Line data Source code
1 : /*-------------------------------------------------------------------------
2 : *
3 : * pgpa_planner.c
4 : * Use planner hooks to observe and modify planner behavior
5 : *
6 : * All interaction with the core planner happens here. Much of it has to
7 : * do with enforcing supplied advice, but we also need these hooks to
8 : * generate advice strings (though the heavy lifting in that case is
9 : * mostly done by pgpa_walker.c).
10 : *
11 : * Copyright (c) 2016-2026, PostgreSQL Global Development Group
12 : *
13 : * contrib/pg_plan_advice/pgpa_planner.c
14 : *
15 : *-------------------------------------------------------------------------
16 : */
17 : #include "postgres.h"
18 :
19 : #include "pg_plan_advice.h"
20 : #include "pgpa_identifier.h"
21 : #include "pgpa_output.h"
22 : #include "pgpa_planner.h"
23 : #include "pgpa_trove.h"
24 : #include "pgpa_walker.h"
25 :
26 : #include "commands/defrem.h"
27 : #include "common/hashfn_unstable.h"
28 : #include "nodes/makefuncs.h"
29 : #include "optimizer/extendplan.h"
30 : #include "optimizer/pathnode.h"
31 : #include "optimizer/paths.h"
32 : #include "optimizer/plancat.h"
33 : #include "optimizer/planner.h"
34 : #include "parser/parsetree.h"
35 : #include "utils/lsyscache.h"
36 :
37 : #ifdef USE_ASSERT_CHECKING
38 :
39 : /*
40 : * When assertions are enabled, we try generating relation identifiers during
41 : * planning, saving them in a hash table, and then cross-checking them against
42 : * the ones generated after planning is complete.
43 : */
44 : typedef struct pgpa_ri_checker_key
45 : {
46 : char *plan_name;
47 : Index rti;
48 : } pgpa_ri_checker_key;
49 :
50 : typedef struct pgpa_ri_checker
51 : {
52 : pgpa_ri_checker_key key;
53 : uint32 status;
54 : const char *rid_string;
55 : } pgpa_ri_checker;
56 :
57 : static uint32 pgpa_ri_checker_hash_key(pgpa_ri_checker_key key);
58 :
59 : static inline bool
60 : pgpa_ri_checker_compare_key(pgpa_ri_checker_key a, pgpa_ri_checker_key b)
61 : {
62 : if (a.rti != b.rti)
63 : return false;
64 : if (a.plan_name == NULL)
65 : return (b.plan_name == NULL);
66 : if (b.plan_name == NULL)
67 : return false;
68 : return strcmp(a.plan_name, b.plan_name) == 0;
69 : }
70 :
71 : #define SH_PREFIX pgpa_ri_check
72 : #define SH_ELEMENT_TYPE pgpa_ri_checker
73 : #define SH_KEY_TYPE pgpa_ri_checker_key
74 : #define SH_KEY key
75 : #define SH_HASH_KEY(tb, key) pgpa_ri_checker_hash_key(key)
76 : #define SH_EQUAL(tb, a, b) pgpa_ri_checker_compare_key(a, b)
77 : #define SH_SCOPE static inline
78 : #define SH_DECLARE
79 : #define SH_DEFINE
80 : #include "lib/simplehash.h"
81 :
82 : #endif
83 :
84 : typedef enum pgpa_jo_outcome
85 : {
86 : PGPA_JO_PERMITTED, /* permit this join order */
87 : PGPA_JO_DENIED, /* deny this join order */
88 : PGPA_JO_INDIFFERENT /* do neither */
89 : } pgpa_jo_outcome;
90 :
91 : typedef struct pgpa_planner_state
92 : {
93 : bool generate_advice_feedback;
94 : bool generate_advice_string;
95 : pgpa_trove *trove;
96 : List *sj_unique_rels;
97 :
98 : #ifdef USE_ASSERT_CHECKING
99 : pgpa_ri_check_hash *ri_check_hash;
100 : #endif
101 : } pgpa_planner_state;
102 :
103 : typedef struct pgpa_join_state
104 : {
105 : /* Most-recently-considered outer rel. */
106 : RelOptInfo *outerrel;
107 :
108 : /* Most-recently-considered inner rel. */
109 : RelOptInfo *innerrel;
110 :
111 : /*
112 : * Array of relation identifiers for all members of this joinrel, with
113 : * outerrel identifiers before innerrel identifiers.
114 : */
115 : pgpa_identifier *rids;
116 :
117 : /* Number of outer rel identifiers. */
118 : int outer_count;
119 :
120 : /* Number of inner rel identifiers. */
121 : int inner_count;
122 :
123 : /*
124 : * Trove lookup results.
125 : *
126 : * join_entries and rel_entries are arrays of entries, and join_indexes
127 : * and rel_indexes are the integer offsets within those arrays of entries
128 : * potentially relevant to us. The "join" fields correspond to a lookup
129 : * using PGPA_TROVE_LOOKUP_JOIN and the "rel" fields to a lookup using
130 : * PGPA_TROVE_LOOKUP_REL.
131 : */
132 : pgpa_trove_entry *join_entries;
133 : Bitmapset *join_indexes;
134 : pgpa_trove_entry *rel_entries;
135 : Bitmapset *rel_indexes;
136 : } pgpa_join_state;
137 :
138 : /* Saved hook values */
139 : static build_simple_rel_hook_type prev_build_simple_rel = NULL;
140 : static join_path_setup_hook_type prev_join_path_setup = NULL;
141 : static joinrel_setup_hook_type prev_joinrel_setup = NULL;
142 : static planner_setup_hook_type prev_planner_setup = NULL;
143 : static planner_shutdown_hook_type prev_planner_shutdown = NULL;
144 :
145 : /* Other global variables */
146 : int pgpa_planner_generate_advice = 0;
147 : static int planner_extension_id = -1;
148 :
149 : /* Function prototypes. */
150 : static void pgpa_planner_setup(PlannerGlobal *glob, Query *parse,
151 : const char *query_string,
152 : int cursorOptions,
153 : double *tuple_fraction,
154 : ExplainState *es);
155 : static void pgpa_planner_shutdown(PlannerGlobal *glob, Query *parse,
156 : const char *query_string, PlannedStmt *pstmt);
157 : static void pgpa_build_simple_rel(PlannerInfo *root,
158 : RelOptInfo *rel,
159 : RangeTblEntry *rte);
160 : static void pgpa_joinrel_setup(PlannerInfo *root,
161 : RelOptInfo *joinrel,
162 : RelOptInfo *outerrel,
163 : RelOptInfo *innerrel,
164 : SpecialJoinInfo *sjinfo,
165 : List *restrictlist);
166 : static void pgpa_join_path_setup(PlannerInfo *root,
167 : RelOptInfo *joinrel,
168 : RelOptInfo *outerrel,
169 : RelOptInfo *innerrel,
170 : JoinType jointype,
171 : JoinPathExtraData *extra);
172 : static pgpa_join_state *pgpa_get_join_state(PlannerInfo *root,
173 : RelOptInfo *joinrel,
174 : RelOptInfo *outerrel,
175 : RelOptInfo *innerrel);
176 : static void pgpa_planner_apply_joinrel_advice(uint64 *pgs_mask_p,
177 : char *plan_name,
178 : pgpa_join_state *pjs);
179 : static void pgpa_planner_apply_join_path_advice(JoinType jointype,
180 : uint64 *pgs_mask_p,
181 : char *plan_name,
182 : pgpa_join_state *pjs);
183 : static void pgpa_planner_apply_scan_advice(RelOptInfo *rel,
184 : pgpa_trove_entry *scan_entries,
185 : Bitmapset *scan_indexes,
186 : pgpa_trove_entry *rel_entries,
187 : Bitmapset *rel_indexes);
188 : static uint64 pgpa_join_strategy_mask_from_advice_tag(pgpa_advice_tag_type tag);
189 : static pgpa_jo_outcome pgpa_join_order_permits_join(int outer_count,
190 : int inner_count,
191 : pgpa_identifier *rids,
192 : pgpa_trove_entry *entry);
193 : static bool pgpa_join_method_permits_join(int outer_count, int inner_count,
194 : pgpa_identifier *rids,
195 : pgpa_trove_entry *entry,
196 : bool *restrict_method);
197 : static bool pgpa_opaque_join_permits_join(int outer_count, int inner_count,
198 : pgpa_identifier *rids,
199 : pgpa_trove_entry *entry,
200 : bool *restrict_method);
201 : static bool pgpa_semijoin_permits_join(int outer_count, int inner_count,
202 : pgpa_identifier *rids,
203 : pgpa_trove_entry *entry,
204 : bool outer_is_nullable,
205 : bool *restrict_method);
206 :
207 : static List *pgpa_planner_append_feedback(List *list, pgpa_trove *trove,
208 : pgpa_trove_lookup_type type,
209 : pgpa_identifier *rt_identifiers,
210 : pgpa_plan_walker_context *walker);
211 : static void pgpa_planner_feedback_warning(List *feedback);
212 :
213 : static inline void pgpa_ri_checker_save(pgpa_planner_state *pps,
214 : PlannerInfo *root,
215 : RelOptInfo *rel);
216 : static void pgpa_ri_checker_validate(pgpa_planner_state *pps,
217 : PlannedStmt *pstmt);
218 :
219 : static char *pgpa_bms_to_cstring(Bitmapset *bms);
220 : static const char *pgpa_jointype_to_cstring(JoinType jointype);
221 :
222 : /*
223 : * Install planner-related hooks.
224 : */
225 : void
226 11 : pgpa_planner_install_hooks(void)
227 : {
228 11 : planner_extension_id = GetPlannerExtensionId("pg_plan_advice");
229 11 : prev_planner_setup = planner_setup_hook;
230 11 : planner_setup_hook = pgpa_planner_setup;
231 11 : prev_planner_shutdown = planner_shutdown_hook;
232 11 : planner_shutdown_hook = pgpa_planner_shutdown;
233 11 : prev_build_simple_rel = build_simple_rel_hook;
234 11 : build_simple_rel_hook = pgpa_build_simple_rel;
235 11 : prev_joinrel_setup = joinrel_setup_hook;
236 11 : joinrel_setup_hook = pgpa_joinrel_setup;
237 11 : prev_join_path_setup = join_path_setup_hook;
238 11 : join_path_setup_hook = pgpa_join_path_setup;
239 11 : }
240 :
241 : /*
242 : * Carry out whatever setup work we need to do before planning.
243 : */
244 : static void
245 174 : pgpa_planner_setup(PlannerGlobal *glob, Query *parse, const char *query_string,
246 : int cursorOptions, double *tuple_fraction,
247 : ExplainState *es)
248 : {
249 174 : pgpa_trove *trove = NULL;
250 : pgpa_planner_state *pps;
251 : char *supplied_advice;
252 174 : bool generate_advice_feedback = false;
253 174 : bool generate_advice_string = false;
254 174 : bool needs_pps = false;
255 :
256 : /*
257 : * Decide whether we need to generate an advice string. We must do this if
258 : * the user has told us to do it categorically, or if another loadable
259 : * module has requested it, or if the user has requested it using the
260 : * EXPLAIN (PLAN_ADVICE) option.
261 : */
262 173 : generate_advice_string = (pg_plan_advice_always_store_advice_details ||
263 347 : pgpa_planner_generate_advice ||
264 173 : pg_plan_advice_should_explain(es));
265 174 : if (generate_advice_string)
266 120 : needs_pps = true;
267 :
268 : /*
269 : * If any advice was provided, build a trove of advice for use during
270 : * planning.
271 : */
272 174 : supplied_advice = pg_plan_advice_get_supplied_query_advice(glob, parse,
273 : query_string,
274 : cursorOptions,
275 : es);
276 174 : if (supplied_advice != NULL && supplied_advice[0] != '\0')
277 : {
278 : List *advice_items;
279 : char *error;
280 :
281 : /*
282 : * If the supplied advice string comes from pg_plan_advice.advice,
283 : * parsing shouldn't fail here, because we must have previously parsed
284 : * successfully in pg_plan_advice_advice_check_hook. However, it might
285 : * also come from a hook registered via pg_plan_advice_add_advisor,
286 : * and we can't be sure whether that's valid. (Plus, having an error
287 : * check here seems like a good idea anyway, just for safety.)
288 : */
289 115 : advice_items = pgpa_parse(supplied_advice, &error);
290 115 : if (error)
291 0 : ereport(WARNING,
292 : errmsg("could not parse supplied advice: %s", error));
293 :
294 : /*
295 : * It's possible that the advice string was non-empty but contained no
296 : * actual advice, e.g. it was all whitespace.
297 : */
298 115 : if (advice_items != NIL)
299 : {
300 113 : trove = pgpa_build_trove(advice_items);
301 113 : needs_pps = true;
302 :
303 : /*
304 : * If we know that we're running under EXPLAIN, or if the user has
305 : * told us to always do the work, generate advice feedback.
306 : */
307 113 : if (es != NULL || pg_plan_advice_feedback_warnings ||
308 : pg_plan_advice_always_store_advice_details)
309 113 : generate_advice_feedback = true;
310 : }
311 : }
312 :
313 : #ifdef USE_ASSERT_CHECKING
314 :
315 : /*
316 : * If asserts are enabled, always build a private state object for
317 : * cross-checks.
318 : */
319 : needs_pps = true;
320 : #endif
321 :
322 : /*
323 : * We only create and initialize a private state object if it's needed for
324 : * some purpose. That could be (1) recording that we will need to generate
325 : * an advice string, (2) storing a trove of supplied advice, or (3)
326 : * facilitating debugging cross-checks when asserts are enabled.
327 : */
328 174 : if (needs_pps)
329 : {
330 132 : pps = palloc0_object(pgpa_planner_state);
331 132 : pps->generate_advice_feedback = generate_advice_feedback;
332 132 : pps->generate_advice_string = generate_advice_string;
333 132 : pps->trove = trove;
334 : #ifdef USE_ASSERT_CHECKING
335 : pps->ri_check_hash =
336 : pgpa_ri_check_create(CurrentMemoryContext, 1024, NULL);
337 : #endif
338 132 : SetPlannerGlobalExtensionState(glob, planner_extension_id, pps);
339 : }
340 :
341 : /* Pass call to previous hook. */
342 174 : if (prev_planner_setup)
343 0 : (*prev_planner_setup) (glob, parse, query_string, cursorOptions,
344 : tuple_fraction, es);
345 174 : }
346 :
347 : /*
348 : * Carry out whatever work we want to do after planning is complete.
349 : */
350 : static void
351 174 : pgpa_planner_shutdown(PlannerGlobal *glob, Query *parse,
352 : const char *query_string, PlannedStmt *pstmt)
353 : {
354 : pgpa_planner_state *pps;
355 174 : pgpa_trove *trove = NULL;
356 174 : pgpa_plan_walker_context walker = {0}; /* placate compiler */
357 174 : bool generate_advice_feedback = false;
358 174 : bool generate_advice_string = false;
359 174 : List *pgpa_items = NIL;
360 174 : pgpa_identifier *rt_identifiers = NULL;
361 :
362 : /* Fetch our private state, set up by pgpa_planner_setup(). */
363 174 : pps = GetPlannerGlobalExtensionState(glob, planner_extension_id);
364 174 : if (pps != NULL)
365 : {
366 132 : trove = pps->trove;
367 132 : generate_advice_feedback = pps->generate_advice_feedback;
368 132 : generate_advice_string = pps->generate_advice_string;
369 : }
370 :
371 : /*
372 : * If we're trying to generate an advice string or if we're trying to
373 : * provide advice feedback, then we will need to create range table
374 : * identifiers.
375 : */
376 174 : if (generate_advice_string || generate_advice_feedback)
377 : {
378 132 : pgpa_plan_walker(&walker, pstmt, pps->sj_unique_rels);
379 132 : rt_identifiers = pgpa_create_identifiers_for_planned_stmt(pstmt);
380 : }
381 :
382 : /* Generate the advice string, if we need to do so. */
383 174 : if (generate_advice_string)
384 : {
385 : char *advice_string;
386 : StringInfoData buf;
387 :
388 : /* Generate a textual advice string. */
389 120 : initStringInfo(&buf);
390 120 : pgpa_output_advice(&buf, &walker, rt_identifiers);
391 120 : advice_string = buf.data;
392 :
393 : /* Save the advice string in the final plan. */
394 120 : pgpa_items = lappend(pgpa_items,
395 120 : makeDefElem("advice_string",
396 120 : (Node *) makeString(advice_string),
397 : -1));
398 : }
399 :
400 : /*
401 : * If we're trying to provide advice feedback, then we will need to
402 : * analyze how successful the advice was.
403 : */
404 174 : if (generate_advice_feedback)
405 : {
406 113 : List *feedback = NIL;
407 :
408 : /*
409 : * Inject a Node-tree representation of all the trove-entry flags into
410 : * the PlannedStmt.
411 : */
412 113 : feedback = pgpa_planner_append_feedback(feedback,
413 : trove,
414 : PGPA_TROVE_LOOKUP_SCAN,
415 : rt_identifiers, &walker);
416 113 : feedback = pgpa_planner_append_feedback(feedback,
417 : trove,
418 : PGPA_TROVE_LOOKUP_JOIN,
419 : rt_identifiers, &walker);
420 113 : feedback = pgpa_planner_append_feedback(feedback,
421 : trove,
422 : PGPA_TROVE_LOOKUP_REL,
423 : rt_identifiers, &walker);
424 :
425 113 : pgpa_items = lappend(pgpa_items, makeDefElem("feedback",
426 : (Node *) feedback, -1));
427 :
428 : /* If we were asked to generate feedback warnings, do so. */
429 113 : if (pg_plan_advice_feedback_warnings)
430 0 : pgpa_planner_feedback_warning(feedback);
431 : }
432 :
433 : /* Push whatever data we're saving into the PlannedStmt. */
434 174 : if (pgpa_items != NIL)
435 132 : pstmt->extension_state =
436 132 : lappend(pstmt->extension_state,
437 132 : makeDefElem("pg_plan_advice", (Node *) pgpa_items, -1));
438 :
439 : /*
440 : * If assertions are enabled, cross-check the generated range table
441 : * identifiers.
442 : */
443 174 : if (pps != NULL)
444 132 : pgpa_ri_checker_validate(pps, pstmt);
445 :
446 : /* Pass call to previous hook. */
447 174 : if (prev_planner_shutdown)
448 0 : (*prev_planner_shutdown) (glob, parse, query_string, pstmt);
449 174 : }
450 :
451 : /*
452 : * Hook function for build_simple_rel().
453 : *
454 : * We can apply scan advice at this point, and we also use this as an
455 : * opportunity to do range-table identifier cross-checking in assert-enabled
456 : * builds.
457 : */
458 : static void
459 382 : pgpa_build_simple_rel(PlannerInfo *root, RelOptInfo *rel, RangeTblEntry *rte)
460 : {
461 : pgpa_planner_state *pps;
462 :
463 : /* Fetch our private state, set up by pgpa_planner_setup(). */
464 382 : pps = GetPlannerGlobalExtensionState(root->glob, planner_extension_id);
465 :
466 : /* Save details needed for range table identifier cross-checking. */
467 382 : if (pps != NULL)
468 297 : pgpa_ri_checker_save(pps, root, rel);
469 :
470 : /* If query advice was provided, search for relevant entries. */
471 382 : if (pps != NULL && pps->trove != NULL)
472 : {
473 : pgpa_identifier rid;
474 : pgpa_trove_result tresult_scan;
475 : pgpa_trove_result tresult_rel;
476 :
477 : /* Search for scan advice and general rel advice. */
478 254 : pgpa_compute_identifier_by_rti(root, rel->relid, &rid);
479 254 : pgpa_trove_lookup(pps->trove, PGPA_TROVE_LOOKUP_SCAN, 1, &rid,
480 : &tresult_scan);
481 254 : pgpa_trove_lookup(pps->trove, PGPA_TROVE_LOOKUP_REL, 1, &rid,
482 : &tresult_rel);
483 :
484 : /* If relevant entries were found, apply them. */
485 254 : if (tresult_scan.indexes != NULL || tresult_rel.indexes != NULL)
486 : {
487 72 : uint64 original_mask = rel->pgs_mask;
488 :
489 72 : pgpa_planner_apply_scan_advice(rel,
490 : tresult_scan.entries,
491 : tresult_scan.indexes,
492 : tresult_rel.entries,
493 : tresult_rel.indexes);
494 :
495 : /* Emit debugging message, if enabled. */
496 72 : if (pg_plan_advice_trace_mask && original_mask != rel->pgs_mask)
497 : {
498 0 : if (root->plan_name != NULL)
499 0 : ereport(WARNING,
500 : (errmsg("strategy mask for RTI %u in subplan \"%s\" changed from 0x%" PRIx64 " to 0x%" PRIx64,
501 : rel->relid, root->plan_name,
502 : original_mask, rel->pgs_mask)));
503 : else
504 0 : ereport(WARNING,
505 : (errmsg("strategy mask for RTI %u changed from 0x%" PRIx64 " to 0x%" PRIx64,
506 : rel->relid, original_mask,
507 : rel->pgs_mask)));
508 : }
509 : }
510 : }
511 :
512 : /* Pass call to previous hook. */
513 382 : if (prev_build_simple_rel)
514 0 : (*prev_build_simple_rel) (root, rel, rte);
515 382 : }
516 :
517 : /*
518 : * Enforce any provided advice that is relevant to any method of implementing
519 : * this join.
520 : *
521 : * Although we're passed the outerrel and innerrel here, those are just
522 : * whatever values happened to prompt the creation of this joinrel; they
523 : * shouldn't really influence our choice of what advice to apply.
524 : */
525 : static void
526 206 : pgpa_joinrel_setup(PlannerInfo *root, RelOptInfo *joinrel,
527 : RelOptInfo *outerrel, RelOptInfo *innerrel,
528 : SpecialJoinInfo *sjinfo, List *restrictlist)
529 : {
530 : pgpa_join_state *pjs;
531 :
532 : Assert(bms_membership(joinrel->relids) == BMS_MULTIPLE);
533 :
534 : /* Get our private state information for this join. */
535 206 : pjs = pgpa_get_join_state(root, joinrel, outerrel, innerrel);
536 :
537 : /* If there is relevant advice, call a helper function to apply it. */
538 206 : if (pjs != NULL)
539 : {
540 105 : uint64 original_mask = joinrel->pgs_mask;
541 :
542 105 : pgpa_planner_apply_joinrel_advice(&joinrel->pgs_mask,
543 : root->plan_name,
544 : pjs);
545 :
546 : /* Emit debugging message, if enabled. */
547 105 : if (pg_plan_advice_trace_mask && original_mask != joinrel->pgs_mask)
548 : {
549 0 : if (root->plan_name != NULL)
550 0 : ereport(WARNING,
551 : (errmsg("strategy mask for join on RTIs %s in subplan \"%s\" changed from 0x%" PRIx64 " to 0x%" PRIx64,
552 : pgpa_bms_to_cstring(joinrel->relids),
553 : root->plan_name,
554 : original_mask,
555 : joinrel->pgs_mask)));
556 : else
557 0 : ereport(WARNING,
558 : (errmsg("strategy mask for join on RTIs %s changed from 0x%" PRIx64 " to 0x%" PRIx64,
559 : pgpa_bms_to_cstring(joinrel->relids),
560 : original_mask,
561 : joinrel->pgs_mask)));
562 : }
563 : }
564 :
565 : /* Pass call to previous hook. */
566 206 : if (prev_joinrel_setup)
567 0 : (*prev_joinrel_setup) (root, joinrel, outerrel, innerrel,
568 : sjinfo, restrictlist);
569 206 : }
570 :
571 : /*
572 : * Enforce any provided advice that is relevant to this particular method of
573 : * implementing this particular join.
574 : */
575 : static void
576 584 : pgpa_join_path_setup(PlannerInfo *root, RelOptInfo *joinrel,
577 : RelOptInfo *outerrel, RelOptInfo *innerrel,
578 : JoinType jointype, JoinPathExtraData *extra)
579 : {
580 : pgpa_join_state *pjs;
581 :
582 : Assert(bms_membership(joinrel->relids) == BMS_MULTIPLE);
583 :
584 : /*
585 : * If we're considering implementing a semijoin by making one side unique,
586 : * make a note of it in the pgpa_planner_state. See comments for
587 : * pgpa_sj_unique_rel for why we do this.
588 : */
589 584 : if (jointype == JOIN_UNIQUE_OUTER || jointype == JOIN_UNIQUE_INNER)
590 : {
591 : pgpa_planner_state *pps;
592 : RelOptInfo *uniquerel;
593 :
594 28 : uniquerel = jointype == JOIN_UNIQUE_OUTER ? outerrel : innerrel;
595 28 : pps = GetPlannerGlobalExtensionState(root->glob, planner_extension_id);
596 28 : if (pps != NULL &&
597 28 : (pps->generate_advice_string || pps->generate_advice_feedback))
598 : {
599 28 : bool found = false;
600 :
601 : /* Avoid adding duplicates. */
602 56 : foreach_ptr(pgpa_sj_unique_rel, ur, pps->sj_unique_rels)
603 : {
604 : /*
605 : * We should always use the same pointer for the same plan
606 : * name, so we need not use strcmp() here.
607 : */
608 28 : if (root->plan_name == ur->plan_name &&
609 14 : bms_equal(uniquerel->relids, ur->relids))
610 : {
611 14 : found = true;
612 14 : break;
613 : }
614 : }
615 :
616 : /* If not a duplicate, append to the list. */
617 28 : if (!found)
618 : {
619 14 : pgpa_sj_unique_rel *ur = palloc_object(pgpa_sj_unique_rel);
620 :
621 14 : ur->plan_name = root->plan_name;
622 14 : ur->relids = uniquerel->relids;
623 14 : pps->sj_unique_rels = lappend(pps->sj_unique_rels, ur);
624 : }
625 : }
626 : }
627 :
628 : /* Get our private state information for this join. */
629 584 : pjs = pgpa_get_join_state(root, joinrel, outerrel, innerrel);
630 :
631 : /* If there is relevant advice, call a helper function to apply it. */
632 584 : if (pjs != NULL)
633 : {
634 288 : uint64 original_mask = extra->pgs_mask;
635 :
636 288 : pgpa_planner_apply_join_path_advice(jointype,
637 : &extra->pgs_mask,
638 : root->plan_name,
639 : pjs);
640 :
641 : /* Emit debugging message, if enabled. */
642 288 : if (pg_plan_advice_trace_mask && original_mask != extra->pgs_mask)
643 : {
644 0 : if (root->plan_name != NULL)
645 0 : ereport(WARNING,
646 : (errmsg("strategy mask for %s join on %s with outer %s and inner %s in subplan \"%s\" changed from 0x%" PRIx64 " to 0x%" PRIx64,
647 : pgpa_jointype_to_cstring(jointype),
648 : pgpa_bms_to_cstring(joinrel->relids),
649 : pgpa_bms_to_cstring(outerrel->relids),
650 : pgpa_bms_to_cstring(innerrel->relids),
651 : root->plan_name,
652 : original_mask,
653 : extra->pgs_mask)));
654 : else
655 0 : ereport(WARNING,
656 : (errmsg("strategy mask for %s join on %s with outer %s and inner %s changed from 0x%" PRIx64 " to 0x%" PRIx64,
657 : pgpa_jointype_to_cstring(jointype),
658 : pgpa_bms_to_cstring(joinrel->relids),
659 : pgpa_bms_to_cstring(outerrel->relids),
660 : pgpa_bms_to_cstring(innerrel->relids),
661 : original_mask,
662 : extra->pgs_mask)));
663 : }
664 : }
665 :
666 : /* Pass call to previous hook. */
667 584 : if (prev_join_path_setup)
668 0 : (*prev_join_path_setup) (root, joinrel, outerrel, innerrel,
669 : jointype, extra);
670 584 : }
671 :
672 : /*
673 : * Search for advice pertaining to a proposed join.
674 : */
675 : static pgpa_join_state *
676 790 : pgpa_get_join_state(PlannerInfo *root, RelOptInfo *joinrel,
677 : RelOptInfo *outerrel, RelOptInfo *innerrel)
678 : {
679 : pgpa_planner_state *pps;
680 : pgpa_join_state *pjs;
681 790 : bool new_pjs = false;
682 :
683 : /* Fetch our private state, set up by pgpa_planner_setup(). */
684 790 : pps = GetPlannerGlobalExtensionState(root->glob, planner_extension_id);
685 790 : if (pps == NULL || pps->trove == NULL)
686 : {
687 : /* No advice applies to this query, hence none to this joinrel. */
688 118 : return NULL;
689 : }
690 :
691 : /*
692 : * See whether we've previously associated a pgpa_join_state with this
693 : * joinrel. If we have not, we need to try to construct one. If we have,
694 : * then there are two cases: (a) if innerrel and outerrel are unchanged,
695 : * we can simply use it, and (b) if they have changed, we need to rejigger
696 : * the array of identifiers but can still skip the trove lookup.
697 : */
698 672 : pjs = GetRelOptInfoExtensionState(joinrel, planner_extension_id);
699 672 : if (pjs != NULL)
700 : {
701 496 : if (pjs->join_indexes == NULL && pjs->rel_indexes == NULL)
702 : {
703 : /*
704 : * If there's no potentially relevant advice, then the presence of
705 : * this pgpa_join_state acts like a negative cache entry: it tells
706 : * us not to bother searching the trove for advice, because we
707 : * will not find any.
708 : */
709 208 : return NULL;
710 : }
711 :
712 288 : if (pjs->outerrel == outerrel && pjs->innerrel == innerrel)
713 : {
714 : /* No updates required, so just return. */
715 : /* XXX. Does this need to do something different under GEQO? */
716 105 : return pjs;
717 : }
718 : }
719 :
720 : /*
721 : * If there's no pgpa_join_state yet, we need to allocate one. Trove keys
722 : * will not get built for RTE_JOIN RTEs, so the array may end up being
723 : * larger than needed. It's not worth trying to compute a perfectly
724 : * accurate count here.
725 : */
726 359 : if (pjs == NULL)
727 : {
728 176 : int pessimistic_count = bms_num_members(joinrel->relids);
729 :
730 176 : pjs = palloc0_object(pgpa_join_state);
731 176 : pjs->rids = palloc_array(pgpa_identifier, pessimistic_count);
732 176 : new_pjs = true;
733 : }
734 :
735 : /*
736 : * Either we just allocated a new pgpa_join_state, or the existing one
737 : * needs reconfiguring for a new innerrel and outerrel. The required array
738 : * size can't change, so we can overwrite the existing one.
739 : */
740 359 : pjs->outerrel = outerrel;
741 359 : pjs->innerrel = innerrel;
742 359 : pjs->outer_count =
743 359 : pgpa_compute_identifiers_by_relids(root, outerrel->relids, pjs->rids);
744 359 : pjs->inner_count =
745 359 : pgpa_compute_identifiers_by_relids(root, innerrel->relids,
746 359 : pjs->rids + pjs->outer_count);
747 :
748 : /*
749 : * If we allocated a new pgpa_join_state, search our trove of advice for
750 : * relevant entries. The trove lookup will return the same results for
751 : * every outerrel/innerrel combination, so we don't need to repeat that
752 : * work every time.
753 : */
754 359 : if (new_pjs)
755 : {
756 : pgpa_trove_result tresult;
757 :
758 : /* Find join entries. */
759 176 : pgpa_trove_lookup(pps->trove, PGPA_TROVE_LOOKUP_JOIN,
760 176 : pjs->outer_count + pjs->inner_count,
761 : pjs->rids, &tresult);
762 176 : pjs->join_entries = tresult.entries;
763 176 : pjs->join_indexes = tresult.indexes;
764 :
765 : /* Find rel entries. */
766 176 : pgpa_trove_lookup(pps->trove, PGPA_TROVE_LOOKUP_REL,
767 176 : pjs->outer_count + pjs->inner_count,
768 : pjs->rids, &tresult);
769 176 : pjs->rel_entries = tresult.entries;
770 176 : pjs->rel_indexes = tresult.indexes;
771 :
772 : /* Now that the new pgpa_join_state is fully valid, save a pointer. */
773 176 : SetRelOptInfoExtensionState(joinrel, planner_extension_id, pjs);
774 :
775 : /*
776 : * If there was no relevant advice found, just return NULL. This
777 : * pgpa_join_state will stick around as a sort of negative cache
778 : * entry, so that future calls for this same joinrel quickly return
779 : * NULL.
780 : */
781 176 : if (pjs->join_indexes == NULL && pjs->rel_indexes == NULL)
782 71 : return NULL;
783 : }
784 :
785 288 : return pjs;
786 : }
787 :
788 : /*
789 : * Enforce overall restrictions on a join relation that apply uniformly
790 : * regardless of the choice of inner and outer rel.
791 : */
792 : static void
793 105 : pgpa_planner_apply_joinrel_advice(uint64 *pgs_mask_p, char *plan_name,
794 : pgpa_join_state *pjs)
795 : {
796 105 : int i = -1;
797 : int flags;
798 105 : bool gather_conflict = false;
799 105 : uint64 gather_mask = 0;
800 105 : Bitmapset *gather_partial_match = NULL;
801 105 : Bitmapset *gather_full_match = NULL;
802 105 : bool partitionwise_conflict = false;
803 105 : int partitionwise_outcome = 0;
804 105 : Bitmapset *partitionwise_partial_match = NULL;
805 105 : Bitmapset *partitionwise_full_match = NULL;
806 :
807 : /* Iterate over all possibly-relevant advice. */
808 151 : while ((i = bms_next_member(pjs->rel_indexes, i)) >= 0)
809 : {
810 46 : pgpa_trove_entry *entry = &pjs->rel_entries[i];
811 : pgpa_itm_type itm;
812 46 : bool full_match = false;
813 46 : uint64 my_gather_mask = 0;
814 46 : int my_partitionwise_outcome = 0; /* >0 yes, <0 no */
815 :
816 : /*
817 : * For GATHER and GATHER_MERGE, if the specified relations exactly
818 : * match this joinrel, do whatever the advice says; otherwise, don't
819 : * allow Gather or Gather Merge at this level. For NO_GATHER, there
820 : * must be a single target relation which must be included in this
821 : * joinrel, so just don't allow Gather or Gather Merge here, full
822 : * stop.
823 : */
824 46 : if (entry->tag == PGPA_TAG_NO_GATHER)
825 : {
826 7 : my_gather_mask = PGS_CONSIDER_NONPARTIAL;
827 7 : full_match = true;
828 : }
829 : else
830 : {
831 : int total_count;
832 :
833 39 : total_count = pjs->outer_count + pjs->inner_count;
834 39 : itm = pgpa_identifiers_match_target(total_count, pjs->rids,
835 : entry->target);
836 : Assert(itm != PGPA_ITM_DISJOINT);
837 :
838 39 : if (itm == PGPA_ITM_EQUAL)
839 : {
840 9 : full_match = true;
841 9 : if (entry->tag == PGPA_TAG_PARTITIONWISE)
842 4 : my_partitionwise_outcome = 1;
843 5 : else if (entry->tag == PGPA_TAG_GATHER)
844 3 : my_gather_mask = PGS_GATHER;
845 2 : else if (entry->tag == PGPA_TAG_GATHER_MERGE)
846 2 : my_gather_mask = PGS_GATHER_MERGE;
847 : else
848 0 : elog(ERROR, "unexpected advice tag: %d",
849 : (int) entry->tag);
850 : }
851 : else
852 : {
853 : /*
854 : * If specified relations don't exactly match this joinrel,
855 : * then we should do the opposite of whatever the advice says.
856 : * For instance, if we have PARTITIONWISE((a b c)) or
857 : * GATHER((a b c)) and this joinrel covers {a, b} or {a, b, c,
858 : * d} or {a, d}, we shouldn't plan it partitionwise or put a
859 : * Gather or Gather Merge on it here.
860 : *
861 : * Also, we can't put a Gather or Gather Merge at this level
862 : * if there is PARTITIONWISE advice that overlaps with it,
863 : * unless the PARTITIONWISE advice covers a subset of the
864 : * relations in the joinrel. To continue the previous example,
865 : * PARTITIONWISE((a b c)) is logically incompatible with
866 : * GATHER((a b)) or GATHER((a d)), but not with GATHER((a b c
867 : * d)).
868 : *
869 : * Conversely, we can't proceed partitionwise at this level if
870 : * there is overlapping GATHER or GATHER_MERGE advice, unless
871 : * that advice covers a superset of the relations in this
872 : * joinrel. This is just the flip side of the preceding point.
873 : */
874 30 : if (entry->tag == PGPA_TAG_PARTITIONWISE)
875 : {
876 21 : my_partitionwise_outcome = -1;
877 21 : if (itm != PGPA_ITM_TARGETS_ARE_SUBSET)
878 6 : my_gather_mask = PGS_CONSIDER_NONPARTIAL;
879 : }
880 9 : else if (entry->tag == PGPA_TAG_GATHER ||
881 4 : entry->tag == PGPA_TAG_GATHER_MERGE)
882 : {
883 9 : my_gather_mask = PGS_CONSIDER_NONPARTIAL;
884 9 : if (itm != PGPA_ITM_KEYS_ARE_SUBSET)
885 9 : my_partitionwise_outcome = -1;
886 : }
887 : else
888 0 : elog(ERROR, "unexpected advice tag: %d",
889 : (int) entry->tag);
890 : }
891 : }
892 :
893 : /*
894 : * If we set my_gather_mask up above, then we (1) make a note if the
895 : * advice conflicted, (2) remember the mask value, and (3) remember
896 : * whether this was a full or partial match.
897 : */
898 46 : if (my_gather_mask != 0)
899 : {
900 27 : if (gather_mask != 0 && gather_mask != my_gather_mask)
901 1 : gather_conflict = true;
902 27 : gather_mask = my_gather_mask;
903 27 : if (full_match)
904 12 : gather_full_match = bms_add_member(gather_full_match, i);
905 : else
906 15 : gather_partial_match = bms_add_member(gather_partial_match, i);
907 : }
908 :
909 : /*
910 : * Likewise, if we set my_partitionwise_outcome up above, then we (1)
911 : * make a note if the advice conflicted, (2) remember what the desired
912 : * outcome was, and (3) remember whether this was a full or partial
913 : * match.
914 : */
915 46 : if (my_partitionwise_outcome != 0)
916 : {
917 34 : if (partitionwise_outcome != 0 &&
918 : partitionwise_outcome != my_partitionwise_outcome)
919 2 : partitionwise_conflict = true;
920 34 : partitionwise_outcome = my_partitionwise_outcome;
921 34 : if (full_match)
922 : partitionwise_full_match =
923 4 : bms_add_member(partitionwise_full_match, i);
924 : else
925 : partitionwise_partial_match =
926 30 : bms_add_member(partitionwise_partial_match, i);
927 : }
928 : }
929 :
930 : /*
931 : * Mark every Gather-related piece of advice as partially matched, and if
932 : * the set of targets exactly matched this relation, fully matched. If
933 : * there was a conflict, mark them all as conflicting.
934 : */
935 105 : flags = PGPA_TE_MATCH_PARTIAL;
936 105 : if (gather_conflict)
937 1 : flags |= PGPA_TE_CONFLICTING;
938 105 : pgpa_trove_set_flags(pjs->rel_entries, gather_partial_match, flags);
939 105 : flags |= PGPA_TE_MATCH_FULL;
940 105 : pgpa_trove_set_flags(pjs->rel_entries, gather_full_match, flags);
941 :
942 : /* Likewise for partitionwise advice. */
943 105 : flags = PGPA_TE_MATCH_PARTIAL;
944 105 : if (partitionwise_conflict)
945 2 : flags |= PGPA_TE_CONFLICTING;
946 105 : pgpa_trove_set_flags(pjs->rel_entries, partitionwise_partial_match, flags);
947 105 : flags |= PGPA_TE_MATCH_FULL;
948 105 : pgpa_trove_set_flags(pjs->rel_entries, partitionwise_full_match, flags);
949 :
950 : /*
951 : * Enforce restrictions on the Gather/Gather Merge. Only clear bits here,
952 : * so that we still respect the enable_* GUCs. Do nothing if the advice
953 : * conflicts.
954 : */
955 105 : if (gather_mask != 0 && !gather_conflict)
956 : {
957 : uint64 all_gather_mask;
958 :
959 17 : all_gather_mask =
960 : PGS_GATHER | PGS_GATHER_MERGE | PGS_CONSIDER_NONPARTIAL;
961 17 : *pgs_mask_p &= ~(all_gather_mask & ~gather_mask);
962 : }
963 :
964 : /*
965 : * As above, but for partitionwise advice.
966 : *
967 : * To induce a partitionwise join, we disable all the ordinary means of
968 : * performing a join, so that an Append or MergeAppend path will hopefully
969 : * be chosen.
970 : *
971 : * To prevent one, we just disable Append and MergeAppend. Note that we
972 : * must not unset PGS_CONSIDER_PARTITIONWISE even when we don't want a
973 : * partitionwise join here, because we might want one at a higher level
974 : * that will construct its own paths using the ones from this level.
975 : */
976 105 : if (partitionwise_outcome != 0 && !partitionwise_conflict)
977 : {
978 18 : if (partitionwise_outcome > 0)
979 2 : *pgs_mask_p = (*pgs_mask_p & ~PGS_JOIN_ANY);
980 : else
981 16 : *pgs_mask_p &= ~(PGS_APPEND | PGS_MERGE_APPEND);
982 : }
983 105 : }
984 :
985 : /*
986 : * Enforce restrictions on the join order or join method.
987 : */
988 : static void
989 288 : pgpa_planner_apply_join_path_advice(JoinType jointype, uint64 *pgs_mask_p,
990 : char *plan_name,
991 : pgpa_join_state *pjs)
992 : {
993 288 : int i = -1;
994 288 : Bitmapset *jo_permit_indexes = NULL;
995 288 : Bitmapset *jo_deny_indexes = NULL;
996 288 : Bitmapset *jo_deny_rel_indexes = NULL;
997 288 : Bitmapset *jm_indexes = NULL;
998 288 : bool jm_conflict = false;
999 288 : uint32 join_mask = 0;
1000 288 : Bitmapset *sj_permit_indexes = NULL;
1001 288 : Bitmapset *sj_deny_indexes = NULL;
1002 :
1003 : /*
1004 : * Reconsider PARTITIONWISE(...) advice.
1005 : *
1006 : * We already thought about this for the joinrel as a whole, but in some
1007 : * cases, partitionwise advice can also constrain the join order. For
1008 : * instance, if the advice says PARTITIONWISE((t1 t2)), we shouldn't build
1009 : * join paths for any joinrel that includes t1 or t2 unless it also
1010 : * includes the other. In general, the partitionwise operation must have
1011 : * already been completed within one side of the current join or the
1012 : * other, else the join order is impermissible.
1013 : *
1014 : * NB: It might seem tempting to try to deal with PARTITIONWISE advice
1015 : * entirely in this function, but that doesn't work. Here, we can only
1016 : * affect the pgs_mask within a particular JoinPathExtraData, that is, for
1017 : * a particular choice of innerrel and outerrel. Partitionwise paths are
1018 : * not built that way, so we must set pgs_mask for the RelOptInfo, which
1019 : * is best done in pgpa_planner_apply_joinrel_advice.
1020 : */
1021 408 : while ((i = bms_next_member(pjs->rel_indexes, i)) >= 0)
1022 : {
1023 120 : pgpa_trove_entry *entry = &pjs->rel_entries[i];
1024 : pgpa_itm_type inner_itm;
1025 : pgpa_itm_type outer_itm;
1026 :
1027 120 : if (entry->tag != PGPA_TAG_PARTITIONWISE)
1028 42 : continue;
1029 :
1030 78 : outer_itm = pgpa_identifiers_match_target(pjs->outer_count,
1031 : pjs->rids, entry->target);
1032 78 : if (outer_itm == PGPA_ITM_EQUAL ||
1033 : outer_itm == PGPA_ITM_TARGETS_ARE_SUBSET)
1034 23 : continue;
1035 :
1036 55 : inner_itm = pgpa_identifiers_match_target(pjs->inner_count,
1037 55 : pjs->rids + pjs->outer_count,
1038 : entry->target);
1039 55 : if (inner_itm == PGPA_ITM_EQUAL ||
1040 : inner_itm == PGPA_ITM_TARGETS_ARE_SUBSET)
1041 23 : continue;
1042 :
1043 32 : jo_deny_rel_indexes = bms_add_member(jo_deny_rel_indexes, i);
1044 : }
1045 :
1046 : /* Iterate over advice that pertains to the join order and method. */
1047 288 : i = -1;
1048 532 : while ((i = bms_next_member(pjs->join_indexes, i)) >= 0)
1049 : {
1050 244 : pgpa_trove_entry *entry = &pjs->join_entries[i];
1051 : uint32 my_join_mask;
1052 :
1053 : /* Handle join order advice. */
1054 244 : if (entry->tag == PGPA_TAG_JOIN_ORDER)
1055 158 : {
1056 : pgpa_jo_outcome jo_outcome;
1057 :
1058 158 : jo_outcome = pgpa_join_order_permits_join(pjs->outer_count,
1059 : pjs->inner_count,
1060 : pjs->rids,
1061 : entry);
1062 158 : if (jo_outcome == PGPA_JO_PERMITTED)
1063 33 : jo_permit_indexes = bms_add_member(jo_permit_indexes, i);
1064 125 : else if (jo_outcome == PGPA_JO_DENIED)
1065 123 : jo_deny_indexes = bms_add_member(jo_deny_indexes, i);
1066 158 : continue;
1067 : }
1068 :
1069 : /* Handle join method advice. */
1070 86 : my_join_mask = pgpa_join_strategy_mask_from_advice_tag(entry->tag);
1071 86 : if (my_join_mask != 0)
1072 36 : {
1073 : bool permit;
1074 : bool restrict_method;
1075 :
1076 36 : if (entry->tag == PGPA_TAG_FOREIGN_JOIN)
1077 2 : permit = pgpa_opaque_join_permits_join(pjs->outer_count,
1078 : pjs->inner_count,
1079 : pjs->rids,
1080 : entry,
1081 : &restrict_method);
1082 : else
1083 34 : permit = pgpa_join_method_permits_join(pjs->outer_count,
1084 : pjs->inner_count,
1085 : pjs->rids,
1086 : entry,
1087 : &restrict_method);
1088 36 : if (!permit)
1089 17 : jo_deny_indexes = bms_add_member(jo_deny_indexes, i);
1090 19 : else if (restrict_method)
1091 : {
1092 19 : jm_indexes = bms_add_member(jm_indexes, i);
1093 19 : if (join_mask != 0 && join_mask != my_join_mask)
1094 1 : jm_conflict = true;
1095 19 : join_mask = my_join_mask;
1096 : }
1097 36 : continue;
1098 : }
1099 :
1100 : /* Handle semijoin uniqueness advice. */
1101 50 : if (entry->tag == PGPA_TAG_SEMIJOIN_UNIQUE ||
1102 24 : entry->tag == PGPA_TAG_SEMIJOIN_NON_UNIQUE)
1103 50 : {
1104 : bool outer_side_nullable;
1105 : bool restrict_method;
1106 :
1107 : /* Planner has nullable side of the semijoin on the outer side? */
1108 50 : outer_side_nullable = (jointype == JOIN_UNIQUE_OUTER ||
1109 : jointype == JOIN_RIGHT_SEMI);
1110 :
1111 50 : if (!pgpa_semijoin_permits_join(pjs->outer_count,
1112 : pjs->inner_count,
1113 : pjs->rids,
1114 : entry,
1115 : outer_side_nullable,
1116 : &restrict_method))
1117 9 : jo_deny_indexes = bms_add_member(jo_deny_indexes, i);
1118 41 : else if (restrict_method)
1119 : {
1120 : bool advice_unique;
1121 : bool jt_unique;
1122 : bool jt_non_unique;
1123 :
1124 : /* Advice wants to unique-ify and use a regular join? */
1125 41 : advice_unique = (entry->tag == PGPA_TAG_SEMIJOIN_UNIQUE);
1126 :
1127 : /* Planner is trying to unique-ify and use a regular join? */
1128 41 : jt_unique = (jointype == JOIN_UNIQUE_INNER ||
1129 : jointype == JOIN_UNIQUE_OUTER);
1130 :
1131 : /* Planner is trying a semi-join, without unique-ifying? */
1132 41 : jt_non_unique = (jointype == JOIN_SEMI ||
1133 : jointype == JOIN_RIGHT_SEMI);
1134 :
1135 41 : if (!jt_unique && !jt_non_unique)
1136 : {
1137 : /*
1138 : * This doesn't seem to be a semijoin to which SJ_UNIQUE
1139 : * or SJ_NON_UNIQUE can be applied.
1140 : */
1141 1 : entry->flags |= PGPA_TE_INAPPLICABLE;
1142 : }
1143 40 : else if (advice_unique != jt_unique)
1144 20 : sj_deny_indexes = bms_add_member(sj_deny_indexes, i);
1145 : else
1146 20 : sj_permit_indexes = bms_add_member(sj_permit_indexes, i);
1147 : }
1148 50 : continue;
1149 : }
1150 : }
1151 :
1152 : /*
1153 : * If the advice indicates both that this join order is permissible and
1154 : * also that it isn't, then mark advice related to the join order as
1155 : * conflicting.
1156 : */
1157 288 : if (jo_permit_indexes != NULL &&
1158 30 : (jo_deny_indexes != NULL || jo_deny_rel_indexes != NULL))
1159 : {
1160 3 : pgpa_trove_set_flags(pjs->join_entries, jo_permit_indexes,
1161 : PGPA_TE_CONFLICTING);
1162 3 : pgpa_trove_set_flags(pjs->join_entries, jo_deny_indexes,
1163 : PGPA_TE_CONFLICTING);
1164 3 : pgpa_trove_set_flags(pjs->rel_entries, jo_deny_rel_indexes,
1165 : PGPA_TE_CONFLICTING);
1166 : }
1167 :
1168 : /*
1169 : * If more than one join method specification is relevant here and they
1170 : * differ, mark them all as conflicting.
1171 : */
1172 288 : if (jm_conflict)
1173 1 : pgpa_trove_set_flags(pjs->join_entries, jm_indexes,
1174 : PGPA_TE_CONFLICTING);
1175 :
1176 : /* If semijoin advice says both yes and no, mark it all as conflicting. */
1177 288 : if (sj_permit_indexes != NULL && sj_deny_indexes != NULL)
1178 : {
1179 4 : pgpa_trove_set_flags(pjs->join_entries, sj_permit_indexes,
1180 : PGPA_TE_CONFLICTING);
1181 4 : pgpa_trove_set_flags(pjs->join_entries, sj_deny_indexes,
1182 : PGPA_TE_CONFLICTING);
1183 : }
1184 :
1185 : /*
1186 : * Enforce restrictions on the join order and join method, and any
1187 : * semijoin-related restrictions. Only clear bits here, so that we still
1188 : * respect the enable_* GUCs. Do nothing in cases where the advice on a
1189 : * single topic conflicts.
1190 : */
1191 288 : if ((jo_deny_indexes != NULL || jo_deny_rel_indexes != NULL) &&
1192 : jo_permit_indexes == NULL)
1193 169 : *pgs_mask_p &= ~PGS_JOIN_ANY;
1194 288 : if (join_mask != 0 && !jm_conflict)
1195 17 : *pgs_mask_p &= ~(PGS_JOIN_ANY & ~join_mask);
1196 288 : if (sj_deny_indexes != NULL && sj_permit_indexes == NULL)
1197 16 : *pgs_mask_p &= ~PGS_JOIN_ANY;
1198 288 : }
1199 :
1200 : /*
1201 : * Translate an advice tag into a path generation strategy mask.
1202 : *
1203 : * This function can be called with tag types that don't represent join
1204 : * strategies. In such cases, we just return 0, which can't be confused with
1205 : * a valid mask.
1206 : */
1207 : static uint64
1208 86 : pgpa_join_strategy_mask_from_advice_tag(pgpa_advice_tag_type tag)
1209 : {
1210 86 : switch (tag)
1211 : {
1212 2 : case PGPA_TAG_FOREIGN_JOIN:
1213 2 : return PGS_FOREIGNJOIN;
1214 4 : case PGPA_TAG_MERGE_JOIN_PLAIN:
1215 4 : return PGS_MERGEJOIN_PLAIN;
1216 4 : case PGPA_TAG_MERGE_JOIN_MATERIALIZE:
1217 4 : return PGS_MERGEJOIN_MATERIALIZE;
1218 10 : case PGPA_TAG_NESTED_LOOP_PLAIN:
1219 10 : return PGS_NESTLOOP_PLAIN;
1220 6 : case PGPA_TAG_NESTED_LOOP_MATERIALIZE:
1221 6 : return PGS_NESTLOOP_MATERIALIZE;
1222 4 : case PGPA_TAG_NESTED_LOOP_MEMOIZE:
1223 4 : return PGS_NESTLOOP_MEMOIZE;
1224 6 : case PGPA_TAG_HASH_JOIN:
1225 6 : return PGS_HASHJOIN;
1226 50 : default:
1227 50 : return 0;
1228 : }
1229 : }
1230 :
1231 : /*
1232 : * Does a certain item of join order advice permit a certain join?
1233 : *
1234 : * Returns PGPA_JO_DENIED if the advice is incompatible with the proposed
1235 : * join order.
1236 : *
1237 : * Returns PGPA_JO_PERMITTED if the advice specifies exactly the proposed
1238 : * join order. This implies that a partitionwise join should not be
1239 : * performed at this level; rather, one of the traditional join methods
1240 : * should be used.
1241 : *
1242 : * Returns PGPA_JO_INDIFFERENT if the advice does not care what happens.
1243 : * We use this for unordered JOIN_ORDER sublists, which are compatible with
1244 : * partitionwise join but do not mandate it.
1245 : */
1246 : static pgpa_jo_outcome
1247 158 : pgpa_join_order_permits_join(int outer_count, int inner_count,
1248 : pgpa_identifier *rids,
1249 : pgpa_trove_entry *entry)
1250 : {
1251 158 : bool loop = true;
1252 158 : bool sublist = false;
1253 : int length;
1254 : int outer_length;
1255 158 : pgpa_advice_target *target = entry->target;
1256 : pgpa_advice_target *prefix_target;
1257 :
1258 : /* We definitely have at least a partial match for this trove entry. */
1259 158 : entry->flags |= PGPA_TE_MATCH_PARTIAL;
1260 :
1261 : /*
1262 : * Find the innermost sublist that contains all keys; if no sublist does,
1263 : * then continue processing with the toplevel list.
1264 : *
1265 : * For example, if the advice says JOIN_ORDER(t1 t2 (t3 t4 t5)), then we
1266 : * should evaluate joins that only involve t3, t4, and/or t5 against the
1267 : * (t3 t4 t5) sublist, and others against the full list.
1268 : *
1269 : * Note that (1) outermost sublist is always ordered and (2) whenever we
1270 : * zoom into an unordered sublist, we instantly return
1271 : * PGPA_JO_INDIFFERENT.
1272 : */
1273 318 : while (loop)
1274 : {
1275 : Assert(target->ttype == PGPA_TARGET_ORDERED_LIST);
1276 :
1277 162 : loop = false;
1278 698 : foreach_ptr(pgpa_advice_target, child_target, target->children)
1279 : {
1280 : pgpa_itm_type itm;
1281 :
1282 382 : if (child_target->ttype == PGPA_TARGET_IDENTIFIER)
1283 346 : continue;
1284 :
1285 36 : itm = pgpa_identifiers_match_target(outer_count + inner_count,
1286 : rids, child_target);
1287 36 : if (itm == PGPA_ITM_EQUAL || itm == PGPA_ITM_KEYS_ARE_SUBSET)
1288 : {
1289 6 : if (child_target->ttype == PGPA_TARGET_ORDERED_LIST)
1290 : {
1291 4 : target = child_target;
1292 4 : sublist = true;
1293 4 : loop = true;
1294 4 : break;
1295 : }
1296 : else
1297 : {
1298 : Assert(child_target->ttype == PGPA_TARGET_UNORDERED_LIST);
1299 2 : return PGPA_JO_INDIFFERENT;
1300 : }
1301 : }
1302 : }
1303 : }
1304 :
1305 : /*
1306 : * Try to find a prefix of the selected join order list that is exactly
1307 : * equal to the outer side of the proposed join.
1308 : */
1309 156 : length = list_length(target->children);
1310 156 : prefix_target = palloc0_object(pgpa_advice_target);
1311 156 : prefix_target->ttype = PGPA_TARGET_ORDERED_LIST;
1312 178 : for (outer_length = 1; outer_length <= length; ++outer_length)
1313 : {
1314 : pgpa_itm_type itm;
1315 :
1316 : /* Avoid leaking memory in every loop iteration. */
1317 175 : if (prefix_target->children != NULL)
1318 19 : list_free(prefix_target->children);
1319 175 : prefix_target->children = list_copy_head(target->children,
1320 : outer_length);
1321 :
1322 : /* Search, hoping to find an exact match. */
1323 175 : itm = pgpa_identifiers_match_target(outer_count, rids, prefix_target);
1324 175 : if (itm == PGPA_ITM_EQUAL)
1325 47 : break;
1326 :
1327 : /*
1328 : * If the prefix of the join order list that we're considering
1329 : * includes some but not all of the outer rels, we can make the prefix
1330 : * longer to find an exact match. But if the advice hasn't mentioned
1331 : * everything that's part of our outer rel yet, but has mentioned
1332 : * things that are not, then this join doesn't match the join order
1333 : * list.
1334 : */
1335 128 : if (itm != PGPA_ITM_TARGETS_ARE_SUBSET)
1336 106 : return PGPA_JO_DENIED;
1337 : }
1338 :
1339 : /*
1340 : * If the previous loop stopped before the prefix_target included the
1341 : * entire join order list, then the next member of the join order list
1342 : * must exactly match the inner side of the join.
1343 : *
1344 : * Example: Given JOIN_ORDER(t1 t2 (t3 t4 t5)), if the outer side of the
1345 : * current join includes only t1, then the inner side must be exactly t2;
1346 : * if the outer side includes both t1 and t2, then the inner side must
1347 : * include exactly t3, t4, and t5.
1348 : */
1349 50 : if (outer_length < length)
1350 : {
1351 : pgpa_advice_target *inner_target;
1352 : pgpa_itm_type itm;
1353 :
1354 36 : inner_target = list_nth(target->children, outer_length);
1355 :
1356 36 : itm = pgpa_identifiers_match_target(inner_count, rids + outer_count,
1357 : inner_target);
1358 :
1359 : /*
1360 : * Before returning, consider whether we need to mark this entry as
1361 : * fully matched. If we're considering the full list rather than a
1362 : * sublist, and if we found every item but one on the outer side of
1363 : * the join and the last item on the inner side of the join, then the
1364 : * answer is yes.
1365 : */
1366 36 : if (!sublist && outer_length + 1 == length && itm == PGPA_ITM_EQUAL)
1367 10 : entry->flags |= PGPA_TE_MATCH_FULL;
1368 :
1369 36 : return (itm == PGPA_ITM_EQUAL) ? PGPA_JO_PERMITTED : PGPA_JO_DENIED;
1370 : }
1371 :
1372 : /*
1373 : * If we get here, then the outer side of the join includes the entirety
1374 : * of the join order list. In this case, we behave differently depending
1375 : * on whether we're looking at the top-level join order list or sublist.
1376 : * At the top-level, we treat the specified list as mandating that the
1377 : * actual join order has the given list as a prefix, but a sublist
1378 : * requires an exact match.
1379 : *
1380 : * Example: Given JOIN_ORDER(t1 t2 (t3 t4 t5)), we must start by joining
1381 : * all five of those relations and in that sequence, but once that is
1382 : * done, it's OK to join any other rels that are part of the join problem.
1383 : * This allows a user to specify the driving table and perhaps the first
1384 : * few things to which it should be joined while leaving the rest of the
1385 : * join order up the optimizer. But it seems like it would be surprising,
1386 : * given that specification, if the user could add t6 to the (t3 t4 t5)
1387 : * sub-join, so we don't allow that. If we did want to allow it, the logic
1388 : * earlier in this function would require substantial adjustment: we could
1389 : * allow the t3-t4-t5-t6 join to be built here, but the next step of
1390 : * joining t1-t2 to the result would still be rejected.
1391 : */
1392 14 : if (!sublist)
1393 14 : entry->flags |= PGPA_TE_MATCH_FULL;
1394 14 : return sublist ? PGPA_JO_DENIED : PGPA_JO_PERMITTED;
1395 : }
1396 :
1397 : /*
1398 : * Does a certain item of join method advice permit a certain join?
1399 : *
1400 : * Advice such as HASH_JOIN((x y)) means that there should be a hash join with
1401 : * exactly x and y on the inner side. Obviously, this means that if we are
1402 : * considering a join with exactly x and y on the inner side, we should enforce
1403 : * the use of a hash join. However, it also means that we must reject some
1404 : * incompatible join orders entirely. For example, a join with exactly x
1405 : * and y on the outer side shouldn't be allowed, because such paths might win
1406 : * over the advice-driven path on cost.
1407 : *
1408 : * To accommodate these requirements, this function returns true if the join
1409 : * should be allowed and false if it should not. Furthermore, *restrict_method
1410 : * is set to true if the join method should be enforced and false if not.
1411 : */
1412 : static bool
1413 34 : pgpa_join_method_permits_join(int outer_count, int inner_count,
1414 : pgpa_identifier *rids,
1415 : pgpa_trove_entry *entry,
1416 : bool *restrict_method)
1417 : {
1418 34 : pgpa_advice_target *target = entry->target;
1419 : pgpa_itm_type inner_itm;
1420 : pgpa_itm_type outer_itm;
1421 : pgpa_itm_type join_itm;
1422 :
1423 : /* We definitely have at least a partial match for this trove entry. */
1424 34 : entry->flags |= PGPA_TE_MATCH_PARTIAL;
1425 :
1426 34 : *restrict_method = false;
1427 :
1428 : /*
1429 : * If our inner rel mentions exactly the same relations as the advice
1430 : * target, allow the join and enforce the join method restriction.
1431 : *
1432 : * If our inner rel mentions a superset of the target relations, allow the
1433 : * join. The join we care about has already taken place, and this advice
1434 : * imposes no further restrictions.
1435 : */
1436 34 : inner_itm = pgpa_identifiers_match_target(inner_count,
1437 34 : rids + outer_count,
1438 : target);
1439 34 : if (inner_itm == PGPA_ITM_EQUAL)
1440 : {
1441 17 : entry->flags |= PGPA_TE_MATCH_FULL;
1442 17 : *restrict_method = true;
1443 17 : return true;
1444 : }
1445 17 : else if (inner_itm == PGPA_ITM_TARGETS_ARE_SUBSET)
1446 0 : return true;
1447 :
1448 : /*
1449 : * If our outer rel mentions a superset of the relations in the advice
1450 : * target, no restrictions apply, because the join we care about has
1451 : * already taken place.
1452 : *
1453 : * On the other hand, if our outer rel mentions exactly the relations
1454 : * mentioned in the advice target, the planner is trying to reverse the
1455 : * sides of the join as compared with our desired outcome. Reject that.
1456 : */
1457 17 : outer_itm = pgpa_identifiers_match_target(outer_count,
1458 : rids, target);
1459 17 : if (outer_itm == PGPA_ITM_TARGETS_ARE_SUBSET)
1460 0 : return true;
1461 17 : else if (outer_itm == PGPA_ITM_EQUAL)
1462 17 : return false;
1463 :
1464 : /*
1465 : * If the advice target mentions only a single relation, the test below
1466 : * cannot ever pass, so save some work by exiting now.
1467 : */
1468 0 : if (target->ttype == PGPA_TARGET_IDENTIFIER)
1469 0 : return false;
1470 :
1471 : /*
1472 : * If everything in the joinrel appears in the advice target, we're below
1473 : * the level of the join we want to control.
1474 : *
1475 : * For example, HASH_JOIN((x y)) doesn't restrict how x and y can be
1476 : * joined.
1477 : *
1478 : * This lookup shouldn't return PGPA_ITM_DISJOINT, because any such advice
1479 : * should not have been returned from the trove in the first place.
1480 : */
1481 0 : join_itm = pgpa_identifiers_match_target(outer_count + inner_count,
1482 : rids, target);
1483 : Assert(join_itm != PGPA_ITM_DISJOINT);
1484 0 : if (join_itm == PGPA_ITM_KEYS_ARE_SUBSET ||
1485 : join_itm == PGPA_ITM_EQUAL)
1486 0 : return true;
1487 :
1488 : /*
1489 : * We've already permitted all allowable cases, so reject this.
1490 : *
1491 : * If we reach this point, then the advice overlaps with this join but
1492 : * isn't entirely contained within either side, and there's also at least
1493 : * one relation present in the join that isn't mentioned by the advice.
1494 : *
1495 : * For instance, in the HASH_JOIN((x y)) example, we would reach here if x
1496 : * were on one side of the join, y on the other, and at least one of the
1497 : * two sides also included some other relation, say t. In that case,
1498 : * accepting this join would allow the (x y t) joinrel to contain
1499 : * non-disabled paths that do not put (x y) on the inner side of a hash
1500 : * join; we could instead end up with something like (x JOIN t) JOIN y.
1501 : */
1502 0 : return false;
1503 : }
1504 :
1505 : /*
1506 : * Does advice concerning an opaque join permit a certain join?
1507 : *
1508 : * By an opaque join, we mean one where the exact mechanism by which the
1509 : * join is performed is not visible to PostgreSQL. Currently this is the
1510 : * case only for foreign joins: FOREIGN_JOIN((x y z)) means that x, y, and
1511 : * z are joined on the remote side, but we know nothing about the join order
1512 : * or join methods used over there.
1513 : *
1514 : * The logic here needs to differ from pgpa_join_method_permits_join because,
1515 : * for other join types, the advice target is the set of inner rels; here, it
1516 : * includes both inner and outer rels.
1517 : */
1518 : static bool
1519 2 : pgpa_opaque_join_permits_join(int outer_count, int inner_count,
1520 : pgpa_identifier *rids,
1521 : pgpa_trove_entry *entry,
1522 : bool *restrict_method)
1523 : {
1524 2 : pgpa_advice_target *target = entry->target;
1525 : pgpa_itm_type join_itm;
1526 :
1527 : /* We definitely have at least a partial match for this trove entry. */
1528 2 : entry->flags |= PGPA_TE_MATCH_PARTIAL;
1529 :
1530 2 : *restrict_method = false;
1531 :
1532 2 : join_itm = pgpa_identifiers_match_target(outer_count + inner_count,
1533 : rids, target);
1534 2 : if (join_itm == PGPA_ITM_EQUAL)
1535 : {
1536 : /*
1537 : * We have an exact match, and should therefore allow the join and
1538 : * enforce the use of the relevant opaque join method.
1539 : */
1540 2 : entry->flags |= PGPA_TE_MATCH_FULL;
1541 2 : *restrict_method = true;
1542 2 : return true;
1543 : }
1544 :
1545 0 : if (join_itm == PGPA_ITM_KEYS_ARE_SUBSET ||
1546 : join_itm == PGPA_ITM_TARGETS_ARE_SUBSET)
1547 : {
1548 : /*
1549 : * If join_itm == PGPA_ITM_TARGETS_ARE_SUBSET, then the join we care
1550 : * about has already taken place and no further restrictions apply.
1551 : *
1552 : * If join_itm == PGPA_ITM_KEYS_ARE_SUBSET, we're still building up to
1553 : * the join we care about and have not introduced any extraneous
1554 : * relations not named in the advice. Note that ForeignScan paths for
1555 : * joins are built up from ForeignScan paths from underlying joins and
1556 : * scans, so we must not disable this join when considering a subset
1557 : * of the relations we ultimately want.
1558 : */
1559 0 : return true;
1560 : }
1561 :
1562 : /*
1563 : * The advice overlaps the join, but at least one relation is present in
1564 : * the join that isn't mentioned by the advice. We want to disable such
1565 : * paths so that we actually push down the join as intended.
1566 : */
1567 0 : return false;
1568 : }
1569 :
1570 : /*
1571 : * Does advice concerning a semijoin permit a certain join?
1572 : *
1573 : * Unlike join method advice, which lists the rels on the inner side of the
1574 : * join, semijoin uniqueness advice lists the rels on the nullable side of the
1575 : * join. Those can be the same, if the join type is JOIN_UNIQUE_INNER or
1576 : * JOIN_SEMI, or they can be different, in case of JOIN_UNIQUE_OUTER or
1577 : * JOIN_RIGHT_SEMI.
1578 : *
1579 : * We don't know here whether the caller specified SEMIJOIN_UNIQUE or
1580 : * SEMIJOIN_NON_UNIQUE. The caller should check the join type against the
1581 : * advice type if and only if we set *restrict_method to true.
1582 : */
1583 : static bool
1584 50 : pgpa_semijoin_permits_join(int outer_count, int inner_count,
1585 : pgpa_identifier *rids,
1586 : pgpa_trove_entry *entry,
1587 : bool outer_is_nullable,
1588 : bool *restrict_method)
1589 : {
1590 50 : pgpa_advice_target *target = entry->target;
1591 : pgpa_itm_type join_itm;
1592 : pgpa_itm_type inner_itm;
1593 : pgpa_itm_type outer_itm;
1594 :
1595 50 : *restrict_method = false;
1596 :
1597 : /* We definitely have at least a partial match for this trove entry. */
1598 50 : entry->flags |= PGPA_TE_MATCH_PARTIAL;
1599 :
1600 : /*
1601 : * If outer rel is the nullable side and contains exactly the same
1602 : * relations as the advice target, then the join order is allowable, but
1603 : * the caller must check whether the advice tag (either SEMIJOIN_UNIQUE or
1604 : * SEMIJOIN_NON_UNIQUE) matches the join type.
1605 : *
1606 : * If the outer rel is a superset of the target relations, the join we
1607 : * care about has already taken place, so we should impose no further
1608 : * restrictions.
1609 : */
1610 50 : outer_itm = pgpa_identifiers_match_target(outer_count,
1611 : rids, target);
1612 50 : if (outer_itm == PGPA_ITM_EQUAL)
1613 : {
1614 25 : entry->flags |= PGPA_TE_MATCH_FULL;
1615 25 : if (outer_is_nullable)
1616 : {
1617 20 : *restrict_method = true;
1618 20 : return true;
1619 : }
1620 : }
1621 25 : else if (outer_itm == PGPA_ITM_TARGETS_ARE_SUBSET)
1622 0 : return true;
1623 :
1624 : /* As above, but for the inner rel. */
1625 30 : inner_itm = pgpa_identifiers_match_target(inner_count,
1626 30 : rids + outer_count,
1627 : target);
1628 30 : if (inner_itm == PGPA_ITM_EQUAL)
1629 : {
1630 25 : entry->flags |= PGPA_TE_MATCH_FULL;
1631 25 : if (!outer_is_nullable)
1632 : {
1633 21 : *restrict_method = true;
1634 21 : return true;
1635 : }
1636 : }
1637 5 : else if (inner_itm == PGPA_ITM_TARGETS_ARE_SUBSET)
1638 0 : return true;
1639 :
1640 : /*
1641 : * If everything in the joinrel appears in the advice target, we're below
1642 : * the level of the join we want to control.
1643 : */
1644 9 : join_itm = pgpa_identifiers_match_target(outer_count + inner_count,
1645 : rids, target);
1646 : Assert(join_itm != PGPA_ITM_DISJOINT);
1647 9 : if (join_itm == PGPA_ITM_KEYS_ARE_SUBSET ||
1648 : join_itm == PGPA_ITM_EQUAL)
1649 0 : return true;
1650 :
1651 : /*
1652 : * We've tested for all allowable possibilities, and so must reject this
1653 : * join order. This can happen in two ways.
1654 : *
1655 : * First, we might be considering a semijoin that overlaps incompletely
1656 : * with one or both sides of the join. For example, if the user has
1657 : * specified SEMIJOIN_UNIQUE((t1 t2)) or SEMIJOIN_NON_UNIQUE((t1 t2)), we
1658 : * should reject a proposed t2-t3 join, since that could not result in a
1659 : * final plan compatible with the advice.
1660 : *
1661 : * Second, we might be considering a semijoin where the advice target
1662 : * perfectly matches one side of the join, but it's the wrong one. For
1663 : * example, in the example above, we might see a 3-way join between t1,
1664 : * t2, and t3, with (t1 t2) on the non-nullable side. That, too, would be
1665 : * incompatible with the advice.
1666 : */
1667 9 : return false;
1668 : }
1669 :
1670 : /*
1671 : * Apply scan advice to a RelOptInfo.
1672 : */
1673 : static void
1674 72 : pgpa_planner_apply_scan_advice(RelOptInfo *rel,
1675 : pgpa_trove_entry *scan_entries,
1676 : Bitmapset *scan_indexes,
1677 : pgpa_trove_entry *rel_entries,
1678 : Bitmapset *rel_indexes)
1679 : {
1680 72 : bool gather_conflict = false;
1681 72 : Bitmapset *gather_partial_match = NULL;
1682 72 : Bitmapset *gather_full_match = NULL;
1683 72 : int i = -1;
1684 72 : pgpa_trove_entry *scan_entry = NULL;
1685 : int flags;
1686 72 : bool scan_type_conflict = false;
1687 72 : Bitmapset *scan_type_indexes = NULL;
1688 72 : Bitmapset *scan_type_rel_indexes = NULL;
1689 72 : uint64 gather_mask = 0;
1690 72 : uint64 scan_type = 0;
1691 :
1692 : /* Scrutinize available scan advice. */
1693 110 : while ((i = bms_next_member(scan_indexes, i)) >= 0)
1694 : {
1695 38 : pgpa_trove_entry *my_entry = &scan_entries[i];
1696 38 : uint64 my_scan_type = 0;
1697 :
1698 : /* Translate our advice tags to a scan strategy advice value. */
1699 38 : if (my_entry->tag == PGPA_TAG_BITMAP_HEAP_SCAN)
1700 : {
1701 : /*
1702 : * Currently, PGS_CONSIDER_INDEXONLY can suppress Bitmap Heap
1703 : * Scans, so don't clear it when such a scan is requested. This
1704 : * happens because build_index_scan() thinks that the possibility
1705 : * of an index-only scan is a sufficient reason to consider using
1706 : * an otherwise-useless index, and get_index_paths() thinks that
1707 : * the same paths that are useful for index or index-only scans
1708 : * should also be considered for bitmap scans. Perhaps that logic
1709 : * should be tightened up, but until then we need to include
1710 : * PGS_CONSIDER_INDEXONLY in my_scan_type here.
1711 : */
1712 3 : my_scan_type = PGS_BITMAPSCAN | PGS_CONSIDER_INDEXONLY;
1713 : }
1714 35 : else if (my_entry->tag == PGPA_TAG_INDEX_ONLY_SCAN)
1715 5 : my_scan_type = PGS_INDEXONLYSCAN | PGS_CONSIDER_INDEXONLY;
1716 30 : else if (my_entry->tag == PGPA_TAG_INDEX_SCAN)
1717 14 : my_scan_type = PGS_INDEXSCAN;
1718 16 : else if (my_entry->tag == PGPA_TAG_SEQ_SCAN)
1719 12 : my_scan_type = PGS_SEQSCAN;
1720 4 : else if (my_entry->tag == PGPA_TAG_TID_SCAN)
1721 4 : my_scan_type = PGS_TIDSCAN;
1722 :
1723 : /*
1724 : * If this is understandable scan advice, hang on to the entry, the
1725 : * inferred scan type, and the index at which we found it.
1726 : *
1727 : * Also make a note if we see conflicting scan type advice. Note that
1728 : * we regard two index specifications as conflicting unless they match
1729 : * exactly. In theory, perhaps we could regard INDEX_SCAN(a c) and
1730 : * INDEX_SCAN(a b.c) as non-conflicting if it happens that the only
1731 : * index named c is in schema b, but it doesn't seem worth the code.
1732 : */
1733 38 : if (my_scan_type != 0)
1734 : {
1735 38 : if (scan_type != 0 && scan_type != my_scan_type)
1736 0 : scan_type_conflict = true;
1737 38 : if (!scan_type_conflict && scan_entry != NULL &&
1738 2 : my_entry->target->itarget != NULL &&
1739 2 : scan_entry->target->itarget != NULL &&
1740 2 : !pgpa_index_targets_equal(scan_entry->target->itarget,
1741 2 : my_entry->target->itarget))
1742 1 : scan_type_conflict = true;
1743 38 : scan_entry = my_entry;
1744 38 : scan_type = my_scan_type;
1745 38 : scan_type_indexes = bms_add_member(scan_type_indexes, i);
1746 : }
1747 : }
1748 :
1749 : /* Scrutinize available gather-related and partitionwise advice. */
1750 72 : i = -1;
1751 110 : while ((i = bms_next_member(rel_indexes, i)) >= 0)
1752 : {
1753 38 : pgpa_trove_entry *my_entry = &rel_entries[i];
1754 38 : uint64 my_gather_mask = 0;
1755 : bool just_one_rel;
1756 :
1757 76 : just_one_rel = my_entry->target->ttype == PGPA_TARGET_IDENTIFIER
1758 38 : || list_length(my_entry->target->children) == 1;
1759 :
1760 : /*
1761 : * PARTITIONWISE behaves like a scan type, except that if there's more
1762 : * than one relation targeted, it has no effect at this level.
1763 : */
1764 38 : if (my_entry->tag == PGPA_TAG_PARTITIONWISE)
1765 : {
1766 12 : if (just_one_rel)
1767 : {
1768 4 : const uint64 my_scan_type = PGS_APPEND | PGS_MERGE_APPEND;
1769 :
1770 4 : if (scan_type != 0 && scan_type != my_scan_type)
1771 0 : scan_type_conflict = true;
1772 4 : scan_entry = my_entry;
1773 4 : scan_type = my_scan_type;
1774 : scan_type_rel_indexes =
1775 4 : bms_add_member(scan_type_rel_indexes, i);
1776 : }
1777 12 : continue;
1778 : }
1779 :
1780 : /*
1781 : * GATHER and GATHER_MERGE applied to a single rel mean that we should
1782 : * use the corresponding strategy here, while applying either to more
1783 : * than one rel means we should not use those strategies here, but
1784 : * rather at the level of the joinrel that corresponds to what was
1785 : * specified. NO_GATHER can only be applied to single rels.
1786 : *
1787 : * Note that setting PGS_CONSIDER_NONPARTIAL in my_gather_mask is
1788 : * equivalent to allowing the non-use of either form of Gather here.
1789 : */
1790 26 : if (my_entry->tag == PGPA_TAG_GATHER ||
1791 15 : my_entry->tag == PGPA_TAG_GATHER_MERGE)
1792 : {
1793 19 : if (!just_one_rel)
1794 11 : my_gather_mask = PGS_CONSIDER_NONPARTIAL;
1795 8 : else if (my_entry->tag == PGPA_TAG_GATHER)
1796 4 : my_gather_mask = PGS_GATHER;
1797 : else
1798 4 : my_gather_mask = PGS_GATHER_MERGE;
1799 : }
1800 7 : else if (my_entry->tag == PGPA_TAG_NO_GATHER)
1801 : {
1802 : Assert(just_one_rel);
1803 7 : my_gather_mask = PGS_CONSIDER_NONPARTIAL;
1804 : }
1805 :
1806 : /*
1807 : * If we set my_gather_mask up above, then we (1) make a note if the
1808 : * advice conflicted, (2) remember the mask value, and (3) remember
1809 : * whether this was a full or partial match.
1810 : */
1811 26 : if (my_gather_mask != 0)
1812 : {
1813 26 : if (gather_mask != 0 && gather_mask != my_gather_mask)
1814 0 : gather_conflict = true;
1815 26 : gather_mask = my_gather_mask;
1816 26 : if (just_one_rel)
1817 15 : gather_full_match = bms_add_member(gather_full_match, i);
1818 : else
1819 11 : gather_partial_match = bms_add_member(gather_partial_match, i);
1820 : }
1821 : }
1822 :
1823 : /* Enforce choice of index. */
1824 72 : if (scan_entry != NULL && !scan_type_conflict &&
1825 39 : (scan_entry->tag == PGPA_TAG_INDEX_SCAN ||
1826 28 : scan_entry->tag == PGPA_TAG_INDEX_ONLY_SCAN))
1827 : {
1828 16 : pgpa_index_target *itarget = scan_entry->target->itarget;
1829 16 : IndexOptInfo *matched_index = NULL;
1830 :
1831 49 : foreach_node(IndexOptInfo, index, rel->indexlist)
1832 : {
1833 30 : char *relname = get_rel_name(index->indexoid);
1834 30 : Oid nspoid = get_rel_namespace(index->indexoid);
1835 30 : char *relnamespace = get_namespace_name_or_temp(nspoid);
1836 :
1837 30 : if (strcmp(itarget->indname, relname) == 0 &&
1838 14 : (itarget->indnamespace == NULL ||
1839 3 : strcmp(itarget->indnamespace, relnamespace) == 0))
1840 : {
1841 13 : matched_index = index;
1842 13 : break;
1843 : }
1844 : }
1845 :
1846 16 : if (matched_index == NULL)
1847 : {
1848 : /* Don't force the scan type if the index doesn't exist. */
1849 3 : scan_type = 0;
1850 :
1851 : /* Mark advice as inapplicable. */
1852 3 : pgpa_trove_set_flags(scan_entries, scan_type_indexes,
1853 : PGPA_TE_INAPPLICABLE);
1854 : }
1855 : else
1856 : {
1857 : /* Disable every other index. */
1858 52 : foreach_node(IndexOptInfo, index, rel->indexlist)
1859 : {
1860 26 : if (index != matched_index)
1861 13 : index->disabled = true;
1862 : }
1863 : }
1864 : }
1865 :
1866 : /*
1867 : * Mark all the scan method entries as fully matched; and if they specify
1868 : * different things, mark them all as conflicting.
1869 : */
1870 72 : flags = PGPA_TE_MATCH_PARTIAL | PGPA_TE_MATCH_FULL;
1871 72 : if (scan_type_conflict)
1872 1 : flags |= PGPA_TE_CONFLICTING;
1873 72 : pgpa_trove_set_flags(scan_entries, scan_type_indexes, flags);
1874 72 : pgpa_trove_set_flags(rel_entries, scan_type_rel_indexes, flags);
1875 :
1876 : /*
1877 : * Mark every Gather-related piece of advice as partially matched. Mark
1878 : * the ones that included this relation as a target by itself as fully
1879 : * matched. If there was a conflict, mark them all as conflicting.
1880 : */
1881 72 : flags = PGPA_TE_MATCH_PARTIAL;
1882 72 : if (gather_conflict)
1883 0 : flags |= PGPA_TE_CONFLICTING;
1884 72 : pgpa_trove_set_flags(rel_entries, gather_partial_match, flags);
1885 72 : flags |= PGPA_TE_MATCH_FULL;
1886 72 : pgpa_trove_set_flags(rel_entries, gather_full_match, flags);
1887 :
1888 : /*
1889 : * Enforce restrictions on the scan type and use of Gather/Gather Merge.
1890 : * Only clear bits here, so that we still respect the enable_* GUCs. Do
1891 : * nothing in cases where the advice on a single topic conflicts.
1892 : */
1893 72 : if (scan_type != 0 && !scan_type_conflict)
1894 : {
1895 : uint64 all_scan_mask;
1896 :
1897 36 : all_scan_mask = PGS_SCAN_ANY | PGS_APPEND | PGS_MERGE_APPEND |
1898 : PGS_CONSIDER_INDEXONLY;
1899 36 : rel->pgs_mask &= ~(all_scan_mask & ~scan_type);
1900 : }
1901 72 : if (gather_mask != 0 && !gather_conflict)
1902 : {
1903 : uint64 all_gather_mask;
1904 :
1905 25 : all_gather_mask =
1906 : PGS_GATHER | PGS_GATHER_MERGE | PGS_CONSIDER_NONPARTIAL;
1907 25 : rel->pgs_mask &= ~(all_gather_mask & ~gather_mask);
1908 : }
1909 72 : }
1910 :
1911 : /*
1912 : * Add feedback entries for one trove slice to the provided list and
1913 : * return the resulting list.
1914 : *
1915 : * Feedback entries are generated from the trove entry's flags. It's assumed
1916 : * that the caller has already set all relevant flags with the exception of
1917 : * PGPA_TE_FAILED. We set that flag here if appropriate.
1918 : */
1919 : static List *
1920 339 : pgpa_planner_append_feedback(List *list, pgpa_trove *trove,
1921 : pgpa_trove_lookup_type type,
1922 : pgpa_identifier *rt_identifiers,
1923 : pgpa_plan_walker_context *walker)
1924 : {
1925 : pgpa_trove_entry *entries;
1926 : int nentries;
1927 :
1928 339 : pgpa_trove_lookup_all(trove, type, &entries, &nentries);
1929 472 : for (int i = 0; i < nentries; ++i)
1930 : {
1931 133 : pgpa_trove_entry *entry = &entries[i];
1932 : DefElem *item;
1933 :
1934 : /*
1935 : * If this entry was fully matched, check whether generating advice
1936 : * from this plan would produce such an entry. If not, label the entry
1937 : * as failed.
1938 : */
1939 133 : if ((entry->flags & PGPA_TE_MATCH_FULL) != 0 &&
1940 113 : !pgpa_walker_would_advise(walker, rt_identifiers,
1941 : entry->tag, entry->target))
1942 30 : entry->flags |= PGPA_TE_FAILED;
1943 :
1944 133 : item = makeDefElem(pgpa_cstring_trove_entry(entry),
1945 133 : (Node *) makeInteger(entry->flags), -1);
1946 133 : list = lappend(list, item);
1947 : }
1948 :
1949 339 : return list;
1950 : }
1951 :
1952 : /*
1953 : * Emit a WARNING to tell the user about a problem with the supplied plan
1954 : * advice.
1955 : */
1956 : static void
1957 0 : pgpa_planner_feedback_warning(List *feedback)
1958 : {
1959 : StringInfoData detailbuf;
1960 : StringInfoData flagbuf;
1961 :
1962 : /* Quick exit if there's no feedback. */
1963 0 : if (feedback == NIL)
1964 0 : return;
1965 :
1966 : /* Initialize buffers. */
1967 0 : initStringInfo(&detailbuf);
1968 0 : initStringInfo(&flagbuf);
1969 :
1970 : /* Main loop. */
1971 0 : foreach_node(DefElem, item, feedback)
1972 : {
1973 0 : int flags = defGetInt32(item);
1974 :
1975 : /*
1976 : * Don't emit anything if it was fully matched with no problems found.
1977 : *
1978 : * NB: Feedback should never be marked fully matched without also
1979 : * being marked partially matched.
1980 : */
1981 0 : if (flags == (PGPA_TE_MATCH_PARTIAL | PGPA_TE_MATCH_FULL))
1982 0 : continue;
1983 :
1984 : /*
1985 : * Terminate each detail line except the last with a newline. This is
1986 : * also a convenient place to reset flagbuf.
1987 : */
1988 0 : if (detailbuf.len > 0)
1989 : {
1990 0 : appendStringInfoChar(&detailbuf, '\n');
1991 0 : resetStringInfo(&flagbuf);
1992 : }
1993 :
1994 : /* Generate output. */
1995 0 : pgpa_trove_append_flags(&flagbuf, flags);
1996 0 : appendStringInfo(&detailbuf, "advice %s feedback is \"%s\"",
1997 : item->defname, flagbuf.data);
1998 : }
1999 :
2000 : /* Emit the warning, if any problems were found. */
2001 0 : if (detailbuf.len > 0)
2002 0 : ereport(WARNING,
2003 : errmsg("supplied plan advice was not enforced"),
2004 : errdetail("%s", detailbuf.data));
2005 : }
2006 :
2007 : #ifdef USE_ASSERT_CHECKING
2008 :
2009 : /*
2010 : * Fast hash function for a key consisting of an RTI and plan name.
2011 : */
2012 : static uint32
2013 : pgpa_ri_checker_hash_key(pgpa_ri_checker_key key)
2014 : {
2015 : fasthash_state hs;
2016 : int sp_len;
2017 :
2018 : fasthash_init(&hs, 0);
2019 :
2020 : hs.accum = key.rti;
2021 : fasthash_combine(&hs);
2022 :
2023 : /* plan_name can be NULL */
2024 : if (key.plan_name == NULL)
2025 : sp_len = 0;
2026 : else
2027 : sp_len = fasthash_accum_cstring(&hs, key.plan_name);
2028 :
2029 : /* hashfn_unstable.h recommends using string length as tweak */
2030 : return fasthash_final32(&hs, sp_len);
2031 : }
2032 :
2033 : #endif
2034 :
2035 : /*
2036 : * Save the range table identifier for one relation for future cross-checking.
2037 : */
2038 : static void
2039 297 : pgpa_ri_checker_save(pgpa_planner_state *pps, PlannerInfo *root,
2040 : RelOptInfo *rel)
2041 : {
2042 : #ifdef USE_ASSERT_CHECKING
2043 : pgpa_ri_checker_key key;
2044 : pgpa_ri_checker *check;
2045 : pgpa_identifier rid;
2046 : const char *rid_string;
2047 : bool found;
2048 :
2049 : key.rti = bms_singleton_member(rel->relids);
2050 : key.plan_name = root->plan_name;
2051 : pgpa_compute_identifier_by_rti(root, key.rti, &rid);
2052 : rid_string = pgpa_identifier_string(&rid);
2053 : check = pgpa_ri_check_insert(pps->ri_check_hash, key, &found);
2054 : Assert(!found || strcmp(check->rid_string, rid_string) == 0);
2055 : check->rid_string = rid_string;
2056 : #endif
2057 297 : }
2058 :
2059 : /*
2060 : * Validate that the range table identifiers we were able to generate during
2061 : * planning match the ones we generated from the final plan.
2062 : */
2063 : static void
2064 132 : pgpa_ri_checker_validate(pgpa_planner_state *pps, PlannedStmt *pstmt)
2065 : {
2066 : #ifdef USE_ASSERT_CHECKING
2067 : pgpa_identifier *rt_identifiers;
2068 : pgpa_ri_check_iterator it;
2069 : pgpa_ri_checker *check;
2070 :
2071 : /* Create identifiers from the planned statement. */
2072 : rt_identifiers = pgpa_create_identifiers_for_planned_stmt(pstmt);
2073 :
2074 : /* Iterate over identifiers created during planning, so we can compare. */
2075 : pgpa_ri_check_start_iterate(pps->ri_check_hash, &it);
2076 : while ((check = pgpa_ri_check_iterate(pps->ri_check_hash, &it)) != NULL)
2077 : {
2078 : int rtoffset = 0;
2079 : const char *rid_string;
2080 : Index flat_rti;
2081 :
2082 : /*
2083 : * If there's no plan name associated with this entry, then the
2084 : * rtoffset is 0. Otherwise, we can search the SubPlanRTInfo list to
2085 : * find the rtoffset.
2086 : */
2087 : if (check->key.plan_name != NULL)
2088 : {
2089 : foreach_node(SubPlanRTInfo, rtinfo, pstmt->subrtinfos)
2090 : {
2091 : /*
2092 : * If rtinfo->dummy is set, then the subquery's range table
2093 : * will only have been partially copied to the final range
2094 : * table. Specifically, only RTE_RELATION entries and
2095 : * RTE_SUBQUERY entries that were once RTE_RELATION entries
2096 : * will be copied, as per add_rtes_to_flat_rtable. Therefore,
2097 : * there's no fixed rtoffset that we can apply to the RTIs
2098 : * used during planning to locate the corresponding relations
2099 : * in the final rtable.
2100 : *
2101 : * With more complex logic, we could work around that problem
2102 : * by remembering the whole contents of the subquery's rtable
2103 : * during planning, determining which of those would have been
2104 : * copied to the final rtable, and matching them up. But it
2105 : * doesn't seem like a worthwhile endeavor for right now,
2106 : * because RTIs from such subqueries won't appear in the plan
2107 : * tree itself, just in the range table. Hence, we can neither
2108 : * generate nor accept advice for them.
2109 : */
2110 : if (strcmp(check->key.plan_name, rtinfo->plan_name) == 0
2111 : && !rtinfo->dummy)
2112 : {
2113 : rtoffset = rtinfo->rtoffset;
2114 : Assert(rtoffset > 0);
2115 : break;
2116 : }
2117 : }
2118 :
2119 : /*
2120 : * It's not an error if we don't find the plan name: that just
2121 : * means that we planned a subplan by this name but it ended up
2122 : * being a dummy subplan and so wasn't included in the final plan
2123 : * tree.
2124 : */
2125 : if (rtoffset == 0)
2126 : continue;
2127 : }
2128 :
2129 : /*
2130 : * check->key.rti is the RTI that we saw prior to range-table
2131 : * flattening, so we must add the appropriate RT offset to get the
2132 : * final RTI.
2133 : */
2134 : flat_rti = check->key.rti + rtoffset;
2135 : Assert(flat_rti <= list_length(pstmt->rtable));
2136 :
2137 : /* Assert that the string we compute now matches the previous one. */
2138 : rid_string = pgpa_identifier_string(&rt_identifiers[flat_rti - 1]);
2139 : Assert(strcmp(rid_string, check->rid_string) == 0);
2140 : }
2141 : #endif
2142 132 : }
2143 :
2144 : /*
2145 : * Convert a bitmapset to a C string of comma-separated integers.
2146 : */
2147 : static char *
2148 0 : pgpa_bms_to_cstring(Bitmapset *bms)
2149 : {
2150 : StringInfoData buf;
2151 0 : int x = -1;
2152 :
2153 0 : if (bms_is_empty(bms))
2154 0 : return "none";
2155 :
2156 0 : initStringInfo(&buf);
2157 0 : while ((x = bms_next_member(bms, x)) >= 0)
2158 : {
2159 0 : if (buf.len > 0)
2160 0 : appendStringInfo(&buf, ", %d", x);
2161 : else
2162 0 : appendStringInfo(&buf, "%d", x);
2163 : }
2164 :
2165 0 : return buf.data;
2166 : }
2167 :
2168 : /*
2169 : * Convert a JoinType to a C string.
2170 : */
2171 : static const char *
2172 0 : pgpa_jointype_to_cstring(JoinType jointype)
2173 : {
2174 0 : switch (jointype)
2175 : {
2176 0 : case JOIN_INNER:
2177 0 : return "inner";
2178 0 : case JOIN_LEFT:
2179 0 : return "left";
2180 0 : case JOIN_FULL:
2181 0 : return "full";
2182 0 : case JOIN_RIGHT:
2183 0 : return "right";
2184 0 : case JOIN_SEMI:
2185 0 : return "semi";
2186 0 : case JOIN_ANTI:
2187 0 : return "anti";
2188 0 : case JOIN_RIGHT_SEMI:
2189 0 : return "right semi";
2190 0 : case JOIN_RIGHT_ANTI:
2191 0 : return "right anti";
2192 0 : case JOIN_UNIQUE_OUTER:
2193 0 : return "unique outer";
2194 0 : case JOIN_UNIQUE_INNER:
2195 0 : return "unique inner";
2196 : }
2197 0 : return "???";
2198 : }
|
