such clauses must compare expressions that
5707 : * aren't just Vars, so they cannot match the FK anyway.
5708 : */
5709 304 : if (fkinfo->eclass[i] == rinfo->parent_ec)
5710 : {
5711 298 : remove_it = true;
5712 298 : break;
5713 : }
5714 : }
5715 : else
5716 : {
5717 : /*
5718 : * Otherwise, see if rinfo was previously matched to FK as
5719 : * a "loose" clause.
5720 : */
5721 1614 : if (list_member_ptr(fkinfo->rinfos[i], rinfo))
5722 : {
5723 1234 : remove_it = true;
5724 1234 : break;
5725 : }
5726 : }
5727 : }
5728 1562 : if (remove_it)
5729 : {
5730 1532 : worklist = foreach_delete_current(worklist, cell);
5731 1532 : removedlist = lappend(removedlist, rinfo);
5732 : }
5733 : }
5734 :
5735 : /*
5736 : * If we failed to remove all the matching clauses we expected to
5737 : * find, chicken out and ignore this FK; applying its selectivity
5738 : * might result in double-counting. Put any clauses we did manage to
5739 : * remove back into the worklist.
5740 : *
5741 : * Since the matching clauses are known not outerjoin-delayed, they
5742 : * would normally have appeared in the initial joinclause list. If we
5743 : * didn't find them, there are two possibilities:
5744 : *
5745 : * 1. If the FK match is based on an EC that is ec_has_const, it won't
5746 : * have generated any join clauses at all. We discount such ECs while
5747 : * checking to see if we have "all" the clauses. (Below, we'll adjust
5748 : * the selectivity estimate for this case.)
5749 : *
5750 : * 2. The clauses were matched to some other FK in a previous
5751 : * iteration of this loop, and thus removed from worklist. (A likely
5752 : * case is that two FKs are matched to the same EC; there will be only
5753 : * one EC-derived clause in the initial list, so the first FK will
5754 : * consume it.) Applying both FKs' selectivity independently risks
5755 : * underestimating the join size; in particular, this would undo one
5756 : * of the main things that ECs were invented for, namely to avoid
5757 : * double-counting the selectivity of redundant equality conditions.
5758 : * Later we might think of a reasonable way to combine the estimates,
5759 : * but for now, just punt, since this is a fairly uncommon situation.
5760 : */
5761 1486 : if (removedlist == NIL ||
5762 1200 : list_length(removedlist) !=
5763 1200 : (fkinfo->nmatched_ec - fkinfo->nconst_ec + fkinfo->nmatched_ri))
5764 : {
5765 286 : worklist = list_concat(worklist, removedlist);
5766 286 : continue;
5767 : }
5768 :
5769 : /*
5770 : * Finally we get to the payoff: estimate selectivity using the
5771 : * knowledge that each referencing row will match exactly one row in
5772 : * the referenced table.
5773 : *
5774 : * XXX that's not true in the presence of nulls in the referencing
5775 : * column(s), so in principle we should derate the estimate for those.
5776 : * However (1) if there are any strict restriction clauses for the
5777 : * referencing column(s) elsewhere in the query, derating here would
5778 : * be double-counting the null fraction, and (2) it's not very clear
5779 : * how to combine null fractions for multiple referencing columns. So
5780 : * we do nothing for now about correcting for nulls.
5781 : *
5782 : * XXX another point here is that if either side of an FK constraint
5783 : * is an inheritance parent, we estimate as though the constraint
5784 : * covers all its children as well. This is not an unreasonable
5785 : * assumption for a referencing table, ie the user probably applied
5786 : * identical constraints to all child tables (though perhaps we ought
5787 : * to check that). But it's not possible to have done that for a
5788 : * referenced table. Fortunately, precisely because that doesn't
5789 : * work, it is uncommon in practice to have an FK referencing a parent
5790 : * table. So, at least for now, disregard inheritance here.
5791 : */
5792 1200 : if (jointype == JOIN_SEMI || jointype == JOIN_ANTI)
5793 752 : {
5794 : /*
5795 : * For JOIN_SEMI and JOIN_ANTI, we only get here when the FK's
5796 : * referenced table is exactly the inside of the join. The join
5797 : * selectivity is defined as the fraction of LHS rows that have
5798 : * matches. The FK implies that every LHS row has a match *in the
5799 : * referenced table*; but any restriction clauses on it will
5800 : * reduce the number of matches. Hence we take the join
5801 : * selectivity as equal to the selectivity of the table's
5802 : * restriction clauses, which is rows / tuples; but we must guard
5803 : * against tuples == 0.
5804 : */
5805 752 : RelOptInfo *ref_rel = find_base_rel(root, fkinfo->ref_relid);
5806 752 : double ref_tuples = Max(ref_rel->tuples, 1.0);
5807 :
5808 752 : fkselec *= ref_rel->rows / ref_tuples;
5809 : }
5810 : else
5811 : {
5812 : /*
5813 : * Otherwise, selectivity is exactly 1/referenced-table-size; but
5814 : * guard against tuples == 0. Note we should use the raw table
5815 : * tuple count, not any estimate of its filtered or joined size.
5816 : */
5817 448 : RelOptInfo *ref_rel = find_base_rel(root, fkinfo->ref_relid);
5818 448 : double ref_tuples = Max(ref_rel->tuples, 1.0);
5819 :
5820 448 : fkselec *= 1.0 / ref_tuples;
5821 : }
5822 :
5823 : /*
5824 : * If any of the FK columns participated in ec_has_const ECs, then
5825 : * equivclass.c will have generated "var = const" restrictions for
5826 : * each side of the join, thus reducing the sizes of both input
5827 : * relations. Taking the fkselec at face value would amount to
5828 : * double-counting the selectivity of the constant restriction for the
5829 : * referencing Var. Hence, look for the restriction clause(s) that
5830 : * were applied to the referencing Var(s), and divide out their
5831 : * selectivity to correct for this.
5832 : */
5833 1200 : if (fkinfo->nconst_ec > 0)
5834 : {
5835 24 : for (int i = 0; i < fkinfo->nkeys; i++)
5836 : {
5837 18 : EquivalenceClass *ec = fkinfo->eclass[i];
5838 :
5839 18 : if (ec && ec->ec_has_const)
5840 : {
5841 6 : EquivalenceMember *em = fkinfo->fk_eclass_member[i];
5842 6 : RestrictInfo *rinfo = find_derived_clause_for_ec_member(ec,
5843 : em);
5844 :
5845 6 : if (rinfo)
5846 : {
5847 : Selectivity s0;
5848 :
5849 6 : s0 = clause_selectivity(root,
5850 : (Node *) rinfo,
5851 : 0,
5852 : jointype,
5853 : sjinfo);
5854 6 : if (s0 > 0)
5855 6 : fkselec /= s0;
5856 : }
5857 : }
5858 : }
5859 : }
5860 : }
5861 :
5862 192710 : *restrictlist = worklist;
5863 192710 : CLAMP_PROBABILITY(fkselec);
5864 192710 : return fkselec;
5865 : }
5866 :
5867 : /*
5868 : * set_subquery_size_estimates
5869 : * Set the size estimates for a base relation that is a subquery.
5870 : *
5871 : * The rel's targetlist and restrictinfo list must have been constructed
5872 : * already, and the Paths for the subquery must have been completed.
5873 : * We look at the subquery's PlannerInfo to extract data.
5874 : *
5875 : * We set the same fields as set_baserel_size_estimates.
5876 : */
5877 : void
5878 22652 : set_subquery_size_estimates(PlannerInfo *root, RelOptInfo *rel)
5879 : {
5880 22652 : PlannerInfo *subroot = rel->subroot;
5881 : RelOptInfo *sub_final_rel;
5882 : ListCell *lc;
5883 :
5884 : /* Should only be applied to base relations that are subqueries */
5885 : Assert(rel->relid > 0);
5886 : Assert(planner_rt_fetch(rel->relid, root)->rtekind == RTE_SUBQUERY);
5887 :
5888 : /*
5889 : * Copy raw number of output rows from subquery. All of its paths should
5890 : * have the same output rowcount, so just look at cheapest-total.
5891 : */
5892 22652 : sub_final_rel = fetch_upper_rel(subroot, UPPERREL_FINAL, NULL);
5893 22652 : rel->tuples = sub_final_rel->cheapest_total_path->rows;
5894 :
5895 : /*
5896 : * Compute per-output-column width estimates by examining the subquery's
5897 : * targetlist. For any output that is a plain Var, get the width estimate
5898 : * that was made while planning the subquery. Otherwise, we leave it to
5899 : * set_rel_width to fill in a datatype-based default estimate.
5900 : */
5901 96916 : foreach(lc, subroot->parse->targetList)
5902 : {
5903 74264 : TargetEntry *te = lfirst_node(TargetEntry, lc);
5904 74264 : Node *texpr = (Node *) te->expr;
5905 74264 : int32 item_width = 0;
5906 :
5907 : /* junk columns aren't visible to upper query */
5908 74264 : if (te->resjunk)
5909 1372 : continue;
5910 :
5911 : /*
5912 : * The subquery could be an expansion of a view that's had columns
5913 : * added to it since the current query was parsed, so that there are
5914 : * non-junk tlist columns in it that don't correspond to any column
5915 : * visible at our query level. Ignore such columns.
5916 : */
5917 72892 : if (te->resno < rel->min_attr || te->resno > rel->max_attr)
5918 0 : continue;
5919 :
5920 : /*
5921 : * XXX This currently doesn't work for subqueries containing set
5922 : * operations, because the Vars in their tlists are bogus references
5923 : * to the first leaf subquery, which wouldn't give the right answer
5924 : * even if we could still get to its PlannerInfo.
5925 : *
5926 : * Also, the subquery could be an appendrel for which all branches are
5927 : * known empty due to constraint exclusion, in which case
5928 : * set_append_rel_pathlist will have left the attr_widths set to zero.
5929 : *
5930 : * In either case, we just leave the width estimate zero until
5931 : * set_rel_width fixes it.
5932 : */
5933 72892 : if (IsA(texpr, Var) &&
5934 33982 : subroot->parse->setOperations == NULL)
5935 : {
5936 32330 : Var *var = (Var *) texpr;
5937 32330 : RelOptInfo *subrel = find_base_rel(subroot, var->varno);
5938 :
5939 32330 : item_width = subrel->attr_widths[var->varattno - subrel->min_attr];
5940 : }
5941 72892 : rel->attr_widths[te->resno - rel->min_attr] = item_width;
5942 : }
5943 :
5944 : /* Now estimate number of output rows, etc */
5945 22652 : set_baserel_size_estimates(root, rel);
5946 22652 : }
5947 :
5948 : /*
5949 : * set_function_size_estimates
5950 : * Set the size estimates for a base relation that is a function call.
5951 : *
5952 : * The rel's targetlist and restrictinfo list must have been constructed
5953 : * already.
5954 : *
5955 : * We set the same fields as set_baserel_size_estimates.
5956 : */
5957 : void
5958 44176 : set_function_size_estimates(PlannerInfo *root, RelOptInfo *rel)
5959 : {
5960 : RangeTblEntry *rte;
5961 : ListCell *lc;
5962 :
5963 : /* Should only be applied to base relations that are functions */
5964 : Assert(rel->relid > 0);
5965 44176 : rte = planner_rt_fetch(rel->relid, root);
5966 : Assert(rte->rtekind == RTE_FUNCTION);
5967 :
5968 : /*
5969 : * Estimate number of rows the functions will return. The rowcount of the
5970 : * node is that of the largest function result.
5971 : */
5972 44176 : rel->tuples = 0;
5973 88664 : foreach(lc, rte->functions)
5974 : {
5975 44488 : RangeTblFunction *rtfunc = (RangeTblFunction *) lfirst(lc);
5976 44488 : double ntup = expression_returns_set_rows(root, rtfunc->funcexpr);
5977 :
5978 44488 : if (ntup > rel->tuples)
5979 44200 : rel->tuples = ntup;
5980 : }
5981 :
5982 : /* Now estimate number of output rows, etc */
5983 44176 : set_baserel_size_estimates(root, rel);
5984 44176 : }
5985 :
5986 : /*
5987 : * set_function_size_estimates
5988 : * Set the size estimates for a base relation that is a function call.
5989 : *
5990 : * The rel's targetlist and restrictinfo list must have been constructed
5991 : * already.
5992 : *
5993 : * We set the same fields as set_tablefunc_size_estimates.
5994 : */
5995 : void
5996 626 : set_tablefunc_size_estimates(PlannerInfo *root, RelOptInfo *rel)
5997 : {
5998 : /* Should only be applied to base relations that are functions */
5999 : Assert(rel->relid > 0);
6000 : Assert(planner_rt_fetch(rel->relid, root)->rtekind == RTE_TABLEFUNC);
6001 :
6002 626 : rel->tuples = 100;
6003 :
6004 : /* Now estimate number of output rows, etc */
6005 626 : set_baserel_size_estimates(root, rel);
6006 626 : }
6007 :
6008 : /*
6009 : * set_values_size_estimates
6010 : * Set the size estimates for a base relation that is a values list.
6011 : *
6012 : * The rel's targetlist and restrictinfo list must have been constructed
6013 : * already.
6014 : *
6015 : * We set the same fields as set_baserel_size_estimates.
6016 : */
6017 : void
6018 7980 : set_values_size_estimates(PlannerInfo *root, RelOptInfo *rel)
6019 : {
6020 : RangeTblEntry *rte;
6021 :
6022 : /* Should only be applied to base relations that are values lists */
6023 : Assert(rel->relid > 0);
6024 7980 : rte = planner_rt_fetch(rel->relid, root);
6025 : Assert(rte->rtekind == RTE_VALUES);
6026 :
6027 : /*
6028 : * Estimate number of rows the values list will return. We know this
6029 : * precisely based on the list length (well, barring set-returning
6030 : * functions in list items, but that's a refinement not catered for
6031 : * anywhere else either).
6032 : */
6033 7980 : rel->tuples = list_length(rte->values_lists);
6034 :
6035 : /* Now estimate number of output rows, etc */
6036 7980 : set_baserel_size_estimates(root, rel);
6037 7980 : }
6038 :
6039 : /*
6040 : * set_cte_size_estimates
6041 : * Set the size estimates for a base relation that is a CTE reference.
6042 : *
6043 : * The rel's targetlist and restrictinfo list must have been constructed
6044 : * already, and we need an estimate of the number of rows returned by the CTE
6045 : * (if a regular CTE) or the non-recursive term (if a self-reference).
6046 : *
6047 : * We set the same fields as set_baserel_size_estimates.
6048 : */
6049 : void
6050 4016 : set_cte_size_estimates(PlannerInfo *root, RelOptInfo *rel, double cte_rows)
6051 : {
6052 : RangeTblEntry *rte;
6053 :
6054 : /* Should only be applied to base relations that are CTE references */
6055 : Assert(rel->relid > 0);
6056 4016 : rte = planner_rt_fetch(rel->relid, root);
6057 : Assert(rte->rtekind == RTE_CTE);
6058 :
6059 4016 : if (rte->self_reference)
6060 : {
6061 : /*
6062 : * In a self-reference, we assume the average worktable size is a
6063 : * multiple of the nonrecursive term's size. The best multiplier will
6064 : * vary depending on query "fan-out", so make its value adjustable.
6065 : */
6066 822 : rel->tuples = clamp_row_est(recursive_worktable_factor * cte_rows);
6067 : }
6068 : else
6069 : {
6070 : /* Otherwise just believe the CTE's rowcount estimate */
6071 3194 : rel->tuples = cte_rows;
6072 : }
6073 :
6074 : /* Now estimate number of output rows, etc */
6075 4016 : set_baserel_size_estimates(root, rel);
6076 4016 : }
6077 :
6078 : /*
6079 : * set_namedtuplestore_size_estimates
6080 : * Set the size estimates for a base relation that is a tuplestore reference.
6081 : *
6082 : * The rel's targetlist and restrictinfo list must have been constructed
6083 : * already.
6084 : *
6085 : * We set the same fields as set_baserel_size_estimates.
6086 : */
6087 : void
6088 438 : set_namedtuplestore_size_estimates(PlannerInfo *root, RelOptInfo *rel)
6089 : {
6090 : RangeTblEntry *rte;
6091 :
6092 : /* Should only be applied to base relations that are tuplestore references */
6093 : Assert(rel->relid > 0);
6094 438 : rte = planner_rt_fetch(rel->relid, root);
6095 : Assert(rte->rtekind == RTE_NAMEDTUPLESTORE);
6096 :
6097 : /*
6098 : * Use the estimate provided by the code which is generating the named
6099 : * tuplestore. In some cases, the actual number might be available; in
6100 : * others the same plan will be re-used, so a "typical" value might be
6101 : * estimated and used.
6102 : */
6103 438 : rel->tuples = rte->enrtuples;
6104 438 : if (rel->tuples < 0)
6105 0 : rel->tuples = 1000;
6106 :
6107 : /* Now estimate number of output rows, etc */
6108 438 : set_baserel_size_estimates(root, rel);
6109 438 : }
6110 :
6111 : /*
6112 : * set_result_size_estimates
6113 : * Set the size estimates for an RTE_RESULT base relation
6114 : *
6115 : * The rel's targetlist and restrictinfo list must have been constructed
6116 : * already.
6117 : *
6118 : * We set the same fields as set_baserel_size_estimates.
6119 : */
6120 : void
6121 1598 : set_result_size_estimates(PlannerInfo *root, RelOptInfo *rel)
6122 : {
6123 : /* Should only be applied to RTE_RESULT base relations */
6124 : Assert(rel->relid > 0);
6125 : Assert(planner_rt_fetch(rel->relid, root)->rtekind == RTE_RESULT);
6126 :
6127 : /* RTE_RESULT always generates a single row, natively */
6128 1598 : rel->tuples = 1;
6129 :
6130 : /* Now estimate number of output rows, etc */
6131 1598 : set_baserel_size_estimates(root, rel);
6132 1598 : }
6133 :
6134 : /*
6135 : * set_foreign_size_estimates
6136 : * Set the size estimates for a base relation that is a foreign table.
6137 : *
6138 : * There is not a whole lot that we can do here; the foreign-data wrapper
6139 : * is responsible for producing useful estimates. We can do a decent job
6140 : * of estimating baserestrictcost, so we set that, and we also set up width
6141 : * using what will be purely datatype-driven estimates from the targetlist.
6142 : * There is no way to do anything sane with the rows value, so we just put
6143 : * a default estimate and hope that the wrapper can improve on it. The
6144 : * wrapper's GetForeignRelSize function will be called momentarily.
6145 : *
6146 : * The rel's targetlist and restrictinfo list must have been constructed
6147 : * already.
6148 : */
6149 : void
6150 2392 : set_foreign_size_estimates(PlannerInfo *root, RelOptInfo *rel)
6151 : {
6152 : /* Should only be applied to base relations */
6153 : Assert(rel->relid > 0);
6154 :
6155 2392 : rel->rows = 1000; /* entirely bogus default estimate */
6156 :
6157 2392 : cost_qual_eval(&rel->baserestrictcost, rel->baserestrictinfo, root);
6158 :
6159 2392 : set_rel_width(root, rel);
6160 2392 : }
6161 :
6162 :
6163 : /*
6164 : * set_rel_width
6165 : * Set the estimated output width of a base relation.
6166 : *
6167 : * The estimated output width is the sum of the per-attribute width estimates
6168 : * for the actually-referenced columns, plus any PHVs or other expressions
6169 : * that have to be calculated at this relation. This is the amount of data
6170 : * we'd need to pass upwards in case of a sort, hash, etc.
6171 : *
6172 : * This function also sets reltarget->cost, so it's a bit misnamed now.
6173 : *
6174 : * NB: this works best on plain relations because it prefers to look at
6175 : * real Vars. For subqueries, set_subquery_size_estimates will already have
6176 : * copied up whatever per-column estimates were made within the subquery,
6177 : * and for other types of rels there isn't much we can do anyway. We fall
6178 : * back on (fairly stupid) datatype-based width estimates if we can't get
6179 : * any better number.
6180 : *
6181 : * The per-attribute width estimates are cached for possible re-use while
6182 : * building join relations or post-scan/join pathtargets.
6183 : */
6184 : static void
6185 446752 : set_rel_width(PlannerInfo *root, RelOptInfo *rel)
6186 : {
6187 446752 : Oid reloid = planner_rt_fetch(rel->relid, root)->relid;
6188 446752 : int64 tuple_width = 0;
6189 446752 : bool have_wholerow_var = false;
6190 : ListCell *lc;
6191 :
6192 : /* Vars are assumed to have cost zero, but other exprs do not */
6193 446752 : rel->reltarget->cost.startup = 0;
6194 446752 : rel->reltarget->cost.per_tuple = 0;
6195 :
6196 1588906 : foreach(lc, rel->reltarget->exprs)
6197 : {
6198 1142154 : Node *node = (Node *) lfirst(lc);
6199 :
6200 : /*
6201 : * Ordinarily, a Var in a rel's targetlist must belong to that rel;
6202 : * but there are corner cases involving LATERAL references where that
6203 : * isn't so. If the Var has the wrong varno, fall through to the
6204 : * generic case (it doesn't seem worth the trouble to be any smarter).
6205 : */
6206 1142154 : if (IsA(node, Var) &&
6207 1122604 : ((Var *) node)->varno == rel->relid)
6208 298408 : {
6209 1122538 : Var *var = (Var *) node;
6210 : int ndx;
6211 : int32 item_width;
6212 :
6213 : Assert(var->varattno >= rel->min_attr);
6214 : Assert(var->varattno <= rel->max_attr);
6215 :
6216 1122538 : ndx = var->varattno - rel->min_attr;
6217 :
6218 : /*
6219 : * If it's a whole-row Var, we'll deal with it below after we have
6220 : * already cached as many attr widths as possible.
6221 : */
6222 1122538 : if (var->varattno == 0)
6223 : {
6224 2976 : have_wholerow_var = true;
6225 2976 : continue;
6226 : }
6227 :
6228 : /*
6229 : * The width may have been cached already (especially if it's a
6230 : * subquery), so don't duplicate effort.
6231 : */
6232 1119562 : if (rel->attr_widths[ndx] > 0)
6233 : {
6234 226338 : tuple_width += rel->attr_widths[ndx];
6235 226338 : continue;
6236 : }
6237 :
6238 : /* Try to get column width from statistics */
6239 893224 : if (reloid != InvalidOid && var->varattno > 0)
6240 : {
6241 700086 : item_width = get_attavgwidth(reloid, var->varattno);
6242 700086 : if (item_width > 0)
6243 : {
6244 594816 : rel->attr_widths[ndx] = item_width;
6245 594816 : tuple_width += item_width;
6246 594816 : continue;
6247 : }
6248 : }
6249 :
6250 : /*
6251 : * Not a plain relation, or can't find statistics for it. Estimate
6252 : * using just the type info.
6253 : */
6254 298408 : item_width = get_typavgwidth(var->vartype, var->vartypmod);
6255 : Assert(item_width > 0);
6256 298408 : rel->attr_widths[ndx] = item_width;
6257 298408 : tuple_width += item_width;
6258 : }
6259 19616 : else if (IsA(node, PlaceHolderVar))
6260 : {
6261 : /*
6262 : * We will need to evaluate the PHV's contained expression while
6263 : * scanning this rel, so be sure to include it in reltarget->cost.
6264 : */
6265 1258 : PlaceHolderVar *phv = (PlaceHolderVar *) node;
6266 1258 : PlaceHolderInfo *phinfo = find_placeholder_info(root, phv);
6267 : QualCost cost;
6268 :
6269 1258 : tuple_width += phinfo->ph_width;
6270 1258 : cost_qual_eval_node(&cost, (Node *) phv->phexpr, root);
6271 1258 : rel->reltarget->cost.startup += cost.startup;
6272 1258 : rel->reltarget->cost.per_tuple += cost.per_tuple;
6273 : }
6274 : else
6275 : {
6276 : /*
6277 : * We could be looking at an expression pulled up from a subquery,
6278 : * or a ROW() representing a whole-row child Var, etc. Do what we
6279 : * can using the expression type information.
6280 : */
6281 : int32 item_width;
6282 : QualCost cost;
6283 :
6284 18358 : item_width = get_typavgwidth(exprType(node), exprTypmod(node));
6285 : Assert(item_width > 0);
6286 18358 : tuple_width += item_width;
6287 : /* Not entirely clear if we need to account for cost, but do so */
6288 18358 : cost_qual_eval_node(&cost, node, root);
6289 18358 : rel->reltarget->cost.startup += cost.startup;
6290 18358 : rel->reltarget->cost.per_tuple += cost.per_tuple;
6291 : }
6292 : }
6293 :
6294 : /*
6295 : * If we have a whole-row reference, estimate its width as the sum of
6296 : * per-column widths plus heap tuple header overhead.
6297 : */
6298 446752 : if (have_wholerow_var)
6299 : {
6300 2976 : int64 wholerow_width = MAXALIGN(SizeofHeapTupleHeader);
6301 :
6302 2976 : if (reloid != InvalidOid)
6303 : {
6304 : /* Real relation, so estimate true tuple width */
6305 2338 : wholerow_width += get_relation_data_width(reloid,
6306 2338 : rel->attr_widths - rel->min_attr);
6307 : }
6308 : else
6309 : {
6310 : /* Do what we can with info for a phony rel */
6311 : AttrNumber i;
6312 :
6313 1708 : for (i = 1; i <= rel->max_attr; i++)
6314 1070 : wholerow_width += rel->attr_widths[i - rel->min_attr];
6315 : }
6316 :
6317 2976 : rel->attr_widths[0 - rel->min_attr] = clamp_width_est(wholerow_width);
6318 :
6319 : /*
6320 : * Include the whole-row Var as part of the output tuple. Yes, that
6321 : * really is what happens at runtime.
6322 : */
6323 2976 : tuple_width += wholerow_width;
6324 : }
6325 :
6326 446752 : rel->reltarget->width = clamp_width_est(tuple_width);
6327 446752 : }
6328 :
6329 : /*
6330 : * set_pathtarget_cost_width
6331 : * Set the estimated eval cost and output width of a PathTarget tlist.
6332 : *
6333 : * As a notational convenience, returns the same PathTarget pointer passed in.
6334 : *
6335 : * Most, though not quite all, uses of this function occur after we've run
6336 : * set_rel_width() for base relations; so we can usually obtain cached width
6337 : * estimates for Vars. If we can't, fall back on datatype-based width
6338 : * estimates. Present early-planning uses of PathTargets don't need accurate
6339 : * widths badly enough to justify going to the catalogs for better data.
6340 : */
6341 : PathTarget *
6342 577302 : set_pathtarget_cost_width(PlannerInfo *root, PathTarget *target)
6343 : {
6344 577302 : int64 tuple_width = 0;
6345 : ListCell *lc;
6346 :
6347 : /* Vars are assumed to have cost zero, but other exprs do not */
6348 577302 : target->cost.startup = 0;
6349 577302 : target->cost.per_tuple = 0;
6350 :
6351 1935090 : foreach(lc, target->exprs)
6352 : {
6353 1357788 : Node *node = (Node *) lfirst(lc);
6354 :
6355 1357788 : tuple_width += get_expr_width(root, node);
6356 :
6357 : /* For non-Vars, account for evaluation cost */
6358 1357788 : if (!IsA(node, Var))
6359 : {
6360 : QualCost cost;
6361 :
6362 599072 : cost_qual_eval_node(&cost, node, root);
6363 599072 : target->cost.startup += cost.startup;
6364 599072 : target->cost.per_tuple += cost.per_tuple;
6365 : }
6366 : }
6367 :
6368 577302 : target->width = clamp_width_est(tuple_width);
6369 :
6370 577302 : return target;
6371 : }
6372 :
6373 : /*
6374 : * get_expr_width
6375 : * Estimate the width of the given expr attempting to use the width
6376 : * cached in a Var's owning RelOptInfo, else fallback on the type's
6377 : * average width when unable to or when the given Node is not a Var.
6378 : */
6379 : static int32
6380 1616940 : get_expr_width(PlannerInfo *root, const Node *expr)
6381 : {
6382 : int32 width;
6383 :
6384 1616940 : if (IsA(expr, Var))
6385 : {
6386 1007036 : const Var *var = (const Var *) expr;
6387 :
6388 : /* We should not see any upper-level Vars here */
6389 : Assert(var->varlevelsup == 0);
6390 :
6391 : /* Try to get data from RelOptInfo cache */
6392 1007036 : if (!IS_SPECIAL_VARNO(var->varno) &&
6393 1001534 : var->varno < root->simple_rel_array_size)
6394 : {
6395 1001534 : RelOptInfo *rel = root->simple_rel_array[var->varno];
6396 :
6397 1001534 : if (rel != NULL &&
6398 974720 : var->varattno >= rel->min_attr &&
6399 974720 : var->varattno <= rel->max_attr)
6400 : {
6401 974720 : int ndx = var->varattno - rel->min_attr;
6402 :
6403 974720 : if (rel->attr_widths[ndx] > 0)
6404 944146 : return rel->attr_widths[ndx];
6405 : }
6406 : }
6407 :
6408 : /*
6409 : * No cached data available, so estimate using just the type info.
6410 : */
6411 62890 : width = get_typavgwidth(var->vartype, var->vartypmod);
6412 : Assert(width > 0);
6413 :
6414 62890 : return width;
6415 : }
6416 :
6417 609904 : width = get_typavgwidth(exprType(expr), exprTypmod(expr));
6418 : Assert(width > 0);
6419 609904 : return width;
6420 : }
6421 :
6422 : /*
6423 : * relation_byte_size
6424 : * Estimate the storage space in bytes for a given number of tuples
6425 : * of a given width (size in bytes).
6426 : */
6427 : static double
6428 3278828 : relation_byte_size(double tuples, int width)
6429 : {
6430 3278828 : return tuples * (MAXALIGN(width) + MAXALIGN(SizeofHeapTupleHeader));
6431 : }
6432 :
6433 : /*
6434 : * page_size
6435 : * Returns an estimate of the number of pages covered by a given
6436 : * number of tuples of a given width (size in bytes).
6437 : */
6438 : static double
6439 8908 : page_size(double tuples, int width)
6440 : {
6441 8908 : return ceil(relation_byte_size(tuples, width) / BLCKSZ);
6442 : }
6443 :
6444 : /*
6445 : * Estimate the fraction of the work that each worker will do given the
6446 : * number of workers budgeted for the path.
6447 : */
6448 : static double
6449 170302 : get_parallel_divisor(Path *path)
6450 : {
6451 170302 : double parallel_divisor = path->parallel_workers;
6452 :
6453 : /*
6454 : * Early experience with parallel query suggests that when there is only
6455 : * one worker, the leader often makes a very substantial contribution to
6456 : * executing the parallel portion of the plan, but as more workers are
6457 : * added, it does less and less, because it's busy reading tuples from the
6458 : * workers and doing whatever non-parallel post-processing is needed. By
6459 : * the time we reach 4 workers, the leader no longer makes a meaningful
6460 : * contribution. Thus, for now, estimate that the leader spends 30% of
6461 : * its time servicing each worker, and the remainder executing the
6462 : * parallel plan.
6463 : */
6464 170302 : if (parallel_leader_participation)
6465 : {
6466 : double leader_contribution;
6467 :
6468 169024 : leader_contribution = 1.0 - (0.3 * path->parallel_workers);
6469 169024 : if (leader_contribution > 0)
6470 166708 : parallel_divisor += leader_contribution;
6471 : }
6472 :
6473 170302 : return parallel_divisor;
6474 : }
6475 :
6476 : /*
6477 : * compute_bitmap_pages
6478 : * Estimate number of pages fetched from heap in a bitmap heap scan.
6479 : *
6480 : * 'baserel' is the relation to be scanned
6481 : * 'bitmapqual' is a tree of IndexPaths, BitmapAndPaths, and BitmapOrPaths
6482 : * 'loop_count' is the number of repetitions of the indexscan to factor into
6483 : * estimates of caching behavior
6484 : *
6485 : * If cost_p isn't NULL, the indexTotalCost estimate is returned in *cost_p.
6486 : * If tuples_p isn't NULL, the tuples_fetched estimate is returned in *tuples_p.
6487 : */
6488 : double
6489 600644 : compute_bitmap_pages(PlannerInfo *root, RelOptInfo *baserel,
6490 : Path *bitmapqual, double loop_count,
6491 : Cost *cost_p, double *tuples_p)
6492 : {
6493 : Cost indexTotalCost;
6494 : Selectivity indexSelectivity;
6495 : double T;
6496 : double pages_fetched;
6497 : double tuples_fetched;
6498 : double heap_pages;
6499 : long maxentries;
6500 :
6501 : /*
6502 : * Fetch total cost of obtaining the bitmap, as well as its total
6503 : * selectivity.
6504 : */
6505 600644 : cost_bitmap_tree_node(bitmapqual, &indexTotalCost, &indexSelectivity);
6506 :
6507 : /*
6508 : * Estimate number of main-table pages fetched.
6509 : */
6510 600644 : tuples_fetched = clamp_row_est(indexSelectivity * baserel->tuples);
6511 :
6512 600644 : T = (baserel->pages > 1) ? (double) baserel->pages : 1.0;
6513 :
6514 : /*
6515 : * For a single scan, the number of heap pages that need to be fetched is
6516 : * the same as the Mackert and Lohman formula for the case T <= b (ie, no
6517 : * re-reads needed).
6518 : */
6519 600644 : pages_fetched = (2.0 * T * tuples_fetched) / (2.0 * T + tuples_fetched);
6520 :
6521 : /*
6522 : * Calculate the number of pages fetched from the heap. Then based on
6523 : * current work_mem estimate get the estimated maxentries in the bitmap.
6524 : * (Note that we always do this calculation based on the number of pages
6525 : * that would be fetched in a single iteration, even if loop_count > 1.
6526 : * That's correct, because only that number of entries will be stored in
6527 : * the bitmap at one time.)
6528 : */
6529 600644 : heap_pages = Min(pages_fetched, baserel->pages);
6530 600644 : maxentries = tbm_calculate_entries(work_mem * 1024L);
6531 :
6532 600644 : if (loop_count > 1)
6533 : {
6534 : /*
6535 : * For repeated bitmap scans, scale up the number of tuples fetched in
6536 : * the Mackert and Lohman formula by the number of scans, so that we
6537 : * estimate the number of pages fetched by all the scans. Then
6538 : * pro-rate for one scan.
6539 : */
6540 118382 : pages_fetched = index_pages_fetched(tuples_fetched * loop_count,
6541 : baserel->pages,
6542 : get_indexpath_pages(bitmapqual),
6543 : root);
6544 118382 : pages_fetched /= loop_count;
6545 : }
6546 :
6547 600644 : if (pages_fetched >= T)
6548 52214 : pages_fetched = T;
6549 : else
6550 548430 : pages_fetched = ceil(pages_fetched);
6551 :
6552 600644 : if (maxentries < heap_pages)
6553 : {
6554 : double exact_pages;
6555 : double lossy_pages;
6556 :
6557 : /*
6558 : * Crude approximation of the number of lossy pages. Because of the
6559 : * way tbm_lossify() is coded, the number of lossy pages increases
6560 : * very sharply as soon as we run short of memory; this formula has
6561 : * that property and seems to perform adequately in testing, but it's
6562 : * possible we could do better somehow.
6563 : */
6564 18 : lossy_pages = Max(0, heap_pages - maxentries / 2);
6565 18 : exact_pages = heap_pages - lossy_pages;
6566 :
6567 : /*
6568 : * If there are lossy pages then recompute the number of tuples
6569 : * processed by the bitmap heap node. We assume here that the chance
6570 : * of a given tuple coming from an exact page is the same as the
6571 : * chance that a given page is exact. This might not be true, but
6572 : * it's not clear how we can do any better.
6573 : */
6574 18 : if (lossy_pages > 0)
6575 : tuples_fetched =
6576 18 : clamp_row_est(indexSelectivity *
6577 18 : (exact_pages / heap_pages) * baserel->tuples +
6578 18 : (lossy_pages / heap_pages) * baserel->tuples);
6579 : }
6580 :
6581 600644 : if (cost_p)
6582 470930 : *cost_p = indexTotalCost;
6583 600644 : if (tuples_p)
6584 470930 : *tuples_p = tuples_fetched;
6585 :
6586 600644 : return pages_fetched;
6587 : }
6588 :
6589 : /*
6590 : * compute_gather_rows
6591 : * Estimate number of rows for gather (merge) nodes.
6592 : *
6593 : * In a parallel plan, each worker's row estimate is determined by dividing the
6594 : * total number of rows by parallel_divisor, which accounts for the leader's
6595 : * contribution in addition to the number of workers. Accordingly, when
6596 : * estimating the number of rows for gather (merge) nodes, we multiply the rows
6597 : * per worker by the same parallel_divisor to undo the division.
6598 : */
6599 : double
6600 26252 : compute_gather_rows(Path *path)
6601 : {
6602 : Assert(path->parallel_workers > 0);
6603 :
6604 26252 : return clamp_row_est(path->rows * get_parallel_divisor(path));
6605 : }
|
