| 1 |
/*------------------------------------------------------------------------- |
| 2 |
* |
| 3 |
* planner.c |
| 4 |
* The query optimizer external interface. |
| 5 |
* |
| 6 |
* Portions Copyright (c) 1996-2025, PostgreSQL Global Development Group |
| 7 |
* Portions Copyright (c) 1994, Regents of the University of California |
| 8 |
* |
| 9 |
* |
| 10 |
* IDENTIFICATION |
| 11 |
* src/backend/optimizer/plan/planner.c |
| 12 |
* |
| 13 |
*------------------------------------------------------------------------- |
| 14 |
*/ |
| 15 |
|
| 16 |
#include "postgres.h" |
| 17 |
|
| 18 |
#include <limits.h> |
| 19 |
#include <math.h> |
| 20 |
|
| 21 |
#include "access/genam.h" |
| 22 |
#include "access/parallel.h" |
| 23 |
#include "access/sysattr.h" |
| 24 |
#include "access/table.h" |
| 25 |
#include "catalog/pg_aggregate.h" |
| 26 |
#include "catalog/pg_inherits.h" |
| 27 |
#include "catalog/pg_proc.h" |
| 28 |
#include "catalog/pg_type.h" |
| 29 |
#include "executor/executor.h" |
| 30 |
#include "foreign/fdwapi.h" |
| 31 |
#include "jit/jit.h" |
| 32 |
#include "lib/bipartite_match.h" |
| 33 |
#include "lib/knapsack.h" |
| 34 |
#include "miscadmin.h" |
| 35 |
#include "nodes/makefuncs.h" |
| 36 |
#include "nodes/nodeFuncs.h" |
| 37 |
#ifdef OPTIMIZER_DEBUG |
| 38 |
#include "nodes/print.h" |
| 39 |
#endif |
| 40 |
#include "nodes/supportnodes.h" |
| 41 |
#include "optimizer/appendinfo.h" |
| 42 |
#include "optimizer/clauses.h" |
| 43 |
#include "optimizer/cost.h" |
| 44 |
#include "optimizer/optimizer.h" |
| 45 |
#include "optimizer/paramassign.h" |
| 46 |
#include "optimizer/pathnode.h" |
| 47 |
#include "optimizer/paths.h" |
| 48 |
#include "optimizer/plancat.h" |
| 49 |
#include "optimizer/planmain.h" |
| 50 |
#include "optimizer/planner.h" |
| 51 |
#include "optimizer/prep.h" |
| 52 |
#include "optimizer/subselect.h" |
| 53 |
#include "optimizer/tlist.h" |
| 54 |
#include "parser/analyze.h" |
| 55 |
#include "parser/parse_agg.h" |
| 56 |
#include "parser/parse_clause.h" |
| 57 |
#include "parser/parse_relation.h" |
| 58 |
#include "parser/parsetree.h" |
| 59 |
#include "partitioning/partdesc.h" |
| 60 |
#include "rewrite/rewriteManip.h" |
| 61 |
#include "utils/acl.h" |
| 62 |
#include "utils/backend_status.h" |
| 63 |
#include "utils/lsyscache.h" |
| 64 |
#include "utils/rel.h" |
| 65 |
#include "utils/selfuncs.h" |
| 66 |
|
| 67 |
/* GUC parameters */ |
| 68 |
double cursor_tuple_fraction = DEFAULT_CURSOR_TUPLE_FRACTION; |
| 69 |
int debug_parallel_query = DEBUG_PARALLEL_OFF; |
| 70 |
bool parallel_leader_participation = true; |
| 71 |
bool enable_distinct_reordering = true; |
| 72 |
|
| 73 |
/* Hook for plugins to get control in planner() */ |
| 74 |
planner_hook_type planner_hook = NULL; |
| 75 |
|
| 76 |
/* Hook for plugins to get control when grouping_planner() plans upper rels */ |
| 77 |
create_upper_paths_hook_type create_upper_paths_hook = NULL; |
| 78 |
|
| 79 |
|
| 80 |
/* Expression kind codes for preprocess_expression */ |
| 81 |
#define EXPRKIND_QUAL 0 |
| 82 |
#define EXPRKIND_TARGET 1 |
| 83 |
#define EXPRKIND_RTFUNC 2 |
| 84 |
#define EXPRKIND_RTFUNC_LATERAL 3 |
| 85 |
#define EXPRKIND_VALUES 4 |
| 86 |
#define EXPRKIND_VALUES_LATERAL 5 |
| 87 |
#define EXPRKIND_LIMIT 6 |
| 88 |
#define EXPRKIND_APPINFO 7 |
| 89 |
#define EXPRKIND_PHV 8 |
| 90 |
#define EXPRKIND_TABLESAMPLE 9 |
| 91 |
#define EXPRKIND_ARBITER_ELEM 10 |
| 92 |
#define EXPRKIND_TABLEFUNC 11 |
| 93 |
#define EXPRKIND_TABLEFUNC_LATERAL 12 |
| 94 |
#define EXPRKIND_GROUPEXPR 13 |
| 95 |
|
| 96 |
/* |
| 97 |
* Data specific to grouping sets |
| 98 |
*/ |
| 99 |
typedef struct |
| 100 |
{ |
| 101 |
List *rollups; |
| 102 |
List *hash_sets_idx; |
| 103 |
double dNumHashGroups; |
| 104 |
bool any_hashable; |
| 105 |
Bitmapset *unsortable_refs; |
| 106 |
Bitmapset *unhashable_refs; |
| 107 |
List *unsortable_sets; |
| 108 |
int *tleref_to_colnum_map; |
| 109 |
} grouping_sets_data; |
| 110 |
|
| 111 |
/* |
| 112 |
* Temporary structure for use during WindowClause reordering in order to be |
| 113 |
* able to sort WindowClauses on partitioning/ordering prefix. |
| 114 |
*/ |
| 115 |
typedef struct |
| 116 |
{ |
| 117 |
WindowClause *wc; |
| 118 |
List *uniqueOrder; /* A List of unique ordering/partitioning |
| 119 |
* clauses per Window */ |
| 120 |
} WindowClauseSortData; |
| 121 |
|
| 122 |
/* Passthrough data for standard_qp_callback */ |
| 123 |
typedef struct |
| 124 |
{ |
| 125 |
List *activeWindows; /* active windows, if any */ |
| 126 |
grouping_sets_data *gset_data; /* grouping sets data, if any */ |
| 127 |
SetOperationStmt *setop; /* parent set operation or NULL if not a |
| 128 |
* subquery belonging to a set operation */ |
| 129 |
} standard_qp_extra; |
| 130 |
|
| 131 |
/* |
| 132 |
* Context for find_having_conflicts. This is the callback context passed to |
| 133 |
* expression_has_grouping_conflict in clauses.c. |
| 134 |
*/ |
| 135 |
typedef struct |
| 136 |
{ |
| 137 |
Query *parse; |
| 138 |
Index group_rtindex; |
| 139 |
} having_grouping_ctx; |
| 140 |
|
| 141 |
/* Local functions */ |
| 142 |
static Node *preprocess_expression(PlannerInfo *root, Node *expr, int kind); |
| 143 |
static void preprocess_qual_conditions(PlannerInfo *root, Node *jtnode); |
| 144 |
static Bitmapset *find_having_conflicts(Query *parse, Index group_rtindex); |
| 145 |
static Oid having_var_grouping_eqop(Var *var, void *context); |
| 146 |
static Oid group_var_eqop(Query *parse, Var *var); |
| 147 |
static void grouping_planner(PlannerInfo *root, double tuple_fraction, |
| 148 |
SetOperationStmt *setops); |
| 149 |
static grouping_sets_data *preprocess_grouping_sets(PlannerInfo *root); |
| 150 |
static List *remap_to_groupclause_idx(List *groupClause, List *gsets, |
| 151 |
int *tleref_to_colnum_map); |
| 152 |
static void preprocess_rowmarks(PlannerInfo *root); |
| 153 |
static double preprocess_limit(PlannerInfo *root, |
| 154 |
double tuple_fraction, |
| 155 |
int64 *offset_est, int64 *count_est); |
| 156 |
static List *preprocess_groupclause(PlannerInfo *root, List *force); |
| 157 |
static List *extract_rollup_sets(List *groupingSets); |
| 158 |
static List *reorder_grouping_sets(List *groupingSets, List *sortclause); |
| 159 |
static void standard_qp_callback(PlannerInfo *root, void *extra); |
| 160 |
static double get_number_of_groups(PlannerInfo *root, |
| 161 |
double path_rows, |
| 162 |
grouping_sets_data *gd, |
| 163 |
List *target_list); |
| 164 |
static RelOptInfo *create_grouping_paths(PlannerInfo *root, |
| 165 |
RelOptInfo *input_rel, |
| 166 |
PathTarget *target, |
| 167 |
bool target_parallel_safe, |
| 168 |
grouping_sets_data *gd); |
| 169 |
static bool is_degenerate_grouping(PlannerInfo *root); |
| 170 |
static void create_degenerate_grouping_paths(PlannerInfo *root, |
| 171 |
RelOptInfo *input_rel, |
| 172 |
RelOptInfo *grouped_rel); |
| 173 |
static RelOptInfo *make_grouping_rel(PlannerInfo *root, RelOptInfo *input_rel, |
| 174 |
PathTarget *target, bool target_parallel_safe, |
| 175 |
Node *havingQual); |
| 176 |
static void create_ordinary_grouping_paths(PlannerInfo *root, |
| 177 |
RelOptInfo *input_rel, |
| 178 |
RelOptInfo *grouped_rel, |
| 179 |
const AggClauseCosts *agg_costs, |
| 180 |
grouping_sets_data *gd, |
| 181 |
GroupPathExtraData *extra, |
| 182 |
RelOptInfo **partially_grouped_rel_p); |
| 183 |
static void consider_groupingsets_paths(PlannerInfo *root, |
| 184 |
RelOptInfo *grouped_rel, |
| 185 |
Path *path, |
| 186 |
bool is_sorted, |
| 187 |
bool can_hash, |
| 188 |
grouping_sets_data *gd, |
| 189 |
const AggClauseCosts *agg_costs, |
| 190 |
double dNumGroups); |
| 191 |
static RelOptInfo *create_window_paths(PlannerInfo *root, |
| 192 |
RelOptInfo *input_rel, |
| 193 |
PathTarget *input_target, |
| 194 |
PathTarget *output_target, |
| 195 |
bool output_target_parallel_safe, |
| 196 |
WindowFuncLists *wflists, |
| 197 |
List *activeWindows); |
| 198 |
static void create_one_window_path(PlannerInfo *root, |
| 199 |
RelOptInfo *window_rel, |
| 200 |
Path *path, |
| 201 |
PathTarget *input_target, |
| 202 |
PathTarget *output_target, |
| 203 |
WindowFuncLists *wflists, |
| 204 |
List *activeWindows); |
| 205 |
static RelOptInfo *create_distinct_paths(PlannerInfo *root, |
| 206 |
RelOptInfo *input_rel, |
| 207 |
PathTarget *target); |
| 208 |
static void create_partial_distinct_paths(PlannerInfo *root, |
| 209 |
RelOptInfo *input_rel, |
| 210 |
RelOptInfo *final_distinct_rel, |
| 211 |
PathTarget *target); |
| 212 |
static RelOptInfo *create_final_distinct_paths(PlannerInfo *root, |
| 213 |
RelOptInfo *input_rel, |
| 214 |
RelOptInfo *distinct_rel); |
| 215 |
static List *get_useful_pathkeys_for_distinct(PlannerInfo *root, |
| 216 |
List *needed_pathkeys, |
| 217 |
List *path_pathkeys); |
| 218 |
static RelOptInfo *create_ordered_paths(PlannerInfo *root, |
| 219 |
RelOptInfo *input_rel, |
| 220 |
PathTarget *target, |
| 221 |
bool target_parallel_safe, |
| 222 |
double limit_tuples); |
| 223 |
static PathTarget *make_group_input_target(PlannerInfo *root, |
| 224 |
PathTarget *final_target); |
| 225 |
static PathTarget *make_partial_grouping_target(PlannerInfo *root, |
| 226 |
PathTarget *grouping_target, |
| 227 |
Node *havingQual); |
| 228 |
static List *postprocess_setop_tlist(List *new_tlist, List *orig_tlist); |
| 229 |
static void optimize_window_clauses(PlannerInfo *root, |
| 230 |
WindowFuncLists *wflists); |
| 231 |
static List *select_active_windows(PlannerInfo *root, WindowFuncLists *wflists); |
| 232 |
static void name_active_windows(List *activeWindows); |
| 233 |
static PathTarget *make_window_input_target(PlannerInfo *root, |
| 234 |
PathTarget *final_target, |
| 235 |
List *activeWindows); |
| 236 |
static List *make_pathkeys_for_window(PlannerInfo *root, WindowClause *wc, |
| 237 |
List *tlist); |
| 238 |
static PathTarget *make_sort_input_target(PlannerInfo *root, |
| 239 |
PathTarget *final_target, |
| 240 |
bool *have_postponed_srfs); |
| 241 |
static void adjust_paths_for_srfs(PlannerInfo *root, RelOptInfo *rel, |
| 242 |
List *targets, List *targets_contain_srfs); |
| 243 |
static void add_paths_to_grouping_rel(PlannerInfo *root, RelOptInfo *input_rel, |
| 244 |
RelOptInfo *grouped_rel, |
| 245 |
RelOptInfo *partially_grouped_rel, |
| 246 |
const AggClauseCosts *agg_costs, |
| 247 |
grouping_sets_data *gd, |
| 248 |
double dNumGroups, |
| 249 |
GroupPathExtraData *extra); |
| 250 |
static RelOptInfo *create_partial_grouping_paths(PlannerInfo *root, |
| 251 |
RelOptInfo *grouped_rel, |
| 252 |
RelOptInfo *input_rel, |
| 253 |
grouping_sets_data *gd, |
| 254 |
GroupPathExtraData *extra, |
| 255 |
bool force_rel_creation); |
| 256 |
static Path *make_ordered_path(PlannerInfo *root, |
| 257 |
RelOptInfo *rel, |
| 258 |
Path *path, |
| 259 |
Path *cheapest_path, |
| 260 |
List *pathkeys, |
| 261 |
double limit_tuples); |
| 262 |
static void gather_grouping_paths(PlannerInfo *root, RelOptInfo *rel); |
| 263 |
static bool can_partial_agg(PlannerInfo *root); |
| 264 |
static void apply_scanjoin_target_to_paths(PlannerInfo *root, |
| 265 |
RelOptInfo *rel, |
| 266 |
List *scanjoin_targets, |
| 267 |
List *scanjoin_targets_contain_srfs, |
| 268 |
bool scanjoin_target_parallel_safe, |
| 269 |
bool tlist_same_exprs); |
| 270 |
static void create_partitionwise_grouping_paths(PlannerInfo *root, |
| 271 |
RelOptInfo *input_rel, |
| 272 |
RelOptInfo *grouped_rel, |
| 273 |
RelOptInfo *partially_grouped_rel, |
| 274 |
const AggClauseCosts *agg_costs, |
| 275 |
grouping_sets_data *gd, |
| 276 |
PartitionwiseAggregateType patype, |
| 277 |
GroupPathExtraData *extra); |
| 278 |
static bool group_by_has_partkey(RelOptInfo *input_rel, |
| 279 |
List *targetList, |
| 280 |
List *groupClause); |
| 281 |
static int common_prefix_cmp(const void *a, const void *b); |
| 282 |
static List *generate_setop_child_grouplist(SetOperationStmt *op, |
| 283 |
List *targetlist); |
| 284 |
|
| 285 |
|
| 286 |
/***************************************************************************** |
| 287 |
* |
| 288 |
* Query optimizer entry point |
| 289 |
* |
| 290 |
* To support loadable plugins that monitor or modify planner behavior, |
| 291 |
* we provide a hook variable that lets a plugin get control before and |
| 292 |
* after the standard planning process. The plugin would normally call |
| 293 |
* standard_planner(). |
| 294 |
* |
| 295 |
* Note to plugin authors: standard_planner() scribbles on its Query input, |
| 296 |
* so you'd better copy that data structure if you want to plan more than once. |
| 297 |
* |
| 298 |
*****************************************************************************/ |
| 299 |
PlannedStmt * |
| 300 |
planner(Query *parse, const char *query_string, int cursorOptions, |
| 301 |
ParamListInfo boundParams) |
| 302 |
{ |
| 303 |
PlannedStmt *result; |
| 304 |
|
| 305 |
if (planner_hook) |
| 306 |
result = (*planner_hook) (parse, query_string, cursorOptions, boundParams); |
| 307 |
else |
| 308 |
result = standard_planner(parse, query_string, cursorOptions, boundParams); |
| 309 |
|
| 310 |
pgstat_report_plan_id(result->planId, false); |
| 311 |
|
| 312 |
return result; |
| 313 |
} |
| 314 |
|
| 315 |
PlannedStmt * |
| 316 |
standard_planner(Query *parse, const char *query_string, int cursorOptions, |
| 317 |
ParamListInfo boundParams) |
| 318 |
{ |
| 319 |
PlannedStmt *result; |
| 320 |
PlannerGlobal *glob; |
| 321 |
double tuple_fraction; |
| 322 |
PlannerInfo *root; |
| 323 |
RelOptInfo *final_rel; |
| 324 |
Path *best_path; |
| 325 |
Plan *top_plan; |
| 326 |
ListCell *lp, |
| 327 |
*lr; |
| 328 |
|
| 329 |
/* |
| 330 |
* Set up global state for this planner invocation. This data is needed |
| 331 |
* across all levels of sub-Query that might exist in the given command, |
| 332 |
* so we keep it in a separate struct that's linked to by each per-Query |
| 333 |
* PlannerInfo. |
| 334 |
*/ |
| 335 |
glob = makeNode(PlannerGlobal); |
| 336 |
|
| 337 |
glob->boundParams = boundParams; |
| 338 |
glob->subplans = NIL; |
| 339 |
glob->subpaths = NIL; |
| 340 |
glob->subroots = NIL; |
| 341 |
glob->rewindPlanIDs = NULL; |
| 342 |
glob->finalrtable = NIL; |
| 343 |
glob->allRelids = NULL; |
| 344 |
glob->prunableRelids = NULL; |
| 345 |
glob->finalrteperminfos = NIL; |
| 346 |
glob->finalrowmarks = NIL; |
| 347 |
glob->resultRelations = NIL; |
| 348 |
glob->appendRelations = NIL; |
| 349 |
glob->partPruneInfos = NIL; |
| 350 |
glob->relationOids = NIL; |
| 351 |
glob->invalItems = NIL; |
| 352 |
glob->paramExecTypes = NIL; |
| 353 |
glob->lastPHId = 0; |
| 354 |
glob->lastRowMarkId = 0; |
| 355 |
glob->lastPlanNodeId = 0; |
| 356 |
glob->transientPlan = false; |
| 357 |
glob->dependsOnRole = false; |
| 358 |
glob->partition_directory = NULL; |
| 359 |
|
| 360 |
/* |
| 361 |
* Assess whether it's feasible to use parallel mode for this query. We |
| 362 |
* can't do this in a standalone backend, or if the command will try to |
| 363 |
* modify any data, or if this is a cursor operation, or if GUCs are set |
| 364 |
* to values that don't permit parallelism, or if parallel-unsafe |
| 365 |
* functions are present in the query tree. |
| 366 |
* |
| 367 |
* (Note that we do allow CREATE TABLE AS, SELECT INTO, and CREATE |
| 368 |
* MATERIALIZED VIEW to use parallel plans, but this is safe only because |
| 369 |
* the command is writing into a completely new table which workers won't |
| 370 |
* be able to see. If the workers could see the table, the fact that |
| 371 |
* group locking would cause them to ignore the leader's heavyweight GIN |
| 372 |
* page locks would make this unsafe. We'll have to fix that somehow if |
| 373 |
* we want to allow parallel inserts in general; updates and deletes have |
| 374 |
* additional problems especially around combo CIDs.) |
| 375 |
* |
| 376 |
* For now, we don't try to use parallel mode if we're running inside a |
| 377 |
* parallel worker. We might eventually be able to relax this |
| 378 |
* restriction, but for now it seems best not to have parallel workers |
| 379 |
* trying to create their own parallel workers. |
| 380 |
*/ |
| 381 |
if ((cursorOptions & CURSOR_OPT_PARALLEL_OK) != 0 && |
| 382 |
IsUnderPostmaster && |
| 383 |
parse->commandType == CMD_SELECT && |
| 384 |
!parse->hasModifyingCTE && |
| 385 |
max_parallel_workers_per_gather > 0 && |
| 386 |
!IsParallelWorker()) |
| 387 |
{ |
| 388 |
/* all the cheap tests pass, so scan the query tree */ |
| 389 |
glob->maxParallelHazard = max_parallel_hazard(parse); |
| 390 |
glob->parallelModeOK = (glob->maxParallelHazard != PROPARALLEL_UNSAFE); |
| 391 |
} |
| 392 |
else |
| 393 |
{ |
| 394 |
/* skip the query tree scan, just assume it's unsafe */ |
| 395 |
glob->maxParallelHazard = PROPARALLEL_UNSAFE; |
| 396 |
glob->parallelModeOK = false; |
| 397 |
} |
| 398 |
|
| 399 |
/* |
| 400 |
* glob->parallelModeNeeded is normally set to false here and changed to |
| 401 |
* true during plan creation if a Gather or Gather Merge plan is actually |
| 402 |
* created (cf. create_gather_plan, create_gather_merge_plan). |
| 403 |
* |
| 404 |
* However, if debug_parallel_query = on or debug_parallel_query = |
| 405 |
* regress, then we impose parallel mode whenever it's safe to do so, even |
| 406 |
* if the final plan doesn't use parallelism. It's not safe to do so if |
| 407 |
* the query contains anything parallel-unsafe; parallelModeOK will be |
| 408 |
* false in that case. Note that parallelModeOK can't change after this |
| 409 |
* point. Otherwise, everything in the query is either parallel-safe or |
| 410 |
* parallel-restricted, and in either case it should be OK to impose |
| 411 |
* parallel-mode restrictions. If that ends up breaking something, then |
| 412 |
* either some function the user included in the query is incorrectly |
| 413 |
* labeled as parallel-safe or parallel-restricted when in reality it's |
| 414 |
* parallel-unsafe, or else the query planner itself has a bug. |
| 415 |
*/ |
| 416 |
glob->parallelModeNeeded = glob->parallelModeOK && |
| 417 |
(debug_parallel_query != DEBUG_PARALLEL_OFF); |
| 418 |
|
| 419 |
/* Determine what fraction of the plan is likely to be scanned */ |
| 420 |
if (cursorOptions & CURSOR_OPT_FAST_PLAN) |
| 421 |
{ |
| 422 |
/* |
| 423 |
* We have no real idea how many tuples the user will ultimately FETCH |
| 424 |
* from a cursor, but it is often the case that he doesn't want 'em |
| 425 |
* all, or would prefer a fast-start plan anyway so that he can |
| 426 |
* process some of the tuples sooner. Use a GUC parameter to decide |
| 427 |
* what fraction to optimize for. |
| 428 |
*/ |
| 429 |
tuple_fraction = cursor_tuple_fraction; |
| 430 |
|
| 431 |
/* |
| 432 |
* We document cursor_tuple_fraction as simply being a fraction, which |
| 433 |
* means the edge cases 0 and 1 have to be treated specially here. We |
| 434 |
* convert 1 to 0 ("all the tuples") and 0 to a very small fraction. |
| 435 |
*/ |
| 436 |
if (tuple_fraction >= 1.0) |
| 437 |
tuple_fraction = 0.0; |
| 438 |
else if (tuple_fraction <= 0.0) |
| 439 |
tuple_fraction = 1e-10; |
| 440 |
} |
| 441 |
else |
| 442 |
{ |
| 443 |
/* Default assumption is we need all the tuples */ |
| 444 |
tuple_fraction = 0.0; |
| 445 |
} |
| 446 |
|
| 447 |
/* primary planning entry point (may recurse for subqueries) */ |
| 448 |
root = subquery_planner(glob, parse, NULL, false, tuple_fraction, NULL); |
| 449 |
|
| 450 |
/* Select best Path and turn it into a Plan */ |
| 451 |
final_rel = fetch_upper_rel(root, UPPERREL_FINAL, NULL); |
| 452 |
best_path = get_cheapest_fractional_path(final_rel, tuple_fraction); |
| 453 |
|
| 454 |
top_plan = create_plan(root, best_path); |
| 455 |
|
| 456 |
/* |
| 457 |
* If creating a plan for a scrollable cursor, make sure it can run |
| 458 |
* backwards on demand. Add a Material node at the top at need. |
| 459 |
*/ |
| 460 |
if (cursorOptions & CURSOR_OPT_SCROLL) |
| 461 |
{ |
| 462 |
if (!ExecSupportsBackwardScan(top_plan)) |
| 463 |
top_plan = materialize_finished_plan(top_plan); |
| 464 |
} |
| 465 |
|
| 466 |
/* |
| 467 |
* Optionally add a Gather node for testing purposes, provided this is |
| 468 |
* actually a safe thing to do. |
| 469 |
* |
| 470 |
* We can add Gather even when top_plan has parallel-safe initPlans, but |
| 471 |
* then we have to move the initPlans to the Gather node because of |
| 472 |
* SS_finalize_plan's limitations. That would cause cosmetic breakage of |
| 473 |
* regression tests when debug_parallel_query = regress, because initPlans |
| 474 |
* that would normally appear on the top_plan move to the Gather, causing |
| 475 |
* them to disappear from EXPLAIN output. That doesn't seem worth kluging |
| 476 |
* EXPLAIN to hide, so skip it when debug_parallel_query = regress. |
| 477 |
*/ |
| 478 |
if (debug_parallel_query != DEBUG_PARALLEL_OFF && |
| 479 |
top_plan->parallel_safe && |
| 480 |
(top_plan->initPlan == NIL || |
| 481 |
debug_parallel_query != DEBUG_PARALLEL_REGRESS)) |
| 482 |
{ |
| 483 |
Gather *gather = makeNode(Gather); |
| 484 |
Cost initplan_cost; |
| 485 |
bool unsafe_initplans; |
| 486 |
|
| 487 |
gather->plan.targetlist = top_plan->targetlist; |
| 488 |
gather->plan.qual = NIL; |
| 489 |
gather->plan.lefttree = top_plan; |
| 490 |
gather->plan.righttree = NULL; |
| 491 |
gather->num_workers = 1; |
| 492 |
gather->single_copy = true; |
| 493 |
gather->invisible = (debug_parallel_query == DEBUG_PARALLEL_REGRESS); |
| 494 |
|
| 495 |
/* Transfer any initPlans to the new top node */ |
| 496 |
gather->plan.initPlan = top_plan->initPlan; |
| 497 |
top_plan->initPlan = NIL; |
| 498 |
|
| 499 |
/* |
| 500 |
* Since this Gather has no parallel-aware descendants to signal to, |
| 501 |
* we don't need a rescan Param. |
| 502 |
*/ |
| 503 |
gather->rescan_param = -1; |
| 504 |
|
| 505 |
/* |
| 506 |
* Ideally we'd use cost_gather here, but setting up dummy path data |
| 507 |
* to satisfy it doesn't seem much cleaner than knowing what it does. |
| 508 |
*/ |
| 509 |
gather->plan.startup_cost = top_plan->startup_cost + |
| 510 |
parallel_setup_cost; |
| 511 |
gather->plan.total_cost = top_plan->total_cost + |
| 512 |
parallel_setup_cost + parallel_tuple_cost * top_plan->plan_rows; |
| 513 |
gather->plan.plan_rows = top_plan->plan_rows; |
| 514 |
gather->plan.plan_width = top_plan->plan_width; |
| 515 |
gather->plan.parallel_aware = false; |
| 516 |
gather->plan.parallel_safe = false; |
| 517 |
|
| 518 |
/* |
| 519 |
* Delete the initplans' cost from top_plan. We needn't add it to the |
| 520 |
* Gather node, since the above coding already included it there. |
| 521 |
*/ |
| 522 |
SS_compute_initplan_cost(gather->plan.initPlan, |
| 523 |
&initplan_cost, &unsafe_initplans); |
| 524 |
top_plan->startup_cost -= initplan_cost; |
| 525 |
top_plan->total_cost -= initplan_cost; |
| 526 |
|
| 527 |
/* use parallel mode for parallel plans. */ |
| 528 |
root->glob->parallelModeNeeded = true; |
| 529 |
|
| 530 |
top_plan = &gather->plan; |
| 531 |
} |
| 532 |
|
| 533 |
/* |
| 534 |
* If any Params were generated, run through the plan tree and compute |
| 535 |
* each plan node's extParam/allParam sets. Ideally we'd merge this into |
| 536 |
* set_plan_references' tree traversal, but for now it has to be separate |
| 537 |
* because we need to visit subplans before not after main plan. |
| 538 |
*/ |
| 539 |
if (glob->paramExecTypes != NIL) |
| 540 |
{ |
| 541 |
Assert(list_length(glob->subplans) == list_length(glob->subroots)); |
| 542 |
forboth(lp, glob->subplans, lr, glob->subroots) |
| 543 |
{ |
| 544 |
Plan *subplan = (Plan *) lfirst(lp); |
| 545 |
PlannerInfo *subroot = lfirst_node(PlannerInfo, lr); |
| 546 |
|
| 547 |
SS_finalize_plan(subroot, subplan); |
| 548 |
} |
| 549 |
SS_finalize_plan(root, top_plan); |
| 550 |
} |
| 551 |
|
| 552 |
/* final cleanup of the plan */ |
| 553 |
Assert(glob->finalrtable == NIL); |
| 554 |
Assert(glob->finalrteperminfos == NIL); |
| 555 |
Assert(glob->finalrowmarks == NIL); |
| 556 |
Assert(glob->resultRelations == NIL); |
| 557 |
Assert(glob->appendRelations == NIL); |
| 558 |
top_plan = set_plan_references(root, top_plan); |
| 559 |
/* ... and the subplans (both regular subplans and initplans) */ |
| 560 |
Assert(list_length(glob->subplans) == list_length(glob->subroots)); |
| 561 |
forboth(lp, glob->subplans, lr, glob->subroots) |
| 562 |
{ |
| 563 |
Plan *subplan = (Plan *) lfirst(lp); |
| 564 |
PlannerInfo *subroot = lfirst_node(PlannerInfo, lr); |
| 565 |
|
| 566 |
lfirst(lp) = set_plan_references(subroot, subplan); |
| 567 |
} |
| 568 |
|
| 569 |
/* build the PlannedStmt result */ |
| 570 |
result = makeNode(PlannedStmt); |
| 571 |
|
| 572 |
result->commandType = parse->commandType; |
| 573 |
result->queryId = parse->queryId; |
| 574 |
result->hasReturning = (parse->returningList != NIL); |
| 575 |
result->hasModifyingCTE = parse->hasModifyingCTE; |
| 576 |
result->canSetTag = parse->canSetTag; |
| 577 |
result->transientPlan = glob->transientPlan; |
| 578 |
result->dependsOnRole = glob->dependsOnRole; |
| 579 |
result->parallelModeNeeded = glob->parallelModeNeeded; |
| 580 |
result->planTree = top_plan; |
| 581 |
result->partPruneInfos = glob->partPruneInfos; |
| 582 |
result->rtable = glob->finalrtable; |
| 583 |
result->unprunableRelids = bms_difference(glob->allRelids, |
| 584 |
glob->prunableRelids); |
| 585 |
result->permInfos = glob->finalrteperminfos; |
| 586 |
result->resultRelations = glob->resultRelations; |
| 587 |
result->appendRelations = glob->appendRelations; |
| 588 |
result->subplans = glob->subplans; |
| 589 |
result->rewindPlanIDs = glob->rewindPlanIDs; |
| 590 |
result->rowMarks = glob->finalrowmarks; |
| 591 |
result->relationOids = glob->relationOids; |
| 592 |
result->invalItems = glob->invalItems; |
| 593 |
result->paramExecTypes = glob->paramExecTypes; |
| 594 |
/* utilityStmt should be null, but we might as well copy it */ |
| 595 |
result->utilityStmt = parse->utilityStmt; |
| 596 |
result->stmt_location = parse->stmt_location; |
| 597 |
result->stmt_len = parse->stmt_len; |
| 598 |
|
| 599 |
result->jitFlags = PGJIT_NONE; |
| 600 |
if (jit_enabled && jit_above_cost >= 0 && |
| 601 |
top_plan->total_cost > jit_above_cost) |
| 602 |
{ |
| 603 |
result->jitFlags |= PGJIT_PERFORM; |
| 604 |
|
| 605 |
/* |
| 606 |
* Decide how much effort should be put into generating better code. |
| 607 |
*/ |
| 608 |
if (jit_optimize_above_cost >= 0 && |
| 609 |
top_plan->total_cost > jit_optimize_above_cost) |
| 610 |
result->jitFlags |= PGJIT_OPT3; |
| 611 |
if (jit_inline_above_cost >= 0 && |
| 612 |
top_plan->total_cost > jit_inline_above_cost) |
| 613 |
result->jitFlags |= PGJIT_INLINE; |
| 614 |
|
| 615 |
/* |
| 616 |
* Decide which operations should be JITed. |
| 617 |
*/ |
| 618 |
if (jit_expressions) |
| 619 |
result->jitFlags |= PGJIT_EXPR; |
| 620 |
if (jit_tuple_deforming) |
| 621 |
result->jitFlags |= PGJIT_DEFORM; |
| 622 |
} |
| 623 |
|
| 624 |
if (glob->partition_directory != NULL) |
| 625 |
DestroyPartitionDirectory(glob->partition_directory); |
| 626 |
|
| 627 |
return result; |
| 628 |
} |
| 629 |
|
| 630 |
|
| 631 |
/*-------------------- |
| 632 |
* subquery_planner |
| 633 |
* Invokes the planner on a subquery. We recurse to here for each |
| 634 |
* sub-SELECT found in the query tree. |
| 635 |
* |
| 636 |
* glob is the global state for the current planner run. |
| 637 |
* parse is the querytree produced by the parser & rewriter. |
| 638 |
* parent_root is the immediate parent Query's info (NULL at the top level). |
| 639 |
* hasRecursion is true if this is a recursive WITH query. |
| 640 |
* tuple_fraction is the fraction of tuples we expect will be retrieved. |
| 641 |
* tuple_fraction is interpreted as explained for grouping_planner, below. |
| 642 |
* setops is used for set operation subqueries to provide the subquery with |
| 643 |
* the context in which it's being used so that Paths correctly sorted for the |
| 644 |
* set operation can be generated. NULL when not planning a set operation |
| 645 |
* child, or when a child of a set op that isn't interested in sorted input. |
| 646 |
* |
| 647 |
* Basically, this routine does the stuff that should only be done once |
| 648 |
* per Query object. It then calls grouping_planner. At one time, |
| 649 |
* grouping_planner could be invoked recursively on the same Query object; |
| 650 |
* that's not currently true, but we keep the separation between the two |
| 651 |
* routines anyway, in case we need it again someday. |
| 652 |
* |
| 653 |
* subquery_planner will be called recursively to handle sub-Query nodes |
| 654 |
* found within the query's expressions and rangetable. |
| 655 |
* |
| 656 |
* Returns the PlannerInfo struct ("root") that contains all data generated |
| 657 |
* while planning the subquery. In particular, the Path(s) attached to |
| 658 |
* the (UPPERREL_FINAL, NULL) upperrel represent our conclusions about the |
| 659 |
* cheapest way(s) to implement the query. The top level will select the |
| 660 |
* best Path and pass it through createplan.c to produce a finished Plan. |
| 661 |
*-------------------- |
| 662 |
*/ |
| 663 |
PlannerInfo * |
| 664 |
subquery_planner(PlannerGlobal *glob, Query *parse, PlannerInfo *parent_root, |
| 665 |
bool hasRecursion, double tuple_fraction, |
| 666 |
SetOperationStmt *setops) |
| 667 |
{ |
| 668 |
PlannerInfo *root; |
| 669 |
List *newWithCheckOptions; |
| 670 |
List *newHaving; |
| 671 |
Bitmapset *havingPushdownConflicts; |
| 672 |
int havingIdx; |
| 673 |
bool hasOuterJoins; |
| 674 |
bool hasResultRTEs; |
| 675 |
RelOptInfo *final_rel; |
| 676 |
ListCell *l; |
| 677 |
|
| 678 |
/* Create a PlannerInfo data structure for this subquery */ |
| 679 |
root = makeNode(PlannerInfo); |
| 680 |
root->parse = parse; |
| 681 |
root->glob = glob; |
| 682 |
root->query_level = parent_root ? parent_root->query_level + 1 : 1; |
| 683 |
root->parent_root = parent_root; |
| 684 |
root->plan_params = NIL; |
| 685 |
root->outer_params = NULL; |
| 686 |
root->planner_cxt = CurrentMemoryContext; |
| 687 |
root->init_plans = NIL; |
| 688 |
root->cte_plan_ids = NIL; |
| 689 |
root->multiexpr_params = NIL; |
| 690 |
root->join_domains = NIL; |
| 691 |
root->eq_classes = NIL; |
| 692 |
root->ec_merging_done = false; |
| 693 |
root->last_rinfo_serial = 0; |
| 694 |
root->all_result_relids = |
| 695 |
parse->resultRelation ? bms_make_singleton(parse->resultRelation) : NULL; |
| 696 |
root->leaf_result_relids = NULL; /* we'll find out leaf-ness later */ |
| 697 |
root->append_rel_list = NIL; |
| 698 |
root->row_identity_vars = NIL; |
| 699 |
root->rowMarks = NIL; |
| 700 |
memset(root->upper_rels, 0, sizeof(root->upper_rels)); |
| 701 |
memset(root->upper_targets, 0, sizeof(root->upper_targets)); |
| 702 |
root->processed_groupClause = NIL; |
| 703 |
root->processed_distinctClause = NIL; |
| 704 |
root->processed_tlist = NIL; |
| 705 |
root->update_colnos = NIL; |
| 706 |
root->grouping_map = NULL; |
| 707 |
root->minmax_aggs = NIL; |
| 708 |
root->qual_security_level = 0; |
| 709 |
root->hasPseudoConstantQuals = false; |
| 710 |
root->hasAlternativeSubPlans = false; |
| 711 |
root->placeholdersFrozen = false; |
| 712 |
root->hasRecursion = hasRecursion; |
| 713 |
if (hasRecursion) |
| 714 |
root->wt_param_id = assign_special_exec_param(root); |
| 715 |
else |
| 716 |
root->wt_param_id = -1; |
| 717 |
root->non_recursive_path = NULL; |
| 718 |
root->partColsUpdated = false; |
| 719 |
|
| 720 |
/* |
| 721 |
* Create the top-level join domain. This won't have valid contents until |
| 722 |
* deconstruct_jointree fills it in, but the node needs to exist before |
| 723 |
* that so we can build EquivalenceClasses referencing it. |
| 724 |
*/ |
| 725 |
root->join_domains = list_make1(makeNode(JoinDomain)); |
| 726 |
|
| 727 |
/* |
| 728 |
* If there is a WITH list, process each WITH query and either convert it |
| 729 |
* to RTE_SUBQUERY RTE(s) or build an initplan SubPlan structure for it. |
| 730 |
*/ |
| 731 |
if (parse->cteList) |
| 732 |
SS_process_ctes(root); |
| 733 |
|
| 734 |
/* |
| 735 |
* If it's a MERGE command, transform the joinlist as appropriate. |
| 736 |
*/ |
| 737 |
transform_MERGE_to_join(parse); |
| 738 |
|
| 739 |
/* |
| 740 |
* If the FROM clause is empty, replace it with a dummy RTE_RESULT RTE, so |
| 741 |
* that we don't need so many special cases to deal with that situation. |
| 742 |
*/ |
| 743 |
replace_empty_jointree(parse); |
| 744 |
|
| 745 |
/* |
| 746 |
* Look for ANY and EXISTS SubLinks in WHERE and JOIN/ON clauses, and try |
| 747 |
* to transform them into joins. Note that this step does not descend |
| 748 |
* into subqueries; if we pull up any subqueries below, their SubLinks are |
| 749 |
* processed just before pulling them up. |
| 750 |
*/ |
| 751 |
if (parse->hasSubLinks) |
| 752 |
pull_up_sublinks(root); |
| 753 |
|
| 754 |
/* |
| 755 |
* Scan the rangetable for function RTEs, do const-simplification on them, |
| 756 |
* and then inline them if possible (producing subqueries that might get |
| 757 |
* pulled up next). Recursion issues here are handled in the same way as |
| 758 |
* for SubLinks. |
| 759 |
*/ |
| 760 |
preprocess_function_rtes(root); |
| 761 |
|
| 762 |
/* |
| 763 |
* Scan the rangetable for relations with virtual generated columns, and |
| 764 |
* replace all Var nodes in the query that reference these columns with |
| 765 |
* the generation expressions. Recursion issues here are handled in the |
| 766 |
* same way as for SubLinks. |
| 767 |
*/ |
| 768 |
parse = root->parse = expand_virtual_generated_columns(root); |
| 769 |
|
| 770 |
/* |
| 771 |
* Check to see if any subqueries in the jointree can be merged into this |
| 772 |
* query. |
| 773 |
*/ |
| 774 |
pull_up_subqueries(root); |
| 775 |
|
| 776 |
/* |
| 777 |
* If this is a simple UNION ALL query, flatten it into an appendrel. We |
| 778 |
* do this now because it requires applying pull_up_subqueries to the leaf |
| 779 |
* queries of the UNION ALL, which weren't touched above because they |
| 780 |
* weren't referenced by the jointree (they will be after we do this). |
| 781 |
*/ |
| 782 |
if (parse->setOperations) |
| 783 |
flatten_simple_union_all(root); |
| 784 |
|
| 785 |
/* |
| 786 |
* Survey the rangetable to see what kinds of entries are present. We can |
| 787 |
* skip some later processing if relevant SQL features are not used; for |
| 788 |
* example if there are no JOIN RTEs we can avoid the expense of doing |
| 789 |
* flatten_join_alias_vars(). This must be done after we have finished |
| 790 |
* adding rangetable entries, of course. (Note: actually, processing of |
| 791 |
* inherited or partitioned rels can cause RTEs for their child tables to |
| 792 |
* get added later; but those must all be RTE_RELATION entries, so they |
| 793 |
* don't invalidate the conclusions drawn here.) |
| 794 |
*/ |
| 795 |
root->hasJoinRTEs = false; |
| 796 |
root->hasLateralRTEs = false; |
| 797 |
root->group_rtindex = 0; |
| 798 |
hasOuterJoins = false; |
| 799 |
hasResultRTEs = false; |
| 800 |
foreach(l, parse->rtable) |
| 801 |
{ |
| 802 |
RangeTblEntry *rte = lfirst_node(RangeTblEntry, l); |
| 803 |
|
| 804 |
switch (rte->rtekind) |
| 805 |
{ |
| 806 |
case RTE_RELATION: |
| 807 |
if (rte->inh) |
| 808 |
{ |
| 809 |
/* |
| 810 |
* Check to see if the relation actually has any children; |
| 811 |
* if not, clear the inh flag so we can treat it as a |
| 812 |
* plain base relation. |
| 813 |
* |
| 814 |
* Note: this could give a false-positive result, if the |
| 815 |
* rel once had children but no longer does. We used to |
| 816 |
* be able to clear rte->inh later on when we discovered |
| 817 |
* that, but no more; we have to handle such cases as |
| 818 |
* full-fledged inheritance. |
| 819 |
*/ |
| 820 |
rte->inh = has_subclass(rte->relid); |
| 821 |
} |
| 822 |
break; |
| 823 |
case RTE_JOIN: |
| 824 |
root->hasJoinRTEs = true; |
| 825 |
if (IS_OUTER_JOIN(rte->jointype)) |
| 826 |
hasOuterJoins = true; |
| 827 |
break; |
| 828 |
case RTE_RESULT: |
| 829 |
hasResultRTEs = true; |
| 830 |
break; |
| 831 |
case RTE_GROUP: |
| 832 |
Assert(parse->hasGroupRTE); |
| 833 |
root->group_rtindex = list_cell_number(parse->rtable, l) + 1; |
| 834 |
break; |
| 835 |
default: |
| 836 |
/* No work here for other RTE types */ |
| 837 |
break; |
| 838 |
} |
| 839 |
|
| 840 |
if (rte->lateral) |
| 841 |
root->hasLateralRTEs = true; |
| 842 |
|
| 843 |
/* |
| 844 |
* We can also determine the maximum security level required for any |
| 845 |
* securityQuals now. Addition of inheritance-child RTEs won't affect |
| 846 |
* this, because child tables don't have their own securityQuals; see |
| 847 |
* expand_single_inheritance_child(). |
| 848 |
*/ |
| 849 |
if (rte->securityQuals) |
| 850 |
root->qual_security_level = Max(root->qual_security_level, |
| 851 |
list_length(rte->securityQuals)); |
| 852 |
} |
| 853 |
|
| 854 |
/* |
| 855 |
* If we have now verified that the query target relation is |
| 856 |
* non-inheriting, mark it as a leaf target. |
| 857 |
*/ |
| 858 |
if (parse->resultRelation) |
| 859 |
{ |
| 860 |
RangeTblEntry *rte = rt_fetch(parse->resultRelation, parse->rtable); |
| 861 |
|
| 862 |
if (!rte->inh) |
| 863 |
root->leaf_result_relids = |
| 864 |
bms_make_singleton(parse->resultRelation); |
| 865 |
} |
| 866 |
|
| 867 |
/* |
| 868 |
* This would be a convenient time to check access permissions for all |
| 869 |
* relations mentioned in the query, since it would be better to fail now, |
| 870 |
* before doing any detailed planning. However, for historical reasons, |
| 871 |
* we leave this to be done at executor startup. |
| 872 |
* |
| 873 |
* Note, however, that we do need to check access permissions for any view |
| 874 |
* relations mentioned in the query, in order to prevent information being |
| 875 |
* leaked by selectivity estimation functions, which only check view owner |
| 876 |
* permissions on underlying tables (see all_rows_selectable() and its |
| 877 |
* callers). This is a little ugly, because it means that access |
| 878 |
* permissions for views will be checked twice, which is another reason |
| 879 |
* why it would be better to do all the ACL checks here. |
| 880 |
*/ |
| 881 |
foreach(l, parse->rtable) |
| 882 |
{ |
| 883 |
RangeTblEntry *rte = lfirst_node(RangeTblEntry, l); |
| 884 |
|
| 885 |
if (rte->perminfoindex != 0 && |
| 886 |
rte->relkind == RELKIND_VIEW) |
| 887 |
{ |
| 888 |
RTEPermissionInfo *perminfo; |
| 889 |
bool result; |
| 890 |
|
| 891 |
perminfo = getRTEPermissionInfo(parse->rteperminfos, rte); |
| 892 |
result = ExecCheckOneRelPerms(perminfo); |
| 893 |
if (!result) |
| 894 |
aclcheck_error(ACLCHECK_NO_PRIV, OBJECT_VIEW, |
| 895 |
get_rel_name(perminfo->relid)); |
| 896 |
} |
| 897 |
} |
| 898 |
|
| 899 |
/* |
| 900 |
* Preprocess RowMark information. We need to do this after subquery |
| 901 |
* pullup, so that all base relations are present. |
| 902 |
*/ |
| 903 |
preprocess_rowmarks(root); |
| 904 |
|
| 905 |
/* |
| 906 |
* Set hasHavingQual to remember if HAVING clause is present. Needed |
| 907 |
* because preprocess_expression will reduce a constant-true condition to |
| 908 |
* an empty qual list ... but "HAVING TRUE" is not a semantic no-op. |
| 909 |
*/ |
| 910 |
root->hasHavingQual = (parse->havingQual != NULL); |
| 911 |
|
| 912 |
/* |
| 913 |
* Do expression preprocessing on targetlist and quals, as well as other |
| 914 |
* random expressions in the querytree. Note that we do not need to |
| 915 |
* handle sort/group expressions explicitly, because they are actually |
| 916 |
* part of the targetlist. |
| 917 |
*/ |
| 918 |
parse->targetList = (List *) |
| 919 |
preprocess_expression(root, (Node *) parse->targetList, |
| 920 |
EXPRKIND_TARGET); |
| 921 |
|
| 922 |
newWithCheckOptions = NIL; |
| 923 |
foreach(l, parse->withCheckOptions) |
| 924 |
{ |
| 925 |
WithCheckOption *wco = lfirst_node(WithCheckOption, l); |
| 926 |
|
| 927 |
wco->qual = preprocess_expression(root, wco->qual, |
| 928 |
EXPRKIND_QUAL); |
| 929 |
if (wco->qual != NULL) |
| 930 |
newWithCheckOptions = lappend(newWithCheckOptions, wco); |
| 931 |
} |
| 932 |
parse->withCheckOptions = newWithCheckOptions; |
| 933 |
|
| 934 |
parse->returningList = (List *) |
| 935 |
preprocess_expression(root, (Node *) parse->returningList, |
| 936 |
EXPRKIND_TARGET); |
| 937 |
|
| 938 |
preprocess_qual_conditions(root, (Node *) parse->jointree); |
| 939 |
|
| 940 |
parse->havingQual = preprocess_expression(root, parse->havingQual, |
| 941 |
EXPRKIND_QUAL); |
| 942 |
|
| 943 |
foreach(l, parse->windowClause) |
| 944 |
{ |
| 945 |
WindowClause *wc = lfirst_node(WindowClause, l); |
| 946 |
|
| 947 |
/* partitionClause/orderClause are sort/group expressions */ |
| 948 |
wc->startOffset = preprocess_expression(root, wc->startOffset, |
| 949 |
EXPRKIND_LIMIT); |
| 950 |
wc->endOffset = preprocess_expression(root, wc->endOffset, |
| 951 |
EXPRKIND_LIMIT); |
| 952 |
} |
| 953 |
|
| 954 |
parse->limitOffset = preprocess_expression(root, parse->limitOffset, |
| 955 |
EXPRKIND_LIMIT); |
| 956 |
parse->limitCount = preprocess_expression(root, parse->limitCount, |
| 957 |
EXPRKIND_LIMIT); |
| 958 |
|
| 959 |
if (parse->onConflict) |
| 960 |
{ |
| 961 |
parse->onConflict->arbiterElems = (List *) |
| 962 |
preprocess_expression(root, |
| 963 |
(Node *) parse->onConflict->arbiterElems, |
| 964 |
EXPRKIND_ARBITER_ELEM); |
| 965 |
parse->onConflict->arbiterWhere = |
| 966 |
preprocess_expression(root, |
| 967 |
parse->onConflict->arbiterWhere, |
| 968 |
EXPRKIND_QUAL); |
| 969 |
parse->onConflict->onConflictSet = (List *) |
| 970 |
preprocess_expression(root, |
| 971 |
(Node *) parse->onConflict->onConflictSet, |
| 972 |
EXPRKIND_TARGET); |
| 973 |
parse->onConflict->onConflictWhere = |
| 974 |
preprocess_expression(root, |
| 975 |
parse->onConflict->onConflictWhere, |
| 976 |
EXPRKIND_QUAL); |
| 977 |
/* exclRelTlist contains only Vars, so no preprocessing needed */ |
| 978 |
} |
| 979 |
|
| 980 |
foreach(l, parse->mergeActionList) |
| 981 |
{ |
| 982 |
MergeAction *action = (MergeAction *) lfirst(l); |
| 983 |
|
| 984 |
action->targetList = (List *) |
| 985 |
preprocess_expression(root, |
| 986 |
(Node *) action->targetList, |
| 987 |
EXPRKIND_TARGET); |
| 988 |
action->qual = |
| 989 |
preprocess_expression(root, |
| 990 |
(Node *) action->qual, |
| 991 |
EXPRKIND_QUAL); |
| 992 |
} |
| 993 |
|
| 994 |
parse->mergeJoinCondition = |
| 995 |
preprocess_expression(root, parse->mergeJoinCondition, EXPRKIND_QUAL); |
| 996 |
|
| 997 |
root->append_rel_list = (List *) |
| 998 |
preprocess_expression(root, (Node *) root->append_rel_list, |
| 999 |
EXPRKIND_APPINFO); |
| 1000 |
|
| 1001 |
/* Also need to preprocess expressions within RTEs */ |
| 1002 |
foreach(l, parse->rtable) |
| 1003 |
{ |
| 1004 |
RangeTblEntry *rte = lfirst_node(RangeTblEntry, l); |
| 1005 |
int kind; |
| 1006 |
ListCell *lcsq; |
| 1007 |
|
| 1008 |
if (rte->rtekind == RTE_RELATION) |
| 1009 |
{ |
| 1010 |
if (rte->tablesample) |
| 1011 |
rte->tablesample = (TableSampleClause *) |
| 1012 |
preprocess_expression(root, |
| 1013 |
(Node *) rte->tablesample, |
| 1014 |
EXPRKIND_TABLESAMPLE); |
| 1015 |
} |
| 1016 |
else if (rte->rtekind == RTE_SUBQUERY) |
| 1017 |
{ |
| 1018 |
/* |
| 1019 |
* We don't want to do all preprocessing yet on the subquery's |
| 1020 |
* expressions, since that will happen when we plan it. But if it |
| 1021 |
* contains any join aliases of our level, those have to get |
| 1022 |
* expanded now, because planning of the subquery won't do it. |
| 1023 |
* That's only possible if the subquery is LATERAL. |
| 1024 |
*/ |
| 1025 |
if (rte->lateral && root->hasJoinRTEs) |
| 1026 |
rte->subquery = (Query *) |
| 1027 |
flatten_join_alias_vars(root, root->parse, |
| 1028 |
(Node *) rte->subquery); |
| 1029 |
} |
| 1030 |
else if (rte->rtekind == RTE_FUNCTION) |
| 1031 |
{ |
| 1032 |
/* Preprocess the function expression(s) fully */ |
| 1033 |
kind = rte->lateral ? EXPRKIND_RTFUNC_LATERAL : EXPRKIND_RTFUNC; |
| 1034 |
rte->functions = (List *) |
| 1035 |
preprocess_expression(root, (Node *) rte->functions, kind); |
| 1036 |
} |
| 1037 |
else if (rte->rtekind == RTE_TABLEFUNC) |
| 1038 |
{ |
| 1039 |
/* Preprocess the function expression(s) fully */ |
| 1040 |
kind = rte->lateral ? EXPRKIND_TABLEFUNC_LATERAL : EXPRKIND_TABLEFUNC; |
| 1041 |
rte->tablefunc = (TableFunc *) |
| 1042 |
preprocess_expression(root, (Node *) rte->tablefunc, kind); |
| 1043 |
} |
| 1044 |
else if (rte->rtekind == RTE_VALUES) |
| 1045 |
{ |
| 1046 |
/* Preprocess the values lists fully */ |
| 1047 |
kind = rte->lateral ? EXPRKIND_VALUES_LATERAL : EXPRKIND_VALUES; |
| 1048 |
rte->values_lists = (List *) |
| 1049 |
preprocess_expression(root, (Node *) rte->values_lists, kind); |
| 1050 |
} |
| 1051 |
else if (rte->rtekind == RTE_GROUP) |
| 1052 |
{ |
| 1053 |
/* Preprocess the groupexprs list fully */ |
| 1054 |
rte->groupexprs = (List *) |
| 1055 |
preprocess_expression(root, (Node *) rte->groupexprs, |
| 1056 |
EXPRKIND_GROUPEXPR); |
| 1057 |
} |
| 1058 |
|
| 1059 |
/* |
| 1060 |
* Process each element of the securityQuals list as if it were a |
| 1061 |
* separate qual expression (as indeed it is). We need to do it this |
| 1062 |
* way to get proper canonicalization of AND/OR structure. Note that |
| 1063 |
* this converts each element into an implicit-AND sublist. |
| 1064 |
*/ |
| 1065 |
foreach(lcsq, rte->securityQuals) |
| 1066 |
{ |
| 1067 |
lfirst(lcsq) = preprocess_expression(root, |
| 1068 |
(Node *) lfirst(lcsq), |
| 1069 |
EXPRKIND_QUAL); |
| 1070 |
} |
| 1071 |
} |
| 1072 |
|
| 1073 |
/* |
| 1074 |
* Now that we are done preprocessing expressions, and in particular done |
| 1075 |
* flattening join alias variables, get rid of the joinaliasvars lists. |
| 1076 |
* They no longer match what expressions in the rest of the tree look |
| 1077 |
* like, because we have not preprocessed expressions in those lists (and |
| 1078 |
* do not want to; for example, expanding a SubLink there would result in |
| 1079 |
* a useless unreferenced subplan). Leaving them in place simply creates |
| 1080 |
* a hazard for later scans of the tree. We could try to prevent that by |
| 1081 |
* using QTW_IGNORE_JOINALIASES in every tree scan done after this point, |
| 1082 |
* but that doesn't sound very reliable. |
| 1083 |
*/ |
| 1084 |
if (root->hasJoinRTEs) |
| 1085 |
{ |
| 1086 |
foreach(l, parse->rtable) |
| 1087 |
{ |
| 1088 |
RangeTblEntry *rte = lfirst_node(RangeTblEntry, l); |
| 1089 |
|
| 1090 |
rte->joinaliasvars = NIL; |
| 1091 |
} |
| 1092 |
} |
| 1093 |
|
| 1094 |
/* |
| 1095 |
* Before we flatten GROUP Vars, identify HAVING clauses whose equality |
| 1096 |
* semantics disagree with the GROUP BY's. See find_having_conflicts. |
| 1097 |
*/ |
| 1098 |
if (parse->hasGroupRTE) |
| 1099 |
havingPushdownConflicts = find_having_conflicts(parse, |
| 1100 |
root->group_rtindex); |
| 1101 |
else |
| 1102 |
havingPushdownConflicts = NULL; |
| 1103 |
|
| 1104 |
/* |
| 1105 |
* Replace any Vars in the subquery's targetlist and havingQual that |
| 1106 |
* reference GROUP outputs with the underlying grouping expressions. |
| 1107 |
* |
| 1108 |
* Note that we need to perform this replacement after we've preprocessed |
| 1109 |
* the grouping expressions. This is to ensure that there is only one |
| 1110 |
* instance of SubPlan for each SubLink contained within the grouping |
| 1111 |
* expressions. |
| 1112 |
*/ |
| 1113 |
if (parse->hasGroupRTE) |
| 1114 |
{ |
| 1115 |
parse->targetList = (List *) |
| 1116 |
flatten_group_exprs(root, root->parse, (Node *) parse->targetList); |
| 1117 |
parse->havingQual = |
| 1118 |
flatten_group_exprs(root, root->parse, parse->havingQual); |
| 1119 |
} |
| 1120 |
|
| 1121 |
/* Constant-folding might have removed all set-returning functions */ |
| 1122 |
if (parse->hasTargetSRFs) |
| 1123 |
parse->hasTargetSRFs = expression_returns_set((Node *) parse->targetList); |
| 1124 |
|
| 1125 |
/* |
| 1126 |
* If we have grouping sets, expand the groupingSets tree of this query to |
| 1127 |
* a flat list of grouping sets. We need to do this before optimizing |
| 1128 |
* HAVING, since we can't easily tell if there's an empty grouping set |
| 1129 |
* until we have this representation. |
| 1130 |
*/ |
| 1131 |
if (parse->groupingSets) |
| 1132 |
{ |
| 1133 |
parse->groupingSets = |
| 1134 |
expand_grouping_sets(parse->groupingSets, parse->groupDistinct, -1); |
| 1135 |
} |
| 1136 |
|
| 1137 |
/* |
| 1138 |
* In some cases we may want to transfer a HAVING clause into WHERE. We |
| 1139 |
* cannot do so if the HAVING clause contains aggregates (obviously) or |
| 1140 |
* volatile functions (since a HAVING clause is supposed to be executed |
| 1141 |
* only once per group). We also can't do this if there are any grouping |
| 1142 |
* sets and the clause references any columns that are nullable by the |
| 1143 |
* grouping sets; the nulled values of those columns are not available |
| 1144 |
* before the grouping step. (The test on groupClause might seem wrong, |
| 1145 |
* but it's okay: it's just an optimization to avoid running pull_varnos |
| 1146 |
* when there cannot be any Vars in the HAVING clause.) |
| 1147 |
* |
| 1148 |
* We also cannot do this for HAVING clauses that conflict with GROUP BY |
| 1149 |
* on collation or operator family. Both kinds of conflict are detected |
| 1150 |
* before flatten_group_exprs (see find_having_conflicts above) and |
| 1151 |
* recorded in the havingPushdownConflicts bitmapset. The bitmapset |
| 1152 |
* indexes remain valid here because flatten_group_exprs uses |
| 1153 |
* expression_tree_mutator, which preserves the list length and ordering |
| 1154 |
* of havingQual. |
| 1155 |
* |
| 1156 |
* Also, it may be that the clause is so expensive to execute that we're |
| 1157 |
* better off doing it only once per group, despite the loss of |
| 1158 |
* selectivity. This is hard to estimate short of doing the entire |
| 1159 |
* planning process twice, so we use a heuristic: clauses containing |
| 1160 |
* subplans are left in HAVING. Otherwise, we move or copy the HAVING |
| 1161 |
* clause into WHERE, in hopes of eliminating tuples before aggregation |
| 1162 |
* instead of after. |
| 1163 |
* |
| 1164 |
* If the query has no empty grouping set then we can simply move such a |
| 1165 |
* clause into WHERE; any group that fails the clause will not be in the |
| 1166 |
* output because none of its tuples will reach the grouping or |
| 1167 |
* aggregation stage. Otherwise we have to keep the clause in HAVING to |
| 1168 |
* ensure that we don't emit a bogus aggregated row. But then the HAVING |
| 1169 |
* clause must be degenerate (variable-free), so we can copy it into WHERE |
| 1170 |
* so that query_planner() can use it in a gating Result node. (This could |
| 1171 |
* be done better, but it seems not worth optimizing.) |
| 1172 |
* |
| 1173 |
* Note that a HAVING clause may contain expressions that are not fully |
| 1174 |
* preprocessed. This can happen if these expressions are part of |
| 1175 |
* grouping items. In such cases, they are replaced with GROUP Vars in |
| 1176 |
* the parser and then replaced back after we're done with expression |
| 1177 |
* preprocessing on havingQual. This is not an issue if the clause |
| 1178 |
* remains in HAVING, because these expressions will be matched to lower |
| 1179 |
* target items in setrefs.c. However, if the clause is moved or copied |
| 1180 |
* into WHERE, we need to ensure that these expressions are fully |
| 1181 |
* preprocessed. |
| 1182 |
* |
| 1183 |
* Note that both havingQual and parse->jointree->quals are in |
| 1184 |
* implicitly-ANDed-list form at this point, even though they are declared |
| 1185 |
* as Node *. |
| 1186 |
*/ |
| 1187 |
newHaving = NIL; |
| 1188 |
havingIdx = 0; |
| 1189 |
foreach(l, (List *) parse->havingQual) |
| 1190 |
{ |
| 1191 |
Node *havingclause = (Node *) lfirst(l); |
| 1192 |
|
| 1193 |
if (contain_agg_clause(havingclause) || |
| 1194 |
contain_volatile_functions(havingclause) || |
| 1195 |
contain_subplans(havingclause) || |
| 1196 |
bms_is_member(havingIdx, havingPushdownConflicts) || |
| 1197 |
(parse->groupClause && parse->groupingSets && |
| 1198 |
bms_is_member(root->group_rtindex, pull_varnos(root, havingclause)))) |
| 1199 |
{ |
| 1200 |
/* keep it in HAVING */ |
| 1201 |
newHaving = lappend(newHaving, havingclause); |
| 1202 |
} |
| 1203 |
else if (parse->groupClause && |
| 1204 |
(parse->groupingSets == NIL || |
| 1205 |
(List *) linitial(parse->groupingSets) != NIL)) |
| 1206 |
{ |
| 1207 |
/* There is GROUP BY, but no empty grouping set */ |
| 1208 |
Node *whereclause; |
| 1209 |
|
| 1210 |
/* Preprocess the HAVING clause fully */ |
| 1211 |
whereclause = preprocess_expression(root, havingclause, |
| 1212 |
EXPRKIND_QUAL); |
| 1213 |
/* ... and move it to WHERE */ |
| 1214 |
parse->jointree->quals = (Node *) |
| 1215 |
list_concat((List *) parse->jointree->quals, |
| 1216 |
(List *) whereclause); |
| 1217 |
} |
| 1218 |
else |
| 1219 |
{ |
| 1220 |
/* There is an empty grouping set (perhaps implicitly) */ |
| 1221 |
Node *whereclause; |
| 1222 |
|
| 1223 |
/* Preprocess the HAVING clause fully */ |
| 1224 |
whereclause = preprocess_expression(root, copyObject(havingclause), |
| 1225 |
EXPRKIND_QUAL); |
| 1226 |
/* ... and put a copy in WHERE */ |
| 1227 |
parse->jointree->quals = (Node *) |
| 1228 |
list_concat((List *) parse->jointree->quals, |
| 1229 |
(List *) whereclause); |
| 1230 |
/* ... and also keep it in HAVING */ |
| 1231 |
newHaving = lappend(newHaving, havingclause); |
| 1232 |
} |
| 1233 |
|
| 1234 |
havingIdx++; |
| 1235 |
} |
| 1236 |
parse->havingQual = (Node *) newHaving; |
| 1237 |
|
| 1238 |
/* |
| 1239 |
* If we have any outer joins, try to reduce them to plain inner joins. |
| 1240 |
* This step is most easily done after we've done expression |
| 1241 |
* preprocessing. |
| 1242 |
*/ |
| 1243 |
if (hasOuterJoins) |
| 1244 |
reduce_outer_joins(root); |
| 1245 |
|
| 1246 |
/* |
| 1247 |
* If we have any RTE_RESULT relations, see if they can be deleted from |
| 1248 |
* the jointree. We also rely on this processing to flatten single-child |
| 1249 |
* FromExprs underneath outer joins. This step is most effectively done |
| 1250 |
* after we've done expression preprocessing and outer join reduction. |
| 1251 |
*/ |
| 1252 |
if (hasResultRTEs || hasOuterJoins) |
| 1253 |
remove_useless_result_rtes(root); |
| 1254 |
|
| 1255 |
/* |
| 1256 |
* Do the main planning. |
| 1257 |
*/ |
| 1258 |
grouping_planner(root, tuple_fraction, setops); |
| 1259 |
|
| 1260 |
/* |
| 1261 |
* Capture the set of outer-level param IDs we have access to, for use in |
| 1262 |
* extParam/allParam calculations later. |
| 1263 |
*/ |
| 1264 |
SS_identify_outer_params(root); |
| 1265 |
|
| 1266 |
/* |
| 1267 |
* If any initPlans were created in this query level, adjust the surviving |
| 1268 |
* Paths' costs and parallel-safety flags to account for them. The |
| 1269 |
* initPlans won't actually get attached to the plan tree till |
| 1270 |
* create_plan() runs, but we must include their effects now. |
| 1271 |
*/ |
| 1272 |
final_rel = fetch_upper_rel(root, UPPERREL_FINAL, NULL); |
| 1273 |
SS_charge_for_initplans(root, final_rel); |
| 1274 |
|
| 1275 |
/* |
| 1276 |
* Make sure we've identified the cheapest Path for the final rel. (By |
| 1277 |
* doing this here not in grouping_planner, we include initPlan costs in |
| 1278 |
* the decision, though it's unlikely that will change anything.) |
| 1279 |
*/ |
| 1280 |
set_cheapest(final_rel); |
| 1281 |
|
| 1282 |
return root; |
| 1283 |
} |
| 1284 |
|
| 1285 |
/* |
| 1286 |
* preprocess_expression |
| 1287 |
* Do subquery_planner's preprocessing work for an expression, |
| 1288 |
* which can be a targetlist, a WHERE clause (including JOIN/ON |
| 1289 |
* conditions), a HAVING clause, or a few other things. |
| 1290 |
*/ |
| 1291 |
static Node * |
| 1292 |
preprocess_expression(PlannerInfo *root, Node *expr, int kind) |
| 1293 |
{ |
| 1294 |
/* |
| 1295 |
* Fall out quickly if expression is empty. This occurs often enough to |
| 1296 |
* be worth checking. Note that null->null is the correct conversion for |
| 1297 |
* implicit-AND result format, too. |
| 1298 |
*/ |
| 1299 |
if (expr == NULL) |
| 1300 |
return NULL; |
| 1301 |
|
| 1302 |
/* |
| 1303 |
* If the query has any join RTEs, replace join alias variables with |
| 1304 |
* base-relation variables. We must do this first, since any expressions |
| 1305 |
* we may extract from the joinaliasvars lists have not been preprocessed. |
| 1306 |
* For example, if we did this after sublink processing, sublinks expanded |
| 1307 |
* out from join aliases would not get processed. But we can skip this in |
| 1308 |
* non-lateral RTE functions, VALUES lists, and TABLESAMPLE clauses, since |
| 1309 |
* they can't contain any Vars of the current query level. |
| 1310 |
*/ |
| 1311 |
if (root->hasJoinRTEs && |
| 1312 |
!(kind == EXPRKIND_RTFUNC || |
| 1313 |
kind == EXPRKIND_VALUES || |
| 1314 |
kind == EXPRKIND_TABLESAMPLE || |
| 1315 |
kind == EXPRKIND_TABLEFUNC)) |
| 1316 |
expr = flatten_join_alias_vars(root, root->parse, expr); |
| 1317 |
|
| 1318 |
/* |
| 1319 |
* Simplify constant expressions. For function RTEs, this was already |
| 1320 |
* done by preprocess_function_rtes. (But note we must do it again for |
| 1321 |
* EXPRKIND_RTFUNC_LATERAL, because those might by now contain |
| 1322 |
* un-simplified subexpressions inserted by flattening of subqueries or |
| 1323 |
* join alias variables.) |
| 1324 |
* |
| 1325 |
* Note: an essential effect of this is to convert named-argument function |
| 1326 |
* calls to positional notation and insert the current actual values of |
| 1327 |
* any default arguments for functions. To ensure that happens, we *must* |
| 1328 |
* process all expressions here. Previous PG versions sometimes skipped |
| 1329 |
* const-simplification if it didn't seem worth the trouble, but we can't |
| 1330 |
* do that anymore. |
| 1331 |
* |
| 1332 |
* Note: this also flattens nested AND and OR expressions into N-argument |
| 1333 |
* form. All processing of a qual expression after this point must be |
| 1334 |
* careful to maintain AND/OR flatness --- that is, do not generate a tree |
| 1335 |
* with AND directly under AND, nor OR directly under OR. |
| 1336 |
*/ |
| 1337 |
if (kind != EXPRKIND_RTFUNC) |
| 1338 |
expr = eval_const_expressions(root, expr); |
| 1339 |
|
| 1340 |
/* |
| 1341 |
* If it's a qual or havingQual, canonicalize it. |
| 1342 |
*/ |
| 1343 |
if (kind == EXPRKIND_QUAL) |
| 1344 |
{ |
| 1345 |
expr = (Node *) canonicalize_qual((Expr *) expr, false); |
| 1346 |
|
| 1347 |
#ifdef OPTIMIZER_DEBUG |
| 1348 |
printf("After canonicalize_qual()\n"); |
| 1349 |
pprint(expr); |
| 1350 |
#endif |
| 1351 |
} |
| 1352 |
|
| 1353 |
/* |
| 1354 |
* Check for ANY ScalarArrayOpExpr with Const arrays and set the |
| 1355 |
* hashfuncid of any that might execute more quickly by using hash lookups |
| 1356 |
* instead of a linear search. |
| 1357 |
*/ |
| 1358 |
if (kind == EXPRKIND_QUAL || kind == EXPRKIND_TARGET) |
| 1359 |
{ |
| 1360 |
convert_saop_to_hashed_saop(expr); |
| 1361 |
} |
| 1362 |
|
| 1363 |
/* Expand SubLinks to SubPlans */ |
| 1364 |
if (root->parse->hasSubLinks) |
| 1365 |
expr = SS_process_sublinks(root, expr, (kind == EXPRKIND_QUAL)); |
| 1366 |
|
| 1367 |
/* |
| 1368 |
* XXX do not insert anything here unless you have grokked the comments in |
| 1369 |
* SS_replace_correlation_vars ... |
| 1370 |
*/ |
| 1371 |
|
| 1372 |
/* Replace uplevel vars with Param nodes (this IS possible in VALUES) */ |
| 1373 |
if (root->query_level > 1) |
| 1374 |
expr = SS_replace_correlation_vars(root, expr); |
| 1375 |
|
| 1376 |
/* |
| 1377 |
* If it's a qual or havingQual, convert it to implicit-AND format. (We |
| 1378 |
* don't want to do this before eval_const_expressions, since the latter |
| 1379 |
* would be unable to simplify a top-level AND correctly. Also, |
| 1380 |
* SS_process_sublinks expects explicit-AND format.) |
| 1381 |
*/ |
| 1382 |
if (kind == EXPRKIND_QUAL) |
| 1383 |
expr = (Node *) make_ands_implicit((Expr *) expr); |
| 1384 |
|
| 1385 |
return expr; |
| 1386 |
} |
| 1387 |
|
| 1388 |
/* |
| 1389 |
* preprocess_qual_conditions |
| 1390 |
* Recursively scan the query's jointree and do subquery_planner's |
| 1391 |
* preprocessing work on each qual condition found therein. |
| 1392 |
*/ |
| 1393 |
static void |
| 1394 |
preprocess_qual_conditions(PlannerInfo *root, Node *jtnode) |
| 1395 |
{ |
| 1396 |
if (jtnode == NULL) |
| 1397 |
return; |
| 1398 |
if (IsA(jtnode, RangeTblRef)) |
| 1399 |
{ |
| 1400 |
/* nothing to do here */ |
| 1401 |
} |
| 1402 |
else if (IsA(jtnode, FromExpr)) |
| 1403 |
{ |
| 1404 |
FromExpr *f = (FromExpr *) jtnode; |
| 1405 |
ListCell *l; |
| 1406 |
|
| 1407 |
foreach(l, f->fromlist) |
| 1408 |
preprocess_qual_conditions(root, lfirst(l)); |
| 1409 |
|
| 1410 |
f->quals = preprocess_expression(root, f->quals, EXPRKIND_QUAL); |
| 1411 |
} |
| 1412 |
else if (IsA(jtnode, JoinExpr)) |
| 1413 |
{ |
| 1414 |
JoinExpr *j = (JoinExpr *) jtnode; |
| 1415 |
|
| 1416 |
preprocess_qual_conditions(root, j->larg); |
| 1417 |
preprocess_qual_conditions(root, j->rarg); |
| 1418 |
|
| 1419 |
j->quals = preprocess_expression(root, j->quals, EXPRKIND_QUAL); |
| 1420 |
} |
| 1421 |
else |
| 1422 |
elog(ERROR, "unrecognized node type: %d", |
| 1423 |
(int) nodeTag(jtnode)); |
| 1424 |
} |
| 1425 |
|
| 1426 |
/* |
| 1427 |
* find_having_conflicts |
| 1428 |
* Identify HAVING clauses that must not be moved to WHERE because they |
| 1429 |
* apply a different equivalence relation than GROUP BY. Pushing such a |
| 1430 |
* clause to WHERE would filter individual rows before grouping happens, |
| 1431 |
* eliminating rows that GROUP BY would have merged into a single group |
| 1432 |
* and thereby changing aggregate results. |
| 1433 |
* |
| 1434 |
* The actual walking is done by expression_has_grouping_conflict; see that |
| 1435 |
* function for the kinds of conflict it looks for. We just iterate over |
| 1436 |
* havingQual and supply a HAVING-specific callback that identifies GROUP |
| 1437 |
* Vars. |
| 1438 |
* |
| 1439 |
* This must be called before flatten_group_exprs, while the HAVING clause |
| 1440 |
* still contains GROUP Vars (Vars referencing RTE_GROUP). These GROUP Vars |
| 1441 |
* carry the GROUP BY collation as their varcollid and let us recover the |
| 1442 |
* grouping eqop via varattno. After flattening, those Vars are replaced by |
| 1443 |
* the underlying expressions, and matching back to grouping expressions is |
| 1444 |
* much harder. |
| 1445 |
* |
| 1446 |
* Returns a Bitmapset of zero-based indexes into the havingQual list for |
| 1447 |
* clauses that conflict and must stay in HAVING. |
| 1448 |
*/ |
| 1449 |
static Bitmapset * |
| 1450 |
find_having_conflicts(Query *parse, Index group_rtindex) |
| 1451 |
{ |
| 1452 |
Bitmapset *result = NULL; |
| 1453 |
having_grouping_ctx ctx; |
| 1454 |
int idx; |
| 1455 |
|
| 1456 |
if (parse->havingQual == NULL) |
| 1457 |
return NULL; |
| 1458 |
|
| 1459 |
ctx.parse = parse; |
| 1460 |
ctx.group_rtindex = group_rtindex; |
| 1461 |
|
| 1462 |
idx = 0; |
| 1463 |
foreach_ptr(Node, clause, (List *) parse->havingQual) |
| 1464 |
{ |
| 1465 |
if (expression_has_grouping_conflict(clause, having_var_grouping_eqop, |
| 1466 |
&ctx)) |
| 1467 |
result = bms_add_member(result, idx); |
| 1468 |
idx++; |
| 1469 |
} |
| 1470 |
|
| 1471 |
return result; |
| 1472 |
} |
| 1473 |
|
| 1474 |
/* |
| 1475 |
* having_var_grouping_eqop |
| 1476 |
* grouping_eqop_callback for find_having_conflicts. |
| 1477 |
* |
| 1478 |
* Returns the GROUP BY equality operator for 'var' if it references the |
| 1479 |
* query's RTE_GROUP, or InvalidOid otherwise. |
| 1480 |
*/ |
| 1481 |
static Oid |
| 1482 |
having_var_grouping_eqop(Var *var, void *context) |
| 1483 |
{ |
| 1484 |
having_grouping_ctx *ctx = (having_grouping_ctx *) context; |
| 1485 |
|
| 1486 |
if (var->varno != ctx->group_rtindex || var->varlevelsup != 0) |
| 1487 |
return InvalidOid; |
| 1488 |
|
| 1489 |
return group_var_eqop(ctx->parse, var); |
| 1490 |
} |
| 1491 |
|
| 1492 |
/* |
| 1493 |
* group_var_eqop |
| 1494 |
* Return the equality operator that GROUP BY uses for the given GROUP Var. |
| 1495 |
* |
| 1496 |
* A GROUP Var's varattno is its 1-based position in the RTE_GROUP's groupexprs |
| 1497 |
* list, which addRangeTableEntryForGroup built by iterating parse->groupClause |
| 1498 |
* and including every SortGroupClause whose TLE was present in the targetlist. |
| 1499 |
* Replay that traversal here to recover the SortGroupClause for the given |
| 1500 |
* varattno. |
| 1501 |
*/ |
| 1502 |
static Oid |
| 1503 |
group_var_eqop(Query *parse, Var *var) |
| 1504 |
{ |
| 1505 |
int counter = 0; |
| 1506 |
|
| 1507 |
Assert(var->varlevelsup == 0); |
| 1508 |
|
| 1509 |
foreach_node(SortGroupClause, sgc, parse->groupClause) |
| 1510 |
{ |
| 1511 |
if (get_sortgroupclause_tle(sgc, parse->targetList) == NULL) |
| 1512 |
continue; |
| 1513 |
if (++counter == var->varattno) |
| 1514 |
return sgc->eqop; |
| 1515 |
} |
| 1516 |
|
| 1517 |
elog(ERROR, "could not find GROUP clause for GROUP Var attno %d", |
| 1518 |
var->varattno); |
| 1519 |
return InvalidOid; /* keep compiler quiet */ |
| 1520 |
} |
| 1521 |
|
| 1522 |
/* |
| 1523 |
* preprocess_phv_expression |
| 1524 |
* Do preprocessing on a PlaceHolderVar expression that's been pulled up. |
| 1525 |
* |
| 1526 |
* If a LATERAL subquery references an output of another subquery, and that |
| 1527 |
* output must be wrapped in a PlaceHolderVar because of an intermediate outer |
| 1528 |
* join, then we'll push the PlaceHolderVar expression down into the subquery |
| 1529 |
* and later pull it back up during find_lateral_references, which runs after |
| 1530 |
* subquery_planner has preprocessed all the expressions that were in the |
| 1531 |
* current query level to start with. So we need to preprocess it then. |
| 1532 |
*/ |
| 1533 |
Expr * |
| 1534 |
preprocess_phv_expression(PlannerInfo *root, Expr *expr) |
| 1535 |
{ |
| 1536 |
return (Expr *) preprocess_expression(root, (Node *) expr, EXPRKIND_PHV); |
| 1537 |
} |
| 1538 |
|
| 1539 |
/*-------------------- |
| 1540 |
* grouping_planner |
| 1541 |
* Perform planning steps related to grouping, aggregation, etc. |
| 1542 |
* |
| 1543 |
* This function adds all required top-level processing to the scan/join |
| 1544 |
* Path(s) produced by query_planner. |
| 1545 |
* |
| 1546 |
* tuple_fraction is the fraction of tuples we expect will be retrieved. |
| 1547 |
* tuple_fraction is interpreted as follows: |
| 1548 |
* 0: expect all tuples to be retrieved (normal case) |
| 1549 |
* 0 < tuple_fraction < 1: expect the given fraction of tuples available |
| 1550 |
* from the plan to be retrieved |
| 1551 |
* tuple_fraction >= 1: tuple_fraction is the absolute number of tuples |
| 1552 |
* expected to be retrieved (ie, a LIMIT specification). |
| 1553 |
* setops is used for set operation subqueries to provide the subquery with |
| 1554 |
* the context in which it's being used so that Paths correctly sorted for the |
| 1555 |
* set operation can be generated. NULL when not planning a set operation |
| 1556 |
* child, or when a child of a set op that isn't interested in sorted input. |
| 1557 |
* |
| 1558 |
* Returns nothing; the useful output is in the Paths we attach to the |
| 1559 |
* (UPPERREL_FINAL, NULL) upperrel in *root. In addition, |
| 1560 |
* root->processed_tlist contains the final processed targetlist. |
| 1561 |
* |
| 1562 |
* Note that we have not done set_cheapest() on the final rel; it's convenient |
| 1563 |
* to leave this to the caller. |
| 1564 |
*-------------------- |
| 1565 |
*/ |
| 1566 |
static void |
| 1567 |
grouping_planner(PlannerInfo *root, double tuple_fraction, |
| 1568 |
SetOperationStmt *setops) |
| 1569 |
{ |
| 1570 |
Query *parse = root->parse; |
| 1571 |
int64 offset_est = 0; |
| 1572 |
int64 count_est = 0; |
| 1573 |
double limit_tuples = -1.0; |
| 1574 |
bool have_postponed_srfs = false; |
| 1575 |
PathTarget *final_target; |
| 1576 |
List *final_targets; |
| 1577 |
List *final_targets_contain_srfs; |
| 1578 |
bool final_target_parallel_safe; |
| 1579 |
RelOptInfo *current_rel; |
| 1580 |
RelOptInfo *final_rel; |
| 1581 |
FinalPathExtraData extra; |
| 1582 |
ListCell *lc; |
| 1583 |
|
| 1584 |
/* Tweak caller-supplied tuple_fraction if have LIMIT/OFFSET */ |
| 1585 |
if (parse->limitCount || parse->limitOffset) |
| 1586 |
{ |
| 1587 |
tuple_fraction = preprocess_limit(root, tuple_fraction, |
| 1588 |
&offset_est, &count_est); |
| 1589 |
|
| 1590 |
/* |
| 1591 |
* If we have a known LIMIT, and don't have an unknown OFFSET, we can |
| 1592 |
* estimate the effects of using a bounded sort. |
| 1593 |
*/ |
| 1594 |
if (count_est > 0 && offset_est >= 0) |
| 1595 |
limit_tuples = (double) count_est + (double) offset_est; |
| 1596 |
} |
| 1597 |
|
| 1598 |
/* Make tuple_fraction accessible to lower-level routines */ |
| 1599 |
root->tuple_fraction = tuple_fraction; |
| 1600 |
|
| 1601 |
if (parse->setOperations) |
| 1602 |
{ |
| 1603 |
/* |
| 1604 |
* Construct Paths for set operations. The results will not need any |
| 1605 |
* work except perhaps a top-level sort and/or LIMIT. Note that any |
| 1606 |
* special work for recursive unions is the responsibility of |
| 1607 |
* plan_set_operations. |
| 1608 |
*/ |
| 1609 |
current_rel = plan_set_operations(root); |
| 1610 |
|
| 1611 |
/* |
| 1612 |
* We should not need to call preprocess_targetlist, since we must be |
| 1613 |
* in a SELECT query node. Instead, use the processed_tlist returned |
| 1614 |
* by plan_set_operations (since this tells whether it returned any |
| 1615 |
* resjunk columns!), and transfer any sort key information from the |
| 1616 |
* original tlist. |
| 1617 |
*/ |
| 1618 |
Assert(parse->commandType == CMD_SELECT); |
| 1619 |
|
| 1620 |
/* for safety, copy processed_tlist instead of modifying in-place */ |
| 1621 |
root->processed_tlist = |
| 1622 |
postprocess_setop_tlist(copyObject(root->processed_tlist), |
| 1623 |
parse->targetList); |
| 1624 |
|
| 1625 |
/* Also extract the PathTarget form of the setop result tlist */ |
| 1626 |
final_target = current_rel->cheapest_total_path->pathtarget; |
| 1627 |
|
| 1628 |
/* And check whether it's parallel safe */ |
| 1629 |
final_target_parallel_safe = |
| 1630 |
is_parallel_safe(root, (Node *) final_target->exprs); |
| 1631 |
|
| 1632 |
/* The setop result tlist couldn't contain any SRFs */ |
| 1633 |
Assert(!parse->hasTargetSRFs); |
| 1634 |
final_targets = final_targets_contain_srfs = NIL; |
| 1635 |
|
| 1636 |
/* |
| 1637 |
* Can't handle FOR [KEY] UPDATE/SHARE here (parser should have |
| 1638 |
* checked already, but let's make sure). |
| 1639 |
*/ |
| 1640 |
if (parse->rowMarks) |
| 1641 |
ereport(ERROR, |
| 1642 |
(errcode(ERRCODE_FEATURE_NOT_SUPPORTED), |
| 1643 |
/*------ |
| 1644 |
translator: %s is a SQL row locking clause such as FOR UPDATE */ |
| 1645 |
errmsg("%s is not allowed with UNION/INTERSECT/EXCEPT", |
| 1646 |
LCS_asString(linitial_node(RowMarkClause, |
| 1647 |
parse->rowMarks)->strength)))); |
| 1648 |
|
| 1649 |
/* |
| 1650 |
* Calculate pathkeys that represent result ordering requirements |
| 1651 |
*/ |
| 1652 |
Assert(parse->distinctClause == NIL); |
| 1653 |
root->sort_pathkeys = make_pathkeys_for_sortclauses(root, |
| 1654 |
parse->sortClause, |
| 1655 |
root->processed_tlist); |
| 1656 |
} |
| 1657 |
else |
| 1658 |
{ |
| 1659 |
/* No set operations, do regular planning */ |
| 1660 |
PathTarget *sort_input_target; |
| 1661 |
List *sort_input_targets; |
| 1662 |
List *sort_input_targets_contain_srfs; |
| 1663 |
bool sort_input_target_parallel_safe; |
| 1664 |
PathTarget *grouping_target; |
| 1665 |
List *grouping_targets; |
| 1666 |
List *grouping_targets_contain_srfs; |
| 1667 |
bool grouping_target_parallel_safe; |
| 1668 |
PathTarget *scanjoin_target; |
| 1669 |
List *scanjoin_targets; |
| 1670 |
List *scanjoin_targets_contain_srfs; |
| 1671 |
bool scanjoin_target_parallel_safe; |
| 1672 |
bool scanjoin_target_same_exprs; |
| 1673 |
bool have_grouping; |
| 1674 |
WindowFuncLists *wflists = NULL; |
| 1675 |
List *activeWindows = NIL; |
| 1676 |
grouping_sets_data *gset_data = NULL; |
| 1677 |
standard_qp_extra qp_extra; |
| 1678 |
|
| 1679 |
/* A recursive query should always have setOperations */ |
| 1680 |
Assert(!root->hasRecursion); |
| 1681 |
|
| 1682 |
/* Preprocess grouping sets and GROUP BY clause, if any */ |
| 1683 |
if (parse->groupingSets) |
| 1684 |
{ |
| 1685 |
gset_data = preprocess_grouping_sets(root); |
| 1686 |
} |
| 1687 |
else if (parse->groupClause) |
| 1688 |
{ |
| 1689 |
/* Preprocess regular GROUP BY clause, if any */ |
| 1690 |
root->processed_groupClause = preprocess_groupclause(root, NIL); |
| 1691 |
} |
| 1692 |
|
| 1693 |
/* |
| 1694 |
* Preprocess targetlist. Note that much of the remaining planning |
| 1695 |
* work will be done with the PathTarget representation of tlists, but |
| 1696 |
* we must also maintain the full representation of the final tlist so |
| 1697 |
* that we can transfer its decoration (resnames etc) to the topmost |
| 1698 |
* tlist of the finished Plan. This is kept in processed_tlist. |
| 1699 |
*/ |
| 1700 |
preprocess_targetlist(root); |
| 1701 |
|
| 1702 |
/* |
| 1703 |
* Mark all the aggregates with resolved aggtranstypes, and detect |
| 1704 |
* aggregates that are duplicates or can share transition state. We |
| 1705 |
* must do this before slicing and dicing the tlist into various |
| 1706 |
* pathtargets, else some copies of the Aggref nodes might escape |
| 1707 |
* being marked. |
| 1708 |
*/ |
| 1709 |
if (parse->hasAggs) |
| 1710 |
{ |
| 1711 |
preprocess_aggrefs(root, (Node *) root->processed_tlist); |
| 1712 |
preprocess_aggrefs(root, (Node *) parse->havingQual); |
| 1713 |
} |
| 1714 |
|
| 1715 |
/* |
| 1716 |
* Locate any window functions in the tlist. (We don't need to look |
| 1717 |
* anywhere else, since expressions used in ORDER BY will be in there |
| 1718 |
* too.) Note that they could all have been eliminated by constant |
| 1719 |
* folding, in which case we don't need to do any more work. |
| 1720 |
*/ |
| 1721 |
if (parse->hasWindowFuncs) |
| 1722 |
{ |
| 1723 |
wflists = find_window_functions((Node *) root->processed_tlist, |
| 1724 |
list_length(parse->windowClause)); |
| 1725 |
if (wflists->numWindowFuncs > 0) |
| 1726 |
{ |
| 1727 |
/* |
| 1728 |
* See if any modifications can be made to each WindowClause |
| 1729 |
* to allow the executor to execute the WindowFuncs more |
| 1730 |
* quickly. |
| 1731 |
*/ |
| 1732 |
optimize_window_clauses(root, wflists); |
| 1733 |
|
| 1734 |
/* Extract the list of windows actually in use. */ |
| 1735 |
activeWindows = select_active_windows(root, wflists); |
| 1736 |
|
| 1737 |
/* Make sure they all have names, for EXPLAIN's use. */ |
| 1738 |
name_active_windows(activeWindows); |
| 1739 |
} |
| 1740 |
else |
| 1741 |
parse->hasWindowFuncs = false; |
| 1742 |
} |
| 1743 |
|
| 1744 |
/* |
| 1745 |
* Preprocess MIN/MAX aggregates, if any. Note: be careful about |
| 1746 |
* adding logic between here and the query_planner() call. Anything |
| 1747 |
* that is needed in MIN/MAX-optimizable cases will have to be |
| 1748 |
* duplicated in planagg.c. |
| 1749 |
*/ |
| 1750 |
if (parse->hasAggs) |
| 1751 |
preprocess_minmax_aggregates(root); |
| 1752 |
|
| 1753 |
/* |
| 1754 |
* Figure out whether there's a hard limit on the number of rows that |
| 1755 |
* query_planner's result subplan needs to return. Even if we know a |
| 1756 |
* hard limit overall, it doesn't apply if the query has any |
| 1757 |
* grouping/aggregation operations, or SRFs in the tlist. |
| 1758 |
*/ |
| 1759 |
if (parse->groupClause || |
| 1760 |
parse->groupingSets || |
| 1761 |
parse->distinctClause || |
| 1762 |
parse->hasAggs || |
| 1763 |
parse->hasWindowFuncs || |
| 1764 |
parse->hasTargetSRFs || |
| 1765 |
root->hasHavingQual) |
| 1766 |
root->limit_tuples = -1.0; |
| 1767 |
else |
| 1768 |
root->limit_tuples = limit_tuples; |
| 1769 |
|
| 1770 |
/* Set up data needed by standard_qp_callback */ |
| 1771 |
qp_extra.activeWindows = activeWindows; |
| 1772 |
qp_extra.gset_data = gset_data; |
| 1773 |
|
| 1774 |
/* |
| 1775 |
* If we're a subquery for a set operation, store the SetOperationStmt |
| 1776 |
* in qp_extra. |
| 1777 |
*/ |
| 1778 |
qp_extra.setop = setops; |
| 1779 |
|
| 1780 |
/* |
| 1781 |
* Generate the best unsorted and presorted paths for the scan/join |
| 1782 |
* portion of this Query, ie the processing represented by the |
| 1783 |
* FROM/WHERE clauses. (Note there may not be any presorted paths.) |
| 1784 |
* We also generate (in standard_qp_callback) pathkey representations |
| 1785 |
* of the query's sort clause, distinct clause, etc. |
| 1786 |
*/ |
| 1787 |
current_rel = query_planner(root, standard_qp_callback, &qp_extra); |
| 1788 |
|
| 1789 |
/* |
| 1790 |
* Convert the query's result tlist into PathTarget format. |
| 1791 |
* |
| 1792 |
* Note: this cannot be done before query_planner() has performed |
| 1793 |
* appendrel expansion, because that might add resjunk entries to |
| 1794 |
* root->processed_tlist. Waiting till afterwards is also helpful |
| 1795 |
* because the target width estimates can use per-Var width numbers |
| 1796 |
* that were obtained within query_planner(). |
| 1797 |
*/ |
| 1798 |
final_target = create_pathtarget(root, root->processed_tlist); |
| 1799 |
final_target_parallel_safe = |
| 1800 |
is_parallel_safe(root, (Node *) final_target->exprs); |
| 1801 |
|
| 1802 |
/* |
| 1803 |
* If ORDER BY was given, consider whether we should use a post-sort |
| 1804 |
* projection, and compute the adjusted target for preceding steps if |
| 1805 |
* so. |
| 1806 |
*/ |
| 1807 |
if (parse->sortClause) |
| 1808 |
{ |
| 1809 |
sort_input_target = make_sort_input_target(root, |
| 1810 |
final_target, |
| 1811 |
&have_postponed_srfs); |
| 1812 |
sort_input_target_parallel_safe = |
| 1813 |
is_parallel_safe(root, (Node *) sort_input_target->exprs); |
| 1814 |
} |
| 1815 |
else |
| 1816 |
{ |
| 1817 |
sort_input_target = final_target; |
| 1818 |
sort_input_target_parallel_safe = final_target_parallel_safe; |
| 1819 |
} |
| 1820 |
|
| 1821 |
/* |
| 1822 |
* If we have window functions to deal with, the output from any |
| 1823 |
* grouping step needs to be what the window functions want; |
| 1824 |
* otherwise, it should be sort_input_target. |
| 1825 |
*/ |
| 1826 |
if (activeWindows) |
| 1827 |
{ |
| 1828 |
grouping_target = make_window_input_target(root, |
| 1829 |
final_target, |
| 1830 |
activeWindows); |
| 1831 |
grouping_target_parallel_safe = |
| 1832 |
is_parallel_safe(root, (Node *) grouping_target->exprs); |
| 1833 |
} |
| 1834 |
else |
| 1835 |
{ |
| 1836 |
grouping_target = sort_input_target; |
| 1837 |
grouping_target_parallel_safe = sort_input_target_parallel_safe; |
| 1838 |
} |
| 1839 |
|
| 1840 |
/* |
| 1841 |
* If we have grouping or aggregation to do, the topmost scan/join |
| 1842 |
* plan node must emit what the grouping step wants; otherwise, it |
| 1843 |
* should emit grouping_target. |
| 1844 |
*/ |
| 1845 |
have_grouping = (parse->groupClause || parse->groupingSets || |
| 1846 |
parse->hasAggs || root->hasHavingQual); |
| 1847 |
if (have_grouping) |
| 1848 |
{ |
| 1849 |
scanjoin_target = make_group_input_target(root, final_target); |
| 1850 |
scanjoin_target_parallel_safe = |
| 1851 |
is_parallel_safe(root, (Node *) scanjoin_target->exprs); |
| 1852 |
} |
| 1853 |
else |
| 1854 |
{ |
| 1855 |
scanjoin_target = grouping_target; |
| 1856 |
scanjoin_target_parallel_safe = grouping_target_parallel_safe; |
| 1857 |
} |
| 1858 |
|
| 1859 |
/* |
| 1860 |
* If there are any SRFs in the targetlist, we must separate each of |
| 1861 |
* these PathTargets into SRF-computing and SRF-free targets. Replace |
| 1862 |
* each of the named targets with a SRF-free version, and remember the |
| 1863 |
* list of additional projection steps we need to add afterwards. |
| 1864 |
*/ |
| 1865 |
if (parse->hasTargetSRFs) |
| 1866 |
{ |
| 1867 |
/* final_target doesn't recompute any SRFs in sort_input_target */ |
| 1868 |
split_pathtarget_at_srfs(root, final_target, sort_input_target, |
| 1869 |
&final_targets, |
| 1870 |
&final_targets_contain_srfs); |
| 1871 |
final_target = linitial_node(PathTarget, final_targets); |
| 1872 |
Assert(!linitial_int(final_targets_contain_srfs)); |
| 1873 |
/* likewise for sort_input_target vs. grouping_target */ |
| 1874 |
split_pathtarget_at_srfs(root, sort_input_target, grouping_target, |
| 1875 |
&sort_input_targets, |
| 1876 |
&sort_input_targets_contain_srfs); |
| 1877 |
sort_input_target = linitial_node(PathTarget, sort_input_targets); |
| 1878 |
Assert(!linitial_int(sort_input_targets_contain_srfs)); |
| 1879 |
/* likewise for grouping_target vs. scanjoin_target */ |
| 1880 |
split_pathtarget_at_srfs_grouping(root, |
| 1881 |
grouping_target, scanjoin_target, |
| 1882 |
&grouping_targets, |
| 1883 |
&grouping_targets_contain_srfs); |
| 1884 |
grouping_target = linitial_node(PathTarget, grouping_targets); |
| 1885 |
Assert(!linitial_int(grouping_targets_contain_srfs)); |
| 1886 |
/* scanjoin_target will not have any SRFs precomputed for it */ |
| 1887 |
split_pathtarget_at_srfs(root, scanjoin_target, NULL, |
| 1888 |
&scanjoin_targets, |
| 1889 |
&scanjoin_targets_contain_srfs); |
| 1890 |
scanjoin_target = linitial_node(PathTarget, scanjoin_targets); |
| 1891 |
Assert(!linitial_int(scanjoin_targets_contain_srfs)); |
| 1892 |
} |
| 1893 |
else |
| 1894 |
{ |
| 1895 |
/* initialize lists; for most of these, dummy values are OK */ |
| 1896 |
final_targets = final_targets_contain_srfs = NIL; |
| 1897 |
sort_input_targets = sort_input_targets_contain_srfs = NIL; |
| 1898 |
grouping_targets = grouping_targets_contain_srfs = NIL; |
| 1899 |
scanjoin_targets = list_make1(scanjoin_target); |
| 1900 |
scanjoin_targets_contain_srfs = NIL; |
| 1901 |
} |
| 1902 |
|
| 1903 |
/* Apply scan/join target. */ |
| 1904 |
scanjoin_target_same_exprs = list_length(scanjoin_targets) == 1 |
| 1905 |
&& equal(scanjoin_target->exprs, current_rel->reltarget->exprs); |
| 1906 |
apply_scanjoin_target_to_paths(root, current_rel, scanjoin_targets, |
| 1907 |
scanjoin_targets_contain_srfs, |
| 1908 |
scanjoin_target_parallel_safe, |
| 1909 |
scanjoin_target_same_exprs); |
| 1910 |
|
| 1911 |
/* |
| 1912 |
* Save the various upper-rel PathTargets we just computed into |
| 1913 |
* root->upper_targets[]. The core code doesn't use this, but it |
| 1914 |
* provides a convenient place for extensions to get at the info. For |
| 1915 |
* consistency, we save all the intermediate targets, even though some |
| 1916 |
* of the corresponding upperrels might not be needed for this query. |
| 1917 |
*/ |
| 1918 |
root->upper_targets[UPPERREL_FINAL] = final_target; |
| 1919 |
root->upper_targets[UPPERREL_ORDERED] = final_target; |
| 1920 |
root->upper_targets[UPPERREL_DISTINCT] = sort_input_target; |
| 1921 |
root->upper_targets[UPPERREL_PARTIAL_DISTINCT] = sort_input_target; |
| 1922 |
root->upper_targets[UPPERREL_WINDOW] = sort_input_target; |
| 1923 |
root->upper_targets[UPPERREL_GROUP_AGG] = grouping_target; |
| 1924 |
|
| 1925 |
/* |
| 1926 |
* If we have grouping and/or aggregation, consider ways to implement |
| 1927 |
* that. We build a new upperrel representing the output of this |
| 1928 |
* phase. |
| 1929 |
*/ |
| 1930 |
if (have_grouping) |
| 1931 |
{ |
| 1932 |
current_rel = create_grouping_paths(root, |
| 1933 |
current_rel, |
| 1934 |
grouping_target, |
| 1935 |
grouping_target_parallel_safe, |
| 1936 |
gset_data); |
| 1937 |
/* Fix things up if grouping_target contains SRFs */ |
| 1938 |
if (parse->hasTargetSRFs) |
| 1939 |
adjust_paths_for_srfs(root, current_rel, |
| 1940 |
grouping_targets, |
| 1941 |
grouping_targets_contain_srfs); |
| 1942 |
} |
| 1943 |
|
| 1944 |
/* |
| 1945 |
* If we have window functions, consider ways to implement those. We |
| 1946 |
* build a new upperrel representing the output of this phase. |
| 1947 |
*/ |
| 1948 |
if (activeWindows) |
| 1949 |
{ |
| 1950 |
current_rel = create_window_paths(root, |
| 1951 |
current_rel, |
| 1952 |
grouping_target, |
| 1953 |
sort_input_target, |
| 1954 |
sort_input_target_parallel_safe, |
| 1955 |
wflists, |
| 1956 |
activeWindows); |
| 1957 |
/* Fix things up if sort_input_target contains SRFs */ |
| 1958 |
if (parse->hasTargetSRFs) |
| 1959 |
adjust_paths_for_srfs(root, current_rel, |
| 1960 |
sort_input_targets, |
| 1961 |
sort_input_targets_contain_srfs); |
| 1962 |
} |
| 1963 |
|
| 1964 |
/* |
| 1965 |
* If there is a DISTINCT clause, consider ways to implement that. We |
| 1966 |
* build a new upperrel representing the output of this phase. |
| 1967 |
*/ |
| 1968 |
if (parse->distinctClause) |
| 1969 |
{ |
| 1970 |
current_rel = create_distinct_paths(root, |
| 1971 |
current_rel, |
| 1972 |
sort_input_target); |
| 1973 |
} |
| 1974 |
} /* end of if (setOperations) */ |
| 1975 |
|
| 1976 |
/* |
| 1977 |
* If ORDER BY was given, consider ways to implement that, and generate a |
| 1978 |
* new upperrel containing only paths that emit the correct ordering and |
| 1979 |
* project the correct final_target. We can apply the original |
| 1980 |
* limit_tuples limit in sort costing here, but only if there are no |
| 1981 |
* postponed SRFs. |
| 1982 |
*/ |
| 1983 |
if (parse->sortClause) |
| 1984 |
{ |
| 1985 |
current_rel = create_ordered_paths(root, |
| 1986 |
current_rel, |
| 1987 |
final_target, |
| 1988 |
final_target_parallel_safe, |
| 1989 |
have_postponed_srfs ? -1.0 : |
| 1990 |
limit_tuples); |
| 1991 |
/* Fix things up if final_target contains SRFs */ |
| 1992 |
if (parse->hasTargetSRFs) |
| 1993 |
adjust_paths_for_srfs(root, current_rel, |
| 1994 |
final_targets, |
| 1995 |
final_targets_contain_srfs); |
| 1996 |
} |
| 1997 |
|
| 1998 |
/* |
| 1999 |
* Now we are prepared to build the final-output upperrel. |
| 2000 |
*/ |
| 2001 |
final_rel = fetch_upper_rel(root, UPPERREL_FINAL, NULL); |
| 2002 |
|
| 2003 |
/* |
| 2004 |
* If the input rel is marked consider_parallel and there's nothing that's |
| 2005 |
* not parallel-safe in the LIMIT clause, then the final_rel can be marked |
| 2006 |
* consider_parallel as well. Note that if the query has rowMarks or is |
| 2007 |
* not a SELECT, consider_parallel will be false for every relation in the |
| 2008 |
* query. |
| 2009 |
*/ |
| 2010 |
if (current_rel->consider_parallel && |
| 2011 |
is_parallel_safe(root, parse->limitOffset) && |
| 2012 |
is_parallel_safe(root, parse->limitCount)) |
| 2013 |
final_rel->consider_parallel = true; |
| 2014 |
|
| 2015 |
/* |
| 2016 |
* If the current_rel belongs to a single FDW, so does the final_rel. |
| 2017 |
*/ |
| 2018 |
final_rel->serverid = current_rel->serverid; |
| 2019 |
final_rel->userid = current_rel->userid; |
| 2020 |
final_rel->useridiscurrent = current_rel->useridiscurrent; |
| 2021 |
final_rel->fdwroutine = current_rel->fdwroutine; |
| 2022 |
|
| 2023 |
/* |
| 2024 |
* Generate paths for the final_rel. Insert all surviving paths, with |
| 2025 |
* LockRows, Limit, and/or ModifyTable steps added if needed. |
| 2026 |
*/ |
| 2027 |
foreach(lc, current_rel->pathlist) |
| 2028 |
{ |
| 2029 |
Path *path = (Path *) lfirst(lc); |
| 2030 |
|
| 2031 |
/* |
| 2032 |
* If there is a FOR [KEY] UPDATE/SHARE clause, add the LockRows node. |
| 2033 |
* (Note: we intentionally test parse->rowMarks not root->rowMarks |
| 2034 |
* here. If there are only non-locking rowmarks, they should be |
| 2035 |
* handled by the ModifyTable node instead. However, root->rowMarks |
| 2036 |
* is what goes into the LockRows node.) |
| 2037 |
*/ |
| 2038 |
if (parse->rowMarks) |
| 2039 |
{ |
| 2040 |
path = (Path *) create_lockrows_path(root, final_rel, path, |
| 2041 |
root->rowMarks, |
| 2042 |
assign_special_exec_param(root)); |
| 2043 |
} |
| 2044 |
|
| 2045 |
/* |
| 2046 |
* If there is a LIMIT/OFFSET clause, add the LIMIT node. |
| 2047 |
*/ |
| 2048 |
if (limit_needed(parse)) |
| 2049 |
{ |
| 2050 |
path = (Path *) create_limit_path(root, final_rel, path, |
| 2051 |
parse->limitOffset, |
| 2052 |
parse->limitCount, |
| 2053 |
parse->limitOption, |
| 2054 |
offset_est, count_est); |
| 2055 |
} |
| 2056 |
|
| 2057 |
/* |
| 2058 |
* If this is an INSERT/UPDATE/DELETE/MERGE, add the ModifyTable node. |
| 2059 |
*/ |
| 2060 |
if (parse->commandType != CMD_SELECT) |
| 2061 |
{ |
| 2062 |
Index rootRelation; |
| 2063 |
List *resultRelations = NIL; |
| 2064 |
List *updateColnosLists = NIL; |
| 2065 |
List *withCheckOptionLists = NIL; |
| 2066 |
List *returningLists = NIL; |
| 2067 |
List *mergeActionLists = NIL; |
| 2068 |
List *mergeJoinConditions = NIL; |
| 2069 |
List *rowMarks; |
| 2070 |
|
| 2071 |
if (bms_membership(root->all_result_relids) == BMS_MULTIPLE) |
| 2072 |
{ |
| 2073 |
/* Inherited UPDATE/DELETE/MERGE */ |
| 2074 |
RelOptInfo *top_result_rel = find_base_rel(root, |
| 2075 |
parse->resultRelation); |
| 2076 |
int resultRelation = -1; |
| 2077 |
|
| 2078 |
/* Pass the root result rel forward to the executor. */ |
| 2079 |
rootRelation = parse->resultRelation; |
| 2080 |
|
| 2081 |
/* Add only leaf children to ModifyTable. */ |
| 2082 |
while ((resultRelation = bms_next_member(root->leaf_result_relids, |
| 2083 |
resultRelation)) >= 0) |
| 2084 |
{ |
| 2085 |
RelOptInfo *this_result_rel = find_base_rel(root, |
| 2086 |
resultRelation); |
| 2087 |
|
| 2088 |
/* |
| 2089 |
* Also exclude any leaf rels that have turned dummy since |
| 2090 |
* being added to the list, for example, by being excluded |
| 2091 |
* by constraint exclusion. |
| 2092 |
*/ |
| 2093 |
if (IS_DUMMY_REL(this_result_rel)) |
| 2094 |
continue; |
| 2095 |
|
| 2096 |
/* Build per-target-rel lists needed by ModifyTable */ |
| 2097 |
resultRelations = lappend_int(resultRelations, |
| 2098 |
resultRelation); |
| 2099 |
if (parse->commandType == CMD_UPDATE) |
| 2100 |
{ |
| 2101 |
List *update_colnos = root->update_colnos; |
| 2102 |
|
| 2103 |
if (this_result_rel != top_result_rel) |
| 2104 |
update_colnos = |
| 2105 |
adjust_inherited_attnums_multilevel(root, |
| 2106 |
update_colnos, |
| 2107 |
this_result_rel->relid, |
| 2108 |
top_result_rel->relid); |
| 2109 |
updateColnosLists = lappend(updateColnosLists, |
| 2110 |
update_colnos); |
| 2111 |
} |
| 2112 |
if (parse->withCheckOptions) |
| 2113 |
{ |
| 2114 |
List *withCheckOptions = parse->withCheckOptions; |
| 2115 |
|
| 2116 |
if (this_result_rel != top_result_rel) |
| 2117 |
withCheckOptions = (List *) |
| 2118 |
adjust_appendrel_attrs_multilevel(root, |
| 2119 |
(Node *) withCheckOptions, |
| 2120 |
this_result_rel, |
| 2121 |
top_result_rel); |
| 2122 |
withCheckOptionLists = lappend(withCheckOptionLists, |
| 2123 |
withCheckOptions); |
| 2124 |
} |
| 2125 |
if (parse->returningList) |
| 2126 |
{ |
| 2127 |
List *returningList = parse->returningList; |
| 2128 |
|
| 2129 |
if (this_result_rel != top_result_rel) |
| 2130 |
returningList = (List *) |
| 2131 |
adjust_appendrel_attrs_multilevel(root, |
| 2132 |
(Node *) returningList, |
| 2133 |
this_result_rel, |
| 2134 |
top_result_rel); |
| 2135 |
returningLists = lappend(returningLists, |
| 2136 |
returningList); |
| 2137 |
} |
| 2138 |
if (parse->mergeActionList) |
| 2139 |
{ |
| 2140 |
ListCell *l; |
| 2141 |
List *mergeActionList = NIL; |
| 2142 |
|
| 2143 |
/* |
| 2144 |
* Copy MergeActions and translate stuff that |
| 2145 |
* references attribute numbers. |
| 2146 |
*/ |
| 2147 |
foreach(l, parse->mergeActionList) |
| 2148 |
{ |
| 2149 |
MergeAction *action = lfirst(l), |
| 2150 |
*leaf_action = copyObject(action); |
| 2151 |
|
| 2152 |
leaf_action->qual = |
| 2153 |
adjust_appendrel_attrs_multilevel(root, |
| 2154 |
(Node *) action->qual, |
| 2155 |
this_result_rel, |
| 2156 |
top_result_rel); |
| 2157 |
leaf_action->targetList = (List *) |
| 2158 |
adjust_appendrel_attrs_multilevel(root, |
| 2159 |
(Node *) action->targetList, |
| 2160 |
this_result_rel, |
| 2161 |
top_result_rel); |
| 2162 |
if (leaf_action->commandType == CMD_UPDATE) |
| 2163 |
leaf_action->updateColnos = |
| 2164 |
adjust_inherited_attnums_multilevel(root, |
| 2165 |
action->updateColnos, |
| 2166 |
this_result_rel->relid, |
| 2167 |
top_result_rel->relid); |
| 2168 |
mergeActionList = lappend(mergeActionList, |
| 2169 |
leaf_action); |
| 2170 |
} |
| 2171 |
|
| 2172 |
mergeActionLists = lappend(mergeActionLists, |
| 2173 |
mergeActionList); |
| 2174 |
} |
| 2175 |
if (parse->commandType == CMD_MERGE) |
| 2176 |
{ |
| 2177 |
Node *mergeJoinCondition = parse->mergeJoinCondition; |
| 2178 |
|
| 2179 |
if (this_result_rel != top_result_rel) |
| 2180 |
mergeJoinCondition = |
| 2181 |
adjust_appendrel_attrs_multilevel(root, |
| 2182 |
mergeJoinCondition, |
| 2183 |
this_result_rel, |
| 2184 |
top_result_rel); |
| 2185 |
mergeJoinConditions = lappend(mergeJoinConditions, |
| 2186 |
mergeJoinCondition); |
| 2187 |
} |
| 2188 |
} |
| 2189 |
|
| 2190 |
if (resultRelations == NIL) |
| 2191 |
{ |
| 2192 |
/* |
| 2193 |
* We managed to exclude every child rel, so generate a |
| 2194 |
* dummy one-relation plan using info for the top target |
| 2195 |
* rel (even though that may not be a leaf target). |
| 2196 |
* Although it's clear that no data will be updated or |
| 2197 |
* deleted, we still need to have a ModifyTable node so |
| 2198 |
* that any statement triggers will be executed. (This |
| 2199 |
* could be cleaner if we fixed nodeModifyTable.c to allow |
| 2200 |
* zero target relations, but that probably wouldn't be a |
| 2201 |
* net win.) |
| 2202 |
*/ |
| 2203 |
resultRelations = list_make1_int(parse->resultRelation); |
| 2204 |
if (parse->commandType == CMD_UPDATE) |
| 2205 |
updateColnosLists = list_make1(root->update_colnos); |
| 2206 |
if (parse->withCheckOptions) |
| 2207 |
withCheckOptionLists = list_make1(parse->withCheckOptions); |
| 2208 |
if (parse->returningList) |
| 2209 |
returningLists = list_make1(parse->returningList); |
| 2210 |
if (parse->mergeActionList) |
| 2211 |
mergeActionLists = list_make1(parse->mergeActionList); |
| 2212 |
if (parse->commandType == CMD_MERGE) |
| 2213 |
mergeJoinConditions = list_make1(parse->mergeJoinCondition); |
| 2214 |
} |
| 2215 |
} |
| 2216 |
else |
| 2217 |
{ |
| 2218 |
/* Single-relation INSERT/UPDATE/DELETE/MERGE. */ |
| 2219 |
rootRelation = 0; /* there's no separate root rel */ |
| 2220 |
resultRelations = list_make1_int(parse->resultRelation); |
| 2221 |
if (parse->commandType == CMD_UPDATE) |
| 2222 |
updateColnosLists = list_make1(root->update_colnos); |
| 2223 |
if (parse->withCheckOptions) |
| 2224 |
withCheckOptionLists = list_make1(parse->withCheckOptions); |
| 2225 |
if (parse->returningList) |
| 2226 |
returningLists = list_make1(parse->returningList); |
| 2227 |
if (parse->mergeActionList) |
| 2228 |
mergeActionLists = list_make1(parse->mergeActionList); |
| 2229 |
if (parse->commandType == CMD_MERGE) |
| 2230 |
mergeJoinConditions = list_make1(parse->mergeJoinCondition); |
| 2231 |
} |
| 2232 |
|
| 2233 |
/* |
| 2234 |
* If there was a FOR [KEY] UPDATE/SHARE clause, the LockRows node |
| 2235 |
* will have dealt with fetching non-locked marked rows, else we |
| 2236 |
* need to have ModifyTable do that. |
| 2237 |
*/ |
| 2238 |
if (parse->rowMarks) |
| 2239 |
rowMarks = NIL; |
| 2240 |
else |
| 2241 |
rowMarks = root->rowMarks; |
| 2242 |
|
| 2243 |
path = (Path *) |
| 2244 |
create_modifytable_path(root, final_rel, |
| 2245 |
path, |
| 2246 |
parse->commandType, |
| 2247 |
parse->canSetTag, |
| 2248 |
parse->resultRelation, |
| 2249 |
rootRelation, |
| 2250 |
root->partColsUpdated, |
| 2251 |
resultRelations, |
| 2252 |
updateColnosLists, |
| 2253 |
withCheckOptionLists, |
| 2254 |
returningLists, |
| 2255 |
rowMarks, |
| 2256 |
parse->onConflict, |
| 2257 |
mergeActionLists, |
| 2258 |
mergeJoinConditions, |
| 2259 |
assign_special_exec_param(root)); |
| 2260 |
} |
| 2261 |
|
| 2262 |
/* And shove it into final_rel */ |
| 2263 |
add_path(final_rel, path); |
| 2264 |
} |
| 2265 |
|
| 2266 |
/* |
| 2267 |
* Generate partial paths for final_rel, too, if outer query levels might |
| 2268 |
* be able to make use of them. |
| 2269 |
*/ |
| 2270 |
if (final_rel->consider_parallel && root->query_level > 1 && |
| 2271 |
!limit_needed(parse)) |
| 2272 |
{ |
| 2273 |
Assert(!parse->rowMarks && parse->commandType == CMD_SELECT); |
| 2274 |
foreach(lc, current_rel->partial_pathlist) |
| 2275 |
{ |
| 2276 |
Path *partial_path = (Path *) lfirst(lc); |
| 2277 |
|
| 2278 |
add_partial_path(final_rel, partial_path); |
| 2279 |
} |
| 2280 |
} |
| 2281 |
|
| 2282 |
extra.limit_needed = limit_needed(parse); |
| 2283 |
extra.limit_tuples = limit_tuples; |
| 2284 |
extra.count_est = count_est; |
| 2285 |
extra.offset_est = offset_est; |
| 2286 |
|
| 2287 |
/* |
| 2288 |
* If there is an FDW that's responsible for all baserels of the query, |
| 2289 |
* let it consider adding ForeignPaths. |
| 2290 |
*/ |
| 2291 |
if (final_rel->fdwroutine && |
| 2292 |
final_rel->fdwroutine->GetForeignUpperPaths) |
| 2293 |
final_rel->fdwroutine->GetForeignUpperPaths(root, UPPERREL_FINAL, |
| 2294 |
current_rel, final_rel, |
| 2295 |
&extra); |
| 2296 |
|
| 2297 |
/* Let extensions possibly add some more paths */ |
| 2298 |
if (create_upper_paths_hook) |
| 2299 |
(*create_upper_paths_hook) (root, UPPERREL_FINAL, |
| 2300 |
current_rel, final_rel, &extra); |
| 2301 |
|
| 2302 |
/* Note: currently, we leave it to callers to do set_cheapest() */ |
| 2303 |
} |
| 2304 |
|
| 2305 |
/* |
| 2306 |
* Do preprocessing for groupingSets clause and related data. |
| 2307 |
* |
| 2308 |
* We expect that parse->groupingSets has already been expanded into a flat |
| 2309 |
* list of grouping sets (that is, just integer Lists of ressortgroupref |
| 2310 |
* numbers) by expand_grouping_sets(). This function handles the preliminary |
| 2311 |
* steps of organizing the grouping sets into lists of rollups, and preparing |
| 2312 |
* annotations which will later be filled in with size estimates. |
| 2313 |
*/ |
| 2314 |
static grouping_sets_data * |
| 2315 |
preprocess_grouping_sets(PlannerInfo *root) |
| 2316 |
{ |
| 2317 |
Query *parse = root->parse; |
| 2318 |
List *sets; |
| 2319 |
int maxref = 0; |
| 2320 |
ListCell *lc_set; |
| 2321 |
grouping_sets_data *gd = palloc0(sizeof(grouping_sets_data)); |
| 2322 |
|
| 2323 |
/* |
| 2324 |
* We don't currently make any attempt to optimize the groupClause when |
| 2325 |
* there are grouping sets, so just duplicate it in processed_groupClause. |
| 2326 |
*/ |
| 2327 |
root->processed_groupClause = parse->groupClause; |
| 2328 |
|
| 2329 |
/* Detect unhashable and unsortable grouping expressions */ |
| 2330 |
gd->any_hashable = false; |
| 2331 |
gd->unhashable_refs = NULL; |
| 2332 |
gd->unsortable_refs = NULL; |
| 2333 |
gd->unsortable_sets = NIL; |
| 2334 |
|
| 2335 |
if (parse->groupClause) |
| 2336 |
{ |
| 2337 |
ListCell *lc; |
| 2338 |
|
| 2339 |
foreach(lc, parse->groupClause) |
| 2340 |
{ |
| 2341 |
SortGroupClause *gc = lfirst_node(SortGroupClause, lc); |
| 2342 |
Index ref = gc->tleSortGroupRef; |
| 2343 |
|
| 2344 |
if (ref > maxref) |
| 2345 |
maxref = ref; |
| 2346 |
|
| 2347 |
if (!gc->hashable) |
| 2348 |
gd->unhashable_refs = bms_add_member(gd->unhashable_refs, ref); |
| 2349 |
|
| 2350 |
if (!OidIsValid(gc->sortop)) |
| 2351 |
gd->unsortable_refs = bms_add_member(gd->unsortable_refs, ref); |
| 2352 |
} |
| 2353 |
} |
| 2354 |
|
| 2355 |
/* Allocate workspace array for remapping */ |
| 2356 |
gd->tleref_to_colnum_map = (int *) palloc((maxref + 1) * sizeof(int)); |
| 2357 |
|
| 2358 |
/* |
| 2359 |
* If we have any unsortable sets, we must extract them before trying to |
| 2360 |
* prepare rollups. Unsortable sets don't go through |
| 2361 |
* reorder_grouping_sets, so we must apply the GroupingSetData annotation |
| 2362 |
* here. |
| 2363 |
*/ |
| 2364 |
if (!bms_is_empty(gd->unsortable_refs)) |
| 2365 |
{ |
| 2366 |
List *sortable_sets = NIL; |
| 2367 |
ListCell *lc; |
| 2368 |
|
| 2369 |
foreach(lc, parse->groupingSets) |
| 2370 |
{ |
| 2371 |
List *gset = (List *) lfirst(lc); |
| 2372 |
|
| 2373 |
if (bms_overlap_list(gd->unsortable_refs, gset)) |
| 2374 |
{ |
| 2375 |
GroupingSetData *gs = makeNode(GroupingSetData); |
| 2376 |
|
| 2377 |
gs->set = gset; |
| 2378 |
gd->unsortable_sets = lappend(gd->unsortable_sets, gs); |
| 2379 |
|
| 2380 |
/* |
| 2381 |
* We must enforce here that an unsortable set is hashable; |
| 2382 |
* later code assumes this. Parse analysis only checks that |
| 2383 |
* every individual column is either hashable or sortable. |
| 2384 |
* |
| 2385 |
* Note that passing this test doesn't guarantee we can |
| 2386 |
* generate a plan; there might be other showstoppers. |
| 2387 |
*/ |
| 2388 |
if (bms_overlap_list(gd->unhashable_refs, gset)) |
| 2389 |
ereport(ERROR, |
| 2390 |
(errcode(ERRCODE_FEATURE_NOT_SUPPORTED), |
| 2391 |
errmsg("could not implement GROUP BY"), |
| 2392 |
errdetail("Some of the datatypes only support hashing, while others only support sorting."))); |
| 2393 |
} |
| 2394 |
else |
| 2395 |
sortable_sets = lappend(sortable_sets, gset); |
| 2396 |
} |
| 2397 |
|
| 2398 |
if (sortable_sets) |
| 2399 |
sets = extract_rollup_sets(sortable_sets); |
| 2400 |
else |
| 2401 |
sets = NIL; |
| 2402 |
} |
| 2403 |
else |
| 2404 |
sets = extract_rollup_sets(parse->groupingSets); |
| 2405 |
|
| 2406 |
foreach(lc_set, sets) |
| 2407 |
{ |
| 2408 |
List *current_sets = (List *) lfirst(lc_set); |
| 2409 |
RollupData *rollup = makeNode(RollupData); |
| 2410 |
GroupingSetData *gs; |
| 2411 |
|
| 2412 |
/* |
| 2413 |
* Reorder the current list of grouping sets into correct prefix |
| 2414 |
* order. If only one aggregation pass is needed, try to make the |
| 2415 |
* list match the ORDER BY clause; if more than one pass is needed, we |
| 2416 |
* don't bother with that. |
| 2417 |
* |
| 2418 |
* Note that this reorders the sets from smallest-member-first to |
| 2419 |
* largest-member-first, and applies the GroupingSetData annotations, |
| 2420 |
* though the data will be filled in later. |
| 2421 |
*/ |
| 2422 |
current_sets = reorder_grouping_sets(current_sets, |
| 2423 |
(list_length(sets) == 1 |
| 2424 |
? parse->sortClause |
| 2425 |
: NIL)); |
| 2426 |
|
| 2427 |
/* |
| 2428 |
* Get the initial (and therefore largest) grouping set. |
| 2429 |
*/ |
| 2430 |
gs = linitial_node(GroupingSetData, current_sets); |
| 2431 |
|
| 2432 |
/* |
| 2433 |
* Order the groupClause appropriately. If the first grouping set is |
| 2434 |
* empty, then the groupClause must also be empty; otherwise we have |
| 2435 |
* to force the groupClause to match that grouping set's order. |
| 2436 |
* |
| 2437 |
* (The first grouping set can be empty even though parse->groupClause |
| 2438 |
* is not empty only if all non-empty grouping sets are unsortable. |
| 2439 |
* The groupClauses for hashed grouping sets are built later on.) |
| 2440 |
*/ |
| 2441 |
if (gs->set) |
| 2442 |
rollup->groupClause = preprocess_groupclause(root, gs->set); |
| 2443 |
else |
| 2444 |
rollup->groupClause = NIL; |
| 2445 |
|
| 2446 |
/* |
| 2447 |
* Is it hashable? We pretend empty sets are hashable even though we |
| 2448 |
* actually force them not to be hashed later. But don't bother if |
| 2449 |
* there's nothing but empty sets (since in that case we can't hash |
| 2450 |
* anything). |
| 2451 |
*/ |
| 2452 |
if (gs->set && |
| 2453 |
!bms_overlap_list(gd->unhashable_refs, gs->set)) |
| 2454 |
{ |
| 2455 |
rollup->hashable = true; |
| 2456 |
gd->any_hashable = true; |
| 2457 |
} |
| 2458 |
|
| 2459 |
/* |
| 2460 |
* Now that we've pinned down an order for the groupClause for this |
| 2461 |
* list of grouping sets, we need to remap the entries in the grouping |
| 2462 |
* sets from sortgrouprefs to plain indices (0-based) into the |
| 2463 |
* groupClause for this collection of grouping sets. We keep the |
| 2464 |
* original form for later use, though. |
| 2465 |
*/ |
| 2466 |
rollup->gsets = remap_to_groupclause_idx(rollup->groupClause, |
| 2467 |
current_sets, |
| 2468 |
gd->tleref_to_colnum_map); |
| 2469 |
rollup->gsets_data = current_sets; |
| 2470 |
|
| 2471 |
gd->rollups = lappend(gd->rollups, rollup); |
| 2472 |
} |
| 2473 |
|
| 2474 |
if (gd->unsortable_sets) |
| 2475 |
{ |
| 2476 |
/* |
| 2477 |
* We have not yet pinned down a groupclause for this, but we will |
| 2478 |
* need index-based lists for estimation purposes. Construct |
| 2479 |
* hash_sets_idx based on the entire original groupclause for now. |
| 2480 |
*/ |
| 2481 |
gd->hash_sets_idx = remap_to_groupclause_idx(parse->groupClause, |
| 2482 |
gd->unsortable_sets, |
| 2483 |
gd->tleref_to_colnum_map); |
| 2484 |
gd->any_hashable = true; |
| 2485 |
} |
| 2486 |
|
| 2487 |
return gd; |
| 2488 |
} |
| 2489 |
|
| 2490 |
/* |
| 2491 |
* Given a groupclause and a list of GroupingSetData, return equivalent sets |
| 2492 |
* (without annotation) mapped to indexes into the given groupclause. |
| 2493 |
*/ |
| 2494 |
static List * |
| 2495 |
remap_to_groupclause_idx(List *groupClause, |
| 2496 |
List *gsets, |
| 2497 |
int *tleref_to_colnum_map) |
| 2498 |
{ |
| 2499 |
int ref = 0; |
| 2500 |
List *result = NIL; |
| 2501 |
ListCell *lc; |
| 2502 |
|
| 2503 |
foreach(lc, groupClause) |
| 2504 |
{ |
| 2505 |
SortGroupClause *gc = lfirst_node(SortGroupClause, lc); |
| 2506 |
|
| 2507 |
tleref_to_colnum_map[gc->tleSortGroupRef] = ref++; |
| 2508 |
} |
| 2509 |
|
| 2510 |
foreach(lc, gsets) |
| 2511 |
{ |
| 2512 |
List *set = NIL; |
| 2513 |
ListCell *lc2; |
| 2514 |
GroupingSetData *gs = lfirst_node(GroupingSetData, lc); |
| 2515 |
|
| 2516 |
foreach(lc2, gs->set) |
| 2517 |
{ |
| 2518 |
set = lappend_int(set, tleref_to_colnum_map[lfirst_int(lc2)]); |
| 2519 |
} |
| 2520 |
|
| 2521 |
result = lappend(result, set); |
| 2522 |
} |
| 2523 |
|
| 2524 |
return result; |
| 2525 |
} |
| 2526 |
|
| 2527 |
|
| 2528 |
/* |
| 2529 |
* preprocess_rowmarks - set up PlanRowMarks if needed |
| 2530 |
*/ |
| 2531 |
static void |
| 2532 |
preprocess_rowmarks(PlannerInfo *root) |
| 2533 |
{ |
| 2534 |
Query *parse = root->parse; |
| 2535 |
Bitmapset *rels; |
| 2536 |
List *prowmarks; |
| 2537 |
ListCell *l; |
| 2538 |
int i; |
| 2539 |
|
| 2540 |
if (parse->rowMarks) |
| 2541 |
{ |
| 2542 |
/* |
| 2543 |
* We've got trouble if FOR [KEY] UPDATE/SHARE appears inside |
| 2544 |
* grouping, since grouping renders a reference to individual tuple |
| 2545 |
* CTIDs invalid. This is also checked at parse time, but that's |
| 2546 |
* insufficient because of rule substitution, query pullup, etc. |
| 2547 |
*/ |
| 2548 |
CheckSelectLocking(parse, linitial_node(RowMarkClause, |
| 2549 |
parse->rowMarks)->strength); |
| 2550 |
} |
| 2551 |
else |
| 2552 |
{ |
| 2553 |
/* |
| 2554 |
* We only need rowmarks for UPDATE, DELETE, MERGE, or FOR [KEY] |
| 2555 |
* UPDATE/SHARE. |
| 2556 |
*/ |
| 2557 |
if (parse->commandType != CMD_UPDATE && |
| 2558 |
parse->commandType != CMD_DELETE && |
| 2559 |
parse->commandType != CMD_MERGE) |
| 2560 |
return; |
| 2561 |
} |
| 2562 |
|
| 2563 |
/* |
| 2564 |
* We need to have rowmarks for all base relations except the target. We |
| 2565 |
* make a bitmapset of all base rels and then remove the items we don't |
| 2566 |
* need or have FOR [KEY] UPDATE/SHARE marks for. |
| 2567 |
*/ |
| 2568 |
rels = get_relids_in_jointree((Node *) parse->jointree, false, false); |
| 2569 |
if (parse->resultRelation) |
| 2570 |
rels = bms_del_member(rels, parse->resultRelation); |
| 2571 |
|
| 2572 |
/* |
| 2573 |
* Convert RowMarkClauses to PlanRowMark representation. |
| 2574 |
*/ |
| 2575 |
prowmarks = NIL; |
| 2576 |
foreach(l, parse->rowMarks) |
| 2577 |
{ |
| 2578 |
RowMarkClause *rc = lfirst_node(RowMarkClause, l); |
| 2579 |
RangeTblEntry *rte = rt_fetch(rc->rti, parse->rtable); |
| 2580 |
PlanRowMark *newrc; |
| 2581 |
|
| 2582 |
/* |
| 2583 |
* Currently, it is syntactically impossible to have FOR UPDATE et al |
| 2584 |
* applied to an update/delete target rel. If that ever becomes |
| 2585 |
* possible, we should drop the target from the PlanRowMark list. |
| 2586 |
*/ |
| 2587 |
Assert(rc->rti != parse->resultRelation); |
| 2588 |
|
| 2589 |
/* |
| 2590 |
* Ignore RowMarkClauses for subqueries; they aren't real tables and |
| 2591 |
* can't support true locking. Subqueries that got flattened into the |
| 2592 |
* main query should be ignored completely. Any that didn't will get |
| 2593 |
* ROW_MARK_COPY items in the next loop. |
| 2594 |
*/ |
| 2595 |
if (rte->rtekind != RTE_RELATION) |
| 2596 |
continue; |
| 2597 |
|
| 2598 |
rels = bms_del_member(rels, rc->rti); |
| 2599 |
|
| 2600 |
newrc = makeNode(PlanRowMark); |
| 2601 |
newrc->rti = newrc->prti = rc->rti; |
| 2602 |
newrc->rowmarkId = ++(root->glob->lastRowMarkId); |
| 2603 |
newrc->markType = select_rowmark_type(rte, rc->strength); |
| 2604 |
newrc->allMarkTypes = (1 << newrc->markType); |
| 2605 |
newrc->strength = rc->strength; |
| 2606 |
newrc->waitPolicy = rc->waitPolicy; |
| 2607 |
newrc->isParent = false; |
| 2608 |
|
| 2609 |
prowmarks = lappend(prowmarks, newrc); |
| 2610 |
} |
| 2611 |
|
| 2612 |
/* |
| 2613 |
* Now, add rowmarks for any non-target, non-locked base relations. |
| 2614 |
*/ |
| 2615 |
i = 0; |
| 2616 |
foreach(l, parse->rtable) |
| 2617 |
{ |
| 2618 |
RangeTblEntry *rte = lfirst_node(RangeTblEntry, l); |
| 2619 |
PlanRowMark *newrc; |
| 2620 |
|
| 2621 |
i++; |
| 2622 |
if (!bms_is_member(i, rels)) |
| 2623 |
continue; |
| 2624 |
|
| 2625 |
newrc = makeNode(PlanRowMark); |
| 2626 |
newrc->rti = newrc->prti = i; |
| 2627 |
newrc->rowmarkId = ++(root->glob->lastRowMarkId); |
| 2628 |
newrc->markType = select_rowmark_type(rte, LCS_NONE); |
| 2629 |
newrc->allMarkTypes = (1 << newrc->markType); |
| 2630 |
newrc->strength = LCS_NONE; |
| 2631 |
newrc->waitPolicy = LockWaitBlock; /* doesn't matter */ |
| 2632 |
newrc->isParent = false; |
| 2633 |
|
| 2634 |
prowmarks = lappend(prowmarks, newrc); |
| 2635 |
} |
| 2636 |
|
| 2637 |
root->rowMarks = prowmarks; |
| 2638 |
} |
| 2639 |
|
| 2640 |
/* |
| 2641 |
* Select RowMarkType to use for a given table |
| 2642 |
*/ |
| 2643 |
RowMarkType |
| 2644 |
select_rowmark_type(RangeTblEntry *rte, LockClauseStrength strength) |
| 2645 |
{ |
| 2646 |
if (rte->rtekind != RTE_RELATION) |
| 2647 |
{ |
| 2648 |
/* If it's not a table at all, use ROW_MARK_COPY */ |
| 2649 |
return ROW_MARK_COPY; |
| 2650 |
} |
| 2651 |
else if (rte->relkind == RELKIND_FOREIGN_TABLE) |
| 2652 |
{ |
| 2653 |
/* Let the FDW select the rowmark type, if it wants to */ |
| 2654 |
FdwRoutine *fdwroutine = GetFdwRoutineByRelId(rte->relid); |
| 2655 |
|
| 2656 |
if (fdwroutine->GetForeignRowMarkType != NULL) |
| 2657 |
return fdwroutine->GetForeignRowMarkType(rte, strength); |
| 2658 |
/* Otherwise, use ROW_MARK_COPY by default */ |
| 2659 |
return ROW_MARK_COPY; |
| 2660 |
} |
| 2661 |
else |
| 2662 |
{ |
| 2663 |
/* Regular table, apply the appropriate lock type */ |
| 2664 |
switch (strength) |
| 2665 |
{ |
| 2666 |
case LCS_NONE: |
| 2667 |
|
| 2668 |
/* |
| 2669 |
* We don't need a tuple lock, only the ability to re-fetch |
| 2670 |
* the row. |
| 2671 |
*/ |
| 2672 |
return ROW_MARK_REFERENCE; |
| 2673 |
break; |
| 2674 |
case LCS_FORKEYSHARE: |
| 2675 |
return ROW_MARK_KEYSHARE; |
| 2676 |
break; |
| 2677 |
case LCS_FORSHARE: |
| 2678 |
return ROW_MARK_SHARE; |
| 2679 |
break; |
| 2680 |
case LCS_FORNOKEYUPDATE: |
| 2681 |
return ROW_MARK_NOKEYEXCLUSIVE; |
| 2682 |
break; |
| 2683 |
case LCS_FORUPDATE: |
| 2684 |
return ROW_MARK_EXCLUSIVE; |
| 2685 |
break; |
| 2686 |
} |
| 2687 |
elog(ERROR, "unrecognized LockClauseStrength %d", (int) strength); |
| 2688 |
return ROW_MARK_EXCLUSIVE; /* keep compiler quiet */ |
| 2689 |
} |
| 2690 |
} |
| 2691 |
|
| 2692 |
/* |
| 2693 |
* preprocess_limit - do pre-estimation for LIMIT and/or OFFSET clauses |
| 2694 |
* |
| 2695 |
* We try to estimate the values of the LIMIT/OFFSET clauses, and pass the |
| 2696 |
* results back in *count_est and *offset_est. These variables are set to |
| 2697 |
* 0 if the corresponding clause is not present, and -1 if it's present |
| 2698 |
* but we couldn't estimate the value for it. (The "0" convention is OK |
| 2699 |
* for OFFSET but a little bit bogus for LIMIT: effectively we estimate |
| 2700 |
* LIMIT 0 as though it were LIMIT 1. But this is in line with the planner's |
| 2701 |
* usual practice of never estimating less than one row.) These values will |
| 2702 |
* be passed to create_limit_path, which see if you change this code. |
| 2703 |
* |
| 2704 |
* The return value is the suitably adjusted tuple_fraction to use for |
| 2705 |
* planning the query. This adjustment is not overridable, since it reflects |
| 2706 |
* plan actions that grouping_planner() will certainly take, not assumptions |
| 2707 |
* about context. |
| 2708 |
*/ |
| 2709 |
static double |
| 2710 |
preprocess_limit(PlannerInfo *root, double tuple_fraction, |
| 2711 |
int64 *offset_est, int64 *count_est) |
| 2712 |
{ |
| 2713 |
Query *parse = root->parse; |
| 2714 |
Node *est; |
| 2715 |
double limit_fraction; |
| 2716 |
|
| 2717 |
/* Should not be called unless LIMIT or OFFSET */ |
| 2718 |
Assert(parse->limitCount || parse->limitOffset); |
| 2719 |
|
| 2720 |
/* |
| 2721 |
* Try to obtain the clause values. We use estimate_expression_value |
| 2722 |
* primarily because it can sometimes do something useful with Params. |
| 2723 |
*/ |
| 2724 |
if (parse->limitCount) |
| 2725 |
{ |
| 2726 |
est = estimate_expression_value(root, parse->limitCount); |
| 2727 |
if (est && IsA(est, Const)) |
| 2728 |
{ |
| 2729 |
if (((Const *) est)->constisnull) |
| 2730 |
{ |
| 2731 |
/* NULL indicates LIMIT ALL, ie, no limit */ |
| 2732 |
*count_est = 0; /* treat as not present */ |
| 2733 |
} |
| 2734 |
else |
| 2735 |
{ |
| 2736 |
*count_est = DatumGetInt64(((Const *) est)->constvalue); |
| 2737 |
if (*count_est <= 0) |
| 2738 |
*count_est = 1; /* force to at least 1 */ |
| 2739 |
} |
| 2740 |
} |
| 2741 |
else |
| 2742 |
*count_est = -1; /* can't estimate */ |
| 2743 |
} |
| 2744 |
else |
| 2745 |
*count_est = 0; /* not present */ |
| 2746 |
|
| 2747 |
if (parse->limitOffset) |
| 2748 |
{ |
| 2749 |
est = estimate_expression_value(root, parse->limitOffset); |
| 2750 |
if (est && IsA(est, Const)) |
| 2751 |
{ |
| 2752 |
if (((Const *) est)->constisnull) |
| 2753 |
{ |
| 2754 |
/* Treat NULL as no offset; the executor will too */ |
| 2755 |
*offset_est = 0; /* treat as not present */ |
| 2756 |
} |
| 2757 |
else |
| 2758 |
{ |
| 2759 |
*offset_est = DatumGetInt64(((Const *) est)->constvalue); |
| 2760 |
if (*offset_est < 0) |
| 2761 |
*offset_est = 0; /* treat as not present */ |
| 2762 |
} |
| 2763 |
} |
| 2764 |
else |
| 2765 |
*offset_est = -1; /* can't estimate */ |
| 2766 |
} |
| 2767 |
else |
| 2768 |
*offset_est = 0; /* not present */ |
| 2769 |
|
| 2770 |
if (*count_est != 0) |
| 2771 |
{ |
| 2772 |
/* |
| 2773 |
* A LIMIT clause limits the absolute number of tuples returned. |
| 2774 |
* However, if it's not a constant LIMIT then we have to guess; for |
| 2775 |
* lack of a better idea, assume 10% of the plan's result is wanted. |
| 2776 |
*/ |
| 2777 |
if (*count_est < 0 || *offset_est < 0) |
| 2778 |
{ |
| 2779 |
/* LIMIT or OFFSET is an expression ... punt ... */ |
| 2780 |
limit_fraction = 0.10; |
| 2781 |
} |
| 2782 |
else |
| 2783 |
{ |
| 2784 |
/* LIMIT (plus OFFSET, if any) is max number of tuples needed */ |
| 2785 |
limit_fraction = (double) *count_est + (double) *offset_est; |
| 2786 |
} |
| 2787 |
|
| 2788 |
/* |
| 2789 |
* If we have absolute limits from both caller and LIMIT, use the |
| 2790 |
* smaller value; likewise if they are both fractional. If one is |
| 2791 |
* fractional and the other absolute, we can't easily determine which |
| 2792 |
* is smaller, but we use the heuristic that the absolute will usually |
| 2793 |
* be smaller. |
| 2794 |
*/ |
| 2795 |
if (tuple_fraction >= 1.0) |
| 2796 |
{ |
| 2797 |
if (limit_fraction >= 1.0) |
| 2798 |
{ |
| 2799 |
/* both absolute */ |
| 2800 |
tuple_fraction = Min(tuple_fraction, limit_fraction); |
| 2801 |
} |
| 2802 |
else |
| 2803 |
{ |
| 2804 |
/* caller absolute, limit fractional; use caller's value */ |
| 2805 |
} |
| 2806 |
} |
| 2807 |
else if (tuple_fraction > 0.0) |
| 2808 |
{ |
| 2809 |
if (limit_fraction >= 1.0) |
| 2810 |
{ |
| 2811 |
/* caller fractional, limit absolute; use limit */ |
| 2812 |
tuple_fraction = limit_fraction; |
| 2813 |
} |
| 2814 |
else |
| 2815 |
{ |
| 2816 |
/* both fractional */ |
| 2817 |
tuple_fraction = Min(tuple_fraction, limit_fraction); |
| 2818 |
} |
| 2819 |
} |
| 2820 |
else |
| 2821 |
{ |
| 2822 |
/* no info from caller, just use limit */ |
| 2823 |
tuple_fraction = limit_fraction; |
| 2824 |
} |
| 2825 |
} |
| 2826 |
else if (*offset_est != 0 && tuple_fraction > 0.0) |
| 2827 |
{ |
| 2828 |
/* |
| 2829 |
* We have an OFFSET but no LIMIT. This acts entirely differently |
| 2830 |
* from the LIMIT case: here, we need to increase rather than decrease |
| 2831 |
* the caller's tuple_fraction, because the OFFSET acts to cause more |
| 2832 |
* tuples to be fetched instead of fewer. This only matters if we got |
| 2833 |
* a tuple_fraction > 0, however. |
| 2834 |
* |
| 2835 |
* As above, use 10% if OFFSET is present but unestimatable. |
| 2836 |
*/ |
| 2837 |
if (*offset_est < 0) |
| 2838 |
limit_fraction = 0.10; |
| 2839 |
else |
| 2840 |
limit_fraction = (double) *offset_est; |
| 2841 |
|
| 2842 |
/* |
| 2843 |
* If we have absolute counts from both caller and OFFSET, add them |
| 2844 |
* together; likewise if they are both fractional. If one is |
| 2845 |
* fractional and the other absolute, we want to take the larger, and |
| 2846 |
* we heuristically assume that's the fractional one. |
| 2847 |
*/ |
| 2848 |
if (tuple_fraction >= 1.0) |
| 2849 |
{ |
| 2850 |
if (limit_fraction >= 1.0) |
| 2851 |
{ |
| 2852 |
/* both absolute, so add them together */ |
| 2853 |
tuple_fraction += limit_fraction; |
| 2854 |
} |
| 2855 |
else |
| 2856 |
{ |
| 2857 |
/* caller absolute, limit fractional; use limit */ |
| 2858 |
tuple_fraction = limit_fraction; |
| 2859 |
} |
| 2860 |
} |
| 2861 |
else |
| 2862 |
{ |
| 2863 |
if (limit_fraction >= 1.0) |
| 2864 |
{ |
| 2865 |
/* caller fractional, limit absolute; use caller's value */ |
| 2866 |
} |
| 2867 |
else |
| 2868 |
{ |
| 2869 |
/* both fractional, so add them together */ |
| 2870 |
tuple_fraction += limit_fraction; |
| 2871 |
if (tuple_fraction >= 1.0) |
| 2872 |
tuple_fraction = 0.0; /* assume fetch all */ |
| 2873 |
} |
| 2874 |
} |
| 2875 |
} |
| 2876 |
|
| 2877 |
return tuple_fraction; |
| 2878 |
} |
| 2879 |
|
| 2880 |
/* |
| 2881 |
* limit_needed - do we actually need a Limit plan node? |
| 2882 |
* |
| 2883 |
* If we have constant-zero OFFSET and constant-null LIMIT, we can skip adding |
| 2884 |
* a Limit node. This is worth checking for because "OFFSET 0" is a common |
| 2885 |
* locution for an optimization fence. (Because other places in the planner |
| 2886 |
* merely check whether parse->limitOffset isn't NULL, it will still work as |
| 2887 |
* an optimization fence --- we're just suppressing unnecessary run-time |
| 2888 |
* overhead.) |
| 2889 |
* |
| 2890 |
* This might look like it could be merged into preprocess_limit, but there's |
| 2891 |
* a key distinction: here we need hard constants in OFFSET/LIMIT, whereas |
| 2892 |
* in preprocess_limit it's good enough to consider estimated values. |
| 2893 |
*/ |
| 2894 |
bool |
| 2895 |
limit_needed(Query *parse) |
| 2896 |
{ |
| 2897 |
Node *node; |
| 2898 |
|
| 2899 |
node = parse->limitCount; |
| 2900 |
if (node) |
| 2901 |
{ |
| 2902 |
if (IsA(node, Const)) |
| 2903 |
{ |
| 2904 |
/* NULL indicates LIMIT ALL, ie, no limit */ |
| 2905 |
if (!((Const *) node)->constisnull) |
| 2906 |
return true; /* LIMIT with a constant value */ |
| 2907 |
} |
| 2908 |
else |
| 2909 |
return true; /* non-constant LIMIT */ |
| 2910 |
} |
| 2911 |
|
| 2912 |
node = parse->limitOffset; |
| 2913 |
if (node) |
| 2914 |
{ |
| 2915 |
if (IsA(node, Const)) |
| 2916 |
{ |
| 2917 |
/* Treat NULL as no offset; the executor would too */ |
| 2918 |
if (!((Const *) node)->constisnull) |
| 2919 |
{ |
| 2920 |
int64 offset = DatumGetInt64(((Const *) node)->constvalue); |
| 2921 |
|
| 2922 |
if (offset != 0) |
| 2923 |
return true; /* OFFSET with a nonzero value */ |
| 2924 |
} |
| 2925 |
} |
| 2926 |
else |
| 2927 |
return true; /* non-constant OFFSET */ |
| 2928 |
} |
| 2929 |
|
| 2930 |
return false; /* don't need a Limit plan node */ |
| 2931 |
} |
| 2932 |
|
| 2933 |
/* |
| 2934 |
* preprocess_groupclause - do preparatory work on GROUP BY clause |
| 2935 |
* |
| 2936 |
* The idea here is to adjust the ordering of the GROUP BY elements |
| 2937 |
* (which in itself is semantically insignificant) to match ORDER BY, |
| 2938 |
* thereby allowing a single sort operation to both implement the ORDER BY |
| 2939 |
* requirement and set up for a Unique step that implements GROUP BY. |
| 2940 |
* We also consider partial match between GROUP BY and ORDER BY elements, |
| 2941 |
* which could allow to implement ORDER BY using the incremental sort. |
| 2942 |
* |
| 2943 |
* We also consider other orderings of the GROUP BY elements, which could |
| 2944 |
* match the sort ordering of other possible plans (eg an indexscan) and |
| 2945 |
* thereby reduce cost. This is implemented during the generation of grouping |
| 2946 |
* paths. See get_useful_group_keys_orderings() for details. |
| 2947 |
* |
| 2948 |
* Note: we need no comparable processing of the distinctClause because |
| 2949 |
* the parser already enforced that that matches ORDER BY. |
| 2950 |
* |
| 2951 |
* Note: we return a fresh List, but its elements are the same |
| 2952 |
* SortGroupClauses appearing in parse->groupClause. This is important |
| 2953 |
* because later processing may modify the processed_groupClause list. |
| 2954 |
* |
| 2955 |
* For grouping sets, the order of items is instead forced to agree with that |
| 2956 |
* of the grouping set (and items not in the grouping set are skipped). The |
| 2957 |
* work of sorting the order of grouping set elements to match the ORDER BY if |
| 2958 |
* possible is done elsewhere. |
| 2959 |
*/ |
| 2960 |
static List * |
| 2961 |
preprocess_groupclause(PlannerInfo *root, List *force) |
| 2962 |
{ |
| 2963 |
Query *parse = root->parse; |
| 2964 |
List *new_groupclause = NIL; |
| 2965 |
ListCell *sl; |
| 2966 |
ListCell *gl; |
| 2967 |
|
| 2968 |
/* For grouping sets, we need to force the ordering */ |
| 2969 |
if (force) |
| 2970 |
{ |
| 2971 |
foreach(sl, force) |
| 2972 |
{ |
| 2973 |
Index ref = lfirst_int(sl); |
| 2974 |
SortGroupClause *cl = get_sortgroupref_clause(ref, parse->groupClause); |
| 2975 |
|
| 2976 |
new_groupclause = lappend(new_groupclause, cl); |
| 2977 |
} |
| 2978 |
|
| 2979 |
return new_groupclause; |
| 2980 |
} |
| 2981 |
|
| 2982 |
/* If no ORDER BY, nothing useful to do here */ |
| 2983 |
if (parse->sortClause == NIL) |
| 2984 |
return list_copy(parse->groupClause); |
| 2985 |
|
| 2986 |
/* |
| 2987 |
* Scan the ORDER BY clause and construct a list of matching GROUP BY |
| 2988 |
* items, but only as far as we can make a matching prefix. |
| 2989 |
* |
| 2990 |
* This code assumes that the sortClause contains no duplicate items. |
| 2991 |
*/ |
| 2992 |
foreach(sl, parse->sortClause) |
| 2993 |
{ |
| 2994 |
SortGroupClause *sc = lfirst_node(SortGroupClause, sl); |
| 2995 |
|
| 2996 |
foreach(gl, parse->groupClause) |
| 2997 |
{ |
| 2998 |
SortGroupClause *gc = lfirst_node(SortGroupClause, gl); |
| 2999 |
|
| 3000 |
if (equal(gc, sc)) |
| 3001 |
{ |
| 3002 |
new_groupclause = lappend(new_groupclause, gc); |
| 3003 |
break; |
| 3004 |
} |
| 3005 |
} |
| 3006 |
if (gl == NULL) |
| 3007 |
break; /* no match, so stop scanning */ |
| 3008 |
} |
| 3009 |
|
| 3010 |
|
| 3011 |
/* If no match at all, no point in reordering GROUP BY */ |
| 3012 |
if (new_groupclause == NIL) |
| 3013 |
return list_copy(parse->groupClause); |
| 3014 |
|
| 3015 |
/* |
| 3016 |
* Add any remaining GROUP BY items to the new list. We don't require a |
| 3017 |
* complete match, because even partial match allows ORDER BY to be |
| 3018 |
* implemented using incremental sort. Also, give up if there are any |
| 3019 |
* non-sortable GROUP BY items, since then there's no hope anyway. |
| 3020 |
*/ |
| 3021 |
foreach(gl, parse->groupClause) |
| 3022 |
{ |
| 3023 |
SortGroupClause *gc = lfirst_node(SortGroupClause, gl); |
| 3024 |
|
| 3025 |
if (list_member_ptr(new_groupclause, gc)) |
| 3026 |
continue; /* it matched an ORDER BY item */ |
| 3027 |
if (!OidIsValid(gc->sortop)) /* give up, GROUP BY can't be sorted */ |
| 3028 |
return list_copy(parse->groupClause); |
| 3029 |
new_groupclause = lappend(new_groupclause, gc); |
| 3030 |
} |
| 3031 |
|
| 3032 |
/* Success --- install the rearranged GROUP BY list */ |
| 3033 |
Assert(list_length(parse->groupClause) == list_length(new_groupclause)); |
| 3034 |
return new_groupclause; |
| 3035 |
} |
| 3036 |
|
| 3037 |
/* |
| 3038 |
* Extract lists of grouping sets that can be implemented using a single |
| 3039 |
* rollup-type aggregate pass each. Returns a list of lists of grouping sets. |
| 3040 |
* |
| 3041 |
* Input must be sorted with smallest sets first. Result has each sublist |
| 3042 |
* sorted with smallest sets first. |
| 3043 |
* |
| 3044 |
* We want to produce the absolute minimum possible number of lists here to |
| 3045 |
* avoid excess sorts. Fortunately, there is an algorithm for this; the problem |
| 3046 |
* of finding the minimal partition of a partially-ordered set into chains |
| 3047 |
* (which is what we need, taking the list of grouping sets as a poset ordered |
| 3048 |
* by set inclusion) can be mapped to the problem of finding the maximum |
| 3049 |
* cardinality matching on a bipartite graph, which is solvable in polynomial |
| 3050 |
* time with a worst case of no worse than O(n^2.5) and usually much |
| 3051 |
* better. Since our N is at most 4096, we don't need to consider fallbacks to |
| 3052 |
* heuristic or approximate methods. (Planning time for a 12-d cube is under |
| 3053 |
* half a second on my modest system even with optimization off and assertions |
| 3054 |
* on.) |
| 3055 |
*/ |
| 3056 |
static List * |
| 3057 |
extract_rollup_sets(List *groupingSets) |
| 3058 |
{ |
| 3059 |
int num_sets_raw = list_length(groupingSets); |
| 3060 |
int num_empty = 0; |
| 3061 |
int num_sets = 0; /* distinct sets */ |
| 3062 |
int num_chains = 0; |
| 3063 |
List *result = NIL; |
| 3064 |
List **results; |
| 3065 |
List **orig_sets; |
| 3066 |
Bitmapset **set_masks; |
| 3067 |
int *chains; |
| 3068 |
short **adjacency; |
| 3069 |
short *adjacency_buf; |
| 3070 |
BipartiteMatchState *state; |
| 3071 |
int i; |
| 3072 |
int j; |
| 3073 |
int j_size; |
| 3074 |
ListCell *lc1 = list_head(groupingSets); |
| 3075 |
ListCell *lc; |
| 3076 |
|
| 3077 |
/* |
| 3078 |
* Start by stripping out empty sets. The algorithm doesn't require this, |
| 3079 |
* but the planner currently needs all empty sets to be returned in the |
| 3080 |
* first list, so we strip them here and add them back after. |
| 3081 |
*/ |
| 3082 |
while (lc1 && lfirst(lc1) == NIL) |
| 3083 |
{ |
| 3084 |
++num_empty; |
| 3085 |
lc1 = lnext(groupingSets, lc1); |
| 3086 |
} |
| 3087 |
|
| 3088 |
/* bail out now if it turns out that all we had were empty sets. */ |
| 3089 |
if (!lc1) |
| 3090 |
return list_make1(groupingSets); |
| 3091 |
|
| 3092 |
/*---------- |
| 3093 |
* We don't strictly need to remove duplicate sets here, but if we don't, |
| 3094 |
* they tend to become scattered through the result, which is a bit |
| 3095 |
* confusing (and irritating if we ever decide to optimize them out). |
| 3096 |
* So we remove them here and add them back after. |
| 3097 |
* |
| 3098 |
* For each non-duplicate set, we fill in the following: |
| 3099 |
* |
| 3100 |
* orig_sets[i] = list of the original set lists |
| 3101 |
* set_masks[i] = bitmapset for testing inclusion |
| 3102 |
* adjacency[i] = array [n, v1, v2, ... vn] of adjacency indices |
| 3103 |
* |
| 3104 |
* chains[i] will be the result group this set is assigned to. |
| 3105 |
* |
| 3106 |
* We index all of these from 1 rather than 0 because it is convenient |
| 3107 |
* to leave 0 free for the NIL node in the graph algorithm. |
| 3108 |
*---------- |
| 3109 |
*/ |
| 3110 |
orig_sets = palloc0((num_sets_raw + 1) * sizeof(List *)); |
| 3111 |
set_masks = palloc0((num_sets_raw + 1) * sizeof(Bitmapset *)); |
| 3112 |
adjacency = palloc0((num_sets_raw + 1) * sizeof(short *)); |
| 3113 |
adjacency_buf = palloc((num_sets_raw + 1) * sizeof(short)); |
| 3114 |
|
| 3115 |
j_size = 0; |
| 3116 |
j = 0; |
| 3117 |
i = 1; |
| 3118 |
|
| 3119 |
for_each_cell(lc, groupingSets, lc1) |
| 3120 |
{ |
| 3121 |
List *candidate = (List *) lfirst(lc); |
| 3122 |
Bitmapset *candidate_set = NULL; |
| 3123 |
ListCell *lc2; |
| 3124 |
int dup_of = 0; |
| 3125 |
|
| 3126 |
foreach(lc2, candidate) |
| 3127 |
{ |
| 3128 |
candidate_set = bms_add_member(candidate_set, lfirst_int(lc2)); |
| 3129 |
} |
| 3130 |
|
| 3131 |
/* we can only be a dup if we're the same length as a previous set */ |
| 3132 |
if (j_size == list_length(candidate)) |
| 3133 |
{ |
| 3134 |
int k; |
| 3135 |
|
| 3136 |
for (k = j; k < i; ++k) |
| 3137 |
{ |
| 3138 |
if (bms_equal(set_masks[k], candidate_set)) |
| 3139 |
{ |
| 3140 |
dup_of = k; |
| 3141 |
break; |
| 3142 |
} |
| 3143 |
} |
| 3144 |
} |
| 3145 |
else if (j_size < list_length(candidate)) |
| 3146 |
{ |
| 3147 |
j_size = list_length(candidate); |
| 3148 |
j = i; |
| 3149 |
} |
| 3150 |
|
| 3151 |
if (dup_of > 0) |
| 3152 |
{ |
| 3153 |
orig_sets[dup_of] = lappend(orig_sets[dup_of], candidate); |
| 3154 |
bms_free(candidate_set); |
| 3155 |
} |
| 3156 |
else |
| 3157 |
{ |
| 3158 |
int k; |
| 3159 |
int n_adj = 0; |
| 3160 |
|
| 3161 |
orig_sets[i] = list_make1(candidate); |
| 3162 |
set_masks[i] = candidate_set; |
| 3163 |
|
| 3164 |
/* fill in adjacency list; no need to compare equal-size sets */ |
| 3165 |
|
| 3166 |
for (k = j - 1; k > 0; --k) |
| 3167 |
{ |
| 3168 |
if (bms_is_subset(set_masks[k], candidate_set)) |
| 3169 |
adjacency_buf[++n_adj] = k; |
| 3170 |
} |
| 3171 |
|
| 3172 |
if (n_adj > 0) |
| 3173 |
{ |
| 3174 |
adjacency_buf[0] = n_adj; |
| 3175 |
adjacency[i] = palloc((n_adj + 1) * sizeof(short)); |
| 3176 |
memcpy(adjacency[i], adjacency_buf, (n_adj + 1) * sizeof(short)); |
| 3177 |
} |
| 3178 |
else |
| 3179 |
adjacency[i] = NULL; |
| 3180 |
|
| 3181 |
++i; |
| 3182 |
} |
| 3183 |
} |
| 3184 |
|
| 3185 |
num_sets = i - 1; |
| 3186 |
|
| 3187 |
/* |
| 3188 |
* Apply the graph matching algorithm to do the work. |
| 3189 |
*/ |
| 3190 |
state = BipartiteMatch(num_sets, num_sets, adjacency); |
| 3191 |
|
| 3192 |
/* |
| 3193 |
* Now, the state->pair* fields have the info we need to assign sets to |
| 3194 |
* chains. Two sets (u,v) belong to the same chain if pair_uv[u] = v or |
| 3195 |
* pair_vu[v] = u (both will be true, but we check both so that we can do |
| 3196 |
* it in one pass) |
| 3197 |
*/ |
| 3198 |
chains = palloc0((num_sets + 1) * sizeof(int)); |
| 3199 |
|
| 3200 |
for (i = 1; i <= num_sets; ++i) |
| 3201 |
{ |
| 3202 |
int u = state->pair_vu[i]; |
| 3203 |
int v = state->pair_uv[i]; |
| 3204 |
|
| 3205 |
if (u > 0 && u < i) |
| 3206 |
chains[i] = chains[u]; |
| 3207 |
else if (v > 0 && v < i) |
| 3208 |
chains[i] = chains[v]; |
| 3209 |
else |
| 3210 |
chains[i] = ++num_chains; |
| 3211 |
} |
| 3212 |
|
| 3213 |
/* build result lists. */ |
| 3214 |
results = palloc0((num_chains + 1) * sizeof(List *)); |
| 3215 |
|
| 3216 |
for (i = 1; i <= num_sets; ++i) |
| 3217 |
{ |
| 3218 |
int c = chains[i]; |
| 3219 |
|
| 3220 |
Assert(c > 0); |
| 3221 |
|
| 3222 |
results[c] = list_concat(results[c], orig_sets[i]); |
| 3223 |
} |
| 3224 |
|
| 3225 |
/* push any empty sets back on the first list. */ |
| 3226 |
while (num_empty-- > 0) |
| 3227 |
results[1] = lcons(NIL, results[1]); |
| 3228 |
|
| 3229 |
/* make result list */ |
| 3230 |
for (i = 1; i <= num_chains; ++i) |
| 3231 |
result = lappend(result, results[i]); |
| 3232 |
|
| 3233 |
/* |
| 3234 |
* Free all the things. |
| 3235 |
* |
| 3236 |
* (This is over-fussy for small sets but for large sets we could have |
| 3237 |
* tied up a nontrivial amount of memory.) |
| 3238 |
*/ |
| 3239 |
BipartiteMatchFree(state); |
| 3240 |
pfree(results); |
| 3241 |
pfree(chains); |
| 3242 |
for (i = 1; i <= num_sets; ++i) |
| 3243 |
if (adjacency[i]) |
| 3244 |
pfree(adjacency[i]); |
| 3245 |
pfree(adjacency); |
| 3246 |
pfree(adjacency_buf); |
| 3247 |
pfree(orig_sets); |
| 3248 |
for (i = 1; i <= num_sets; ++i) |
| 3249 |
bms_free(set_masks[i]); |
| 3250 |
pfree(set_masks); |
| 3251 |
|
| 3252 |
return result; |
| 3253 |
} |
| 3254 |
|
| 3255 |
/* |
| 3256 |
* Reorder the elements of a list of grouping sets such that they have correct |
| 3257 |
* prefix relationships. Also inserts the GroupingSetData annotations. |
| 3258 |
* |
| 3259 |
* The input must be ordered with smallest sets first; the result is returned |
| 3260 |
* with largest sets first. Note that the result shares no list substructure |
| 3261 |
* with the input, so it's safe for the caller to modify it later. |
| 3262 |
* |
| 3263 |
* If we're passed in a sortclause, we follow its order of columns to the |
| 3264 |
* extent possible, to minimize the chance that we add unnecessary sorts. |
| 3265 |
* (We're trying here to ensure that GROUPING SETS ((a,b,c),(c)) ORDER BY c,b,a |
| 3266 |
* gets implemented in one pass.) |
| 3267 |
*/ |
| 3268 |
static List * |
| 3269 |
reorder_grouping_sets(List *groupingSets, List *sortclause) |
| 3270 |
{ |
| 3271 |
ListCell *lc; |
| 3272 |
List *previous = NIL; |
| 3273 |
List *result = NIL; |
| 3274 |
|
| 3275 |
foreach(lc, groupingSets) |
| 3276 |
{ |
| 3277 |
List *candidate = (List *) lfirst(lc); |
| 3278 |
List *new_elems = list_difference_int(candidate, previous); |
| 3279 |
GroupingSetData *gs = makeNode(GroupingSetData); |
| 3280 |
|
| 3281 |
while (list_length(sortclause) > list_length(previous) && |
| 3282 |
new_elems != NIL) |
| 3283 |
{ |
| 3284 |
SortGroupClause *sc = list_nth(sortclause, list_length(previous)); |
| 3285 |
int ref = sc->tleSortGroupRef; |
| 3286 |
|
| 3287 |
if (list_member_int(new_elems, ref)) |
| 3288 |
{ |
| 3289 |
previous = lappend_int(previous, ref); |
| 3290 |
new_elems = list_delete_int(new_elems, ref); |
| 3291 |
} |
| 3292 |
else |
| 3293 |
{ |
| 3294 |
/* diverged from the sortclause; give up on it */ |
| 3295 |
sortclause = NIL; |
| 3296 |
break; |
| 3297 |
} |
| 3298 |
} |
| 3299 |
|
| 3300 |
previous = list_concat(previous, new_elems); |
| 3301 |
|
| 3302 |
gs->set = list_copy(previous); |
| 3303 |
result = lcons(gs, result); |
| 3304 |
} |
| 3305 |
|
| 3306 |
list_free(previous); |
| 3307 |
|
| 3308 |
return result; |
| 3309 |
} |
| 3310 |
|
| 3311 |
/* |
| 3312 |
* has_volatile_pathkey |
| 3313 |
* Returns true if any PathKey in 'keys' has an EquivalenceClass |
| 3314 |
* containing a volatile function. Otherwise returns false. |
| 3315 |
*/ |
| 3316 |
static bool |
| 3317 |
has_volatile_pathkey(List *keys) |
| 3318 |
{ |
| 3319 |
ListCell *lc; |
| 3320 |
|
| 3321 |
foreach(lc, keys) |
| 3322 |
{ |
| 3323 |
PathKey *pathkey = lfirst_node(PathKey, lc); |
| 3324 |
|
| 3325 |
if (pathkey->pk_eclass->ec_has_volatile) |
| 3326 |
return true; |
| 3327 |
} |
| 3328 |
|
| 3329 |
return false; |
| 3330 |
} |
| 3331 |
|
| 3332 |
/* |
| 3333 |
* adjust_group_pathkeys_for_groupagg |
| 3334 |
* Add pathkeys to root->group_pathkeys to reflect the best set of |
| 3335 |
* pre-ordered input for ordered aggregates. |
| 3336 |
* |
| 3337 |
* We define "best" as the pathkeys that suit the largest number of |
| 3338 |
* aggregate functions. We find these by looking at the first ORDER BY / |
| 3339 |
* DISTINCT aggregate and take the pathkeys for that before searching for |
| 3340 |
* other aggregates that require the same or a more strict variation of the |
| 3341 |
* same pathkeys. We then repeat that process for any remaining aggregates |
| 3342 |
* with different pathkeys and if we find another set of pathkeys that suits a |
| 3343 |
* larger number of aggregates then we select those pathkeys instead. |
| 3344 |
* |
| 3345 |
* When the best pathkeys are found we also mark each Aggref that can use |
| 3346 |
* those pathkeys as aggpresorted = true. |
| 3347 |
* |
| 3348 |
* Note: When an aggregate function's ORDER BY / DISTINCT clause contains any |
| 3349 |
* volatile functions, we never make use of these pathkeys. We want to ensure |
| 3350 |
* that sorts using volatile functions are done independently in each Aggref |
| 3351 |
* rather than once at the query level. If we were to allow this then Aggrefs |
| 3352 |
* with compatible sort orders would all transition their rows in the same |
| 3353 |
* order if those pathkeys were deemed to be the best pathkeys to sort on. |
| 3354 |
* Whereas, if some other set of Aggref's pathkeys happened to be deemed |
| 3355 |
* better pathkeys to sort on, then the volatile function Aggrefs would be |
| 3356 |
* left to perform their sorts individually. To avoid this inconsistent |
| 3357 |
* behavior which could make Aggref results depend on what other Aggrefs the |
| 3358 |
* query contains, we always force Aggrefs with volatile functions to perform |
| 3359 |
* their own sorts. |
| 3360 |
*/ |
| 3361 |
static void |
| 3362 |
adjust_group_pathkeys_for_groupagg(PlannerInfo *root) |
| 3363 |
{ |
| 3364 |
List *grouppathkeys = root->group_pathkeys; |
| 3365 |
List *bestpathkeys; |
| 3366 |
Bitmapset *bestaggs; |
| 3367 |
Bitmapset *unprocessed_aggs; |
| 3368 |
ListCell *lc; |
| 3369 |
int i; |
| 3370 |
|
| 3371 |
/* Shouldn't be here if there are grouping sets */ |
| 3372 |
Assert(root->parse->groupingSets == NIL); |
| 3373 |
/* Shouldn't be here unless there are some ordered aggregates */ |
| 3374 |
Assert(root->numOrderedAggs > 0); |
| 3375 |
|
| 3376 |
/* Do nothing if disabled */ |
| 3377 |
if (!enable_presorted_aggregate) |
| 3378 |
return; |
| 3379 |
|
| 3380 |
/* |
| 3381 |
* Make a first pass over all AggInfos to collect a Bitmapset containing |
| 3382 |
* the indexes of all AggInfos to be processed below. |
| 3383 |
*/ |
| 3384 |
unprocessed_aggs = NULL; |
| 3385 |
foreach(lc, root->agginfos) |
| 3386 |
{ |
| 3387 |
AggInfo *agginfo = lfirst_node(AggInfo, lc); |
| 3388 |
Aggref *aggref = linitial_node(Aggref, agginfo->aggrefs); |
| 3389 |
|
| 3390 |
if (AGGKIND_IS_ORDERED_SET(aggref->aggkind)) |
| 3391 |
continue; |
| 3392 |
|
| 3393 |
/* Skip unless there's a DISTINCT or ORDER BY clause */ |
| 3394 |
if (aggref->aggdistinct == NIL && aggref->aggorder == NIL) |
| 3395 |
continue; |
| 3396 |
|
| 3397 |
/* Additional safety checks are needed if there's a FILTER clause */ |
| 3398 |
if (aggref->aggfilter != NULL) |
| 3399 |
{ |
| 3400 |
ListCell *lc2; |
| 3401 |
bool allow_presort = true; |
| 3402 |
|
| 3403 |
/* |
| 3404 |
* When the Aggref has a FILTER clause, it's possible that the |
| 3405 |
* filter removes rows that cannot be sorted because the |
| 3406 |
* expression to sort by results in an error during its |
| 3407 |
* evaluation. This is a problem for presorting as that happens |
| 3408 |
* before the FILTER, whereas without presorting, the Aggregate |
| 3409 |
* node will apply the FILTER *before* sorting. So that we never |
| 3410 |
* try to sort anything that might error, here we aim to skip over |
| 3411 |
* any Aggrefs with arguments with expressions which, when |
| 3412 |
* evaluated, could cause an ERROR. Vars and Consts are ok. There |
| 3413 |
* may be more cases that should be allowed, but more thought |
| 3414 |
* needs to be given. Err on the side of caution. |
| 3415 |
*/ |
| 3416 |
foreach(lc2, aggref->args) |
| 3417 |
{ |
| 3418 |
TargetEntry *tle = (TargetEntry *) lfirst(lc2); |
| 3419 |
Expr *expr = tle->expr; |
| 3420 |
|
| 3421 |
while (IsA(expr, RelabelType)) |
| 3422 |
expr = (Expr *) (castNode(RelabelType, expr))->arg; |
| 3423 |
|
| 3424 |
/* Common case, Vars and Consts are ok */ |
| 3425 |
if (IsA(expr, Var) || IsA(expr, Const)) |
| 3426 |
continue; |
| 3427 |
|
| 3428 |
/* Unsupported. Don't try to presort for this Aggref */ |
| 3429 |
allow_presort = false; |
| 3430 |
break; |
| 3431 |
} |
| 3432 |
|
| 3433 |
/* Skip unsupported Aggrefs */ |
| 3434 |
if (!allow_presort) |
| 3435 |
continue; |
| 3436 |
} |
| 3437 |
|
| 3438 |
unprocessed_aggs = bms_add_member(unprocessed_aggs, |
| 3439 |
foreach_current_index(lc)); |
| 3440 |
} |
| 3441 |
|
| 3442 |
/* |
| 3443 |
* Now process all the unprocessed_aggs to find the best set of pathkeys |
| 3444 |
* for the given set of aggregates. |
| 3445 |
* |
| 3446 |
* On the first outer loop here 'bestaggs' will be empty. We'll populate |
| 3447 |
* this during the first loop using the pathkeys for the very first |
| 3448 |
* AggInfo then taking any stronger pathkeys from any other AggInfos with |
| 3449 |
* a more strict set of compatible pathkeys. Once the outer loop is |
| 3450 |
* complete, we mark off all the aggregates with compatible pathkeys then |
| 3451 |
* remove those from the unprocessed_aggs and repeat the process to try to |
| 3452 |
* find another set of pathkeys that are suitable for a larger number of |
| 3453 |
* aggregates. The outer loop will stop when there are not enough |
| 3454 |
* unprocessed aggregates for it to be possible to find a set of pathkeys |
| 3455 |
* to suit a larger number of aggregates. |
| 3456 |
*/ |
| 3457 |
bestpathkeys = NIL; |
| 3458 |
bestaggs = NULL; |
| 3459 |
while (bms_num_members(unprocessed_aggs) > bms_num_members(bestaggs)) |
| 3460 |
{ |
| 3461 |
Bitmapset *aggindexes = NULL; |
| 3462 |
List *currpathkeys = NIL; |
| 3463 |
|
| 3464 |
i = -1; |
| 3465 |
while ((i = bms_next_member(unprocessed_aggs, i)) >= 0) |
| 3466 |
{ |
| 3467 |
AggInfo *agginfo = list_nth_node(AggInfo, root->agginfos, i); |
| 3468 |
Aggref *aggref = linitial_node(Aggref, agginfo->aggrefs); |
| 3469 |
List *sortlist; |
| 3470 |
List *pathkeys; |
| 3471 |
|
| 3472 |
if (aggref->aggdistinct != NIL) |
| 3473 |
sortlist = aggref->aggdistinct; |
| 3474 |
else |
| 3475 |
sortlist = aggref->aggorder; |
| 3476 |
|
| 3477 |
pathkeys = make_pathkeys_for_sortclauses(root, sortlist, |
| 3478 |
aggref->args); |
| 3479 |
|
| 3480 |
/* |
| 3481 |
* Ignore Aggrefs which have volatile functions in their ORDER BY |
| 3482 |
* or DISTINCT clause. |
| 3483 |
*/ |
| 3484 |
if (has_volatile_pathkey(pathkeys)) |
| 3485 |
{ |
| 3486 |
unprocessed_aggs = bms_del_member(unprocessed_aggs, i); |
| 3487 |
continue; |
| 3488 |
} |
| 3489 |
|
| 3490 |
/* |
| 3491 |
* When not set yet, take the pathkeys from the first unprocessed |
| 3492 |
* aggregate. |
| 3493 |
*/ |
| 3494 |
if (currpathkeys == NIL) |
| 3495 |
{ |
| 3496 |
currpathkeys = pathkeys; |
| 3497 |
|
| 3498 |
/* include the GROUP BY pathkeys, if they exist */ |
| 3499 |
if (grouppathkeys != NIL) |
| 3500 |
currpathkeys = append_pathkeys(list_copy(grouppathkeys), |
| 3501 |
currpathkeys); |
| 3502 |
|
| 3503 |
/* record that we found pathkeys for this aggregate */ |
| 3504 |
aggindexes = bms_add_member(aggindexes, i); |
| 3505 |
} |
| 3506 |
else |
| 3507 |
{ |
| 3508 |
/* now look for a stronger set of matching pathkeys */ |
| 3509 |
|
| 3510 |
/* include the GROUP BY pathkeys, if they exist */ |
| 3511 |
if (grouppathkeys != NIL) |
| 3512 |
pathkeys = append_pathkeys(list_copy(grouppathkeys), |
| 3513 |
pathkeys); |
| 3514 |
|
| 3515 |
/* are 'pathkeys' compatible or better than 'currpathkeys'? */ |
| 3516 |
switch (compare_pathkeys(currpathkeys, pathkeys)) |
| 3517 |
{ |
| 3518 |
case PATHKEYS_BETTER2: |
| 3519 |
/* 'pathkeys' are stronger, use these ones instead */ |
| 3520 |
currpathkeys = pathkeys; |
| 3521 |
/* FALLTHROUGH */ |
| 3522 |
|
| 3523 |
case PATHKEYS_BETTER1: |
| 3524 |
/* 'pathkeys' are less strict */ |
| 3525 |
/* FALLTHROUGH */ |
| 3526 |
|
| 3527 |
case PATHKEYS_EQUAL: |
| 3528 |
/* mark this aggregate as covered by 'currpathkeys' */ |
| 3529 |
aggindexes = bms_add_member(aggindexes, i); |
| 3530 |
break; |
| 3531 |
|
| 3532 |
case PATHKEYS_DIFFERENT: |
| 3533 |
break; |
| 3534 |
} |
| 3535 |
} |
| 3536 |
} |
| 3537 |
|
| 3538 |
/* remove the aggregates that we've just processed */ |
| 3539 |
unprocessed_aggs = bms_del_members(unprocessed_aggs, aggindexes); |
| 3540 |
|
| 3541 |
/* |
| 3542 |
* If this pass included more aggregates than the previous best then |
| 3543 |
* use these ones as the best set. |
| 3544 |
*/ |
| 3545 |
if (bms_num_members(aggindexes) > bms_num_members(bestaggs)) |
| 3546 |
{ |
| 3547 |
bestaggs = aggindexes; |
| 3548 |
bestpathkeys = currpathkeys; |
| 3549 |
} |
| 3550 |
} |
| 3551 |
|
| 3552 |
/* |
| 3553 |
* If we found any ordered aggregates, update root->group_pathkeys to add |
| 3554 |
* the best set of aggregate pathkeys. Note that bestpathkeys includes |
| 3555 |
* the original GROUP BY pathkeys already. |
| 3556 |
*/ |
| 3557 |
if (bestpathkeys != NIL) |
| 3558 |
root->group_pathkeys = bestpathkeys; |
| 3559 |
|
| 3560 |
/* |
| 3561 |
* Now that we've found the best set of aggregates we can set the |
| 3562 |
* presorted flag to indicate to the executor that it needn't bother |
| 3563 |
* performing a sort for these Aggrefs. We're able to do this now as |
| 3564 |
* there's no chance of a Hash Aggregate plan as create_grouping_paths |
| 3565 |
* will not mark the GROUP BY as GROUPING_CAN_USE_HASH due to the presence |
| 3566 |
* of ordered aggregates. |
| 3567 |
*/ |
| 3568 |
i = -1; |
| 3569 |
while ((i = bms_next_member(bestaggs, i)) >= 0) |
| 3570 |
{ |
| 3571 |
AggInfo *agginfo = list_nth_node(AggInfo, root->agginfos, i); |
| 3572 |
|
| 3573 |
foreach(lc, agginfo->aggrefs) |
| 3574 |
{ |
| 3575 |
Aggref *aggref = lfirst_node(Aggref, lc); |
| 3576 |
|
| 3577 |
aggref->aggpresorted = true; |
| 3578 |
} |
| 3579 |
} |
| 3580 |
} |
| 3581 |
|
| 3582 |
/* |
| 3583 |
* Compute query_pathkeys and other pathkeys during plan generation |
| 3584 |
*/ |
| 3585 |
static void |
| 3586 |
standard_qp_callback(PlannerInfo *root, void *extra) |
| 3587 |
{ |
| 3588 |
Query *parse = root->parse; |
| 3589 |
standard_qp_extra *qp_extra = (standard_qp_extra *) extra; |
| 3590 |
List *tlist = root->processed_tlist; |
| 3591 |
List *activeWindows = qp_extra->activeWindows; |
| 3592 |
|
| 3593 |
/* |
| 3594 |
* Calculate pathkeys that represent grouping/ordering and/or ordered |
| 3595 |
* aggregate requirements. |
| 3596 |
*/ |
| 3597 |
if (qp_extra->gset_data) |
| 3598 |
{ |
| 3599 |
/* |
| 3600 |
* With grouping sets, just use the first RollupData's groupClause. We |
| 3601 |
* don't make any effort to optimize grouping clauses when there are |
| 3602 |
* grouping sets, nor can we combine aggregate ordering keys with |
| 3603 |
* grouping. |
| 3604 |
*/ |
| 3605 |
List *rollups = qp_extra->gset_data->rollups; |
| 3606 |
List *groupClause = (rollups ? linitial_node(RollupData, rollups)->groupClause : NIL); |
| 3607 |
|
| 3608 |
if (grouping_is_sortable(groupClause)) |
| 3609 |
{ |
| 3610 |
bool sortable; |
| 3611 |
|
| 3612 |
/* |
| 3613 |
* The groupClause is logically below the grouping step. So if |
| 3614 |
* there is an RTE entry for the grouping step, we need to remove |
| 3615 |
* its RT index from the sort expressions before we make PathKeys |
| 3616 |
* for them. |
| 3617 |
*/ |
| 3618 |
root->group_pathkeys = |
| 3619 |
make_pathkeys_for_sortclauses_extended(root, |
| 3620 |
&groupClause, |
| 3621 |
tlist, |
| 3622 |
false, |
| 3623 |
parse->hasGroupRTE, |
| 3624 |
&sortable, |
| 3625 |
false); |
| 3626 |
Assert(sortable); |
| 3627 |
root->num_groupby_pathkeys = list_length(root->group_pathkeys); |
| 3628 |
} |
| 3629 |
else |
| 3630 |
{ |
| 3631 |
root->group_pathkeys = NIL; |
| 3632 |
root->num_groupby_pathkeys = 0; |
| 3633 |
} |
| 3634 |
} |
| 3635 |
else if (parse->groupClause || root->numOrderedAggs > 0) |
| 3636 |
{ |
| 3637 |
/* |
| 3638 |
* With a plain GROUP BY list, we can remove any grouping items that |
| 3639 |
* are proven redundant by EquivalenceClass processing. For example, |
| 3640 |
* we can remove y given "WHERE x = y GROUP BY x, y". These aren't |
| 3641 |
* especially common cases, but they're nearly free to detect. Note |
| 3642 |
* that we remove redundant items from processed_groupClause but not |
| 3643 |
* the original parse->groupClause. |
| 3644 |
*/ |
| 3645 |
bool sortable; |
| 3646 |
|
| 3647 |
/* |
| 3648 |
* Convert group clauses into pathkeys. Set the ec_sortref field of |
| 3649 |
* EquivalenceClass'es if it's not set yet. |
| 3650 |
*/ |
| 3651 |
root->group_pathkeys = |
| 3652 |
make_pathkeys_for_sortclauses_extended(root, |
| 3653 |
&root->processed_groupClause, |
| 3654 |
tlist, |
| 3655 |
true, |
| 3656 |
false, |
| 3657 |
&sortable, |
| 3658 |
true); |
| 3659 |
if (!sortable) |
| 3660 |
{ |
| 3661 |
/* Can't sort; no point in considering aggregate ordering either */ |
| 3662 |
root->group_pathkeys = NIL; |
| 3663 |
root->num_groupby_pathkeys = 0; |
| 3664 |
} |
| 3665 |
else |
| 3666 |
{ |
| 3667 |
root->num_groupby_pathkeys = list_length(root->group_pathkeys); |
| 3668 |
/* If we have ordered aggs, consider adding onto group_pathkeys */ |
| 3669 |
if (root->numOrderedAggs > 0) |
| 3670 |
adjust_group_pathkeys_for_groupagg(root); |
| 3671 |
} |
| 3672 |
} |
| 3673 |
else |
| 3674 |
{ |
| 3675 |
root->group_pathkeys = NIL; |
| 3676 |
root->num_groupby_pathkeys = 0; |
| 3677 |
} |
| 3678 |
|
| 3679 |
/* We consider only the first (bottom) window in pathkeys logic */ |
| 3680 |
if (activeWindows != NIL) |
| 3681 |
{ |
| 3682 |
WindowClause *wc = linitial_node(WindowClause, activeWindows); |
| 3683 |
|
| 3684 |
root->window_pathkeys = make_pathkeys_for_window(root, |
| 3685 |
wc, |
| 3686 |
tlist); |
| 3687 |
} |
| 3688 |
else |
| 3689 |
root->window_pathkeys = NIL; |
| 3690 |
|
| 3691 |
/* |
| 3692 |
* As with GROUP BY, we can discard any DISTINCT items that are proven |
| 3693 |
* redundant by EquivalenceClass processing. The non-redundant list is |
| 3694 |
* kept in root->processed_distinctClause, leaving the original |
| 3695 |
* parse->distinctClause alone. |
| 3696 |
*/ |
| 3697 |
if (parse->distinctClause) |
| 3698 |
{ |
| 3699 |
bool sortable; |
| 3700 |
|
| 3701 |
/* Make a copy since pathkey processing can modify the list */ |
| 3702 |
root->processed_distinctClause = list_copy(parse->distinctClause); |
| 3703 |
root->distinct_pathkeys = |
| 3704 |
make_pathkeys_for_sortclauses_extended(root, |
| 3705 |
&root->processed_distinctClause, |
| 3706 |
tlist, |
| 3707 |
true, |
| 3708 |
false, |
| 3709 |
&sortable, |
| 3710 |
false); |
| 3711 |
if (!sortable) |
| 3712 |
root->distinct_pathkeys = NIL; |
| 3713 |
} |
| 3714 |
else |
| 3715 |
root->distinct_pathkeys = NIL; |
| 3716 |
|
| 3717 |
root->sort_pathkeys = |
| 3718 |
make_pathkeys_for_sortclauses(root, |
| 3719 |
parse->sortClause, |
| 3720 |
tlist); |
| 3721 |
|
| 3722 |
/* setting setop_pathkeys might be useful to the union planner */ |
| 3723 |
if (qp_extra->setop != NULL) |
| 3724 |
{ |
| 3725 |
List *groupClauses; |
| 3726 |
bool sortable; |
| 3727 |
|
| 3728 |
groupClauses = generate_setop_child_grouplist(qp_extra->setop, tlist); |
| 3729 |
|
| 3730 |
root->setop_pathkeys = |
| 3731 |
make_pathkeys_for_sortclauses_extended(root, |
| 3732 |
&groupClauses, |
| 3733 |
tlist, |
| 3734 |
false, |
| 3735 |
false, |
| 3736 |
&sortable, |
| 3737 |
false); |
| 3738 |
if (!sortable) |
| 3739 |
root->setop_pathkeys = NIL; |
| 3740 |
} |
| 3741 |
else |
| 3742 |
root->setop_pathkeys = NIL; |
| 3743 |
|
| 3744 |
/* |
| 3745 |
* Figure out whether we want a sorted result from query_planner. |
| 3746 |
* |
| 3747 |
* If we have a sortable GROUP BY clause, then we want a result sorted |
| 3748 |
* properly for grouping. Otherwise, if we have window functions to |
| 3749 |
* evaluate, we try to sort for the first window. Otherwise, if there's a |
| 3750 |
* sortable DISTINCT clause that's more rigorous than the ORDER BY clause, |
| 3751 |
* we try to produce output that's sufficiently well sorted for the |
| 3752 |
* DISTINCT. Otherwise, if there is an ORDER BY clause, we want to sort |
| 3753 |
* by the ORDER BY clause. Otherwise, if we're a subquery being planned |
| 3754 |
* for a set operation which can benefit from presorted results and have a |
| 3755 |
* sortable targetlist, we want to sort by the target list. |
| 3756 |
* |
| 3757 |
* Note: if we have both ORDER BY and GROUP BY, and ORDER BY is a superset |
| 3758 |
* of GROUP BY, it would be tempting to request sort by ORDER BY --- but |
| 3759 |
* that might just leave us failing to exploit an available sort order at |
| 3760 |
* all. Needs more thought. The choice for DISTINCT versus ORDER BY is |
| 3761 |
* much easier, since we know that the parser ensured that one is a |
| 3762 |
* superset of the other. |
| 3763 |
*/ |
| 3764 |
if (root->group_pathkeys) |
| 3765 |
root->query_pathkeys = root->group_pathkeys; |
| 3766 |
else if (root->window_pathkeys) |
| 3767 |
root->query_pathkeys = root->window_pathkeys; |
| 3768 |
else if (list_length(root->distinct_pathkeys) > |
| 3769 |
list_length(root->sort_pathkeys)) |
| 3770 |
root->query_pathkeys = root->distinct_pathkeys; |
| 3771 |
else if (root->sort_pathkeys) |
| 3772 |
root->query_pathkeys = root->sort_pathkeys; |
| 3773 |
else if (root->setop_pathkeys != NIL) |
| 3774 |
root->query_pathkeys = root->setop_pathkeys; |
| 3775 |
else |
| 3776 |
root->query_pathkeys = NIL; |
| 3777 |
} |
| 3778 |
|
| 3779 |
/* |
| 3780 |
* Estimate number of groups produced by grouping clauses (1 if not grouping) |
| 3781 |
* |
| 3782 |
* path_rows: number of output rows from scan/join step |
| 3783 |
* gd: grouping sets data including list of grouping sets and their clauses |
| 3784 |
* target_list: target list containing group clause references |
| 3785 |
* |
| 3786 |
* If doing grouping sets, we also annotate the gsets data with the estimates |
| 3787 |
* for each set and each individual rollup list, with a view to later |
| 3788 |
* determining whether some combination of them could be hashed instead. |
| 3789 |
*/ |
| 3790 |
static double |
| 3791 |
get_number_of_groups(PlannerInfo *root, |
| 3792 |
double path_rows, |
| 3793 |
grouping_sets_data *gd, |
| 3794 |
List *target_list) |
| 3795 |
{ |
| 3796 |
Query *parse = root->parse; |
| 3797 |
double dNumGroups; |
| 3798 |
|
| 3799 |
if (parse->groupClause) |
| 3800 |
{ |
| 3801 |
List *groupExprs; |
| 3802 |
|
| 3803 |
if (parse->groupingSets) |
| 3804 |
{ |
| 3805 |
/* Add up the estimates for each grouping set */ |
| 3806 |
ListCell *lc; |
| 3807 |
|
| 3808 |
Assert(gd); /* keep Coverity happy */ |
| 3809 |
|
| 3810 |
dNumGroups = 0; |
| 3811 |
|
| 3812 |
foreach(lc, gd->rollups) |
| 3813 |
{ |
| 3814 |
RollupData *rollup = lfirst_node(RollupData, lc); |
| 3815 |
ListCell *lc2; |
| 3816 |
ListCell *lc3; |
| 3817 |
|
| 3818 |
groupExprs = get_sortgrouplist_exprs(rollup->groupClause, |
| 3819 |
target_list); |
| 3820 |
|
| 3821 |
rollup->numGroups = 0.0; |
| 3822 |
|
| 3823 |
forboth(lc2, rollup->gsets, lc3, rollup->gsets_data) |
| 3824 |
{ |
| 3825 |
List *gset = (List *) lfirst(lc2); |
| 3826 |
GroupingSetData *gs = lfirst_node(GroupingSetData, lc3); |
| 3827 |
double numGroups = estimate_num_groups(root, |
| 3828 |
groupExprs, |
| 3829 |
path_rows, |
| 3830 |
&gset, |
| 3831 |
NULL); |
| 3832 |
|
| 3833 |
gs->numGroups = numGroups; |
| 3834 |
rollup->numGroups += numGroups; |
| 3835 |
} |
| 3836 |
|
| 3837 |
dNumGroups += rollup->numGroups; |
| 3838 |
} |
| 3839 |
|
| 3840 |
if (gd->hash_sets_idx) |
| 3841 |
{ |
| 3842 |
ListCell *lc2; |
| 3843 |
|
| 3844 |
gd->dNumHashGroups = 0; |
| 3845 |
|
| 3846 |
groupExprs = get_sortgrouplist_exprs(parse->groupClause, |
| 3847 |
target_list); |
| 3848 |
|
| 3849 |
forboth(lc, gd->hash_sets_idx, lc2, gd->unsortable_sets) |
| 3850 |
{ |
| 3851 |
List *gset = (List *) lfirst(lc); |
| 3852 |
GroupingSetData *gs = lfirst_node(GroupingSetData, lc2); |
| 3853 |
double numGroups = estimate_num_groups(root, |
| 3854 |
groupExprs, |
| 3855 |
path_rows, |
| 3856 |
&gset, |
| 3857 |
NULL); |
| 3858 |
|
| 3859 |
gs->numGroups = numGroups; |
| 3860 |
gd->dNumHashGroups += numGroups; |
| 3861 |
} |
| 3862 |
|
| 3863 |
dNumGroups += gd->dNumHashGroups; |
| 3864 |
} |
| 3865 |
} |
| 3866 |
else |
| 3867 |
{ |
| 3868 |
/* Plain GROUP BY -- estimate based on optimized groupClause */ |
| 3869 |
groupExprs = get_sortgrouplist_exprs(root->processed_groupClause, |
| 3870 |
target_list); |
| 3871 |
|
| 3872 |
dNumGroups = estimate_num_groups(root, groupExprs, path_rows, |
| 3873 |
NULL, NULL); |
| 3874 |
} |
| 3875 |
} |
| 3876 |
else if (parse->groupingSets) |
| 3877 |
{ |
| 3878 |
/* Empty grouping sets ... one result row for each one */ |
| 3879 |
dNumGroups = list_length(parse->groupingSets); |
| 3880 |
} |
| 3881 |
else if (parse->hasAggs || root->hasHavingQual) |
| 3882 |
{ |
| 3883 |
/* Plain aggregation, one result row */ |
| 3884 |
dNumGroups = 1; |
| 3885 |
} |
| 3886 |
else |
| 3887 |
{ |
| 3888 |
/* Not grouping */ |
| 3889 |
dNumGroups = 1; |
| 3890 |
} |
| 3891 |
|
| 3892 |
return dNumGroups; |
| 3893 |
} |
| 3894 |
|
| 3895 |
/* |
| 3896 |
* create_grouping_paths |
| 3897 |
* |
| 3898 |
* Build a new upperrel containing Paths for grouping and/or aggregation. |
| 3899 |
* Along the way, we also build an upperrel for Paths which are partially |
| 3900 |
* grouped and/or aggregated. A partially grouped and/or aggregated path |
| 3901 |
* needs a FinalizeAggregate node to complete the aggregation. Currently, |
| 3902 |
* the only partially grouped paths we build are also partial paths; that |
| 3903 |
* is, they need a Gather and then a FinalizeAggregate. |
| 3904 |
* |
| 3905 |
* input_rel: contains the source-data Paths |
| 3906 |
* target: the pathtarget for the result Paths to compute |
| 3907 |
* gd: grouping sets data including list of grouping sets and their clauses |
| 3908 |
* |
| 3909 |
* Note: all Paths in input_rel are expected to return the target computed |
| 3910 |
* by make_group_input_target. |
| 3911 |
*/ |
| 3912 |
static RelOptInfo * |
| 3913 |
create_grouping_paths(PlannerInfo *root, |
| 3914 |
RelOptInfo *input_rel, |
| 3915 |
PathTarget *target, |
| 3916 |
bool target_parallel_safe, |
| 3917 |
grouping_sets_data *gd) |
| 3918 |
{ |
| 3919 |
Query *parse = root->parse; |
| 3920 |
RelOptInfo *grouped_rel; |
| 3921 |
RelOptInfo *partially_grouped_rel; |
| 3922 |
AggClauseCosts agg_costs; |
| 3923 |
|
| 3924 |
MemSet(&agg_costs, 0, sizeof(AggClauseCosts)); |
| 3925 |
get_agg_clause_costs(root, AGGSPLIT_SIMPLE, &agg_costs); |
| 3926 |
|
| 3927 |
/* |
| 3928 |
* Create grouping relation to hold fully aggregated grouping and/or |
| 3929 |
* aggregation paths. |
| 3930 |
*/ |
| 3931 |
grouped_rel = make_grouping_rel(root, input_rel, target, |
| 3932 |
target_parallel_safe, parse->havingQual); |
| 3933 |
|
| 3934 |
/* |
| 3935 |
* Create either paths for a degenerate grouping or paths for ordinary |
| 3936 |
* grouping, as appropriate. |
| 3937 |
*/ |
| 3938 |
if (is_degenerate_grouping(root)) |
| 3939 |
create_degenerate_grouping_paths(root, input_rel, grouped_rel); |
| 3940 |
else |
| 3941 |
{ |
| 3942 |
int flags = 0; |
| 3943 |
GroupPathExtraData extra; |
| 3944 |
|
| 3945 |
/* |
| 3946 |
* Determine whether it's possible to perform sort-based |
| 3947 |
* implementations of grouping. (Note that if processed_groupClause |
| 3948 |
* is empty, grouping_is_sortable() is trivially true, and all the |
| 3949 |
* pathkeys_contained_in() tests will succeed too, so that we'll |
| 3950 |
* consider every surviving input path.) |
| 3951 |
* |
| 3952 |
* If we have grouping sets, we might be able to sort some but not all |
| 3953 |
* of them; in this case, we need can_sort to be true as long as we |
| 3954 |
* must consider any sorted-input plan. |
| 3955 |
*/ |
| 3956 |
if ((gd && gd->rollups != NIL) |
| 3957 |
|| grouping_is_sortable(root->processed_groupClause)) |
| 3958 |
flags |= GROUPING_CAN_USE_SORT; |
| 3959 |
|
| 3960 |
/* |
| 3961 |
* Determine whether we should consider hash-based implementations of |
| 3962 |
* grouping. |
| 3963 |
* |
| 3964 |
* Hashed aggregation only applies if we're grouping. If we have |
| 3965 |
* grouping sets, some groups might be hashable but others not; in |
| 3966 |
* this case we set can_hash true as long as there is nothing globally |
| 3967 |
* preventing us from hashing (and we should therefore consider plans |
| 3968 |
* with hashes). |
| 3969 |
* |
| 3970 |
* Executor doesn't support hashed aggregation with DISTINCT or ORDER |
| 3971 |
* BY aggregates. (Doing so would imply storing *all* the input |
| 3972 |
* values in the hash table, and/or running many sorts in parallel, |
| 3973 |
* either of which seems like a certain loser.) We similarly don't |
| 3974 |
* support ordered-set aggregates in hashed aggregation, but that case |
| 3975 |
* is also included in the numOrderedAggs count. |
| 3976 |
* |
| 3977 |
* Note: grouping_is_hashable() is much more expensive to check than |
| 3978 |
* the other gating conditions, so we want to do it last. |
| 3979 |
*/ |
| 3980 |
if ((parse->groupClause != NIL && |
| 3981 |
root->numOrderedAggs == 0 && |
| 3982 |
(gd ? gd->any_hashable : grouping_is_hashable(root->processed_groupClause)))) |
| 3983 |
flags |= GROUPING_CAN_USE_HASH; |
| 3984 |
|
| 3985 |
/* |
| 3986 |
* Determine whether partial aggregation is possible. |
| 3987 |
*/ |
| 3988 |
if (can_partial_agg(root)) |
| 3989 |
flags |= GROUPING_CAN_PARTIAL_AGG; |
| 3990 |
|
| 3991 |
extra.flags = flags; |
| 3992 |
extra.target_parallel_safe = target_parallel_safe; |
| 3993 |
extra.havingQual = parse->havingQual; |
| 3994 |
extra.targetList = parse->targetList; |
| 3995 |
extra.partial_costs_set = false; |
| 3996 |
|
| 3997 |
/* |
| 3998 |
* Determine whether partitionwise aggregation is in theory possible. |
| 3999 |
* It can be disabled by the user, and for now, we don't try to |
| 4000 |
* support grouping sets. create_ordinary_grouping_paths() will check |
| 4001 |
* additional conditions, such as whether input_rel is partitioned. |
| 4002 |
*/ |
| 4003 |
if (enable_partitionwise_aggregate && !parse->groupingSets) |
| 4004 |
extra.patype = PARTITIONWISE_AGGREGATE_FULL; |
| 4005 |
else |
| 4006 |
extra.patype = PARTITIONWISE_AGGREGATE_NONE; |
| 4007 |
|
| 4008 |
create_ordinary_grouping_paths(root, input_rel, grouped_rel, |
| 4009 |
&agg_costs, gd, &extra, |
| 4010 |
&partially_grouped_rel); |
| 4011 |
} |
| 4012 |
|
| 4013 |
set_cheapest(grouped_rel); |
| 4014 |
return grouped_rel; |
| 4015 |
} |
| 4016 |
|
| 4017 |
/* |
| 4018 |
* make_grouping_rel |
| 4019 |
* |
| 4020 |
* Create a new grouping rel and set basic properties. |
| 4021 |
* |
| 4022 |
* input_rel represents the underlying scan/join relation. |
| 4023 |
* target is the output expected from the grouping relation. |
| 4024 |
*/ |
| 4025 |
static RelOptInfo * |
| 4026 |
make_grouping_rel(PlannerInfo *root, RelOptInfo *input_rel, |
| 4027 |
PathTarget *target, bool target_parallel_safe, |
| 4028 |
Node *havingQual) |
| 4029 |
{ |
| 4030 |
RelOptInfo *grouped_rel; |
| 4031 |
|
| 4032 |
if (IS_OTHER_REL(input_rel)) |
| 4033 |
{ |
| 4034 |
grouped_rel = fetch_upper_rel(root, UPPERREL_GROUP_AGG, |
| 4035 |
input_rel->relids); |
| 4036 |
grouped_rel->reloptkind = RELOPT_OTHER_UPPER_REL; |
| 4037 |
} |
| 4038 |
else |
| 4039 |
{ |
| 4040 |
/* |
| 4041 |
* By tradition, the relids set for the main grouping relation is |
| 4042 |
* NULL. (This could be changed, but might require adjustments |
| 4043 |
* elsewhere.) |
| 4044 |
*/ |
| 4045 |
grouped_rel = fetch_upper_rel(root, UPPERREL_GROUP_AGG, NULL); |
| 4046 |
} |
| 4047 |
|
| 4048 |
/* Set target. */ |
| 4049 |
grouped_rel->reltarget = target; |
| 4050 |
|
| 4051 |
/* |
| 4052 |
* If the input relation is not parallel-safe, then the grouped relation |
| 4053 |
* can't be parallel-safe, either. Otherwise, it's parallel-safe if the |
| 4054 |
* target list and HAVING quals are parallel-safe. |
| 4055 |
*/ |
| 4056 |
if (input_rel->consider_parallel && target_parallel_safe && |
| 4057 |
is_parallel_safe(root, (Node *) havingQual)) |
| 4058 |
grouped_rel->consider_parallel = true; |
| 4059 |
|
| 4060 |
/* |
| 4061 |
* If the input rel belongs to a single FDW, so does the grouped rel. |
| 4062 |
*/ |
| 4063 |
grouped_rel->serverid = input_rel->serverid; |
| 4064 |
grouped_rel->userid = input_rel->userid; |
| 4065 |
grouped_rel->useridiscurrent = input_rel->useridiscurrent; |
| 4066 |
grouped_rel->fdwroutine = input_rel->fdwroutine; |
| 4067 |
|
| 4068 |
return grouped_rel; |
| 4069 |
} |
| 4070 |
|
| 4071 |
/* |
| 4072 |
* is_degenerate_grouping |
| 4073 |
* |
| 4074 |
* A degenerate grouping is one in which the query has a HAVING qual and/or |
| 4075 |
* grouping sets, but no aggregates and no GROUP BY (which implies that the |
| 4076 |
* grouping sets are all empty). |
| 4077 |
*/ |
| 4078 |
static bool |
| 4079 |
is_degenerate_grouping(PlannerInfo *root) |
| 4080 |
{ |
| 4081 |
Query *parse = root->parse; |
| 4082 |
|
| 4083 |
return (root->hasHavingQual || parse->groupingSets) && |
| 4084 |
!parse->hasAggs && parse->groupClause == NIL; |
| 4085 |
} |
| 4086 |
|
| 4087 |
/* |
| 4088 |
* create_degenerate_grouping_paths |
| 4089 |
* |
| 4090 |
* When the grouping is degenerate (see is_degenerate_grouping), we are |
| 4091 |
* supposed to emit either zero or one row for each grouping set depending on |
| 4092 |
* whether HAVING succeeds. Furthermore, there cannot be any variables in |
| 4093 |
* either HAVING or the targetlist, so we actually do not need the FROM table |
| 4094 |
* at all! We can just throw away the plan-so-far and generate a Result node. |
| 4095 |
* This is a sufficiently unusual corner case that it's not worth contorting |
| 4096 |
* the structure of this module to avoid having to generate the earlier paths |
| 4097 |
* in the first place. |
| 4098 |
*/ |
| 4099 |
static void |
| 4100 |
create_degenerate_grouping_paths(PlannerInfo *root, RelOptInfo *input_rel, |
| 4101 |
RelOptInfo *grouped_rel) |
| 4102 |
{ |
| 4103 |
Query *parse = root->parse; |
| 4104 |
int nrows; |
| 4105 |
Path *path; |
| 4106 |
|
| 4107 |
nrows = list_length(parse->groupingSets); |
| 4108 |
if (nrows > 1) |
| 4109 |
{ |
| 4110 |
/* |
| 4111 |
* Doesn't seem worthwhile writing code to cons up a generate_series |
| 4112 |
* or a values scan to emit multiple rows. Instead just make N clones |
| 4113 |
* and append them. (With a volatile HAVING clause, this means you |
| 4114 |
* might get between 0 and N output rows. Offhand I think that's |
| 4115 |
* desired.) |
| 4116 |
*/ |
| 4117 |
List *paths = NIL; |
| 4118 |
|
| 4119 |
while (--nrows >= 0) |
| 4120 |
{ |
| 4121 |
path = (Path *) |
| 4122 |
create_group_result_path(root, grouped_rel, |
| 4123 |
grouped_rel->reltarget, |
| 4124 |
(List *) parse->havingQual); |
| 4125 |
paths = lappend(paths, path); |
| 4126 |
} |
| 4127 |
path = (Path *) |
| 4128 |
create_append_path(root, |
| 4129 |
grouped_rel, |
| 4130 |
paths, |
| 4131 |
NIL, |
| 4132 |
NIL, |
| 4133 |
NULL, |
| 4134 |
0, |
| 4135 |
false, |
| 4136 |
-1); |
| 4137 |
} |
| 4138 |
else |
| 4139 |
{ |
| 4140 |
/* No grouping sets, or just one, so one output row */ |
| 4141 |
path = (Path *) |
| 4142 |
create_group_result_path(root, grouped_rel, |
| 4143 |
grouped_rel->reltarget, |
| 4144 |
(List *) parse->havingQual); |
| 4145 |
} |
| 4146 |
|
| 4147 |
add_path(grouped_rel, path); |
| 4148 |
} |
| 4149 |
|
| 4150 |
/* |
| 4151 |
* create_ordinary_grouping_paths |
| 4152 |
* |
| 4153 |
* Create grouping paths for the ordinary (that is, non-degenerate) case. |
| 4154 |
* |
| 4155 |
* We need to consider sorted and hashed aggregation in the same function, |
| 4156 |
* because otherwise (1) it would be harder to throw an appropriate error |
| 4157 |
* message if neither way works, and (2) we should not allow hashtable size |
| 4158 |
* considerations to dissuade us from using hashing if sorting is not possible. |
| 4159 |
* |
| 4160 |
* *partially_grouped_rel_p will be set to the partially grouped rel which this |
| 4161 |
* function creates, or to NULL if it doesn't create one. |
| 4162 |
*/ |
| 4163 |
static void |
| 4164 |
create_ordinary_grouping_paths(PlannerInfo *root, RelOptInfo *input_rel, |
| 4165 |
RelOptInfo *grouped_rel, |
| 4166 |
const AggClauseCosts *agg_costs, |
| 4167 |
grouping_sets_data *gd, |
| 4168 |
GroupPathExtraData *extra, |
| 4169 |
RelOptInfo **partially_grouped_rel_p) |
| 4170 |
{ |
| 4171 |
Path *cheapest_path = input_rel->cheapest_total_path; |
| 4172 |
RelOptInfo *partially_grouped_rel = NULL; |
| 4173 |
double dNumGroups; |
| 4174 |
PartitionwiseAggregateType patype = PARTITIONWISE_AGGREGATE_NONE; |
| 4175 |
|
| 4176 |
/* |
| 4177 |
* If this is the topmost grouping relation or if the parent relation is |
| 4178 |
* doing some form of partitionwise aggregation, then we may be able to do |
| 4179 |
* it at this level also. However, if the input relation is not |
| 4180 |
* partitioned, partitionwise aggregate is impossible. |
| 4181 |
*/ |
| 4182 |
if (extra->patype != PARTITIONWISE_AGGREGATE_NONE && |
| 4183 |
IS_PARTITIONED_REL(input_rel)) |
| 4184 |
{ |
| 4185 |
/* |
| 4186 |
* If this is the topmost relation or if the parent relation is doing |
| 4187 |
* full partitionwise aggregation, then we can do full partitionwise |
| 4188 |
* aggregation provided that the GROUP BY clause contains all of the |
| 4189 |
* partitioning columns at this level and the collation used by GROUP |
| 4190 |
* BY matches the partitioning collation. Otherwise, we can do at |
| 4191 |
* most partial partitionwise aggregation. But if partial aggregation |
| 4192 |
* is not supported in general then we can't use it for partitionwise |
| 4193 |
* aggregation either. |
| 4194 |
* |
| 4195 |
* Check parse->groupClause not processed_groupClause, because it's |
| 4196 |
* okay if some of the partitioning columns were proved redundant. |
| 4197 |
*/ |
| 4198 |
if (extra->patype == PARTITIONWISE_AGGREGATE_FULL && |
| 4199 |
group_by_has_partkey(input_rel, extra->targetList, |
| 4200 |
root->parse->groupClause)) |
| 4201 |
patype = PARTITIONWISE_AGGREGATE_FULL; |
| 4202 |
else if ((extra->flags & GROUPING_CAN_PARTIAL_AGG) != 0) |
| 4203 |
patype = PARTITIONWISE_AGGREGATE_PARTIAL; |
| 4204 |
else |
| 4205 |
patype = PARTITIONWISE_AGGREGATE_NONE; |
| 4206 |
} |
| 4207 |
|
| 4208 |
/* |
| 4209 |
* Before generating paths for grouped_rel, we first generate any possible |
| 4210 |
* partially grouped paths; that way, later code can easily consider both |
| 4211 |
* parallel and non-parallel approaches to grouping. |
| 4212 |
*/ |
| 4213 |
if ((extra->flags & GROUPING_CAN_PARTIAL_AGG) != 0) |
| 4214 |
{ |
| 4215 |
bool force_rel_creation; |
| 4216 |
|
| 4217 |
/* |
| 4218 |
* If we're doing partitionwise aggregation at this level, force |
| 4219 |
* creation of a partially_grouped_rel so we can add partitionwise |
| 4220 |
* paths to it. |
| 4221 |
*/ |
| 4222 |
force_rel_creation = (patype == PARTITIONWISE_AGGREGATE_PARTIAL); |
| 4223 |
|
| 4224 |
partially_grouped_rel = |
| 4225 |
create_partial_grouping_paths(root, |
| 4226 |
grouped_rel, |
| 4227 |
input_rel, |
| 4228 |
gd, |
| 4229 |
extra, |
| 4230 |
force_rel_creation); |
| 4231 |
} |
| 4232 |
|
| 4233 |
/* Set out parameter. */ |
| 4234 |
*partially_grouped_rel_p = partially_grouped_rel; |
| 4235 |
|
| 4236 |
/* Apply partitionwise aggregation technique, if possible. */ |
| 4237 |
if (patype != PARTITIONWISE_AGGREGATE_NONE) |
| 4238 |
create_partitionwise_grouping_paths(root, input_rel, grouped_rel, |
| 4239 |
partially_grouped_rel, agg_costs, |
| 4240 |
gd, patype, extra); |
| 4241 |
|
| 4242 |
/* If we are doing partial aggregation only, return. */ |
| 4243 |
if (extra->patype == PARTITIONWISE_AGGREGATE_PARTIAL) |
| 4244 |
{ |
| 4245 |
Assert(partially_grouped_rel); |
| 4246 |
|
| 4247 |
if (partially_grouped_rel->pathlist) |
| 4248 |
set_cheapest(partially_grouped_rel); |
| 4249 |
|
| 4250 |
return; |
| 4251 |
} |
| 4252 |
|
| 4253 |
/* Gather any partially grouped partial paths. */ |
| 4254 |
if (partially_grouped_rel && partially_grouped_rel->partial_pathlist) |
| 4255 |
{ |
| 4256 |
gather_grouping_paths(root, partially_grouped_rel); |
| 4257 |
set_cheapest(partially_grouped_rel); |
| 4258 |
} |
| 4259 |
|
| 4260 |
/* |
| 4261 |
* Estimate number of groups. |
| 4262 |
*/ |
| 4263 |
dNumGroups = get_number_of_groups(root, |
| 4264 |
cheapest_path->rows, |
| 4265 |
gd, |
| 4266 |
extra->targetList); |
| 4267 |
|
| 4268 |
/* Build final grouping paths */ |
| 4269 |
add_paths_to_grouping_rel(root, input_rel, grouped_rel, |
| 4270 |
partially_grouped_rel, agg_costs, gd, |
| 4271 |
dNumGroups, extra); |
| 4272 |
|
| 4273 |
/* Give a helpful error if we failed to find any implementation */ |
| 4274 |
if (grouped_rel->pathlist == NIL) |
| 4275 |
ereport(ERROR, |
| 4276 |
(errcode(ERRCODE_FEATURE_NOT_SUPPORTED), |
| 4277 |
errmsg("could not implement GROUP BY"), |
| 4278 |
errdetail("Some of the datatypes only support hashing, while others only support sorting."))); |
| 4279 |
|
| 4280 |
/* |
| 4281 |
* If there is an FDW that's responsible for all baserels of the query, |
| 4282 |
* let it consider adding ForeignPaths. |
| 4283 |
*/ |
| 4284 |
if (grouped_rel->fdwroutine && |
| 4285 |
grouped_rel->fdwroutine->GetForeignUpperPaths) |
| 4286 |
grouped_rel->fdwroutine->GetForeignUpperPaths(root, UPPERREL_GROUP_AGG, |
| 4287 |
input_rel, grouped_rel, |
| 4288 |
extra); |
| 4289 |
|
| 4290 |
/* Let extensions possibly add some more paths */ |
| 4291 |
if (create_upper_paths_hook) |
| 4292 |
(*create_upper_paths_hook) (root, UPPERREL_GROUP_AGG, |
| 4293 |
input_rel, grouped_rel, |
| 4294 |
extra); |
| 4295 |
} |
| 4296 |
|
| 4297 |
/* |
| 4298 |
* For a given input path, consider the possible ways of doing grouping sets on |
| 4299 |
* it, by combinations of hashing and sorting. This can be called multiple |
| 4300 |
* times, so it's important that it not scribble on input. No result is |
| 4301 |
* returned, but any generated paths are added to grouped_rel. |
| 4302 |
*/ |
| 4303 |
static void |
| 4304 |
consider_groupingsets_paths(PlannerInfo *root, |
| 4305 |
RelOptInfo *grouped_rel, |
| 4306 |
Path *path, |
| 4307 |
bool is_sorted, |
| 4308 |
bool can_hash, |
| 4309 |
grouping_sets_data *gd, |
| 4310 |
const AggClauseCosts *agg_costs, |
| 4311 |
double dNumGroups) |
| 4312 |
{ |
| 4313 |
Query *parse = root->parse; |
| 4314 |
Size hash_mem_limit = get_hash_memory_limit(); |
| 4315 |
|
| 4316 |
/* |
| 4317 |
* If we're not being offered sorted input, then only consider plans that |
| 4318 |
* can be done entirely by hashing. |
| 4319 |
* |
| 4320 |
* We can hash everything if it looks like it'll fit in hash_mem. But if |
| 4321 |
* the input is actually sorted despite not being advertised as such, we |
| 4322 |
* prefer to make use of that in order to use less memory. |
| 4323 |
* |
| 4324 |
* If none of the grouping sets are sortable, then ignore the hash_mem |
| 4325 |
* limit and generate a path anyway, since otherwise we'll just fail. |
| 4326 |
*/ |
| 4327 |
if (!is_sorted) |
| 4328 |
{ |
| 4329 |
List *new_rollups = NIL; |
| 4330 |
RollupData *unhashed_rollup = NULL; |
| 4331 |
List *sets_data; |
| 4332 |
List *empty_sets_data = NIL; |
| 4333 |
List *empty_sets = NIL; |
| 4334 |
ListCell *lc; |
| 4335 |
ListCell *l_start = list_head(gd->rollups); |
| 4336 |
AggStrategy strat = AGG_HASHED; |
| 4337 |
double hashsize; |
| 4338 |
double exclude_groups = 0.0; |
| 4339 |
|
| 4340 |
Assert(can_hash); |
| 4341 |
|
| 4342 |
/* |
| 4343 |
* If the input is coincidentally sorted usefully (which can happen |
| 4344 |
* even if is_sorted is false, since that only means that our caller |
| 4345 |
* has set up the sorting for us), then save some hashtable space by |
| 4346 |
* making use of that. But we need to watch out for degenerate cases: |
| 4347 |
* |
| 4348 |
* 1) If there are any empty grouping sets, then group_pathkeys might |
| 4349 |
* be NIL if all non-empty grouping sets are unsortable. In this case, |
| 4350 |
* there will be a rollup containing only empty groups, and the |
| 4351 |
* pathkeys_contained_in test is vacuously true; this is ok. |
| 4352 |
* |
| 4353 |
* XXX: the above relies on the fact that group_pathkeys is generated |
| 4354 |
* from the first rollup. If we add the ability to consider multiple |
| 4355 |
* sort orders for grouping input, this assumption might fail. |
| 4356 |
* |
| 4357 |
* 2) If there are no empty sets and only unsortable sets, then the |
| 4358 |
* rollups list will be empty (and thus l_start == NULL), and |
| 4359 |
* group_pathkeys will be NIL; we must ensure that the vacuously-true |
| 4360 |
* pathkeys_contained_in test doesn't cause us to crash. |
| 4361 |
*/ |
| 4362 |
if (l_start != NULL && |
| 4363 |
pathkeys_contained_in(root->group_pathkeys, path->pathkeys)) |
| 4364 |
{ |
| 4365 |
unhashed_rollup = lfirst_node(RollupData, l_start); |
| 4366 |
exclude_groups = unhashed_rollup->numGroups; |
| 4367 |
l_start = lnext(gd->rollups, l_start); |
| 4368 |
} |
| 4369 |
|
| 4370 |
hashsize = estimate_hashagg_tablesize(root, |
| 4371 |
path, |
| 4372 |
agg_costs, |
| 4373 |
dNumGroups - exclude_groups); |
| 4374 |
|
| 4375 |
/* |
| 4376 |
* gd->rollups is empty if we have only unsortable columns to work |
| 4377 |
* with. Override hash_mem in that case; otherwise, we'll rely on the |
| 4378 |
* sorted-input case to generate usable mixed paths. |
| 4379 |
*/ |
| 4380 |
if (hashsize > hash_mem_limit && gd->rollups) |
| 4381 |
return; /* nope, won't fit */ |
| 4382 |
|
| 4383 |
/* |
| 4384 |
* We need to burst the existing rollups list into individual grouping |
| 4385 |
* sets and recompute a groupClause for each set. |
| 4386 |
*/ |
| 4387 |
sets_data = list_copy(gd->unsortable_sets); |
| 4388 |
|
| 4389 |
for_each_cell(lc, gd->rollups, l_start) |
| 4390 |
{ |
| 4391 |
RollupData *rollup = lfirst_node(RollupData, lc); |
| 4392 |
|
| 4393 |
/* |
| 4394 |
* If we find an unhashable rollup that's not been skipped by the |
| 4395 |
* "actually sorted" check above, we can't cope; we'd need sorted |
| 4396 |
* input (with a different sort order) but we can't get that here. |
| 4397 |
* So bail out; we'll get a valid path from the is_sorted case |
| 4398 |
* instead. |
| 4399 |
* |
| 4400 |
* The mere presence of empty grouping sets doesn't make a rollup |
| 4401 |
* unhashable (see preprocess_grouping_sets), we handle those |
| 4402 |
* specially below. |
| 4403 |
*/ |
| 4404 |
if (!rollup->hashable) |
| 4405 |
return; |
| 4406 |
|
| 4407 |
sets_data = list_concat(sets_data, rollup->gsets_data); |
| 4408 |
} |
| 4409 |
foreach(lc, sets_data) |
| 4410 |
{ |
| 4411 |
GroupingSetData *gs = lfirst_node(GroupingSetData, lc); |
| 4412 |
List *gset = gs->set; |
| 4413 |
RollupData *rollup; |
| 4414 |
|
| 4415 |
if (gset == NIL) |
| 4416 |
{ |
| 4417 |
/* Empty grouping sets can't be hashed. */ |
| 4418 |
empty_sets_data = lappend(empty_sets_data, gs); |
| 4419 |
empty_sets = lappend(empty_sets, NIL); |
| 4420 |
} |
| 4421 |
else |
| 4422 |
{ |
| 4423 |
rollup = makeNode(RollupData); |
| 4424 |
|
| 4425 |
rollup->groupClause = preprocess_groupclause(root, gset); |
| 4426 |
rollup->gsets_data = list_make1(gs); |
| 4427 |
rollup->gsets = remap_to_groupclause_idx(rollup->groupClause, |
| 4428 |
rollup->gsets_data, |
| 4429 |
gd->tleref_to_colnum_map); |
| 4430 |
rollup->numGroups = gs->numGroups; |
| 4431 |
rollup->hashable = true; |
| 4432 |
rollup->is_hashed = true; |
| 4433 |
new_rollups = lappend(new_rollups, rollup); |
| 4434 |
} |
| 4435 |
} |
| 4436 |
|
| 4437 |
/* |
| 4438 |
* If we didn't find anything nonempty to hash, then bail. We'll |
| 4439 |
* generate a path from the is_sorted case. |
| 4440 |
*/ |
| 4441 |
if (new_rollups == NIL) |
| 4442 |
return; |
| 4443 |
|
| 4444 |
/* |
| 4445 |
* If there were empty grouping sets they should have been in the |
| 4446 |
* first rollup. |
| 4447 |
*/ |
| 4448 |
Assert(!unhashed_rollup || !empty_sets); |
| 4449 |
|
| 4450 |
if (unhashed_rollup) |
| 4451 |
{ |
| 4452 |
new_rollups = lappend(new_rollups, unhashed_rollup); |
| 4453 |
strat = AGG_MIXED; |
| 4454 |
} |
| 4455 |
else if (empty_sets) |
| 4456 |
{ |
| 4457 |
RollupData *rollup = makeNode(RollupData); |
| 4458 |
|
| 4459 |
rollup->groupClause = NIL; |
| 4460 |
rollup->gsets_data = empty_sets_data; |
| 4461 |
rollup->gsets = empty_sets; |
| 4462 |
rollup->numGroups = list_length(empty_sets); |
| 4463 |
rollup->hashable = false; |
| 4464 |
rollup->is_hashed = false; |
| 4465 |
new_rollups = lappend(new_rollups, rollup); |
| 4466 |
strat = AGG_MIXED; |
| 4467 |
} |
| 4468 |
|
| 4469 |
add_path(grouped_rel, (Path *) |
| 4470 |
create_groupingsets_path(root, |
| 4471 |
grouped_rel, |
| 4472 |
path, |
| 4473 |
(List *) parse->havingQual, |
| 4474 |
strat, |
| 4475 |
new_rollups, |
| 4476 |
agg_costs)); |
| 4477 |
return; |
| 4478 |
} |
| 4479 |
|
| 4480 |
/* |
| 4481 |
* If we have sorted input but nothing we can do with it, bail. |
| 4482 |
*/ |
| 4483 |
if (gd->rollups == NIL) |
| 4484 |
return; |
| 4485 |
|
| 4486 |
/* |
| 4487 |
* Given sorted input, we try and make two paths: one sorted and one mixed |
| 4488 |
* sort/hash. (We need to try both because hashagg might be disabled, or |
| 4489 |
* some columns might not be sortable.) |
| 4490 |
* |
| 4491 |
* can_hash is passed in as false if some obstacle elsewhere (such as |
| 4492 |
* ordered aggs) means that we shouldn't consider hashing at all. |
| 4493 |
*/ |
| 4494 |
if (can_hash && gd->any_hashable) |
| 4495 |
{ |
| 4496 |
List *rollups = NIL; |
| 4497 |
List *hash_sets = list_copy(gd->unsortable_sets); |
| 4498 |
double availspace = hash_mem_limit; |
| 4499 |
ListCell *lc; |
| 4500 |
|
| 4501 |
/* |
| 4502 |
* Account first for space needed for groups we can't sort at all. |
| 4503 |
*/ |
| 4504 |
availspace -= estimate_hashagg_tablesize(root, |
| 4505 |
path, |
| 4506 |
agg_costs, |
| 4507 |
gd->dNumHashGroups); |
| 4508 |
|
| 4509 |
if (availspace > 0 && list_length(gd->rollups) > 1) |
| 4510 |
{ |
| 4511 |
double scale; |
| 4512 |
int num_rollups = list_length(gd->rollups); |
| 4513 |
int k_capacity; |
| 4514 |
int *k_weights = palloc(num_rollups * sizeof(int)); |
| 4515 |
Bitmapset *hash_items = NULL; |
| 4516 |
int i; |
| 4517 |
|
| 4518 |
/* |
| 4519 |
* We treat this as a knapsack problem: the knapsack capacity |
| 4520 |
* represents hash_mem, the item weights are the estimated memory |
| 4521 |
* usage of the hashtables needed to implement a single rollup, |
| 4522 |
* and we really ought to use the cost saving as the item value; |
| 4523 |
* however, currently the costs assigned to sort nodes don't |
| 4524 |
* reflect the comparison costs well, and so we treat all items as |
| 4525 |
* of equal value (each rollup we hash instead saves us one sort). |
| 4526 |
* |
| 4527 |
* To use the discrete knapsack, we need to scale the values to a |
| 4528 |
* reasonably small bounded range. We choose to allow a 5% error |
| 4529 |
* margin; we have no more than 4096 rollups in the worst possible |
| 4530 |
* case, which with a 5% error margin will require a bit over 42MB |
| 4531 |
* of workspace. (Anyone wanting to plan queries that complex had |
| 4532 |
* better have the memory for it. In more reasonable cases, with |
| 4533 |
* no more than a couple of dozen rollups, the memory usage will |
| 4534 |
* be negligible.) |
| 4535 |
* |
| 4536 |
* k_capacity is naturally bounded, but we clamp the values for |
| 4537 |
* scale and weight (below) to avoid overflows or underflows (or |
| 4538 |
* uselessly trying to use a scale factor less than 1 byte). |
| 4539 |
*/ |
| 4540 |
scale = Max(availspace / (20.0 * num_rollups), 1.0); |
| 4541 |
k_capacity = (int) floor(availspace / scale); |
| 4542 |
|
| 4543 |
/* |
| 4544 |
* We leave the first rollup out of consideration since it's the |
| 4545 |
* one that matches the input sort order. We assign indexes "i" |
| 4546 |
* to only those entries considered for hashing; the second loop, |
| 4547 |
* below, must use the same condition. |
| 4548 |
*/ |
| 4549 |
i = 0; |
| 4550 |
for_each_from(lc, gd->rollups, 1) |
| 4551 |
{ |
| 4552 |
RollupData *rollup = lfirst_node(RollupData, lc); |
| 4553 |
|
| 4554 |
if (rollup->hashable) |
| 4555 |
{ |
| 4556 |
double sz = estimate_hashagg_tablesize(root, |
| 4557 |
path, |
| 4558 |
agg_costs, |
| 4559 |
rollup->numGroups); |
| 4560 |
|
| 4561 |
/* |
| 4562 |
* If sz is enormous, but hash_mem (and hence scale) is |
| 4563 |
* small, avoid integer overflow here. |
| 4564 |
*/ |
| 4565 |
k_weights[i] = (int) Min(floor(sz / scale), |
| 4566 |
k_capacity + 1.0); |
| 4567 |
++i; |
| 4568 |
} |
| 4569 |
} |
| 4570 |
|
| 4571 |
/* |
| 4572 |
* Apply knapsack algorithm; compute the set of items which |
| 4573 |
* maximizes the value stored (in this case the number of sorts |
| 4574 |
* saved) while keeping the total size (approximately) within |
| 4575 |
* capacity. |
| 4576 |
*/ |
| 4577 |
if (i > 0) |
| 4578 |
hash_items = DiscreteKnapsack(k_capacity, i, k_weights, NULL); |
| 4579 |
|
| 4580 |
if (!bms_is_empty(hash_items)) |
| 4581 |
{ |
| 4582 |
rollups = list_make1(linitial(gd->rollups)); |
| 4583 |
|
| 4584 |
i = 0; |
| 4585 |
for_each_from(lc, gd->rollups, 1) |
| 4586 |
{ |
| 4587 |
RollupData *rollup = lfirst_node(RollupData, lc); |
| 4588 |
|
| 4589 |
if (rollup->hashable) |
| 4590 |
{ |
| 4591 |
if (bms_is_member(i, hash_items)) |
| 4592 |
hash_sets = list_concat(hash_sets, |
| 4593 |
rollup->gsets_data); |
| 4594 |
else |
| 4595 |
rollups = lappend(rollups, rollup); |
| 4596 |
++i; |
| 4597 |
} |
| 4598 |
else |
| 4599 |
rollups = lappend(rollups, rollup); |
| 4600 |
} |
| 4601 |
} |
| 4602 |
} |
| 4603 |
|
| 4604 |
if (!rollups && hash_sets) |
| 4605 |
rollups = list_copy(gd->rollups); |
| 4606 |
|
| 4607 |
foreach(lc, hash_sets) |
| 4608 |
{ |
| 4609 |
GroupingSetData *gs = lfirst_node(GroupingSetData, lc); |
| 4610 |
RollupData *rollup = makeNode(RollupData); |
| 4611 |
|
| 4612 |
Assert(gs->set != NIL); |
| 4613 |
|
| 4614 |
rollup->groupClause = preprocess_groupclause(root, gs->set); |
| 4615 |
rollup->gsets_data = list_make1(gs); |
| 4616 |
rollup->gsets = remap_to_groupclause_idx(rollup->groupClause, |
| 4617 |
rollup->gsets_data, |
| 4618 |
gd->tleref_to_colnum_map); |
| 4619 |
rollup->numGroups = gs->numGroups; |
| 4620 |
rollup->hashable = true; |
| 4621 |
rollup->is_hashed = true; |
| 4622 |
rollups = lcons(rollup, rollups); |
| 4623 |
} |
| 4624 |
|
| 4625 |
if (rollups) |
| 4626 |
{ |
| 4627 |
add_path(grouped_rel, (Path *) |
| 4628 |
create_groupingsets_path(root, |
| 4629 |
grouped_rel, |
| 4630 |
path, |
| 4631 |
(List *) parse->havingQual, |
| 4632 |
AGG_MIXED, |
| 4633 |
rollups, |
| 4634 |
agg_costs)); |
| 4635 |
} |
| 4636 |
} |
| 4637 |
|
| 4638 |
/* |
| 4639 |
* Now try the simple sorted case. |
| 4640 |
*/ |
| 4641 |
if (!gd->unsortable_sets) |
| 4642 |
add_path(grouped_rel, (Path *) |
| 4643 |
create_groupingsets_path(root, |
| 4644 |
grouped_rel, |
| 4645 |
path, |
| 4646 |
(List *) parse->havingQual, |
| 4647 |
AGG_SORTED, |
| 4648 |
gd->rollups, |
| 4649 |
agg_costs)); |
| 4650 |
} |
| 4651 |
|
| 4652 |
/* |
| 4653 |
* create_window_paths |
| 4654 |
* |
| 4655 |
* Build a new upperrel containing Paths for window-function evaluation. |
| 4656 |
* |
| 4657 |
* input_rel: contains the source-data Paths |
| 4658 |
* input_target: result of make_window_input_target |
| 4659 |
* output_target: what the topmost WindowAggPath should return |
| 4660 |
* wflists: result of find_window_functions |
| 4661 |
* activeWindows: result of select_active_windows |
| 4662 |
* |
| 4663 |
* Note: all Paths in input_rel are expected to return input_target. |
| 4664 |
*/ |
| 4665 |
static RelOptInfo * |
| 4666 |
create_window_paths(PlannerInfo *root, |
| 4667 |
RelOptInfo *input_rel, |
| 4668 |
PathTarget *input_target, |
| 4669 |
PathTarget *output_target, |
| 4670 |
bool output_target_parallel_safe, |
| 4671 |
WindowFuncLists *wflists, |
| 4672 |
List *activeWindows) |
| 4673 |
{ |
| 4674 |
RelOptInfo *window_rel; |
| 4675 |
ListCell *lc; |
| 4676 |
|
| 4677 |
/* For now, do all work in the (WINDOW, NULL) upperrel */ |
| 4678 |
window_rel = fetch_upper_rel(root, UPPERREL_WINDOW, NULL); |
| 4679 |
|
| 4680 |
/* |
| 4681 |
* If the input relation is not parallel-safe, then the window relation |
| 4682 |
* can't be parallel-safe, either. Otherwise, we need to examine the |
| 4683 |
* target list and active windows for non-parallel-safe constructs. |
| 4684 |
*/ |
| 4685 |
if (input_rel->consider_parallel && output_target_parallel_safe && |
| 4686 |
is_parallel_safe(root, (Node *) activeWindows)) |
| 4687 |
window_rel->consider_parallel = true; |
| 4688 |
|
| 4689 |
/* |
| 4690 |
* If the input rel belongs to a single FDW, so does the window rel. |
| 4691 |
*/ |
| 4692 |
window_rel->serverid = input_rel->serverid; |
| 4693 |
window_rel->userid = input_rel->userid; |
| 4694 |
window_rel->useridiscurrent = input_rel->useridiscurrent; |
| 4695 |
window_rel->fdwroutine = input_rel->fdwroutine; |
| 4696 |
|
| 4697 |
/* |
| 4698 |
* Consider computing window functions starting from the existing |
| 4699 |
* cheapest-total path (which will likely require a sort) as well as any |
| 4700 |
* existing paths that satisfy or partially satisfy root->window_pathkeys. |
| 4701 |
*/ |
| 4702 |
foreach(lc, input_rel->pathlist) |
| 4703 |
{ |
| 4704 |
Path *path = (Path *) lfirst(lc); |
| 4705 |
int presorted_keys; |
| 4706 |
|
| 4707 |
if (path == input_rel->cheapest_total_path || |
| 4708 |
pathkeys_count_contained_in(root->window_pathkeys, path->pathkeys, |
| 4709 |
&presorted_keys) || |
| 4710 |
presorted_keys > 0) |
| 4711 |
create_one_window_path(root, |
| 4712 |
window_rel, |
| 4713 |
path, |
| 4714 |
input_target, |
| 4715 |
output_target, |
| 4716 |
wflists, |
| 4717 |
activeWindows); |
| 4718 |
} |
| 4719 |
|
| 4720 |
/* |
| 4721 |
* If there is an FDW that's responsible for all baserels of the query, |
| 4722 |
* let it consider adding ForeignPaths. |
| 4723 |
*/ |
| 4724 |
if (window_rel->fdwroutine && |
| 4725 |
window_rel->fdwroutine->GetForeignUpperPaths) |
| 4726 |
window_rel->fdwroutine->GetForeignUpperPaths(root, UPPERREL_WINDOW, |
| 4727 |
input_rel, window_rel, |
| 4728 |
NULL); |
| 4729 |
|
| 4730 |
/* Let extensions possibly add some more paths */ |
| 4731 |
if (create_upper_paths_hook) |
| 4732 |
(*create_upper_paths_hook) (root, UPPERREL_WINDOW, |
| 4733 |
input_rel, window_rel, NULL); |
| 4734 |
|
| 4735 |
/* Now choose the best path(s) */ |
| 4736 |
set_cheapest(window_rel); |
| 4737 |
|
| 4738 |
return window_rel; |
| 4739 |
} |
| 4740 |
|
| 4741 |
/* |
| 4742 |
* Stack window-function implementation steps atop the given Path, and |
| 4743 |
* add the result to window_rel. |
| 4744 |
* |
| 4745 |
* window_rel: upperrel to contain result |
| 4746 |
* path: input Path to use (must return input_target) |
| 4747 |
* input_target: result of make_window_input_target |
| 4748 |
* output_target: what the topmost WindowAggPath should return |
| 4749 |
* wflists: result of find_window_functions |
| 4750 |
* activeWindows: result of select_active_windows |
| 4751 |
*/ |
| 4752 |
static void |
| 4753 |
create_one_window_path(PlannerInfo *root, |
| 4754 |
RelOptInfo *window_rel, |
| 4755 |
Path *path, |
| 4756 |
PathTarget *input_target, |
| 4757 |
PathTarget *output_target, |
| 4758 |
WindowFuncLists *wflists, |
| 4759 |
List *activeWindows) |
| 4760 |
{ |
| 4761 |
PathTarget *window_target; |
| 4762 |
ListCell *l; |
| 4763 |
List *topqual = NIL; |
| 4764 |
|
| 4765 |
/* |
| 4766 |
* Since each window clause could require a different sort order, we stack |
| 4767 |
* up a WindowAgg node for each clause, with sort steps between them as |
| 4768 |
* needed. (We assume that select_active_windows chose a good order for |
| 4769 |
* executing the clauses in.) |
| 4770 |
* |
| 4771 |
* input_target should contain all Vars and Aggs needed for the result. |
| 4772 |
* (In some cases we wouldn't need to propagate all of these all the way |
| 4773 |
* to the top, since they might only be needed as inputs to WindowFuncs. |
| 4774 |
* It's probably not worth trying to optimize that though.) It must also |
| 4775 |
* contain all window partitioning and sorting expressions, to ensure |
| 4776 |
* they're computed only once at the bottom of the stack (that's critical |
| 4777 |
* for volatile functions). As we climb up the stack, we'll add outputs |
| 4778 |
* for the WindowFuncs computed at each level. |
| 4779 |
*/ |
| 4780 |
window_target = input_target; |
| 4781 |
|
| 4782 |
foreach(l, activeWindows) |
| 4783 |
{ |
| 4784 |
WindowClause *wc = lfirst_node(WindowClause, l); |
| 4785 |
List *window_pathkeys; |
| 4786 |
List *runcondition = NIL; |
| 4787 |
int presorted_keys; |
| 4788 |
bool is_sorted; |
| 4789 |
bool topwindow; |
| 4790 |
ListCell *lc2; |
| 4791 |
|
| 4792 |
window_pathkeys = make_pathkeys_for_window(root, |
| 4793 |
wc, |
| 4794 |
root->processed_tlist); |
| 4795 |
|
| 4796 |
is_sorted = pathkeys_count_contained_in(window_pathkeys, |
| 4797 |
path->pathkeys, |
| 4798 |
&presorted_keys); |
| 4799 |
|
| 4800 |
/* Sort if necessary */ |
| 4801 |
if (!is_sorted) |
| 4802 |
{ |
| 4803 |
/* |
| 4804 |
* No presorted keys or incremental sort disabled, just perform a |
| 4805 |
* complete sort. |
| 4806 |
*/ |
| 4807 |
if (presorted_keys == 0 || !enable_incremental_sort) |
| 4808 |
path = (Path *) create_sort_path(root, window_rel, |
| 4809 |
path, |
| 4810 |
window_pathkeys, |
| 4811 |
-1.0); |
| 4812 |
else |
| 4813 |
{ |
| 4814 |
/* |
| 4815 |
* Since we have presorted keys and incremental sort is |
| 4816 |
* enabled, just use incremental sort. |
| 4817 |
*/ |
| 4818 |
path = (Path *) create_incremental_sort_path(root, |
| 4819 |
window_rel, |
| 4820 |
path, |
| 4821 |
window_pathkeys, |
| 4822 |
presorted_keys, |
| 4823 |
-1.0); |
| 4824 |
} |
| 4825 |
} |
| 4826 |
|
| 4827 |
if (lnext(activeWindows, l)) |
| 4828 |
{ |
| 4829 |
/* |
| 4830 |
* Add the current WindowFuncs to the output target for this |
| 4831 |
* intermediate WindowAggPath. We must copy window_target to |
| 4832 |
* avoid changing the previous path's target. |
| 4833 |
* |
| 4834 |
* Note: a WindowFunc adds nothing to the target's eval costs; but |
| 4835 |
* we do need to account for the increase in tlist width. |
| 4836 |
*/ |
| 4837 |
int64 tuple_width = window_target->width; |
| 4838 |
|
| 4839 |
window_target = copy_pathtarget(window_target); |
| 4840 |
foreach(lc2, wflists->windowFuncs[wc->winref]) |
| 4841 |
{ |
| 4842 |
WindowFunc *wfunc = lfirst_node(WindowFunc, lc2); |
| 4843 |
|
| 4844 |
add_column_to_pathtarget(window_target, (Expr *) wfunc, 0); |
| 4845 |
tuple_width += get_typavgwidth(wfunc->wintype, -1); |
| 4846 |
} |
| 4847 |
window_target->width = clamp_width_est(tuple_width); |
| 4848 |
} |
| 4849 |
else |
| 4850 |
{ |
| 4851 |
/* Install the goal target in the topmost WindowAgg */ |
| 4852 |
window_target = output_target; |
| 4853 |
} |
| 4854 |
|
| 4855 |
/* mark the final item in the list as the top-level window */ |
| 4856 |
topwindow = foreach_current_index(l) == list_length(activeWindows) - 1; |
| 4857 |
|
| 4858 |
/* |
| 4859 |
* Collect the WindowFuncRunConditions from each WindowFunc and |
| 4860 |
* convert them into OpExprs |
| 4861 |
*/ |
| 4862 |
foreach(lc2, wflists->windowFuncs[wc->winref]) |
| 4863 |
{ |
| 4864 |
ListCell *lc3; |
| 4865 |
WindowFunc *wfunc = lfirst_node(WindowFunc, lc2); |
| 4866 |
|
| 4867 |
foreach(lc3, wfunc->runCondition) |
| 4868 |
{ |
| 4869 |
WindowFuncRunCondition *wfuncrc = |
| 4870 |
lfirst_node(WindowFuncRunCondition, lc3); |
| 4871 |
Expr *opexpr; |
| 4872 |
Expr *leftop; |
| 4873 |
Expr *rightop; |
| 4874 |
|
| 4875 |
if (wfuncrc->wfunc_left) |
| 4876 |
{ |
| 4877 |
leftop = (Expr *) copyObject(wfunc); |
| 4878 |
rightop = copyObject(wfuncrc->arg); |
| 4879 |
} |
| 4880 |
else |
| 4881 |
{ |
| 4882 |
leftop = copyObject(wfuncrc->arg); |
| 4883 |
rightop = (Expr *) copyObject(wfunc); |
| 4884 |
} |
| 4885 |
|
| 4886 |
opexpr = make_opclause(wfuncrc->opno, |
| 4887 |
BOOLOID, |
| 4888 |
false, |
| 4889 |
leftop, |
| 4890 |
rightop, |
| 4891 |
InvalidOid, |
| 4892 |
wfuncrc->inputcollid); |
| 4893 |
|
| 4894 |
runcondition = lappend(runcondition, opexpr); |
| 4895 |
|
| 4896 |
if (!topwindow) |
| 4897 |
topqual = lappend(topqual, opexpr); |
| 4898 |
} |
| 4899 |
} |
| 4900 |
|
| 4901 |
path = (Path *) |
| 4902 |
create_windowagg_path(root, window_rel, path, window_target, |
| 4903 |
wflists->windowFuncs[wc->winref], |
| 4904 |
runcondition, wc, |
| 4905 |
topwindow ? topqual : NIL, topwindow); |
| 4906 |
} |
| 4907 |
|
| 4908 |
add_path(window_rel, path); |
| 4909 |
} |
| 4910 |
|
| 4911 |
/* |
| 4912 |
* create_distinct_paths |
| 4913 |
* |
| 4914 |
* Build a new upperrel containing Paths for SELECT DISTINCT evaluation. |
| 4915 |
* |
| 4916 |
* input_rel: contains the source-data Paths |
| 4917 |
* target: the pathtarget for the result Paths to compute |
| 4918 |
* |
| 4919 |
* Note: input paths should already compute the desired pathtarget, since |
| 4920 |
* Sort/Unique won't project anything. |
| 4921 |
*/ |
| 4922 |
static RelOptInfo * |
| 4923 |
create_distinct_paths(PlannerInfo *root, RelOptInfo *input_rel, |
| 4924 |
PathTarget *target) |
| 4925 |
{ |
| 4926 |
RelOptInfo *distinct_rel; |
| 4927 |
|
| 4928 |
/* For now, do all work in the (DISTINCT, NULL) upperrel */ |
| 4929 |
distinct_rel = fetch_upper_rel(root, UPPERREL_DISTINCT, NULL); |
| 4930 |
|
| 4931 |
/* |
| 4932 |
* We don't compute anything at this level, so distinct_rel will be |
| 4933 |
* parallel-safe if the input rel is parallel-safe. In particular, if |
| 4934 |
* there is a DISTINCT ON (...) clause, any path for the input_rel will |
| 4935 |
* output those expressions, and will not be parallel-safe unless those |
| 4936 |
* expressions are parallel-safe. |
| 4937 |
*/ |
| 4938 |
distinct_rel->consider_parallel = input_rel->consider_parallel; |
| 4939 |
|
| 4940 |
/* |
| 4941 |
* If the input rel belongs to a single FDW, so does the distinct_rel. |
| 4942 |
*/ |
| 4943 |
distinct_rel->serverid = input_rel->serverid; |
| 4944 |
distinct_rel->userid = input_rel->userid; |
| 4945 |
distinct_rel->useridiscurrent = input_rel->useridiscurrent; |
| 4946 |
distinct_rel->fdwroutine = input_rel->fdwroutine; |
| 4947 |
|
| 4948 |
/* build distinct paths based on input_rel's pathlist */ |
| 4949 |
create_final_distinct_paths(root, input_rel, distinct_rel); |
| 4950 |
|
| 4951 |
/* now build distinct paths based on input_rel's partial_pathlist */ |
| 4952 |
create_partial_distinct_paths(root, input_rel, distinct_rel, target); |
| 4953 |
|
| 4954 |
/* Give a helpful error if we failed to create any paths */ |
| 4955 |
if (distinct_rel->pathlist == NIL) |
| 4956 |
ereport(ERROR, |
| 4957 |
(errcode(ERRCODE_FEATURE_NOT_SUPPORTED), |
| 4958 |
errmsg("could not implement DISTINCT"), |
| 4959 |
errdetail("Some of the datatypes only support hashing, while others only support sorting."))); |
| 4960 |
|
| 4961 |
/* |
| 4962 |
* If there is an FDW that's responsible for all baserels of the query, |
| 4963 |
* let it consider adding ForeignPaths. |
| 4964 |
*/ |
| 4965 |
if (distinct_rel->fdwroutine && |
| 4966 |
distinct_rel->fdwroutine->GetForeignUpperPaths) |
| 4967 |
distinct_rel->fdwroutine->GetForeignUpperPaths(root, |
| 4968 |
UPPERREL_DISTINCT, |
| 4969 |
input_rel, |
| 4970 |
distinct_rel, |
| 4971 |
NULL); |
| 4972 |
|
| 4973 |
/* Let extensions possibly add some more paths */ |
| 4974 |
if (create_upper_paths_hook) |
| 4975 |
(*create_upper_paths_hook) (root, UPPERREL_DISTINCT, input_rel, |
| 4976 |
distinct_rel, NULL); |
| 4977 |
|
| 4978 |
/* Now choose the best path(s) */ |
| 4979 |
set_cheapest(distinct_rel); |
| 4980 |
|
| 4981 |
return distinct_rel; |
| 4982 |
} |
| 4983 |
|
| 4984 |
/* |
| 4985 |
* create_partial_distinct_paths |
| 4986 |
* |
| 4987 |
* Process 'input_rel' partial paths and add unique/aggregate paths to the |
| 4988 |
* UPPERREL_PARTIAL_DISTINCT rel. For paths created, add Gather/GatherMerge |
| 4989 |
* paths on top and add a final unique/aggregate path to remove any duplicate |
| 4990 |
* produced from combining rows from parallel workers. |
| 4991 |
*/ |
| 4992 |
static void |
| 4993 |
create_partial_distinct_paths(PlannerInfo *root, RelOptInfo *input_rel, |
| 4994 |
RelOptInfo *final_distinct_rel, |
| 4995 |
PathTarget *target) |
| 4996 |
{ |
| 4997 |
RelOptInfo *partial_distinct_rel; |
| 4998 |
Query *parse; |
| 4999 |
List *distinctExprs; |
| 5000 |
double numDistinctRows; |
| 5001 |
Path *cheapest_partial_path; |
| 5002 |
ListCell *lc; |
| 5003 |
|
| 5004 |
/* nothing to do when there are no partial paths in the input rel */ |
| 5005 |
if (!input_rel->consider_parallel || input_rel->partial_pathlist == NIL) |
| 5006 |
return; |
| 5007 |
|
| 5008 |
parse = root->parse; |
| 5009 |
|
| 5010 |
/* can't do parallel DISTINCT ON */ |
| 5011 |
if (parse->hasDistinctOn) |
| 5012 |
return; |
| 5013 |
|
| 5014 |
partial_distinct_rel = fetch_upper_rel(root, UPPERREL_PARTIAL_DISTINCT, |
| 5015 |
NULL); |
| 5016 |
partial_distinct_rel->reltarget = target; |
| 5017 |
partial_distinct_rel->consider_parallel = input_rel->consider_parallel; |
| 5018 |
|
| 5019 |
/* |
| 5020 |
* If input_rel belongs to a single FDW, so does the partial_distinct_rel. |
| 5021 |
*/ |
| 5022 |
partial_distinct_rel->serverid = input_rel->serverid; |
| 5023 |
partial_distinct_rel->userid = input_rel->userid; |
| 5024 |
partial_distinct_rel->useridiscurrent = input_rel->useridiscurrent; |
| 5025 |
partial_distinct_rel->fdwroutine = input_rel->fdwroutine; |
| 5026 |
|
| 5027 |
cheapest_partial_path = linitial(input_rel->partial_pathlist); |
| 5028 |
|
| 5029 |
distinctExprs = get_sortgrouplist_exprs(root->processed_distinctClause, |
| 5030 |
parse->targetList); |
| 5031 |
|
| 5032 |
/* estimate how many distinct rows we'll get from each worker */ |
| 5033 |
numDistinctRows = estimate_num_groups(root, distinctExprs, |
| 5034 |
cheapest_partial_path->rows, |
| 5035 |
NULL, NULL); |
| 5036 |
|
| 5037 |
/* |
| 5038 |
* Try sorting the cheapest path and incrementally sorting any paths with |
| 5039 |
* presorted keys and put a unique paths atop of those. We'll also |
| 5040 |
* attempt to reorder the required pathkeys to match the input path's |
| 5041 |
* pathkeys as much as possible, in hopes of avoiding a possible need to |
| 5042 |
* re-sort. |
| 5043 |
*/ |
| 5044 |
if (grouping_is_sortable(root->processed_distinctClause)) |
| 5045 |
{ |
| 5046 |
foreach(lc, input_rel->partial_pathlist) |
| 5047 |
{ |
| 5048 |
Path *input_path = (Path *) lfirst(lc); |
| 5049 |
Path *sorted_path; |
| 5050 |
List *useful_pathkeys_list = NIL; |
| 5051 |
|
| 5052 |
useful_pathkeys_list = |
| 5053 |
get_useful_pathkeys_for_distinct(root, |
| 5054 |
root->distinct_pathkeys, |
| 5055 |
input_path->pathkeys); |
| 5056 |
Assert(list_length(useful_pathkeys_list) > 0); |
| 5057 |
|
| 5058 |
foreach_node(List, useful_pathkeys, useful_pathkeys_list) |
| 5059 |
{ |
| 5060 |
sorted_path = make_ordered_path(root, |
| 5061 |
partial_distinct_rel, |
| 5062 |
input_path, |
| 5063 |
cheapest_partial_path, |
| 5064 |
useful_pathkeys, |
| 5065 |
-1.0); |
| 5066 |
|
| 5067 |
if (sorted_path == NULL) |
| 5068 |
continue; |
| 5069 |
|
| 5070 |
/* |
| 5071 |
* An empty distinct_pathkeys means all tuples have the same |
| 5072 |
* value for the DISTINCT clause. See |
| 5073 |
* create_final_distinct_paths() |
| 5074 |
*/ |
| 5075 |
if (root->distinct_pathkeys == NIL) |
| 5076 |
{ |
| 5077 |
Node *limitCount; |
| 5078 |
|
| 5079 |
limitCount = (Node *) makeConst(INT8OID, -1, InvalidOid, |
| 5080 |
sizeof(int64), |
| 5081 |
Int64GetDatum(1), false, |
| 5082 |
FLOAT8PASSBYVAL); |
| 5083 |
|
| 5084 |
/* |
| 5085 |
* Apply a LimitPath onto the partial path to restrict the |
| 5086 |
* tuples from each worker to 1. |
| 5087 |
* create_final_distinct_paths will need to apply an |
| 5088 |
* additional LimitPath to restrict this to a single row |
| 5089 |
* after the Gather node. If the query already has a |
| 5090 |
* LIMIT clause, then we could end up with three Limit |
| 5091 |
* nodes in the final plan. Consolidating the top two of |
| 5092 |
* these could be done, but does not seem worth troubling |
| 5093 |
* over. |
| 5094 |
*/ |
| 5095 |
add_partial_path(partial_distinct_rel, (Path *) |
| 5096 |
create_limit_path(root, partial_distinct_rel, |
| 5097 |
sorted_path, |
| 5098 |
NULL, |
| 5099 |
limitCount, |
| 5100 |
LIMIT_OPTION_COUNT, |
| 5101 |
0, 1)); |
| 5102 |
} |
| 5103 |
else |
| 5104 |
{ |
| 5105 |
add_partial_path(partial_distinct_rel, (Path *) |
| 5106 |
create_upper_unique_path(root, partial_distinct_rel, |
| 5107 |
sorted_path, |
| 5108 |
list_length(root->distinct_pathkeys), |
| 5109 |
numDistinctRows)); |
| 5110 |
} |
| 5111 |
} |
| 5112 |
} |
| 5113 |
} |
| 5114 |
|
| 5115 |
/* |
| 5116 |
* Now try hash aggregate paths, if enabled and hashing is possible. Since |
| 5117 |
* we're not on the hook to ensure we do our best to create at least one |
| 5118 |
* path here, we treat enable_hashagg as a hard off-switch rather than the |
| 5119 |
* slightly softer variant in create_final_distinct_paths. |
| 5120 |
*/ |
| 5121 |
if (enable_hashagg && grouping_is_hashable(root->processed_distinctClause)) |
| 5122 |
{ |
| 5123 |
add_partial_path(partial_distinct_rel, (Path *) |
| 5124 |
create_agg_path(root, |
| 5125 |
partial_distinct_rel, |
| 5126 |
cheapest_partial_path, |
| 5127 |
cheapest_partial_path->pathtarget, |
| 5128 |
AGG_HASHED, |
| 5129 |
AGGSPLIT_SIMPLE, |
| 5130 |
root->processed_distinctClause, |
| 5131 |
NIL, |
| 5132 |
NULL, |
| 5133 |
numDistinctRows)); |
| 5134 |
} |
| 5135 |
|
| 5136 |
/* |
| 5137 |
* If there is an FDW that's responsible for all baserels of the query, |
| 5138 |
* let it consider adding ForeignPaths. |
| 5139 |
*/ |
| 5140 |
if (partial_distinct_rel->fdwroutine && |
| 5141 |
partial_distinct_rel->fdwroutine->GetForeignUpperPaths) |
| 5142 |
partial_distinct_rel->fdwroutine->GetForeignUpperPaths(root, |
| 5143 |
UPPERREL_PARTIAL_DISTINCT, |
| 5144 |
input_rel, |
| 5145 |
partial_distinct_rel, |
| 5146 |
NULL); |
| 5147 |
|
| 5148 |
/* Let extensions possibly add some more partial paths */ |
| 5149 |
if (create_upper_paths_hook) |
| 5150 |
(*create_upper_paths_hook) (root, UPPERREL_PARTIAL_DISTINCT, |
| 5151 |
input_rel, partial_distinct_rel, NULL); |
| 5152 |
|
| 5153 |
if (partial_distinct_rel->partial_pathlist != NIL) |
| 5154 |
{ |
| 5155 |
generate_useful_gather_paths(root, partial_distinct_rel, true); |
| 5156 |
set_cheapest(partial_distinct_rel); |
| 5157 |
|
| 5158 |
/* |
| 5159 |
* Finally, create paths to distinctify the final result. This step |
| 5160 |
* is needed to remove any duplicates due to combining rows from |
| 5161 |
* parallel workers. |
| 5162 |
*/ |
| 5163 |
create_final_distinct_paths(root, partial_distinct_rel, |
| 5164 |
final_distinct_rel); |
| 5165 |
} |
| 5166 |
} |
| 5167 |
|
| 5168 |
/* |
| 5169 |
* create_final_distinct_paths |
| 5170 |
* Create distinct paths in 'distinct_rel' based on 'input_rel' pathlist |
| 5171 |
* |
| 5172 |
* input_rel: contains the source-data paths |
| 5173 |
* distinct_rel: destination relation for storing created paths |
| 5174 |
*/ |
| 5175 |
static RelOptInfo * |
| 5176 |
create_final_distinct_paths(PlannerInfo *root, RelOptInfo *input_rel, |
| 5177 |
RelOptInfo *distinct_rel) |
| 5178 |
{ |
| 5179 |
Query *parse = root->parse; |
| 5180 |
Path *cheapest_input_path = input_rel->cheapest_total_path; |
| 5181 |
double numDistinctRows; |
| 5182 |
bool allow_hash; |
| 5183 |
|
| 5184 |
/* Estimate number of distinct rows there will be */ |
| 5185 |
if (parse->groupClause || parse->groupingSets || parse->hasAggs || |
| 5186 |
root->hasHavingQual) |
| 5187 |
{ |
| 5188 |
/* |
| 5189 |
* If there was grouping or aggregation, use the number of input rows |
| 5190 |
* as the estimated number of DISTINCT rows (ie, assume the input is |
| 5191 |
* already mostly unique). |
| 5192 |
*/ |
| 5193 |
numDistinctRows = cheapest_input_path->rows; |
| 5194 |
} |
| 5195 |
else |
| 5196 |
{ |
| 5197 |
/* |
| 5198 |
* Otherwise, the UNIQUE filter has effects comparable to GROUP BY. |
| 5199 |
*/ |
| 5200 |
List *distinctExprs; |
| 5201 |
|
| 5202 |
distinctExprs = get_sortgrouplist_exprs(root->processed_distinctClause, |
| 5203 |
parse->targetList); |
| 5204 |
numDistinctRows = estimate_num_groups(root, distinctExprs, |
| 5205 |
cheapest_input_path->rows, |
| 5206 |
NULL, NULL); |
| 5207 |
} |
| 5208 |
|
| 5209 |
/* |
| 5210 |
* Consider sort-based implementations of DISTINCT, if possible. |
| 5211 |
*/ |
| 5212 |
if (grouping_is_sortable(root->processed_distinctClause)) |
| 5213 |
{ |
| 5214 |
/* |
| 5215 |
* Firstly, if we have any adequately-presorted paths, just stick a |
| 5216 |
* Unique node on those. We also, consider doing an explicit sort of |
| 5217 |
* the cheapest input path and Unique'ing that. If any paths have |
| 5218 |
* presorted keys then we'll create an incremental sort atop of those |
| 5219 |
* before adding a unique node on the top. We'll also attempt to |
| 5220 |
* reorder the required pathkeys to match the input path's pathkeys as |
| 5221 |
* much as possible, in hopes of avoiding a possible need to re-sort. |
| 5222 |
* |
| 5223 |
* When we have DISTINCT ON, we must sort by the more rigorous of |
| 5224 |
* DISTINCT and ORDER BY, else it won't have the desired behavior. |
| 5225 |
* Also, if we do have to do an explicit sort, we might as well use |
| 5226 |
* the more rigorous ordering to avoid a second sort later. (Note |
| 5227 |
* that the parser will have ensured that one clause is a prefix of |
| 5228 |
* the other.) |
| 5229 |
*/ |
| 5230 |
List *needed_pathkeys; |
| 5231 |
ListCell *lc; |
| 5232 |
double limittuples = root->distinct_pathkeys == NIL ? 1.0 : -1.0; |
| 5233 |
|
| 5234 |
if (parse->hasDistinctOn && |
| 5235 |
list_length(root->distinct_pathkeys) < |
| 5236 |
list_length(root->sort_pathkeys)) |
| 5237 |
needed_pathkeys = root->sort_pathkeys; |
| 5238 |
else |
| 5239 |
needed_pathkeys = root->distinct_pathkeys; |
| 5240 |
|
| 5241 |
foreach(lc, input_rel->pathlist) |
| 5242 |
{ |
| 5243 |
Path *input_path = (Path *) lfirst(lc); |
| 5244 |
Path *sorted_path; |
| 5245 |
List *useful_pathkeys_list = NIL; |
| 5246 |
|
| 5247 |
useful_pathkeys_list = |
| 5248 |
get_useful_pathkeys_for_distinct(root, |
| 5249 |
needed_pathkeys, |
| 5250 |
input_path->pathkeys); |
| 5251 |
Assert(list_length(useful_pathkeys_list) > 0); |
| 5252 |
|
| 5253 |
foreach_node(List, useful_pathkeys, useful_pathkeys_list) |
| 5254 |
{ |
| 5255 |
sorted_path = make_ordered_path(root, |
| 5256 |
distinct_rel, |
| 5257 |
input_path, |
| 5258 |
cheapest_input_path, |
| 5259 |
useful_pathkeys, |
| 5260 |
limittuples); |
| 5261 |
|
| 5262 |
if (sorted_path == NULL) |
| 5263 |
continue; |
| 5264 |
|
| 5265 |
/* |
| 5266 |
* distinct_pathkeys may have become empty if all of the |
| 5267 |
* pathkeys were determined to be redundant. If all of the |
| 5268 |
* pathkeys are redundant then each DISTINCT target must only |
| 5269 |
* allow a single value, therefore all resulting tuples must |
| 5270 |
* be identical (or at least indistinguishable by an equality |
| 5271 |
* check). We can uniquify these tuples simply by just taking |
| 5272 |
* the first tuple. All we do here is add a path to do "LIMIT |
| 5273 |
* 1" atop of 'sorted_path'. When doing a DISTINCT ON we may |
| 5274 |
* still have a non-NIL sort_pathkeys list, so we must still |
| 5275 |
* only do this with paths which are correctly sorted by |
| 5276 |
* sort_pathkeys. |
| 5277 |
*/ |
| 5278 |
if (root->distinct_pathkeys == NIL) |
| 5279 |
{ |
| 5280 |
Node *limitCount; |
| 5281 |
|
| 5282 |
limitCount = (Node *) makeConst(INT8OID, -1, InvalidOid, |
| 5283 |
sizeof(int64), |
| 5284 |
Int64GetDatum(1), false, |
| 5285 |
FLOAT8PASSBYVAL); |
| 5286 |
|
| 5287 |
/* |
| 5288 |
* If the query already has a LIMIT clause, then we could |
| 5289 |
* end up with a duplicate LimitPath in the final plan. |
| 5290 |
* That does not seem worth troubling over too much. |
| 5291 |
*/ |
| 5292 |
add_path(distinct_rel, (Path *) |
| 5293 |
create_limit_path(root, distinct_rel, sorted_path, |
| 5294 |
NULL, limitCount, |
| 5295 |
LIMIT_OPTION_COUNT, 0, 1)); |
| 5296 |
} |
| 5297 |
else |
| 5298 |
{ |
| 5299 |
add_path(distinct_rel, (Path *) |
| 5300 |
create_upper_unique_path(root, distinct_rel, |
| 5301 |
sorted_path, |
| 5302 |
list_length(root->distinct_pathkeys), |
| 5303 |
numDistinctRows)); |
| 5304 |
} |
| 5305 |
} |
| 5306 |
} |
| 5307 |
} |
| 5308 |
|
| 5309 |
/* |
| 5310 |
* Consider hash-based implementations of DISTINCT, if possible. |
| 5311 |
* |
| 5312 |
* If we were not able to make any other types of path, we *must* hash or |
| 5313 |
* die trying. If we do have other choices, there are two things that |
| 5314 |
* should prevent selection of hashing: if the query uses DISTINCT ON |
| 5315 |
* (because it won't really have the expected behavior if we hash), or if |
| 5316 |
* enable_hashagg is off. |
| 5317 |
* |
| 5318 |
* Note: grouping_is_hashable() is much more expensive to check than the |
| 5319 |
* other gating conditions, so we want to do it last. |
| 5320 |
*/ |
| 5321 |
if (distinct_rel->pathlist == NIL) |
| 5322 |
allow_hash = true; /* we have no alternatives */ |
| 5323 |
else if (parse->hasDistinctOn || !enable_hashagg) |
| 5324 |
allow_hash = false; /* policy-based decision not to hash */ |
| 5325 |
else |
| 5326 |
allow_hash = true; /* default */ |
| 5327 |
|
| 5328 |
if (allow_hash && grouping_is_hashable(root->processed_distinctClause)) |
| 5329 |
{ |
| 5330 |
/* Generate hashed aggregate path --- no sort needed */ |
| 5331 |
add_path(distinct_rel, (Path *) |
| 5332 |
create_agg_path(root, |
| 5333 |
distinct_rel, |
| 5334 |
cheapest_input_path, |
| 5335 |
cheapest_input_path->pathtarget, |
| 5336 |
AGG_HASHED, |
| 5337 |
AGGSPLIT_SIMPLE, |
| 5338 |
root->processed_distinctClause, |
| 5339 |
NIL, |
| 5340 |
NULL, |
| 5341 |
numDistinctRows)); |
| 5342 |
} |
| 5343 |
|
| 5344 |
return distinct_rel; |
| 5345 |
} |
| 5346 |
|
| 5347 |
/* |
| 5348 |
* get_useful_pathkeys_for_distinct |
| 5349 |
* Get useful orderings of pathkeys for distinctClause by reordering |
| 5350 |
* 'needed_pathkeys' to match the given 'path_pathkeys' as much as possible. |
| 5351 |
* |
| 5352 |
* This returns a list of pathkeys that can be useful for DISTINCT or DISTINCT |
| 5353 |
* ON clause. For convenience, it always includes the given 'needed_pathkeys'. |
| 5354 |
*/ |
| 5355 |
static List * |
| 5356 |
get_useful_pathkeys_for_distinct(PlannerInfo *root, List *needed_pathkeys, |
| 5357 |
List *path_pathkeys) |
| 5358 |
{ |
| 5359 |
List *useful_pathkeys_list = NIL; |
| 5360 |
List *useful_pathkeys = NIL; |
| 5361 |
|
| 5362 |
/* always include the given 'needed_pathkeys' */ |
| 5363 |
useful_pathkeys_list = lappend(useful_pathkeys_list, |
| 5364 |
needed_pathkeys); |
| 5365 |
|
| 5366 |
if (!enable_distinct_reordering) |
| 5367 |
return useful_pathkeys_list; |
| 5368 |
|
| 5369 |
/* |
| 5370 |
* Scan the given 'path_pathkeys' and construct a list of PathKey nodes |
| 5371 |
* that match 'needed_pathkeys', but only up to the longest matching |
| 5372 |
* prefix. |
| 5373 |
* |
| 5374 |
* When we have DISTINCT ON, we must ensure that the resulting pathkey |
| 5375 |
* list matches initial distinctClause pathkeys; otherwise, it won't have |
| 5376 |
* the desired behavior. |
| 5377 |
*/ |
| 5378 |
foreach_node(PathKey, pathkey, path_pathkeys) |
| 5379 |
{ |
| 5380 |
/* |
| 5381 |
* The PathKey nodes are canonical, so they can be checked for |
| 5382 |
* equality by simple pointer comparison. |
| 5383 |
*/ |
| 5384 |
if (!list_member_ptr(needed_pathkeys, pathkey)) |
| 5385 |
break; |
| 5386 |
if (root->parse->hasDistinctOn && |
| 5387 |
!list_member_ptr(root->distinct_pathkeys, pathkey)) |
| 5388 |
break; |
| 5389 |
|
| 5390 |
useful_pathkeys = lappend(useful_pathkeys, pathkey); |
| 5391 |
} |
| 5392 |
|
| 5393 |
/* If no match at all, no point in reordering needed_pathkeys */ |
| 5394 |
if (useful_pathkeys == NIL) |
| 5395 |
return useful_pathkeys_list; |
| 5396 |
|
| 5397 |
/* |
| 5398 |
* If not full match, the resulting pathkey list is not useful without |
| 5399 |
* incremental sort. |
| 5400 |
*/ |
| 5401 |
if (list_length(useful_pathkeys) < list_length(needed_pathkeys) && |
| 5402 |
!enable_incremental_sort) |
| 5403 |
return useful_pathkeys_list; |
| 5404 |
|
| 5405 |
/* Append the remaining PathKey nodes in needed_pathkeys */ |
| 5406 |
useful_pathkeys = list_concat_unique_ptr(useful_pathkeys, |
| 5407 |
needed_pathkeys); |
| 5408 |
|
| 5409 |
/* |
| 5410 |
* If the resulting pathkey list is the same as the 'needed_pathkeys', |
| 5411 |
* just drop it. |
| 5412 |
*/ |
| 5413 |
if (compare_pathkeys(needed_pathkeys, |
| 5414 |
useful_pathkeys) == PATHKEYS_EQUAL) |
| 5415 |
return useful_pathkeys_list; |
| 5416 |
|
| 5417 |
useful_pathkeys_list = lappend(useful_pathkeys_list, |
| 5418 |
useful_pathkeys); |
| 5419 |
|
| 5420 |
return useful_pathkeys_list; |
| 5421 |
} |
| 5422 |
|
| 5423 |
/* |
| 5424 |
* create_ordered_paths |
| 5425 |
* |
| 5426 |
* Build a new upperrel containing Paths for ORDER BY evaluation. |
| 5427 |
* |
| 5428 |
* All paths in the result must satisfy the ORDER BY ordering. |
| 5429 |
* The only new paths we need consider are an explicit full sort |
| 5430 |
* and incremental sort on the cheapest-total existing path. |
| 5431 |
* |
| 5432 |
* input_rel: contains the source-data Paths |
| 5433 |
* target: the output tlist the result Paths must emit |
| 5434 |
* limit_tuples: estimated bound on the number of output tuples, |
| 5435 |
* or -1 if no LIMIT or couldn't estimate |
| 5436 |
* |
| 5437 |
* XXX This only looks at sort_pathkeys. I wonder if it needs to look at the |
| 5438 |
* other pathkeys (grouping, ...) like generate_useful_gather_paths. |
| 5439 |
*/ |
| 5440 |
static RelOptInfo * |
| 5441 |
create_ordered_paths(PlannerInfo *root, |
| 5442 |
RelOptInfo *input_rel, |
| 5443 |
PathTarget *target, |
| 5444 |
bool target_parallel_safe, |
| 5445 |
double limit_tuples) |
| 5446 |
{ |
| 5447 |
Path *cheapest_input_path = input_rel->cheapest_total_path; |
| 5448 |
RelOptInfo *ordered_rel; |
| 5449 |
ListCell *lc; |
| 5450 |
|
| 5451 |
/* For now, do all work in the (ORDERED, NULL) upperrel */ |
| 5452 |
ordered_rel = fetch_upper_rel(root, UPPERREL_ORDERED, NULL); |
| 5453 |
|
| 5454 |
/* |
| 5455 |
* If the input relation is not parallel-safe, then the ordered relation |
| 5456 |
* can't be parallel-safe, either. Otherwise, it's parallel-safe if the |
| 5457 |
* target list is parallel-safe. |
| 5458 |
*/ |
| 5459 |
if (input_rel->consider_parallel && target_parallel_safe) |
| 5460 |
ordered_rel->consider_parallel = true; |
| 5461 |
|
| 5462 |
/* |
| 5463 |
* If the input rel belongs to a single FDW, so does the ordered_rel. |
| 5464 |
*/ |
| 5465 |
ordered_rel->serverid = input_rel->serverid; |
| 5466 |
ordered_rel->userid = input_rel->userid; |
| 5467 |
ordered_rel->useridiscurrent = input_rel->useridiscurrent; |
| 5468 |
ordered_rel->fdwroutine = input_rel->fdwroutine; |
| 5469 |
|
| 5470 |
foreach(lc, input_rel->pathlist) |
| 5471 |
{ |
| 5472 |
Path *input_path = (Path *) lfirst(lc); |
| 5473 |
Path *sorted_path; |
| 5474 |
bool is_sorted; |
| 5475 |
int presorted_keys; |
| 5476 |
|
| 5477 |
is_sorted = pathkeys_count_contained_in(root->sort_pathkeys, |
| 5478 |
input_path->pathkeys, &presorted_keys); |
| 5479 |
|
| 5480 |
if (is_sorted) |
| 5481 |
sorted_path = input_path; |
| 5482 |
else |
| 5483 |
{ |
| 5484 |
/* |
| 5485 |
* Try at least sorting the cheapest path and also try |
| 5486 |
* incrementally sorting any path which is partially sorted |
| 5487 |
* already (no need to deal with paths which have presorted keys |
| 5488 |
* when incremental sort is disabled unless it's the cheapest |
| 5489 |
* input path). |
| 5490 |
*/ |
| 5491 |
if (input_path != cheapest_input_path && |
| 5492 |
(presorted_keys == 0 || !enable_incremental_sort)) |
| 5493 |
continue; |
| 5494 |
|
| 5495 |
/* |
| 5496 |
* We've no need to consider both a sort and incremental sort. |
| 5497 |
* We'll just do a sort if there are no presorted keys and an |
| 5498 |
* incremental sort when there are presorted keys. |
| 5499 |
*/ |
| 5500 |
if (presorted_keys == 0 || !enable_incremental_sort) |
| 5501 |
sorted_path = (Path *) create_sort_path(root, |
| 5502 |
ordered_rel, |
| 5503 |
input_path, |
| 5504 |
root->sort_pathkeys, |
| 5505 |
limit_tuples); |
| 5506 |
else |
| 5507 |
sorted_path = (Path *) create_incremental_sort_path(root, |
| 5508 |
ordered_rel, |
| 5509 |
input_path, |
| 5510 |
root->sort_pathkeys, |
| 5511 |
presorted_keys, |
| 5512 |
limit_tuples); |
| 5513 |
} |
| 5514 |
|
| 5515 |
/* |
| 5516 |
* If the pathtarget of the result path has different expressions from |
| 5517 |
* the target to be applied, a projection step is needed. |
| 5518 |
*/ |
| 5519 |
if (!equal(sorted_path->pathtarget->exprs, target->exprs)) |
| 5520 |
sorted_path = apply_projection_to_path(root, ordered_rel, |
| 5521 |
sorted_path, target); |
| 5522 |
|
| 5523 |
add_path(ordered_rel, sorted_path); |
| 5524 |
} |
| 5525 |
|
| 5526 |
/* |
| 5527 |
* generate_gather_paths() will have already generated a simple Gather |
| 5528 |
* path for the best parallel path, if any, and the loop above will have |
| 5529 |
* considered sorting it. Similarly, generate_gather_paths() will also |
| 5530 |
* have generated order-preserving Gather Merge plans which can be used |
| 5531 |
* without sorting if they happen to match the sort_pathkeys, and the loop |
| 5532 |
* above will have handled those as well. However, there's one more |
| 5533 |
* possibility: it may make sense to sort the cheapest partial path or |
| 5534 |
* incrementally sort any partial path that is partially sorted according |
| 5535 |
* to the required output order and then use Gather Merge. |
| 5536 |
*/ |
| 5537 |
if (ordered_rel->consider_parallel && root->sort_pathkeys != NIL && |
| 5538 |
input_rel->partial_pathlist != NIL) |
| 5539 |
{ |
| 5540 |
Path *cheapest_partial_path; |
| 5541 |
|
| 5542 |
cheapest_partial_path = linitial(input_rel->partial_pathlist); |
| 5543 |
|
| 5544 |
foreach(lc, input_rel->partial_pathlist) |
| 5545 |
{ |
| 5546 |
Path *input_path = (Path *) lfirst(lc); |
| 5547 |
Path *sorted_path; |
| 5548 |
bool is_sorted; |
| 5549 |
int presorted_keys; |
| 5550 |
double total_groups; |
| 5551 |
|
| 5552 |
is_sorted = pathkeys_count_contained_in(root->sort_pathkeys, |
| 5553 |
input_path->pathkeys, |
| 5554 |
&presorted_keys); |
| 5555 |
|
| 5556 |
if (is_sorted) |
| 5557 |
continue; |
| 5558 |
|
| 5559 |
/* |
| 5560 |
* Try at least sorting the cheapest path and also try |
| 5561 |
* incrementally sorting any path which is partially sorted |
| 5562 |
* already (no need to deal with paths which have presorted keys |
| 5563 |
* when incremental sort is disabled unless it's the cheapest |
| 5564 |
* partial path). |
| 5565 |
*/ |
| 5566 |
if (input_path != cheapest_partial_path && |
| 5567 |
(presorted_keys == 0 || !enable_incremental_sort)) |
| 5568 |
continue; |
| 5569 |
|
| 5570 |
/* |
| 5571 |
* We've no need to consider both a sort and incremental sort. |
| 5572 |
* We'll just do a sort if there are no presorted keys and an |
| 5573 |
* incremental sort when there are presorted keys. |
| 5574 |
*/ |
| 5575 |
if (presorted_keys == 0 || !enable_incremental_sort) |
| 5576 |
sorted_path = (Path *) create_sort_path(root, |
| 5577 |
ordered_rel, |
| 5578 |
input_path, |
| 5579 |
root->sort_pathkeys, |
| 5580 |
limit_tuples); |
| 5581 |
else |
| 5582 |
sorted_path = (Path *) create_incremental_sort_path(root, |
| 5583 |
ordered_rel, |
| 5584 |
input_path, |
| 5585 |
root->sort_pathkeys, |
| 5586 |
presorted_keys, |
| 5587 |
limit_tuples); |
| 5588 |
total_groups = compute_gather_rows(sorted_path); |
| 5589 |
sorted_path = (Path *) |
| 5590 |
create_gather_merge_path(root, ordered_rel, |
| 5591 |
sorted_path, |
| 5592 |
sorted_path->pathtarget, |
| 5593 |
root->sort_pathkeys, NULL, |
| 5594 |
&total_groups); |
| 5595 |
|
| 5596 |
/* |
| 5597 |
* If the pathtarget of the result path has different expressions |
| 5598 |
* from the target to be applied, a projection step is needed. |
| 5599 |
*/ |
| 5600 |
if (!equal(sorted_path->pathtarget->exprs, target->exprs)) |
| 5601 |
sorted_path = apply_projection_to_path(root, ordered_rel, |
| 5602 |
sorted_path, target); |
| 5603 |
|
| 5604 |
add_path(ordered_rel, sorted_path); |
| 5605 |
} |
| 5606 |
} |
| 5607 |
|
| 5608 |
/* |
| 5609 |
* If there is an FDW that's responsible for all baserels of the query, |
| 5610 |
* let it consider adding ForeignPaths. |
| 5611 |
*/ |
| 5612 |
if (ordered_rel->fdwroutine && |
| 5613 |
ordered_rel->fdwroutine->GetForeignUpperPaths) |
| 5614 |
ordered_rel->fdwroutine->GetForeignUpperPaths(root, UPPERREL_ORDERED, |
| 5615 |
input_rel, ordered_rel, |
| 5616 |
NULL); |
| 5617 |
|
| 5618 |
/* Let extensions possibly add some more paths */ |
| 5619 |
if (create_upper_paths_hook) |
| 5620 |
(*create_upper_paths_hook) (root, UPPERREL_ORDERED, |
| 5621 |
input_rel, ordered_rel, NULL); |
| 5622 |
|
| 5623 |
/* |
| 5624 |
* No need to bother with set_cheapest here; grouping_planner does not |
| 5625 |
* need us to do it. |
| 5626 |
*/ |
| 5627 |
Assert(ordered_rel->pathlist != NIL); |
| 5628 |
|
| 5629 |
return ordered_rel; |
| 5630 |
} |
| 5631 |
|
| 5632 |
|
| 5633 |
/* |
| 5634 |
* make_group_input_target |
| 5635 |
* Generate appropriate PathTarget for initial input to grouping nodes. |
| 5636 |
* |
| 5637 |
* If there is grouping or aggregation, the scan/join subplan cannot emit |
| 5638 |
* the query's final targetlist; for example, it certainly can't emit any |
| 5639 |
* aggregate function calls. This routine generates the correct target |
| 5640 |
* for the scan/join subplan. |
| 5641 |
* |
| 5642 |
* The query target list passed from the parser already contains entries |
| 5643 |
* for all ORDER BY and GROUP BY expressions, but it will not have entries |
| 5644 |
* for variables used only in HAVING clauses; so we need to add those |
| 5645 |
* variables to the subplan target list. Also, we flatten all expressions |
| 5646 |
* except GROUP BY items into their component variables; other expressions |
| 5647 |
* will be computed by the upper plan nodes rather than by the subplan. |
| 5648 |
* For example, given a query like |
| 5649 |
* SELECT a+b,SUM(c+d) FROM table GROUP BY a+b; |
| 5650 |
* we want to pass this targetlist to the subplan: |
| 5651 |
* a+b,c,d |
| 5652 |
* where the a+b target will be used by the Sort/Group steps, and the |
| 5653 |
* other targets will be used for computing the final results. |
| 5654 |
* |
| 5655 |
* 'final_target' is the query's final target list (in PathTarget form) |
| 5656 |
* |
| 5657 |
* The result is the PathTarget to be computed by the Paths returned from |
| 5658 |
* query_planner(). |
| 5659 |
*/ |
| 5660 |
static PathTarget * |
| 5661 |
make_group_input_target(PlannerInfo *root, PathTarget *final_target) |
| 5662 |
{ |
| 5663 |
Query *parse = root->parse; |
| 5664 |
PathTarget *input_target; |
| 5665 |
List *non_group_cols; |
| 5666 |
List *non_group_vars; |
| 5667 |
int i; |
| 5668 |
ListCell *lc; |
| 5669 |
|
| 5670 |
/* |
| 5671 |
* We must build a target containing all grouping columns, plus any other |
| 5672 |
* Vars mentioned in the query's targetlist and HAVING qual. |
| 5673 |
*/ |
| 5674 |
input_target = create_empty_pathtarget(); |
| 5675 |
non_group_cols = NIL; |
| 5676 |
|
| 5677 |
i = 0; |
| 5678 |
foreach(lc, final_target->exprs) |
| 5679 |
{ |
| 5680 |
Expr *expr = (Expr *) lfirst(lc); |
| 5681 |
Index sgref = get_pathtarget_sortgroupref(final_target, i); |
| 5682 |
|
| 5683 |
if (sgref && root->processed_groupClause && |
| 5684 |
get_sortgroupref_clause_noerr(sgref, |
| 5685 |
root->processed_groupClause) != NULL) |
| 5686 |
{ |
| 5687 |
/* |
| 5688 |
* It's a grouping column, so add it to the input target as-is. |
| 5689 |
* |
| 5690 |
* Note that the target is logically below the grouping step. So |
| 5691 |
* with grouping sets we need to remove the RT index of the |
| 5692 |
* grouping step if there is any from the target expression. |
| 5693 |
*/ |
| 5694 |
if (parse->hasGroupRTE && parse->groupingSets != NIL) |
| 5695 |
{ |
| 5696 |
Assert(root->group_rtindex > 0); |
| 5697 |
expr = (Expr *) |
| 5698 |
remove_nulling_relids((Node *) expr, |
| 5699 |
bms_make_singleton(root->group_rtindex), |
| 5700 |
NULL); |
| 5701 |
} |
| 5702 |
add_column_to_pathtarget(input_target, expr, sgref); |
| 5703 |
} |
| 5704 |
else |
| 5705 |
{ |
| 5706 |
/* |
| 5707 |
* Non-grouping column, so just remember the expression for later |
| 5708 |
* call to pull_var_clause. |
| 5709 |
*/ |
| 5710 |
non_group_cols = lappend(non_group_cols, expr); |
| 5711 |
} |
| 5712 |
|
| 5713 |
i++; |
| 5714 |
} |
| 5715 |
|
| 5716 |
/* |
| 5717 |
* If there's a HAVING clause, we'll need the Vars it uses, too. |
| 5718 |
*/ |
| 5719 |
if (parse->havingQual) |
| 5720 |
non_group_cols = lappend(non_group_cols, parse->havingQual); |
| 5721 |
|
| 5722 |
/* |
| 5723 |
* Pull out all the Vars mentioned in non-group cols (plus HAVING), and |
| 5724 |
* add them to the input target if not already present. (A Var used |
| 5725 |
* directly as a GROUP BY item will be present already.) Note this |
| 5726 |
* includes Vars used in resjunk items, so we are covering the needs of |
| 5727 |
* ORDER BY and window specifications. Vars used within Aggrefs and |
| 5728 |
* WindowFuncs will be pulled out here, too. |
| 5729 |
* |
| 5730 |
* Note that the target is logically below the grouping step. So with |
| 5731 |
* grouping sets we need to remove the RT index of the grouping step if |
| 5732 |
* there is any from the non-group Vars. |
| 5733 |
*/ |
| 5734 |
non_group_vars = pull_var_clause((Node *) non_group_cols, |
| 5735 |
PVC_RECURSE_AGGREGATES | |
| 5736 |
PVC_RECURSE_WINDOWFUNCS | |
| 5737 |
PVC_INCLUDE_PLACEHOLDERS); |
| 5738 |
if (parse->hasGroupRTE && parse->groupingSets != NIL) |
| 5739 |
{ |
| 5740 |
Assert(root->group_rtindex > 0); |
| 5741 |
non_group_vars = (List *) |
| 5742 |
remove_nulling_relids((Node *) non_group_vars, |
| 5743 |
bms_make_singleton(root->group_rtindex), |
| 5744 |
NULL); |
| 5745 |
} |
| 5746 |
add_new_columns_to_pathtarget(input_target, non_group_vars); |
| 5747 |
|
| 5748 |
/* clean up cruft */ |
| 5749 |
list_free(non_group_vars); |
| 5750 |
list_free(non_group_cols); |
| 5751 |
|
| 5752 |
/* XXX this causes some redundant cost calculation ... */ |
| 5753 |
return set_pathtarget_cost_width(root, input_target); |
| 5754 |
} |
| 5755 |
|
| 5756 |
/* |
| 5757 |
* make_partial_grouping_target |
| 5758 |
* Generate appropriate PathTarget for output of partial aggregate |
| 5759 |
* (or partial grouping, if there are no aggregates) nodes. |
| 5760 |
* |
| 5761 |
* A partial aggregation node needs to emit all the same aggregates that |
| 5762 |
* a regular aggregation node would, plus any aggregates used in HAVING; |
| 5763 |
* except that the Aggref nodes should be marked as partial aggregates. |
| 5764 |
* |
| 5765 |
* In addition, we'd better emit any Vars and PlaceHolderVars that are |
| 5766 |
* used outside of Aggrefs in the aggregation tlist and HAVING. (Presumably, |
| 5767 |
* these would be Vars that are grouped by or used in grouping expressions.) |
| 5768 |
* |
| 5769 |
* grouping_target is the tlist to be emitted by the topmost aggregation step. |
| 5770 |
* havingQual represents the HAVING clause. |
| 5771 |
*/ |
| 5772 |
static PathTarget * |
| 5773 |
make_partial_grouping_target(PlannerInfo *root, |
| 5774 |
PathTarget *grouping_target, |
| 5775 |
Node *havingQual) |
| 5776 |
{ |
| 5777 |
PathTarget *partial_target; |
| 5778 |
List *non_group_cols; |
| 5779 |
List *non_group_exprs; |
| 5780 |
int i; |
| 5781 |
ListCell *lc; |
| 5782 |
|
| 5783 |
partial_target = create_empty_pathtarget(); |
| 5784 |
non_group_cols = NIL; |
| 5785 |
|
| 5786 |
i = 0; |
| 5787 |
foreach(lc, grouping_target->exprs) |
| 5788 |
{ |
| 5789 |
Expr *expr = (Expr *) lfirst(lc); |
| 5790 |
Index sgref = get_pathtarget_sortgroupref(grouping_target, i); |
| 5791 |
|
| 5792 |
if (sgref && root->processed_groupClause && |
| 5793 |
get_sortgroupref_clause_noerr(sgref, |
| 5794 |
root->processed_groupClause) != NULL) |
| 5795 |
{ |
| 5796 |
/* |
| 5797 |
* It's a grouping column, so add it to the partial_target as-is. |
| 5798 |
* (This allows the upper agg step to repeat the grouping calcs.) |
| 5799 |
*/ |
| 5800 |
add_column_to_pathtarget(partial_target, expr, sgref); |
| 5801 |
} |
| 5802 |
else |
| 5803 |
{ |
| 5804 |
/* |
| 5805 |
* Non-grouping column, so just remember the expression for later |
| 5806 |
* call to pull_var_clause. |
| 5807 |
*/ |
| 5808 |
non_group_cols = lappend(non_group_cols, expr); |
| 5809 |
} |
| 5810 |
|
| 5811 |
i++; |
| 5812 |
} |
| 5813 |
|
| 5814 |
/* |
| 5815 |
* If there's a HAVING clause, we'll need the Vars/Aggrefs it uses, too. |
| 5816 |
*/ |
| 5817 |
if (havingQual) |
| 5818 |
non_group_cols = lappend(non_group_cols, havingQual); |
| 5819 |
|
| 5820 |
/* |
| 5821 |
* Pull out all the Vars, PlaceHolderVars, and Aggrefs mentioned in |
| 5822 |
* non-group cols (plus HAVING), and add them to the partial_target if not |
| 5823 |
* already present. (An expression used directly as a GROUP BY item will |
| 5824 |
* be present already.) Note this includes Vars used in resjunk items, so |
| 5825 |
* we are covering the needs of ORDER BY and window specifications. |
| 5826 |
*/ |
| 5827 |
non_group_exprs = pull_var_clause((Node *) non_group_cols, |
| 5828 |
PVC_INCLUDE_AGGREGATES | |
| 5829 |
PVC_RECURSE_WINDOWFUNCS | |
| 5830 |
PVC_INCLUDE_PLACEHOLDERS); |
| 5831 |
|
| 5832 |
add_new_columns_to_pathtarget(partial_target, non_group_exprs); |
| 5833 |
|
| 5834 |
/* |
| 5835 |
* Adjust Aggrefs to put them in partial mode. At this point all Aggrefs |
| 5836 |
* are at the top level of the target list, so we can just scan the list |
| 5837 |
* rather than recursing through the expression trees. |
| 5838 |
*/ |
| 5839 |
foreach(lc, partial_target->exprs) |
| 5840 |
{ |
| 5841 |
Aggref *aggref = (Aggref *) lfirst(lc); |
| 5842 |
|
| 5843 |
if (IsA(aggref, Aggref)) |
| 5844 |
{ |
| 5845 |
Aggref *newaggref; |
| 5846 |
|
| 5847 |
/* |
| 5848 |
* We shouldn't need to copy the substructure of the Aggref node, |
| 5849 |
* but flat-copy the node itself to avoid damaging other trees. |
| 5850 |
*/ |
| 5851 |
newaggref = makeNode(Aggref); |
| 5852 |
memcpy(newaggref, aggref, sizeof(Aggref)); |
| 5853 |
|
| 5854 |
/* For now, assume serialization is required */ |
| 5855 |
mark_partial_aggref(newaggref, AGGSPLIT_INITIAL_SERIAL); |
| 5856 |
|
| 5857 |
lfirst(lc) = newaggref; |
| 5858 |
} |
| 5859 |
} |
| 5860 |
|
| 5861 |
/* clean up cruft */ |
| 5862 |
list_free(non_group_exprs); |
| 5863 |
list_free(non_group_cols); |
| 5864 |
|
| 5865 |
/* XXX this causes some redundant cost calculation ... */ |
| 5866 |
return set_pathtarget_cost_width(root, partial_target); |
| 5867 |
} |
| 5868 |
|
| 5869 |
/* |
| 5870 |
* mark_partial_aggref |
| 5871 |
* Adjust an Aggref to make it represent a partial-aggregation step. |
| 5872 |
* |
| 5873 |
* The Aggref node is modified in-place; caller must do any copying required. |
| 5874 |
*/ |
| 5875 |
void |
| 5876 |
mark_partial_aggref(Aggref *agg, AggSplit aggsplit) |
| 5877 |
{ |
| 5878 |
/* aggtranstype should be computed by this point */ |
| 5879 |
Assert(OidIsValid(agg->aggtranstype)); |
| 5880 |
/* ... but aggsplit should still be as the parser left it */ |
| 5881 |
Assert(agg->aggsplit == AGGSPLIT_SIMPLE); |
| 5882 |
|
| 5883 |
/* Mark the Aggref with the intended partial-aggregation mode */ |
| 5884 |
agg->aggsplit = aggsplit; |
| 5885 |
|
| 5886 |
/* |
| 5887 |
* Adjust result type if needed. Normally, a partial aggregate returns |
| 5888 |
* the aggregate's transition type; but if that's INTERNAL and we're |
| 5889 |
* serializing, it returns BYTEA instead. |
| 5890 |
*/ |
| 5891 |
if (DO_AGGSPLIT_SKIPFINAL(aggsplit)) |
| 5892 |
{ |
| 5893 |
if (agg->aggtranstype == INTERNALOID && DO_AGGSPLIT_SERIALIZE(aggsplit)) |
| 5894 |
agg->aggtype = BYTEAOID; |
| 5895 |
else |
| 5896 |
agg->aggtype = agg->aggtranstype; |
| 5897 |
} |
| 5898 |
} |
| 5899 |
|
| 5900 |
/* |
| 5901 |
* postprocess_setop_tlist |
| 5902 |
* Fix up targetlist returned by plan_set_operations(). |
| 5903 |
* |
| 5904 |
* We need to transpose sort key info from the orig_tlist into new_tlist. |
| 5905 |
* NOTE: this would not be good enough if we supported resjunk sort keys |
| 5906 |
* for results of set operations --- then, we'd need to project a whole |
| 5907 |
* new tlist to evaluate the resjunk columns. For now, just ereport if we |
| 5908 |
* find any resjunk columns in orig_tlist. |
| 5909 |
*/ |
| 5910 |
static List * |
| 5911 |
postprocess_setop_tlist(List *new_tlist, List *orig_tlist) |
| 5912 |
{ |
| 5913 |
ListCell *l; |
| 5914 |
ListCell *orig_tlist_item = list_head(orig_tlist); |
| 5915 |
|
| 5916 |
foreach(l, new_tlist) |
| 5917 |
{ |
| 5918 |
TargetEntry *new_tle = lfirst_node(TargetEntry, l); |
| 5919 |
TargetEntry *orig_tle; |
| 5920 |
|
| 5921 |
/* ignore resjunk columns in setop result */ |
| 5922 |
if (new_tle->resjunk) |
| 5923 |
continue; |
| 5924 |
|
| 5925 |
Assert(orig_tlist_item != NULL); |
| 5926 |
orig_tle = lfirst_node(TargetEntry, orig_tlist_item); |
| 5927 |
orig_tlist_item = lnext(orig_tlist, orig_tlist_item); |
| 5928 |
if (orig_tle->resjunk) /* should not happen */ |
| 5929 |
elog(ERROR, "resjunk output columns are not implemented"); |
| 5930 |
Assert(new_tle->resno == orig_tle->resno); |
| 5931 |
new_tle->ressortgroupref = orig_tle->ressortgroupref; |
| 5932 |
} |
| 5933 |
if (orig_tlist_item != NULL) |
| 5934 |
elog(ERROR, "resjunk output columns are not implemented"); |
| 5935 |
return new_tlist; |
| 5936 |
} |
| 5937 |
|
| 5938 |
/* |
| 5939 |
* optimize_window_clauses |
| 5940 |
* Call each WindowFunc's prosupport function to see if we're able to |
| 5941 |
* make any adjustments to any of the WindowClause's so that the executor |
| 5942 |
* can execute the window functions in a more optimal way. |
| 5943 |
* |
| 5944 |
* Currently we only allow adjustments to the WindowClause's frameOptions. We |
| 5945 |
* may allow more things to be done here in the future. |
| 5946 |
*/ |
| 5947 |
static void |
| 5948 |
optimize_window_clauses(PlannerInfo *root, WindowFuncLists *wflists) |
| 5949 |
{ |
| 5950 |
List *windowClause = root->parse->windowClause; |
| 5951 |
ListCell *lc; |
| 5952 |
|
| 5953 |
foreach(lc, windowClause) |
| 5954 |
{ |
| 5955 |
WindowClause *wc = lfirst_node(WindowClause, lc); |
| 5956 |
ListCell *lc2; |
| 5957 |
int optimizedFrameOptions = 0; |
| 5958 |
|
| 5959 |
Assert(wc->winref <= wflists->maxWinRef); |
| 5960 |
|
| 5961 |
/* skip any WindowClauses that have no WindowFuncs */ |
| 5962 |
if (wflists->windowFuncs[wc->winref] == NIL) |
| 5963 |
continue; |
| 5964 |
|
| 5965 |
foreach(lc2, wflists->windowFuncs[wc->winref]) |
| 5966 |
{ |
| 5967 |
SupportRequestOptimizeWindowClause req; |
| 5968 |
SupportRequestOptimizeWindowClause *res; |
| 5969 |
WindowFunc *wfunc = lfirst_node(WindowFunc, lc2); |
| 5970 |
Oid prosupport; |
| 5971 |
|
| 5972 |
prosupport = get_func_support(wfunc->winfnoid); |
| 5973 |
|
| 5974 |
/* Check if there's a support function for 'wfunc' */ |
| 5975 |
if (!OidIsValid(prosupport)) |
| 5976 |
break; /* can't optimize this WindowClause */ |
| 5977 |
|
| 5978 |
req.type = T_SupportRequestOptimizeWindowClause; |
| 5979 |
req.window_clause = wc; |
| 5980 |
req.window_func = wfunc; |
| 5981 |
req.frameOptions = wc->frameOptions; |
| 5982 |
|
| 5983 |
/* call the support function */ |
| 5984 |
res = (SupportRequestOptimizeWindowClause *) |
| 5985 |
DatumGetPointer(OidFunctionCall1(prosupport, |
| 5986 |
PointerGetDatum(&req))); |
| 5987 |
|
| 5988 |
/* |
| 5989 |
* Skip to next WindowClause if the support function does not |
| 5990 |
* support this request type. |
| 5991 |
*/ |
| 5992 |
if (res == NULL) |
| 5993 |
break; |
| 5994 |
|
| 5995 |
/* |
| 5996 |
* Save these frameOptions for the first WindowFunc for this |
| 5997 |
* WindowClause. |
| 5998 |
*/ |
| 5999 |
if (foreach_current_index(lc2) == 0) |
| 6000 |
optimizedFrameOptions = res->frameOptions; |
| 6001 |
|
| 6002 |
/* |
| 6003 |
* On subsequent WindowFuncs, if the frameOptions are not the same |
| 6004 |
* then we're unable to optimize the frameOptions for this |
| 6005 |
* WindowClause. |
| 6006 |
*/ |
| 6007 |
else if (optimizedFrameOptions != res->frameOptions) |
| 6008 |
break; /* skip to the next WindowClause, if any */ |
| 6009 |
} |
| 6010 |
|
| 6011 |
/* adjust the frameOptions if all WindowFunc's agree that it's ok */ |
| 6012 |
if (lc2 == NULL && wc->frameOptions != optimizedFrameOptions) |
| 6013 |
{ |
| 6014 |
ListCell *lc3; |
| 6015 |
|
| 6016 |
/* apply the new frame options */ |
| 6017 |
wc->frameOptions = optimizedFrameOptions; |
| 6018 |
|
| 6019 |
/* |
| 6020 |
* We now check to see if changing the frameOptions has caused |
| 6021 |
* this WindowClause to be a duplicate of some other WindowClause. |
| 6022 |
* This can only happen if we have multiple WindowClauses, so |
| 6023 |
* don't bother if there's only 1. |
| 6024 |
*/ |
| 6025 |
if (list_length(windowClause) == 1) |
| 6026 |
continue; |
| 6027 |
|
| 6028 |
/* |
| 6029 |
* Do the duplicate check and reuse the existing WindowClause if |
| 6030 |
* we find a duplicate. |
| 6031 |
*/ |
| 6032 |
foreach(lc3, windowClause) |
| 6033 |
{ |
| 6034 |
WindowClause *existing_wc = lfirst_node(WindowClause, lc3); |
| 6035 |
|
| 6036 |
/* skip over the WindowClause we're currently editing */ |
| 6037 |
if (existing_wc == wc) |
| 6038 |
continue; |
| 6039 |
|
| 6040 |
/* |
| 6041 |
* Perform the same duplicate check that is done in |
| 6042 |
* transformWindowFuncCall. |
| 6043 |
*/ |
| 6044 |
if (equal(wc->partitionClause, existing_wc->partitionClause) && |
| 6045 |
equal(wc->orderClause, existing_wc->orderClause) && |
| 6046 |
wc->frameOptions == existing_wc->frameOptions && |
| 6047 |
equal(wc->startOffset, existing_wc->startOffset) && |
| 6048 |
equal(wc->endOffset, existing_wc->endOffset)) |
| 6049 |
{ |
| 6050 |
ListCell *lc4; |
| 6051 |
|
| 6052 |
/* |
| 6053 |
* Now move each WindowFunc in 'wc' into 'existing_wc'. |
| 6054 |
* This required adjusting each WindowFunc's winref and |
| 6055 |
* moving the WindowFuncs in 'wc' to the list of |
| 6056 |
* WindowFuncs in 'existing_wc'. |
| 6057 |
*/ |
| 6058 |
foreach(lc4, wflists->windowFuncs[wc->winref]) |
| 6059 |
{ |
| 6060 |
WindowFunc *wfunc = lfirst_node(WindowFunc, lc4); |
| 6061 |
|
| 6062 |
wfunc->winref = existing_wc->winref; |
| 6063 |
} |
| 6064 |
|
| 6065 |
/* move list items */ |
| 6066 |
wflists->windowFuncs[existing_wc->winref] = list_concat(wflists->windowFuncs[existing_wc->winref], |
| 6067 |
wflists->windowFuncs[wc->winref]); |
| 6068 |
wflists->windowFuncs[wc->winref] = NIL; |
| 6069 |
|
| 6070 |
/* |
| 6071 |
* transformWindowFuncCall() should have made sure there |
| 6072 |
* are no other duplicates, so we needn't bother looking |
| 6073 |
* any further. |
| 6074 |
*/ |
| 6075 |
break; |
| 6076 |
} |
| 6077 |
} |
| 6078 |
} |
| 6079 |
} |
| 6080 |
|
| 6081 |
/* |
| 6082 |
* XXX remove any duplicate WindowFuncs from each WindowClause. This has |
| 6083 |
* been done only in the back branches. Previously, the deduplication was |
| 6084 |
* done in find_window_functions(), but that caused issues with the code |
| 6085 |
* above when moving a WindowFunc to another WindowClause as any duplicate |
| 6086 |
* WindowFuncs won't receive the adjusted winref when merging |
| 6087 |
* WindowClauses. The deduplication below has been done only so that we |
| 6088 |
* maintain the same cost calculations. As it turns out, the previous |
| 6089 |
* deduplication code thought it was saving effort during execution by |
| 6090 |
* getting rid of duplicates, but that was not true as the expression |
| 6091 |
* evaluation code will evaluate each WindowFunc mentioned in the |
| 6092 |
* targetlist. |
| 6093 |
*/ |
| 6094 |
foreach(lc, windowClause) |
| 6095 |
{ |
| 6096 |
WindowClause *wc = lfirst_node(WindowClause, lc); |
| 6097 |
ListCell *lc2; |
| 6098 |
List *list = wflists->windowFuncs[wc->winref]; |
| 6099 |
List *newlist = NIL; |
| 6100 |
|
| 6101 |
if (list == NIL) |
| 6102 |
continue; |
| 6103 |
|
| 6104 |
foreach(lc2, list) |
| 6105 |
{ |
| 6106 |
if (!list_member(newlist, lfirst(lc2))) |
| 6107 |
newlist = lappend(newlist, lfirst(lc2)); |
| 6108 |
else |
| 6109 |
wflists->numWindowFuncs--; |
| 6110 |
} |
| 6111 |
list_free(list); |
| 6112 |
|
| 6113 |
wflists->windowFuncs[wc->winref] = newlist; |
| 6114 |
} |
| 6115 |
} |
| 6116 |
|
| 6117 |
/* |
| 6118 |
* select_active_windows |
| 6119 |
* Create a list of the "active" window clauses (ie, those referenced |
| 6120 |
* by non-deleted WindowFuncs) in the order they are to be executed. |
| 6121 |
*/ |
| 6122 |
static List * |
| 6123 |
select_active_windows(PlannerInfo *root, WindowFuncLists *wflists) |
| 6124 |
{ |
| 6125 |
List *windowClause = root->parse->windowClause; |
| 6126 |
List *result = NIL; |
| 6127 |
ListCell *lc; |
| 6128 |
int nActive = 0; |
| 6129 |
WindowClauseSortData *actives = palloc(sizeof(WindowClauseSortData) |
| 6130 |
* list_length(windowClause)); |
| 6131 |
|
| 6132 |
/* First, construct an array of the active windows */ |
| 6133 |
foreach(lc, windowClause) |
| 6134 |
{ |
| 6135 |
WindowClause *wc = lfirst_node(WindowClause, lc); |
| 6136 |
|
| 6137 |
/* It's only active if wflists shows some related WindowFuncs */ |
| 6138 |
Assert(wc->winref <= wflists->maxWinRef); |
| 6139 |
if (wflists->windowFuncs[wc->winref] == NIL) |
| 6140 |
continue; |
| 6141 |
|
| 6142 |
actives[nActive].wc = wc; /* original clause */ |
| 6143 |
|
| 6144 |
/* |
| 6145 |
* For sorting, we want the list of partition keys followed by the |
| 6146 |
* list of sort keys. But pathkeys construction will remove duplicates |
| 6147 |
* between the two, so we can as well (even though we can't detect all |
| 6148 |
* of the duplicates, since some may come from ECs - that might mean |
| 6149 |
* we miss optimization chances here). We must, however, ensure that |
| 6150 |
* the order of entries is preserved with respect to the ones we do |
| 6151 |
* keep. |
| 6152 |
* |
| 6153 |
* partitionClause and orderClause had their own duplicates removed in |
| 6154 |
* parse analysis, so we're only concerned here with removing |
| 6155 |
* orderClause entries that also appear in partitionClause. |
| 6156 |
*/ |
| 6157 |
actives[nActive].uniqueOrder = |
| 6158 |
list_concat_unique(list_copy(wc->partitionClause), |
| 6159 |
wc->orderClause); |
| 6160 |
nActive++; |
| 6161 |
} |
| 6162 |
|
| 6163 |
/* |
| 6164 |
* Sort active windows by their partitioning/ordering clauses, ignoring |
| 6165 |
* any framing clauses, so that the windows that need the same sorting are |
| 6166 |
* adjacent in the list. When we come to generate paths, this will avoid |
| 6167 |
* inserting additional Sort nodes. |
| 6168 |
* |
| 6169 |
* This is how we implement a specific requirement from the SQL standard, |
| 6170 |
* which says that when two or more windows are order-equivalent (i.e. |
| 6171 |
* have matching partition and order clauses, even if their names or |
| 6172 |
* framing clauses differ), then all peer rows must be presented in the |
| 6173 |
* same order in all of them. If we allowed multiple sort nodes for such |
| 6174 |
* cases, we'd risk having the peer rows end up in different orders in |
| 6175 |
* equivalent windows due to sort instability. (See General Rule 4 of |
| 6176 |
* <window clause> in SQL2008 - SQL2016.) |
| 6177 |
* |
| 6178 |
* Additionally, if the entire list of clauses of one window is a prefix |
| 6179 |
* of another, put first the window with stronger sorting requirements. |
| 6180 |
* This way we will first sort for stronger window, and won't have to sort |
| 6181 |
* again for the weaker one. |
| 6182 |
*/ |
| 6183 |
qsort(actives, nActive, sizeof(WindowClauseSortData), common_prefix_cmp); |
| 6184 |
|
| 6185 |
/* build ordered list of the original WindowClause nodes */ |
| 6186 |
for (int i = 0; i < nActive; i++) |
| 6187 |
result = lappend(result, actives[i].wc); |
| 6188 |
|
| 6189 |
pfree(actives); |
| 6190 |
|
| 6191 |
return result; |
| 6192 |
} |
| 6193 |
|
| 6194 |
/* |
| 6195 |
* name_active_windows |
| 6196 |
* Ensure all active windows have unique names. |
| 6197 |
* |
| 6198 |
* The parser will have checked that user-assigned window names are unique |
| 6199 |
* within the Query. Here we assign made-up names to any unnamed |
| 6200 |
* WindowClauses for the benefit of EXPLAIN. (We don't want to do this |
| 6201 |
* at parse time, because it'd mess up decompilation of views.) |
| 6202 |
* |
| 6203 |
* activeWindows: result of select_active_windows |
| 6204 |
*/ |
| 6205 |
static void |
| 6206 |
name_active_windows(List *activeWindows) |
| 6207 |
{ |
| 6208 |
int next_n = 1; |
| 6209 |
char newname[16]; |
| 6210 |
ListCell *lc; |
| 6211 |
|
| 6212 |
foreach(lc, activeWindows) |
| 6213 |
{ |
| 6214 |
WindowClause *wc = lfirst_node(WindowClause, lc); |
| 6215 |
|
| 6216 |
/* Nothing to do if it has a name already. */ |
| 6217 |
if (wc->name) |
| 6218 |
continue; |
| 6219 |
|
| 6220 |
/* Select a name not currently present in the list. */ |
| 6221 |
for (;;) |
| 6222 |
{ |
| 6223 |
ListCell *lc2; |
| 6224 |
|
| 6225 |
snprintf(newname, sizeof(newname), "w%d", next_n++); |
| 6226 |
foreach(lc2, activeWindows) |
| 6227 |
{ |
| 6228 |
WindowClause *wc2 = lfirst_node(WindowClause, lc2); |
| 6229 |
|
| 6230 |
if (wc2->name && strcmp(wc2->name, newname) == 0) |
| 6231 |
break; /* matched */ |
| 6232 |
} |
| 6233 |
if (lc2 == NULL) |
| 6234 |
break; /* reached the end with no match */ |
| 6235 |
} |
| 6236 |
wc->name = pstrdup(newname); |
| 6237 |
} |
| 6238 |
} |
| 6239 |
|
| 6240 |
/* |
| 6241 |
* common_prefix_cmp |
| 6242 |
* QSort comparison function for WindowClauseSortData |
| 6243 |
* |
| 6244 |
* Sort the windows by the required sorting clauses. First, compare the sort |
| 6245 |
* clauses themselves. Second, if one window's clauses are a prefix of another |
| 6246 |
* one's clauses, put the window with more sort clauses first. |
| 6247 |
* |
| 6248 |
* We purposefully sort by the highest tleSortGroupRef first. Since |
| 6249 |
* tleSortGroupRefs are assigned for the query's DISTINCT and ORDER BY first |
| 6250 |
* and because here we sort the lowest tleSortGroupRefs last, if a |
| 6251 |
* WindowClause is sharing a tleSortGroupRef with the query's DISTINCT or |
| 6252 |
* ORDER BY clause, this makes it more likely that the final WindowAgg will |
| 6253 |
* provide presorted input for the query's DISTINCT or ORDER BY clause, thus |
| 6254 |
* reducing the total number of sorts required for the query. |
| 6255 |
*/ |
| 6256 |
static int |
| 6257 |
common_prefix_cmp(const void *a, const void *b) |
| 6258 |
{ |
| 6259 |
const WindowClauseSortData *wcsa = a; |
| 6260 |
const WindowClauseSortData *wcsb = b; |
| 6261 |
ListCell *item_a; |
| 6262 |
ListCell *item_b; |
| 6263 |
|
| 6264 |
forboth(item_a, wcsa->uniqueOrder, item_b, wcsb->uniqueOrder) |
| 6265 |
{ |
| 6266 |
SortGroupClause *sca = lfirst_node(SortGroupClause, item_a); |
| 6267 |
SortGroupClause *scb = lfirst_node(SortGroupClause, item_b); |
| 6268 |
|
| 6269 |
if (sca->tleSortGroupRef > scb->tleSortGroupRef) |
| 6270 |
return -1; |
| 6271 |
else if (sca->tleSortGroupRef < scb->tleSortGroupRef) |
| 6272 |
return 1; |
| 6273 |
else if (sca->sortop > scb->sortop) |
| 6274 |
return -1; |
| 6275 |
else if (sca->sortop < scb->sortop) |
| 6276 |
return 1; |
| 6277 |
else if (sca->nulls_first && !scb->nulls_first) |
| 6278 |
return -1; |
| 6279 |
else if (!sca->nulls_first && scb->nulls_first) |
| 6280 |
return 1; |
| 6281 |
/* no need to compare eqop, since it is fully determined by sortop */ |
| 6282 |
} |
| 6283 |
|
| 6284 |
if (list_length(wcsa->uniqueOrder) > list_length(wcsb->uniqueOrder)) |
| 6285 |
return -1; |
| 6286 |
else if (list_length(wcsa->uniqueOrder) < list_length(wcsb->uniqueOrder)) |
| 6287 |
return 1; |
| 6288 |
|
| 6289 |
return 0; |
| 6290 |
} |
| 6291 |
|
| 6292 |
/* |
| 6293 |
* make_window_input_target |
| 6294 |
* Generate appropriate PathTarget for initial input to WindowAgg nodes. |
| 6295 |
* |
| 6296 |
* When the query has window functions, this function computes the desired |
| 6297 |
* target to be computed by the node just below the first WindowAgg. |
| 6298 |
* This tlist must contain all values needed to evaluate the window functions, |
| 6299 |
* compute the final target list, and perform any required final sort step. |
| 6300 |
* If multiple WindowAggs are needed, each intermediate one adds its window |
| 6301 |
* function results onto this base tlist; only the topmost WindowAgg computes |
| 6302 |
* the actual desired target list. |
| 6303 |
* |
| 6304 |
* This function is much like make_group_input_target, though not quite enough |
| 6305 |
* like it to share code. As in that function, we flatten most expressions |
| 6306 |
* into their component variables. But we do not want to flatten window |
| 6307 |
* PARTITION BY/ORDER BY clauses, since that might result in multiple |
| 6308 |
* evaluations of them, which would be bad (possibly even resulting in |
| 6309 |
* inconsistent answers, if they contain volatile functions). |
| 6310 |
* Also, we must not flatten GROUP BY clauses that were left unflattened by |
| 6311 |
* make_group_input_target, because we may no longer have access to the |
| 6312 |
* individual Vars in them. |
| 6313 |
* |
| 6314 |
* Another key difference from make_group_input_target is that we don't |
| 6315 |
* flatten Aggref expressions, since those are to be computed below the |
| 6316 |
* window functions and just referenced like Vars above that. |
| 6317 |
* |
| 6318 |
* 'final_target' is the query's final target list (in PathTarget form) |
| 6319 |
* 'activeWindows' is the list of active windows previously identified by |
| 6320 |
* select_active_windows. |
| 6321 |
* |
| 6322 |
* The result is the PathTarget to be computed by the plan node immediately |
| 6323 |
* below the first WindowAgg node. |
| 6324 |
*/ |
| 6325 |
static PathTarget * |
| 6326 |
make_window_input_target(PlannerInfo *root, |
| 6327 |
PathTarget *final_target, |
| 6328 |
List *activeWindows) |
| 6329 |
{ |
| 6330 |
PathTarget *input_target; |
| 6331 |
Bitmapset *sgrefs; |
| 6332 |
List *flattenable_cols; |
| 6333 |
List *flattenable_vars; |
| 6334 |
int i; |
| 6335 |
ListCell *lc; |
| 6336 |
|
| 6337 |
Assert(root->parse->hasWindowFuncs); |
| 6338 |
|
| 6339 |
/* |
| 6340 |
* Collect the sortgroupref numbers of window PARTITION/ORDER BY clauses |
| 6341 |
* into a bitmapset for convenient reference below. |
| 6342 |
*/ |
| 6343 |
sgrefs = NULL; |
| 6344 |
foreach(lc, activeWindows) |
| 6345 |
{ |
| 6346 |
WindowClause *wc = lfirst_node(WindowClause, lc); |
| 6347 |
ListCell *lc2; |
| 6348 |
|
| 6349 |
foreach(lc2, wc->partitionClause) |
| 6350 |
{ |
| 6351 |
SortGroupClause *sortcl = lfirst_node(SortGroupClause, lc2); |
| 6352 |
|
| 6353 |
sgrefs = bms_add_member(sgrefs, sortcl->tleSortGroupRef); |
| 6354 |
} |
| 6355 |
foreach(lc2, wc->orderClause) |
| 6356 |
{ |
| 6357 |
SortGroupClause *sortcl = lfirst_node(SortGroupClause, lc2); |
| 6358 |
|
| 6359 |
sgrefs = bms_add_member(sgrefs, sortcl->tleSortGroupRef); |
| 6360 |
} |
| 6361 |
} |
| 6362 |
|
| 6363 |
/* Add in sortgroupref numbers of GROUP BY clauses, too */ |
| 6364 |
foreach(lc, root->processed_groupClause) |
| 6365 |
{ |
| 6366 |
SortGroupClause *grpcl = lfirst_node(SortGroupClause, lc); |
| 6367 |
|
| 6368 |
sgrefs = bms_add_member(sgrefs, grpcl->tleSortGroupRef); |
| 6369 |
} |
| 6370 |
|
| 6371 |
/* |
| 6372 |
* Construct a target containing all the non-flattenable targetlist items, |
| 6373 |
* and save aside the others for a moment. |
| 6374 |
*/ |
| 6375 |
input_target = create_empty_pathtarget(); |
| 6376 |
flattenable_cols = NIL; |
| 6377 |
|
| 6378 |
i = 0; |
| 6379 |
foreach(lc, final_target->exprs) |
| 6380 |
{ |
| 6381 |
Expr *expr = (Expr *) lfirst(lc); |
| 6382 |
Index sgref = get_pathtarget_sortgroupref(final_target, i); |
| 6383 |
|
| 6384 |
/* |
| 6385 |
* Don't want to deconstruct window clauses or GROUP BY items. (Note |
| 6386 |
* that such items can't contain window functions, so it's okay to |
| 6387 |
* compute them below the WindowAgg nodes.) |
| 6388 |
*/ |
| 6389 |
if (sgref != 0 && bms_is_member(sgref, sgrefs)) |
| 6390 |
{ |
| 6391 |
/* |
| 6392 |
* Don't want to deconstruct this value, so add it to the input |
| 6393 |
* target as-is. |
| 6394 |
*/ |
| 6395 |
add_column_to_pathtarget(input_target, expr, sgref); |
| 6396 |
} |
| 6397 |
else |
| 6398 |
{ |
| 6399 |
/* |
| 6400 |
* Column is to be flattened, so just remember the expression for |
| 6401 |
* later call to pull_var_clause. |
| 6402 |
*/ |
| 6403 |
flattenable_cols = lappend(flattenable_cols, expr); |
| 6404 |
} |
| 6405 |
|
| 6406 |
i++; |
| 6407 |
} |
| 6408 |
|
| 6409 |
/* |
| 6410 |
* Pull out all the Vars and Aggrefs mentioned in flattenable columns, and |
| 6411 |
* add them to the input target if not already present. (Some might be |
| 6412 |
* there already because they're used directly as window/group clauses.) |
| 6413 |
* |
| 6414 |
* Note: it's essential to use PVC_INCLUDE_AGGREGATES here, so that any |
| 6415 |
* Aggrefs are placed in the Agg node's tlist and not left to be computed |
| 6416 |
* at higher levels. On the other hand, we should recurse into |
| 6417 |
* WindowFuncs to make sure their input expressions are available. |
| 6418 |
*/ |
| 6419 |
flattenable_vars = pull_var_clause((Node *) flattenable_cols, |
| 6420 |
PVC_INCLUDE_AGGREGATES | |
| 6421 |
PVC_RECURSE_WINDOWFUNCS | |
| 6422 |
PVC_INCLUDE_PLACEHOLDERS); |
| 6423 |
add_new_columns_to_pathtarget(input_target, flattenable_vars); |
| 6424 |
|
| 6425 |
/* clean up cruft */ |
| 6426 |
list_free(flattenable_vars); |
| 6427 |
list_free(flattenable_cols); |
| 6428 |
|
| 6429 |
/* XXX this causes some redundant cost calculation ... */ |
| 6430 |
return set_pathtarget_cost_width(root, input_target); |
| 6431 |
} |
| 6432 |
|
| 6433 |
/* |
| 6434 |
* make_pathkeys_for_window |
| 6435 |
* Create a pathkeys list describing the required input ordering |
| 6436 |
* for the given WindowClause. |
| 6437 |
* |
| 6438 |
* Modifies wc's partitionClause to remove any clauses which are deemed |
| 6439 |
* redundant by the pathkey logic. |
| 6440 |
* |
| 6441 |
* The required ordering is first the PARTITION keys, then the ORDER keys. |
| 6442 |
* In the future we might try to implement windowing using hashing, in which |
| 6443 |
* case the ordering could be relaxed, but for now we always sort. |
| 6444 |
*/ |
| 6445 |
static List * |
| 6446 |
make_pathkeys_for_window(PlannerInfo *root, WindowClause *wc, |
| 6447 |
List *tlist) |
| 6448 |
{ |
| 6449 |
List *window_pathkeys = NIL; |
| 6450 |
|
| 6451 |
/* Throw error if can't sort */ |
| 6452 |
if (!grouping_is_sortable(wc->partitionClause)) |
| 6453 |
ereport(ERROR, |
| 6454 |
(errcode(ERRCODE_FEATURE_NOT_SUPPORTED), |
| 6455 |
errmsg("could not implement window PARTITION BY"), |
| 6456 |
errdetail("Window partitioning columns must be of sortable datatypes."))); |
| 6457 |
if (!grouping_is_sortable(wc->orderClause)) |
| 6458 |
ereport(ERROR, |
| 6459 |
(errcode(ERRCODE_FEATURE_NOT_SUPPORTED), |
| 6460 |
errmsg("could not implement window ORDER BY"), |
| 6461 |
errdetail("Window ordering columns must be of sortable datatypes."))); |
| 6462 |
|
| 6463 |
/* |
| 6464 |
* First fetch the pathkeys for the PARTITION BY clause. We can safely |
| 6465 |
* remove any clauses from the wc->partitionClause for redundant pathkeys. |
| 6466 |
*/ |
| 6467 |
if (wc->partitionClause != NIL) |
| 6468 |
{ |
| 6469 |
bool sortable; |
| 6470 |
|
| 6471 |
window_pathkeys = make_pathkeys_for_sortclauses_extended(root, |
| 6472 |
&wc->partitionClause, |
| 6473 |
tlist, |
| 6474 |
true, |
| 6475 |
false, |
| 6476 |
&sortable, |
| 6477 |
false); |
| 6478 |
|
| 6479 |
Assert(sortable); |
| 6480 |
} |
| 6481 |
|
| 6482 |
/* |
| 6483 |
* In principle, we could also consider removing redundant ORDER BY items |
| 6484 |
* too as doing so does not alter the result of peer row checks done by |
| 6485 |
* the executor. However, we must *not* remove the ordering column for |
| 6486 |
* RANGE OFFSET cases, as the executor needs that for in_range tests even |
| 6487 |
* if it's known to be equal to some partitioning column. |
| 6488 |
*/ |
| 6489 |
if (wc->orderClause != NIL) |
| 6490 |
{ |
| 6491 |
List *orderby_pathkeys; |
| 6492 |
|
| 6493 |
orderby_pathkeys = make_pathkeys_for_sortclauses(root, |
| 6494 |
wc->orderClause, |
| 6495 |
tlist); |
| 6496 |
|
| 6497 |
/* Okay, make the combined pathkeys */ |
| 6498 |
if (window_pathkeys != NIL) |
| 6499 |
window_pathkeys = append_pathkeys(window_pathkeys, orderby_pathkeys); |
| 6500 |
else |
| 6501 |
window_pathkeys = orderby_pathkeys; |
| 6502 |
} |
| 6503 |
|
| 6504 |
return window_pathkeys; |
| 6505 |
} |
| 6506 |
|
| 6507 |
/* |
| 6508 |
* make_sort_input_target |
| 6509 |
* Generate appropriate PathTarget for initial input to Sort step. |
| 6510 |
* |
| 6511 |
* If the query has ORDER BY, this function chooses the target to be computed |
| 6512 |
* by the node just below the Sort (and DISTINCT, if any, since Unique can't |
| 6513 |
* project) steps. This might or might not be identical to the query's final |
| 6514 |
* output target. |
| 6515 |
* |
| 6516 |
* The main argument for keeping the sort-input tlist the same as the final |
| 6517 |
* is that we avoid a separate projection node (which will be needed if |
| 6518 |
* they're different, because Sort can't project). However, there are also |
| 6519 |
* advantages to postponing tlist evaluation till after the Sort: it ensures |
| 6520 |
* a consistent order of evaluation for any volatile functions in the tlist, |
| 6521 |
* and if there's also a LIMIT, we can stop the query without ever computing |
| 6522 |
* tlist functions for later rows, which is beneficial for both volatile and |
| 6523 |
* expensive functions. |
| 6524 |
* |
| 6525 |
* Our current policy is to postpone volatile expressions till after the sort |
| 6526 |
* unconditionally (assuming that that's possible, ie they are in plain tlist |
| 6527 |
* columns and not ORDER BY/GROUP BY/DISTINCT columns). We also prefer to |
| 6528 |
* postpone set-returning expressions, because running them beforehand would |
| 6529 |
* bloat the sort dataset, and because it might cause unexpected output order |
| 6530 |
* if the sort isn't stable. However there's a constraint on that: all SRFs |
| 6531 |
* in the tlist should be evaluated at the same plan step, so that they can |
| 6532 |
* run in sync in nodeProjectSet. So if any SRFs are in sort columns, we |
| 6533 |
* mustn't postpone any SRFs. (Note that in principle that policy should |
| 6534 |
* probably get applied to the group/window input targetlists too, but we |
| 6535 |
* have not done that historically.) Lastly, expensive expressions are |
| 6536 |
* postponed if there is a LIMIT, or if root->tuple_fraction shows that |
| 6537 |
* partial evaluation of the query is possible (if neither is true, we expect |
| 6538 |
* to have to evaluate the expressions for every row anyway), or if there are |
| 6539 |
* any volatile or set-returning expressions (since once we've put in a |
| 6540 |
* projection at all, it won't cost any more to postpone more stuff). |
| 6541 |
* |
| 6542 |
* Another issue that could potentially be considered here is that |
| 6543 |
* evaluating tlist expressions could result in data that's either wider |
| 6544 |
* or narrower than the input Vars, thus changing the volume of data that |
| 6545 |
* has to go through the Sort. However, we usually have only a very bad |
| 6546 |
* idea of the output width of any expression more complex than a Var, |
| 6547 |
* so for now it seems too risky to try to optimize on that basis. |
| 6548 |
* |
| 6549 |
* Note that if we do produce a modified sort-input target, and then the |
| 6550 |
* query ends up not using an explicit Sort, no particular harm is done: |
| 6551 |
* we'll initially use the modified target for the preceding path nodes, |
| 6552 |
* but then change them to the final target with apply_projection_to_path. |
| 6553 |
* Moreover, in such a case the guarantees about evaluation order of |
| 6554 |
* volatile functions still hold, since the rows are sorted already. |
| 6555 |
* |
| 6556 |
* This function has some things in common with make_group_input_target and |
| 6557 |
* make_window_input_target, though the detailed rules for what to do are |
| 6558 |
* different. We never flatten/postpone any grouping or ordering columns; |
| 6559 |
* those are needed before the sort. If we do flatten a particular |
| 6560 |
* expression, we leave Aggref and WindowFunc nodes alone, since those were |
| 6561 |
* computed earlier. |
| 6562 |
* |
| 6563 |
* 'final_target' is the query's final target list (in PathTarget form) |
| 6564 |
* 'have_postponed_srfs' is an output argument, see below |
| 6565 |
* |
| 6566 |
* The result is the PathTarget to be computed by the plan node immediately |
| 6567 |
* below the Sort step (and the Distinct step, if any). This will be |
| 6568 |
* exactly final_target if we decide a projection step wouldn't be helpful. |
| 6569 |
* |
| 6570 |
* In addition, *have_postponed_srfs is set to true if we choose to postpone |
| 6571 |
* any set-returning functions to after the Sort. |
| 6572 |
*/ |
| 6573 |
static PathTarget * |
| 6574 |
make_sort_input_target(PlannerInfo *root, |
| 6575 |
PathTarget *final_target, |
| 6576 |
bool *have_postponed_srfs) |
| 6577 |
{ |
| 6578 |
Query *parse = root->parse; |
| 6579 |
PathTarget *input_target; |
| 6580 |
int ncols; |
| 6581 |
bool *col_is_srf; |
| 6582 |
bool *postpone_col; |
| 6583 |
bool have_srf; |
| 6584 |
bool have_volatile; |
| 6585 |
bool have_expensive; |
| 6586 |
bool have_srf_sortcols; |
| 6587 |
bool postpone_srfs; |
| 6588 |
List *postponable_cols; |
| 6589 |
List *postponable_vars; |
| 6590 |
int i; |
| 6591 |
ListCell *lc; |
| 6592 |
|
| 6593 |
/* Shouldn't get here unless query has ORDER BY */ |
| 6594 |
Assert(parse->sortClause); |
| 6595 |
|
| 6596 |
*have_postponed_srfs = false; /* default result */ |
| 6597 |
|
| 6598 |
/* Inspect tlist and collect per-column information */ |
| 6599 |
ncols = list_length(final_target->exprs); |
| 6600 |
col_is_srf = (bool *) palloc0(ncols * sizeof(bool)); |
| 6601 |
postpone_col = (bool *) palloc0(ncols * sizeof(bool)); |
| 6602 |
have_srf = have_volatile = have_expensive = have_srf_sortcols = false; |
| 6603 |
|
| 6604 |
i = 0; |
| 6605 |
foreach(lc, final_target->exprs) |
| 6606 |
{ |
| 6607 |
Expr *expr = (Expr *) lfirst(lc); |
| 6608 |
|
| 6609 |
/* |
| 6610 |
* If the column has a sortgroupref, assume it has to be evaluated |
| 6611 |
* before sorting. Generally such columns would be ORDER BY, GROUP |
| 6612 |
* BY, etc targets. One exception is columns that were removed from |
| 6613 |
* GROUP BY by remove_useless_groupby_columns() ... but those would |
| 6614 |
* only be Vars anyway. There don't seem to be any cases where it |
| 6615 |
* would be worth the trouble to double-check. |
| 6616 |
*/ |
| 6617 |
if (get_pathtarget_sortgroupref(final_target, i) == 0) |
| 6618 |
{ |
| 6619 |
/* |
| 6620 |
* Check for SRF or volatile functions. Check the SRF case first |
| 6621 |
* because we must know whether we have any postponed SRFs. |
| 6622 |
*/ |
| 6623 |
if (parse->hasTargetSRFs && |
| 6624 |
expression_returns_set((Node *) expr)) |
| 6625 |
{ |
| 6626 |
/* We'll decide below whether these are postponable */ |
| 6627 |
col_is_srf[i] = true; |
| 6628 |
have_srf = true; |
| 6629 |
} |
| 6630 |
else if (contain_volatile_functions((Node *) expr)) |
| 6631 |
{ |
| 6632 |
/* Unconditionally postpone */ |
| 6633 |
postpone_col[i] = true; |
| 6634 |
have_volatile = true; |
| 6635 |
} |
| 6636 |
else |
| 6637 |
{ |
| 6638 |
/* |
| 6639 |
* Else check the cost. XXX it's annoying to have to do this |
| 6640 |
* when set_pathtarget_cost_width() just did it. Refactor to |
| 6641 |
* allow sharing the work? |
| 6642 |
*/ |
| 6643 |
QualCost cost; |
| 6644 |
|
| 6645 |
cost_qual_eval_node(&cost, (Node *) expr, root); |
| 6646 |
|
| 6647 |
/* |
| 6648 |
* We arbitrarily define "expensive" as "more than 10X |
| 6649 |
* cpu_operator_cost". Note this will take in any PL function |
| 6650 |
* with default cost. |
| 6651 |
*/ |
| 6652 |
if (cost.per_tuple > 10 * cpu_operator_cost) |
| 6653 |
{ |
| 6654 |
postpone_col[i] = true; |
| 6655 |
have_expensive = true; |
| 6656 |
} |
| 6657 |
} |
| 6658 |
} |
| 6659 |
else |
| 6660 |
{ |
| 6661 |
/* For sortgroupref cols, just check if any contain SRFs */ |
| 6662 |
if (!have_srf_sortcols && |
| 6663 |
parse->hasTargetSRFs && |
| 6664 |
expression_returns_set((Node *) expr)) |
| 6665 |
have_srf_sortcols = true; |
| 6666 |
} |
| 6667 |
|
| 6668 |
i++; |
| 6669 |
} |
| 6670 |
|
| 6671 |
/* |
| 6672 |
* We can postpone SRFs if we have some but none are in sortgroupref cols. |
| 6673 |
*/ |
| 6674 |
postpone_srfs = (have_srf && !have_srf_sortcols); |
| 6675 |
|
| 6676 |
/* |
| 6677 |
* If we don't need a post-sort projection, just return final_target. |
| 6678 |
*/ |
| 6679 |
if (!(postpone_srfs || have_volatile || |
| 6680 |
(have_expensive && |
| 6681 |
(parse->limitCount || root->tuple_fraction > 0)))) |
| 6682 |
return final_target; |
| 6683 |
|
| 6684 |
/* |
| 6685 |
* Report whether the post-sort projection will contain set-returning |
| 6686 |
* functions. This is important because it affects whether the Sort can |
| 6687 |
* rely on the query's LIMIT (if any) to bound the number of rows it needs |
| 6688 |
* to return. |
| 6689 |
*/ |
| 6690 |
*have_postponed_srfs = postpone_srfs; |
| 6691 |
|
| 6692 |
/* |
| 6693 |
* Construct the sort-input target, taking all non-postponable columns and |
| 6694 |
* then adding Vars, PlaceHolderVars, Aggrefs, and WindowFuncs found in |
| 6695 |
* the postponable ones. |
| 6696 |
*/ |
| 6697 |
input_target = create_empty_pathtarget(); |
| 6698 |
postponable_cols = NIL; |
| 6699 |
|
| 6700 |
i = 0; |
| 6701 |
foreach(lc, final_target->exprs) |
| 6702 |
{ |
| 6703 |
Expr *expr = (Expr *) lfirst(lc); |
| 6704 |
|
| 6705 |
if (postpone_col[i] || (postpone_srfs && col_is_srf[i])) |
| 6706 |
postponable_cols = lappend(postponable_cols, expr); |
| 6707 |
else |
| 6708 |
add_column_to_pathtarget(input_target, expr, |
| 6709 |
get_pathtarget_sortgroupref(final_target, i)); |
| 6710 |
|
| 6711 |
i++; |
| 6712 |
} |
| 6713 |
|
| 6714 |
/* |
| 6715 |
* Pull out all the Vars, Aggrefs, and WindowFuncs mentioned in |
| 6716 |
* postponable columns, and add them to the sort-input target if not |
| 6717 |
* already present. (Some might be there already.) We mustn't |
| 6718 |
* deconstruct Aggrefs or WindowFuncs here, since the projection node |
| 6719 |
* would be unable to recompute them. |
| 6720 |
*/ |
| 6721 |
postponable_vars = pull_var_clause((Node *) postponable_cols, |
| 6722 |
PVC_INCLUDE_AGGREGATES | |
| 6723 |
PVC_INCLUDE_WINDOWFUNCS | |
| 6724 |
PVC_INCLUDE_PLACEHOLDERS); |
| 6725 |
add_new_columns_to_pathtarget(input_target, postponable_vars); |
| 6726 |
|
| 6727 |
/* clean up cruft */ |
| 6728 |
list_free(postponable_vars); |
| 6729 |
list_free(postponable_cols); |
| 6730 |
|
| 6731 |
/* XXX this represents even more redundant cost calculation ... */ |
| 6732 |
return set_pathtarget_cost_width(root, input_target); |
| 6733 |
} |
| 6734 |
|
| 6735 |
/* |
| 6736 |
* get_cheapest_fractional_path |
| 6737 |
* Find the cheapest path for retrieving a specified fraction of all |
| 6738 |
* the tuples expected to be returned by the given relation. |
| 6739 |
* |
| 6740 |
* Do not consider parameterized paths. If the caller needs a path for upper |
| 6741 |
* rel, it can't have parameterized paths. If the caller needs an append |
| 6742 |
* subpath, it could become limited by the treatment of similar |
| 6743 |
* parameterization of all the subpaths. |
| 6744 |
* |
| 6745 |
* We interpret tuple_fraction the same way as grouping_planner. |
| 6746 |
* |
| 6747 |
* We assume set_cheapest() has been run on the given rel. |
| 6748 |
*/ |
| 6749 |
Path * |
| 6750 |
get_cheapest_fractional_path(RelOptInfo *rel, double tuple_fraction) |
| 6751 |
{ |
| 6752 |
Path *best_path = rel->cheapest_total_path; |
| 6753 |
ListCell *l; |
| 6754 |
|
| 6755 |
/* If all tuples will be retrieved, just return the cheapest-total path */ |
| 6756 |
if (tuple_fraction <= 0.0) |
| 6757 |
return best_path; |
| 6758 |
|
| 6759 |
/* Convert absolute # of tuples to a fraction; no need to clamp to 0..1 */ |
| 6760 |
if (tuple_fraction >= 1.0 && best_path->rows > 0) |
| 6761 |
tuple_fraction /= best_path->rows; |
| 6762 |
|
| 6763 |
foreach(l, rel->pathlist) |
| 6764 |
{ |
| 6765 |
Path *path = (Path *) lfirst(l); |
| 6766 |
|
| 6767 |
if (path->param_info) |
| 6768 |
continue; |
| 6769 |
|
| 6770 |
if (path == rel->cheapest_total_path || |
| 6771 |
compare_fractional_path_costs(best_path, path, tuple_fraction) <= 0) |
| 6772 |
continue; |
| 6773 |
|
| 6774 |
best_path = path; |
| 6775 |
} |
| 6776 |
|
| 6777 |
return best_path; |
| 6778 |
} |
| 6779 |
|
| 6780 |
/* |
| 6781 |
* adjust_paths_for_srfs |
| 6782 |
* Fix up the Paths of the given upperrel to handle tSRFs properly. |
| 6783 |
* |
| 6784 |
* The executor can only handle set-returning functions that appear at the |
| 6785 |
* top level of the targetlist of a ProjectSet plan node. If we have any SRFs |
| 6786 |
* that are not at top level, we need to split up the evaluation into multiple |
| 6787 |
* plan levels in which each level satisfies this constraint. This function |
| 6788 |
* modifies each Path of an upperrel that (might) compute any SRFs in its |
| 6789 |
* output tlist to insert appropriate projection steps. |
| 6790 |
* |
| 6791 |
* The given targets and targets_contain_srfs lists are from |
| 6792 |
* split_pathtarget_at_srfs(). We assume the existing Paths emit the first |
| 6793 |
* target in targets. |
| 6794 |
*/ |
| 6795 |
static void |
| 6796 |
adjust_paths_for_srfs(PlannerInfo *root, RelOptInfo *rel, |
| 6797 |
List *targets, List *targets_contain_srfs) |
| 6798 |
{ |
| 6799 |
ListCell *lc; |
| 6800 |
|
| 6801 |
Assert(list_length(targets) == list_length(targets_contain_srfs)); |
| 6802 |
Assert(!linitial_int(targets_contain_srfs)); |
| 6803 |
|
| 6804 |
/* If no SRFs appear at this plan level, nothing to do */ |
| 6805 |
if (list_length(targets) == 1) |
| 6806 |
return; |
| 6807 |
|
| 6808 |
/* |
| 6809 |
* Stack SRF-evaluation nodes atop each path for the rel. |
| 6810 |
* |
| 6811 |
* In principle we should re-run set_cheapest() here to identify the |
| 6812 |
* cheapest path, but it seems unlikely that adding the same tlist eval |
| 6813 |
* costs to all the paths would change that, so we don't bother. Instead, |
| 6814 |
* just assume that the cheapest-startup and cheapest-total paths remain |
| 6815 |
* so. (There should be no parameterized paths anymore, so we needn't |
| 6816 |
* worry about updating cheapest_parameterized_paths.) |
| 6817 |
*/ |
| 6818 |
foreach(lc, rel->pathlist) |
| 6819 |
{ |
| 6820 |
Path *subpath = (Path *) lfirst(lc); |
| 6821 |
Path *newpath = subpath; |
| 6822 |
ListCell *lc1, |
| 6823 |
*lc2; |
| 6824 |
|
| 6825 |
Assert(subpath->param_info == NULL); |
| 6826 |
forboth(lc1, targets, lc2, targets_contain_srfs) |
| 6827 |
{ |
| 6828 |
PathTarget *thistarget = lfirst_node(PathTarget, lc1); |
| 6829 |
bool contains_srfs = (bool) lfirst_int(lc2); |
| 6830 |
|
| 6831 |
/* If this level doesn't contain SRFs, do regular projection */ |
| 6832 |
if (contains_srfs) |
| 6833 |
newpath = (Path *) create_set_projection_path(root, |
| 6834 |
rel, |
| 6835 |
newpath, |
| 6836 |
thistarget); |
| 6837 |
else |
| 6838 |
newpath = (Path *) apply_projection_to_path(root, |
| 6839 |
rel, |
| 6840 |
newpath, |
| 6841 |
thistarget); |
| 6842 |
} |
| 6843 |
lfirst(lc) = newpath; |
| 6844 |
if (subpath == rel->cheapest_startup_path) |
| 6845 |
rel->cheapest_startup_path = newpath; |
| 6846 |
if (subpath == rel->cheapest_total_path) |
| 6847 |
rel->cheapest_total_path = newpath; |
| 6848 |
} |
| 6849 |
|
| 6850 |
/* Likewise for partial paths, if any */ |
| 6851 |
foreach(lc, rel->partial_pathlist) |
| 6852 |
{ |
| 6853 |
Path *subpath = (Path *) lfirst(lc); |
| 6854 |
Path *newpath = subpath; |
| 6855 |
ListCell *lc1, |
| 6856 |
*lc2; |
| 6857 |
|
| 6858 |
Assert(subpath->param_info == NULL); |
| 6859 |
forboth(lc1, targets, lc2, targets_contain_srfs) |
| 6860 |
{ |
| 6861 |
PathTarget *thistarget = lfirst_node(PathTarget, lc1); |
| 6862 |
bool contains_srfs = (bool) lfirst_int(lc2); |
| 6863 |
|
| 6864 |
/* If this level doesn't contain SRFs, do regular projection */ |
| 6865 |
if (contains_srfs) |
| 6866 |
newpath = (Path *) create_set_projection_path(root, |
| 6867 |
rel, |
| 6868 |
newpath, |
| 6869 |
thistarget); |
| 6870 |
else |
| 6871 |
{ |
| 6872 |
/* avoid apply_projection_to_path, in case of multiple refs */ |
| 6873 |
newpath = (Path *) create_projection_path(root, |
| 6874 |
rel, |
| 6875 |
newpath, |
| 6876 |
thistarget); |
| 6877 |
} |
| 6878 |
} |
| 6879 |
lfirst(lc) = newpath; |
| 6880 |
} |
| 6881 |
} |
| 6882 |
|
| 6883 |
/* |
| 6884 |
* expression_planner |
| 6885 |
* Perform planner's transformations on a standalone expression. |
| 6886 |
* |
| 6887 |
* Various utility commands need to evaluate expressions that are not part |
| 6888 |
* of a plannable query. They can do so using the executor's regular |
| 6889 |
* expression-execution machinery, but first the expression has to be fed |
| 6890 |
* through here to transform it from parser output to something executable. |
| 6891 |
* |
| 6892 |
* Currently, we disallow sublinks in standalone expressions, so there's no |
| 6893 |
* real "planning" involved here. (That might not always be true though.) |
| 6894 |
* What we must do is run eval_const_expressions to ensure that any function |
| 6895 |
* calls are converted to positional notation and function default arguments |
| 6896 |
* get inserted. The fact that constant subexpressions get simplified is a |
| 6897 |
* side-effect that is useful when the expression will get evaluated more than |
| 6898 |
* once. Also, we must fix operator function IDs. |
| 6899 |
* |
| 6900 |
* This does not return any information about dependencies of the expression. |
| 6901 |
* Hence callers should use the results only for the duration of the current |
| 6902 |
* query. Callers that would like to cache the results for longer should use |
| 6903 |
* expression_planner_with_deps, probably via the plancache. |
| 6904 |
* |
| 6905 |
* Note: this must not make any damaging changes to the passed-in expression |
| 6906 |
* tree. (It would actually be okay to apply fix_opfuncids to it, but since |
| 6907 |
* we first do an expression_tree_mutator-based walk, what is returned will |
| 6908 |
* be a new node tree.) The result is constructed in the current memory |
| 6909 |
* context; beware that this can leak a lot of additional stuff there, too. |
| 6910 |
*/ |
| 6911 |
Expr * |
| 6912 |
expression_planner(Expr *expr) |
| 6913 |
{ |
| 6914 |
Node *result; |
| 6915 |
|
| 6916 |
/* |
| 6917 |
* Convert named-argument function calls, insert default arguments and |
| 6918 |
* simplify constant subexprs |
| 6919 |
*/ |
| 6920 |
result = eval_const_expressions(NULL, (Node *) expr); |
| 6921 |
|
| 6922 |
/* Fill in opfuncid values if missing */ |
| 6923 |
fix_opfuncids(result); |
| 6924 |
|
| 6925 |
return (Expr *) result; |
| 6926 |
} |
| 6927 |
|
| 6928 |
/* |
| 6929 |
* expression_planner_with_deps |
| 6930 |
* Perform planner's transformations on a standalone expression, |
| 6931 |
* returning expression dependency information along with the result. |
| 6932 |
* |
| 6933 |
* This is identical to expression_planner() except that it also returns |
| 6934 |
* information about possible dependencies of the expression, ie identities of |
| 6935 |
* objects whose definitions affect the result. As in a PlannedStmt, these |
| 6936 |
* are expressed as a list of relation Oids and a list of PlanInvalItems. |
| 6937 |
*/ |
| 6938 |
Expr * |
| 6939 |
expression_planner_with_deps(Expr *expr, |
| 6940 |
List **relationOids, |
| 6941 |
List **invalItems) |
| 6942 |
{ |
| 6943 |
Node *result; |
| 6944 |
PlannerGlobal glob; |
| 6945 |
PlannerInfo root; |
| 6946 |
|
| 6947 |
/* Make up dummy planner state so we can use setrefs machinery */ |
| 6948 |
MemSet(&glob, 0, sizeof(glob)); |
| 6949 |
glob.type = T_PlannerGlobal; |
| 6950 |
glob.relationOids = NIL; |
| 6951 |
glob.invalItems = NIL; |
| 6952 |
|
| 6953 |
MemSet(&root, 0, sizeof(root)); |
| 6954 |
root.type = T_PlannerInfo; |
| 6955 |
root.glob = &glob; |
| 6956 |
|
| 6957 |
/* |
| 6958 |
* Convert named-argument function calls, insert default arguments and |
| 6959 |
* simplify constant subexprs. Collect identities of inlined functions |
| 6960 |
* and elided domains, too. |
| 6961 |
*/ |
| 6962 |
result = eval_const_expressions(&root, (Node *) expr); |
| 6963 |
|
| 6964 |
/* Fill in opfuncid values if missing */ |
| 6965 |
fix_opfuncids(result); |
| 6966 |
|
| 6967 |
/* |
| 6968 |
* Now walk the finished expression to find anything else we ought to |
| 6969 |
* record as an expression dependency. |
| 6970 |
*/ |
| 6971 |
(void) extract_query_dependencies_walker(result, &root); |
| 6972 |
|
| 6973 |
*relationOids = glob.relationOids; |
| 6974 |
*invalItems = glob.invalItems; |
| 6975 |
|
| 6976 |
return (Expr *) result; |
| 6977 |
} |
| 6978 |
|
| 6979 |
|
| 6980 |
/* |
| 6981 |
* plan_cluster_use_sort |
| 6982 |
* Use the planner to decide how CLUSTER should implement sorting |
| 6983 |
* |
| 6984 |
* tableOid is the OID of a table to be clustered on its index indexOid |
| 6985 |
* (which is already known to be a btree index). Decide whether it's |
| 6986 |
* cheaper to do an indexscan or a seqscan-plus-sort to execute the CLUSTER. |
| 6987 |
* Return true to use sorting, false to use an indexscan. |
| 6988 |
* |
| 6989 |
* Note: caller had better already hold some type of lock on the table. |
| 6990 |
*/ |
| 6991 |
bool |
| 6992 |
plan_cluster_use_sort(Oid tableOid, Oid indexOid) |
| 6993 |
{ |
| 6994 |
PlannerInfo *root; |
| 6995 |
Query *query; |
| 6996 |
PlannerGlobal *glob; |
| 6997 |
RangeTblEntry *rte; |
| 6998 |
RelOptInfo *rel; |
| 6999 |
IndexOptInfo *indexInfo; |
| 7000 |
QualCost indexExprCost; |
| 7001 |
Cost comparisonCost; |
| 7002 |
Path *seqScanPath; |
| 7003 |
Path seqScanAndSortPath; |
| 7004 |
IndexPath *indexScanPath; |
| 7005 |
ListCell *lc; |
| 7006 |
|
| 7007 |
/* We can short-circuit the cost comparison if indexscans are disabled */ |
| 7008 |
if (!enable_indexscan) |
| 7009 |
return true; /* use sort */ |
| 7010 |
|
| 7011 |
/* Set up mostly-dummy planner state */ |
| 7012 |
query = makeNode(Query); |
| 7013 |
query->commandType = CMD_SELECT; |
| 7014 |
|
| 7015 |
glob = makeNode(PlannerGlobal); |
| 7016 |
|
| 7017 |
root = makeNode(PlannerInfo); |
| 7018 |
root->parse = query; |
| 7019 |
root->glob = glob; |
| 7020 |
root->query_level = 1; |
| 7021 |
root->planner_cxt = CurrentMemoryContext; |
| 7022 |
root->wt_param_id = -1; |
| 7023 |
root->join_domains = list_make1(makeNode(JoinDomain)); |
| 7024 |
|
| 7025 |
/* Build a minimal RTE for the rel */ |
| 7026 |
rte = makeNode(RangeTblEntry); |
| 7027 |
rte->rtekind = RTE_RELATION; |
| 7028 |
rte->relid = tableOid; |
| 7029 |
rte->relkind = RELKIND_RELATION; /* Don't be too picky. */ |
| 7030 |
rte->rellockmode = AccessShareLock; |
| 7031 |
rte->lateral = false; |
| 7032 |
rte->inh = false; |
| 7033 |
rte->inFromCl = true; |
| 7034 |
query->rtable = list_make1(rte); |
| 7035 |
addRTEPermissionInfo(&query->rteperminfos, rte); |
| 7036 |
|
| 7037 |
/* Set up RTE/RelOptInfo arrays */ |
| 7038 |
setup_simple_rel_arrays(root); |
| 7039 |
|
| 7040 |
/* Build RelOptInfo */ |
| 7041 |
rel = build_simple_rel(root, 1, NULL); |
| 7042 |
|
| 7043 |
/* Locate IndexOptInfo for the target index */ |
| 7044 |
indexInfo = NULL; |
| 7045 |
foreach(lc, rel->indexlist) |
| 7046 |
{ |
| 7047 |
indexInfo = lfirst_node(IndexOptInfo, lc); |
| 7048 |
if (indexInfo->indexoid == indexOid) |
| 7049 |
break; |
| 7050 |
} |
| 7051 |
|
| 7052 |
/* |
| 7053 |
* It's possible that get_relation_info did not generate an IndexOptInfo |
| 7054 |
* for the desired index; this could happen if it's not yet reached its |
| 7055 |
* indcheckxmin usability horizon, or if it's a system index and we're |
| 7056 |
* ignoring system indexes. In such cases we should tell CLUSTER to not |
| 7057 |
* trust the index contents but use seqscan-and-sort. |
| 7058 |
*/ |
| 7059 |
if (lc == NULL) /* not in the list? */ |
| 7060 |
return true; /* use sort */ |
| 7061 |
|
| 7062 |
/* |
| 7063 |
* Rather than doing all the pushups that would be needed to use |
| 7064 |
* set_baserel_size_estimates, just do a quick hack for rows and width. |
| 7065 |
*/ |
| 7066 |
rel->rows = rel->tuples; |
| 7067 |
rel->reltarget->width = get_relation_data_width(tableOid, NULL); |
| 7068 |
|
| 7069 |
root->total_table_pages = rel->pages; |
| 7070 |
|
| 7071 |
/* |
| 7072 |
* Determine eval cost of the index expressions, if any. We need to |
| 7073 |
* charge twice that amount for each tuple comparison that happens during |
| 7074 |
* the sort, since tuplesort.c will have to re-evaluate the index |
| 7075 |
* expressions each time. (XXX that's pretty inefficient...) |
| 7076 |
*/ |
| 7077 |
cost_qual_eval(&indexExprCost, indexInfo->indexprs, root); |
| 7078 |
comparisonCost = 2.0 * (indexExprCost.startup + indexExprCost.per_tuple); |
| 7079 |
|
| 7080 |
/* Estimate the cost of seq scan + sort */ |
| 7081 |
seqScanPath = create_seqscan_path(root, rel, NULL, 0); |
| 7082 |
cost_sort(&seqScanAndSortPath, root, NIL, |
| 7083 |
seqScanPath->disabled_nodes, |
| 7084 |
seqScanPath->total_cost, rel->tuples, rel->reltarget->width, |
| 7085 |
comparisonCost, maintenance_work_mem, -1.0); |
| 7086 |
|
| 7087 |
/* Estimate the cost of index scan */ |
| 7088 |
indexScanPath = create_index_path(root, indexInfo, |
| 7089 |
NIL, NIL, NIL, NIL, |
| 7090 |
ForwardScanDirection, false, |
| 7091 |
NULL, 1.0, false); |
| 7092 |
|
| 7093 |
return (seqScanAndSortPath.total_cost < indexScanPath->path.total_cost); |
| 7094 |
} |
| 7095 |
|
| 7096 |
/* |
| 7097 |
* plan_create_index_workers |
| 7098 |
* Use the planner to decide how many parallel worker processes |
| 7099 |
* CREATE INDEX should request for use |
| 7100 |
* |
| 7101 |
* tableOid is the table on which the index is to be built. indexOid is the |
| 7102 |
* OID of an index to be created or reindexed (which must be an index with |
| 7103 |
* support for parallel builds - currently btree, GIN, or BRIN). |
| 7104 |
* |
| 7105 |
* Return value is the number of parallel worker processes to request. It |
| 7106 |
* may be unsafe to proceed if this is 0. Note that this does not include the |
| 7107 |
* leader participating as a worker (value is always a number of parallel |
| 7108 |
* worker processes). |
| 7109 |
* |
| 7110 |
* Note: caller had better already hold some type of lock on the table and |
| 7111 |
* index. |
| 7112 |
*/ |
| 7113 |
int |
| 7114 |
plan_create_index_workers(Oid tableOid, Oid indexOid) |
| 7115 |
{ |
| 7116 |
PlannerInfo *root; |
| 7117 |
Query *query; |
| 7118 |
PlannerGlobal *glob; |
| 7119 |
RangeTblEntry *rte; |
| 7120 |
Relation heap; |
| 7121 |
Relation index; |
| 7122 |
RelOptInfo *rel; |
| 7123 |
int parallel_workers; |
| 7124 |
BlockNumber heap_blocks; |
| 7125 |
double reltuples; |
| 7126 |
double allvisfrac; |
| 7127 |
|
| 7128 |
/* |
| 7129 |
* We don't allow performing parallel operation in standalone backend or |
| 7130 |
* when parallelism is disabled. |
| 7131 |
*/ |
| 7132 |
if (!IsUnderPostmaster || max_parallel_maintenance_workers == 0) |
| 7133 |
return 0; |
| 7134 |
|
| 7135 |
/* Set up largely-dummy planner state */ |
| 7136 |
query = makeNode(Query); |
| 7137 |
query->commandType = CMD_SELECT; |
| 7138 |
|
| 7139 |
glob = makeNode(PlannerGlobal); |
| 7140 |
|
| 7141 |
root = makeNode(PlannerInfo); |
| 7142 |
root->parse = query; |
| 7143 |
root->glob = glob; |
| 7144 |
root->query_level = 1; |
| 7145 |
root->planner_cxt = CurrentMemoryContext; |
| 7146 |
root->wt_param_id = -1; |
| 7147 |
root->join_domains = list_make1(makeNode(JoinDomain)); |
| 7148 |
|
| 7149 |
/* |
| 7150 |
* Build a minimal RTE. |
| 7151 |
* |
| 7152 |
* Mark the RTE with inh = true. This is a kludge to prevent |
| 7153 |
* get_relation_info() from fetching index info, which is necessary |
| 7154 |
* because it does not expect that any IndexOptInfo is currently |
| 7155 |
* undergoing REINDEX. |
| 7156 |
*/ |
| 7157 |
rte = makeNode(RangeTblEntry); |
| 7158 |
rte->rtekind = RTE_RELATION; |
| 7159 |
rte->relid = tableOid; |
| 7160 |
rte->relkind = RELKIND_RELATION; /* Don't be too picky. */ |
| 7161 |
rte->rellockmode = AccessShareLock; |
| 7162 |
rte->lateral = false; |
| 7163 |
rte->inh = true; |
| 7164 |
rte->inFromCl = true; |
| 7165 |
query->rtable = list_make1(rte); |
| 7166 |
addRTEPermissionInfo(&query->rteperminfos, rte); |
| 7167 |
|
| 7168 |
/* Set up RTE/RelOptInfo arrays */ |
| 7169 |
setup_simple_rel_arrays(root); |
| 7170 |
|
| 7171 |
/* Build RelOptInfo */ |
| 7172 |
rel = build_simple_rel(root, 1, NULL); |
| 7173 |
|
| 7174 |
/* Rels are assumed already locked by the caller */ |
| 7175 |
heap = table_open(tableOid, NoLock); |
| 7176 |
index = index_open(indexOid, NoLock); |
| 7177 |
|
| 7178 |
/* |
| 7179 |
* Determine if it's safe to proceed. |
| 7180 |
* |
| 7181 |
* Currently, parallel workers can't access the leader's temporary tables. |
| 7182 |
* Furthermore, any index predicate or index expressions must be parallel |
| 7183 |
* safe. |
| 7184 |
*/ |
| 7185 |
if (heap->rd_rel->relpersistence == RELPERSISTENCE_TEMP || |
| 7186 |
!is_parallel_safe(root, (Node *) RelationGetIndexExpressions(index)) || |
| 7187 |
!is_parallel_safe(root, (Node *) RelationGetIndexPredicate(index))) |
| 7188 |
{ |
| 7189 |
parallel_workers = 0; |
| 7190 |
goto done; |
| 7191 |
} |
| 7192 |
|
| 7193 |
/* |
| 7194 |
* If parallel_workers storage parameter is set for the table, accept that |
| 7195 |
* as the number of parallel worker processes to launch (though still cap |
| 7196 |
* at max_parallel_maintenance_workers). Note that we deliberately do not |
| 7197 |
* consider any other factor when parallel_workers is set. (e.g., memory |
| 7198 |
* use by workers.) |
| 7199 |
*/ |
| 7200 |
if (rel->rel_parallel_workers != -1) |
| 7201 |
{ |
| 7202 |
parallel_workers = Min(rel->rel_parallel_workers, |
| 7203 |
max_parallel_maintenance_workers); |
| 7204 |
goto done; |
| 7205 |
} |
| 7206 |
|
| 7207 |
/* |
| 7208 |
* Estimate heap relation size ourselves, since rel->pages cannot be |
| 7209 |
* trusted (heap RTE was marked as inheritance parent) |
| 7210 |
*/ |
| 7211 |
estimate_rel_size(heap, NULL, &heap_blocks, &reltuples, &allvisfrac); |
| 7212 |
|
| 7213 |
/* |
| 7214 |
* Determine number of workers to scan the heap relation using generic |
| 7215 |
* model |
| 7216 |
*/ |
| 7217 |
parallel_workers = compute_parallel_worker(rel, heap_blocks, -1, |
| 7218 |
max_parallel_maintenance_workers); |
| 7219 |
|
| 7220 |
/* |
| 7221 |
* Cap workers based on available maintenance_work_mem as needed. |
| 7222 |
* |
| 7223 |
* Note that each tuplesort participant receives an even share of the |
| 7224 |
* total maintenance_work_mem budget. Aim to leave participants |
| 7225 |
* (including the leader as a participant) with no less than 32MB of |
| 7226 |
* memory. This leaves cases where maintenance_work_mem is set to 64MB |
| 7227 |
* immediately past the threshold of being capable of launching a single |
| 7228 |
* parallel worker to sort. |
| 7229 |
*/ |
| 7230 |
while (parallel_workers > 0 && |
| 7231 |
maintenance_work_mem / (parallel_workers + 1) < 32 * 1024) |
| 7232 |
parallel_workers--; |
| 7233 |
|
| 7234 |
done: |
| 7235 |
index_close(index, NoLock); |
| 7236 |
table_close(heap, NoLock); |
| 7237 |
|
| 7238 |
return parallel_workers; |
| 7239 |
} |
| 7240 |
|
| 7241 |
/* |
| 7242 |
* add_paths_to_grouping_rel |
| 7243 |
* |
| 7244 |
* Add non-partial paths to grouping relation. |
| 7245 |
*/ |
| 7246 |
static void |
| 7247 |
add_paths_to_grouping_rel(PlannerInfo *root, RelOptInfo *input_rel, |
| 7248 |
RelOptInfo *grouped_rel, |
| 7249 |
RelOptInfo *partially_grouped_rel, |
| 7250 |
const AggClauseCosts *agg_costs, |
| 7251 |
grouping_sets_data *gd, double dNumGroups, |
| 7252 |
GroupPathExtraData *extra) |
| 7253 |
{ |
| 7254 |
Query *parse = root->parse; |
| 7255 |
Path *cheapest_path = input_rel->cheapest_total_path; |
| 7256 |
ListCell *lc; |
| 7257 |
bool can_hash = (extra->flags & GROUPING_CAN_USE_HASH) != 0; |
| 7258 |
bool can_sort = (extra->flags & GROUPING_CAN_USE_SORT) != 0; |
| 7259 |
List *havingQual = (List *) extra->havingQual; |
| 7260 |
AggClauseCosts *agg_final_costs = &extra->agg_final_costs; |
| 7261 |
|
| 7262 |
if (can_sort) |
| 7263 |
{ |
| 7264 |
/* |
| 7265 |
* Use any available suitably-sorted path as input, and also consider |
| 7266 |
* sorting the cheapest-total path and incremental sort on any paths |
| 7267 |
* with presorted keys. |
| 7268 |
*/ |
| 7269 |
foreach(lc, input_rel->pathlist) |
| 7270 |
{ |
| 7271 |
ListCell *lc2; |
| 7272 |
Path *path = (Path *) lfirst(lc); |
| 7273 |
Path *path_save = path; |
| 7274 |
List *pathkey_orderings = NIL; |
| 7275 |
|
| 7276 |
/* generate alternative group orderings that might be useful */ |
| 7277 |
pathkey_orderings = get_useful_group_keys_orderings(root, path); |
| 7278 |
|
| 7279 |
Assert(list_length(pathkey_orderings) > 0); |
| 7280 |
|
| 7281 |
foreach(lc2, pathkey_orderings) |
| 7282 |
{ |
| 7283 |
GroupByOrdering *info = (GroupByOrdering *) lfirst(lc2); |
| 7284 |
|
| 7285 |
/* restore the path (we replace it in the loop) */ |
| 7286 |
path = path_save; |
| 7287 |
|
| 7288 |
path = make_ordered_path(root, |
| 7289 |
grouped_rel, |
| 7290 |
path, |
| 7291 |
cheapest_path, |
| 7292 |
info->pathkeys, |
| 7293 |
-1.0); |
| 7294 |
if (path == NULL) |
| 7295 |
continue; |
| 7296 |
|
| 7297 |
/* Now decide what to stick atop it */ |
| 7298 |
if (parse->groupingSets) |
| 7299 |
{ |
| 7300 |
consider_groupingsets_paths(root, grouped_rel, |
| 7301 |
path, true, can_hash, |
| 7302 |
gd, agg_costs, dNumGroups); |
| 7303 |
} |
| 7304 |
else if (parse->hasAggs) |
| 7305 |
{ |
| 7306 |
/* |
| 7307 |
* We have aggregation, possibly with plain GROUP BY. Make |
| 7308 |
* an AggPath. |
| 7309 |
*/ |
| 7310 |
add_path(grouped_rel, (Path *) |
| 7311 |
create_agg_path(root, |
| 7312 |
grouped_rel, |
| 7313 |
path, |
| 7314 |
grouped_rel->reltarget, |
| 7315 |
parse->groupClause ? AGG_SORTED : AGG_PLAIN, |
| 7316 |
AGGSPLIT_SIMPLE, |
| 7317 |
info->clauses, |
| 7318 |
havingQual, |
| 7319 |
agg_costs, |
| 7320 |
dNumGroups)); |
| 7321 |
} |
| 7322 |
else if (parse->groupClause) |
| 7323 |
{ |
| 7324 |
/* |
| 7325 |
* We have GROUP BY without aggregation or grouping sets. |
| 7326 |
* Make a GroupPath. |
| 7327 |
*/ |
| 7328 |
add_path(grouped_rel, (Path *) |
| 7329 |
create_group_path(root, |
| 7330 |
grouped_rel, |
| 7331 |
path, |
| 7332 |
info->clauses, |
| 7333 |
havingQual, |
| 7334 |
dNumGroups)); |
| 7335 |
} |
| 7336 |
else |
| 7337 |
{ |
| 7338 |
/* Other cases should have been handled above */ |
| 7339 |
Assert(false); |
| 7340 |
} |
| 7341 |
} |
| 7342 |
} |
| 7343 |
|
| 7344 |
/* |
| 7345 |
* Instead of operating directly on the input relation, we can |
| 7346 |
* consider finalizing a partially aggregated path. |
| 7347 |
*/ |
| 7348 |
if (partially_grouped_rel != NULL) |
| 7349 |
{ |
| 7350 |
foreach(lc, partially_grouped_rel->pathlist) |
| 7351 |
{ |
| 7352 |
ListCell *lc2; |
| 7353 |
Path *path = (Path *) lfirst(lc); |
| 7354 |
Path *path_save = path; |
| 7355 |
List *pathkey_orderings = NIL; |
| 7356 |
|
| 7357 |
/* generate alternative group orderings that might be useful */ |
| 7358 |
pathkey_orderings = get_useful_group_keys_orderings(root, path); |
| 7359 |
|
| 7360 |
Assert(list_length(pathkey_orderings) > 0); |
| 7361 |
|
| 7362 |
/* process all potentially interesting grouping reorderings */ |
| 7363 |
foreach(lc2, pathkey_orderings) |
| 7364 |
{ |
| 7365 |
GroupByOrdering *info = (GroupByOrdering *) lfirst(lc2); |
| 7366 |
|
| 7367 |
/* restore the path (we replace it in the loop) */ |
| 7368 |
path = path_save; |
| 7369 |
|
| 7370 |
path = make_ordered_path(root, |
| 7371 |
grouped_rel, |
| 7372 |
path, |
| 7373 |
partially_grouped_rel->cheapest_total_path, |
| 7374 |
info->pathkeys, |
| 7375 |
-1.0); |
| 7376 |
|
| 7377 |
if (path == NULL) |
| 7378 |
continue; |
| 7379 |
|
| 7380 |
if (parse->hasAggs) |
| 7381 |
add_path(grouped_rel, (Path *) |
| 7382 |
create_agg_path(root, |
| 7383 |
grouped_rel, |
| 7384 |
path, |
| 7385 |
grouped_rel->reltarget, |
| 7386 |
parse->groupClause ? AGG_SORTED : AGG_PLAIN, |
| 7387 |
AGGSPLIT_FINAL_DESERIAL, |
| 7388 |
info->clauses, |
| 7389 |
havingQual, |
| 7390 |
agg_final_costs, |
| 7391 |
dNumGroups)); |
| 7392 |
else |
| 7393 |
add_path(grouped_rel, (Path *) |
| 7394 |
create_group_path(root, |
| 7395 |
grouped_rel, |
| 7396 |
path, |
| 7397 |
info->clauses, |
| 7398 |
havingQual, |
| 7399 |
dNumGroups)); |
| 7400 |
|
| 7401 |
} |
| 7402 |
} |
| 7403 |
} |
| 7404 |
} |
| 7405 |
|
| 7406 |
if (can_hash) |
| 7407 |
{ |
| 7408 |
if (parse->groupingSets) |
| 7409 |
{ |
| 7410 |
/* |
| 7411 |
* Try for a hash-only groupingsets path over unsorted input. |
| 7412 |
*/ |
| 7413 |
consider_groupingsets_paths(root, grouped_rel, |
| 7414 |
cheapest_path, false, true, |
| 7415 |
gd, agg_costs, dNumGroups); |
| 7416 |
} |
| 7417 |
else |
| 7418 |
{ |
| 7419 |
/* |
| 7420 |
* Generate a HashAgg Path. We just need an Agg over the |
| 7421 |
* cheapest-total input path, since input order won't matter. |
| 7422 |
*/ |
| 7423 |
add_path(grouped_rel, (Path *) |
| 7424 |
create_agg_path(root, grouped_rel, |
| 7425 |
cheapest_path, |
| 7426 |
grouped_rel->reltarget, |
| 7427 |
AGG_HASHED, |
| 7428 |
AGGSPLIT_SIMPLE, |
| 7429 |
root->processed_groupClause, |
| 7430 |
havingQual, |
| 7431 |
agg_costs, |
| 7432 |
dNumGroups)); |
| 7433 |
} |
| 7434 |
|
| 7435 |
/* |
| 7436 |
* Generate a Finalize HashAgg Path atop of the cheapest partially |
| 7437 |
* grouped path, assuming there is one |
| 7438 |
*/ |
| 7439 |
if (partially_grouped_rel && partially_grouped_rel->pathlist) |
| 7440 |
{ |
| 7441 |
Path *path = partially_grouped_rel->cheapest_total_path; |
| 7442 |
|
| 7443 |
add_path(grouped_rel, (Path *) |
| 7444 |
create_agg_path(root, |
| 7445 |
grouped_rel, |
| 7446 |
path, |
| 7447 |
grouped_rel->reltarget, |
| 7448 |
AGG_HASHED, |
| 7449 |
AGGSPLIT_FINAL_DESERIAL, |
| 7450 |
root->processed_groupClause, |
| 7451 |
havingQual, |
| 7452 |
agg_final_costs, |
| 7453 |
dNumGroups)); |
| 7454 |
} |
| 7455 |
} |
| 7456 |
|
| 7457 |
/* |
| 7458 |
* When partitionwise aggregate is used, we might have fully aggregated |
| 7459 |
* paths in the partial pathlist, because add_paths_to_append_rel() will |
| 7460 |
* consider a path for grouped_rel consisting of a Parallel Append of |
| 7461 |
* non-partial paths from each child. |
| 7462 |
*/ |
| 7463 |
if (grouped_rel->partial_pathlist != NIL) |
| 7464 |
gather_grouping_paths(root, grouped_rel); |
| 7465 |
} |
| 7466 |
|
| 7467 |
/* |
| 7468 |
* create_partial_grouping_paths |
| 7469 |
* |
| 7470 |
* Create a new upper relation representing the result of partial aggregation |
| 7471 |
* and populate it with appropriate paths. Note that we don't finalize the |
| 7472 |
* lists of paths here, so the caller can add additional partial or non-partial |
| 7473 |
* paths and must afterward call gather_grouping_paths and set_cheapest on |
| 7474 |
* the returned upper relation. |
| 7475 |
* |
| 7476 |
* All paths for this new upper relation -- both partial and non-partial -- |
| 7477 |
* have been partially aggregated but require a subsequent FinalizeAggregate |
| 7478 |
* step. |
| 7479 |
* |
| 7480 |
* NB: This function is allowed to return NULL if it determines that there is |
| 7481 |
* no real need to create a new RelOptInfo. |
| 7482 |
*/ |
| 7483 |
static RelOptInfo * |
| 7484 |
create_partial_grouping_paths(PlannerInfo *root, |
| 7485 |
RelOptInfo *grouped_rel, |
| 7486 |
RelOptInfo *input_rel, |
| 7487 |
grouping_sets_data *gd, |
| 7488 |
GroupPathExtraData *extra, |
| 7489 |
bool force_rel_creation) |
| 7490 |
{ |
| 7491 |
Query *parse = root->parse; |
| 7492 |
RelOptInfo *partially_grouped_rel; |
| 7493 |
AggClauseCosts *agg_partial_costs = &extra->agg_partial_costs; |
| 7494 |
AggClauseCosts *agg_final_costs = &extra->agg_final_costs; |
| 7495 |
Path *cheapest_partial_path = NULL; |
| 7496 |
Path *cheapest_total_path = NULL; |
| 7497 |
double dNumPartialGroups = 0; |
| 7498 |
double dNumPartialPartialGroups = 0; |
| 7499 |
ListCell *lc; |
| 7500 |
bool can_hash = (extra->flags & GROUPING_CAN_USE_HASH) != 0; |
| 7501 |
bool can_sort = (extra->flags & GROUPING_CAN_USE_SORT) != 0; |
| 7502 |
|
| 7503 |
/* |
| 7504 |
* Consider whether we should generate partially aggregated non-partial |
| 7505 |
* paths. We can only do this if we have a non-partial path, and only if |
| 7506 |
* the parent of the input rel is performing partial partitionwise |
| 7507 |
* aggregation. (Note that extra->patype is the type of partitionwise |
| 7508 |
* aggregation being used at the parent level, not this level.) |
| 7509 |
*/ |
| 7510 |
if (input_rel->pathlist != NIL && |
| 7511 |
extra->patype == PARTITIONWISE_AGGREGATE_PARTIAL) |
| 7512 |
cheapest_total_path = input_rel->cheapest_total_path; |
| 7513 |
|
| 7514 |
/* |
| 7515 |
* If parallelism is possible for grouped_rel, then we should consider |
| 7516 |
* generating partially-grouped partial paths. However, if the input rel |
| 7517 |
* has no partial paths, then we can't. |
| 7518 |
*/ |
| 7519 |
if (grouped_rel->consider_parallel && input_rel->partial_pathlist != NIL) |
| 7520 |
cheapest_partial_path = linitial(input_rel->partial_pathlist); |
| 7521 |
|
| 7522 |
/* |
| 7523 |
* If we can't partially aggregate partial paths, and we can't partially |
| 7524 |
* aggregate non-partial paths, then don't bother creating the new |
| 7525 |
* RelOptInfo at all, unless the caller specified force_rel_creation. |
| 7526 |
*/ |
| 7527 |
if (cheapest_total_path == NULL && |
| 7528 |
cheapest_partial_path == NULL && |
| 7529 |
!force_rel_creation) |
| 7530 |
return NULL; |
| 7531 |
|
| 7532 |
/* |
| 7533 |
* Build a new upper relation to represent the result of partially |
| 7534 |
* aggregating the rows from the input relation. |
| 7535 |
*/ |
| 7536 |
partially_grouped_rel = fetch_upper_rel(root, |
| 7537 |
UPPERREL_PARTIAL_GROUP_AGG, |
| 7538 |
grouped_rel->relids); |
| 7539 |
partially_grouped_rel->consider_parallel = |
| 7540 |
grouped_rel->consider_parallel; |
| 7541 |
partially_grouped_rel->reloptkind = grouped_rel->reloptkind; |
| 7542 |
partially_grouped_rel->serverid = grouped_rel->serverid; |
| 7543 |
partially_grouped_rel->userid = grouped_rel->userid; |
| 7544 |
partially_grouped_rel->useridiscurrent = grouped_rel->useridiscurrent; |
| 7545 |
partially_grouped_rel->fdwroutine = grouped_rel->fdwroutine; |
| 7546 |
|
| 7547 |
/* |
| 7548 |
* Build target list for partial aggregate paths. These paths cannot just |
| 7549 |
* emit the same tlist as regular aggregate paths, because (1) we must |
| 7550 |
* include Vars and Aggrefs needed in HAVING, which might not appear in |
| 7551 |
* the result tlist, and (2) the Aggrefs must be set in partial mode. |
| 7552 |
*/ |
| 7553 |
partially_grouped_rel->reltarget = |
| 7554 |
make_partial_grouping_target(root, grouped_rel->reltarget, |
| 7555 |
extra->havingQual); |
| 7556 |
|
| 7557 |
if (!extra->partial_costs_set) |
| 7558 |
{ |
| 7559 |
/* |
| 7560 |
* Collect statistics about aggregates for estimating costs of |
| 7561 |
* performing aggregation in parallel. |
| 7562 |
*/ |
| 7563 |
MemSet(agg_partial_costs, 0, sizeof(AggClauseCosts)); |
| 7564 |
MemSet(agg_final_costs, 0, sizeof(AggClauseCosts)); |
| 7565 |
if (parse->hasAggs) |
| 7566 |
{ |
| 7567 |
/* partial phase */ |
| 7568 |
get_agg_clause_costs(root, AGGSPLIT_INITIAL_SERIAL, |
| 7569 |
agg_partial_costs); |
| 7570 |
|
| 7571 |
/* final phase */ |
| 7572 |
get_agg_clause_costs(root, AGGSPLIT_FINAL_DESERIAL, |
| 7573 |
agg_final_costs); |
| 7574 |
} |
| 7575 |
|
| 7576 |
extra->partial_costs_set = true; |
| 7577 |
} |
| 7578 |
|
| 7579 |
/* Estimate number of partial groups. */ |
| 7580 |
if (cheapest_total_path != NULL) |
| 7581 |
dNumPartialGroups = |
| 7582 |
get_number_of_groups(root, |
| 7583 |
cheapest_total_path->rows, |
| 7584 |
gd, |
| 7585 |
extra->targetList); |
| 7586 |
if (cheapest_partial_path != NULL) |
| 7587 |
dNumPartialPartialGroups = |
| 7588 |
get_number_of_groups(root, |
| 7589 |
cheapest_partial_path->rows, |
| 7590 |
gd, |
| 7591 |
extra->targetList); |
| 7592 |
|
| 7593 |
if (can_sort && cheapest_total_path != NULL) |
| 7594 |
{ |
| 7595 |
/* This should have been checked previously */ |
| 7596 |
Assert(parse->hasAggs || parse->groupClause); |
| 7597 |
|
| 7598 |
/* |
| 7599 |
* Use any available suitably-sorted path as input, and also consider |
| 7600 |
* sorting the cheapest partial path. |
| 7601 |
*/ |
| 7602 |
foreach(lc, input_rel->pathlist) |
| 7603 |
{ |
| 7604 |
ListCell *lc2; |
| 7605 |
Path *path = (Path *) lfirst(lc); |
| 7606 |
Path *path_save = path; |
| 7607 |
List *pathkey_orderings = NIL; |
| 7608 |
|
| 7609 |
/* generate alternative group orderings that might be useful */ |
| 7610 |
pathkey_orderings = get_useful_group_keys_orderings(root, path); |
| 7611 |
|
| 7612 |
Assert(list_length(pathkey_orderings) > 0); |
| 7613 |
|
| 7614 |
/* process all potentially interesting grouping reorderings */ |
| 7615 |
foreach(lc2, pathkey_orderings) |
| 7616 |
{ |
| 7617 |
GroupByOrdering *info = (GroupByOrdering *) lfirst(lc2); |
| 7618 |
|
| 7619 |
/* restore the path (we replace it in the loop) */ |
| 7620 |
path = path_save; |
| 7621 |
|
| 7622 |
path = make_ordered_path(root, |
| 7623 |
partially_grouped_rel, |
| 7624 |
path, |
| 7625 |
cheapest_total_path, |
| 7626 |
info->pathkeys, |
| 7627 |
-1.0); |
| 7628 |
|
| 7629 |
if (path == NULL) |
| 7630 |
continue; |
| 7631 |
|
| 7632 |
if (parse->hasAggs) |
| 7633 |
add_path(partially_grouped_rel, (Path *) |
| 7634 |
create_agg_path(root, |
| 7635 |
partially_grouped_rel, |
| 7636 |
path, |
| 7637 |
partially_grouped_rel->reltarget, |
| 7638 |
parse->groupClause ? AGG_SORTED : AGG_PLAIN, |
| 7639 |
AGGSPLIT_INITIAL_SERIAL, |
| 7640 |
info->clauses, |
| 7641 |
NIL, |
| 7642 |
agg_partial_costs, |
| 7643 |
dNumPartialGroups)); |
| 7644 |
else |
| 7645 |
add_path(partially_grouped_rel, (Path *) |
| 7646 |
create_group_path(root, |
| 7647 |
partially_grouped_rel, |
| 7648 |
path, |
| 7649 |
info->clauses, |
| 7650 |
NIL, |
| 7651 |
dNumPartialGroups)); |
| 7652 |
} |
| 7653 |
} |
| 7654 |
} |
| 7655 |
|
| 7656 |
if (can_sort && cheapest_partial_path != NULL) |
| 7657 |
{ |
| 7658 |
/* Similar to above logic, but for partial paths. */ |
| 7659 |
foreach(lc, input_rel->partial_pathlist) |
| 7660 |
{ |
| 7661 |
ListCell *lc2; |
| 7662 |
Path *path = (Path *) lfirst(lc); |
| 7663 |
Path *path_save = path; |
| 7664 |
List *pathkey_orderings = NIL; |
| 7665 |
|
| 7666 |
/* generate alternative group orderings that might be useful */ |
| 7667 |
pathkey_orderings = get_useful_group_keys_orderings(root, path); |
| 7668 |
|
| 7669 |
Assert(list_length(pathkey_orderings) > 0); |
| 7670 |
|
| 7671 |
/* process all potentially interesting grouping reorderings */ |
| 7672 |
foreach(lc2, pathkey_orderings) |
| 7673 |
{ |
| 7674 |
GroupByOrdering *info = (GroupByOrdering *) lfirst(lc2); |
| 7675 |
|
| 7676 |
|
| 7677 |
/* restore the path (we replace it in the loop) */ |
| 7678 |
path = path_save; |
| 7679 |
|
| 7680 |
path = make_ordered_path(root, |
| 7681 |
partially_grouped_rel, |
| 7682 |
path, |
| 7683 |
cheapest_partial_path, |
| 7684 |
info->pathkeys, |
| 7685 |
-1.0); |
| 7686 |
|
| 7687 |
if (path == NULL) |
| 7688 |
continue; |
| 7689 |
|
| 7690 |
if (parse->hasAggs) |
| 7691 |
add_partial_path(partially_grouped_rel, (Path *) |
| 7692 |
create_agg_path(root, |
| 7693 |
partially_grouped_rel, |
| 7694 |
path, |
| 7695 |
partially_grouped_rel->reltarget, |
| 7696 |
parse->groupClause ? AGG_SORTED : AGG_PLAIN, |
| 7697 |
AGGSPLIT_INITIAL_SERIAL, |
| 7698 |
info->clauses, |
| 7699 |
NIL, |
| 7700 |
agg_partial_costs, |
| 7701 |
dNumPartialPartialGroups)); |
| 7702 |
else |
| 7703 |
add_partial_path(partially_grouped_rel, (Path *) |
| 7704 |
create_group_path(root, |
| 7705 |
partially_grouped_rel, |
| 7706 |
path, |
| 7707 |
info->clauses, |
| 7708 |
NIL, |
| 7709 |
dNumPartialPartialGroups)); |
| 7710 |
} |
| 7711 |
} |
| 7712 |
} |
| 7713 |
|
| 7714 |
/* |
| 7715 |
* Add a partially-grouped HashAgg Path where possible |
| 7716 |
*/ |
| 7717 |
if (can_hash && cheapest_total_path != NULL) |
| 7718 |
{ |
| 7719 |
/* Checked above */ |
| 7720 |
Assert(parse->hasAggs || parse->groupClause); |
| 7721 |
|
| 7722 |
add_path(partially_grouped_rel, (Path *) |
| 7723 |
create_agg_path(root, |
| 7724 |
partially_grouped_rel, |
| 7725 |
cheapest_total_path, |
| 7726 |
partially_grouped_rel->reltarget, |
| 7727 |
AGG_HASHED, |
| 7728 |
AGGSPLIT_INITIAL_SERIAL, |
| 7729 |
root->processed_groupClause, |
| 7730 |
NIL, |
| 7731 |
agg_partial_costs, |
| 7732 |
dNumPartialGroups)); |
| 7733 |
} |
| 7734 |
|
| 7735 |
/* |
| 7736 |
* Now add a partially-grouped HashAgg partial Path where possible |
| 7737 |
*/ |
| 7738 |
if (can_hash && cheapest_partial_path != NULL) |
| 7739 |
{ |
| 7740 |
add_partial_path(partially_grouped_rel, (Path *) |
| 7741 |
create_agg_path(root, |
| 7742 |
partially_grouped_rel, |
| 7743 |
cheapest_partial_path, |
| 7744 |
partially_grouped_rel->reltarget, |
| 7745 |
AGG_HASHED, |
| 7746 |
AGGSPLIT_INITIAL_SERIAL, |
| 7747 |
root->processed_groupClause, |
| 7748 |
NIL, |
| 7749 |
agg_partial_costs, |
| 7750 |
dNumPartialPartialGroups)); |
| 7751 |
} |
| 7752 |
|
| 7753 |
/* |
| 7754 |
* If there is an FDW that's responsible for all baserels of the query, |
| 7755 |
* let it consider adding partially grouped ForeignPaths. |
| 7756 |
*/ |
| 7757 |
if (partially_grouped_rel->fdwroutine && |
| 7758 |
partially_grouped_rel->fdwroutine->GetForeignUpperPaths) |
| 7759 |
{ |
| 7760 |
FdwRoutine *fdwroutine = partially_grouped_rel->fdwroutine; |
| 7761 |
|
| 7762 |
fdwroutine->GetForeignUpperPaths(root, |
| 7763 |
UPPERREL_PARTIAL_GROUP_AGG, |
| 7764 |
input_rel, partially_grouped_rel, |
| 7765 |
extra); |
| 7766 |
} |
| 7767 |
|
| 7768 |
return partially_grouped_rel; |
| 7769 |
} |
| 7770 |
|
| 7771 |
/* |
| 7772 |
* make_ordered_path |
| 7773 |
* Return a path ordered by 'pathkeys' based on the given 'path'. May |
| 7774 |
* return NULL if it doesn't make sense to generate an ordered path in |
| 7775 |
* this case. |
| 7776 |
*/ |
| 7777 |
static Path * |
| 7778 |
make_ordered_path(PlannerInfo *root, RelOptInfo *rel, Path *path, |
| 7779 |
Path *cheapest_path, List *pathkeys, double limit_tuples) |
| 7780 |
{ |
| 7781 |
bool is_sorted; |
| 7782 |
int presorted_keys; |
| 7783 |
|
| 7784 |
is_sorted = pathkeys_count_contained_in(pathkeys, |
| 7785 |
path->pathkeys, |
| 7786 |
&presorted_keys); |
| 7787 |
|
| 7788 |
if (!is_sorted) |
| 7789 |
{ |
| 7790 |
/* |
| 7791 |
* Try at least sorting the cheapest path and also try incrementally |
| 7792 |
* sorting any path which is partially sorted already (no need to deal |
| 7793 |
* with paths which have presorted keys when incremental sort is |
| 7794 |
* disabled unless it's the cheapest input path). |
| 7795 |
*/ |
| 7796 |
if (path != cheapest_path && |
| 7797 |
(presorted_keys == 0 || !enable_incremental_sort)) |
| 7798 |
return NULL; |
| 7799 |
|
| 7800 |
/* |
| 7801 |
* We've no need to consider both a sort and incremental sort. We'll |
| 7802 |
* just do a sort if there are no presorted keys and an incremental |
| 7803 |
* sort when there are presorted keys. |
| 7804 |
*/ |
| 7805 |
if (presorted_keys == 0 || !enable_incremental_sort) |
| 7806 |
path = (Path *) create_sort_path(root, |
| 7807 |
rel, |
| 7808 |
path, |
| 7809 |
pathkeys, |
| 7810 |
limit_tuples); |
| 7811 |
else |
| 7812 |
path = (Path *) create_incremental_sort_path(root, |
| 7813 |
rel, |
| 7814 |
path, |
| 7815 |
pathkeys, |
| 7816 |
presorted_keys, |
| 7817 |
limit_tuples); |
| 7818 |
} |
| 7819 |
|
| 7820 |
return path; |
| 7821 |
} |
| 7822 |
|
| 7823 |
/* |
| 7824 |
* Generate Gather and Gather Merge paths for a grouping relation or partial |
| 7825 |
* grouping relation. |
| 7826 |
* |
| 7827 |
* generate_useful_gather_paths does most of the work, but we also consider a |
| 7828 |
* special case: we could try sorting the data by the group_pathkeys and then |
| 7829 |
* applying Gather Merge. |
| 7830 |
* |
| 7831 |
* NB: This function shouldn't be used for anything other than a grouped or |
| 7832 |
* partially grouped relation not only because of the fact that it explicitly |
| 7833 |
* references group_pathkeys but we pass "true" as the third argument to |
| 7834 |
* generate_useful_gather_paths(). |
| 7835 |
*/ |
| 7836 |
static void |
| 7837 |
gather_grouping_paths(PlannerInfo *root, RelOptInfo *rel) |
| 7838 |
{ |
| 7839 |
ListCell *lc; |
| 7840 |
Path *cheapest_partial_path; |
| 7841 |
List *groupby_pathkeys; |
| 7842 |
|
| 7843 |
/* |
| 7844 |
* This occurs after any partial aggregation has taken place, so trim off |
| 7845 |
* any pathkeys added for ORDER BY / DISTINCT aggregates. |
| 7846 |
*/ |
| 7847 |
if (list_length(root->group_pathkeys) > root->num_groupby_pathkeys) |
| 7848 |
groupby_pathkeys = list_copy_head(root->group_pathkeys, |
| 7849 |
root->num_groupby_pathkeys); |
| 7850 |
else |
| 7851 |
groupby_pathkeys = root->group_pathkeys; |
| 7852 |
|
| 7853 |
/* Try Gather for unordered paths and Gather Merge for ordered ones. */ |
| 7854 |
generate_useful_gather_paths(root, rel, true); |
| 7855 |
|
| 7856 |
cheapest_partial_path = linitial(rel->partial_pathlist); |
| 7857 |
|
| 7858 |
/* XXX Shouldn't this also consider the group-key-reordering? */ |
| 7859 |
foreach(lc, rel->partial_pathlist) |
| 7860 |
{ |
| 7861 |
Path *path = (Path *) lfirst(lc); |
| 7862 |
bool is_sorted; |
| 7863 |
int presorted_keys; |
| 7864 |
double total_groups; |
| 7865 |
|
| 7866 |
is_sorted = pathkeys_count_contained_in(groupby_pathkeys, |
| 7867 |
path->pathkeys, |
| 7868 |
&presorted_keys); |
| 7869 |
|
| 7870 |
if (is_sorted) |
| 7871 |
continue; |
| 7872 |
|
| 7873 |
/* |
| 7874 |
* Try at least sorting the cheapest path and also try incrementally |
| 7875 |
* sorting any path which is partially sorted already (no need to deal |
| 7876 |
* with paths which have presorted keys when incremental sort is |
| 7877 |
* disabled unless it's the cheapest input path). |
| 7878 |
*/ |
| 7879 |
if (path != cheapest_partial_path && |
| 7880 |
(presorted_keys == 0 || !enable_incremental_sort)) |
| 7881 |
continue; |
| 7882 |
|
| 7883 |
/* |
| 7884 |
* We've no need to consider both a sort and incremental sort. We'll |
| 7885 |
* just do a sort if there are no presorted keys and an incremental |
| 7886 |
* sort when there are presorted keys. |
| 7887 |
*/ |
| 7888 |
if (presorted_keys == 0 || !enable_incremental_sort) |
| 7889 |
path = (Path *) create_sort_path(root, rel, path, |
| 7890 |
groupby_pathkeys, |
| 7891 |
-1.0); |
| 7892 |
else |
| 7893 |
path = (Path *) create_incremental_sort_path(root, |
| 7894 |
rel, |
| 7895 |
path, |
| 7896 |
groupby_pathkeys, |
| 7897 |
presorted_keys, |
| 7898 |
-1.0); |
| 7899 |
total_groups = compute_gather_rows(path); |
| 7900 |
path = (Path *) |
| 7901 |
create_gather_merge_path(root, |
| 7902 |
rel, |
| 7903 |
path, |
| 7904 |
rel->reltarget, |
| 7905 |
groupby_pathkeys, |
| 7906 |
NULL, |
| 7907 |
&total_groups); |
| 7908 |
|
| 7909 |
add_path(rel, path); |
| 7910 |
} |
| 7911 |
} |
| 7912 |
|
| 7913 |
/* |
| 7914 |
* can_partial_agg |
| 7915 |
* |
| 7916 |
* Determines whether or not partial grouping and/or aggregation is possible. |
| 7917 |
* Returns true when possible, false otherwise. |
| 7918 |
*/ |
| 7919 |
static bool |
| 7920 |
can_partial_agg(PlannerInfo *root) |
| 7921 |
{ |
| 7922 |
Query *parse = root->parse; |
| 7923 |
|
| 7924 |
if (!parse->hasAggs && parse->groupClause == NIL) |
| 7925 |
{ |
| 7926 |
/* |
| 7927 |
* We don't know how to do parallel aggregation unless we have either |
| 7928 |
* some aggregates or a grouping clause. |
| 7929 |
*/ |
| 7930 |
return false; |
| 7931 |
} |
| 7932 |
else if (parse->groupingSets) |
| 7933 |
{ |
| 7934 |
/* We don't know how to do grouping sets in parallel. */ |
| 7935 |
return false; |
| 7936 |
} |
| 7937 |
else if (root->hasNonPartialAggs || root->hasNonSerialAggs) |
| 7938 |
{ |
| 7939 |
/* Insufficient support for partial mode. */ |
| 7940 |
return false; |
| 7941 |
} |
| 7942 |
|
| 7943 |
/* Everything looks good. */ |
| 7944 |
return true; |
| 7945 |
} |
| 7946 |
|
| 7947 |
/* |
| 7948 |
* apply_scanjoin_target_to_paths |
| 7949 |
* |
| 7950 |
* Adjust the final scan/join relation, and recursively all of its children, |
| 7951 |
* to generate the final scan/join target. It would be more correct to model |
| 7952 |
* this as a separate planning step with a new RelOptInfo at the toplevel and |
| 7953 |
* for each child relation, but doing it this way is noticeably cheaper. |
| 7954 |
* Maybe that problem can be solved at some point, but for now we do this. |
| 7955 |
* |
| 7956 |
* If tlist_same_exprs is true, then the scan/join target to be applied has |
| 7957 |
* the same expressions as the existing reltarget, so we need only insert the |
| 7958 |
* appropriate sortgroupref information. By avoiding the creation of |
| 7959 |
* projection paths we save effort both immediately and at plan creation time. |
| 7960 |
*/ |
| 7961 |
static void |
| 7962 |
apply_scanjoin_target_to_paths(PlannerInfo *root, |
| 7963 |
RelOptInfo *rel, |
| 7964 |
List *scanjoin_targets, |
| 7965 |
List *scanjoin_targets_contain_srfs, |
| 7966 |
bool scanjoin_target_parallel_safe, |
| 7967 |
bool tlist_same_exprs) |
| 7968 |
{ |
| 7969 |
bool rel_is_partitioned = IS_PARTITIONED_REL(rel); |
| 7970 |
PathTarget *scanjoin_target; |
| 7971 |
ListCell *lc; |
| 7972 |
|
| 7973 |
/* This recurses, so be paranoid. */ |
| 7974 |
check_stack_depth(); |
| 7975 |
|
| 7976 |
/* |
| 7977 |
* If the rel is partitioned, we want to drop its existing paths and |
| 7978 |
* generate new ones. This function would still be correct if we kept the |
| 7979 |
* existing paths: we'd modify them to generate the correct target above |
| 7980 |
* the partitioning Append, and then they'd compete on cost with paths |
| 7981 |
* generating the target below the Append. However, in our current cost |
| 7982 |
* model the latter way is always the same or cheaper cost, so modifying |
| 7983 |
* the existing paths would just be useless work. Moreover, when the cost |
| 7984 |
* is the same, varying roundoff errors might sometimes allow an existing |
| 7985 |
* path to be picked, resulting in undesirable cross-platform plan |
| 7986 |
* variations. So we drop old paths and thereby force the work to be done |
| 7987 |
* below the Append, except in the case of a non-parallel-safe target. |
| 7988 |
* |
| 7989 |
* Some care is needed, because we have to allow |
| 7990 |
* generate_useful_gather_paths to see the old partial paths in the next |
| 7991 |
* stanza. Hence, zap the main pathlist here, then allow |
| 7992 |
* generate_useful_gather_paths to add path(s) to the main list, and |
| 7993 |
* finally zap the partial pathlist. |
| 7994 |
*/ |
| 7995 |
if (rel_is_partitioned) |
| 7996 |
rel->pathlist = NIL; |
| 7997 |
|
| 7998 |
/* |
| 7999 |
* If the scan/join target is not parallel-safe, partial paths cannot |
| 8000 |
* generate it. |
| 8001 |
*/ |
| 8002 |
if (!scanjoin_target_parallel_safe) |
| 8003 |
{ |
| 8004 |
/* |
| 8005 |
* Since we can't generate the final scan/join target in parallel |
| 8006 |
* workers, this is our last opportunity to use any partial paths that |
| 8007 |
* exist; so build Gather path(s) that use them and emit whatever the |
| 8008 |
* current reltarget is. We don't do this in the case where the |
| 8009 |
* target is parallel-safe, since we will be able to generate superior |
| 8010 |
* paths by doing it after the final scan/join target has been |
| 8011 |
* applied. |
| 8012 |
*/ |
| 8013 |
generate_useful_gather_paths(root, rel, false); |
| 8014 |
|
| 8015 |
/* Can't use parallel query above this level. */ |
| 8016 |
rel->partial_pathlist = NIL; |
| 8017 |
rel->consider_parallel = false; |
| 8018 |
} |
| 8019 |
|
| 8020 |
/* Finish dropping old paths for a partitioned rel, per comment above */ |
| 8021 |
if (rel_is_partitioned) |
| 8022 |
rel->partial_pathlist = NIL; |
| 8023 |
|
| 8024 |
/* Extract SRF-free scan/join target. */ |
| 8025 |
scanjoin_target = linitial_node(PathTarget, scanjoin_targets); |
| 8026 |
|
| 8027 |
/* |
| 8028 |
* Apply the SRF-free scan/join target to each existing path. |
| 8029 |
* |
| 8030 |
* If the tlist exprs are the same, we can just inject the sortgroupref |
| 8031 |
* information into the existing pathtargets. Otherwise, replace each |
| 8032 |
* path with a projection path that generates the SRF-free scan/join |
| 8033 |
* target. This can't change the ordering of paths within rel->pathlist, |
| 8034 |
* so we just modify the list in place. |
| 8035 |
*/ |
| 8036 |
foreach(lc, rel->pathlist) |
| 8037 |
{ |
| 8038 |
Path *subpath = (Path *) lfirst(lc); |
| 8039 |
|
| 8040 |
/* Shouldn't have any parameterized paths anymore */ |
| 8041 |
Assert(subpath->param_info == NULL); |
| 8042 |
|
| 8043 |
if (tlist_same_exprs) |
| 8044 |
subpath->pathtarget->sortgrouprefs = |
| 8045 |
scanjoin_target->sortgrouprefs; |
| 8046 |
else |
| 8047 |
{ |
| 8048 |
Path *newpath; |
| 8049 |
|
| 8050 |
newpath = (Path *) create_projection_path(root, rel, subpath, |
| 8051 |
scanjoin_target); |
| 8052 |
lfirst(lc) = newpath; |
| 8053 |
} |
| 8054 |
} |
| 8055 |
|
| 8056 |
/* Likewise adjust the targets for any partial paths. */ |
| 8057 |
foreach(lc, rel->partial_pathlist) |
| 8058 |
{ |
| 8059 |
Path *subpath = (Path *) lfirst(lc); |
| 8060 |
|
| 8061 |
/* Shouldn't have any parameterized paths anymore */ |
| 8062 |
Assert(subpath->param_info == NULL); |
| 8063 |
|
| 8064 |
if (tlist_same_exprs) |
| 8065 |
subpath->pathtarget->sortgrouprefs = |
| 8066 |
scanjoin_target->sortgrouprefs; |
| 8067 |
else |
| 8068 |
{ |
| 8069 |
Path *newpath; |
| 8070 |
|
| 8071 |
newpath = (Path *) create_projection_path(root, rel, subpath, |
| 8072 |
scanjoin_target); |
| 8073 |
lfirst(lc) = newpath; |
| 8074 |
} |
| 8075 |
} |
| 8076 |
|
| 8077 |
/* |
| 8078 |
* Now, if final scan/join target contains SRFs, insert ProjectSetPath(s) |
| 8079 |
* atop each existing path. (Note that this function doesn't look at the |
| 8080 |
* cheapest-path fields, which is a good thing because they're bogus right |
| 8081 |
* now.) |
| 8082 |
*/ |
| 8083 |
if (root->parse->hasTargetSRFs) |
| 8084 |
adjust_paths_for_srfs(root, rel, |
| 8085 |
scanjoin_targets, |
| 8086 |
scanjoin_targets_contain_srfs); |
| 8087 |
|
| 8088 |
/* |
| 8089 |
* Update the rel's target to be the final (with SRFs) scan/join target. |
| 8090 |
* This now matches the actual output of all the paths, and we might get |
| 8091 |
* confused in createplan.c if they don't agree. We must do this now so |
| 8092 |
* that any append paths made in the next part will use the correct |
| 8093 |
* pathtarget (cf. create_append_path). |
| 8094 |
* |
| 8095 |
* Note that this is also necessary if GetForeignUpperPaths() gets called |
| 8096 |
* on the final scan/join relation or on any of its children, since the |
| 8097 |
* FDW might look at the rel's target to create ForeignPaths. |
| 8098 |
*/ |
| 8099 |
rel->reltarget = llast_node(PathTarget, scanjoin_targets); |
| 8100 |
|
| 8101 |
/* |
| 8102 |
* If the relation is partitioned, recursively apply the scan/join target |
| 8103 |
* to all partitions, and generate brand-new Append paths in which the |
| 8104 |
* scan/join target is computed below the Append rather than above it. |
| 8105 |
* Since Append is not projection-capable, that might save a separate |
| 8106 |
* Result node, and it also is important for partitionwise aggregate. |
| 8107 |
*/ |
| 8108 |
if (rel_is_partitioned) |
| 8109 |
{ |
| 8110 |
List *live_children = NIL; |
| 8111 |
int i; |
| 8112 |
|
| 8113 |
/* Adjust each partition. */ |
| 8114 |
i = -1; |
| 8115 |
while ((i = bms_next_member(rel->live_parts, i)) >= 0) |
| 8116 |
{ |
| 8117 |
RelOptInfo *child_rel = rel->part_rels[i]; |
| 8118 |
AppendRelInfo **appinfos; |
| 8119 |
int nappinfos; |
| 8120 |
List *child_scanjoin_targets = NIL; |
| 8121 |
|
| 8122 |
Assert(child_rel != NULL); |
| 8123 |
|
| 8124 |
/* Dummy children can be ignored. */ |
| 8125 |
if (IS_DUMMY_REL(child_rel)) |
| 8126 |
continue; |
| 8127 |
|
| 8128 |
/* Translate scan/join targets for this child. */ |
| 8129 |
appinfos = find_appinfos_by_relids(root, child_rel->relids, |
| 8130 |
&nappinfos); |
| 8131 |
foreach(lc, scanjoin_targets) |
| 8132 |
{ |
| 8133 |
PathTarget *target = lfirst_node(PathTarget, lc); |
| 8134 |
|
| 8135 |
target = copy_pathtarget(target); |
| 8136 |
target->exprs = (List *) |
| 8137 |
adjust_appendrel_attrs(root, |
| 8138 |
(Node *) target->exprs, |
| 8139 |
nappinfos, appinfos); |
| 8140 |
child_scanjoin_targets = lappend(child_scanjoin_targets, |
| 8141 |
target); |
| 8142 |
} |
| 8143 |
pfree(appinfos); |
| 8144 |
|
| 8145 |
/* Recursion does the real work. */ |
| 8146 |
apply_scanjoin_target_to_paths(root, child_rel, |
| 8147 |
child_scanjoin_targets, |
| 8148 |
scanjoin_targets_contain_srfs, |
| 8149 |
scanjoin_target_parallel_safe, |
| 8150 |
tlist_same_exprs); |
| 8151 |
|
| 8152 |
/* Save non-dummy children for Append paths. */ |
| 8153 |
if (!IS_DUMMY_REL(child_rel)) |
| 8154 |
live_children = lappend(live_children, child_rel); |
| 8155 |
} |
| 8156 |
|
| 8157 |
/* Build new paths for this relation by appending child paths. */ |
| 8158 |
add_paths_to_append_rel(root, rel, live_children); |
| 8159 |
} |
| 8160 |
|
| 8161 |
/* |
| 8162 |
* Consider generating Gather or Gather Merge paths. We must only do this |
| 8163 |
* if the relation is parallel safe, and we don't do it for child rels to |
| 8164 |
* avoid creating multiple Gather nodes within the same plan. We must do |
| 8165 |
* this after all paths have been generated and before set_cheapest, since |
| 8166 |
* one of the generated paths may turn out to be the cheapest one. |
| 8167 |
*/ |
| 8168 |
if (rel->consider_parallel && !IS_OTHER_REL(rel)) |
| 8169 |
generate_useful_gather_paths(root, rel, false); |
| 8170 |
|
| 8171 |
/* |
| 8172 |
* Reassess which paths are the cheapest, now that we've potentially added |
| 8173 |
* new Gather (or Gather Merge) and/or Append (or MergeAppend) paths to |
| 8174 |
* this relation. |
| 8175 |
*/ |
| 8176 |
set_cheapest(rel); |
| 8177 |
} |
| 8178 |
|
| 8179 |
/* |
| 8180 |
* create_partitionwise_grouping_paths |
| 8181 |
* |
| 8182 |
* If the partition keys of input relation are part of the GROUP BY clause, all |
| 8183 |
* the rows belonging to a given group come from a single partition. This |
| 8184 |
* allows aggregation/grouping over a partitioned relation to be broken down |
| 8185 |
* into aggregation/grouping on each partition. This should be no worse, and |
| 8186 |
* often better, than the normal approach. |
| 8187 |
* |
| 8188 |
* However, if the GROUP BY clause does not contain all the partition keys, |
| 8189 |
* rows from a given group may be spread across multiple partitions. In that |
| 8190 |
* case, we perform partial aggregation for each group, append the results, |
| 8191 |
* and then finalize aggregation. This is less certain to win than the |
| 8192 |
* previous case. It may win if the PartialAggregate stage greatly reduces |
| 8193 |
* the number of groups, because fewer rows will pass through the Append node. |
| 8194 |
* It may lose if we have lots of small groups. |
| 8195 |
*/ |
| 8196 |
static void |
| 8197 |
create_partitionwise_grouping_paths(PlannerInfo *root, |
| 8198 |
RelOptInfo *input_rel, |
| 8199 |
RelOptInfo *grouped_rel, |
| 8200 |
RelOptInfo *partially_grouped_rel, |
| 8201 |
const AggClauseCosts *agg_costs, |
| 8202 |
grouping_sets_data *gd, |
| 8203 |
PartitionwiseAggregateType patype, |
| 8204 |
GroupPathExtraData *extra) |
| 8205 |
{ |
| 8206 |
List *grouped_live_children = NIL; |
| 8207 |
List *partially_grouped_live_children = NIL; |
| 8208 |
PathTarget *target = grouped_rel->reltarget; |
| 8209 |
bool partial_grouping_valid = true; |
| 8210 |
int i; |
| 8211 |
|
| 8212 |
Assert(patype != PARTITIONWISE_AGGREGATE_NONE); |
| 8213 |
Assert(patype != PARTITIONWISE_AGGREGATE_PARTIAL || |
| 8214 |
partially_grouped_rel != NULL); |
| 8215 |
|
| 8216 |
/* Add paths for partitionwise aggregation/grouping. */ |
| 8217 |
i = -1; |
| 8218 |
while ((i = bms_next_member(input_rel->live_parts, i)) >= 0) |
| 8219 |
{ |
| 8220 |
RelOptInfo *child_input_rel = input_rel->part_rels[i]; |
| 8221 |
PathTarget *child_target; |
| 8222 |
AppendRelInfo **appinfos; |
| 8223 |
int nappinfos; |
| 8224 |
GroupPathExtraData child_extra; |
| 8225 |
RelOptInfo *child_grouped_rel; |
| 8226 |
RelOptInfo *child_partially_grouped_rel; |
| 8227 |
|
| 8228 |
Assert(child_input_rel != NULL); |
| 8229 |
|
| 8230 |
/* Dummy children can be ignored. */ |
| 8231 |
if (IS_DUMMY_REL(child_input_rel)) |
| 8232 |
continue; |
| 8233 |
|
| 8234 |
child_target = copy_pathtarget(target); |
| 8235 |
|
| 8236 |
/* |
| 8237 |
* Copy the given "extra" structure as is and then override the |
| 8238 |
* members specific to this child. |
| 8239 |
*/ |
| 8240 |
memcpy(&child_extra, extra, sizeof(child_extra)); |
| 8241 |
|
| 8242 |
appinfos = find_appinfos_by_relids(root, child_input_rel->relids, |
| 8243 |
&nappinfos); |
| 8244 |
|
| 8245 |
child_target->exprs = (List *) |
| 8246 |
adjust_appendrel_attrs(root, |
| 8247 |
(Node *) target->exprs, |
| 8248 |
nappinfos, appinfos); |
| 8249 |
|
| 8250 |
/* Translate havingQual and targetList. */ |
| 8251 |
child_extra.havingQual = (Node *) |
| 8252 |
adjust_appendrel_attrs(root, |
| 8253 |
extra->havingQual, |
| 8254 |
nappinfos, appinfos); |
| 8255 |
child_extra.targetList = (List *) |
| 8256 |
adjust_appendrel_attrs(root, |
| 8257 |
(Node *) extra->targetList, |
| 8258 |
nappinfos, appinfos); |
| 8259 |
|
| 8260 |
/* |
| 8261 |
* extra->patype was the value computed for our parent rel; patype is |
| 8262 |
* the value for this relation. For the child, our value is its |
| 8263 |
* parent rel's value. |
| 8264 |
*/ |
| 8265 |
child_extra.patype = patype; |
| 8266 |
|
| 8267 |
/* |
| 8268 |
* Create grouping relation to hold fully aggregated grouping and/or |
| 8269 |
* aggregation paths for the child. |
| 8270 |
*/ |
| 8271 |
child_grouped_rel = make_grouping_rel(root, child_input_rel, |
| 8272 |
child_target, |
| 8273 |
extra->target_parallel_safe, |
| 8274 |
child_extra.havingQual); |
| 8275 |
|
| 8276 |
/* Create grouping paths for this child relation. */ |
| 8277 |
create_ordinary_grouping_paths(root, child_input_rel, |
| 8278 |
child_grouped_rel, |
| 8279 |
agg_costs, gd, &child_extra, |
| 8280 |
&child_partially_grouped_rel); |
| 8281 |
|
| 8282 |
if (child_partially_grouped_rel) |
| 8283 |
{ |
| 8284 |
partially_grouped_live_children = |
| 8285 |
lappend(partially_grouped_live_children, |
| 8286 |
child_partially_grouped_rel); |
| 8287 |
} |
| 8288 |
else |
| 8289 |
partial_grouping_valid = false; |
| 8290 |
|
| 8291 |
if (patype == PARTITIONWISE_AGGREGATE_FULL) |
| 8292 |
{ |
| 8293 |
set_cheapest(child_grouped_rel); |
| 8294 |
grouped_live_children = lappend(grouped_live_children, |
| 8295 |
child_grouped_rel); |
| 8296 |
} |
| 8297 |
|
| 8298 |
pfree(appinfos); |
| 8299 |
} |
| 8300 |
|
| 8301 |
/* |
| 8302 |
* Try to create append paths for partially grouped children. For full |
| 8303 |
* partitionwise aggregation, we might have paths in the partial_pathlist |
| 8304 |
* if parallel aggregation is possible. For partial partitionwise |
| 8305 |
* aggregation, we may have paths in both pathlist and partial_pathlist. |
| 8306 |
* |
| 8307 |
* NB: We must have a partially grouped path for every child in order to |
| 8308 |
* generate a partially grouped path for this relation. |
| 8309 |
*/ |
| 8310 |
if (partially_grouped_rel && partial_grouping_valid) |
| 8311 |
{ |
| 8312 |
Assert(partially_grouped_live_children != NIL); |
| 8313 |
|
| 8314 |
add_paths_to_append_rel(root, partially_grouped_rel, |
| 8315 |
partially_grouped_live_children); |
| 8316 |
|
| 8317 |
/* |
| 8318 |
* We need call set_cheapest, since the finalization step will use the |
| 8319 |
* cheapest path from the rel. |
| 8320 |
*/ |
| 8321 |
if (partially_grouped_rel->pathlist) |
| 8322 |
set_cheapest(partially_grouped_rel); |
| 8323 |
} |
| 8324 |
|
| 8325 |
/* If possible, create append paths for fully grouped children. */ |
| 8326 |
if (patype == PARTITIONWISE_AGGREGATE_FULL) |
| 8327 |
{ |
| 8328 |
Assert(grouped_live_children != NIL); |
| 8329 |
|
| 8330 |
add_paths_to_append_rel(root, grouped_rel, grouped_live_children); |
| 8331 |
} |
| 8332 |
} |
| 8333 |
|
| 8334 |
/* |
| 8335 |
* group_by_has_partkey |
| 8336 |
* |
| 8337 |
* Returns true if all the partition keys of the given relation are part of |
| 8338 |
* the GROUP BY clauses, including having matching collation, false otherwise. |
| 8339 |
*/ |
| 8340 |
static bool |
| 8341 |
group_by_has_partkey(RelOptInfo *input_rel, |
| 8342 |
List *targetList, |
| 8343 |
List *groupClause) |
| 8344 |
{ |
| 8345 |
List *groupexprs = get_sortgrouplist_exprs(groupClause, targetList); |
| 8346 |
int cnt = 0; |
| 8347 |
int partnatts; |
| 8348 |
|
| 8349 |
/* Input relation should be partitioned. */ |
| 8350 |
Assert(input_rel->part_scheme); |
| 8351 |
|
| 8352 |
/* Rule out early, if there are no partition keys present. */ |
| 8353 |
if (!input_rel->partexprs) |
| 8354 |
return false; |
| 8355 |
|
| 8356 |
partnatts = input_rel->part_scheme->partnatts; |
| 8357 |
|
| 8358 |
for (cnt = 0; cnt < partnatts; cnt++) |
| 8359 |
{ |
| 8360 |
List *partexprs = input_rel->partexprs[cnt]; |
| 8361 |
ListCell *lc; |
| 8362 |
bool found = false; |
| 8363 |
|
| 8364 |
foreach(lc, partexprs) |
| 8365 |
{ |
| 8366 |
ListCell *lg; |
| 8367 |
Expr *partexpr = lfirst(lc); |
| 8368 |
Oid partcoll = input_rel->part_scheme->partcollation[cnt]; |
| 8369 |
|
| 8370 |
foreach(lg, groupexprs) |
| 8371 |
{ |
| 8372 |
Expr *groupexpr = lfirst(lg); |
| 8373 |
Oid groupcoll = exprCollation((Node *) groupexpr); |
| 8374 |
|
| 8375 |
/* |
| 8376 |
* Note: we can assume there is at most one RelabelType node; |
| 8377 |
* eval_const_expressions() will have simplified if more than |
| 8378 |
* one. |
| 8379 |
*/ |
| 8380 |
if (IsA(groupexpr, RelabelType)) |
| 8381 |
groupexpr = ((RelabelType *) groupexpr)->arg; |
| 8382 |
|
| 8383 |
if (equal(groupexpr, partexpr)) |
| 8384 |
{ |
| 8385 |
/* |
| 8386 |
* Reject a match if the grouping collation does not match |
| 8387 |
* the partitioning collation. |
| 8388 |
*/ |
| 8389 |
if (OidIsValid(partcoll) && OidIsValid(groupcoll) && |
| 8390 |
partcoll != groupcoll) |
| 8391 |
return false; |
| 8392 |
|
| 8393 |
found = true; |
| 8394 |
break; |
| 8395 |
} |
| 8396 |
} |
| 8397 |
|
| 8398 |
if (found) |
| 8399 |
break; |
| 8400 |
} |
| 8401 |
|
| 8402 |
/* |
| 8403 |
* If none of the partition key expressions match with any of the |
| 8404 |
* GROUP BY expression, return false. |
| 8405 |
*/ |
| 8406 |
if (!found) |
| 8407 |
return false; |
| 8408 |
} |
| 8409 |
|
| 8410 |
return true; |
| 8411 |
} |
| 8412 |
|
| 8413 |
/* |
| 8414 |
* generate_setop_child_grouplist |
| 8415 |
* Build a SortGroupClause list defining the sort/grouping properties |
| 8416 |
* of the child of a set operation. |
| 8417 |
* |
| 8418 |
* This is similar to generate_setop_grouplist() but differs as the setop |
| 8419 |
* child query's targetlist entries may already have a tleSortGroupRef |
| 8420 |
* assigned for other purposes, such as GROUP BYs. Here we keep the |
| 8421 |
* SortGroupClause list in the same order as 'op' groupClauses and just adjust |
| 8422 |
* the tleSortGroupRef to reference the TargetEntry's 'ressortgroupref'. If |
| 8423 |
* any of the columns in the targetlist don't match to the setop's colTypes |
| 8424 |
* then we return an empty list. This may leave some TLEs with unreferenced |
| 8425 |
* ressortgroupref markings, but that's harmless. |
| 8426 |
*/ |
| 8427 |
static List * |
| 8428 |
generate_setop_child_grouplist(SetOperationStmt *op, List *targetlist) |
| 8429 |
{ |
| 8430 |
List *grouplist = copyObject(op->groupClauses); |
| 8431 |
ListCell *lg; |
| 8432 |
ListCell *lt; |
| 8433 |
ListCell *ct; |
| 8434 |
|
| 8435 |
lg = list_head(grouplist); |
| 8436 |
ct = list_head(op->colTypes); |
| 8437 |
foreach(lt, targetlist) |
| 8438 |
{ |
| 8439 |
TargetEntry *tle = (TargetEntry *) lfirst(lt); |
| 8440 |
SortGroupClause *sgc; |
| 8441 |
Oid coltype; |
| 8442 |
|
| 8443 |
/* resjunk columns could have sortgrouprefs. Leave these alone */ |
| 8444 |
if (tle->resjunk) |
| 8445 |
continue; |
| 8446 |
|
| 8447 |
/* |
| 8448 |
* We expect every non-resjunk target to have a SortGroupClause and |
| 8449 |
* colTypes. |
| 8450 |
*/ |
| 8451 |
Assert(lg != NULL); |
| 8452 |
Assert(ct != NULL); |
| 8453 |
sgc = (SortGroupClause *) lfirst(lg); |
| 8454 |
coltype = lfirst_oid(ct); |
| 8455 |
|
| 8456 |
/* reject if target type isn't the same as the setop target type */ |
| 8457 |
if (coltype != exprType((Node *) tle->expr)) |
| 8458 |
return NIL; |
| 8459 |
|
| 8460 |
lg = lnext(grouplist, lg); |
| 8461 |
ct = lnext(op->colTypes, ct); |
| 8462 |
|
| 8463 |
/* assign a tleSortGroupRef, or reuse the existing one */ |
| 8464 |
sgc->tleSortGroupRef = assignSortGroupRef(tle, targetlist); |
| 8465 |
} |
| 8466 |
|
| 8467 |
Assert(lg == NULL); |
| 8468 |
Assert(ct == NULL); |
| 8469 |
|
| 8470 |
return grouplist; |
| 8471 |
} |