granicus.if.org Git - postgresql/blob - src/backend/optimizer/plan/planner.c

   1 /*-------------------------------------------------------------------------
   2  *
   3  * planner.c
   4  *        The query optimizer external interface.
   5  *
   6  * Portions Copyright (c) 1996-2019, 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/htup_details.h"
  23 #include "access/parallel.h"
  24 #include "access/sysattr.h"
  25 #include "access/table.h"
  26 #include "access/xact.h"
  27 #include "catalog/pg_constraint.h"
  28 #include "catalog/pg_inherits.h"
  29 #include "catalog/pg_proc.h"
  30 #include "catalog/pg_type.h"
  31 #include "executor/executor.h"
  32 #include "executor/nodeAgg.h"
  33 #include "foreign/fdwapi.h"
  34 #include "miscadmin.h"
  35 #include "jit/jit.h"
  36 #include "lib/bipartite_match.h"
  37 #include "lib/knapsack.h"
  38 #include "nodes/makefuncs.h"
  39 #include "nodes/nodeFuncs.h"
  40 #ifdef OPTIMIZER_DEBUG
  41 #include "nodes/print.h"
  42 #endif
  43 #include "optimizer/appendinfo.h"
  44 #include "optimizer/clauses.h"
  45 #include "optimizer/cost.h"
  46 #include "optimizer/inherit.h"
  47 #include "optimizer/optimizer.h"
  48 #include "optimizer/paramassign.h"
  49 #include "optimizer/pathnode.h"
  50 #include "optimizer/paths.h"
  51 #include "optimizer/plancat.h"
  52 #include "optimizer/planmain.h"
  53 #include "optimizer/planner.h"
  54 #include "optimizer/prep.h"
  55 #include "optimizer/subselect.h"
  56 #include "optimizer/tlist.h"
  57 #include "parser/analyze.h"
  58 #include "parser/parsetree.h"
  59 #include "parser/parse_agg.h"
  60 #include "partitioning/partdesc.h"
  61 #include "rewrite/rewriteManip.h"
  62 #include "storage/dsm_impl.h"
  63 #include "utils/rel.h"
  64 #include "utils/selfuncs.h"
  65 #include "utils/lsyscache.h"
  66 #include "utils/syscache.h"
  67
  68
  69 /* GUC parameters */
  70 double          cursor_tuple_fraction = DEFAULT_CURSOR_TUPLE_FRACTION;
  71 int                     force_parallel_mode = FORCE_PARALLEL_OFF;
  72 bool            parallel_leader_participation = true;
  73
  74 /* Hook for plugins to get control in planner() */
  75 planner_hook_type planner_hook = NULL;
  76
  77 /* Hook for plugins to get control when grouping_planner() plans upper rels */
  78 create_upper_paths_hook_type create_upper_paths_hook = NULL;
  79
  80
  81 /* Expression kind codes for preprocess_expression */
  82 #define EXPRKIND_QUAL                           0
  83 #define EXPRKIND_TARGET                         1
  84 #define EXPRKIND_RTFUNC                         2
  85 #define EXPRKIND_RTFUNC_LATERAL         3
  86 #define EXPRKIND_VALUES                         4
  87 #define EXPRKIND_VALUES_LATERAL         5
  88 #define EXPRKIND_LIMIT                          6
  89 #define EXPRKIND_APPINFO                        7
  90 #define EXPRKIND_PHV                            8
  91 #define EXPRKIND_TABLESAMPLE            9
  92 #define EXPRKIND_ARBITER_ELEM           10
  93 #define EXPRKIND_TABLEFUNC                      11
  94 #define EXPRKIND_TABLEFUNC_LATERAL      12
  95
  96 /* Passthrough data for standard_qp_callback */
  97 typedef struct
  98 {
  99         List       *activeWindows;      /* active windows, if any */
 100         List       *groupClause;        /* overrides parse->groupClause */
 101 } standard_qp_extra;
 102
 103 /*
 104  * Data specific to grouping sets
 105  */
 106
 107 typedef struct
 108 {
 109         List       *rollups;
 110         List       *hash_sets_idx;
 111         double          dNumHashGroups;
 112         bool            any_hashable;
 113         Bitmapset  *unsortable_refs;
 114         Bitmapset  *unhashable_refs;
 115         List       *unsortable_sets;
 116         int                *tleref_to_colnum_map;
 117 } grouping_sets_data;
 118
 119 /*
 120  * Temporary structure for use during WindowClause reordering in order to be
 121  * able to sort WindowClauses on partitioning/ordering prefix.
 122  */
 123 typedef struct
 124 {
 125         WindowClause *wc;
 126         List       *uniqueOrder;        /* A List of unique ordering/partitioning
 127                                                                  * clauses per Window */
 128 } WindowClauseSortData;
 129
 130 /* Local functions */
 131 static Node *preprocess_expression(PlannerInfo *root, Node *expr, int kind);
 132 static void preprocess_qual_conditions(PlannerInfo *root, Node *jtnode);
 133 static void inheritance_planner(PlannerInfo *root);
 134 static void grouping_planner(PlannerInfo *root, bool inheritance_update,
 135                                                          double tuple_fraction);
 136 static grouping_sets_data *preprocess_grouping_sets(PlannerInfo *root);
 137 static List *remap_to_groupclause_idx(List *groupClause, List *gsets,
 138                                                                           int *tleref_to_colnum_map);
 139 static void preprocess_rowmarks(PlannerInfo *root);
 140 static double preprocess_limit(PlannerInfo *root,
 141                                                            double tuple_fraction,
 142                                                            int64 *offset_est, int64 *count_est);
 143 static void remove_useless_groupby_columns(PlannerInfo *root);
 144 static List *preprocess_groupclause(PlannerInfo *root, List *force);
 145 static List *extract_rollup_sets(List *groupingSets);
 146 static List *reorder_grouping_sets(List *groupingSets, List *sortclause);
 147 static void standard_qp_callback(PlannerInfo *root, void *extra);
 148 static double get_number_of_groups(PlannerInfo *root,
 149                                                                    double path_rows,
 150                                                                    grouping_sets_data *gd,
 151                                                                    List *target_list);
 152 static RelOptInfo *create_grouping_paths(PlannerInfo *root,
 153                                                                                  RelOptInfo *input_rel,
 154                                                                                  PathTarget *target,
 155                                                                                  bool target_parallel_safe,
 156                                                                                  const AggClauseCosts *agg_costs,
 157                                                                                  grouping_sets_data *gd);
 158 static bool is_degenerate_grouping(PlannerInfo *root);
 159 static void create_degenerate_grouping_paths(PlannerInfo *root,
 160                                                                                          RelOptInfo *input_rel,
 161                                                                                          RelOptInfo *grouped_rel);
 162 static RelOptInfo *make_grouping_rel(PlannerInfo *root, RelOptInfo *input_rel,
 163                                                                          PathTarget *target, bool target_parallel_safe,
 164                                                                          Node *havingQual);
 165 static void create_ordinary_grouping_paths(PlannerInfo *root,
 166                                                                                    RelOptInfo *input_rel,
 167                                                                                    RelOptInfo *grouped_rel,
 168                                                                                    const AggClauseCosts *agg_costs,
 169                                                                                    grouping_sets_data *gd,
 170                                                                                    GroupPathExtraData *extra,
 171                                                                                    RelOptInfo **partially_grouped_rel_p);
 172 static void consider_groupingsets_paths(PlannerInfo *root,
 173                                                                                 RelOptInfo *grouped_rel,
 174                                                                                 Path *path,
 175                                                                                 bool is_sorted,
 176                                                                                 bool can_hash,
 177                                                                                 grouping_sets_data *gd,
 178                                                                                 const AggClauseCosts *agg_costs,
 179                                                                                 double dNumGroups);
 180 static RelOptInfo *create_window_paths(PlannerInfo *root,
 181                                                                            RelOptInfo *input_rel,
 182                                                                            PathTarget *input_target,
 183                                                                            PathTarget *output_target,
 184                                                                            bool output_target_parallel_safe,
 185                                                                            WindowFuncLists *wflists,
 186                                                                            List *activeWindows);
 187 static void create_one_window_path(PlannerInfo *root,
 188                                                                    RelOptInfo *window_rel,
 189                                                                    Path *path,
 190                                                                    PathTarget *input_target,
 191                                                                    PathTarget *output_target,
 192                                                                    WindowFuncLists *wflists,
 193                                                                    List *activeWindows);
 194 static RelOptInfo *create_distinct_paths(PlannerInfo *root,
 195                                                                                  RelOptInfo *input_rel);
 196 static RelOptInfo *create_ordered_paths(PlannerInfo *root,
 197                                                                                 RelOptInfo *input_rel,
 198                                                                                 PathTarget *target,
 199                                                                                 bool target_parallel_safe,
 200                                                                                 double limit_tuples);
 201 static PathTarget *make_group_input_target(PlannerInfo *root,
 202                                                                                    PathTarget *final_target);
 203 static PathTarget *make_partial_grouping_target(PlannerInfo *root,
 204                                                                                                 PathTarget *grouping_target,
 205                                                                                                 Node *havingQual);
 206 static List *postprocess_setop_tlist(List *new_tlist, List *orig_tlist);
 207 static List *select_active_windows(PlannerInfo *root, WindowFuncLists *wflists);
 208 static PathTarget *make_window_input_target(PlannerInfo *root,
 209                                                                                         PathTarget *final_target,
 210                                                                                         List *activeWindows);
 211 static List *make_pathkeys_for_window(PlannerInfo *root, WindowClause *wc,
 212                                                                           List *tlist);
 213 static PathTarget *make_sort_input_target(PlannerInfo *root,
 214                                                                                   PathTarget *final_target,
 215                                                                                   bool *have_postponed_srfs);
 216 static void adjust_paths_for_srfs(PlannerInfo *root, RelOptInfo *rel,
 217                                                                   List *targets, List *targets_contain_srfs);
 218 static void add_paths_to_grouping_rel(PlannerInfo *root, RelOptInfo *input_rel,
 219                                                                           RelOptInfo *grouped_rel,
 220                                                                           RelOptInfo *partially_grouped_rel,
 221                                                                           const AggClauseCosts *agg_costs,
 222                                                                           grouping_sets_data *gd,
 223                                                                           double dNumGroups,
 224                                                                           GroupPathExtraData *extra);
 225 static RelOptInfo *create_partial_grouping_paths(PlannerInfo *root,
 226                                                                                                  RelOptInfo *grouped_rel,
 227                                                                                                  RelOptInfo *input_rel,
 228                                                                                                  grouping_sets_data *gd,
 229                                                                                                  GroupPathExtraData *extra,
 230                                                                                                  bool force_rel_creation);
 231 static void gather_grouping_paths(PlannerInfo *root, RelOptInfo *rel);
 232 static bool can_partial_agg(PlannerInfo *root,
 233                                                         const AggClauseCosts *agg_costs);
 234 static void apply_scanjoin_target_to_paths(PlannerInfo *root,
 235                                                                                    RelOptInfo *rel,
 236                                                                                    List *scanjoin_targets,
 237                                                                                    List *scanjoin_targets_contain_srfs,
 238                                                                                    bool scanjoin_target_parallel_safe,
 239                                                                                    bool tlist_same_exprs);
 240 static void create_partitionwise_grouping_paths(PlannerInfo *root,
 241                                                                                                 RelOptInfo *input_rel,
 242                                                                                                 RelOptInfo *grouped_rel,
 243                                                                                                 RelOptInfo *partially_grouped_rel,
 244                                                                                                 const AggClauseCosts *agg_costs,
 245                                                                                                 grouping_sets_data *gd,
 246                                                                                                 PartitionwiseAggregateType patype,
 247                                                                                                 GroupPathExtraData *extra);
 248 static bool group_by_has_partkey(RelOptInfo *input_rel,
 249                                                                  List *targetList,
 250                                                                  List *groupClause);
 251 static int      common_prefix_cmp(const void *a, const void *b);
 252
 253
 254 /*****************************************************************************
 255  *
 256  *         Query optimizer entry point
 257  *
 258  * To support loadable plugins that monitor or modify planner behavior,
 259  * we provide a hook variable that lets a plugin get control before and
 260  * after the standard planning process.  The plugin would normally call
 261  * standard_planner().
 262  *
 263  * Note to plugin authors: standard_planner() scribbles on its Query input,
 264  * so you'd better copy that data structure if you want to plan more than once.
 265  *
 266  *****************************************************************************/
 267 PlannedStmt *
 268 planner(Query *parse, int cursorOptions, ParamListInfo boundParams)
 269 {
 270         PlannedStmt *result;
 271
 272         if (planner_hook)
 273                 result = (*planner_hook) (parse, cursorOptions, boundParams);
 274         else
 275                 result = standard_planner(parse, cursorOptions, boundParams);
 276         return result;
 277 }
 278
 279 PlannedStmt *
 280 standard_planner(Query *parse, int cursorOptions, ParamListInfo boundParams)
 281 {
 282         PlannedStmt *result;
 283         PlannerGlobal *glob;
 284         double          tuple_fraction;
 285         PlannerInfo *root;
 286         RelOptInfo *final_rel;
 287         Path       *best_path;
 288         Plan       *top_plan;
 289         ListCell   *lp,
 290                            *lr;
 291
 292         /*
 293          * Set up global state for this planner invocation.  This data is needed
 294          * across all levels of sub-Query that might exist in the given command,
 295          * so we keep it in a separate struct that's linked to by each per-Query
 296          * PlannerInfo.
 297          */
 298         glob = makeNode(PlannerGlobal);
 299
 300         glob->boundParams = boundParams;
 301         glob->subplans = NIL;
 302         glob->subroots = NIL;
 303         glob->rewindPlanIDs = NULL;
 304         glob->finalrtable = NIL;
 305         glob->finalrowmarks = NIL;
 306         glob->resultRelations = NIL;
 307         glob->rootResultRelations = NIL;
 308         glob->relationOids = NIL;
 309         glob->invalItems = NIL;
 310         glob->paramExecTypes = NIL;
 311         glob->lastPHId = 0;
 312         glob->lastRowMarkId = 0;
 313         glob->lastPlanNodeId = 0;
 314         glob->transientPlan = false;
 315         glob->dependsOnRole = false;
 316
 317         /*
 318          * Assess whether it's feasible to use parallel mode for this query. We
 319          * can't do this in a standalone backend, or if the command will try to
 320          * modify any data, or if this is a cursor operation, or if GUCs are set
 321          * to values that don't permit parallelism, or if parallel-unsafe
 322          * functions are present in the query tree.
 323          *
 324          * (Note that we do allow CREATE TABLE AS, SELECT INTO, and CREATE
 325          * MATERIALIZED VIEW to use parallel plans, but this is safe only because
 326          * the command is writing into a completely new table which workers won't
 327          * be able to see.  If the workers could see the table, the fact that
 328          * group locking would cause them to ignore the leader's heavyweight
 329          * relation extension lock and GIN page locks would make this unsafe.
 330          * We'll have to fix that somehow if we want to allow parallel inserts in
 331          * general; updates and deletes have additional problems especially around
 332          * combo CIDs.)
 333          *
 334          * For now, we don't try to use parallel mode if we're running inside a
 335          * parallel worker.  We might eventually be able to relax this
 336          * restriction, but for now it seems best not to have parallel workers
 337          * trying to create their own parallel workers.
 338          */
 339         if ((cursorOptions & CURSOR_OPT_PARALLEL_OK) != 0 &&
 340                 IsUnderPostmaster &&
 341                 parse->commandType == CMD_SELECT &&
 342                 !parse->hasModifyingCTE &&
 343                 max_parallel_workers_per_gather > 0 &&
 344                 !IsParallelWorker())
 345         {
 346                 /* all the cheap tests pass, so scan the query tree */
 347                 glob->maxParallelHazard = max_parallel_hazard(parse);
 348                 glob->parallelModeOK = (glob->maxParallelHazard != PROPARALLEL_UNSAFE);
 349         }
 350         else
 351         {
 352                 /* skip the query tree scan, just assume it's unsafe */
 353                 glob->maxParallelHazard = PROPARALLEL_UNSAFE;
 354                 glob->parallelModeOK = false;
 355         }
 356
 357         /*
 358          * glob->parallelModeNeeded is normally set to false here and changed to
 359          * true during plan creation if a Gather or Gather Merge plan is actually
 360          * created (cf. create_gather_plan, create_gather_merge_plan).
 361          *
 362          * However, if force_parallel_mode = on or force_parallel_mode = regress,
 363          * then we impose parallel mode whenever it's safe to do so, even if the
 364          * final plan doesn't use parallelism.  It's not safe to do so if the
 365          * query contains anything parallel-unsafe; parallelModeOK will be false
 366          * in that case.  Note that parallelModeOK can't change after this point.
 367          * Otherwise, everything in the query is either parallel-safe or
 368          * parallel-restricted, and in either case it should be OK to impose
 369          * parallel-mode restrictions.  If that ends up breaking something, then
 370          * either some function the user included in the query is incorrectly
 371          * labelled as parallel-safe or parallel-restricted when in reality it's
 372          * parallel-unsafe, or else the query planner itself has a bug.
 373          */
 374         glob->parallelModeNeeded = glob->parallelModeOK &&
 375                 (force_parallel_mode != FORCE_PARALLEL_OFF);
 376
 377         /* Determine what fraction of the plan is likely to be scanned */
 378         if (cursorOptions & CURSOR_OPT_FAST_PLAN)
 379         {
 380                 /*
 381                  * We have no real idea how many tuples the user will ultimately FETCH
 382                  * from a cursor, but it is often the case that he doesn't want 'em
 383                  * all, or would prefer a fast-start plan anyway so that he can
 384                  * process some of the tuples sooner.  Use a GUC parameter to decide
 385                  * what fraction to optimize for.
 386                  */
 387                 tuple_fraction = cursor_tuple_fraction;
 388
 389                 /*
 390                  * We document cursor_tuple_fraction as simply being a fraction, which
 391                  * means the edge cases 0 and 1 have to be treated specially here.  We
 392                  * convert 1 to 0 ("all the tuples") and 0 to a very small fraction.
 393                  */
 394                 if (tuple_fraction >= 1.0)
 395                         tuple_fraction = 0.0;
 396                 else if (tuple_fraction <= 0.0)
 397                         tuple_fraction = 1e-10;
 398         }
 399         else
 400         {
 401                 /* Default assumption is we need all the tuples */
 402                 tuple_fraction = 0.0;
 403         }
 404
 405         /* primary planning entry point (may recurse for subqueries) */
 406         root = subquery_planner(glob, parse, NULL,
 407                                                         false, tuple_fraction);
 408
 409         /* Select best Path and turn it into a Plan */
 410         final_rel = fetch_upper_rel(root, UPPERREL_FINAL, NULL);
 411         best_path = get_cheapest_fractional_path(final_rel, tuple_fraction);
 412
 413         top_plan = create_plan(root, best_path);
 414
 415         /*
 416          * If creating a plan for a scrollable cursor, make sure it can run
 417          * backwards on demand.  Add a Material node at the top at need.
 418          */
 419         if (cursorOptions & CURSOR_OPT_SCROLL)
 420         {
 421                 if (!ExecSupportsBackwardScan(top_plan))
 422                         top_plan = materialize_finished_plan(top_plan);
 423         }
 424
 425         /*
 426          * Optionally add a Gather node for testing purposes, provided this is
 427          * actually a safe thing to do.
 428          */
 429         if (force_parallel_mode != FORCE_PARALLEL_OFF && top_plan->parallel_safe)
 430         {
 431                 Gather     *gather = makeNode(Gather);
 432
 433                 /*
 434                  * If there are any initPlans attached to the formerly-top plan node,
 435                  * move them up to the Gather node; same as we do for Material node in
 436                  * materialize_finished_plan.
 437                  */
 438                 gather->plan.initPlan = top_plan->initPlan;
 439                 top_plan->initPlan = NIL;
 440
 441                 gather->plan.targetlist = top_plan->targetlist;
 442                 gather->plan.qual = NIL;
 443                 gather->plan.lefttree = top_plan;
 444                 gather->plan.righttree = NULL;
 445                 gather->num_workers = 1;
 446                 gather->single_copy = true;
 447                 gather->invisible = (force_parallel_mode == FORCE_PARALLEL_REGRESS);
 448
 449                 /*
 450                  * Since this Gather has no parallel-aware descendants to signal to,
 451                  * we don't need a rescan Param.
 452                  */
 453                 gather->rescan_param = -1;
 454
 455                 /*
 456                  * Ideally we'd use cost_gather here, but setting up dummy path data
 457                  * to satisfy it doesn't seem much cleaner than knowing what it does.
 458                  */
 459                 gather->plan.startup_cost = top_plan->startup_cost +
 460                         parallel_setup_cost;
 461                 gather->plan.total_cost = top_plan->total_cost +
 462                         parallel_setup_cost + parallel_tuple_cost * top_plan->plan_rows;
 463                 gather->plan.plan_rows = top_plan->plan_rows;
 464                 gather->plan.plan_width = top_plan->plan_width;
 465                 gather->plan.parallel_aware = false;
 466                 gather->plan.parallel_safe = false;
 467
 468                 /* use parallel mode for parallel plans. */
 469                 root->glob->parallelModeNeeded = true;
 470
 471                 top_plan = &gather->plan;
 472         }
 473
 474         /*
 475          * If any Params were generated, run through the plan tree and compute
 476          * each plan node's extParam/allParam sets.  Ideally we'd merge this into
 477          * set_plan_references' tree traversal, but for now it has to be separate
 478          * because we need to visit subplans before not after main plan.
 479          */
 480         if (glob->paramExecTypes != NIL)
 481         {
 482                 Assert(list_length(glob->subplans) == list_length(glob->subroots));
 483                 forboth(lp, glob->subplans, lr, glob->subroots)
 484                 {
 485                         Plan       *subplan = (Plan *) lfirst(lp);
 486                         PlannerInfo *subroot = lfirst_node(PlannerInfo, lr);
 487
 488                         SS_finalize_plan(subroot, subplan);
 489                 }
 490                 SS_finalize_plan(root, top_plan);
 491         }
 492
 493         /* final cleanup of the plan */
 494         Assert(glob->finalrtable == NIL);
 495         Assert(glob->finalrowmarks == NIL);
 496         Assert(glob->resultRelations == NIL);
 497         Assert(glob->rootResultRelations == NIL);
 498         top_plan = set_plan_references(root, top_plan);
 499         /* ... and the subplans (both regular subplans and initplans) */
 500         Assert(list_length(glob->subplans) == list_length(glob->subroots));
 501         forboth(lp, glob->subplans, lr, glob->subroots)
 502         {
 503                 Plan       *subplan = (Plan *) lfirst(lp);
 504                 PlannerInfo *subroot = lfirst_node(PlannerInfo, lr);
 505
 506                 lfirst(lp) = set_plan_references(subroot, subplan);
 507         }
 508
 509         /* build the PlannedStmt result */
 510         result = makeNode(PlannedStmt);
 511
 512         result->commandType = parse->commandType;
 513         result->queryId = parse->queryId;
 514         result->hasReturning = (parse->returningList != NIL);
 515         result->hasModifyingCTE = parse->hasModifyingCTE;
 516         result->canSetTag = parse->canSetTag;
 517         result->transientPlan = glob->transientPlan;
 518         result->dependsOnRole = glob->dependsOnRole;
 519         result->parallelModeNeeded = glob->parallelModeNeeded;
 520         result->planTree = top_plan;
 521         result->rtable = glob->finalrtable;
 522         result->resultRelations = glob->resultRelations;
 523         result->rootResultRelations = glob->rootResultRelations;
 524         result->subplans = glob->subplans;
 525         result->rewindPlanIDs = glob->rewindPlanIDs;
 526         result->rowMarks = glob->finalrowmarks;
 527         result->relationOids = glob->relationOids;
 528         result->invalItems = glob->invalItems;
 529         result->paramExecTypes = glob->paramExecTypes;
 530         /* utilityStmt should be null, but we might as well copy it */
 531         result->utilityStmt = parse->utilityStmt;
 532         result->stmt_location = parse->stmt_location;
 533         result->stmt_len = parse->stmt_len;
 534
 535         result->jitFlags = PGJIT_NONE;
 536         if (jit_enabled && jit_above_cost >= 0 &&
 537                 top_plan->total_cost > jit_above_cost)
 538         {
 539                 result->jitFlags |= PGJIT_PERFORM;
 540
 541                 /*
 542                  * Decide how much effort should be put into generating better code.
 543                  */
 544                 if (jit_optimize_above_cost >= 0 &&
 545                         top_plan->total_cost > jit_optimize_above_cost)
 546                         result->jitFlags |= PGJIT_OPT3;
 547                 if (jit_inline_above_cost >= 0 &&
 548                         top_plan->total_cost > jit_inline_above_cost)
 549                         result->jitFlags |= PGJIT_INLINE;
 550
 551                 /*
 552                  * Decide which operations should be JITed.
 553                  */
 554                 if (jit_expressions)
 555                         result->jitFlags |= PGJIT_EXPR;
 556                 if (jit_tuple_deforming)
 557                         result->jitFlags |= PGJIT_DEFORM;
 558         }
 559
 560         if (glob->partition_directory != NULL)
 561                 DestroyPartitionDirectory(glob->partition_directory);
 562
 563         return result;
 564 }
 565
 566
 567 /*--------------------
 568  * subquery_planner
 569  *        Invokes the planner on a subquery.  We recurse to here for each
 570  *        sub-SELECT found in the query tree.
 571  *
 572  * glob is the global state for the current planner run.
 573  * parse is the querytree produced by the parser & rewriter.
 574  * parent_root is the immediate parent Query's info (NULL at the top level).
 575  * hasRecursion is true if this is a recursive WITH query.
 576  * tuple_fraction is the fraction of tuples we expect will be retrieved.
 577  * tuple_fraction is interpreted as explained for grouping_planner, below.
 578  *
 579  * Basically, this routine does the stuff that should only be done once
 580  * per Query object.  It then calls grouping_planner.  At one time,
 581  * grouping_planner could be invoked recursively on the same Query object;
 582  * that's not currently true, but we keep the separation between the two
 583  * routines anyway, in case we need it again someday.
 584  *
 585  * subquery_planner will be called recursively to handle sub-Query nodes
 586  * found within the query's expressions and rangetable.
 587  *
 588  * Returns the PlannerInfo struct ("root") that contains all data generated
 589  * while planning the subquery.  In particular, the Path(s) attached to
 590  * the (UPPERREL_FINAL, NULL) upperrel represent our conclusions about the
 591  * cheapest way(s) to implement the query.  The top level will select the
 592  * best Path and pass it through createplan.c to produce a finished Plan.
 593  *--------------------
 594  */
 595 PlannerInfo *
 596 subquery_planner(PlannerGlobal *glob, Query *parse,
 597                                  PlannerInfo *parent_root,
 598                                  bool hasRecursion, double tuple_fraction)
 599 {
 600         PlannerInfo *root;
 601         List       *newWithCheckOptions;
 602         List       *newHaving;
 603         bool            hasOuterJoins;
 604         bool            hasResultRTEs;
 605         RelOptInfo *final_rel;
 606         ListCell   *l;
 607
 608         /* Create a PlannerInfo data structure for this subquery */
 609         root = makeNode(PlannerInfo);
 610         root->parse = parse;
 611         root->glob = glob;
 612         root->query_level = parent_root ? parent_root->query_level + 1 : 1;
 613         root->parent_root = parent_root;
 614         root->plan_params = NIL;
 615         root->outer_params = NULL;
 616         root->planner_cxt = CurrentMemoryContext;
 617         root->init_plans = NIL;
 618         root->cte_plan_ids = NIL;
 619         root->multiexpr_params = NIL;
 620         root->eq_classes = NIL;
 621         root->ec_merging_done = false;
 622         root->append_rel_list = NIL;
 623         root->rowMarks = NIL;
 624         memset(root->upper_rels, 0, sizeof(root->upper_rels));
 625         memset(root->upper_targets, 0, sizeof(root->upper_targets));
 626         root->processed_tlist = NIL;
 627         root->grouping_map = NULL;
 628         root->minmax_aggs = NIL;
 629         root->qual_security_level = 0;
 630         root->inhTargetKind = INHKIND_NONE;
 631         root->hasRecursion = hasRecursion;
 632         if (hasRecursion)
 633                 root->wt_param_id = assign_special_exec_param(root);
 634         else
 635                 root->wt_param_id = -1;
 636         root->non_recursive_path = NULL;
 637         root->partColsUpdated = false;
 638
 639         /*
 640          * If there is a WITH list, process each WITH query and either convert it
 641          * to RTE_SUBQUERY RTE(s) or build an initplan SubPlan structure for it.
 642          */
 643         if (parse->cteList)
 644                 SS_process_ctes(root);
 645
 646         /*
 647          * If the FROM clause is empty, replace it with a dummy RTE_RESULT RTE, so
 648          * that we don't need so many special cases to deal with that situation.
 649          */
 650         replace_empty_jointree(parse);
 651
 652         /*
 653          * Look for ANY and EXISTS SubLinks in WHERE and JOIN/ON clauses, and try
 654          * to transform them into joins.  Note that this step does not descend
 655          * into subqueries; if we pull up any subqueries below, their SubLinks are
 656          * processed just before pulling them up.
 657          */
 658         if (parse->hasSubLinks)
 659                 pull_up_sublinks(root);
 660
 661         /*
 662          * Scan the rangetable for function RTEs, do const-simplification on them,
 663          * and then inline them if possible (producing subqueries that might get
 664          * pulled up next).  Recursion issues here are handled in the same way as
 665          * for SubLinks.
 666          */
 667         preprocess_function_rtes(root);
 668
 669         /*
 670          * Check to see if any subqueries in the jointree can be merged into this
 671          * query.
 672          */
 673         pull_up_subqueries(root);
 674
 675         /*
 676          * If this is a simple UNION ALL query, flatten it into an appendrel. We
 677          * do this now because it requires applying pull_up_subqueries to the leaf
 678          * queries of the UNION ALL, which weren't touched above because they
 679          * weren't referenced by the jointree (they will be after we do this).
 680          */
 681         if (parse->setOperations)
 682                 flatten_simple_union_all(root);
 683
 684         /*
 685          * Survey the rangetable to see what kinds of entries are present.  We can
 686          * skip some later processing if relevant SQL features are not used; for
 687          * example if there are no JOIN RTEs we can avoid the expense of doing
 688          * flatten_join_alias_vars().  This must be done after we have finished
 689          * adding rangetable entries, of course.  (Note: actually, processing of
 690          * inherited or partitioned rels can cause RTEs for their child tables to
 691          * get added later; but those must all be RTE_RELATION entries, so they
 692          * don't invalidate the conclusions drawn here.)
 693          */
 694         root->hasJoinRTEs = false;
 695         root->hasLateralRTEs = false;
 696         hasOuterJoins = false;
 697         hasResultRTEs = false;
 698         foreach(l, parse->rtable)
 699         {
 700                 RangeTblEntry *rte = lfirst_node(RangeTblEntry, l);
 701
 702                 switch (rte->rtekind)
 703                 {
 704                         case RTE_RELATION:
 705                                 if (rte->inh)
 706                                 {
 707                                         /*
 708                                          * Check to see if the relation actually has any children;
 709                                          * if not, clear the inh flag so we can treat it as a
 710                                          * plain base relation.
 711                                          *
 712                                          * Note: this could give a false-positive result, if the
 713                                          * rel once had children but no longer does.  We used to
 714                                          * be able to clear rte->inh later on when we discovered
 715                                          * that, but no more; we have to handle such cases as
 716                                          * full-fledged inheritance.
 717                                          */
 718                                         rte->inh = has_subclass(rte->relid);
 719                                 }
 720                                 break;
 721                         case RTE_JOIN:
 722                                 root->hasJoinRTEs = true;
 723                                 if (IS_OUTER_JOIN(rte->jointype))
 724                                         hasOuterJoins = true;
 725                                 break;
 726                         case RTE_RESULT:
 727                                 hasResultRTEs = true;
 728                                 break;
 729                         default:
 730                                 /* No work here for other RTE types */
 731                                 break;
 732                 }
 733
 734                 if (rte->lateral)
 735                         root->hasLateralRTEs = true;
 736
 737                 /*
 738                  * We can also determine the maximum security level required for any
 739                  * securityQuals now.  Addition of inheritance-child RTEs won't affect
 740                  * this, because child tables don't have their own securityQuals; see
 741                  * expand_single_inheritance_child().
 742                  */
 743                 if (rte->securityQuals)
 744                         root->qual_security_level = Max(root->qual_security_level,
 745                                                                                         list_length(rte->securityQuals));
 746         }
 747
 748         /*
 749          * Preprocess RowMark information.  We need to do this after subquery
 750          * pullup, so that all base relations are present.
 751          */
 752         preprocess_rowmarks(root);
 753
 754         /*
 755          * Set hasHavingQual to remember if HAVING clause is present.  Needed
 756          * because preprocess_expression will reduce a constant-true condition to
 757          * an empty qual list ... but "HAVING TRUE" is not a semantic no-op.
 758          */
 759         root->hasHavingQual = (parse->havingQual != NULL);
 760
 761         /* Clear this flag; might get set in distribute_qual_to_rels */
 762         root->hasPseudoConstantQuals = false;
 763
 764         /*
 765          * Do expression preprocessing on targetlist and quals, as well as other
 766          * random expressions in the querytree.  Note that we do not need to
 767          * handle sort/group expressions explicitly, because they are actually
 768          * part of the targetlist.
 769          */
 770         parse->targetList = (List *)
 771                 preprocess_expression(root, (Node *) parse->targetList,
 772                                                           EXPRKIND_TARGET);
 773
 774         /* Constant-folding might have removed all set-returning functions */
 775         if (parse->hasTargetSRFs)
 776                 parse->hasTargetSRFs = expression_returns_set((Node *) parse->targetList);
 777
 778         newWithCheckOptions = NIL;
 779         foreach(l, parse->withCheckOptions)
 780         {
 781                 WithCheckOption *wco = lfirst_node(WithCheckOption, l);
 782
 783                 wco->qual = preprocess_expression(root, wco->qual,
 784                                                                                   EXPRKIND_QUAL);
 785                 if (wco->qual != NULL)
 786                         newWithCheckOptions = lappend(newWithCheckOptions, wco);
 787         }
 788         parse->withCheckOptions = newWithCheckOptions;
 789
 790         parse->returningList = (List *)
 791                 preprocess_expression(root, (Node *) parse->returningList,
 792                                                           EXPRKIND_TARGET);
 793
 794         preprocess_qual_conditions(root, (Node *) parse->jointree);
 795
 796         parse->havingQual = preprocess_expression(root, parse->havingQual,
 797                                                                                           EXPRKIND_QUAL);
 798
 799         foreach(l, parse->windowClause)
 800         {
 801                 WindowClause *wc = lfirst_node(WindowClause, l);
 802
 803                 /* partitionClause/orderClause are sort/group expressions */
 804                 wc->startOffset = preprocess_expression(root, wc->startOffset,
 805                                                                                                 EXPRKIND_LIMIT);
 806                 wc->endOffset = preprocess_expression(root, wc->endOffset,
 807                                                                                           EXPRKIND_LIMIT);
 808         }
 809
 810         parse->limitOffset = preprocess_expression(root, parse->limitOffset,
 811                                                                                            EXPRKIND_LIMIT);
 812         parse->limitCount = preprocess_expression(root, parse->limitCount,
 813                                                                                           EXPRKIND_LIMIT);
 814
 815         if (parse->onConflict)
 816         {
 817                 parse->onConflict->arbiterElems = (List *)
 818                         preprocess_expression(root,
 819                                                                   (Node *) parse->onConflict->arbiterElems,
 820                                                                   EXPRKIND_ARBITER_ELEM);
 821                 parse->onConflict->arbiterWhere =
 822                         preprocess_expression(root,
 823                                                                   parse->onConflict->arbiterWhere,
 824                                                                   EXPRKIND_QUAL);
 825                 parse->onConflict->onConflictSet = (List *)
 826                         preprocess_expression(root,
 827                                                                   (Node *) parse->onConflict->onConflictSet,
 828                                                                   EXPRKIND_TARGET);
 829                 parse->onConflict->onConflictWhere =
 830                         preprocess_expression(root,
 831                                                                   parse->onConflict->onConflictWhere,
 832                                                                   EXPRKIND_QUAL);
 833                 /* exclRelTlist contains only Vars, so no preprocessing needed */
 834         }
 835
 836         root->append_rel_list = (List *)
 837                 preprocess_expression(root, (Node *) root->append_rel_list,
 838                                                           EXPRKIND_APPINFO);
 839
 840         /* Also need to preprocess expressions within RTEs */
 841         foreach(l, parse->rtable)
 842         {
 843                 RangeTblEntry *rte = lfirst_node(RangeTblEntry, l);
 844                 int                     kind;
 845                 ListCell   *lcsq;
 846
 847                 if (rte->rtekind == RTE_RELATION)
 848                 {
 849                         if (rte->tablesample)
 850                                 rte->tablesample = (TableSampleClause *)
 851                                         preprocess_expression(root,
 852                                                                                   (Node *) rte->tablesample,
 853                                                                                   EXPRKIND_TABLESAMPLE);
 854                 }
 855                 else if (rte->rtekind == RTE_SUBQUERY)
 856                 {
 857                         /*
 858                          * We don't want to do all preprocessing yet on the subquery's
 859                          * expressions, since that will happen when we plan it.  But if it
 860                          * contains any join aliases of our level, those have to get
 861                          * expanded now, because planning of the subquery won't do it.
 862                          * That's only possible if the subquery is LATERAL.
 863                          */
 864                         if (rte->lateral && root->hasJoinRTEs)
 865                                 rte->subquery = (Query *)
 866                                         flatten_join_alias_vars(root->parse,
 867                                                                                         (Node *) rte->subquery);
 868                 }
 869                 else if (rte->rtekind == RTE_FUNCTION)
 870                 {
 871                         /* Preprocess the function expression(s) fully */
 872                         kind = rte->lateral ? EXPRKIND_RTFUNC_LATERAL : EXPRKIND_RTFUNC;
 873                         rte->functions = (List *)
 874                                 preprocess_expression(root, (Node *) rte->functions, kind);
 875                 }
 876                 else if (rte->rtekind == RTE_TABLEFUNC)
 877                 {
 878                         /* Preprocess the function expression(s) fully */
 879                         kind = rte->lateral ? EXPRKIND_TABLEFUNC_LATERAL : EXPRKIND_TABLEFUNC;
 880                         rte->tablefunc = (TableFunc *)
 881                                 preprocess_expression(root, (Node *) rte->tablefunc, kind);
 882                 }
 883                 else if (rte->rtekind == RTE_VALUES)
 884                 {
 885                         /* Preprocess the values lists fully */
 886                         kind = rte->lateral ? EXPRKIND_VALUES_LATERAL : EXPRKIND_VALUES;
 887                         rte->values_lists = (List *)
 888                                 preprocess_expression(root, (Node *) rte->values_lists, kind);
 889                 }
 890
 891                 /*
 892                  * Process each element of the securityQuals list as if it were a
 893                  * separate qual expression (as indeed it is).  We need to do it this
 894                  * way to get proper canonicalization of AND/OR structure.  Note that
 895                  * this converts each element into an implicit-AND sublist.
 896                  */
 897                 foreach(lcsq, rte->securityQuals)
 898                 {
 899                         lfirst(lcsq) = preprocess_expression(root,
 900                                                                                                  (Node *) lfirst(lcsq),
 901                                                                                                  EXPRKIND_QUAL);
 902                 }
 903         }
 904
 905         /*
 906          * Now that we are done preprocessing expressions, and in particular done
 907          * flattening join alias variables, get rid of the joinaliasvars lists.
 908          * They no longer match what expressions in the rest of the tree look
 909          * like, because we have not preprocessed expressions in those lists (and
 910          * do not want to; for example, expanding a SubLink there would result in
 911          * a useless unreferenced subplan).  Leaving them in place simply creates
 912          * a hazard for later scans of the tree.  We could try to prevent that by
 913          * using QTW_IGNORE_JOINALIASES in every tree scan done after this point,
 914          * but that doesn't sound very reliable.
 915          */
 916         if (root->hasJoinRTEs)
 917         {
 918                 foreach(l, parse->rtable)
 919                 {
 920                         RangeTblEntry *rte = lfirst_node(RangeTblEntry, l);
 921
 922                         rte->joinaliasvars = NIL;
 923                 }
 924         }
 925
 926         /*
 927          * In some cases we may want to transfer a HAVING clause into WHERE. We
 928          * cannot do so if the HAVING clause contains aggregates (obviously) or
 929          * volatile functions (since a HAVING clause is supposed to be executed
 930          * only once per group).  We also can't do this if there are any nonempty
 931          * grouping sets; moving such a clause into WHERE would potentially change
 932          * the results, if any referenced column isn't present in all the grouping
 933          * sets.  (If there are only empty grouping sets, then the HAVING clause
 934          * must be degenerate as discussed below.)
 935          *
 936          * Also, it may be that the clause is so expensive to execute that we're
 937          * better off doing it only once per group, despite the loss of
 938          * selectivity.  This is hard to estimate short of doing the entire
 939          * planning process twice, so we use a heuristic: clauses containing
 940          * subplans are left in HAVING.  Otherwise, we move or copy the HAVING
 941          * clause into WHERE, in hopes of eliminating tuples before aggregation
 942          * instead of after.
 943          *
 944          * If the query has explicit grouping then we can simply move such a
 945          * clause into WHERE; any group that fails the clause will not be in the
 946          * output because none of its tuples will reach the grouping or
 947          * aggregation stage.  Otherwise we must have a degenerate (variable-free)
 948          * HAVING clause, which we put in WHERE so that query_planner() can use it
 949          * in a gating Result node, but also keep in HAVING to ensure that we
 950          * don't emit a bogus aggregated row. (This could be done better, but it
 951          * seems not worth optimizing.)
 952          *
 953          * Note that both havingQual and parse->jointree->quals are in
 954          * implicitly-ANDed-list form at this point, even though they are declared
 955          * as Node *.
 956          */
 957         newHaving = NIL;
 958         foreach(l, (List *) parse->havingQual)
 959         {
 960                 Node       *havingclause = (Node *) lfirst(l);
 961
 962                 if ((parse->groupClause && parse->groupingSets) ||
 963                         contain_agg_clause(havingclause) ||
 964                         contain_volatile_functions(havingclause) ||
 965                         contain_subplans(havingclause))
 966                 {
 967                         /* keep it in HAVING */
 968                         newHaving = lappend(newHaving, havingclause);
 969                 }
 970                 else if (parse->groupClause && !parse->groupingSets)
 971                 {
 972                         /* move it to WHERE */
 973                         parse->jointree->quals = (Node *)
 974                                 lappend((List *) parse->jointree->quals, havingclause);
 975                 }
 976                 else
 977                 {
 978                         /* put a copy in WHERE, keep it in HAVING */
 979                         parse->jointree->quals = (Node *)
 980                                 lappend((List *) parse->jointree->quals,
 981                                                 copyObject(havingclause));
 982                         newHaving = lappend(newHaving, havingclause);
 983                 }
 984         }
 985         parse->havingQual = (Node *) newHaving;
 986
 987         /* Remove any redundant GROUP BY columns */
 988         remove_useless_groupby_columns(root);
 989
 990         /*
 991          * If we have any outer joins, try to reduce them to plain inner joins.
 992          * This step is most easily done after we've done expression
 993          * preprocessing.
 994          */
 995         if (hasOuterJoins)
 996                 reduce_outer_joins(root);
 997
 998         /*
 999          * If we have any RTE_RESULT relations, see if they can be deleted from
1000          * the jointree.  This step is most effectively done after we've done
1001          * expression preprocessing and outer join reduction.
1002          */
1003         if (hasResultRTEs)
1004                 remove_useless_result_rtes(root);
1005
1006         /*
1007          * Do the main planning.  If we have an inherited target relation, that
1008          * needs special processing, else go straight to grouping_planner.
1009          */
1010         if (parse->resultRelation &&
1011                 rt_fetch(parse->resultRelation, parse->rtable)->inh)
1012                 inheritance_planner(root);
1013         else
1014                 grouping_planner(root, false, tuple_fraction);
1015
1016         /*
1017          * Capture the set of outer-level param IDs we have access to, for use in
1018          * extParam/allParam calculations later.
1019          */
1020         SS_identify_outer_params(root);
1021
1022         /*
1023          * If any initPlans were created in this query level, adjust the surviving
1024          * Paths' costs and parallel-safety flags to account for them.  The
1025          * initPlans won't actually get attached to the plan tree till
1026          * create_plan() runs, but we must include their effects now.
1027          */
1028         final_rel = fetch_upper_rel(root, UPPERREL_FINAL, NULL);
1029         SS_charge_for_initplans(root, final_rel);
1030
1031         /*
1032          * Make sure we've identified the cheapest Path for the final rel.  (By
1033          * doing this here not in grouping_planner, we include initPlan costs in
1034          * the decision, though it's unlikely that will change anything.)
1035          */
1036         set_cheapest(final_rel);
1037
1038         return root;
1039 }
1040
1041 /*
1042  * preprocess_expression
1043  *              Do subquery_planner's preprocessing work for an expression,
1044  *              which can be a targetlist, a WHERE clause (including JOIN/ON
1045  *              conditions), a HAVING clause, or a few other things.
1046  */
1047 static Node *
1048 preprocess_expression(PlannerInfo *root, Node *expr, int kind)
1049 {
1050         /*
1051          * Fall out quickly if expression is empty.  This occurs often enough to
1052          * be worth checking.  Note that null->null is the correct conversion for
1053          * implicit-AND result format, too.
1054          */
1055         if (expr == NULL)
1056                 return NULL;
1057
1058         /*
1059          * If the query has any join RTEs, replace join alias variables with
1060          * base-relation variables.  We must do this first, since any expressions
1061          * we may extract from the joinaliasvars lists have not been preprocessed.
1062          * For example, if we did this after sublink processing, sublinks expanded
1063          * out from join aliases would not get processed.  But we can skip this in
1064          * non-lateral RTE functions, VALUES lists, and TABLESAMPLE clauses, since
1065          * they can't contain any Vars of the current query level.
1066          */
1067         if (root->hasJoinRTEs &&
1068                 !(kind == EXPRKIND_RTFUNC ||
1069                   kind == EXPRKIND_VALUES ||
1070                   kind == EXPRKIND_TABLESAMPLE ||
1071                   kind == EXPRKIND_TABLEFUNC))
1072                 expr = flatten_join_alias_vars(root->parse, expr);
1073
1074         /*
1075          * Simplify constant expressions.  For function RTEs, this was already
1076          * done by preprocess_function_rtes ... but we have to do it again if the
1077          * RTE is LATERAL and might have contained join alias variables.
1078          *
1079          * Note: an essential effect of this is to convert named-argument function
1080          * calls to positional notation and insert the current actual values of
1081          * any default arguments for functions.  To ensure that happens, we *must*
1082          * process all expressions here.  Previous PG versions sometimes skipped
1083          * const-simplification if it didn't seem worth the trouble, but we can't
1084          * do that anymore.
1085          *
1086          * Note: this also flattens nested AND and OR expressions into N-argument
1087          * form.  All processing of a qual expression after this point must be
1088          * careful to maintain AND/OR flatness --- that is, do not generate a tree
1089          * with AND directly under AND, nor OR directly under OR.
1090          */
1091         if (!(kind == EXPRKIND_RTFUNC ||
1092                   (kind == EXPRKIND_RTFUNC_LATERAL && !root->hasJoinRTEs)))
1093                 expr = eval_const_expressions(root, expr);
1094
1095         /*
1096          * If it's a qual or havingQual, canonicalize it.
1097          */
1098         if (kind == EXPRKIND_QUAL)
1099         {
1100                 expr = (Node *) canonicalize_qual((Expr *) expr, false);
1101
1102 #ifdef OPTIMIZER_DEBUG
1103                 printf("After canonicalize_qual()\n");
1104                 pprint(expr);
1105 #endif
1106         }
1107
1108         /* Expand SubLinks to SubPlans */
1109         if (root->parse->hasSubLinks)
1110                 expr = SS_process_sublinks(root, expr, (kind == EXPRKIND_QUAL));
1111
1112         /*
1113          * XXX do not insert anything here unless you have grokked the comments in
1114          * SS_replace_correlation_vars ...
1115          */
1116
1117         /* Replace uplevel vars with Param nodes (this IS possible in VALUES) */
1118         if (root->query_level > 1)
1119                 expr = SS_replace_correlation_vars(root, expr);
1120
1121         /*
1122          * If it's a qual or havingQual, convert it to implicit-AND format. (We
1123          * don't want to do this before eval_const_expressions, since the latter
1124          * would be unable to simplify a top-level AND correctly. Also,
1125          * SS_process_sublinks expects explicit-AND format.)
1126          */
1127         if (kind == EXPRKIND_QUAL)
1128                 expr = (Node *) make_ands_implicit((Expr *) expr);
1129
1130         return expr;
1131 }
1132
1133 /*
1134  * preprocess_qual_conditions
1135  *              Recursively scan the query's jointree and do subquery_planner's
1136  *              preprocessing work on each qual condition found therein.
1137  */
1138 static void
1139 preprocess_qual_conditions(PlannerInfo *root, Node *jtnode)
1140 {
1141         if (jtnode == NULL)
1142                 return;
1143         if (IsA(jtnode, RangeTblRef))
1144         {
1145                 /* nothing to do here */
1146         }
1147         else if (IsA(jtnode, FromExpr))
1148         {
1149                 FromExpr   *f = (FromExpr *) jtnode;
1150                 ListCell   *l;
1151
1152                 foreach(l, f->fromlist)
1153                         preprocess_qual_conditions(root, lfirst(l));
1154
1155                 f->quals = preprocess_expression(root, f->quals, EXPRKIND_QUAL);
1156         }
1157         else if (IsA(jtnode, JoinExpr))
1158         {
1159                 JoinExpr   *j = (JoinExpr *) jtnode;
1160
1161                 preprocess_qual_conditions(root, j->larg);
1162                 preprocess_qual_conditions(root, j->rarg);
1163
1164                 j->quals = preprocess_expression(root, j->quals, EXPRKIND_QUAL);
1165         }
1166         else
1167                 elog(ERROR, "unrecognized node type: %d",
1168                          (int) nodeTag(jtnode));
1169 }
1170
1171 /*
1172  * preprocess_phv_expression
1173  *        Do preprocessing on a PlaceHolderVar expression that's been pulled up.
1174  *
1175  * If a LATERAL subquery references an output of another subquery, and that
1176  * output must be wrapped in a PlaceHolderVar because of an intermediate outer
1177  * join, then we'll push the PlaceHolderVar expression down into the subquery
1178  * and later pull it back up during find_lateral_references, which runs after
1179  * subquery_planner has preprocessed all the expressions that were in the
1180  * current query level to start with.  So we need to preprocess it then.
1181  */
1182 Expr *
1183 preprocess_phv_expression(PlannerInfo *root, Expr *expr)
1184 {
1185         return (Expr *) preprocess_expression(root, (Node *) expr, EXPRKIND_PHV);
1186 }
1187
1188 /*
1189  * inheritance_planner
1190  *        Generate Paths in the case where the result relation is an
1191  *        inheritance set.
1192  *
1193  * We have to handle this case differently from cases where a source relation
1194  * is an inheritance set. Source inheritance is expanded at the bottom of the
1195  * plan tree (see allpaths.c), but target inheritance has to be expanded at
1196  * the top.  The reason is that for UPDATE, each target relation needs a
1197  * different targetlist matching its own column set.  Fortunately,
1198  * the UPDATE/DELETE target can never be the nullable side of an outer join,
1199  * so it's OK to generate the plan this way.
1200  *
1201  * Returns nothing; the useful output is in the Paths we attach to
1202  * the (UPPERREL_FINAL, NULL) upperrel stored in *root.
1203  *
1204  * Note that we have not done set_cheapest() on the final rel; it's convenient
1205  * to leave this to the caller.
1206  */
1207 static void
1208 inheritance_planner(PlannerInfo *root)
1209 {
1210         Query      *parse = root->parse;
1211         int                     top_parentRTindex = parse->resultRelation;
1212         List       *select_rtable;
1213         List       *select_appinfos;
1214         List       *child_appinfos;
1215         List       *old_child_rtis;
1216         List       *new_child_rtis;
1217         Bitmapset  *subqueryRTindexes;
1218         Index           next_subquery_rti;
1219         int                     nominalRelation = -1;
1220         Index           rootRelation = 0;
1221         List       *final_rtable = NIL;
1222         List       *final_rowmarks = NIL;
1223         int                     save_rel_array_size = 0;
1224         RelOptInfo **save_rel_array = NULL;
1225         AppendRelInfo **save_append_rel_array = NULL;
1226         List       *subpaths = NIL;
1227         List       *subroots = NIL;
1228         List       *resultRelations = NIL;
1229         List       *withCheckOptionLists = NIL;
1230         List       *returningLists = NIL;
1231         List       *rowMarks;
1232         RelOptInfo *final_rel;
1233         ListCell   *lc;
1234         ListCell   *lc2;
1235         Index           rti;
1236         RangeTblEntry *parent_rte;
1237         Bitmapset  *parent_relids;
1238         Query     **parent_parses;
1239
1240         /* Should only get here for UPDATE or DELETE */
1241         Assert(parse->commandType == CMD_UPDATE ||
1242                    parse->commandType == CMD_DELETE);
1243
1244         /*
1245          * We generate a modified instance of the original Query for each target
1246          * relation, plan that, and put all the plans into a list that will be
1247          * controlled by a single ModifyTable node.  All the instances share the
1248          * same rangetable, but each instance must have its own set of subquery
1249          * RTEs within the finished rangetable because (1) they are likely to get
1250          * scribbled on during planning, and (2) it's not inconceivable that
1251          * subqueries could get planned differently in different cases.  We need
1252          * not create duplicate copies of other RTE kinds, in particular not the
1253          * target relations, because they don't have either of those issues.  Not
1254          * having to duplicate the target relations is important because doing so
1255          * (1) would result in a rangetable of length O(N^2) for N targets, with
1256          * at least O(N^3) work expended here; and (2) would greatly complicate
1257          * management of the rowMarks list.
1258          *
1259          * To begin with, generate a bitmapset of the relids of the subquery RTEs.
1260          */
1261         subqueryRTindexes = NULL;
1262         rti = 1;
1263         foreach(lc, parse->rtable)
1264         {
1265                 RangeTblEntry *rte = lfirst_node(RangeTblEntry, lc);
1266
1267                 if (rte->rtekind == RTE_SUBQUERY)
1268                         subqueryRTindexes = bms_add_member(subqueryRTindexes, rti);
1269                 rti++;
1270         }
1271
1272         /*
1273          * If the parent RTE is a partitioned table, we should use that as the
1274          * nominal target relation, because the RTEs added for partitioned tables
1275          * (including the root parent) as child members of the inheritance set do
1276          * not appear anywhere else in the plan, so the confusion explained below
1277          * for non-partitioning inheritance cases is not possible.
1278          */
1279         parent_rte = rt_fetch(top_parentRTindex, parse->rtable);
1280         Assert(parent_rte->inh);
1281         if (parent_rte->relkind == RELKIND_PARTITIONED_TABLE)
1282         {
1283                 nominalRelation = top_parentRTindex;
1284                 rootRelation = top_parentRTindex;
1285         }
1286
1287         /*
1288          * Before generating the real per-child-relation plans, do a cycle of
1289          * planning as though the query were a SELECT.  The objective here is to
1290          * find out which child relations need to be processed, using the same
1291          * expansion and pruning logic as for a SELECT.  We'll then pull out the
1292          * RangeTblEntry-s generated for the child rels, and make use of the
1293          * AppendRelInfo entries for them to guide the real planning.  (This is
1294          * rather inefficient; we could perhaps stop short of making a full Path
1295          * tree.  But this whole function is inefficient and slated for
1296          * destruction, so let's not contort query_planner for that.)
1297          */
1298         {
1299                 PlannerInfo *subroot;
1300
1301                 /*
1302                  * Flat-copy the PlannerInfo to prevent modification of the original.
1303                  */
1304                 subroot = makeNode(PlannerInfo);
1305                 memcpy(subroot, root, sizeof(PlannerInfo));
1306
1307                 /*
1308                  * Make a deep copy of the parsetree for this planning cycle to mess
1309                  * around with, and change it to look like a SELECT.  (Hack alert: the
1310                  * target RTE still has updatedCols set if this is an UPDATE, so that
1311                  * expand_partitioned_rtentry will correctly update
1312                  * subroot->partColsUpdated.)
1313                  */
1314                 subroot->parse = copyObject(root->parse);
1315
1316                 subroot->parse->commandType = CMD_SELECT;
1317                 subroot->parse->resultRelation = 0;
1318
1319                 /*
1320                  * Ensure the subroot has its own copy of the original
1321                  * append_rel_list, since it'll be scribbled on.  (Note that at this
1322                  * point, the list only contains AppendRelInfos for flattened UNION
1323                  * ALL subqueries.)
1324                  */
1325                 subroot->append_rel_list = copyObject(root->append_rel_list);
1326
1327                 /*
1328                  * Better make a private copy of the rowMarks, too.
1329                  */
1330                 subroot->rowMarks = copyObject(root->rowMarks);
1331
1332                 /* There shouldn't be any OJ info to translate, as yet */
1333                 Assert(subroot->join_info_list == NIL);
1334                 /* and we haven't created PlaceHolderInfos, either */
1335                 Assert(subroot->placeholder_list == NIL);
1336
1337                 /* Generate Path(s) for accessing this result relation */
1338                 grouping_planner(subroot, true, 0.0 /* retrieve all tuples */ );
1339
1340                 /* Extract the info we need. */
1341                 select_rtable = subroot->parse->rtable;
1342                 select_appinfos = subroot->append_rel_list;
1343
1344                 /*
1345                  * We need to propagate partColsUpdated back, too.  (The later
1346                  * planning cycles will not set this because they won't run
1347                  * expand_partitioned_rtentry for the UPDATE target.)
1348                  */
1349                 root->partColsUpdated = subroot->partColsUpdated;
1350         }
1351
1352         /*----------
1353          * Since only one rangetable can exist in the final plan, we need to make
1354          * sure that it contains all the RTEs needed for any child plan.  This is
1355          * complicated by the need to use separate subquery RTEs for each child.
1356          * We arrange the final rtable as follows:
1357          * 1. All original rtable entries (with their original RT indexes).
1358          * 2. All the relation RTEs generated for children of the target table.
1359          * 3. Subquery RTEs for children after the first.  We need N * (K - 1)
1360          *    RT slots for this, if there are N subqueries and K child tables.
1361          * 4. Additional RTEs generated during the child planning runs, such as
1362          *    children of inheritable RTEs other than the target table.
1363          * We assume that each child planning run will create an identical set
1364          * of type-4 RTEs.
1365          *
1366          * So the next thing to do is append the type-2 RTEs (the target table's
1367          * children) to the original rtable.  We look through select_appinfos
1368          * to find them.
1369          *
1370          * To identify which AppendRelInfos are relevant as we thumb through
1371          * select_appinfos, we need to look for both direct and indirect children
1372          * of top_parentRTindex, so we use a bitmap of known parent relids.
1373          * expand_inherited_rtentry() always processes a parent before any of that
1374          * parent's children, so we should see an intermediate parent before its
1375          * children.
1376          *----------
1377          */
1378         child_appinfos = NIL;
1379         old_child_rtis = NIL;
1380         new_child_rtis = NIL;
1381         parent_relids = bms_make_singleton(top_parentRTindex);
1382         foreach(lc, select_appinfos)
1383         {
1384                 AppendRelInfo *appinfo = lfirst_node(AppendRelInfo, lc);
1385                 RangeTblEntry *child_rte;
1386
1387                 /* append_rel_list contains all append rels; ignore others */
1388                 if (!bms_is_member(appinfo->parent_relid, parent_relids))
1389                         continue;
1390
1391                 /* remember relevant AppendRelInfos for use below */
1392                 child_appinfos = lappend(child_appinfos, appinfo);
1393
1394                 /* extract RTE for this child rel */
1395                 child_rte = rt_fetch(appinfo->child_relid, select_rtable);
1396
1397                 /* and append it to the original rtable */
1398                 parse->rtable = lappend(parse->rtable, child_rte);
1399
1400                 /* remember child's index in the SELECT rtable */
1401                 old_child_rtis = lappend_int(old_child_rtis, appinfo->child_relid);
1402
1403                 /* and its new index in the final rtable */
1404                 new_child_rtis = lappend_int(new_child_rtis, list_length(parse->rtable));
1405
1406                 /* if child is itself partitioned, update parent_relids */
1407                 if (child_rte->inh)
1408                 {
1409                         Assert(child_rte->relkind == RELKIND_PARTITIONED_TABLE);
1410                         parent_relids = bms_add_member(parent_relids, appinfo->child_relid);
1411                 }
1412         }
1413
1414         /*
1415          * It's possible that the RTIs we just assigned for the child rels in the
1416          * final rtable are different from what they were in the SELECT query.
1417          * Adjust the AppendRelInfos so that they will correctly map RT indexes to
1418          * the final indexes.  We can do this left-to-right since no child rel's
1419          * final RT index could be greater than what it had in the SELECT query.
1420          */
1421         forboth(lc, old_child_rtis, lc2, new_child_rtis)
1422         {
1423                 int                     old_child_rti = lfirst_int(lc);
1424                 int                     new_child_rti = lfirst_int(lc2);
1425
1426                 if (old_child_rti == new_child_rti)
1427                         continue;                       /* nothing to do */
1428
1429                 Assert(old_child_rti > new_child_rti);
1430
1431                 ChangeVarNodes((Node *) child_appinfos,
1432                                            old_child_rti, new_child_rti, 0);
1433         }
1434
1435         /*
1436          * Now set up rangetable entries for subqueries for additional children
1437          * (the first child will just use the original ones).  These all have to
1438          * look more or less real, or EXPLAIN will get unhappy; so we just make
1439          * them all clones of the original subqueries.
1440          */
1441         next_subquery_rti = list_length(parse->rtable) + 1;
1442         if (subqueryRTindexes != NULL)
1443         {
1444                 int                     n_children = list_length(child_appinfos);
1445
1446                 while (n_children-- > 1)
1447                 {
1448                         int                     oldrti = -1;
1449
1450                         while ((oldrti = bms_next_member(subqueryRTindexes, oldrti)) >= 0)
1451                         {
1452                                 RangeTblEntry *subqrte;
1453
1454                                 subqrte = rt_fetch(oldrti, parse->rtable);
1455                                 parse->rtable = lappend(parse->rtable, copyObject(subqrte));
1456                         }
1457                 }
1458         }
1459
1460         /*
1461          * The query for each child is obtained by translating the query for its
1462          * immediate parent, since the AppendRelInfo data we have shows deltas
1463          * between parents and children.  We use the parent_parses array to
1464          * remember the appropriate query trees.  This is indexed by parent relid.
1465          * Since the maximum number of parents is limited by the number of RTEs in
1466          * the SELECT query, we use that number to allocate the array.  An extra
1467          * entry is needed since relids start from 1.
1468          */
1469         parent_parses = (Query **) palloc0((list_length(select_rtable) + 1) *
1470                                                                            sizeof(Query *));
1471         parent_parses[top_parentRTindex] = parse;
1472
1473         /*
1474          * And now we can get on with generating a plan for each child table.
1475          */
1476         foreach(lc, child_appinfos)
1477         {
1478                 AppendRelInfo *appinfo = lfirst_node(AppendRelInfo, lc);
1479                 Index           this_subquery_rti = next_subquery_rti;
1480                 Query      *parent_parse;
1481                 PlannerInfo *subroot;
1482                 RangeTblEntry *child_rte;
1483                 RelOptInfo *sub_final_rel;
1484                 Path       *subpath;
1485
1486                 /*
1487                  * expand_inherited_rtentry() always processes a parent before any of
1488                  * that parent's children, so the parent query for this relation
1489                  * should already be available.
1490                  */
1491                 parent_parse = parent_parses[appinfo->parent_relid];
1492                 Assert(parent_parse != NULL);
1493
1494                 /*
1495                  * We need a working copy of the PlannerInfo so that we can control
1496                  * propagation of information back to the main copy.
1497                  */
1498                 subroot = makeNode(PlannerInfo);
1499                 memcpy(subroot, root, sizeof(PlannerInfo));
1500
1501                 /*
1502                  * Generate modified query with this rel as target.  We first apply
1503                  * adjust_appendrel_attrs, which copies the Query and changes
1504                  * references to the parent RTE to refer to the current child RTE,
1505                  * then fool around with subquery RTEs.
1506                  */
1507                 subroot->parse = (Query *)
1508                         adjust_appendrel_attrs(subroot,
1509                                                                    (Node *) parent_parse,
1510                                                                    1, &appinfo);
1511
1512                 /*
1513                  * If there are securityQuals attached to the parent, move them to the
1514                  * child rel (they've already been transformed properly for that).
1515                  */
1516                 parent_rte = rt_fetch(appinfo->parent_relid, subroot->parse->rtable);
1517                 child_rte = rt_fetch(appinfo->child_relid, subroot->parse->rtable);
1518                 child_rte->securityQuals = parent_rte->securityQuals;
1519                 parent_rte->securityQuals = NIL;
1520
1521                 /*
1522                  * HACK: setting this to a value other than INHKIND_NONE signals to
1523                  * relation_excluded_by_constraints() to treat the result relation as
1524                  * being an appendrel member.
1525                  */
1526                 subroot->inhTargetKind =
1527                         (rootRelation != 0) ? INHKIND_PARTITIONED : INHKIND_INHERITED;
1528
1529                 /*
1530                  * If this child is further partitioned, remember it as a parent.
1531                  * Since a partitioned table does not have any data, we don't need to
1532                  * create a plan for it, and we can stop processing it here.  We do,
1533                  * however, need to remember its modified PlannerInfo for use when
1534                  * processing its children, since we'll update their varnos based on
1535                  * the delta from immediate parent to child, not from top to child.
1536                  *
1537                  * Note: a very non-obvious point is that we have not yet added
1538                  * duplicate subquery RTEs to the subroot's rtable.  We mustn't,
1539                  * because then its children would have two sets of duplicates,
1540                  * confusing matters.
1541                  */
1542                 if (child_rte->inh)
1543                 {
1544                         Assert(child_rte->relkind == RELKIND_PARTITIONED_TABLE);
1545                         parent_parses[appinfo->child_relid] = subroot->parse;
1546                         continue;
1547                 }
1548
1549                 /*
1550                  * Set the nominal target relation of the ModifyTable node if not
1551                  * already done.  If the target is a partitioned table, we already set
1552                  * nominalRelation to refer to the partition root, above.  For
1553                  * non-partitioned inheritance cases, we'll use the first child
1554                  * relation (even if it's excluded) as the nominal target relation.
1555                  * Because of the way expand_inherited_rtentry works, that should be
1556                  * the RTE representing the parent table in its role as a simple
1557                  * member of the inheritance set.
1558                  *
1559                  * It would be logically cleaner to *always* use the inheritance
1560                  * parent RTE as the nominal relation; but that RTE is not otherwise
1561                  * referenced in the plan in the non-partitioned inheritance case.
1562                  * Instead the duplicate child RTE created by expand_inherited_rtentry
1563                  * is used elsewhere in the plan, so using the original parent RTE
1564                  * would give rise to confusing use of multiple aliases in EXPLAIN
1565                  * output for what the user will think is the "same" table.  OTOH,
1566                  * it's not a problem in the partitioned inheritance case, because
1567                  * there is no duplicate RTE for the parent.
1568                  */
1569                 if (nominalRelation < 0)
1570                         nominalRelation = appinfo->child_relid;
1571
1572                 /*
1573                  * As above, each child plan run needs its own append_rel_list and
1574                  * rowmarks, which should start out as pristine copies of the
1575                  * originals.  There can't be any references to UPDATE/DELETE target
1576                  * rels in them; but there could be subquery references, which we'll
1577                  * fix up in a moment.
1578                  */
1579                 subroot->append_rel_list = copyObject(root->append_rel_list);
1580                 subroot->rowMarks = copyObject(root->rowMarks);
1581
1582                 /*
1583                  * If this isn't the first child Query, adjust Vars and jointree
1584                  * entries to reference the appropriate set of subquery RTEs.
1585                  */
1586                 if (final_rtable != NIL && subqueryRTindexes != NULL)
1587                 {
1588                         int                     oldrti = -1;
1589
1590                         while ((oldrti = bms_next_member(subqueryRTindexes, oldrti)) >= 0)
1591                         {
1592                                 Index           newrti = next_subquery_rti++;
1593
1594                                 ChangeVarNodes((Node *) subroot->parse, oldrti, newrti, 0);
1595                                 ChangeVarNodes((Node *) subroot->append_rel_list,
1596                                                            oldrti, newrti, 0);
1597                                 ChangeVarNodes((Node *) subroot->rowMarks, oldrti, newrti, 0);
1598                         }
1599                 }
1600
1601                 /* There shouldn't be any OJ info to translate, as yet */
1602                 Assert(subroot->join_info_list == NIL);
1603                 /* and we haven't created PlaceHolderInfos, either */
1604                 Assert(subroot->placeholder_list == NIL);
1605
1606                 /* Generate Path(s) for accessing this result relation */
1607                 grouping_planner(subroot, true, 0.0 /* retrieve all tuples */ );
1608
1609                 /*
1610                  * Select cheapest path in case there's more than one.  We always run
1611                  * modification queries to conclusion, so we care only for the
1612                  * cheapest-total path.
1613                  */
1614                 sub_final_rel = fetch_upper_rel(subroot, UPPERREL_FINAL, NULL);
1615                 set_cheapest(sub_final_rel);
1616                 subpath = sub_final_rel->cheapest_total_path;
1617
1618                 /*
1619                  * If this child rel was excluded by constraint exclusion, exclude it
1620                  * from the result plan.
1621                  */
1622                 if (IS_DUMMY_REL(sub_final_rel))
1623                         continue;
1624
1625                 /*
1626                  * If this is the first non-excluded child, its post-planning rtable
1627                  * becomes the initial contents of final_rtable; otherwise, copy its
1628                  * modified subquery RTEs into final_rtable, to ensure we have sane
1629                  * copies of those.  Also save the first non-excluded child's version
1630                  * of the rowmarks list; we assume all children will end up with
1631                  * equivalent versions of that.
1632                  */
1633                 if (final_rtable == NIL)
1634                 {
1635                         final_rtable = subroot->parse->rtable;
1636                         final_rowmarks = subroot->rowMarks;
1637                 }
1638                 else
1639                 {
1640                         Assert(list_length(final_rtable) ==
1641                                    list_length(subroot->parse->rtable));
1642                         if (subqueryRTindexes != NULL)
1643                         {
1644                                 int                     oldrti = -1;
1645
1646                                 while ((oldrti = bms_next_member(subqueryRTindexes, oldrti)) >= 0)
1647                                 {
1648                                         Index           newrti = this_subquery_rti++;
1649                                         RangeTblEntry *subqrte;
1650                                         ListCell   *newrticell;
1651
1652                                         subqrte = rt_fetch(newrti, subroot->parse->rtable);
1653                                         newrticell = list_nth_cell(final_rtable, newrti - 1);
1654                                         lfirst(newrticell) = subqrte;
1655                                 }
1656                         }
1657                 }
1658
1659                 /*
1660                  * We need to collect all the RelOptInfos from all child plans into
1661                  * the main PlannerInfo, since setrefs.c will need them.  We use the
1662                  * last child's simple_rel_array, so we have to propagate forward the
1663                  * RelOptInfos that were already built in previous children.
1664                  */
1665                 Assert(subroot->simple_rel_array_size >= save_rel_array_size);
1666                 for (rti = 1; rti < save_rel_array_size; rti++)
1667                 {
1668                         RelOptInfo *brel = save_rel_array[rti];
1669
1670                         if (brel)
1671                                 subroot->simple_rel_array[rti] = brel;
1672                 }
1673                 save_rel_array_size = subroot->simple_rel_array_size;
1674                 save_rel_array = subroot->simple_rel_array;
1675                 save_append_rel_array = subroot->append_rel_array;
1676
1677                 /*
1678                  * Make sure any initplans from this rel get into the outer list. Note
1679                  * we're effectively assuming all children generate the same
1680                  * init_plans.
1681                  */
1682                 root->init_plans = subroot->init_plans;
1683
1684                 /* Build list of sub-paths */
1685                 subpaths = lappend(subpaths, subpath);
1686
1687                 /* Build list of modified subroots, too */
1688                 subroots = lappend(subroots, subroot);
1689
1690                 /* Build list of target-relation RT indexes */
1691                 resultRelations = lappend_int(resultRelations, appinfo->child_relid);
1692
1693                 /* Build lists of per-relation WCO and RETURNING targetlists */
1694                 if (parse->withCheckOptions)
1695                         withCheckOptionLists = lappend(withCheckOptionLists,
1696                                                                                    subroot->parse->withCheckOptions);
1697                 if (parse->returningList)
1698                         returningLists = lappend(returningLists,
1699                                                                          subroot->parse->returningList);
1700
1701                 Assert(!parse->onConflict);
1702         }
1703
1704         /* Result path must go into outer query's FINAL upperrel */
1705         final_rel = fetch_upper_rel(root, UPPERREL_FINAL, NULL);
1706
1707         /*
1708          * We don't currently worry about setting final_rel's consider_parallel
1709          * flag in this case, nor about allowing FDWs or create_upper_paths_hook
1710          * to get control here.
1711          */
1712
1713         if (subpaths == NIL)
1714         {
1715                 /*
1716                  * We managed to exclude every child rel, so generate a dummy path
1717                  * representing the empty set.  Although it's clear that no data will
1718                  * be updated or deleted, we will still need to have a ModifyTable
1719                  * node so that any statement triggers are executed.  (This could be
1720                  * cleaner if we fixed nodeModifyTable.c to support zero child nodes,
1721                  * but that probably wouldn't be a net win.)
1722                  */
1723                 Path       *dummy_path;
1724
1725                 /* tlist processing never got done, either */
1726                 root->processed_tlist = preprocess_targetlist(root);
1727                 final_rel->reltarget = create_pathtarget(root, root->processed_tlist);
1728
1729                 /* Make a dummy path, cf set_dummy_rel_pathlist() */
1730                 dummy_path = (Path *) create_append_path(NULL, final_rel, NIL, NIL,
1731                                                                                                  NIL, NULL, 0, false,
1732                                                                                                  NIL, -1);
1733
1734                 /* These lists must be nonempty to make a valid ModifyTable node */
1735                 subpaths = list_make1(dummy_path);
1736                 subroots = list_make1(root);
1737                 resultRelations = list_make1_int(parse->resultRelation);
1738                 if (parse->withCheckOptions)
1739                         withCheckOptionLists = list_make1(parse->withCheckOptions);
1740                 if (parse->returningList)
1741                         returningLists = list_make1(parse->returningList);
1742                 /* Disable tuple routing, too, just to be safe */
1743                 root->partColsUpdated = false;
1744         }
1745         else
1746         {
1747                 /*
1748                  * Put back the final adjusted rtable into the master copy of the
1749                  * Query.  (We mustn't do this if we found no non-excluded children,
1750                  * since we never saved an adjusted rtable at all.)
1751                  */
1752                 parse->rtable = final_rtable;
1753                 root->simple_rel_array_size = save_rel_array_size;
1754                 root->simple_rel_array = save_rel_array;
1755                 root->append_rel_array = save_append_rel_array;
1756
1757                 /* Must reconstruct master's simple_rte_array, too */
1758                 root->simple_rte_array = (RangeTblEntry **)
1759                         palloc0((list_length(final_rtable) + 1) * sizeof(RangeTblEntry *));
1760                 rti = 1;
1761                 foreach(lc, final_rtable)
1762                 {
1763                         RangeTblEntry *rte = lfirst_node(RangeTblEntry, lc);
1764
1765                         root->simple_rte_array[rti++] = rte;
1766                 }
1767
1768                 /* Put back adjusted rowmarks, too */
1769                 root->rowMarks = final_rowmarks;
1770         }
1771
1772         /*
1773          * If there was a FOR [KEY] UPDATE/SHARE clause, the LockRows node will
1774          * have dealt with fetching non-locked marked rows, else we need to have
1775          * ModifyTable do that.
1776          */
1777         if (parse->rowMarks)
1778                 rowMarks = NIL;
1779         else
1780                 rowMarks = root->rowMarks;
1781
1782         /* Create Path representing a ModifyTable to do the UPDATE/DELETE work */
1783         add_path(final_rel, (Path *)
1784                          create_modifytable_path(root, final_rel,
1785                                                                          parse->commandType,
1786                                                                          parse->canSetTag,
1787                                                                          nominalRelation,
1788                                                                          rootRelation,
1789                                                                          root->partColsUpdated,
1790                                                                          resultRelations,
1791                                                                          subpaths,
1792                                                                          subroots,
1793                                                                          withCheckOptionLists,
1794                                                                          returningLists,
1795                                                                          rowMarks,
1796                                                                          NULL,
1797                                                                          assign_special_exec_param(root)));
1798 }
1799
1800 /*--------------------
1801  * grouping_planner
1802  *        Perform planning steps related to grouping, aggregation, etc.
1803  *
1804  * This function adds all required top-level processing to the scan/join
1805  * Path(s) produced by query_planner.
1806  *
1807  * If inheritance_update is true, we're being called from inheritance_planner
1808  * and should not include a ModifyTable step in the resulting Path(s).
1809  * (inheritance_planner will create a single ModifyTable node covering all the
1810  * target tables.)
1811  *
1812  * tuple_fraction is the fraction of tuples we expect will be retrieved.
1813  * tuple_fraction is interpreted as follows:
1814  *        0: expect all tuples to be retrieved (normal case)
1815  *        0 < tuple_fraction < 1: expect the given fraction of tuples available
1816  *              from the plan to be retrieved
1817  *        tuple_fraction >= 1: tuple_fraction is the absolute number of tuples
1818  *              expected to be retrieved (ie, a LIMIT specification)
1819  *
1820  * Returns nothing; the useful output is in the Paths we attach to the
1821  * (UPPERREL_FINAL, NULL) upperrel in *root.  In addition,
1822  * root->processed_tlist contains the final processed targetlist.
1823  *
1824  * Note that we have not done set_cheapest() on the final rel; it's convenient
1825  * to leave this to the caller.
1826  *--------------------
1827  */
1828 static void
1829 grouping_planner(PlannerInfo *root, bool inheritance_update,
1830                                  double tuple_fraction)
1831 {
1832         Query      *parse = root->parse;
1833         int64           offset_est = 0;
1834         int64           count_est = 0;
1835         double          limit_tuples = -1.0;
1836         bool            have_postponed_srfs = false;
1837         PathTarget *final_target;
1838         List       *final_targets;
1839         List       *final_targets_contain_srfs;
1840         bool            final_target_parallel_safe;
1841         RelOptInfo *current_rel;
1842         RelOptInfo *final_rel;
1843         FinalPathExtraData extra;
1844         ListCell   *lc;
1845
1846         /* Tweak caller-supplied tuple_fraction if have LIMIT/OFFSET */
1847         if (parse->limitCount || parse->limitOffset)
1848         {
1849                 tuple_fraction = preprocess_limit(root, tuple_fraction,
1850                                                                                   &offset_est, &count_est);
1851
1852                 /*
1853                  * If we have a known LIMIT, and don't have an unknown OFFSET, we can
1854                  * estimate the effects of using a bounded sort.
1855                  */
1856                 if (count_est > 0 && offset_est >= 0)
1857                         limit_tuples = (double) count_est + (double) offset_est;
1858         }
1859
1860         /* Make tuple_fraction accessible to lower-level routines */
1861         root->tuple_fraction = tuple_fraction;
1862
1863         if (parse->setOperations)
1864         {
1865                 /*
1866                  * If there's a top-level ORDER BY, assume we have to fetch all the
1867                  * tuples.  This might be too simplistic given all the hackery below
1868                  * to possibly avoid the sort; but the odds of accurate estimates here
1869                  * are pretty low anyway.  XXX try to get rid of this in favor of
1870                  * letting plan_set_operations generate both fast-start and
1871                  * cheapest-total paths.
1872                  */
1873                 if (parse->sortClause)
1874                         root->tuple_fraction = 0.0;
1875
1876                 /*
1877                  * Construct Paths for set operations.  The results will not need any
1878                  * work except perhaps a top-level sort and/or LIMIT.  Note that any
1879                  * special work for recursive unions is the responsibility of
1880                  * plan_set_operations.
1881                  */
1882                 current_rel = plan_set_operations(root);
1883
1884                 /*
1885                  * We should not need to call preprocess_targetlist, since we must be
1886                  * in a SELECT query node.  Instead, use the processed_tlist returned
1887                  * by plan_set_operations (since this tells whether it returned any
1888                  * resjunk columns!), and transfer any sort key information from the
1889                  * original tlist.
1890                  */
1891                 Assert(parse->commandType == CMD_SELECT);
1892
1893                 /* for safety, copy processed_tlist instead of modifying in-place */
1894                 root->processed_tlist =
1895                         postprocess_setop_tlist(copyObject(root->processed_tlist),
1896                                                                         parse->targetList);
1897
1898                 /* Also extract the PathTarget form of the setop result tlist */
1899                 final_target = current_rel->cheapest_total_path->pathtarget;
1900
1901                 /* And check whether it's parallel safe */
1902                 final_target_parallel_safe =
1903                         is_parallel_safe(root, (Node *) final_target->exprs);
1904
1905                 /* The setop result tlist couldn't contain any SRFs */
1906                 Assert(!parse->hasTargetSRFs);
1907                 final_targets = final_targets_contain_srfs = NIL;
1908
1909                 /*
1910                  * Can't handle FOR [KEY] UPDATE/SHARE here (parser should have
1911                  * checked already, but let's make sure).
1912                  */
1913                 if (parse->rowMarks)
1914                         ereport(ERROR,
1915                                         (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
1916                         /*------
1917                           translator: %s is a SQL row locking clause such as FOR UPDATE */
1918                                          errmsg("%s is not allowed with UNION/INTERSECT/EXCEPT",
1919                                                         LCS_asString(linitial_node(RowMarkClause,
1920                                                                                                            parse->rowMarks)->strength))));
1921
1922                 /*
1923                  * Calculate pathkeys that represent result ordering requirements
1924                  */
1925                 Assert(parse->distinctClause == NIL);
1926                 root->sort_pathkeys = make_pathkeys_for_sortclauses(root,
1927                                                                                                                         parse->sortClause,
1928                                                                                                                         root->processed_tlist);
1929         }
1930         else
1931         {
1932                 /* No set operations, do regular planning */
1933                 PathTarget *sort_input_target;
1934                 List       *sort_input_targets;
1935                 List       *sort_input_targets_contain_srfs;
1936                 bool            sort_input_target_parallel_safe;
1937                 PathTarget *grouping_target;
1938                 List       *grouping_targets;
1939                 List       *grouping_targets_contain_srfs;
1940                 bool            grouping_target_parallel_safe;
1941                 PathTarget *scanjoin_target;
1942                 List       *scanjoin_targets;
1943                 List       *scanjoin_targets_contain_srfs;
1944                 bool            scanjoin_target_parallel_safe;
1945                 bool            scanjoin_target_same_exprs;
1946                 bool            have_grouping;
1947                 AggClauseCosts agg_costs;
1948                 WindowFuncLists *wflists = NULL;
1949                 List       *activeWindows = NIL;
1950                 grouping_sets_data *gset_data = NULL;
1951                 standard_qp_extra qp_extra;
1952
1953                 /* A recursive query should always have setOperations */
1954                 Assert(!root->hasRecursion);
1955
1956                 /* Preprocess grouping sets and GROUP BY clause, if any */
1957                 if (parse->groupingSets)
1958                 {
1959                         gset_data = preprocess_grouping_sets(root);
1960                 }
1961                 else
1962                 {
1963                         /* Preprocess regular GROUP BY clause, if any */
1964                         if (parse->groupClause)
1965                                 parse->groupClause = preprocess_groupclause(root, NIL);
1966                 }
1967
1968                 /*
1969                  * Preprocess targetlist.  Note that much of the remaining planning
1970                  * work will be done with the PathTarget representation of tlists, but
1971                  * we must also maintain the full representation of the final tlist so
1972                  * that we can transfer its decoration (resnames etc) to the topmost
1973                  * tlist of the finished Plan.  This is kept in processed_tlist.
1974                  */
1975                 root->processed_tlist = preprocess_targetlist(root);
1976
1977                 /*
1978                  * Collect statistics about aggregates for estimating costs, and mark
1979                  * all the aggregates with resolved aggtranstypes.  We must do this
1980                  * before slicing and dicing the tlist into various pathtargets, else
1981                  * some copies of the Aggref nodes might escape being marked with the
1982                  * correct transtypes.
1983                  *
1984                  * Note: currently, we do not detect duplicate aggregates here.  This
1985                  * may result in somewhat-overestimated cost, which is fine for our
1986                  * purposes since all Paths will get charged the same.  But at some
1987                  * point we might wish to do that detection in the planner, rather
1988                  * than during executor startup.
1989                  */
1990                 MemSet(&agg_costs, 0, sizeof(AggClauseCosts));
1991                 if (parse->hasAggs)
1992                 {
1993                         get_agg_clause_costs(root, (Node *) root->processed_tlist,
1994                                                                  AGGSPLIT_SIMPLE, &agg_costs);
1995                         get_agg_clause_costs(root, parse->havingQual, AGGSPLIT_SIMPLE,
1996                                                                  &agg_costs);
1997                 }
1998
1999                 /*
2000                  * Locate any window functions in the tlist.  (We don't need to look
2001                  * anywhere else, since expressions used in ORDER BY will be in there
2002                  * too.)  Note that they could all have been eliminated by constant
2003                  * folding, in which case we don't need to do any more work.
2004                  */
2005                 if (parse->hasWindowFuncs)
2006                 {
2007                         wflists = find_window_functions((Node *) root->processed_tlist,
2008                                                                                         list_length(parse->windowClause));
2009                         if (wflists->numWindowFuncs > 0)
2010                                 activeWindows = select_active_windows(root, wflists);
2011                         else
2012                                 parse->hasWindowFuncs = false;
2013                 }
2014
2015                 /*
2016                  * Preprocess MIN/MAX aggregates, if any.  Note: be careful about
2017                  * adding logic between here and the query_planner() call.  Anything
2018                  * that is needed in MIN/MAX-optimizable cases will have to be
2019                  * duplicated in planagg.c.
2020                  */
2021                 if (parse->hasAggs)
2022                         preprocess_minmax_aggregates(root);
2023
2024                 /*
2025                  * Figure out whether there's a hard limit on the number of rows that
2026                  * query_planner's result subplan needs to return.  Even if we know a
2027                  * hard limit overall, it doesn't apply if the query has any
2028                  * grouping/aggregation operations, or SRFs in the tlist.
2029                  */
2030                 if (parse->groupClause ||
2031                         parse->groupingSets ||
2032                         parse->distinctClause ||
2033                         parse->hasAggs ||
2034                         parse->hasWindowFuncs ||
2035                         parse->hasTargetSRFs ||
2036                         root->hasHavingQual)
2037                         root->limit_tuples = -1.0;
2038                 else
2039                         root->limit_tuples = limit_tuples;
2040
2041                 /* Set up data needed by standard_qp_callback */
2042                 qp_extra.activeWindows = activeWindows;
2043                 qp_extra.groupClause = (gset_data
2044                                                                 ? (gset_data->rollups ? linitial_node(RollupData, gset_data->rollups)->groupClause : NIL)
2045                                                                 : parse->groupClause);
2046
2047                 /*
2048                  * Generate the best unsorted and presorted paths for the scan/join
2049                  * portion of this Query, ie the processing represented by the
2050                  * FROM/WHERE clauses.  (Note there may not be any presorted paths.)
2051                  * We also generate (in standard_qp_callback) pathkey representations
2052                  * of the query's sort clause, distinct clause, etc.
2053                  */
2054                 current_rel = query_planner(root, standard_qp_callback, &qp_extra);
2055
2056                 /*
2057                  * Convert the query's result tlist into PathTarget format.
2058                  *
2059                  * Note: this cannot be done before query_planner() has performed
2060                  * appendrel expansion, because that might add resjunk entries to
2061                  * root->processed_tlist.  Waiting till afterwards is also helpful
2062                  * because the target width estimates can use per-Var width numbers
2063                  * that were obtained within query_planner().
2064                  */
2065                 final_target = create_pathtarget(root, root->processed_tlist);
2066                 final_target_parallel_safe =
2067                         is_parallel_safe(root, (Node *) final_target->exprs);
2068
2069                 /*
2070                  * If ORDER BY was given, consider whether we should use a post-sort
2071                  * projection, and compute the adjusted target for preceding steps if
2072                  * so.
2073                  */
2074                 if (parse->sortClause)
2075                 {
2076                         sort_input_target = make_sort_input_target(root,
2077                                                                                                            final_target,
2078                                                                                                            &have_postponed_srfs);
2079                         sort_input_target_parallel_safe =
2080                                 is_parallel_safe(root, (Node *) sort_input_target->exprs);
2081                 }
2082                 else
2083                 {
2084                         sort_input_target = final_target;
2085                         sort_input_target_parallel_safe = final_target_parallel_safe;
2086                 }
2087
2088                 /*
2089                  * If we have window functions to deal with, the output from any
2090                  * grouping step needs to be what the window functions want;
2091                  * otherwise, it should be sort_input_target.
2092                  */
2093                 if (activeWindows)
2094                 {
2095                         grouping_target = make_window_input_target(root,
2096                                                                                                            final_target,
2097                                                                                                            activeWindows);
2098                         grouping_target_parallel_safe =
2099                                 is_parallel_safe(root, (Node *) grouping_target->exprs);
2100                 }
2101                 else
2102                 {
2103                         grouping_target = sort_input_target;
2104                         grouping_target_parallel_safe = sort_input_target_parallel_safe;
2105                 }
2106
2107                 /*
2108                  * If we have grouping or aggregation to do, the topmost scan/join
2109                  * plan node must emit what the grouping step wants; otherwise, it
2110                  * should emit grouping_target.
2111                  */
2112                 have_grouping = (parse->groupClause || parse->groupingSets ||
2113                                                  parse->hasAggs || root->hasHavingQual);
2114                 if (have_grouping)
2115                 {
2116                         scanjoin_target = make_group_input_target(root, final_target);
2117                         scanjoin_target_parallel_safe =
2118                                 is_parallel_safe(root, (Node *) scanjoin_target->exprs);
2119                 }
2120                 else
2121                 {
2122                         scanjoin_target = grouping_target;
2123                         scanjoin_target_parallel_safe = grouping_target_parallel_safe;
2124                 }
2125
2126                 /*
2127                  * If there are any SRFs in the targetlist, we must separate each of
2128                  * these PathTargets into SRF-computing and SRF-free targets.  Replace
2129                  * each of the named targets with a SRF-free version, and remember the
2130                  * list of additional projection steps we need to add afterwards.
2131                  */
2132                 if (parse->hasTargetSRFs)
2133                 {
2134                         /* final_target doesn't recompute any SRFs in sort_input_target */
2135                         split_pathtarget_at_srfs(root, final_target, sort_input_target,
2136                                                                          &final_targets,
2137                                                                          &final_targets_contain_srfs);
2138                         final_target = linitial_node(PathTarget, final_targets);
2139                         Assert(!linitial_int(final_targets_contain_srfs));
2140                         /* likewise for sort_input_target vs. grouping_target */
2141                         split_pathtarget_at_srfs(root, sort_input_target, grouping_target,
2142                                                                          &sort_input_targets,
2143                                                                          &sort_input_targets_contain_srfs);
2144                         sort_input_target = linitial_node(PathTarget, sort_input_targets);
2145                         Assert(!linitial_int(sort_input_targets_contain_srfs));
2146                         /* likewise for grouping_target vs. scanjoin_target */
2147                         split_pathtarget_at_srfs(root, grouping_target, scanjoin_target,
2148                                                                          &grouping_targets,
2149                                                                          &grouping_targets_contain_srfs);
2150                         grouping_target = linitial_node(PathTarget, grouping_targets);
2151                         Assert(!linitial_int(grouping_targets_contain_srfs));
2152                         /* scanjoin_target will not have any SRFs precomputed for it */
2153                         split_pathtarget_at_srfs(root, scanjoin_target, NULL,
2154                                                                          &scanjoin_targets,
2155                                                                          &scanjoin_targets_contain_srfs);
2156                         scanjoin_target = linitial_node(PathTarget, scanjoin_targets);
2157                         Assert(!linitial_int(scanjoin_targets_contain_srfs));
2158                 }
2159                 else
2160                 {
2161                         /* initialize lists; for most of these, dummy values are OK */
2162                         final_targets = final_targets_contain_srfs = NIL;
2163                         sort_input_targets = sort_input_targets_contain_srfs = NIL;
2164                         grouping_targets = grouping_targets_contain_srfs = NIL;
2165                         scanjoin_targets = list_make1(scanjoin_target);
2166                         scanjoin_targets_contain_srfs = NIL;
2167                 }
2168
2169                 /* Apply scan/join target. */
2170                 scanjoin_target_same_exprs = list_length(scanjoin_targets) == 1
2171                         && equal(scanjoin_target->exprs, current_rel->reltarget->exprs);
2172                 apply_scanjoin_target_to_paths(root, current_rel, scanjoin_targets,
2173                                                                            scanjoin_targets_contain_srfs,
2174                                                                            scanjoin_target_parallel_safe,
2175                                                                            scanjoin_target_same_exprs);
2176
2177                 /*
2178                  * Save the various upper-rel PathTargets we just computed into
2179                  * root->upper_targets[].  The core code doesn't use this, but it
2180                  * provides a convenient place for extensions to get at the info.  For
2181                  * consistency, we save all the intermediate targets, even though some
2182                  * of the corresponding upperrels might not be needed for this query.
2183                  */
2184                 root->upper_targets[UPPERREL_FINAL] = final_target;
2185                 root->upper_targets[UPPERREL_ORDERED] = final_target;
2186                 root->upper_targets[UPPERREL_DISTINCT] = sort_input_target;
2187                 root->upper_targets[UPPERREL_WINDOW] = sort_input_target;
2188                 root->upper_targets[UPPERREL_GROUP_AGG] = grouping_target;
2189
2190                 /*
2191                  * If we have grouping and/or aggregation, consider ways to implement
2192                  * that.  We build a new upperrel representing the output of this
2193                  * phase.
2194                  */
2195                 if (have_grouping)
2196                 {
2197                         current_rel = create_grouping_paths(root,
2198                                                                                                 current_rel,
2199                                                                                                 grouping_target,
2200                                                                                                 grouping_target_parallel_safe,
2201                                                                                                 &agg_costs,
2202                                                                                                 gset_data);
2203                         /* Fix things up if grouping_target contains SRFs */
2204                         if (parse->hasTargetSRFs)
2205                                 adjust_paths_for_srfs(root, current_rel,
2206                                                                           grouping_targets,
2207                                                                           grouping_targets_contain_srfs);
2208                 }
2209
2210                 /*
2211                  * If we have window functions, consider ways to implement those.  We
2212                  * build a new upperrel representing the output of this phase.
2213                  */
2214                 if (activeWindows)
2215                 {
2216                         current_rel = create_window_paths(root,
2217                                                                                           current_rel,
2218                                                                                           grouping_target,
2219                                                                                           sort_input_target,
2220                                                                                           sort_input_target_parallel_safe,
2221                                                                                           wflists,
2222                                                                                           activeWindows);
2223                         /* Fix things up if sort_input_target contains SRFs */
2224                         if (parse->hasTargetSRFs)
2225                                 adjust_paths_for_srfs(root, current_rel,
2226                                                                           sort_input_targets,
2227                                                                           sort_input_targets_contain_srfs);
2228                 }
2229
2230                 /*
2231                  * If there is a DISTINCT clause, consider ways to implement that. We
2232                  * build a new upperrel representing the output of this phase.
2233                  */
2234                 if (parse->distinctClause)
2235                 {
2236                         current_rel = create_distinct_paths(root,
2237                                                                                                 current_rel);
2238                 }
2239         }                                                       /* end of if (setOperations) */
2240
2241         /*
2242          * If ORDER BY was given, consider ways to implement that, and generate a
2243          * new upperrel containing only paths that emit the correct ordering and
2244          * project the correct final_target.  We can apply the original
2245          * limit_tuples limit in sort costing here, but only if there are no
2246          * postponed SRFs.
2247          */
2248         if (parse->sortClause)
2249         {
2250                 current_rel = create_ordered_paths(root,
2251                                                                                    current_rel,
2252                                                                                    final_target,
2253                                                                                    final_target_parallel_safe,
2254                                                                                    have_postponed_srfs ? -1.0 :
2255                                                                                    limit_tuples);
2256                 /* Fix things up if final_target contains SRFs */
2257                 if (parse->hasTargetSRFs)
2258                         adjust_paths_for_srfs(root, current_rel,
2259                                                                   final_targets,
2260                                                                   final_targets_contain_srfs);
2261         }
2262
2263         /*
2264          * Now we are prepared to build the final-output upperrel.
2265          */
2266         final_rel = fetch_upper_rel(root, UPPERREL_FINAL, NULL);
2267
2268         /*
2269          * If the input rel is marked consider_parallel and there's nothing that's
2270          * not parallel-safe in the LIMIT clause, then the final_rel can be marked
2271          * consider_parallel as well.  Note that if the query has rowMarks or is
2272          * not a SELECT, consider_parallel will be false for every relation in the
2273          * query.
2274          */
2275         if (current_rel->consider_parallel &&
2276                 is_parallel_safe(root, parse->limitOffset) &&
2277                 is_parallel_safe(root, parse->limitCount))
2278                 final_rel->consider_parallel = true;
2279
2280         /*
2281          * If the current_rel belongs to a single FDW, so does the final_rel.
2282          */
2283         final_rel->serverid = current_rel->serverid;
2284         final_rel->userid = current_rel->userid;
2285         final_rel->useridiscurrent = current_rel->useridiscurrent;
2286         final_rel->fdwroutine = current_rel->fdwroutine;
2287
2288         /*
2289          * Generate paths for the final_rel.  Insert all surviving paths, with
2290          * LockRows, Limit, and/or ModifyTable steps added if needed.
2291          */
2292         foreach(lc, current_rel->pathlist)
2293         {
2294                 Path       *path = (Path *) lfirst(lc);
2295
2296                 /*
2297                  * If there is a FOR [KEY] UPDATE/SHARE clause, add the LockRows node.
2298                  * (Note: we intentionally test parse->rowMarks not root->rowMarks
2299                  * here.  If there are only non-locking rowmarks, they should be
2300                  * handled by the ModifyTable node instead.  However, root->rowMarks
2301                  * is what goes into the LockRows node.)
2302                  */
2303                 if (parse->rowMarks)
2304                 {
2305                         path = (Path *) create_lockrows_path(root, final_rel, path,
2306                                                                                                  root->rowMarks,
2307                                                                                                  assign_special_exec_param(root));
2308                 }
2309
2310                 /*
2311                  * If there is a LIMIT/OFFSET clause, add the LIMIT node.
2312                  */
2313                 if (limit_needed(parse))
2314                 {
2315                         path = (Path *) create_limit_path(root, final_rel, path,
2316                                                                                           parse->limitOffset,
2317                                                                                           parse->limitCount,
2318                                                                                           offset_est, count_est);
2319                 }
2320
2321                 /*
2322                  * If this is an INSERT/UPDATE/DELETE, and we're not being called from
2323                  * inheritance_planner, add the ModifyTable node.
2324                  */
2325                 if (parse->commandType != CMD_SELECT && !inheritance_update)
2326                 {
2327                         Index           rootRelation;
2328                         List       *withCheckOptionLists;
2329                         List       *returningLists;
2330                         List       *rowMarks;
2331
2332                         /*
2333                          * If target is a partition root table, we need to mark the
2334                          * ModifyTable node appropriately for that.
2335                          */
2336                         if (rt_fetch(parse->resultRelation, parse->rtable)->relkind ==
2337                                 RELKIND_PARTITIONED_TABLE)
2338                                 rootRelation = parse->resultRelation;
2339                         else
2340                                 rootRelation = 0;
2341
2342                         /*
2343                          * Set up the WITH CHECK OPTION and RETURNING lists-of-lists, if
2344                          * needed.
2345                          */
2346                         if (parse->withCheckOptions)
2347                                 withCheckOptionLists = list_make1(parse->withCheckOptions);
2348                         else
2349                                 withCheckOptionLists = NIL;
2350
2351                         if (parse->returningList)
2352                                 returningLists = list_make1(parse->returningList);
2353                         else
2354                                 returningLists = NIL;
2355
2356                         /*
2357                          * If there was a FOR [KEY] UPDATE/SHARE clause, the LockRows node
2358                          * will have dealt with fetching non-locked marked rows, else we
2359                          * need to have ModifyTable do that.
2360                          */
2361                         if (parse->rowMarks)
2362                                 rowMarks = NIL;
2363                         else
2364                                 rowMarks = root->rowMarks;
2365
2366                         path = (Path *)
2367                                 create_modifytable_path(root, final_rel,
2368                                                                                 parse->commandType,
2369                                                                                 parse->canSetTag,
2370                                                                                 parse->resultRelation,
2371                                                                                 rootRelation,
2372                                                                                 false,
2373                                                                                 list_make1_int(parse->resultRelation),
2374                                                                                 list_make1(path),
2375                                                                                 list_make1(root),
2376                                                                                 withCheckOptionLists,
2377                                                                                 returningLists,
2378                                                                                 rowMarks,
2379                                                                                 parse->onConflict,
2380                                                                                 assign_special_exec_param(root));
2381                 }
2382
2383                 /* And shove it into final_rel */
2384                 add_path(final_rel, path);
2385         }
2386
2387         /*
2388          * Generate partial paths for final_rel, too, if outer query levels might
2389          * be able to make use of them.
2390          */
2391         if (final_rel->consider_parallel && root->query_level > 1 &&
2392                 !limit_needed(parse))
2393         {
2394                 Assert(!parse->rowMarks && parse->commandType == CMD_SELECT);
2395                 foreach(lc, current_rel->partial_pathlist)
2396                 {
2397                         Path       *partial_path = (Path *) lfirst(lc);
2398
2399                         add_partial_path(final_rel, partial_path);
2400                 }
2401         }
2402
2403         extra.limit_needed = limit_needed(parse);
2404         extra.limit_tuples = limit_tuples;
2405         extra.count_est = count_est;
2406         extra.offset_est = offset_est;
2407
2408         /*
2409          * If there is an FDW that's responsible for all baserels of the query,
2410          * let it consider adding ForeignPaths.
2411          */
2412         if (final_rel->fdwroutine &&
2413                 final_rel->fdwroutine->GetForeignUpperPaths)
2414                 final_rel->fdwroutine->GetForeignUpperPaths(root, UPPERREL_FINAL,
2415                                                                                                         current_rel, final_rel,
2416                                                                                                         &extra);
2417
2418         /* Let extensions possibly add some more paths */
2419         if (create_upper_paths_hook)
2420                 (*create_upper_paths_hook) (root, UPPERREL_FINAL,
2421                                                                         current_rel, final_rel, &extra);
2422
2423         /* Note: currently, we leave it to callers to do set_cheapest() */
2424 }
2425
2426 /*
2427  * Do preprocessing for groupingSets clause and related data.  This handles the
2428  * preliminary steps of expanding the grouping sets, organizing them into lists
2429  * of rollups, and preparing annotations which will later be filled in with
2430  * size estimates.
2431  */
2432 static grouping_sets_data *
2433 preprocess_grouping_sets(PlannerInfo *root)
2434 {
2435         Query      *parse = root->parse;
2436         List       *sets;
2437         int                     maxref = 0;
2438         ListCell   *lc;
2439         ListCell   *lc_set;
2440         grouping_sets_data *gd = palloc0(sizeof(grouping_sets_data));
2441
2442         parse->groupingSets = expand_grouping_sets(parse->groupingSets, -1);
2443
2444         gd->any_hashable = false;
2445         gd->unhashable_refs = NULL;
2446         gd->unsortable_refs = NULL;
2447         gd->unsortable_sets = NIL;
2448
2449         if (parse->groupClause)
2450         {
2451                 ListCell   *lc;
2452
2453                 foreach(lc, parse->groupClause)
2454                 {
2455                         SortGroupClause *gc = lfirst_node(SortGroupClause, lc);
2456                         Index           ref = gc->tleSortGroupRef;
2457
2458                         if (ref > maxref)
2459                                 maxref = ref;
2460
2461                         if (!gc->hashable)
2462                                 gd->unhashable_refs = bms_add_member(gd->unhashable_refs, ref);
2463
2464                         if (!OidIsValid(gc->sortop))
2465                                 gd->unsortable_refs = bms_add_member(gd->unsortable_refs, ref);
2466                 }
2467         }
2468
2469         /* Allocate workspace array for remapping */
2470         gd->tleref_to_colnum_map = (int *) palloc((maxref + 1) * sizeof(int));
2471
2472         /*
2473          * If we have any unsortable sets, we must extract them before trying to
2474          * prepare rollups. Unsortable sets don't go through
2475          * reorder_grouping_sets, so we must apply the GroupingSetData annotation
2476          * here.
2477          */
2478         if (!bms_is_empty(gd->unsortable_refs))
2479         {
2480                 List       *sortable_sets = NIL;
2481
2482                 foreach(lc, parse->groupingSets)
2483                 {
2484                         List       *gset = (List *) lfirst(lc);
2485
2486                         if (bms_overlap_list(gd->unsortable_refs, gset))
2487                         {
2488                                 GroupingSetData *gs = makeNode(GroupingSetData);
2489
2490                                 gs->set = gset;
2491                                 gd->unsortable_sets = lappend(gd->unsortable_sets, gs);
2492
2493                                 /*
2494                                  * We must enforce here that an unsortable set is hashable;
2495                                  * later code assumes this.  Parse analysis only checks that
2496                                  * every individual column is either hashable or sortable.
2497                                  *
2498                                  * Note that passing this test doesn't guarantee we can
2499                                  * generate a plan; there might be other showstoppers.
2500                                  */
2501                                 if (bms_overlap_list(gd->unhashable_refs, gset))
2502                                         ereport(ERROR,
2503                                                         (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
2504                                                          errmsg("could not implement GROUP BY"),
2505                                                          errdetail("Some of the datatypes only support hashing, while others only support sorting.")));
2506                         }
2507                         else
2508                                 sortable_sets = lappend(sortable_sets, gset);
2509                 }
2510
2511                 if (sortable_sets)
2512                         sets = extract_rollup_sets(sortable_sets);
2513                 else
2514                         sets = NIL;
2515         }
2516         else
2517                 sets = extract_rollup_sets(parse->groupingSets);
2518
2519         foreach(lc_set, sets)
2520         {
2521                 List       *current_sets = (List *) lfirst(lc_set);
2522                 RollupData *rollup = makeNode(RollupData);
2523                 GroupingSetData *gs;
2524
2525                 /*
2526                  * Reorder the current list of grouping sets into correct prefix
2527                  * order.  If only one aggregation pass is needed, try to make the
2528                  * list match the ORDER BY clause; if more than one pass is needed, we
2529                  * don't bother with that.
2530                  *
2531                  * Note that this reorders the sets from smallest-member-first to
2532                  * largest-member-first, and applies the GroupingSetData annotations,
2533                  * though the data will be filled in later.
2534                  */
2535                 current_sets = reorder_grouping_sets(current_sets,
2536                                                                                          (list_length(sets) == 1
2537                                                                                           ? parse->sortClause
2538                                                                                           : NIL));
2539
2540                 /*
2541                  * Get the initial (and therefore largest) grouping set.
2542                  */
2543                 gs = linitial_node(GroupingSetData, current_sets);
2544
2545                 /*
2546                  * Order the groupClause appropriately.  If the first grouping set is
2547                  * empty, then the groupClause must also be empty; otherwise we have
2548                  * to force the groupClause to match that grouping set's order.
2549                  *
2550                  * (The first grouping set can be empty even though parse->groupClause
2551                  * is not empty only if all non-empty grouping sets are unsortable.
2552                  * The groupClauses for hashed grouping sets are built later on.)
2553                  */
2554                 if (gs->set)
2555                         rollup->groupClause = preprocess_groupclause(root, gs->set);
2556                 else
2557                         rollup->groupClause = NIL;
2558
2559                 /*
2560                  * Is it hashable? We pretend empty sets are hashable even though we
2561                  * actually force them not to be hashed later. But don't bother if
2562                  * there's nothing but empty sets (since in that case we can't hash
2563                  * anything).
2564                  */
2565                 if (gs->set &&
2566                         !bms_overlap_list(gd->unhashable_refs, gs->set))
2567                 {
2568                         rollup->hashable = true;
2569                         gd->any_hashable = true;
2570                 }
2571
2572                 /*
2573                  * Now that we've pinned down an order for the groupClause for this
2574                  * list of grouping sets, we need to remap the entries in the grouping
2575                  * sets from sortgrouprefs to plain indices (0-based) into the
2576                  * groupClause for this collection of grouping sets. We keep the
2577                  * original form for later use, though.
2578                  */
2579                 rollup->gsets = remap_to_groupclause_idx(rollup->groupClause,
2580                                                                                                  current_sets,
2581                                                                                                  gd->tleref_to_colnum_map);
2582                 rollup->gsets_data = current_sets;
2583
2584                 gd->rollups = lappend(gd->rollups, rollup);
2585         }
2586
2587         if (gd->unsortable_sets)
2588         {
2589                 /*
2590                  * We have not yet pinned down a groupclause for this, but we will
2591                  * need index-based lists for estimation purposes. Construct
2592                  * hash_sets_idx based on the entire original groupclause for now.
2593                  */
2594                 gd->hash_sets_idx = remap_to_groupclause_idx(parse->groupClause,
2595                                                                                                          gd->unsortable_sets,
2596                                                                                                          gd->tleref_to_colnum_map);
2597                 gd->any_hashable = true;
2598         }
2599
2600         return gd;
2601 }
2602
2603 /*
2604  * Given a groupclause and a list of GroupingSetData, return equivalent sets
2605  * (without annotation) mapped to indexes into the given groupclause.
2606  */
2607 static List *
2608 remap_to_groupclause_idx(List *groupClause,
2609                                                  List *gsets,
2610                                                  int *tleref_to_colnum_map)
2611 {
2612         int                     ref = 0;
2613         List       *result = NIL;
2614         ListCell   *lc;
2615
2616         foreach(lc, groupClause)
2617         {
2618                 SortGroupClause *gc = lfirst_node(SortGroupClause, lc);
2619
2620                 tleref_to_colnum_map[gc->tleSortGroupRef] = ref++;
2621         }
2622
2623         foreach(lc, gsets)
2624         {
2625                 List       *set = NIL;
2626                 ListCell   *lc2;
2627                 GroupingSetData *gs = lfirst_node(GroupingSetData, lc);
2628
2629                 foreach(lc2, gs->set)
2630                 {
2631                         set = lappend_int(set, tleref_to_colnum_map[lfirst_int(lc2)]);
2632                 }
2633
2634                 result = lappend(result, set);
2635         }
2636
2637         return result;
2638 }
2639
2640
2641 /*
2642  * preprocess_rowmarks - set up PlanRowMarks if needed
2643  */
2644 static void
2645 preprocess_rowmarks(PlannerInfo *root)
2646 {
2647         Query      *parse = root->parse;
2648         Bitmapset  *rels;
2649         List       *prowmarks;
2650         ListCell   *l;
2651         int                     i;
2652
2653         if (parse->rowMarks)
2654         {
2655                 /*
2656                  * We've got trouble if FOR [KEY] UPDATE/SHARE appears inside
2657                  * grouping, since grouping renders a reference to individual tuple
2658                  * CTIDs invalid.  This is also checked at parse time, but that's
2659                  * insufficient because of rule substitution, query pullup, etc.
2660                  */
2661                 CheckSelectLocking(parse, linitial_node(RowMarkClause,
2662                                                                                                 parse->rowMarks)->strength);
2663         }
2664         else
2665         {
2666                 /*
2667                  * We only need rowmarks for UPDATE, DELETE, or FOR [KEY]
2668                  * UPDATE/SHARE.
2669                  */
2670                 if (parse->commandType != CMD_UPDATE &&
2671                         parse->commandType != CMD_DELETE)
2672                         return;
2673         }
2674
2675         /*
2676          * We need to have rowmarks for all base relations except the target. We
2677          * make a bitmapset of all base rels and then remove the items we don't
2678          * need or have FOR [KEY] UPDATE/SHARE marks for.
2679          */
2680         rels = get_relids_in_jointree((Node *) parse->jointree, false);
2681         if (parse->resultRelation)
2682                 rels = bms_del_member(rels, parse->resultRelation);
2683
2684         /*
2685          * Convert RowMarkClauses to PlanRowMark representation.
2686          */
2687         prowmarks = NIL;
2688         foreach(l, parse->rowMarks)
2689         {
2690                 RowMarkClause *rc = lfirst_node(RowMarkClause, l);
2691                 RangeTblEntry *rte = rt_fetch(rc->rti, parse->rtable);
2692                 PlanRowMark *newrc;
2693
2694                 /*
2695                  * Currently, it is syntactically impossible to have FOR UPDATE et al
2696                  * applied to an update/delete target rel.  If that ever becomes
2697                  * possible, we should drop the target from the PlanRowMark list.
2698                  */
2699                 Assert(rc->rti != parse->resultRelation);
2700
2701                 /*
2702                  * Ignore RowMarkClauses for subqueries; they aren't real tables and
2703                  * can't support true locking.  Subqueries that got flattened into the
2704                  * main query should be ignored completely.  Any that didn't will get
2705                  * ROW_MARK_COPY items in the next loop.
2706                  */
2707                 if (rte->rtekind != RTE_RELATION)
2708                         continue;
2709
2710                 rels = bms_del_member(rels, rc->rti);
2711
2712                 newrc = makeNode(PlanRowMark);
2713                 newrc->rti = newrc->prti = rc->rti;
2714                 newrc->rowmarkId = ++(root->glob->lastRowMarkId);
2715                 newrc->markType = select_rowmark_type(rte, rc->strength);
2716                 newrc->allMarkTypes = (1 << newrc->markType);
2717                 newrc->strength = rc->strength;
2718                 newrc->waitPolicy = rc->waitPolicy;
2719                 newrc->isParent = false;
2720
2721                 prowmarks = lappend(prowmarks, newrc);
2722         }
2723
2724         /*
2725          * Now, add rowmarks for any non-target, non-locked base relations.
2726          */
2727         i = 0;
2728         foreach(l, parse->rtable)
2729         {
2730                 RangeTblEntry *rte = lfirst_node(RangeTblEntry, l);
2731                 PlanRowMark *newrc;
2732
2733                 i++;
2734                 if (!bms_is_member(i, rels))
2735                         continue;
2736
2737                 newrc = makeNode(PlanRowMark);
2738                 newrc->rti = newrc->prti = i;
2739                 newrc->rowmarkId = ++(root->glob->lastRowMarkId);
2740                 newrc->markType = select_rowmark_type(rte, LCS_NONE);
2741                 newrc->allMarkTypes = (1 << newrc->markType);
2742                 newrc->strength = LCS_NONE;
2743                 newrc->waitPolicy = LockWaitBlock;      /* doesn't matter */
2744                 newrc->isParent = false;
2745
2746                 prowmarks = lappend(prowmarks, newrc);
2747         }
2748
2749         root->rowMarks = prowmarks;
2750 }
2751
2752 /*
2753  * Select RowMarkType to use for a given table
2754  */
2755 RowMarkType
2756 select_rowmark_type(RangeTblEntry *rte, LockClauseStrength strength)
2757 {
2758         if (rte->rtekind != RTE_RELATION)
2759         {
2760                 /* If it's not a table at all, use ROW_MARK_COPY */
2761                 return ROW_MARK_COPY;
2762         }
2763         else if (rte->relkind == RELKIND_FOREIGN_TABLE)
2764         {
2765                 /* Let the FDW select the rowmark type, if it wants to */
2766                 FdwRoutine *fdwroutine = GetFdwRoutineByRelId(rte->relid);
2767
2768                 if (fdwroutine->GetForeignRowMarkType != NULL)
2769                         return fdwroutine->GetForeignRowMarkType(rte, strength);
2770                 /* Otherwise, use ROW_MARK_COPY by default */
2771                 return ROW_MARK_COPY;
2772         }
2773         else
2774         {
2775                 /* Regular table, apply the appropriate lock type */
2776                 switch (strength)
2777                 {
2778                         case LCS_NONE:
2779
2780                                 /*
2781                                  * We don't need a tuple lock, only the ability to re-fetch
2782                                  * the row.
2783                                  */
2784                                 return ROW_MARK_REFERENCE;
2785                                 break;
2786                         case LCS_FORKEYSHARE:
2787                                 return ROW_MARK_KEYSHARE;
2788                                 break;
2789                         case LCS_FORSHARE:
2790                                 return ROW_MARK_SHARE;
2791                                 break;
2792                         case LCS_FORNOKEYUPDATE:
2793                                 return ROW_MARK_NOKEYEXCLUSIVE;
2794                                 break;
2795                         case LCS_FORUPDATE:
2796                                 return ROW_MARK_EXCLUSIVE;
2797                                 break;
2798                 }
2799                 elog(ERROR, "unrecognized LockClauseStrength %d", (int) strength);
2800                 return ROW_MARK_EXCLUSIVE;      /* keep compiler quiet */
2801         }
2802 }
2803
2804 /*
2805  * preprocess_limit - do pre-estimation for LIMIT and/or OFFSET clauses
2806  *
2807  * We try to estimate the values of the LIMIT/OFFSET clauses, and pass the
2808  * results back in *count_est and *offset_est.  These variables are set to
2809  * 0 if the corresponding clause is not present, and -1 if it's present
2810  * but we couldn't estimate the value for it.  (The "0" convention is OK
2811  * for OFFSET but a little bit bogus for LIMIT: effectively we estimate
2812  * LIMIT 0 as though it were LIMIT 1.  But this is in line with the planner's
2813  * usual practice of never estimating less than one row.)  These values will
2814  * be passed to create_limit_path, which see if you change this code.
2815  *
2816  * The return value is the suitably adjusted tuple_fraction to use for
2817  * planning the query.  This adjustment is not overridable, since it reflects
2818  * plan actions that grouping_planner() will certainly take, not assumptions
2819  * about context.
2820  */
2821 static double
2822 preprocess_limit(PlannerInfo *root, double tuple_fraction,
2823                                  int64 *offset_est, int64 *count_est)
2824 {
2825         Query      *parse = root->parse;
2826         Node       *est;
2827         double          limit_fraction;
2828
2829         /* Should not be called unless LIMIT or OFFSET */
2830         Assert(parse->limitCount || parse->limitOffset);
2831
2832         /*
2833          * Try to obtain the clause values.  We use estimate_expression_value
2834          * primarily because it can sometimes do something useful with Params.
2835          */
2836         if (parse->limitCount)
2837         {
2838                 est = estimate_expression_value(root, parse->limitCount);
2839                 if (est && IsA(est, Const))
2840                 {
2841                         if (((Const *) est)->constisnull)
2842                         {
2843                                 /* NULL indicates LIMIT ALL, ie, no limit */
2844                                 *count_est = 0; /* treat as not present */
2845                         }
2846                         else
2847                         {
2848                                 *count_est = DatumGetInt64(((Const *) est)->constvalue);
2849                                 if (*count_est <= 0)
2850                                         *count_est = 1; /* force to at least 1 */
2851                         }
2852                 }
2853                 else
2854                         *count_est = -1;        /* can't estimate */
2855         }
2856         else
2857                 *count_est = 0;                 /* not present */
2858
2859         if (parse->limitOffset)
2860         {
2861                 est = estimate_expression_value(root, parse->limitOffset);
2862                 if (est && IsA(est, Const))
2863                 {
2864                         if (((Const *) est)->constisnull)
2865                         {
2866                                 /* Treat NULL as no offset; the executor will too */
2867                                 *offset_est = 0;        /* treat as not present */
2868                         }
2869                         else
2870                         {
2871                                 *offset_est = DatumGetInt64(((Const *) est)->constvalue);
2872                                 if (*offset_est < 0)
2873                                         *offset_est = 0;        /* treat as not present */
2874                         }
2875                 }
2876                 else
2877                         *offset_est = -1;       /* can't estimate */
2878         }
2879         else
2880                 *offset_est = 0;                /* not present */
2881
2882         if (*count_est != 0)
2883         {
2884                 /*
2885                  * A LIMIT clause limits the absolute number of tuples returned.
2886                  * However, if it's not a constant LIMIT then we have to guess; for
2887                  * lack of a better idea, assume 10% of the plan's result is wanted.
2888                  */
2889                 if (*count_est < 0 || *offset_est < 0)
2890                 {
2891                         /* LIMIT or OFFSET is an expression ... punt ... */
2892                         limit_fraction = 0.10;
2893                 }
2894                 else
2895                 {
2896                         /* LIMIT (plus OFFSET, if any) is max number of tuples needed */
2897                         limit_fraction = (double) *count_est + (double) *offset_est;
2898                 }
2899
2900                 /*
2901                  * If we have absolute limits from both caller and LIMIT, use the
2902                  * smaller value; likewise if they are both fractional.  If one is
2903                  * fractional and the other absolute, we can't easily determine which
2904                  * is smaller, but we use the heuristic that the absolute will usually
2905                  * be smaller.
2906                  */
2907                 if (tuple_fraction >= 1.0)
2908                 {
2909                         if (limit_fraction >= 1.0)
2910                         {
2911                                 /* both absolute */
2912                                 tuple_fraction = Min(tuple_fraction, limit_fraction);
2913                         }
2914                         else
2915                         {
2916                                 /* caller absolute, limit fractional; use caller's value */
2917                         }
2918                 }
2919                 else if (tuple_fraction > 0.0)
2920                 {
2921                         if (limit_fraction >= 1.0)
2922                         {
2923                                 /* caller fractional, limit absolute; use limit */
2924                                 tuple_fraction = limit_fraction;
2925                         }
2926                         else
2927                         {
2928                                 /* both fractional */
2929                                 tuple_fraction = Min(tuple_fraction, limit_fraction);
2930                         }
2931                 }
2932                 else
2933                 {
2934                         /* no info from caller, just use limit */
2935                         tuple_fraction = limit_fraction;
2936                 }
2937         }
2938         else if (*offset_est != 0 && tuple_fraction > 0.0)
2939         {
2940                 /*
2941                  * We have an OFFSET but no LIMIT.  This acts entirely differently
2942                  * from the LIMIT case: here, we need to increase rather than decrease
2943                  * the caller's tuple_fraction, because the OFFSET acts to cause more
2944                  * tuples to be fetched instead of fewer.  This only matters if we got
2945                  * a tuple_fraction > 0, however.
2946                  *
2947                  * As above, use 10% if OFFSET is present but unestimatable.
2948                  */
2949                 if (*offset_est < 0)
2950                         limit_fraction = 0.10;
2951                 else
2952                         limit_fraction = (double) *offset_est;
2953
2954                 /*
2955                  * If we have absolute counts from both caller and OFFSET, add them
2956                  * together; likewise if they are both fractional.  If one is
2957                  * fractional and the other absolute, we want to take the larger, and
2958                  * we heuristically assume that's the fractional one.
2959                  */
2960                 if (tuple_fraction >= 1.0)
2961                 {
2962                         if (limit_fraction >= 1.0)
2963                         {
2964                                 /* both absolute, so add them together */
2965                                 tuple_fraction += limit_fraction;
2966                         }
2967                         else
2968                         {
2969                                 /* caller absolute, limit fractional; use limit */
2970                                 tuple_fraction = limit_fraction;
2971                         }
2972                 }
2973                 else
2974                 {
2975                         if (limit_fraction >= 1.0)
2976                         {
2977                                 /* caller fractional, limit absolute; use caller's value */
2978                         }
2979                         else
2980                         {
2981                                 /* both fractional, so add them together */
2982                                 tuple_fraction += limit_fraction;
2983                                 if (tuple_fraction >= 1.0)
2984                                         tuple_fraction = 0.0;   /* assume fetch all */
2985                         }
2986                 }
2987         }
2988
2989         return tuple_fraction;
2990 }
2991
2992 /*
2993  * limit_needed - do we actually need a Limit plan node?
2994  *
2995  * If we have constant-zero OFFSET and constant-null LIMIT, we can skip adding
2996  * a Limit node.  This is worth checking for because "OFFSET 0" is a common
2997  * locution for an optimization fence.  (Because other places in the planner
2998  * merely check whether parse->limitOffset isn't NULL, it will still work as
2999  * an optimization fence --- we're just suppressing unnecessary run-time
3000  * overhead.)
3001  *
3002  * This might look like it could be merged into preprocess_limit, but there's
3003  * a key distinction: here we need hard constants in OFFSET/LIMIT, whereas
3004  * in preprocess_limit it's good enough to consider estimated values.
3005  */
3006 bool
3007 limit_needed(Query *parse)
3008 {
3009         Node       *node;
3010
3011         node = parse->limitCount;
3012         if (node)
3013         {
3014                 if (IsA(node, Const))
3015                 {
3016                         /* NULL indicates LIMIT ALL, ie, no limit */
3017                         if (!((Const *) node)->constisnull)
3018                                 return true;    /* LIMIT with a constant value */
3019                 }
3020                 else
3021                         return true;            /* non-constant LIMIT */
3022         }
3023
3024         node = parse->limitOffset;
3025         if (node)
3026         {
3027                 if (IsA(node, Const))
3028                 {
3029                         /* Treat NULL as no offset; the executor would too */
3030                         if (!((Const *) node)->constisnull)
3031                         {
3032                                 int64           offset = DatumGetInt64(((Const *) node)->constvalue);
3033
3034                                 if (offset != 0)
3035                                         return true;    /* OFFSET with a nonzero value */
3036                         }
3037                 }
3038                 else
3039                         return true;            /* non-constant OFFSET */
3040         }
3041
3042         return false;                           /* don't need a Limit plan node */
3043 }
3044
3045
3046 /*
3047  * remove_useless_groupby_columns
3048  *              Remove any columns in the GROUP BY clause that are redundant due to
3049  *              being functionally dependent on other GROUP BY columns.
3050  *
3051  * Since some other DBMSes do not allow references to ungrouped columns, it's
3052  * not unusual to find all columns listed in GROUP BY even though listing the
3053  * primary-key columns would be sufficient.  Deleting such excess columns
3054  * avoids redundant sorting work, so it's worth doing.  When we do this, we
3055  * must mark the plan as dependent on the pkey constraint (compare the
3056  * parser's check_ungrouped_columns() and check_functional_grouping()).
3057  *
3058  * In principle, we could treat any NOT-NULL columns appearing in a UNIQUE
3059  * index as the determining columns.  But as with check_functional_grouping(),
3060  * there's currently no way to represent dependency on a NOT NULL constraint,
3061  * so we consider only the pkey for now.
3062  */
3063 static void
3064 remove_useless_groupby_columns(PlannerInfo *root)
3065 {
3066         Query      *parse = root->parse;
3067         Bitmapset **groupbyattnos;
3068         Bitmapset **surplusvars;
3069         ListCell   *lc;
3070         int                     relid;
3071
3072         /* No chance to do anything if there are less than two GROUP BY items */
3073         if (list_length(parse->groupClause) < 2)
3074                 return;
3075
3076         /* Don't fiddle with the GROUP BY clause if the query has grouping sets */
3077         if (parse->groupingSets)
3078                 return;
3079
3080         /*
3081          * Scan the GROUP BY clause to find GROUP BY items that are simple Vars.
3082          * Fill groupbyattnos[k] with a bitmapset of the column attnos of RTE k
3083          * that are GROUP BY items.
3084          */
3085         groupbyattnos = (Bitmapset **) palloc0(sizeof(Bitmapset *) *
3086                                                                                    (list_length(parse->rtable) + 1));
3087         foreach(lc, parse->groupClause)
3088         {
3089                 SortGroupClause *sgc = lfirst_node(SortGroupClause, lc);
3090                 TargetEntry *tle = get_sortgroupclause_tle(sgc, parse->targetList);
3091                 Var                *var = (Var *) tle->expr;
3092
3093                 /*
3094                  * Ignore non-Vars and Vars from other query levels.
3095                  *
3096                  * XXX in principle, stable expressions containing Vars could also be
3097                  * removed, if all the Vars are functionally dependent on other GROUP
3098                  * BY items.  But it's not clear that such cases occur often enough to
3099                  * be worth troubling over.
3100                  */
3101                 if (!IsA(var, Var) ||
3102                         var->varlevelsup > 0)
3103                         continue;
3104
3105                 /* OK, remember we have this Var */
3106                 relid = var->varno;
3107                 Assert(relid <= list_length(parse->rtable));
3108                 groupbyattnos[relid] = bms_add_member(groupbyattnos[relid],
3109                                                                                           var->varattno - FirstLowInvalidHeapAttributeNumber);
3110         }
3111
3112         /*
3113          * Consider each relation and see if it is possible to remove some of its
3114          * Vars from GROUP BY.  For simplicity and speed, we do the actual removal
3115          * in a separate pass.  Here, we just fill surplusvars[k] with a bitmapset
3116          * of the column attnos of RTE k that are removable GROUP BY items.
3117          */
3118         surplusvars = NULL;                     /* don't allocate array unless required */
3119         relid = 0;
3120         foreach(lc, parse->rtable)
3121         {
3122                 RangeTblEntry *rte = lfirst_node(RangeTblEntry, lc);
3123                 Bitmapset  *relattnos;
3124                 Bitmapset  *pkattnos;
3125                 Oid                     constraintOid;
3126
3127                 relid++;
3128
3129                 /* Only plain relations could have primary-key constraints */
3130                 if (rte->rtekind != RTE_RELATION)
3131                         continue;
3132
3133                 /*
3134                  * We must skip inheritance parent tables as some of the child rels
3135                  * may cause duplicate rows.  This cannot happen with partitioned
3136                  * tables, however.
3137                  */
3138                 if (rte->inh && rte->relkind != RELKIND_PARTITIONED_TABLE)
3139                         continue;
3140
3141                 /* Nothing to do unless this rel has multiple Vars in GROUP BY */
3142                 relattnos = groupbyattnos[relid];
3143                 if (bms_membership(relattnos) != BMS_MULTIPLE)
3144                         continue;
3145
3146                 /*
3147                  * Can't remove any columns for this rel if there is no suitable
3148                  * (i.e., nondeferrable) primary key constraint.
3149                  */
3150                 pkattnos = get_primary_key_attnos(rte->relid, false, &constraintOid);
3151                 if (pkattnos == NULL)
3152                         continue;
3153
3154                 /*
3155                  * If the primary key is a proper subset of relattnos then we have
3156                  * some items in the GROUP BY that can be removed.
3157                  */
3158                 if (bms_subset_compare(pkattnos, relattnos) == BMS_SUBSET1)
3159                 {
3160                         /*
3161                          * To easily remember whether we've found anything to do, we don't
3162                          * allocate the surplusvars[] array until we find something.
3163                          */
3164                         if (surplusvars == NULL)
3165                                 surplusvars = (Bitmapset **) palloc0(sizeof(Bitmapset *) *
3166                                                                                                          (list_length(parse->rtable) + 1));
3167
3168                         /* Remember the attnos of the removable columns */
3169                         surplusvars[relid] = bms_difference(relattnos, pkattnos);
3170
3171                         /* Also, mark the resulting plan as dependent on this constraint */
3172                         parse->constraintDeps = lappend_oid(parse->constraintDeps,
3173                                                                                                 constraintOid);
3174                 }
3175         }
3176
3177         /*
3178          * If we found any surplus Vars, build a new GROUP BY clause without them.
3179          * (Note: this may leave some TLEs with unreferenced ressortgroupref
3180          * markings, but that's harmless.)
3181          */
3182         if (surplusvars != NULL)
3183         {
3184                 List       *new_groupby = NIL;
3185
3186                 foreach(lc, parse->groupClause)
3187                 {
3188                         SortGroupClause *sgc = lfirst_node(SortGroupClause, lc);
3189                         TargetEntry *tle = get_sortgroupclause_tle(sgc, parse->targetList);
3190                         Var                *var = (Var *) tle->expr;
3191
3192                         /*
3193                          * New list must include non-Vars, outer Vars, and anything not
3194                          * marked as surplus.
3195                          */
3196                         if (!IsA(var, Var) ||
3197                                 var->varlevelsup > 0 ||
3198                                 !bms_is_member(var->varattno - FirstLowInvalidHeapAttributeNumber,
3199                                                            surplusvars[var->varno]))
3200                                 new_groupby = lappend(new_groupby, sgc);
3201                 }
3202
3203                 parse->groupClause = new_groupby;
3204         }
3205 }
3206
3207 /*
3208  * preprocess_groupclause - do preparatory work on GROUP BY clause
3209  *
3210  * The idea here is to adjust the ordering of the GROUP BY elements
3211  * (which in itself is semantically insignificant) to match ORDER BY,
3212  * thereby allowing a single sort operation to both implement the ORDER BY
3213  * requirement and set up for a Unique step that implements GROUP BY.
3214  *
3215  * In principle it might be interesting to consider other orderings of the
3216  * GROUP BY elements, which could match the sort ordering of other
3217  * possible plans (eg an indexscan) and thereby reduce cost.  We don't
3218  * bother with that, though.  Hashed grouping will frequently win anyway.
3219  *
3220  * Note: we need no comparable processing of the distinctClause because
3221  * the parser already enforced that that matches ORDER BY.
3222  *
3223  * For grouping sets, the order of items is instead forced to agree with that
3224  * of the grouping set (and items not in the grouping set are skipped). The
3225  * work of sorting the order of grouping set elements to match the ORDER BY if
3226  * possible is done elsewhere.
3227  */
3228 static List *
3229 preprocess_groupclause(PlannerInfo *root, List *force)
3230 {
3231         Query      *parse = root->parse;
3232         List       *new_groupclause = NIL;
3233         bool            partial_match;
3234         ListCell   *sl;
3235         ListCell   *gl;
3236
3237         /* For grouping sets, we need to force the ordering */
3238         if (force)
3239         {
3240                 foreach(sl, force)
3241                 {
3242                         Index           ref = lfirst_int(sl);
3243                         SortGroupClause *cl = get_sortgroupref_clause(ref, parse->groupClause);
3244
3245                         new_groupclause = lappend(new_groupclause, cl);
3246                 }
3247
3248                 return new_groupclause;
3249         }
3250
3251         /* If no ORDER BY, nothing useful to do here */
3252         if (parse->sortClause == NIL)
3253                 return parse->groupClause;
3254
3255         /*
3256          * Scan the ORDER BY clause and construct a list of matching GROUP BY
3257          * items, but only as far as we can make a matching prefix.
3258          *
3259          * This code assumes that the sortClause contains no duplicate items.
3260          */
3261         foreach(sl, parse->sortClause)
3262         {
3263                 SortGroupClause *sc = lfirst_node(SortGroupClause, sl);
3264
3265                 foreach(gl, parse->groupClause)
3266                 {
3267                         SortGroupClause *gc = lfirst_node(SortGroupClause, gl);
3268
3269                         if (equal(gc, sc))
3270                         {
3271                                 new_groupclause = lappend(new_groupclause, gc);
3272                                 break;
3273                         }
3274                 }
3275                 if (gl == NULL)
3276                         break;                          /* no match, so stop scanning */
3277         }
3278
3279         /* Did we match all of the ORDER BY list, or just some of it? */
3280         partial_match = (sl != NULL);
3281
3282         /* If no match at all, no point in reordering GROUP BY */
3283         if (new_groupclause == NIL)
3284                 return parse->groupClause;
3285
3286         /*
3287          * Add any remaining GROUP BY items to the new list, but only if we were
3288          * able to make a complete match.  In other words, we only rearrange the
3289          * GROUP BY list if the result is that one list is a prefix of the other
3290          * --- otherwise there's no possibility of a common sort.  Also, give up
3291          * if there are any non-sortable GROUP BY items, since then there's no
3292          * hope anyway.
3293          */
3294         foreach(gl, parse->groupClause)
3295         {
3296                 SortGroupClause *gc = lfirst_node(SortGroupClause, gl);
3297
3298                 if (list_member_ptr(new_groupclause, gc))
3299                         continue;                       /* it matched an ORDER BY item */
3300                 if (partial_match)
3301                         return parse->groupClause;      /* give up, no common sort possible */
3302                 if (!OidIsValid(gc->sortop))
3303                         return parse->groupClause;      /* give up, GROUP BY can't be sorted */
3304                 new_groupclause = lappend(new_groupclause, gc);
3305         }
3306
3307         /* Success --- install the rearranged GROUP BY list */
3308         Assert(list_length(parse->groupClause) == list_length(new_groupclause));
3309         return new_groupclause;
3310 }
3311
3312 /*
3313  * Extract lists of grouping sets that can be implemented using a single
3314  * rollup-type aggregate pass each. Returns a list of lists of grouping sets.
3315  *
3316  * Input must be sorted with smallest sets first. Result has each sublist
3317  * sorted with smallest sets first.
3318  *
3319  * We want to produce the absolute minimum possible number of lists here to
3320  * avoid excess sorts. Fortunately, there is an algorithm for this; the problem
3321  * of finding the minimal partition of a partially-ordered set into chains
3322  * (which is what we need, taking the list of grouping sets as a poset ordered
3323  * by set inclusion) can be mapped to the problem of finding the maximum
3324  * cardinality matching on a bipartite graph, which is solvable in polynomial
3325  * time with a worst case of no worse than O(n^2.5) and usually much
3326  * better. Since our N is at most 4096, we don't need to consider fallbacks to
3327  * heuristic or approximate methods.  (Planning time for a 12-d cube is under
3328  * half a second on my modest system even with optimization off and assertions
3329  * on.)
3330  */
3331 static List *
3332 extract_rollup_sets(List *groupingSets)
3333 {
3334         int                     num_sets_raw = list_length(groupingSets);
3335         int                     num_empty = 0;
3336         int                     num_sets = 0;   /* distinct sets */
3337         int                     num_chains = 0;
3338         List       *result = NIL;
3339         List      **results;
3340         List      **orig_sets;
3341         Bitmapset **set_masks;
3342         int                *chains;
3343         short     **adjacency;
3344         short      *adjacency_buf;
3345         BipartiteMatchState *state;
3346         int                     i;
3347         int                     j;
3348         int                     j_size;
3349         ListCell   *lc1 = list_head(groupingSets);
3350         ListCell   *lc;
3351
3352         /*
3353          * Start by stripping out empty sets.  The algorithm doesn't require this,
3354          * but the planner currently needs all empty sets to be returned in the
3355          * first list, so we strip them here and add them back after.
3356          */
3357         while (lc1 && lfirst(lc1) == NIL)
3358         {
3359                 ++num_empty;
3360                 lc1 = lnext(groupingSets, lc1);
3361         }
3362
3363         /* bail out now if it turns out that all we had were empty sets. */
3364         if (!lc1)
3365                 return list_make1(groupingSets);
3366
3367         /*----------
3368          * We don't strictly need to remove duplicate sets here, but if we don't,
3369          * they tend to become scattered through the result, which is a bit
3370          * confusing (and irritating if we ever decide to optimize them out).
3371          * So we remove them here and add them back after.
3372          *
3373          * For each non-duplicate set, we fill in the following:
3374          *
3375          * orig_sets[i] = list of the original set lists
3376          * set_masks[i] = bitmapset for testing inclusion
3377          * adjacency[i] = array [n, v1, v2, ... vn] of adjacency indices
3378          *
3379          * chains[i] will be the result group this set is assigned to.
3380          *
3381          * We index all of these from 1 rather than 0 because it is convenient
3382          * to leave 0 free for the NIL node in the graph algorithm.
3383          *----------
3384          */
3385         orig_sets = palloc0((num_sets_raw + 1) * sizeof(List *));
3386         set_masks = palloc0((num_sets_raw + 1) * sizeof(Bitmapset *));
3387         adjacency = palloc0((num_sets_raw + 1) * sizeof(short *));
3388         adjacency_buf = palloc((num_sets_raw + 1) * sizeof(short));
3389
3390         j_size = 0;
3391         j = 0;
3392         i = 1;
3393
3394         for_each_cell(lc, groupingSets, lc1)
3395         {
3396                 List       *candidate = (List *) lfirst(lc);
3397                 Bitmapset  *candidate_set = NULL;
3398                 ListCell   *lc2;
3399                 int                     dup_of = 0;
3400
3401                 foreach(lc2, candidate)
3402                 {
3403                         candidate_set = bms_add_member(candidate_set, lfirst_int(lc2));
3404                 }
3405
3406                 /* we can only be a dup if we're the same length as a previous set */
3407                 if (j_size == list_length(candidate))
3408                 {
3409                         int                     k;
3410
3411                         for (k = j; k < i; ++k)
3412                         {
3413                                 if (bms_equal(set_masks[k], candidate_set))
3414                                 {
3415                                         dup_of = k;
3416                                         break;
3417                                 }
3418                         }
3419                 }
3420                 else if (j_size < list_length(candidate))
3421                 {
3422                         j_size = list_length(candidate);
3423                         j = i;
3424                 }
3425
3426                 if (dup_of > 0)
3427                 {
3428                         orig_sets[dup_of] = lappend(orig_sets[dup_of], candidate);
3429                         bms_free(candidate_set);
3430                 }
3431                 else
3432                 {
3433                         int                     k;
3434                         int                     n_adj = 0;
3435
3436                         orig_sets[i] = list_make1(candidate);
3437                         set_masks[i] = candidate_set;
3438
3439                         /* fill in adjacency list; no need to compare equal-size sets */
3440
3441                         for (k = j - 1; k > 0; --k)
3442                         {
3443                                 if (bms_is_subset(set_masks[k], candidate_set))
3444                                         adjacency_buf[++n_adj] = k;
3445                         }
3446
3447                         if (n_adj > 0)
3448                         {
3449                                 adjacency_buf[0] = n_adj;
3450                                 adjacency[i] = palloc((n_adj + 1) * sizeof(short));
3451                                 memcpy(adjacency[i], adjacency_buf, (n_adj + 1) * sizeof(short));
3452                         }
3453                         else
3454                                 adjacency[i] = NULL;
3455
3456                         ++i;
3457                 }
3458         }
3459
3460         num_sets = i - 1;
3461
3462         /*
3463          * Apply the graph matching algorithm to do the work.
3464          */
3465         state = BipartiteMatch(num_sets, num_sets, adjacency);
3466
3467         /*
3468          * Now, the state->pair* fields have the info we need to assign sets to
3469          * chains. Two sets (u,v) belong to the same chain if pair_uv[u] = v or
3470          * pair_vu[v] = u (both will be true, but we check both so that we can do
3471          * it in one pass)
3472          */
3473         chains = palloc0((num_sets + 1) * sizeof(int));
3474
3475         for (i = 1; i <= num_sets; ++i)
3476         {
3477                 int                     u = state->pair_vu[i];
3478                 int                     v = state->pair_uv[i];
3479
3480                 if (u > 0 && u < i)
3481                         chains[i] = chains[u];
3482                 else if (v > 0 && v < i)
3483                         chains[i] = chains[v];
3484                 else
3485                         chains[i] = ++num_chains;
3486         }
3487
3488         /* build result lists. */
3489         results = palloc0((num_chains + 1) * sizeof(List *));
3490
3491         for (i = 1; i <= num_sets; ++i)
3492         {
3493                 int                     c = chains[i];
3494
3495                 Assert(c > 0);
3496
3497                 results[c] = list_concat(results[c], orig_sets[i]);
3498         }
3499
3500         /* push any empty sets back on the first list. */
3501         while (num_empty-- > 0)
3502                 results[1] = lcons(NIL, results[1]);
3503
3504         /* make result list */
3505         for (i = 1; i <= num_chains; ++i)
3506                 result = lappend(result, results[i]);
3507
3508         /*
3509          * Free all the things.
3510          *
3511          * (This is over-fussy for small sets but for large sets we could have
3512          * tied up a nontrivial amount of memory.)
3513          */
3514         BipartiteMatchFree(state);
3515         pfree(results);
3516         pfree(chains);
3517         for (i = 1; i <= num_sets; ++i)
3518                 if (adjacency[i])
3519                         pfree(adjacency[i]);
3520         pfree(adjacency);
3521         pfree(adjacency_buf);
3522         pfree(orig_sets);
3523         for (i = 1; i <= num_sets; ++i)
3524                 bms_free(set_masks[i]);
3525         pfree(set_masks);
3526
3527         return result;
3528 }
3529
3530 /*
3531  * Reorder the elements of a list of grouping sets such that they have correct
3532  * prefix relationships. Also inserts the GroupingSetData annotations.
3533  *
3534  * The input must be ordered with smallest sets first; the result is returned
3535  * with largest sets first.  Note that the result shares no list substructure
3536  * with the input, so it's safe for the caller to modify it later.
3537  *
3538  * If we're passed in a sortclause, we follow its order of columns to the
3539  * extent possible, to minimize the chance that we add unnecessary sorts.
3540  * (We're trying here to ensure that GROUPING SETS ((a,b,c),(c)) ORDER BY c,b,a
3541  * gets implemented in one pass.)
3542  */
3543 static List *
3544 reorder_grouping_sets(List *groupingsets, List *sortclause)
3545 {
3546         ListCell   *lc;
3547         List       *previous = NIL;
3548         List       *result = NIL;
3549
3550         foreach(lc, groupingsets)
3551         {
3552                 List       *candidate = (List *) lfirst(lc);
3553                 List       *new_elems = list_difference_int(candidate, previous);
3554                 GroupingSetData *gs = makeNode(GroupingSetData);
3555
3556                 while (list_length(sortclause) > list_length(previous) &&
3557                            list_length(new_elems) > 0)
3558                 {
3559                         SortGroupClause *sc = list_nth(sortclause, list_length(previous));
3560                         int                     ref = sc->tleSortGroupRef;
3561
3562                         if (list_member_int(new_elems, ref))
3563                         {
3564                                 previous = lappend_int(previous, ref);
3565                                 new_elems = list_delete_int(new_elems, ref);
3566                         }
3567                         else
3568                         {
3569                                 /* diverged from the sortclause; give up on it */
3570                                 sortclause = NIL;
3571                                 break;
3572                         }
3573                 }
3574
3575                 /*
3576                  * Safe to use list_concat (which shares cells of the second arg)
3577                  * because we know that new_elems does not share cells with anything.
3578                  */
3579                 previous = list_concat(previous, new_elems);
3580
3581                 gs->set = list_copy(previous);
3582                 result = lcons(gs, result);
3583         }
3584
3585         list_free(previous);
3586
3587         return result;
3588 }
3589
3590 /*
3591  * Compute query_pathkeys and other pathkeys during plan generation
3592  */
3593 static void
3594 standard_qp_callback(PlannerInfo *root, void *extra)
3595 {
3596         Query      *parse = root->parse;
3597         standard_qp_extra *qp_extra = (standard_qp_extra *) extra;
3598         List       *tlist = root->processed_tlist;
3599         List       *activeWindows = qp_extra->activeWindows;
3600
3601         /*
3602          * Calculate pathkeys that represent grouping/ordering requirements.  The
3603          * sortClause is certainly sort-able, but GROUP BY and DISTINCT might not
3604          * be, in which case we just leave their pathkeys empty.
3605          */
3606         if (qp_extra->groupClause &&
3607                 grouping_is_sortable(qp_extra->groupClause))
3608                 root->group_pathkeys =
3609                         make_pathkeys_for_sortclauses(root,
3610                                                                                   qp_extra->groupClause,
3611                                                                                   tlist);
3612         else
3613                 root->group_pathkeys = NIL;
3614
3615         /* We consider only the first (bottom) window in pathkeys logic */
3616         if (activeWindows != NIL)
3617         {
3618                 WindowClause *wc = linitial_node(WindowClause, activeWindows);
3619
3620                 root->window_pathkeys = make_pathkeys_for_window(root,
3621                                                                                                                  wc,
3622                                                                                                                  tlist);
3623         }
3624         else
3625                 root->window_pathkeys = NIL;
3626
3627         if (parse->distinctClause &&
3628                 grouping_is_sortable(parse->distinctClause))
3629                 root->distinct_pathkeys =
3630                         make_pathkeys_for_sortclauses(root,
3631                                                                                   parse->distinctClause,
3632                                                                                   tlist);
3633         else
3634                 root->distinct_pathkeys = NIL;
3635
3636         root->sort_pathkeys =
3637                 make_pathkeys_for_sortclauses(root,
3638                                                                           parse->sortClause,
3639                                                                           tlist);
3640
3641         /*
3642          * Figure out whether we want a sorted result from query_planner.
3643          *
3644          * If we have a sortable GROUP BY clause, then we want a result sorted
3645          * properly for grouping.  Otherwise, if we have window functions to
3646          * evaluate, we try to sort for the first window.  Otherwise, if there's a
3647          * sortable DISTINCT clause that's more rigorous than the ORDER BY clause,
3648          * we try to produce output that's sufficiently well sorted for the
3649          * DISTINCT.  Otherwise, if there is an ORDER BY clause, we want to sort
3650          * by the ORDER BY clause.
3651          *
3652          * Note: if we have both ORDER BY and GROUP BY, and ORDER BY is a superset
3653          * of GROUP BY, it would be tempting to request sort by ORDER BY --- but
3654          * that might just leave us failing to exploit an available sort order at
3655          * all.  Needs more thought.  The choice for DISTINCT versus ORDER BY is
3656          * much easier, since we know that the parser ensured that one is a
3657          * superset of the other.
3658          */
3659         if (root->group_pathkeys)
3660                 root->query_pathkeys = root->group_pathkeys;
3661         else if (root->window_pathkeys)
3662                 root->query_pathkeys = root->window_pathkeys;
3663         else if (list_length(root->distinct_pathkeys) >
3664                          list_length(root->sort_pathkeys))
3665                 root->query_pathkeys = root->distinct_pathkeys;
3666         else if (root->sort_pathkeys)
3667                 root->query_pathkeys = root->sort_pathkeys;
3668         else
3669                 root->query_pathkeys = NIL;
3670 }
3671
3672 /*
3673  * Estimate number of groups produced by grouping clauses (1 if not grouping)
3674  *
3675  * path_rows: number of output rows from scan/join step
3676  * gd: grouping sets data including list of grouping sets and their clauses
3677  * target_list: target list containing group clause references
3678  *
3679  * If doing grouping sets, we also annotate the gsets data with the estimates
3680  * for each set and each individual rollup list, with a view to later
3681  * determining whether some combination of them could be hashed instead.
3682  */
3683 static double
3684 get_number_of_groups(PlannerInfo *root,
3685                                          double path_rows,
3686                                          grouping_sets_data *gd,
3687                                          List *target_list)
3688 {
3689         Query      *parse = root->parse;
3690         double          dNumGroups;
3691
3692         if (parse->groupClause)
3693         {
3694                 List       *groupExprs;
3695
3696                 if (parse->groupingSets)
3697                 {
3698                         /* Add up the estimates for each grouping set */
3699                         ListCell   *lc;
3700                         ListCell   *lc2;
3701
3702                         Assert(gd);                     /* keep Coverity happy */
3703
3704                         dNumGroups = 0;
3705
3706                         foreach(lc, gd->rollups)
3707                         {
3708                                 RollupData *rollup = lfirst_node(RollupData, lc);
3709                                 ListCell   *lc;
3710
3711                                 groupExprs = get_sortgrouplist_exprs(rollup->groupClause,
3712                                                                                                          target_list);
3713
3714                                 rollup->numGroups = 0.0;
3715
3716                                 forboth(lc, rollup->gsets, lc2, rollup->gsets_data)
3717                                 {
3718                                         List       *gset = (List *) lfirst(lc);
3719                                         GroupingSetData *gs = lfirst_node(GroupingSetData, lc2);
3720                                         double          numGroups = estimate_num_groups(root,
3721                                                                                                                                 groupExprs,
3722                                                                                                                                 path_rows,
3723                                                                                                                                 &gset);
3724
3725                                         gs->numGroups = numGroups;
3726                                         rollup->numGroups += numGroups;
3727                                 }
3728
3729                                 dNumGroups += rollup->numGroups;
3730                         }
3731
3732                         if (gd->hash_sets_idx)
3733                         {
3734                                 ListCell   *lc;
3735
3736                                 gd->dNumHashGroups = 0;
3737
3738                                 groupExprs = get_sortgrouplist_exprs(parse->groupClause,
3739                                                                                                          target_list);
3740
3741                                 forboth(lc, gd->hash_sets_idx, lc2, gd->unsortable_sets)
3742                                 {
3743                                         List       *gset = (List *) lfirst(lc);
3744                                         GroupingSetData *gs = lfirst_node(GroupingSetData, lc2);
3745                                         double          numGroups = estimate_num_groups(root,
3746                                                                                                                                 groupExprs,
3747                                                                                                                                 path_rows,
3748                                                                                                                                 &gset);
3749
3750                                         gs->numGroups = numGroups;
3751                                         gd->dNumHashGroups += numGroups;
3752                                 }
3753
3754                                 dNumGroups += gd->dNumHashGroups;
3755                         }
3756                 }
3757                 else
3758                 {
3759                         /* Plain GROUP BY */
3760                         groupExprs = get_sortgrouplist_exprs(parse->groupClause,
3761                                                                                                  target_list);
3762
3763                         dNumGroups = estimate_num_groups(root, groupExprs, path_rows,
3764                                                                                          NULL);
3765                 }
3766         }
3767         else if (parse->groupingSets)
3768         {
3769                 /* Empty grouping sets ... one result row for each one */
3770                 dNumGroups = list_length(parse->groupingSets);
3771         }
3772         else if (parse->hasAggs || root->hasHavingQual)
3773         {
3774                 /* Plain aggregation, one result row */
3775                 dNumGroups = 1;
3776         }
3777         else
3778         {
3779                 /* Not grouping */
3780                 dNumGroups = 1;
3781         }
3782
3783         return dNumGroups;
3784 }
3785
3786 /*
3787  * create_grouping_paths
3788  *
3789  * Build a new upperrel containing Paths for grouping and/or aggregation.
3790  * Along the way, we also build an upperrel for Paths which are partially
3791  * grouped and/or aggregated.  A partially grouped and/or aggregated path
3792  * needs a FinalizeAggregate node to complete the aggregation.  Currently,
3793  * the only partially grouped paths we build are also partial paths; that
3794  * is, they need a Gather and then a FinalizeAggregate.
3795  *
3796  * input_rel: contains the source-data Paths
3797  * target: the pathtarget for the result Paths to compute
3798  * agg_costs: cost info about all aggregates in query (in AGGSPLIT_SIMPLE mode)
3799  * gd: grouping sets data including list of grouping sets and their clauses
3800  *
3801  * Note: all Paths in input_rel are expected to return the target computed
3802  * by make_group_input_target.
3803  */
3804 static RelOptInfo *
3805 create_grouping_paths(PlannerInfo *root,
3806                                           RelOptInfo *input_rel,
3807                                           PathTarget *target,
3808                                           bool target_parallel_safe,
3809                                           const AggClauseCosts *agg_costs,
3810                                           grouping_sets_data *gd)
3811 {
3812         Query      *parse = root->parse;
3813         RelOptInfo *grouped_rel;
3814         RelOptInfo *partially_grouped_rel;
3815
3816         /*
3817          * Create grouping relation to hold fully aggregated grouping and/or
3818          * aggregation paths.
3819          */
3820         grouped_rel = make_grouping_rel(root, input_rel, target,
3821                                                                         target_parallel_safe, parse->havingQual);
3822
3823         /*
3824          * Create either paths for a degenerate grouping or paths for ordinary
3825          * grouping, as appropriate.
3826          */
3827         if (is_degenerate_grouping(root))
3828                 create_degenerate_grouping_paths(root, input_rel, grouped_rel);
3829         else
3830         {
3831                 int                     flags = 0;
3832                 GroupPathExtraData extra;
3833
3834                 /*
3835                  * Determine whether it's possible to perform sort-based
3836                  * implementations of grouping.  (Note that if groupClause is empty,
3837                  * grouping_is_sortable() is trivially true, and all the
3838                  * pathkeys_contained_in() tests will succeed too, so that we'll
3839                  * consider every surviving input path.)
3840                  *
3841                  * If we have grouping sets, we might be able to sort some but not all
3842                  * of them; in this case, we need can_sort to be true as long as we
3843                  * must consider any sorted-input plan.
3844                  */
3845                 if ((gd && gd->rollups != NIL)
3846                         || grouping_is_sortable(parse->groupClause))
3847                         flags |= GROUPING_CAN_USE_SORT;
3848
3849                 /*
3850                  * Determine whether we should consider hash-based implementations of
3851                  * grouping.
3852                  *
3853                  * Hashed aggregation only applies if we're grouping. If we have
3854                  * grouping sets, some groups might be hashable but others not; in
3855                  * this case we set can_hash true as long as there is nothing globally
3856                  * preventing us from hashing (and we should therefore consider plans
3857                  * with hashes).
3858                  *
3859                  * Executor doesn't support hashed aggregation with DISTINCT or ORDER
3860                  * BY aggregates.  (Doing so would imply storing *all* the input
3861                  * values in the hash table, and/or running many sorts in parallel,
3862                  * either of which seems like a certain loser.)  We similarly don't
3863                  * support ordered-set aggregates in hashed aggregation, but that case
3864                  * is also included in the numOrderedAggs count.
3865                  *
3866                  * Note: grouping_is_hashable() is much more expensive to check than
3867                  * the other gating conditions, so we want to do it last.
3868                  */
3869                 if ((parse->groupClause != NIL &&
3870                          agg_costs->numOrderedAggs == 0 &&
3871                          (gd ? gd->any_hashable : grouping_is_hashable(parse->groupClause))))
3872                         flags |= GROUPING_CAN_USE_HASH;
3873
3874                 /*
3875                  * Determine whether partial aggregation is possible.
3876                  */
3877                 if (can_partial_agg(root, agg_costs))
3878                         flags |= GROUPING_CAN_PARTIAL_AGG;
3879
3880                 extra.flags = flags;
3881                 extra.target_parallel_safe = target_parallel_safe;
3882                 extra.havingQual = parse->havingQual;
3883                 extra.targetList = parse->targetList;
3884                 extra.partial_costs_set = false;
3885
3886                 /*
3887                  * Determine whether partitionwise aggregation is in theory possible.
3888                  * It can be disabled by the user, and for now, we don't try to
3889                  * support grouping sets.  create_ordinary_grouping_paths() will check
3890                  * additional conditions, such as whether input_rel is partitioned.
3891                  */
3892                 if (enable_partitionwise_aggregate && !parse->groupingSets)
3893                         extra.patype = PARTITIONWISE_AGGREGATE_FULL;
3894                 else
3895                         extra.patype = PARTITIONWISE_AGGREGATE_NONE;
3896
3897                 create_ordinary_grouping_paths(root, input_rel, grouped_rel,
3898                                                                            agg_costs, gd, &extra,
3899                                                                            &partially_grouped_rel);
3900         }
3901
3902         set_cheapest(grouped_rel);
3903         return grouped_rel;
3904 }
3905
3906 /*
3907  * make_grouping_rel
3908  *
3909  * Create a new grouping rel and set basic properties.
3910  *
3911  * input_rel represents the underlying scan/join relation.
3912  * target is the output expected from the grouping relation.
3913  */
3914 static RelOptInfo *
3915 make_grouping_rel(PlannerInfo *root, RelOptInfo *input_rel,
3916                                   PathTarget *target, bool target_parallel_safe,
3917                                   Node *havingQual)
3918 {
3919         RelOptInfo *grouped_rel;
3920
3921         if (IS_OTHER_REL(input_rel))
3922         {
3923                 grouped_rel = fetch_upper_rel(root, UPPERREL_GROUP_AGG,
3924                                                                           input_rel->relids);
3925                 grouped_rel->reloptkind = RELOPT_OTHER_UPPER_REL;
3926         }
3927         else
3928         {
3929                 /*
3930                  * By tradition, the relids set for the main grouping relation is
3931                  * NULL.  (This could be changed, but might require adjustments
3932                  * elsewhere.)
3933                  */
3934                 grouped_rel = fetch_upper_rel(root, UPPERREL_GROUP_AGG, NULL);
3935         }
3936
3937         /* Set target. */
3938         grouped_rel->reltarget = target;
3939
3940         /*
3941          * If the input relation is not parallel-safe, then the grouped relation
3942          * can't be parallel-safe, either.  Otherwise, it's parallel-safe if the
3943          * target list and HAVING quals are parallel-safe.
3944          */
3945         if (input_rel->consider_parallel && target_parallel_safe &&
3946                 is_parallel_safe(root, (Node *) havingQual))
3947                 grouped_rel->consider_parallel = true;
3948
3949         /*
3950          * If the input rel belongs to a single FDW, so does the grouped rel.
3951          */
3952         grouped_rel->serverid = input_rel->serverid;
3953         grouped_rel->userid = input_rel->userid;
3954         grouped_rel->useridiscurrent = input_rel->useridiscurrent;
3955         grouped_rel->fdwroutine = input_rel->fdwroutine;
3956
3957         return grouped_rel;
3958 }
3959
3960 /*
3961  * is_degenerate_grouping
3962  *
3963  * A degenerate grouping is one in which the query has a HAVING qual and/or
3964  * grouping sets, but no aggregates and no GROUP BY (which implies that the
3965  * grouping sets are all empty).
3966  */
3967 static bool
3968 is_degenerate_grouping(PlannerInfo *root)
3969 {
3970         Query      *parse = root->parse;
3971
3972         return (root->hasHavingQual || parse->groupingSets) &&
3973                 !parse->hasAggs && parse->groupClause == NIL;
3974 }
3975
3976 /*
3977  * create_degenerate_grouping_paths
3978  *
3979  * When the grouping is degenerate (see is_degenerate_grouping), we are
3980  * supposed to emit either zero or one row for each grouping set depending on
3981  * whether HAVING succeeds.  Furthermore, there cannot be any variables in
3982  * either HAVING or the targetlist, so we actually do not need the FROM table
3983  * at all! We can just throw away the plan-so-far and generate a Result node.
3984  * This is a sufficiently unusual corner case that it's not worth contorting
3985  * the structure of this module to avoid having to generate the earlier paths
3986  * in the first place.
3987  */
3988 static void
3989 create_degenerate_grouping_paths(PlannerInfo *root, RelOptInfo *input_rel,
3990                                                                  RelOptInfo *grouped_rel)
3991 {
3992         Query      *parse = root->parse;
3993         int                     nrows;
3994         Path       *path;
3995
3996         nrows = list_length(parse->groupingSets);
3997         if (nrows > 1)
3998         {
3999                 /*
4000                  * Doesn't seem worthwhile writing code to cons up a generate_series
4001                  * or a values scan to emit multiple rows. Instead just make N clones
4002                  * and append them.  (With a volatile HAVING clause, this means you
4003                  * might get between 0 and N output rows. Offhand I think that's
4004                  * desired.)
4005                  */
4006                 List       *paths = NIL;
4007
4008                 while (--nrows >= 0)
4009                 {
4010                         path = (Path *)
4011                                 create_group_result_path(root, grouped_rel,
4012                                                                                  grouped_rel->reltarget,
4013                                                                                  (List *) parse->havingQual);
4014                         paths = lappend(paths, path);
4015                 }
4016                 path = (Path *)
4017                         create_append_path(root,
4018                                                            grouped_rel,
4019                                                            paths,
4020                                                            NIL,
4021                                                            NIL,
4022                                                            NULL,
4023                                                            0,
4024                                                            false,
4025                                                            NIL,
4026                                                            -1);
4027         }
4028         else
4029         {
4030                 /* No grouping sets, or just one, so one output row */
4031                 path = (Path *)
4032                         create_group_result_path(root, grouped_rel,
4033                                                                          grouped_rel->reltarget,
4034                                                                          (List *) parse->havingQual);
4035         }
4036
4037         add_path(grouped_rel, path);
4038 }
4039
4040 /*
4041  * create_ordinary_grouping_paths
4042  *
4043  * Create grouping paths for the ordinary (that is, non-degenerate) case.
4044  *
4045  * We need to consider sorted and hashed aggregation in the same function,
4046  * because otherwise (1) it would be harder to throw an appropriate error
4047  * message if neither way works, and (2) we should not allow hashtable size
4048  * considerations to dissuade us from using hashing if sorting is not possible.
4049  *
4050  * *partially_grouped_rel_p will be set to the partially grouped rel which this
4051  * function creates, or to NULL if it doesn't create one.
4052  */
4053 static void
4054 create_ordinary_grouping_paths(PlannerInfo *root, RelOptInfo *input_rel,
4055                                                            RelOptInfo *grouped_rel,
4056                                                            const AggClauseCosts *agg_costs,
4057                                                            grouping_sets_data *gd,
4058                                                            GroupPathExtraData *extra,
4059                                                            RelOptInfo **partially_grouped_rel_p)
4060 {
4061         Path       *cheapest_path = input_rel->cheapest_total_path;
4062         RelOptInfo *partially_grouped_rel = NULL;
4063         double          dNumGroups;
4064         PartitionwiseAggregateType patype = PARTITIONWISE_AGGREGATE_NONE;
4065
4066         /*
4067          * If this is the topmost grouping relation or if the parent relation is
4068          * doing some form of partitionwise aggregation, then we may be able to do
4069          * it at this level also.  However, if the input relation is not
4070          * partitioned, partitionwise aggregate is impossible.
4071          */
4072         if (extra->patype != PARTITIONWISE_AGGREGATE_NONE &&
4073                 IS_PARTITIONED_REL(input_rel))
4074         {
4075                 /*
4076                  * If this is the topmost relation or if the parent relation is doing
4077                  * full partitionwise aggregation, then we can do full partitionwise
4078                  * aggregation provided that the GROUP BY clause contains all of the
4079                  * partitioning columns at this level.  Otherwise, we can do at most
4080                  * partial partitionwise aggregation.  But if partial aggregation is
4081                  * not supported in general then we can't use it for partitionwise
4082                  * aggregation either.
4083                  */
4084                 if (extra->patype == PARTITIONWISE_AGGREGATE_FULL &&
4085                         group_by_has_partkey(input_rel, extra->targetList,
4086                                                                  root->parse->groupClause))
4087                         patype = PARTITIONWISE_AGGREGATE_FULL;
4088                 else if ((extra->flags & GROUPING_CAN_PARTIAL_AGG) != 0)
4089                         patype = PARTITIONWISE_AGGREGATE_PARTIAL;
4090                 else
4091                         patype = PARTITIONWISE_AGGREGATE_NONE;
4092         }
4093
4094         /*
4095          * Before generating paths for grouped_rel, we first generate any possible
4096          * partially grouped paths; that way, later code can easily consider both
4097          * parallel and non-parallel approaches to grouping.
4098          */
4099         if ((extra->flags & GROUPING_CAN_PARTIAL_AGG) != 0)
4100         {
4101                 bool            force_rel_creation;
4102
4103                 /*
4104                  * If we're doing partitionwise aggregation at this level, force
4105                  * creation of a partially_grouped_rel so we can add partitionwise
4106                  * paths to it.
4107                  */
4108                 force_rel_creation = (patype == PARTITIONWISE_AGGREGATE_PARTIAL);
4109
4110                 partially_grouped_rel =
4111                         create_partial_grouping_paths(root,
4112                                                                                   grouped_rel,
4113                                                                                   input_rel,
4114                                                                                   gd,
4115                                                                                   extra,
4116                                                                                   force_rel_creation);
4117         }
4118
4119         /* Set out parameter. */
4120         *partially_grouped_rel_p = partially_grouped_rel;
4121
4122         /* Apply partitionwise aggregation technique, if possible. */
4123         if (patype != PARTITIONWISE_AGGREGATE_NONE)
4124                 create_partitionwise_grouping_paths(root, input_rel, grouped_rel,
4125                                                                                         partially_grouped_rel, agg_costs,
4126                                                                                         gd, patype, extra);
4127
4128         /* If we are doing partial aggregation only, return. */
4129         if (extra->patype == PARTITIONWISE_AGGREGATE_PARTIAL)
4130         {
4131                 Assert(partially_grouped_rel);
4132
4133                 if (partially_grouped_rel->pathlist)
4134                         set_cheapest(partially_grouped_rel);
4135
4136                 return;
4137         }
4138
4139         /* Gather any partially grouped partial paths. */
4140         if (partially_grouped_rel && partially_grouped_rel->partial_pathlist)
4141         {
4142                 gather_grouping_paths(root, partially_grouped_rel);
4143                 set_cheapest(partially_grouped_rel);
4144         }
4145
4146         /*
4147          * Estimate number of groups.
4148          */
4149         dNumGroups = get_number_of_groups(root,
4150                                                                           cheapest_path->rows,
4151                                                                           gd,
4152                                                                           extra->targetList);
4153
4154         /* Build final grouping paths */
4155         add_paths_to_grouping_rel(root, input_rel, grouped_rel,
4156                                                           partially_grouped_rel, agg_costs, gd,
4157                                                           dNumGroups, extra);
4158
4159         /* Give a helpful error if we failed to find any implementation */
4160         if (grouped_rel->pathlist == NIL)
4161                 ereport(ERROR,
4162                                 (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
4163                                  errmsg("could not implement GROUP BY"),
4164                                  errdetail("Some of the datatypes only support hashing, while others only support sorting.")));
4165
4166         /*
4167          * If there is an FDW that's responsible for all baserels of the query,
4168          * let it consider adding ForeignPaths.
4169          */
4170         if (grouped_rel->fdwroutine &&
4171                 grouped_rel->fdwroutine->GetForeignUpperPaths)
4172                 grouped_rel->fdwroutine->GetForeignUpperPaths(root, UPPERREL_GROUP_AGG,
4173                                                                                                           input_rel, grouped_rel,
4174                                                                                                           extra);
4175
4176         /* Let extensions possibly add some more paths */
4177         if (create_upper_paths_hook)
4178                 (*create_upper_paths_hook) (root, UPPERREL_GROUP_AGG,
4179                                                                         input_rel, grouped_rel,
4180                                                                         extra);
4181 }
4182
4183 /*
4184  * For a given input path, consider the possible ways of doing grouping sets on
4185  * it, by combinations of hashing and sorting.  This can be called multiple
4186  * times, so it's important that it not scribble on input.  No result is
4187  * returned, but any generated paths are added to grouped_rel.
4188  */
4189 static void
4190 consider_groupingsets_paths(PlannerInfo *root,
4191                                                         RelOptInfo *grouped_rel,
4192                                                         Path *path,
4193                                                         bool is_sorted,
4194                                                         bool can_hash,
4195                                                         grouping_sets_data *gd,
4196                                                         const AggClauseCosts *agg_costs,
4197                                                         double dNumGroups)
4198 {
4199         Query      *parse = root->parse;
4200
4201         /*
4202          * If we're not being offered sorted input, then only consider plans that
4203          * can be done entirely by hashing.
4204          *
4205          * We can hash everything if it looks like it'll fit in work_mem. But if
4206          * the input is actually sorted despite not being advertised as such, we
4207          * prefer to make use of that in order to use less memory.
4208          *
4209          * If none of the grouping sets are sortable, then ignore the work_mem
4210          * limit and generate a path anyway, since otherwise we'll just fail.
4211          */
4212         if (!is_sorted)
4213         {
4214                 List       *new_rollups = NIL;
4215                 RollupData *unhashed_rollup = NULL;
4216                 List       *sets_data;
4217                 List       *empty_sets_data = NIL;
4218                 List       *empty_sets = NIL;
4219                 ListCell   *lc;
4220                 ListCell   *l_start = list_head(gd->rollups);
4221                 AggStrategy strat = AGG_HASHED;
4222                 double          hashsize;
4223                 double          exclude_groups = 0.0;
4224
4225                 Assert(can_hash);
4226
4227                 /*
4228                  * If the input is coincidentally sorted usefully (which can happen
4229                  * even if is_sorted is false, since that only means that our caller
4230                  * has set up the sorting for us), then save some hashtable space by
4231                  * making use of that. But we need to watch out for degenerate cases:
4232                  *
4233                  * 1) If there are any empty grouping sets, then group_pathkeys might
4234                  * be NIL if all non-empty grouping sets are unsortable. In this case,
4235                  * there will be a rollup containing only empty groups, and the
4236                  * pathkeys_contained_in test is vacuously true; this is ok.
4237                  *
4238                  * XXX: the above relies on the fact that group_pathkeys is generated
4239                  * from the first rollup. If we add the ability to consider multiple
4240                  * sort orders for grouping input, this assumption might fail.
4241                  *
4242                  * 2) If there are no empty sets and only unsortable sets, then the
4243                  * rollups list will be empty (and thus l_start == NULL), and
4244                  * group_pathkeys will be NIL; we must ensure that the vacuously-true
4245                  * pathkeys_contain_in test doesn't cause us to crash.
4246                  */
4247                 if (l_start != NULL &&
4248                         pathkeys_contained_in(root->group_pathkeys, path->pathkeys))
4249                 {
4250                         unhashed_rollup = lfirst_node(RollupData, l_start);
4251                         exclude_groups = unhashed_rollup->numGroups;
4252                         l_start = lnext(gd->rollups, l_start);
4253                 }
4254
4255                 hashsize = estimate_hashagg_tablesize(path,
4256                                                                                           agg_costs,
4257                                                                                           dNumGroups - exclude_groups);
4258
4259                 /*
4260                  * gd->rollups is empty if we have only unsortable columns to work
4261                  * with.  Override work_mem in that case; otherwise, we'll rely on the
4262                  * sorted-input case to generate usable mixed paths.
4263                  */
4264                 if (hashsize > work_mem * 1024L && gd->rollups)
4265                         return;                         /* nope, won't fit */
4266
4267                 /*
4268                  * We need to burst the existing rollups list into individual grouping
4269                  * sets and recompute a groupClause for each set.
4270                  */
4271                 sets_data = list_copy(gd->unsortable_sets);
4272
4273                 for_each_cell(lc, gd->rollups, l_start)
4274                 {
4275                         RollupData *rollup = lfirst_node(RollupData, lc);
4276
4277                         /*
4278                          * If we find an unhashable rollup that's not been skipped by the
4279                          * "actually sorted" check above, we can't cope; we'd need sorted
4280                          * input (with a different sort order) but we can't get that here.
4281                          * So bail out; we'll get a valid path from the is_sorted case
4282                          * instead.
4283                          *
4284                          * The mere presence of empty grouping sets doesn't make a rollup
4285                          * unhashable (see preprocess_grouping_sets), we handle those
4286                          * specially below.
4287                          */
4288                         if (!rollup->hashable)
4289                                 return;
4290                         else
4291                                 sets_data = list_concat(sets_data, list_copy(rollup->gsets_data));
4292                 }
4293                 foreach(lc, sets_data)
4294                 {
4295                         GroupingSetData *gs = lfirst_node(GroupingSetData, lc);
4296                         List       *gset = gs->set;
4297                         RollupData *rollup;
4298
4299                         if (gset == NIL)
4300                         {
4301                                 /* Empty grouping sets can't be hashed. */
4302                                 empty_sets_data = lappend(empty_sets_data, gs);
4303                                 empty_sets = lappend(empty_sets, NIL);
4304                         }
4305                         else
4306                         {
4307                                 rollup = makeNode(RollupData);
4308
4309                                 rollup->groupClause = preprocess_groupclause(root, gset);
4310                                 rollup->gsets_data = list_make1(gs);
4311                                 rollup->gsets = remap_to_groupclause_idx(rollup->groupClause,
4312                                                                                                                  rollup->gsets_data,
4313                                                                                                                  gd->tleref_to_colnum_map);
4314                                 rollup->numGroups = gs->numGroups;
4315                                 rollup->hashable = true;
4316                                 rollup->is_hashed = true;
4317                                 new_rollups = lappend(new_rollups, rollup);
4318                         }
4319                 }
4320
4321                 /*
4322                  * If we didn't find anything nonempty to hash, then bail.  We'll
4323                  * generate a path from the is_sorted case.
4324                  */
4325                 if (new_rollups == NIL)
4326                         return;
4327
4328                 /*
4329                  * If there were empty grouping sets they should have been in the
4330                  * first rollup.
4331                  */
4332                 Assert(!unhashed_rollup || !empty_sets);
4333
4334                 if (unhashed_rollup)
4335                 {
4336                         new_rollups = lappend(new_rollups, unhashed_rollup);
4337                         strat = AGG_MIXED;
4338                 }
4339                 else if (empty_sets)
4340                 {
4341                         RollupData *rollup = makeNode(RollupData);
4342
4343                         rollup->groupClause = NIL;
4344                         rollup->gsets_data = empty_sets_data;
4345                         rollup->gsets = empty_sets;
4346                         rollup->numGroups = list_length(empty_sets);
4347                         rollup->hashable = false;
4348                         rollup->is_hashed = false;
4349                         new_rollups = lappend(new_rollups, rollup);
4350                         strat = AGG_MIXED;
4351                 }
4352
4353                 add_path(grouped_rel, (Path *)
4354                                  create_groupingsets_path(root,
4355                                                                                   grouped_rel,
4356                                                                                   path,
4357                                                                                   (List *) parse->havingQual,
4358                                                                                   strat,
4359                                                                                   new_rollups,
4360                                                                                   agg_costs,
4361                                                                                   dNumGroups));
4362                 return;
4363         }
4364
4365         /*
4366          * If we have sorted input but nothing we can do with it, bail.
4367          */
4368         if (list_length(gd->rollups) == 0)
4369                 return;
4370
4371         /*
4372          * Given sorted input, we try and make two paths: one sorted and one mixed
4373          * sort/hash. (We need to try both because hashagg might be disabled, or
4374          * some columns might not be sortable.)
4375          *
4376          * can_hash is passed in as false if some obstacle elsewhere (such as
4377          * ordered aggs) means that we shouldn't consider hashing at all.
4378          */
4379         if (can_hash && gd->any_hashable)
4380         {
4381                 List       *rollups = NIL;
4382                 List       *hash_sets = list_copy(gd->unsortable_sets);
4383                 double          availspace = (work_mem * 1024.0);
4384                 ListCell   *lc;
4385
4386                 /*
4387                  * Account first for space needed for groups we can't sort at all.
4388                  */
4389                 availspace -= estimate_hashagg_tablesize(path,
4390                                                                                                  agg_costs,
4391                                                                                                  gd->dNumHashGroups);
4392
4393                 if (availspace > 0 && list_length(gd->rollups) > 1)
4394                 {
4395                         double          scale;
4396                         int                     num_rollups = list_length(gd->rollups);
4397                         int                     k_capacity;
4398                         int                *k_weights = palloc(num_rollups * sizeof(int));
4399                         Bitmapset  *hash_items = NULL;
4400                         int                     i;
4401
4402                         /*
4403                          * We treat this as a knapsack problem: the knapsack capacity
4404                          * represents work_mem, the item weights are the estimated memory
4405                          * usage of the hashtables needed to implement a single rollup,
4406                          * and we really ought to use the cost saving as the item value;
4407                          * however, currently the costs assigned to sort nodes don't
4408                          * reflect the comparison costs well, and so we treat all items as
4409                          * of equal value (each rollup we hash instead saves us one sort).
4410                          *
4411                          * To use the discrete knapsack, we need to scale the values to a
4412                          * reasonably small bounded range.  We choose to allow a 5% error
4413                          * margin; we have no more than 4096 rollups in the worst possible
4414                          * case, which with a 5% error margin will require a bit over 42MB
4415                          * of workspace. (Anyone wanting to plan queries that complex had
4416                          * better have the memory for it.  In more reasonable cases, with
4417                          * no more than a couple of dozen rollups, the memory usage will
4418                          * be negligible.)
4419                          *
4420                          * k_capacity is naturally bounded, but we clamp the values for
4421                          * scale and weight (below) to avoid overflows or underflows (or
4422                          * uselessly trying to use a scale factor less than 1 byte).
4423                          */
4424                         scale = Max(availspace / (20.0 * num_rollups), 1.0);
4425                         k_capacity = (int) floor(availspace / scale);
4426
4427                         /*
4428                          * We leave the first rollup out of consideration since it's the
4429                          * one that matches the input sort order.  We assign indexes "i"
4430                          * to only those entries considered for hashing; the second loop,
4431                          * below, must use the same condition.
4432                          */
4433                         i = 0;
4434                         for_each_cell(lc, gd->rollups, list_second_cell(gd->rollups))
4435                         {
4436                                 RollupData *rollup = lfirst_node(RollupData, lc);
4437
4438                                 if (rollup->hashable)
4439                                 {
4440                                         double          sz = estimate_hashagg_tablesize(path,
4441                                                                                                                                 agg_costs,
4442                                                                                                                                 rollup->numGroups);
4443
4444                                         /*
4445                                          * If sz is enormous, but work_mem (and hence scale) is
4446                                          * small, avoid integer overflow here.
4447                                          */
4448                                         k_weights[i] = (int) Min(floor(sz / scale),
4449                                                                                          k_capacity + 1.0);
4450                                         ++i;
4451                                 }
4452                         }
4453
4454                         /*
4455                          * Apply knapsack algorithm; compute the set of items which
4456                          * maximizes the value stored (in this case the number of sorts
4457                          * saved) while keeping the total size (approximately) within
4458                          * capacity.
4459                          */
4460                         if (i > 0)
4461                                 hash_items = DiscreteKnapsack(k_capacity, i, k_weights, NULL);
4462
4463                         if (!bms_is_empty(hash_items))
4464                         {
4465                                 rollups = list_make1(linitial(gd->rollups));
4466
4467                                 i = 0;
4468                                 for_each_cell(lc, gd->rollups, list_second_cell(gd->rollups))
4469                                 {
4470                                         RollupData *rollup = lfirst_node(RollupData, lc);
4471
4472                                         if (rollup->hashable)
4473                                         {
4474                                                 if (bms_is_member(i, hash_items))
4475                                                         hash_sets = list_concat(hash_sets,
4476                                                                                                         list_copy(rollup->gsets_data));
4477                                                 else
4478                                                         rollups = lappend(rollups, rollup);
4479                                                 ++i;
4480                                         }
4481                                         else
4482                                                 rollups = lappend(rollups, rollup);
4483                                 }
4484                         }
4485                 }
4486
4487                 if (!rollups && hash_sets)
4488                         rollups = list_copy(gd->rollups);
4489
4490                 foreach(lc, hash_sets)
4491                 {
4492                         GroupingSetData *gs = lfirst_node(GroupingSetData, lc);
4493                         RollupData *rollup = makeNode(RollupData);
4494
4495                         Assert(gs->set != NIL);
4496
4497                         rollup->groupClause = preprocess_groupclause(root, gs->set);
4498                         rollup->gsets_data = list_make1(gs);
4499                         rollup->gsets = remap_to_groupclause_idx(rollup->groupClause,
4500                                                                                                          rollup->gsets_data,
4501                                                                                                          gd->tleref_to_colnum_map);
4502                         rollup->numGroups = gs->numGroups;
4503                         rollup->hashable = true;
4504                         rollup->is_hashed = true;
4505                         rollups = lcons(rollup, rollups);
4506                 }
4507
4508                 if (rollups)
4509                 {
4510                         add_path(grouped_rel, (Path *)
4511                                          create_groupingsets_path(root,
4512                                                                                           grouped_rel,
4513                                                                                           path,
4514                                                                                           (List *) parse->havingQual,
4515                                                                                           AGG_MIXED,
4516                                                                                           rollups,
4517                                                                                           agg_costs,
4518                                                                                           dNumGroups));
4519                 }
4520         }
4521
4522         /*
4523          * Now try the simple sorted case.
4524          */
4525         if (!gd->unsortable_sets)
4526                 add_path(grouped_rel, (Path *)
4527                                  create_groupingsets_path(root,
4528                                                                                   grouped_rel,
4529                                                                                   path,
4530                                                                                   (List *) parse->havingQual,
4531                                                                                   AGG_SORTED,
4532                                                                                   gd->rollups,
4533                                                                                   agg_costs,
4534                                                                                   dNumGroups));
4535 }
4536
4537 /*
4538  * create_window_paths
4539  *
4540  * Build a new upperrel containing Paths for window-function evaluation.
4541  *
4542  * input_rel: contains the source-data Paths
4543  * input_target: result of make_window_input_target
4544  * output_target: what the topmost WindowAggPath should return
4545  * wflists: result of find_window_functions
4546  * activeWindows: result of select_active_windows
4547  *
4548  * Note: all Paths in input_rel are expected to return input_target.
4549  */
4550 static RelOptInfo *
4551 create_window_paths(PlannerInfo *root,
4552                                         RelOptInfo *input_rel,
4553                                         PathTarget *input_target,
4554                                         PathTarget *output_target,
4555                                         bool output_target_parallel_safe,
4556                                         WindowFuncLists *wflists,
4557                                         List *activeWindows)
4558 {
4559         RelOptInfo *window_rel;
4560         ListCell   *lc;
4561
4562         /* For now, do all work in the (WINDOW, NULL) upperrel */
4563         window_rel = fetch_upper_rel(root, UPPERREL_WINDOW, NULL);
4564
4565         /*
4566          * If the input relation is not parallel-safe, then the window relation
4567          * can't be parallel-safe, either.  Otherwise, we need to examine the
4568          * target list and active windows for non-parallel-safe constructs.
4569          */
4570         if (input_rel->consider_parallel && output_target_parallel_safe &&
4571                 is_parallel_safe(root, (Node *) activeWindows))
4572                 window_rel->consider_parallel = true;
4573
4574         /*
4575          * If the input rel belongs to a single FDW, so does the window rel.
4576          */
4577         window_rel->serverid = input_rel->serverid;
4578         window_rel->userid = input_rel->userid;
4579         window_rel->useridiscurrent = input_rel->useridiscurrent;
4580         window_rel->fdwroutine = input_rel->fdwroutine;
4581
4582         /*
4583          * Consider computing window functions starting from the existing
4584          * cheapest-total path (which will likely require a sort) as well as any
4585          * existing paths that satisfy root->window_pathkeys (which won't).
4586          */
4587         foreach(lc, input_rel->pathlist)
4588         {
4589                 Path       *path = (Path *) lfirst(lc);
4590
4591                 if (path == input_rel->cheapest_total_path ||
4592                         pathkeys_contained_in(root->window_pathkeys, path->pathkeys))
4593                         create_one_window_path(root,
4594                                                                    window_rel,
4595                                                                    path,
4596                                                                    input_target,
4597                                                                    output_target,
4598                                                                    wflists,
4599                                                                    activeWindows);
4600         }
4601
4602         /*
4603          * If there is an FDW that's responsible for all baserels of the query,
4604          * let it consider adding ForeignPaths.
4605          */
4606         if (window_rel->fdwroutine &&
4607                 window_rel->fdwroutine->GetForeignUpperPaths)
4608                 window_rel->fdwroutine->GetForeignUpperPaths(root, UPPERREL_WINDOW,
4609                                                                                                          input_rel, window_rel,
4610                                                                                                          NULL);
4611
4612         /* Let extensions possibly add some more paths */
4613         if (create_upper_paths_hook)
4614                 (*create_upper_paths_hook) (root, UPPERREL_WINDOW,
4615                                                                         input_rel, window_rel, NULL);
4616
4617         /* Now choose the best path(s) */
4618         set_cheapest(window_rel);
4619
4620         return window_rel;
4621 }
4622
4623 /*
4624  * Stack window-function implementation steps atop the given Path, and
4625  * add the result to window_rel.
4626  *
4627  * window_rel: upperrel to contain result
4628  * path: input Path to use (must return input_target)
4629  * input_target: result of make_window_input_target
4630  * output_target: what the topmost WindowAggPath should return
4631  * wflists: result of find_window_functions
4632  * activeWindows: result of select_active_windows
4633  */
4634 static void
4635 create_one_window_path(PlannerInfo *root,
4636                                            RelOptInfo *window_rel,
4637                                            Path *path,
4638                                            PathTarget *input_target,
4639                                            PathTarget *output_target,
4640                                            WindowFuncLists *wflists,
4641                                            List *activeWindows)
4642 {
4643         PathTarget *window_target;
4644         ListCell   *l;
4645
4646         /*
4647          * Since each window clause could require a different sort order, we stack
4648          * up a WindowAgg node for each clause, with sort steps between them as
4649          * needed.  (We assume that select_active_windows chose a good order for
4650          * executing the clauses in.)
4651          *
4652          * input_target should contain all Vars and Aggs needed for the result.
4653          * (In some cases we wouldn't need to propagate all of these all the way
4654          * to the top, since they might only be needed as inputs to WindowFuncs.
4655          * It's probably not worth trying to optimize that though.)  It must also
4656          * contain all window partitioning and sorting expressions, to ensure
4657          * they're computed only once at the bottom of the stack (that's critical
4658          * for volatile functions).  As we climb up the stack, we'll add outputs
4659          * for the WindowFuncs computed at each level.
4660          */
4661         window_target = input_target;
4662
4663         foreach(l, activeWindows)
4664         {
4665                 WindowClause *wc = lfirst_node(WindowClause, l);
4666                 List       *window_pathkeys;
4667
4668                 window_pathkeys = make_pathkeys_for_window(root,
4669                                                                                                    wc,
4670                                                                                                    root->processed_tlist);
4671
4672                 /* Sort if necessary */
4673                 if (!pathkeys_contained_in(window_pathkeys, path->pathkeys))
4674                 {
4675                         path = (Path *) create_sort_path(root, window_rel,
4676                                                                                          path,
4677                                                                                          window_pathkeys,
4678                                                                                          -1.0);
4679                 }
4680
4681                 if (lnext(activeWindows, l))
4682                 {
4683                         /*
4684                          * Add the current WindowFuncs to the output target for this
4685                          * intermediate WindowAggPath.  We must copy window_target to
4686                          * avoid changing the previous path's target.
4687                          *
4688                          * Note: a WindowFunc adds nothing to the target's eval costs; but
4689                          * we do need to account for the increase in tlist width.
4690                          */
4691                         ListCell   *lc2;
4692
4693                         window_target = copy_pathtarget(window_target);
4694                         foreach(lc2, wflists->windowFuncs[wc->winref])
4695                         {
4696                                 WindowFunc *wfunc = lfirst_node(WindowFunc, lc2);
4697
4698                                 add_column_to_pathtarget(window_target, (Expr *) wfunc, 0);
4699                                 window_target->width += get_typavgwidth(wfunc->wintype, -1);
4700                         }
4701                 }
4702                 else
4703                 {
4704                         /* Install the goal target in the topmost WindowAgg */
4705                         window_target = output_target;
4706                 }
4707
4708                 path = (Path *)
4709                         create_windowagg_path(root, window_rel, path, window_target,
4710                                                                   wflists->windowFuncs[wc->winref],
4711                                                                   wc);
4712         }
4713
4714         add_path(window_rel, path);
4715 }
4716
4717 /*
4718  * create_distinct_paths
4719  *
4720  * Build a new upperrel containing Paths for SELECT DISTINCT evaluation.
4721  *
4722  * input_rel: contains the source-data Paths
4723  *
4724  * Note: input paths should already compute the desired pathtarget, since
4725  * Sort/Unique won't project anything.
4726  */
4727 static RelOptInfo *
4728 create_distinct_paths(PlannerInfo *root,
4729                                           RelOptInfo *input_rel)
4730 {
4731         Query      *parse = root->parse;
4732         Path       *cheapest_input_path = input_rel->cheapest_total_path;
4733         RelOptInfo *distinct_rel;
4734         double          numDistinctRows;
4735         bool            allow_hash;
4736         Path       *path;
4737         ListCell   *lc;
4738
4739         /* For now, do all work in the (DISTINCT, NULL) upperrel */
4740         distinct_rel = fetch_upper_rel(root, UPPERREL_DISTINCT, NULL);
4741
4742         /*
4743          * We don't compute anything at this level, so distinct_rel will be
4744          * parallel-safe if the input rel is parallel-safe.  In particular, if
4745          * there is a DISTINCT ON (...) clause, any path for the input_rel will
4746          * output those expressions, and will not be parallel-safe unless those
4747          * expressions are parallel-safe.
4748          */
4749         distinct_rel->consider_parallel = input_rel->consider_parallel;
4750
4751         /*
4752          * If the input rel belongs to a single FDW, so does the distinct_rel.
4753          */
4754         distinct_rel->serverid = input_rel->serverid;
4755         distinct_rel->userid = input_rel->userid;
4756         distinct_rel->useridiscurrent = input_rel->useridiscurrent;
4757         distinct_rel->fdwroutine = input_rel->fdwroutine;
4758
4759         /* Estimate number of distinct rows there will be */
4760         if (parse->groupClause || parse->groupingSets || parse->hasAggs ||
4761                 root->hasHavingQual)
4762         {
4763                 /*
4764                  * If there was grouping or aggregation, use the number of input rows
4765                  * as the estimated number of DISTINCT rows (ie, assume the input is
4766                  * already mostly unique).
4767                  */
4768                 numDistinctRows = cheapest_input_path->rows;
4769         }
4770         else
4771         {
4772                 /*
4773                  * Otherwise, the UNIQUE filter has effects comparable to GROUP BY.
4774                  */
4775                 List       *distinctExprs;
4776
4777                 distinctExprs = get_sortgrouplist_exprs(parse->distinctClause,
4778                                                                                                 parse->targetList);
4779                 numDistinctRows = estimate_num_groups(root, distinctExprs,
4780                                                                                           cheapest_input_path->rows,
4781                                                                                           NULL);
4782         }
4783
4784         /*
4785          * Consider sort-based implementations of DISTINCT, if possible.
4786          */
4787         if (grouping_is_sortable(parse->distinctClause))
4788         {
4789                 /*
4790                  * First, if we have any adequately-presorted paths, just stick a
4791                  * Unique node on those.  Then consider doing an explicit sort of the
4792                  * cheapest input path and Unique'ing that.
4793                  *
4794                  * When we have DISTINCT ON, we must sort by the more rigorous of
4795                  * DISTINCT and ORDER BY, else it won't have the desired behavior.
4796                  * Also, if we do have to do an explicit sort, we might as well use
4797                  * the more rigorous ordering to avoid a second sort later.  (Note
4798                  * that the parser will have ensured that one clause is a prefix of
4799                  * the other.)
4800                  */
4801                 List       *needed_pathkeys;
4802
4803                 if (parse->hasDistinctOn &&
4804                         list_length(root->distinct_pathkeys) <
4805                         list_length(root->sort_pathkeys))
4806                         needed_pathkeys = root->sort_pathkeys;
4807                 else
4808                         needed_pathkeys = root->distinct_pathkeys;
4809
4810                 foreach(lc, input_rel->pathlist)
4811                 {
4812                         Path       *path = (Path *) lfirst(lc);
4813
4814                         if (pathkeys_contained_in(needed_pathkeys, path->pathkeys))
4815                         {
4816                                 add_path(distinct_rel, (Path *)
4817                                                  create_upper_unique_path(root, distinct_rel,
4818                                                                                                   path,
4819                                                                                                   list_length(root->distinct_pathkeys),
4820                                                                                                   numDistinctRows));
4821                         }
4822                 }
4823
4824                 /* For explicit-sort case, always use the more rigorous clause */
4825                 if (list_length(root->distinct_pathkeys) <
4826                         list_length(root->sort_pathkeys))
4827                 {
4828                         needed_pathkeys = root->sort_pathkeys;
4829                         /* Assert checks that parser didn't mess up... */
4830                         Assert(pathkeys_contained_in(root->distinct_pathkeys,
4831                                                                                  needed_pathkeys));
4832                 }
4833                 else
4834                         needed_pathkeys = root->distinct_pathkeys;
4835
4836                 path = cheapest_input_path;
4837                 if (!pathkeys_contained_in(needed_pathkeys, path->pathkeys))
4838                         path = (Path *) create_sort_path(root, distinct_rel,
4839                                                                                          path,
4840                                                                                          needed_pathkeys,
4841                                                                                          -1.0);
4842
4843                 add_path(distinct_rel, (Path *)
4844                                  create_upper_unique_path(root, distinct_rel,
4845                                                                                   path,
4846                                                                                   list_length(root->distinct_pathkeys),
4847                                                                                   numDistinctRows));
4848         }
4849
4850         /*
4851          * Consider hash-based implementations of DISTINCT, if possible.
4852          *
4853          * If we were not able to make any other types of path, we *must* hash or
4854          * die trying.  If we do have other choices, there are several things that
4855          * should prevent selection of hashing: if the query uses DISTINCT ON
4856          * (because it won't really have the expected behavior if we hash), or if
4857          * enable_hashagg is off, or if it looks like the hashtable will exceed
4858          * work_mem.
4859          *
4860          * Note: grouping_is_hashable() is much more expensive to check than the
4861          * other gating conditions, so we want to do it last.
4862          */
4863         if (distinct_rel->pathlist == NIL)
4864                 allow_hash = true;              /* we have no alternatives */
4865         else if (parse->hasDistinctOn || !enable_hashagg)
4866                 allow_hash = false;             /* policy-based decision not to hash */
4867         else
4868         {
4869                 Size            hashentrysize;
4870
4871                 /* Estimate per-hash-entry space at tuple width... */
4872                 hashentrysize = MAXALIGN(cheapest_input_path->pathtarget->width) +
4873                         MAXALIGN(SizeofMinimalTupleHeader);
4874                 /* plus the per-hash-entry overhead */
4875                 hashentrysize += hash_agg_entry_size(0);
4876
4877                 /* Allow hashing only if hashtable is predicted to fit in work_mem */
4878                 allow_hash = (hashentrysize * numDistinctRows <= work_mem * 1024L);
4879         }
4880
4881         if (allow_hash && grouping_is_hashable(parse->distinctClause))
4882         {
4883                 /* Generate hashed aggregate path --- no sort needed */
4884                 add_path(distinct_rel, (Path *)
4885                                  create_agg_path(root,
4886                                                                  distinct_rel,
4887                                                                  cheapest_input_path,
4888                                                                  cheapest_input_path->pathtarget,
4889                                                                  AGG_HASHED,
4890                                                                  AGGSPLIT_SIMPLE,
4891                                                                  parse->distinctClause,
4892                                                                  NIL,
4893                                                                  NULL,
4894                                                                  numDistinctRows));
4895         }
4896
4897         /* Give a helpful error if we failed to find any implementation */
4898         if (distinct_rel->pathlist == NIL)
4899                 ereport(ERROR,
4900                                 (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
4901                                  errmsg("could not implement DISTINCT"),
4902                                  errdetail("Some of the datatypes only support hashing, while others only support sorting.")));
4903
4904         /*
4905          * If there is an FDW that's responsible for all baserels of the query,
4906          * let it consider adding ForeignPaths.
4907          */
4908         if (distinct_rel->fdwroutine &&
4909                 distinct_rel->fdwroutine->GetForeignUpperPaths)
4910                 distinct_rel->fdwroutine->GetForeignUpperPaths(root, UPPERREL_DISTINCT,
4911                                                                                                            input_rel, distinct_rel,
4912                                                                                                            NULL);
4913
4914         /* Let extensions possibly add some more paths */
4915         if (create_upper_paths_hook)
4916                 (*create_upper_paths_hook) (root, UPPERREL_DISTINCT,
4917                                                                         input_rel, distinct_rel, NULL);
4918
4919         /* Now choose the best path(s) */
4920         set_cheapest(distinct_rel);
4921
4922         return distinct_rel;
4923 }
4924
4925 /*
4926  * create_ordered_paths
4927  *
4928  * Build a new upperrel containing Paths for ORDER BY evaluation.
4929  *
4930  * All paths in the result must satisfy the ORDER BY ordering.
4931  * The only new path we need consider is an explicit sort on the
4932  * cheapest-total existing path.
4933  *
4934  * input_rel: contains the source-data Paths
4935  * target: the output tlist the result Paths must emit
4936  * limit_tuples: estimated bound on the number of output tuples,
4937  *              or -1 if no LIMIT or couldn't estimate
4938  */
4939 static RelOptInfo *
4940 create_ordered_paths(PlannerInfo *root,
4941                                          RelOptInfo *input_rel,
4942                                          PathTarget *target,
4943                                          bool target_parallel_safe,
4944                                          double limit_tuples)
4945 {
4946         Path       *cheapest_input_path = input_rel->cheapest_total_path;
4947         RelOptInfo *ordered_rel;
4948         ListCell   *lc;
4949
4950         /* For now, do all work in the (ORDERED, NULL) upperrel */
4951         ordered_rel = fetch_upper_rel(root, UPPERREL_ORDERED, NULL);
4952
4953         /*
4954          * If the input relation is not parallel-safe, then the ordered relation
4955          * can't be parallel-safe, either.  Otherwise, it's parallel-safe if the
4956          * target list is parallel-safe.
4957          */
4958         if (input_rel->consider_parallel && target_parallel_safe)
4959                 ordered_rel->consider_parallel = true;
4960
4961         /*
4962          * If the input rel belongs to a single FDW, so does the ordered_rel.
4963          */
4964         ordered_rel->serverid = input_rel->serverid;
4965         ordered_rel->userid = input_rel->userid;
4966         ordered_rel->useridiscurrent = input_rel->useridiscurrent;
4967         ordered_rel->fdwroutine = input_rel->fdwroutine;
4968
4969         foreach(lc, input_rel->pathlist)
4970         {
4971                 Path       *path = (Path *) lfirst(lc);
4972                 bool            is_sorted;
4973
4974                 is_sorted = pathkeys_contained_in(root->sort_pathkeys,
4975                                                                                   path->pathkeys);
4976                 if (path == cheapest_input_path || is_sorted)
4977                 {
4978                         if (!is_sorted)
4979                         {
4980                                 /* An explicit sort here can take advantage of LIMIT */
4981                                 path = (Path *) create_sort_path(root,
4982                                                                                                  ordered_rel,
4983                                                                                                  path,
4984                                                                                                  root->sort_pathkeys,
4985                                                                                                  limit_tuples);
4986                         }
4987
4988                         /* Add projection step if needed */
4989                         if (path->pathtarget != target)
4990                                 path = apply_projection_to_path(root, ordered_rel,
4991                                                                                                 path, target);
4992
4993                         add_path(ordered_rel, path);
4994                 }
4995         }
4996
4997         /*
4998          * generate_gather_paths() will have already generated a simple Gather
4999          * path for the best parallel path, if any, and the loop above will have
5000          * considered sorting it.  Similarly, generate_gather_paths() will also
5001          * have generated order-preserving Gather Merge plans which can be used
5002          * without sorting if they happen to match the sort_pathkeys, and the loop
5003          * above will have handled those as well.  However, there's one more
5004          * possibility: it may make sense to sort the cheapest partial path
5005          * according to the required output order and then use Gather Merge.
5006          */
5007         if (ordered_rel->consider_parallel && root->sort_pathkeys != NIL &&
5008                 input_rel->partial_pathlist != NIL)
5009         {
5010                 Path       *cheapest_partial_path;
5011
5012                 cheapest_partial_path = linitial(input_rel->partial_pathlist);
5013
5014                 /*
5015                  * If cheapest partial path doesn't need a sort, this is redundant
5016                  * with what's already been tried.
5017                  */
5018                 if (!pathkeys_contained_in(root->sort_pathkeys,
5019                                                                    cheapest_partial_path->pathkeys))
5020                 {
5021                         Path       *path;
5022                         double          total_groups;
5023
5024                         path = (Path *) create_sort_path(root,
5025                                                                                          ordered_rel,
5026                                                                                          cheapest_partial_path,
5027                                                                                          root->sort_pathkeys,
5028                                                                                          limit_tuples);
5029
5030                         total_groups = cheapest_partial_path->rows *
5031                                 cheapest_partial_path->parallel_workers;
5032                         path = (Path *)
5033                                 create_gather_merge_path(root, ordered_rel,
5034                                                                                  path,
5035                                                                                  path->pathtarget,
5036                                                                                  root->sort_pathkeys, NULL,
5037                                                                                  &total_groups);
5038
5039                         /* Add projection step if needed */
5040                         if (path->pathtarget != target)
5041                                 path = apply_projection_to_path(root, ordered_rel,
5042                                                                                                 path, target);
5043
5044                         add_path(ordered_rel, path);
5045                 }
5046         }
5047
5048         /*
5049          * If there is an FDW that's responsible for all baserels of the query,
5050          * let it consider adding ForeignPaths.
5051          */
5052         if (ordered_rel->fdwroutine &&
5053                 ordered_rel->fdwroutine->GetForeignUpperPaths)
5054                 ordered_rel->fdwroutine->GetForeignUpperPaths(root, UPPERREL_ORDERED,
5055                                                                                                           input_rel, ordered_rel,
5056                                                                                                           NULL);
5057
5058         /* Let extensions possibly add some more paths */
5059         if (create_upper_paths_hook)
5060                 (*create_upper_paths_hook) (root, UPPERREL_ORDERED,
5061                                                                         input_rel, ordered_rel, NULL);
5062
5063         /*
5064          * No need to bother with set_cheapest here; grouping_planner does not
5065          * need us to do it.
5066          */
5067         Assert(ordered_rel->pathlist != NIL);
5068
5069         return ordered_rel;
5070 }
5071
5072
5073 /*
5074  * make_group_input_target
5075  *        Generate appropriate PathTarget for initial input to grouping nodes.
5076  *
5077  * If there is grouping or aggregation, the scan/join subplan cannot emit
5078  * the query's final targetlist; for example, it certainly can't emit any
5079  * aggregate function calls.  This routine generates the correct target
5080  * for the scan/join subplan.
5081  *
5082  * The query target list passed from the parser already contains entries
5083  * for all ORDER BY and GROUP BY expressions, but it will not have entries
5084  * for variables used only in HAVING clauses; so we need to add those
5085  * variables to the subplan target list.  Also, we flatten all expressions
5086  * except GROUP BY items into their component variables; other expressions
5087  * will be computed by the upper plan nodes rather than by the subplan.
5088  * For example, given a query like
5089  *              SELECT a+b,SUM(c+d) FROM table GROUP BY a+b;
5090  * we want to pass this targetlist to the subplan:
5091  *              a+b,c,d
5092  * where the a+b target will be used by the Sort/Group steps, and the
5093  * other targets will be used for computing the final results.
5094  *
5095  * 'final_target' is the query's final target list (in PathTarget form)
5096  *
5097  * The result is the PathTarget to be computed by the Paths returned from
5098  * query_planner().
5099  */
5100 static PathTarget *
5101 make_group_input_target(PlannerInfo *root, PathTarget *final_target)
5102 {
5103         Query      *parse = root->parse;
5104         PathTarget *input_target;
5105         List       *non_group_cols;
5106         List       *non_group_vars;
5107         int                     i;
5108         ListCell   *lc;
5109
5110         /*
5111          * We must build a target containing all grouping columns, plus any other
5112          * Vars mentioned in the query's targetlist and HAVING qual.
5113          */
5114         input_target = create_empty_pathtarget();
5115         non_group_cols = NIL;
5116
5117         i = 0;
5118         foreach(lc, final_target->exprs)
5119         {
5120                 Expr       *expr = (Expr *) lfirst(lc);
5121                 Index           sgref = get_pathtarget_sortgroupref(final_target, i);
5122
5123                 if (sgref && parse->groupClause &&
5124                         get_sortgroupref_clause_noerr(sgref, parse->groupClause) != NULL)
5125                 {
5126                         /*
5127                          * It's a grouping column, so add it to the input target as-is.
5128                          */
5129                         add_column_to_pathtarget(input_target, expr, sgref);
5130                 }
5131                 else
5132                 {
5133                         /*
5134                          * Non-grouping column, so just remember the expression for later
5135                          * call to pull_var_clause.
5136                          */
5137                         non_group_cols = lappend(non_group_cols, expr);
5138                 }
5139
5140                 i++;
5141         }
5142
5143         /*
5144          * If there's a HAVING clause, we'll need the Vars it uses, too.
5145          */
5146         if (parse->havingQual)
5147                 non_group_cols = lappend(non_group_cols, parse->havingQual);
5148
5149         /*
5150          * Pull out all the Vars mentioned in non-group cols (plus HAVING), and
5151          * add them to the input target if not already present.  (A Var used
5152          * directly as a GROUP BY item will be present already.)  Note this
5153          * includes Vars used in resjunk items, so we are covering the needs of
5154          * ORDER BY and window specifications.  Vars used within Aggrefs and
5155          * WindowFuncs will be pulled out here, too.
5156          */
5157         non_group_vars = pull_var_clause((Node *) non_group_cols,
5158                                                                          PVC_RECURSE_AGGREGATES |
5159                                                                          PVC_RECURSE_WINDOWFUNCS |
5160                                                                          PVC_INCLUDE_PLACEHOLDERS);
5161         add_new_columns_to_pathtarget(input_target, non_group_vars);
5162
5163         /* clean up cruft */
5164         list_free(non_group_vars);
5165         list_free(non_group_cols);
5166
5167         /* XXX this causes some redundant cost calculation ... */
5168         return set_pathtarget_cost_width(root, input_target);
5169 }
5170
5171 /*
5172  * make_partial_grouping_target
5173  *        Generate appropriate PathTarget for output of partial aggregate
5174  *        (or partial grouping, if there are no aggregates) nodes.
5175  *
5176  * A partial aggregation node needs to emit all the same aggregates that
5177  * a regular aggregation node would, plus any aggregates used in HAVING;
5178  * except that the Aggref nodes should be marked as partial aggregates.
5179  *
5180  * In addition, we'd better emit any Vars and PlaceholderVars that are
5181  * used outside of Aggrefs in the aggregation tlist and HAVING.  (Presumably,
5182  * these would be Vars that are grouped by or used in grouping expressions.)
5183  *
5184  * grouping_target is the tlist to be emitted by the topmost aggregation step.
5185  * havingQual represents the HAVING clause.
5186  */
5187 static PathTarget *
5188 make_partial_grouping_target(PlannerInfo *root,
5189                                                          PathTarget *grouping_target,
5190                                                          Node *havingQual)
5191 {
5192         Query      *parse = root->parse;
5193         PathTarget *partial_target;
5194         List       *non_group_cols;
5195         List       *non_group_exprs;
5196         int                     i;
5197         ListCell   *lc;
5198
5199         partial_target = create_empty_pathtarget();
5200         non_group_cols = NIL;
5201
5202         i = 0;
5203         foreach(lc, grouping_target->exprs)
5204         {
5205                 Expr       *expr = (Expr *) lfirst(lc);
5206                 Index           sgref = get_pathtarget_sortgroupref(grouping_target, i);
5207
5208                 if (sgref && parse->groupClause &&
5209                         get_sortgroupref_clause_noerr(sgref, parse->groupClause) != NULL)
5210                 {
5211                         /*
5212                          * It's a grouping column, so add it to the partial_target as-is.
5213                          * (This allows the upper agg step to repeat the grouping calcs.)
5214                          */
5215                         add_column_to_pathtarget(partial_target, expr, sgref);
5216                 }
5217                 else
5218                 {
5219                         /*
5220                          * Non-grouping column, so just remember the expression for later
5221                          * call to pull_var_clause.
5222                          */
5223                         non_group_cols = lappend(non_group_cols, expr);
5224                 }
5225
5226                 i++;
5227         }
5228
5229         /*
5230          * If there's a HAVING clause, we'll need the Vars/Aggrefs it uses, too.
5231          */
5232         if (havingQual)
5233                 non_group_cols = lappend(non_group_cols, havingQual);
5234
5235         /*
5236          * Pull out all the Vars, PlaceHolderVars, and Aggrefs mentioned in
5237          * non-group cols (plus HAVING), and add them to the partial_target if not
5238          * already present.  (An expression used directly as a GROUP BY item will
5239          * be present already.)  Note this includes Vars used in resjunk items, so
5240          * we are covering the needs of ORDER BY and window specifications.
5241          */
5242         non_group_exprs = pull_var_clause((Node *) non_group_cols,
5243                                                                           PVC_INCLUDE_AGGREGATES |
5244                                                                           PVC_RECURSE_WINDOWFUNCS |
5245                                                                           PVC_INCLUDE_PLACEHOLDERS);
5246
5247         add_new_columns_to_pathtarget(partial_target, non_group_exprs);
5248
5249         /*
5250          * Adjust Aggrefs to put them in partial mode.  At this point all Aggrefs
5251          * are at the top level of the target list, so we can just scan the list
5252          * rather than recursing through the expression trees.
5253          */
5254         foreach(lc, partial_target->exprs)
5255         {
5256                 Aggref     *aggref = (Aggref *) lfirst(lc);
5257
5258                 if (IsA(aggref, Aggref))
5259                 {
5260                         Aggref     *newaggref;
5261
5262                         /*
5263                          * We shouldn't need to copy the substructure of the Aggref node,
5264                          * but flat-copy the node itself to avoid damaging other trees.
5265                          */
5266                         newaggref = makeNode(Aggref);
5267                         memcpy(newaggref, aggref, sizeof(Aggref));
5268
5269                         /* For now, assume serialization is required */
5270                         mark_partial_aggref(newaggref, AGGSPLIT_INITIAL_SERIAL);
5271
5272                         lfirst(lc) = newaggref;
5273                 }
5274         }
5275
5276         /* clean up cruft */
5277         list_free(non_group_exprs);
5278         list_free(non_group_cols);
5279
5280         /* XXX this causes some redundant cost calculation ... */
5281         return set_pathtarget_cost_width(root, partial_target);
5282 }
5283
5284 /*
5285  * mark_partial_aggref
5286  *        Adjust an Aggref to make it represent a partial-aggregation step.
5287  *
5288  * The Aggref node is modified in-place; caller must do any copying required.
5289  */
5290 void
5291 mark_partial_aggref(Aggref *agg, AggSplit aggsplit)
5292 {
5293         /* aggtranstype should be computed by this point */
5294         Assert(OidIsValid(agg->aggtranstype));
5295         /* ... but aggsplit should still be as the parser left it */
5296         Assert(agg->aggsplit == AGGSPLIT_SIMPLE);
5297
5298         /* Mark the Aggref with the intended partial-aggregation mode */
5299         agg->aggsplit = aggsplit;
5300
5301         /*
5302          * Adjust result type if needed.  Normally, a partial aggregate returns
5303          * the aggregate's transition type; but if that's INTERNAL and we're
5304          * serializing, it returns BYTEA instead.
5305          */
5306         if (DO_AGGSPLIT_SKIPFINAL(aggsplit))
5307         {
5308                 if (agg->aggtranstype == INTERNALOID && DO_AGGSPLIT_SERIALIZE(aggsplit))
5309                         agg->aggtype = BYTEAOID;
5310                 else
5311                         agg->aggtype = agg->aggtranstype;
5312         }
5313 }
5314
5315 /*
5316  * postprocess_setop_tlist
5317  *        Fix up targetlist returned by plan_set_operations().
5318  *
5319  * We need to transpose sort key info from the orig_tlist into new_tlist.
5320  * NOTE: this would not be good enough if we supported resjunk sort keys
5321  * for results of set operations --- then, we'd need to project a whole
5322  * new tlist to evaluate the resjunk columns.  For now, just ereport if we
5323  * find any resjunk columns in orig_tlist.
5324  */
5325 static List *
5326 postprocess_setop_tlist(List *new_tlist, List *orig_tlist)
5327 {
5328         ListCell   *l;
5329         ListCell   *orig_tlist_item = list_head(orig_tlist);
5330
5331         foreach(l, new_tlist)
5332         {
5333                 TargetEntry *new_tle = lfirst_node(TargetEntry, l);
5334                 TargetEntry *orig_tle;
5335
5336                 /* ignore resjunk columns in setop result */
5337                 if (new_tle->resjunk)
5338                         continue;
5339
5340                 Assert(orig_tlist_item != NULL);
5341                 orig_tle = lfirst_node(TargetEntry, orig_tlist_item);
5342                 orig_tlist_item = lnext(orig_tlist, orig_tlist_item);
5343                 if (orig_tle->resjunk)  /* should not happen */
5344                         elog(ERROR, "resjunk output columns are not implemented");
5345                 Assert(new_tle->resno == orig_tle->resno);
5346                 new_tle->ressortgroupref = orig_tle->ressortgroupref;
5347         }
5348         if (orig_tlist_item != NULL)
5349                 elog(ERROR, "resjunk output columns are not implemented");
5350         return new_tlist;
5351 }
5352
5353 /*
5354  * select_active_windows
5355  *              Create a list of the "active" window clauses (ie, those referenced
5356  *              by non-deleted WindowFuncs) in the order they are to be executed.
5357  */
5358 static List *
5359 select_active_windows(PlannerInfo *root, WindowFuncLists *wflists)
5360 {
5361         List       *windowClause = root->parse->windowClause;
5362         List       *result = NIL;
5363         ListCell   *lc;
5364         int                     nActive = 0;
5365         WindowClauseSortData *actives = palloc(sizeof(WindowClauseSortData)
5366                                                                                    * list_length(windowClause));
5367
5368         /* First, construct an array of the active windows */
5369         foreach(lc, windowClause)
5370         {
5371                 WindowClause *wc = lfirst_node(WindowClause, lc);
5372
5373                 /* It's only active if wflists shows some related WindowFuncs */
5374                 Assert(wc->winref <= wflists->maxWinRef);
5375                 if (wflists->windowFuncs[wc->winref] == NIL)
5376                         continue;
5377
5378                 actives[nActive].wc = wc;       /* original clause */
5379
5380                 /*
5381                  * For sorting, we want the list of partition keys followed by the
5382                  * list of sort keys. But pathkeys construction will remove duplicates
5383                  * between the two, so we can as well (even though we can't detect all
5384                  * of the duplicates, since some may come from ECs - that might mean
5385                  * we miss optimization chances here). We must, however, ensure that
5386                  * the order of entries is preserved with respect to the ones we do
5387                  * keep.
5388                  *
5389                  * partitionClause and orderClause had their own duplicates removed in
5390                  * parse analysis, so we're only concerned here with removing
5391                  * orderClause entries that also appear in partitionClause.
5392                  */
5393                 actives[nActive].uniqueOrder =
5394                         list_concat_unique(list_copy(wc->partitionClause),
5395                                                            wc->orderClause);
5396                 nActive++;
5397         }
5398
5399         /*
5400          * Sort active windows by their partitioning/ordering clauses, ignoring
5401          * any framing clauses, so that the windows that need the same sorting are
5402          * adjacent in the list. When we come to generate paths, this will avoid
5403          * inserting additional Sort nodes.
5404          *
5405          * This is how we implement a specific requirement from the SQL standard,
5406          * which says that when two or more windows are order-equivalent (i.e.
5407          * have matching partition and order clauses, even if their names or
5408          * framing clauses differ), then all peer rows must be presented in the
5409          * same order in all of them. If we allowed multiple sort nodes for such
5410          * cases, we'd risk having the peer rows end up in different orders in
5411          * equivalent windows due to sort instability. (See General Rule 4 of
5412          * <window clause> in SQL2008 - SQL2016.)
5413          *
5414          * Additionally, if the entire list of clauses of one window is a prefix
5415          * of another, put first the window with stronger sorting requirements.
5416          * This way we will first sort for stronger window, and won't have to sort
5417          * again for the weaker one.
5418          */
5419         qsort(actives, nActive, sizeof(WindowClauseSortData), common_prefix_cmp);
5420
5421         /* build ordered list of the original WindowClause nodes */
5422         for (int i = 0; i < nActive; i++)
5423                 result = lappend(result, actives[i].wc);
5424
5425         pfree(actives);
5426
5427         return result;
5428 }
5429
5430 /*
5431  * common_prefix_cmp
5432  *        QSort comparison function for WindowClauseSortData
5433  *
5434  * Sort the windows by the required sorting clauses. First, compare the sort
5435  * clauses themselves. Second, if one window's clauses are a prefix of another
5436  * one's clauses, put the window with more sort clauses first.
5437  */
5438 static int
5439 common_prefix_cmp(const void *a, const void *b)
5440 {
5441         const WindowClauseSortData *wcsa = a;
5442         const WindowClauseSortData *wcsb = b;
5443         ListCell   *item_a;
5444         ListCell   *item_b;
5445
5446         forboth(item_a, wcsa->uniqueOrder, item_b, wcsb->uniqueOrder)
5447         {
5448                 SortGroupClause *sca = lfirst_node(SortGroupClause, item_a);
5449                 SortGroupClause *scb = lfirst_node(SortGroupClause, item_b);
5450
5451                 if (sca->tleSortGroupRef > scb->tleSortGroupRef)
5452                         return -1;
5453                 else if (sca->tleSortGroupRef < scb->tleSortGroupRef)
5454                         return 1;
5455                 else if (sca->sortop > scb->sortop)
5456                         return -1;
5457                 else if (sca->sortop < scb->sortop)
5458                         return 1;
5459                 else if (sca->nulls_first && !scb->nulls_first)
5460                         return -1;
5461                 else if (!sca->nulls_first && scb->nulls_first)
5462                         return 1;
5463                 /* no need to compare eqop, since it is fully determined by sortop */
5464         }
5465
5466         if (list_length(wcsa->uniqueOrder) > list_length(wcsb->uniqueOrder))
5467                 return -1;
5468         else if (list_length(wcsa->uniqueOrder) < list_length(wcsb->uniqueOrder))
5469                 return 1;
5470
5471         return 0;
5472 }
5473
5474 /*
5475  * make_window_input_target
5476  *        Generate appropriate PathTarget for initial input to WindowAgg nodes.
5477  *
5478  * When the query has window functions, this function computes the desired
5479  * target to be computed by the node just below the first WindowAgg.
5480  * This tlist must contain all values needed to evaluate the window functions,
5481  * compute the final target list, and perform any required final sort step.
5482  * If multiple WindowAggs are needed, each intermediate one adds its window
5483  * function results onto this base tlist; only the topmost WindowAgg computes
5484  * the actual desired target list.
5485  *
5486  * This function is much like make_group_input_target, though not quite enough
5487  * like it to share code.  As in that function, we flatten most expressions
5488  * into their component variables.  But we do not want to flatten window
5489  * PARTITION BY/ORDER BY clauses, since that might result in multiple
5490  * evaluations of them, which would be bad (possibly even resulting in
5491  * inconsistent answers, if they contain volatile functions).
5492  * Also, we must not flatten GROUP BY clauses that were left unflattened by
5493  * make_group_input_target, because we may no longer have access to the
5494  * individual Vars in them.
5495  *
5496  * Another key difference from make_group_input_target is that we don't
5497  * flatten Aggref expressions, since those are to be computed below the
5498  * window functions and just referenced like Vars above that.
5499  *
5500  * 'final_target' is the query's final target list (in PathTarget form)
5501  * 'activeWindows' is the list of active windows previously identified by
5502  *                      select_active_windows.
5503  *
5504  * The result is the PathTarget to be computed by the plan node immediately
5505  * below the first WindowAgg node.
5506  */
5507 static PathTarget *
5508 make_window_input_target(PlannerInfo *root,
5509                                                  PathTarget *final_target,
5510                                                  List *activeWindows)
5511 {
5512         Query      *parse = root->parse;
5513         PathTarget *input_target;
5514         Bitmapset  *sgrefs;
5515         List       *flattenable_cols;
5516         List       *flattenable_vars;
5517         int                     i;
5518         ListCell   *lc;
5519
5520         Assert(parse->hasWindowFuncs);
5521
5522         /*
5523          * Collect the sortgroupref numbers of window PARTITION/ORDER BY clauses
5524          * into a bitmapset for convenient reference below.
5525          */
5526         sgrefs = NULL;
5527         foreach(lc, activeWindows)
5528         {
5529                 WindowClause *wc = lfirst_node(WindowClause, lc);
5530                 ListCell   *lc2;
5531
5532                 foreach(lc2, wc->partitionClause)
5533                 {
5534                         SortGroupClause *sortcl = lfirst_node(SortGroupClause, lc2);
5535
5536                         sgrefs = bms_add_member(sgrefs, sortcl->tleSortGroupRef);
5537                 }
5538                 foreach(lc2, wc->orderClause)
5539                 {
5540                         SortGroupClause *sortcl = lfirst_node(SortGroupClause, lc2);
5541
5542                         sgrefs = bms_add_member(sgrefs, sortcl->tleSortGroupRef);
5543                 }
5544         }
5545
5546         /* Add in sortgroupref numbers of GROUP BY clauses, too */
5547         foreach(lc, parse->groupClause)
5548         {
5549                 SortGroupClause *grpcl = lfirst_node(SortGroupClause, lc);
5550
5551                 sgrefs = bms_add_member(sgrefs, grpcl->tleSortGroupRef);
5552         }
5553
5554         /*
5555          * Construct a target containing all the non-flattenable targetlist items,
5556          * and save aside the others for a moment.
5557          */
5558         input_target = create_empty_pathtarget();
5559         flattenable_cols = NIL;
5560
5561         i = 0;
5562         foreach(lc, final_target->exprs)
5563         {
5564                 Expr       *expr = (Expr *) lfirst(lc);
5565                 Index           sgref = get_pathtarget_sortgroupref(final_target, i);
5566
5567                 /*
5568                  * Don't want to deconstruct window clauses or GROUP BY items.  (Note
5569                  * that such items can't contain window functions, so it's okay to
5570                  * compute them below the WindowAgg nodes.)
5571                  */
5572                 if (sgref != 0 && bms_is_member(sgref, sgrefs))
5573                 {
5574                         /*
5575                          * Don't want to deconstruct this value, so add it to the input
5576                          * target as-is.
5577                          */
5578                         add_column_to_pathtarget(input_target, expr, sgref);
5579                 }
5580                 else
5581                 {
5582                         /*
5583                          * Column is to be flattened, so just remember the expression for
5584                          * later call to pull_var_clause.
5585                          */
5586                         flattenable_cols = lappend(flattenable_cols, expr);
5587                 }
5588
5589                 i++;
5590         }
5591
5592         /*
5593          * Pull out all the Vars and Aggrefs mentioned in flattenable columns, and
5594          * add them to the input target if not already present.  (Some might be
5595          * there already because they're used directly as window/group clauses.)
5596          *
5597          * Note: it's essential to use PVC_INCLUDE_AGGREGATES here, so that any
5598          * Aggrefs are placed in the Agg node's tlist and not left to be computed
5599          * at higher levels.  On the other hand, we should recurse into
5600          * WindowFuncs to make sure their input expressions are available.
5601          */
5602         flattenable_vars = pull_var_clause((Node *) flattenable_cols,
5603                                                                            PVC_INCLUDE_AGGREGATES |
5604                                                                            PVC_RECURSE_WINDOWFUNCS |
5605                                                                            PVC_INCLUDE_PLACEHOLDERS);
5606         add_new_columns_to_pathtarget(input_target, flattenable_vars);
5607
5608         /* clean up cruft */
5609         list_free(flattenable_vars);
5610         list_free(flattenable_cols);
5611
5612         /* XXX this causes some redundant cost calculation ... */
5613         return set_pathtarget_cost_width(root, input_target);
5614 }
5615
5616 /*
5617  * make_pathkeys_for_window
5618  *              Create a pathkeys list describing the required input ordering
5619  *              for the given WindowClause.
5620  *
5621  * The required ordering is first the PARTITION keys, then the ORDER keys.
5622  * In the future we might try to implement windowing using hashing, in which
5623  * case the ordering could be relaxed, but for now we always sort.
5624  */
5625 static List *
5626 make_pathkeys_for_window(PlannerInfo *root, WindowClause *wc,
5627                                                  List *tlist)
5628 {
5629         List       *window_pathkeys;
5630         List       *window_sortclauses;
5631
5632         /* Throw error if can't sort */
5633         if (!grouping_is_sortable(wc->partitionClause))
5634                 ereport(ERROR,
5635                                 (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
5636                                  errmsg("could not implement window PARTITION BY"),
5637                                  errdetail("Window partitioning columns must be of sortable datatypes.")));
5638         if (!grouping_is_sortable(wc->orderClause))
5639                 ereport(ERROR,
5640                                 (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
5641                                  errmsg("could not implement window ORDER BY"),
5642                                  errdetail("Window ordering columns must be of sortable datatypes.")));
5643
5644         /* Okay, make the combined pathkeys */
5645         window_sortclauses = list_concat(list_copy(wc->partitionClause),
5646                                                                          list_copy(wc->orderClause));
5647         window_pathkeys = make_pathkeys_for_sortclauses(root,
5648                                                                                                         window_sortclauses,
5649                                                                                                         tlist);
5650         list_free(window_sortclauses);
5651         return window_pathkeys;
5652 }
5653
5654 /*
5655  * make_sort_input_target
5656  *        Generate appropriate PathTarget for initial input to Sort step.
5657  *
5658  * If the query has ORDER BY, this function chooses the target to be computed
5659  * by the node just below the Sort (and DISTINCT, if any, since Unique can't
5660  * project) steps.  This might or might not be identical to the query's final
5661  * output target.
5662  *
5663  * The main argument for keeping the sort-input tlist the same as the final
5664  * is that we avoid a separate projection node (which will be needed if
5665  * they're different, because Sort can't project).  However, there are also
5666  * advantages to postponing tlist evaluation till after the Sort: it ensures
5667  * a consistent order of evaluation for any volatile functions in the tlist,
5668  * and if there's also a LIMIT, we can stop the query without ever computing
5669  * tlist functions for later rows, which is beneficial for both volatile and
5670  * expensive functions.
5671  *
5672  * Our current policy is to postpone volatile expressions till after the sort
5673  * unconditionally (assuming that that's possible, ie they are in plain tlist
5674  * columns and not ORDER BY/GROUP BY/DISTINCT columns).  We also prefer to
5675  * postpone set-returning expressions, because running them beforehand would
5676  * bloat the sort dataset, and because it might cause unexpected output order
5677  * if the sort isn't stable.  However there's a constraint on that: all SRFs
5678  * in the tlist should be evaluated at the same plan step, so that they can
5679  * run in sync in nodeProjectSet.  So if any SRFs are in sort columns, we
5680  * mustn't postpone any SRFs.  (Note that in principle that policy should
5681  * probably get applied to the group/window input targetlists too, but we
5682  * have not done that historically.)  Lastly, expensive expressions are
5683  * postponed if there is a LIMIT, or if root->tuple_fraction shows that
5684  * partial evaluation of the query is possible (if neither is true, we expect
5685  * to have to evaluate the expressions for every row anyway), or if there are
5686  * any volatile or set-returning expressions (since once we've put in a
5687  * projection at all, it won't cost any more to postpone more stuff).
5688  *
5689  * Another issue that could potentially be considered here is that
5690  * evaluating tlist expressions could result in data that's either wider
5691  * or narrower than the input Vars, thus changing the volume of data that
5692  * has to go through the Sort.  However, we usually have only a very bad
5693  * idea of the output width of any expression more complex than a Var,
5694  * so for now it seems too risky to try to optimize on that basis.
5695  *
5696  * Note that if we do produce a modified sort-input target, and then the
5697  * query ends up not using an explicit Sort, no particular harm is done:
5698  * we'll initially use the modified target for the preceding path nodes,
5699  * but then change them to the final target with apply_projection_to_path.
5700  * Moreover, in such a case the guarantees about evaluation order of
5701  * volatile functions still hold, since the rows are sorted already.
5702  *
5703  * This function has some things in common with make_group_input_target and
5704  * make_window_input_target, though the detailed rules for what to do are
5705  * different.  We never flatten/postpone any grouping or ordering columns;
5706  * those are needed before the sort.  If we do flatten a particular
5707  * expression, we leave Aggref and WindowFunc nodes alone, since those were
5708  * computed earlier.
5709  *
5710  * 'final_target' is the query's final target list (in PathTarget form)
5711  * 'have_postponed_srfs' is an output argument, see below
5712  *
5713  * The result is the PathTarget to be computed by the plan node immediately
5714  * below the Sort step (and the Distinct step, if any).  This will be
5715  * exactly final_target if we decide a projection step wouldn't be helpful.
5716  *
5717  * In addition, *have_postponed_srfs is set to true if we choose to postpone
5718  * any set-returning functions to after the Sort.
5719  */
5720 static PathTarget *
5721 make_sort_input_target(PlannerInfo *root,
5722                                            PathTarget *final_target,
5723                                            bool *have_postponed_srfs)
5724 {
5725         Query      *parse = root->parse;
5726         PathTarget *input_target;
5727         int                     ncols;
5728         bool       *col_is_srf;
5729         bool       *postpone_col;
5730         bool            have_srf;
5731         bool            have_volatile;
5732         bool            have_expensive;
5733         bool            have_srf_sortcols;
5734         bool            postpone_srfs;
5735         List       *postponable_cols;
5736         List       *postponable_vars;
5737         int                     i;
5738         ListCell   *lc;
5739
5740         /* Shouldn't get here unless query has ORDER BY */
5741         Assert(parse->sortClause);
5742
5743         *have_postponed_srfs = false;   /* default result */
5744
5745         /* Inspect tlist and collect per-column information */
5746         ncols = list_length(final_target->exprs);
5747         col_is_srf = (bool *) palloc0(ncols * sizeof(bool));
5748         postpone_col = (bool *) palloc0(ncols * sizeof(bool));
5749         have_srf = have_volatile = have_expensive = have_srf_sortcols = false;
5750
5751         i = 0;
5752         foreach(lc, final_target->exprs)
5753         {
5754                 Expr       *expr = (Expr *) lfirst(lc);
5755
5756                 /*
5757                  * If the column has a sortgroupref, assume it has to be evaluated
5758                  * before sorting.  Generally such columns would be ORDER BY, GROUP
5759                  * BY, etc targets.  One exception is columns that were removed from
5760                  * GROUP BY by remove_useless_groupby_columns() ... but those would
5761                  * only be Vars anyway.  There don't seem to be any cases where it
5762                  * would be worth the trouble to double-check.
5763                  */
5764                 if (get_pathtarget_sortgroupref(final_target, i) == 0)
5765                 {
5766                         /*
5767                          * Check for SRF or volatile functions.  Check the SRF case first
5768                          * because we must know whether we have any postponed SRFs.
5769                          */
5770                         if (parse->hasTargetSRFs &&
5771                                 expression_returns_set((Node *) expr))
5772                         {
5773                                 /* We'll decide below whether these are postponable */
5774                                 col_is_srf[i] = true;
5775                                 have_srf = true;
5776                         }
5777                         else if (contain_volatile_functions((Node *) expr))
5778                         {
5779                                 /* Unconditionally postpone */
5780                                 postpone_col[i] = true;
5781                                 have_volatile = true;
5782                         }
5783                         else
5784                         {
5785                                 /*
5786                                  * Else check the cost.  XXX it's annoying to have to do this
5787                                  * when set_pathtarget_cost_width() just did it.  Refactor to
5788                                  * allow sharing the work?
5789                                  */
5790                                 QualCost        cost;
5791
5792                                 cost_qual_eval_node(&cost, (Node *) expr, root);
5793
5794                                 /*
5795                                  * We arbitrarily define "expensive" as "more than 10X
5796                                  * cpu_operator_cost".  Note this will take in any PL function
5797                                  * with default cost.
5798                                  */
5799                                 if (cost.per_tuple > 10 * cpu_operator_cost)
5800                                 {
5801                                         postpone_col[i] = true;
5802                                         have_expensive = true;
5803                                 }
5804                         }
5805                 }
5806                 else
5807                 {
5808                         /* For sortgroupref cols, just check if any contain SRFs */
5809                         if (!have_srf_sortcols &&
5810                                 parse->hasTargetSRFs &&
5811                                 expression_returns_set((Node *) expr))
5812                                 have_srf_sortcols = true;
5813                 }
5814
5815                 i++;
5816         }
5817
5818         /*
5819          * We can postpone SRFs if we have some but none are in sortgroupref cols.
5820          */
5821         postpone_srfs = (have_srf && !have_srf_sortcols);
5822
5823         /*
5824          * If we don't need a post-sort projection, just return final_target.
5825          */
5826         if (!(postpone_srfs || have_volatile ||
5827                   (have_expensive &&
5828                    (parse->limitCount || root->tuple_fraction > 0))))
5829                 return final_target;
5830
5831         /*
5832          * Report whether the post-sort projection will contain set-returning
5833          * functions.  This is important because it affects whether the Sort can
5834          * rely on the query's LIMIT (if any) to bound the number of rows it needs
5835          * to return.
5836          */
5837         *have_postponed_srfs = postpone_srfs;
5838
5839         /*
5840          * Construct the sort-input target, taking all non-postponable columns and
5841          * then adding Vars, PlaceHolderVars, Aggrefs, and WindowFuncs found in
5842          * the postponable ones.
5843          */
5844         input_target = create_empty_pathtarget();
5845         postponable_cols = NIL;
5846
5847         i = 0;
5848         foreach(lc, final_target->exprs)
5849         {
5850                 Expr       *expr = (Expr *) lfirst(lc);
5851
5852                 if (postpone_col[i] || (postpone_srfs && col_is_srf[i]))
5853                         postponable_cols = lappend(postponable_cols, expr);
5854                 else
5855                         add_column_to_pathtarget(input_target, expr,
5856                                                                          get_pathtarget_sortgroupref(final_target, i));
5857
5858                 i++;
5859         }
5860
5861         /*
5862          * Pull out all the Vars, Aggrefs, and WindowFuncs mentioned in
5863          * postponable columns, and add them to the sort-input target if not
5864          * already present.  (Some might be there already.)  We mustn't
5865          * deconstruct Aggrefs or WindowFuncs here, since the projection node
5866          * would be unable to recompute them.
5867          */
5868         postponable_vars = pull_var_clause((Node *) postponable_cols,
5869                                                                            PVC_INCLUDE_AGGREGATES |
5870                                                                            PVC_INCLUDE_WINDOWFUNCS |
5871                                                                            PVC_INCLUDE_PLACEHOLDERS);
5872         add_new_columns_to_pathtarget(input_target, postponable_vars);
5873
5874         /* clean up cruft */
5875         list_free(postponable_vars);
5876         list_free(postponable_cols);
5877
5878         /* XXX this represents even more redundant cost calculation ... */
5879         return set_pathtarget_cost_width(root, input_target);
5880 }
5881
5882 /*
5883  * get_cheapest_fractional_path
5884  *        Find the cheapest path for retrieving a specified fraction of all
5885  *        the tuples expected to be returned by the given relation.
5886  *
5887  * We interpret tuple_fraction the same way as grouping_planner.
5888  *
5889  * We assume set_cheapest() has been run on the given rel.
5890  */
5891 Path *
5892 get_cheapest_fractional_path(RelOptInfo *rel, double tuple_fraction)
5893 {
5894         Path       *best_path = rel->cheapest_total_path;
5895         ListCell   *l;
5896
5897         /* If all tuples will be retrieved, just return the cheapest-total path */
5898         if (tuple_fraction <= 0.0)
5899                 return best_path;
5900
5901         /* Convert absolute # of tuples to a fraction; no need to clamp to 0..1 */
5902         if (tuple_fraction >= 1.0 && best_path->rows > 0)
5903                 tuple_fraction /= best_path->rows;
5904
5905         foreach(l, rel->pathlist)
5906         {
5907                 Path       *path = (Path *) lfirst(l);
5908
5909                 if (path == rel->cheapest_total_path ||
5910                         compare_fractional_path_costs(best_path, path, tuple_fraction) <= 0)
5911                         continue;
5912
5913                 best_path = path;
5914         }
5915
5916         return best_path;
5917 }
5918
5919 /*
5920  * adjust_paths_for_srfs
5921  *              Fix up the Paths of the given upperrel to handle tSRFs properly.
5922  *
5923  * The executor can only handle set-returning functions that appear at the
5924  * top level of the targetlist of a ProjectSet plan node.  If we have any SRFs
5925  * that are not at top level, we need to split up the evaluation into multiple
5926  * plan levels in which each level satisfies this constraint.  This function
5927  * modifies each Path of an upperrel that (might) compute any SRFs in its
5928  * output tlist to insert appropriate projection steps.
5929  *
5930  * The given targets and targets_contain_srfs lists are from
5931  * split_pathtarget_at_srfs().  We assume the existing Paths emit the first
5932  * target in targets.
5933  */
5934 static void
5935 adjust_paths_for_srfs(PlannerInfo *root, RelOptInfo *rel,
5936                                           List *targets, List *targets_contain_srfs)
5937 {
5938         ListCell   *lc;
5939
5940         Assert(list_length(targets) == list_length(targets_contain_srfs));
5941         Assert(!linitial_int(targets_contain_srfs));
5942
5943         /* If no SRFs appear at this plan level, nothing to do */
5944         if (list_length(targets) == 1)
5945                 return;
5946
5947         /*
5948          * Stack SRF-evaluation nodes atop each path for the rel.
5949          *
5950          * In principle we should re-run set_cheapest() here to identify the
5951          * cheapest path, but it seems unlikely that adding the same tlist eval
5952          * costs to all the paths would change that, so we don't bother. Instead,
5953          * just assume that the cheapest-startup and cheapest-total paths remain
5954          * so.  (There should be no parameterized paths anymore, so we needn't
5955          * worry about updating cheapest_parameterized_paths.)
5956          */
5957         foreach(lc, rel->pathlist)
5958         {
5959                 Path       *subpath = (Path *) lfirst(lc);
5960                 Path       *newpath = subpath;
5961                 ListCell   *lc1,
5962                                    *lc2;
5963
5964                 Assert(subpath->param_info == NULL);
5965                 forboth(lc1, targets, lc2, targets_contain_srfs)
5966                 {
5967                         PathTarget *thistarget = lfirst_node(PathTarget, lc1);
5968                         bool            contains_srfs = (bool) lfirst_int(lc2);
5969
5970                         /* If this level doesn't contain SRFs, do regular projection */
5971                         if (contains_srfs)
5972                                 newpath = (Path *) create_set_projection_path(root,
5973                                                                                                                           rel,
5974                                                                                                                           newpath,
5975                                                                                                                           thistarget);
5976                         else
5977                                 newpath = (Path *) apply_projection_to_path(root,
5978                                                                                                                         rel,
5979                                                                                                                         newpath,
5980                                                                                                                         thistarget);
5981                 }
5982                 lfirst(lc) = newpath;
5983                 if (subpath == rel->cheapest_startup_path)
5984                         rel->cheapest_startup_path = newpath;
5985                 if (subpath == rel->cheapest_total_path)
5986                         rel->cheapest_total_path = newpath;
5987         }
5988
5989         /* Likewise for partial paths, if any */
5990         foreach(lc, rel->partial_pathlist)
5991         {
5992                 Path       *subpath = (Path *) lfirst(lc);
5993                 Path       *newpath = subpath;
5994                 ListCell   *lc1,
5995                                    *lc2;
5996
5997                 Assert(subpath->param_info == NULL);
5998                 forboth(lc1, targets, lc2, targets_contain_srfs)
5999                 {
6000                         PathTarget *thistarget = lfirst_node(PathTarget, lc1);
6001                         bool            contains_srfs = (bool) lfirst_int(lc2);
6002
6003                         /* If this level doesn't contain SRFs, do regular projection */
6004                         if (contains_srfs)
6005                                 newpath = (Path *) create_set_projection_path(root,
6006                                                                                                                           rel,
6007                                                                                                                           newpath,
6008                                                                                                                           thistarget);
6009                         else
6010                         {
6011                                 /* avoid apply_projection_to_path, in case of multiple refs */
6012                                 newpath = (Path *) create_projection_path(root,
6013                                                                                                                   rel,
6014                                                                                                                   newpath,
6015                                                                                                                   thistarget);
6016                         }
6017                 }
6018                 lfirst(lc) = newpath;
6019         }
6020 }
6021
6022 /*
6023  * expression_planner
6024  *              Perform planner's transformations on a standalone expression.
6025  *
6026  * Various utility commands need to evaluate expressions that are not part
6027  * of a plannable query.  They can do so using the executor's regular
6028  * expression-execution machinery, but first the expression has to be fed
6029  * through here to transform it from parser output to something executable.
6030  *
6031  * Currently, we disallow sublinks in standalone expressions, so there's no
6032  * real "planning" involved here.  (That might not always be true though.)
6033  * What we must do is run eval_const_expressions to ensure that any function
6034  * calls are converted to positional notation and function default arguments
6035  * get inserted.  The fact that constant subexpressions get simplified is a
6036  * side-effect that is useful when the expression will get evaluated more than
6037  * once.  Also, we must fix operator function IDs.
6038  *
6039  * This does not return any information about dependencies of the expression.
6040  * Hence callers should use the results only for the duration of the current
6041  * query.  Callers that would like to cache the results for longer should use
6042  * expression_planner_with_deps, probably via the plancache.
6043  *
6044  * Note: this must not make any damaging changes to the passed-in expression
6045  * tree.  (It would actually be okay to apply fix_opfuncids to it, but since
6046  * we first do an expression_tree_mutator-based walk, what is returned will
6047  * be a new node tree.)  The result is constructed in the current memory
6048  * context; beware that this can leak a lot of additional stuff there, too.
6049  */
6050 Expr *
6051 expression_planner(Expr *expr)
6052 {
6053         Node       *result;
6054
6055         /*
6056          * Convert named-argument function calls, insert default arguments and
6057          * simplify constant subexprs
6058          */
6059         result = eval_const_expressions(NULL, (Node *) expr);
6060
6061         /* Fill in opfuncid values if missing */
6062         fix_opfuncids(result);
6063
6064         return (Expr *) result;
6065 }
6066
6067 /*
6068  * expression_planner_with_deps
6069  *              Perform planner's transformations on a standalone expression,
6070  *              returning expression dependency information along with the result.
6071  *
6072  * This is identical to expression_planner() except that it also returns
6073  * information about possible dependencies of the expression, ie identities of
6074  * objects whose definitions affect the result.  As in a PlannedStmt, these
6075  * are expressed as a list of relation Oids and a list of PlanInvalItems.
6076  */
6077 Expr *
6078 expression_planner_with_deps(Expr *expr,
6079                                                          List **relationOids,
6080                                                          List **invalItems)
6081 {
6082         Node       *result;
6083         PlannerGlobal glob;
6084         PlannerInfo root;
6085
6086         /* Make up dummy planner state so we can use setrefs machinery */
6087         MemSet(&glob, 0, sizeof(glob));
6088         glob.type = T_PlannerGlobal;
6089         glob.relationOids = NIL;
6090         glob.invalItems = NIL;
6091
6092         MemSet(&root, 0, sizeof(root));
6093         root.type = T_PlannerInfo;
6094         root.glob = &glob;
6095
6096         /*
6097          * Convert named-argument function calls, insert default arguments and
6098          * simplify constant subexprs.  Collect identities of inlined functions
6099          * and elided domains, too.
6100          */
6101         result = eval_const_expressions(&root, (Node *) expr);
6102
6103         /* Fill in opfuncid values if missing */
6104         fix_opfuncids(result);
6105
6106         /*
6107          * Now walk the finished expression to find anything else we ought to
6108          * record as an expression dependency.
6109          */
6110         (void) extract_query_dependencies_walker(result, &root);
6111
6112         *relationOids = glob.relationOids;
6113         *invalItems = glob.invalItems;
6114
6115         return (Expr *) result;
6116 }
6117
6118
6119 /*
6120  * plan_cluster_use_sort
6121  *              Use the planner to decide how CLUSTER should implement sorting
6122  *
6123  * tableOid is the OID of a table to be clustered on its index indexOid
6124  * (which is already known to be a btree index).  Decide whether it's
6125  * cheaper to do an indexscan or a seqscan-plus-sort to execute the CLUSTER.
6126  * Return true to use sorting, false to use an indexscan.
6127  *
6128  * Note: caller had better already hold some type of lock on the table.
6129  */
6130 bool
6131 plan_cluster_use_sort(Oid tableOid, Oid indexOid)
6132 {
6133         PlannerInfo *root;
6134         Query      *query;
6135         PlannerGlobal *glob;
6136         RangeTblEntry *rte;
6137         RelOptInfo *rel;
6138         IndexOptInfo *indexInfo;
6139         QualCost        indexExprCost;
6140         Cost            comparisonCost;
6141         Path       *seqScanPath;
6142         Path            seqScanAndSortPath;
6143         IndexPath  *indexScanPath;
6144         ListCell   *lc;
6145
6146         /* We can short-circuit the cost comparison if indexscans are disabled */
6147         if (!enable_indexscan)
6148                 return true;                    /* use sort */
6149
6150         /* Set up mostly-dummy planner state */
6151         query = makeNode(Query);
6152         query->commandType = CMD_SELECT;
6153
6154         glob = makeNode(PlannerGlobal);
6155
6156         root = makeNode(PlannerInfo);
6157         root->parse = query;
6158         root->glob = glob;
6159         root->query_level = 1;
6160         root->planner_cxt = CurrentMemoryContext;
6161         root->wt_param_id = -1;
6162
6163         /* Build a minimal RTE for the rel */
6164         rte = makeNode(RangeTblEntry);
6165         rte->rtekind = RTE_RELATION;
6166         rte->relid = tableOid;
6167         rte->relkind = RELKIND_RELATION;        /* Don't be too picky. */
6168         rte->rellockmode = AccessShareLock;
6169         rte->lateral = false;
6170         rte->inh = false;
6171         rte->inFromCl = true;
6172         query->rtable = list_make1(rte);
6173
6174         /* Set up RTE/RelOptInfo arrays */
6175         setup_simple_rel_arrays(root);
6176
6177         /* Build RelOptInfo */
6178         rel = build_simple_rel(root, 1, NULL);
6179
6180         /* Locate IndexOptInfo for the target index */
6181         indexInfo = NULL;
6182         foreach(lc, rel->indexlist)
6183         {
6184                 indexInfo = lfirst_node(IndexOptInfo, lc);
6185                 if (indexInfo->indexoid == indexOid)
6186                         break;
6187         }
6188
6189         /*
6190          * It's possible that get_relation_info did not generate an IndexOptInfo
6191          * for the desired index; this could happen if it's not yet reached its
6192          * indcheckxmin usability horizon, or if it's a system index and we're
6193          * ignoring system indexes.  In such cases we should tell CLUSTER to not
6194          * trust the index contents but use seqscan-and-sort.
6195          */
6196         if (lc == NULL)                         /* not in the list? */
6197                 return true;                    /* use sort */
6198
6199         /*
6200          * Rather than doing all the pushups that would be needed to use
6201          * set_baserel_size_estimates, just do a quick hack for rows and width.
6202          */
6203         rel->rows = rel->tuples;
6204         rel->reltarget->width = get_relation_data_width(tableOid, NULL);
6205
6206         root->total_table_pages = rel->pages;
6207
6208         /*
6209          * Determine eval cost of the index expressions, if any.  We need to
6210          * charge twice that amount for each tuple comparison that happens during
6211          * the sort, since tuplesort.c will have to re-evaluate the index
6212          * expressions each time.  (XXX that's pretty inefficient...)
6213          */
6214         cost_qual_eval(&indexExprCost, indexInfo->indexprs, root);
6215         comparisonCost = 2.0 * (indexExprCost.startup + indexExprCost.per_tuple);
6216
6217         /* Estimate the cost of seq scan + sort */
6218         seqScanPath = create_seqscan_path(root, rel, NULL, 0);
6219         cost_sort(&seqScanAndSortPath, root, NIL,
6220                           seqScanPath->total_cost, rel->tuples, rel->reltarget->width,
6221                           comparisonCost, maintenance_work_mem, -1.0);
6222
6223         /* Estimate the cost of index scan */
6224         indexScanPath = create_index_path(root, indexInfo,
6225                                                                           NIL, NIL, NIL, NIL,
6226                                                                           ForwardScanDirection, false,
6227                                                                           NULL, 1.0, false);
6228
6229         return (seqScanAndSortPath.total_cost < indexScanPath->path.total_cost);
6230 }
6231
6232 /*
6233  * plan_create_index_workers
6234  *              Use the planner to decide how many parallel worker processes
6235  *              CREATE INDEX should request for use
6236  *
6237  * tableOid is the table on which the index is to be built.  indexOid is the
6238  * OID of an index to be created or reindexed (which must be a btree index).
6239  *
6240  * Return value is the number of parallel worker processes to request.  It
6241  * may be unsafe to proceed if this is 0.  Note that this does not include the
6242  * leader participating as a worker (value is always a number of parallel
6243  * worker processes).
6244  *
6245  * Note: caller had better already hold some type of lock on the table and
6246  * index.
6247  */
6248 int
6249 plan_create_index_workers(Oid tableOid, Oid indexOid)
6250 {
6251         PlannerInfo *root;
6252         Query      *query;
6253         PlannerGlobal *glob;
6254         RangeTblEntry *rte;
6255         Relation        heap;
6256         Relation        index;
6257         RelOptInfo *rel;
6258         int                     parallel_workers;
6259         BlockNumber heap_blocks;
6260         double          reltuples;
6261         double          allvisfrac;
6262
6263         /* Return immediately when parallelism disabled */
6264         if (max_parallel_maintenance_workers == 0)
6265                 return 0;
6266
6267         /* Set up largely-dummy planner state */
6268         query = makeNode(Query);
6269         query->commandType = CMD_SELECT;
6270
6271         glob = makeNode(PlannerGlobal);
6272
6273         root = makeNode(PlannerInfo);
6274         root->parse = query;
6275         root->glob = glob;
6276         root->query_level = 1;
6277         root->planner_cxt = CurrentMemoryContext;
6278         root->wt_param_id = -1;
6279
6280         /*
6281          * Build a minimal RTE.
6282          *
6283          * Mark the RTE with inh = true.  This is a kludge to prevent
6284          * get_relation_info() from fetching index info, which is necessary
6285          * because it does not expect that any IndexOptInfo is currently
6286          * undergoing REINDEX.
6287          */
6288         rte = makeNode(RangeTblEntry);
6289         rte->rtekind = RTE_RELATION;
6290         rte->relid = tableOid;
6291         rte->relkind = RELKIND_RELATION;        /* Don't be too picky. */
6292         rte->rellockmode = AccessShareLock;
6293         rte->lateral = false;
6294         rte->inh = true;
6295         rte->inFromCl = true;
6296         query->rtable = list_make1(rte);
6297
6298         /* Set up RTE/RelOptInfo arrays */
6299         setup_simple_rel_arrays(root);
6300
6301         /* Build RelOptInfo */
6302         rel = build_simple_rel(root, 1, NULL);
6303
6304         /* Rels are assumed already locked by the caller */
6305         heap = table_open(tableOid, NoLock);
6306         index = index_open(indexOid, NoLock);
6307
6308         /*
6309          * Determine if it's safe to proceed.
6310          *
6311          * Currently, parallel workers can't access the leader's temporary tables.
6312          * Furthermore, any index predicate or index expressions must be parallel
6313          * safe.
6314          */
6315         if (heap->rd_rel->relpersistence == RELPERSISTENCE_TEMP ||
6316                 !is_parallel_safe(root, (Node *) RelationGetIndexExpressions(index)) ||
6317                 !is_parallel_safe(root, (Node *) RelationGetIndexPredicate(index)))
6318         {
6319                 parallel_workers = 0;
6320                 goto done;
6321         }
6322
6323         /*
6324          * If parallel_workers storage parameter is set for the table, accept that
6325          * as the number of parallel worker processes to launch (though still cap
6326          * at max_parallel_maintenance_workers).  Note that we deliberately do not
6327          * consider any other factor when parallel_workers is set. (e.g., memory
6328          * use by workers.)
6329          */
6330         if (rel->rel_parallel_workers != -1)
6331         {
6332                 parallel_workers = Min(rel->rel_parallel_workers,
6333                                                            max_parallel_maintenance_workers);
6334                 goto done;
6335         }
6336
6337         /*
6338          * Estimate heap relation size ourselves, since rel->pages cannot be
6339          * trusted (heap RTE was marked as inheritance parent)
6340          */
6341         estimate_rel_size(heap, NULL, &heap_blocks, &reltuples, &allvisfrac);
6342
6343         /*
6344          * Determine number of workers to scan the heap relation using generic
6345          * model
6346          */
6347         parallel_workers = compute_parallel_worker(rel, heap_blocks, -1,
6348                                                                                            max_parallel_maintenance_workers);
6349
6350         /*
6351          * Cap workers based on available maintenance_work_mem as needed.
6352          *
6353          * Note that each tuplesort participant receives an even share of the
6354          * total maintenance_work_mem budget.  Aim to leave participants
6355          * (including the leader as a participant) with no less than 32MB of
6356          * memory.  This leaves cases where maintenance_work_mem is set to 64MB
6357          * immediately past the threshold of being capable of launching a single
6358          * parallel worker to sort.
6359          */
6360         while (parallel_workers > 0 &&
6361                    maintenance_work_mem / (parallel_workers + 1) < 32768L)
6362                 parallel_workers--;
6363
6364 done:
6365         index_close(index, NoLock);
6366         table_close(heap, NoLock);
6367
6368         return parallel_workers;
6369 }
6370
6371 /*
6372  * add_paths_to_grouping_rel
6373  *
6374  * Add non-partial paths to grouping relation.
6375  */
6376 static void
6377 add_paths_to_grouping_rel(PlannerInfo *root, RelOptInfo *input_rel,
6378                                                   RelOptInfo *grouped_rel,
6379                                                   RelOptInfo *partially_grouped_rel,
6380                                                   const AggClauseCosts *agg_costs,
6381                                                   grouping_sets_data *gd, double dNumGroups,
6382                                                   GroupPathExtraData *extra)
6383 {
6384         Query      *parse = root->parse;
6385         Path       *cheapest_path = input_rel->cheapest_total_path;
6386         ListCell   *lc;
6387         bool            can_hash = (extra->flags & GROUPING_CAN_USE_HASH) != 0;
6388         bool            can_sort = (extra->flags & GROUPING_CAN_USE_SORT) != 0;
6389         List       *havingQual = (List *) extra->havingQual;
6390         AggClauseCosts *agg_final_costs = &extra->agg_final_costs;
6391
6392         if (can_sort)
6393         {
6394                 /*
6395                  * Use any available suitably-sorted path as input, and also consider
6396                  * sorting the cheapest-total path.
6397                  */
6398                 foreach(lc, input_rel->pathlist)
6399                 {
6400                         Path       *path = (Path *) lfirst(lc);
6401                         bool            is_sorted;
6402
6403                         is_sorted = pathkeys_contained_in(root->group_pathkeys,
6404                                                                                           path->pathkeys);
6405                         if (path == cheapest_path || is_sorted)
6406                         {
6407                                 /* Sort the cheapest-total path if it isn't already sorted */
6408                                 if (!is_sorted)
6409                                         path = (Path *) create_sort_path(root,
6410                                                                                                          grouped_rel,
6411                                                                                                          path,
6412                                                                                                          root->group_pathkeys,
6413                                                                                                          -1.0);
6414
6415                                 /* Now decide what to stick atop it */
6416                                 if (parse->groupingSets)
6417                                 {
6418                                         consider_groupingsets_paths(root, grouped_rel,
6419                                                                                                 path, true, can_hash,
6420                                                                                                 gd, agg_costs, dNumGroups);
6421                                 }
6422                                 else if (parse->hasAggs)
6423                                 {
6424                                         /*
6425                                          * We have aggregation, possibly with plain GROUP BY. Make
6426                                          * an AggPath.
6427                                          */
6428                                         add_path(grouped_rel, (Path *)
6429                                                          create_agg_path(root,
6430                                                                                          grouped_rel,
6431                                                                                          path,
6432                                                                                          grouped_rel->reltarget,
6433                                                                                          parse->groupClause ? AGG_SORTED : AGG_PLAIN,
6434                                                                                          AGGSPLIT_SIMPLE,
6435                                                                                          parse->groupClause,
6436                                                                                          havingQual,
6437                                                                                          agg_costs,
6438                                                                                          dNumGroups));
6439                                 }
6440                                 else if (parse->groupClause)
6441                                 {
6442                                         /*
6443                                          * We have GROUP BY without aggregation or grouping sets.
6444                                          * Make a GroupPath.
6445                                          */
6446                                         add_path(grouped_rel, (Path *)
6447                                                          create_group_path(root,
6448                                                                                            grouped_rel,
6449                                                                                            path,
6450                                                                                            parse->groupClause,
6451                                                                                            havingQual,
6452                                                                                            dNumGroups));
6453                                 }
6454                                 else
6455                                 {
6456                                         /* Other cases should have been handled above */
6457                                         Assert(false);
6458                                 }
6459                         }
6460                 }
6461
6462                 /*
6463                  * Instead of operating directly on the input relation, we can
6464                  * consider finalizing a partially aggregated path.
6465                  */
6466                 if (partially_grouped_rel != NULL)
6467                 {
6468                         foreach(lc, partially_grouped_rel->pathlist)
6469                         {
6470                                 Path       *path = (Path *) lfirst(lc);
6471
6472                                 /*
6473                                  * Insert a Sort node, if required.  But there's no point in
6474                                  * sorting anything but the cheapest path.
6475                                  */
6476                                 if (!pathkeys_contained_in(root->group_pathkeys, path->pathkeys))
6477                                 {
6478                                         if (path != partially_grouped_rel->cheapest_total_path)
6479                                                 continue;
6480                                         path = (Path *) create_sort_path(root,
6481                                                                                                          grouped_rel,
6482                                                                                                          path,
6483                                                                                                          root->group_pathkeys,
6484                                                                                                          -1.0);
6485                                 }
6486
6487                                 if (parse->hasAggs)
6488                                         add_path(grouped_rel, (Path *)
6489                                                          create_agg_path(root,
6490                                                                                          grouped_rel,
6491                                                                                          path,
6492                                                                                          grouped_rel->reltarget,
6493                                                                                          parse->groupClause ? AGG_SORTED : AGG_PLAIN,
6494                                                                                          AGGSPLIT_FINAL_DESERIAL,
6495                                                                                          parse->groupClause,
6496                                                                                          havingQual,
6497                                                                                          agg_final_costs,
6498                                                                                          dNumGroups));
6499                                 else
6500                                         add_path(grouped_rel, (Path *)
6501                                                          create_group_path(root,
6502                                                                                            grouped_rel,
6503                                                                                            path,
6504                                                                                            parse->groupClause,
6505                                                                                            havingQual,
6506                                                                                            dNumGroups));
6507                         }
6508                 }
6509         }
6510
6511         if (can_hash)
6512         {
6513                 double          hashaggtablesize;
6514
6515                 if (parse->groupingSets)
6516                 {
6517                         /*
6518                          * Try for a hash-only groupingsets path over unsorted input.
6519                          */
6520                         consider_groupingsets_paths(root, grouped_rel,
6521                                                                                 cheapest_path, false, true,
6522                                                                                 gd, agg_costs, dNumGroups);
6523                 }
6524                 else
6525                 {
6526                         hashaggtablesize = estimate_hashagg_tablesize(cheapest_path,
6527                                                                                                                   agg_costs,
6528                                                                                                                   dNumGroups);
6529
6530                         /*
6531                          * Provided that the estimated size of the hashtable does not
6532                          * exceed work_mem, we'll generate a HashAgg Path, although if we
6533                          * were unable to sort above, then we'd better generate a Path, so
6534                          * that we at least have one.
6535                          */
6536                         if (hashaggtablesize < work_mem * 1024L ||
6537                                 grouped_rel->pathlist == NIL)
6538                         {
6539                                 /*
6540                                  * We just need an Agg over the cheapest-total input path,
6541                                  * since input order won't matter.
6542                                  */
6543                                 add_path(grouped_rel, (Path *)
6544                                                  create_agg_path(root, grouped_rel,
6545                                                                                  cheapest_path,
6546                                                                                  grouped_rel->reltarget,
6547                                                                                  AGG_HASHED,
6548                                                                                  AGGSPLIT_SIMPLE,
6549                                                                                  parse->groupClause,
6550                                                                                  havingQual,
6551                                                                                  agg_costs,
6552                                                                                  dNumGroups));
6553                         }
6554                 }
6555
6556                 /*
6557                  * Generate a Finalize HashAgg Path atop of the cheapest partially
6558                  * grouped path, assuming there is one. Once again, we'll only do this
6559                  * if it looks as though the hash table won't exceed work_mem.
6560                  */
6561                 if (partially_grouped_rel && partially_grouped_rel->pathlist)
6562                 {
6563                         Path       *path = partially_grouped_rel->cheapest_total_path;
6564
6565                         hashaggtablesize = estimate_hashagg_tablesize(path,
6566                                                                                                                   agg_final_costs,
6567                                                                                                                   dNumGroups);
6568
6569                         if (hashaggtablesize < work_mem * 1024L)
6570                                 add_path(grouped_rel, (Path *)
6571                                                  create_agg_path(root,
6572                                                                                  grouped_rel,
6573                                                                                  path,
6574                                                                                  grouped_rel->reltarget,
6575                                                                                  AGG_HASHED,
6576                                                                                  AGGSPLIT_FINAL_DESERIAL,
6577                                                                                  parse->groupClause,
6578                                                                                  havingQual,
6579                                                                                  agg_final_costs,
6580                                                                                  dNumGroups));
6581                 }
6582         }
6583
6584         /*
6585          * When partitionwise aggregate is used, we might have fully aggregated
6586          * paths in the partial pathlist, because add_paths_to_append_rel() will
6587          * consider a path for grouped_rel consisting of a Parallel Append of
6588          * non-partial paths from each child.
6589          */
6590         if (grouped_rel->partial_pathlist != NIL)
6591                 gather_grouping_paths(root, grouped_rel);
6592 }
6593
6594 /*
6595  * create_partial_grouping_paths
6596  *
6597  * Create a new upper relation representing the result of partial aggregation
6598  * and populate it with appropriate paths.  Note that we don't finalize the
6599  * lists of paths here, so the caller can add additional partial or non-partial
6600  * paths and must afterward call gather_grouping_paths and set_cheapest on
6601  * the returned upper relation.
6602  *
6603  * All paths for this new upper relation -- both partial and non-partial --
6604  * have been partially aggregated but require a subsequent FinalizeAggregate
6605  * step.
6606  *
6607  * NB: This function is allowed to return NULL if it determines that there is
6608  * no real need to create a new RelOptInfo.
6609  */
6610 static RelOptInfo *
6611 create_partial_grouping_paths(PlannerInfo *root,
6612                                                           RelOptInfo *grouped_rel,
6613                                                           RelOptInfo *input_rel,
6614                                                           grouping_sets_data *gd,
6615                                                           GroupPathExtraData *extra,
6616                                                           bool force_rel_creation)
6617 {
6618         Query      *parse = root->parse;
6619         RelOptInfo *partially_grouped_rel;
6620         AggClauseCosts *agg_partial_costs = &extra->agg_partial_costs;
6621         AggClauseCosts *agg_final_costs = &extra->agg_final_costs;
6622         Path       *cheapest_partial_path = NULL;
6623         Path       *cheapest_total_path = NULL;
6624         double          dNumPartialGroups = 0;
6625         double          dNumPartialPartialGroups = 0;
6626         ListCell   *lc;
6627         bool            can_hash = (extra->flags & GROUPING_CAN_USE_HASH) != 0;
6628         bool            can_sort = (extra->flags & GROUPING_CAN_USE_SORT) != 0;
6629
6630         /*
6631          * Consider whether we should generate partially aggregated non-partial
6632          * paths.  We can only do this if we have a non-partial path, and only if
6633          * the parent of the input rel is performing partial partitionwise
6634          * aggregation.  (Note that extra->patype is the type of partitionwise
6635          * aggregation being used at the parent level, not this level.)
6636          */
6637         if (input_rel->pathlist != NIL &&
6638                 extra->patype == PARTITIONWISE_AGGREGATE_PARTIAL)
6639                 cheapest_total_path = input_rel->cheapest_total_path;
6640
6641         /*
6642          * If parallelism is possible for grouped_rel, then we should consider
6643          * generating partially-grouped partial paths.  However, if the input rel
6644          * has no partial paths, then we can't.
6645          */
6646         if (grouped_rel->consider_parallel && input_rel->partial_pathlist != NIL)
6647                 cheapest_partial_path = linitial(input_rel->partial_pathlist);
6648
6649         /*
6650          * If we can't partially aggregate partial paths, and we can't partially
6651          * aggregate non-partial paths, then don't bother creating the new
6652          * RelOptInfo at all, unless the caller specified force_rel_creation.
6653          */
6654         if (cheapest_total_path == NULL &&
6655                 cheapest_partial_path == NULL &&
6656                 !force_rel_creation)
6657                 return NULL;
6658
6659         /*
6660          * Build a new upper relation to represent the result of partially
6661          * aggregating the rows from the input relation.
6662          */
6663         partially_grouped_rel = fetch_upper_rel(root,
6664                                                                                         UPPERREL_PARTIAL_GROUP_AGG,
6665                                                                                         grouped_rel->relids);
6666         partially_grouped_rel->consider_parallel =
6667                 grouped_rel->consider_parallel;
6668         partially_grouped_rel->reloptkind = grouped_rel->reloptkind;
6669         partially_grouped_rel->serverid = grouped_rel->serverid;
6670         partially_grouped_rel->userid = grouped_rel->userid;
6671         partially_grouped_rel->useridiscurrent = grouped_rel->useridiscurrent;
6672         partially_grouped_rel->fdwroutine = grouped_rel->fdwroutine;
6673
6674         /*
6675          * Build target list for partial aggregate paths.  These paths cannot just
6676          * emit the same tlist as regular aggregate paths, because (1) we must
6677          * include Vars and Aggrefs needed in HAVING, which might not appear in
6678          * the result tlist, and (2) the Aggrefs must be set in partial mode.
6679          */
6680         partially_grouped_rel->reltarget =
6681                 make_partial_grouping_target(root, grouped_rel->reltarget,
6682                                                                          extra->havingQual);
6683
6684         if (!extra->partial_costs_set)
6685         {
6686                 /*
6687                  * Collect statistics about aggregates for estimating costs of
6688                  * performing aggregation in parallel.
6689                  */
6690                 MemSet(agg_partial_costs, 0, sizeof(AggClauseCosts));
6691                 MemSet(agg_final_costs, 0, sizeof(AggClauseCosts));
6692                 if (parse->hasAggs)
6693                 {
6694                         List       *partial_target_exprs;
6695
6696                         /* partial phase */
6697                         partial_target_exprs = partially_grouped_rel->reltarget->exprs;
6698                         get_agg_clause_costs(root, (Node *) partial_target_exprs,
6699                                                                  AGGSPLIT_INITIAL_SERIAL,
6700                                                                  agg_partial_costs);
6701
6702                         /* final phase */
6703                         get_agg_clause_costs(root, (Node *) grouped_rel->reltarget->exprs,
6704                                                                  AGGSPLIT_FINAL_DESERIAL,
6705                                                                  agg_final_costs);
6706                         get_agg_clause_costs(root, extra->havingQual,
6707                                                                  AGGSPLIT_FINAL_DESERIAL,
6708                                                                  agg_final_costs);
6709                 }
6710
6711                 extra->partial_costs_set = true;
6712         }
6713
6714         /* Estimate number of partial groups. */
6715         if (cheapest_total_path != NULL)
6716                 dNumPartialGroups =
6717                         get_number_of_groups(root,
6718                                                                  cheapest_total_path->rows,
6719                                                                  gd,
6720                                                                  extra->targetList);
6721         if (cheapest_partial_path != NULL)
6722                 dNumPartialPartialGroups =
6723                         get_number_of_groups(root,
6724                                                                  cheapest_partial_path->rows,
6725                                                                  gd,
6726                                                                  extra->targetList);
6727
6728         if (can_sort && cheapest_total_path != NULL)
6729         {
6730                 /* This should have been checked previously */
6731                 Assert(parse->hasAggs || parse->groupClause);
6732
6733                 /*
6734                  * Use any available suitably-sorted path as input, and also consider
6735                  * sorting the cheapest partial path.
6736                  */
6737                 foreach(lc, input_rel->pathlist)
6738                 {
6739                         Path       *path = (Path *) lfirst(lc);
6740                         bool            is_sorted;
6741
6742                         is_sorted = pathkeys_contained_in(root->group_pathkeys,
6743                                                                                           path->pathkeys);
6744                         if (path == cheapest_total_path || is_sorted)
6745                         {
6746                                 /* Sort the cheapest partial path, if it isn't already */
6747                                 if (!is_sorted)
6748                                         path = (Path *) create_sort_path(root,
6749                                                                                                          partially_grouped_rel,
6750                                                                                                          path,
6751                                                                                                          root->group_pathkeys,
6752                                                                                                          -1.0);
6753
6754                                 if (parse->hasAggs)
6755                                         add_path(partially_grouped_rel, (Path *)
6756                                                          create_agg_path(root,
6757                                                                                          partially_grouped_rel,
6758                                                                                          path,
6759                                                                                          partially_grouped_rel->reltarget,
6760                                                                                          parse->groupClause ? AGG_SORTED : AGG_PLAIN,
6761                                                                                          AGGSPLIT_INITIAL_SERIAL,
6762                                                                                          parse->groupClause,
6763                                                                                          NIL,
6764                                                                                          agg_partial_costs,
6765                                                                                          dNumPartialGroups));
6766                                 else
6767                                         add_path(partially_grouped_rel, (Path *)
6768                                                          create_group_path(root,
6769                                                                                            partially_grouped_rel,
6770                                                                                            path,
6771                                                                                            parse->groupClause,
6772                                                                                            NIL,
6773                                                                                            dNumPartialGroups));
6774                         }
6775                 }
6776         }
6777
6778         if (can_sort && cheapest_partial_path != NULL)
6779         {
6780                 /* Similar to above logic, but for partial paths. */
6781                 foreach(lc, input_rel->partial_pathlist)
6782                 {
6783                         Path       *path = (Path *) lfirst(lc);
6784                         bool            is_sorted;
6785
6786                         is_sorted = pathkeys_contained_in(root->group_pathkeys,
6787                                                                                           path->pathkeys);
6788                         if (path == cheapest_partial_path || is_sorted)
6789                         {
6790                                 /* Sort the cheapest partial path, if it isn't already */
6791                                 if (!is_sorted)
6792                                         path = (Path *) create_sort_path(root,
6793                                                                                                          partially_grouped_rel,
6794                                                                                                          path,
6795                                                                                                          root->group_pathkeys,
6796                                                                                                          -1.0);
6797
6798                                 if (parse->hasAggs)
6799                                         add_partial_path(partially_grouped_rel, (Path *)
6800                                                                          create_agg_path(root,
6801                                                                                                          partially_grouped_rel,
6802                                                                                                          path,
6803                                                                                                          partially_grouped_rel->reltarget,
6804                                                                                                          parse->groupClause ? AGG_SORTED : AGG_PLAIN,
6805                                                                                                          AGGSPLIT_INITIAL_SERIAL,
6806                                                                                                          parse->groupClause,
6807                                                                                                          NIL,
6808                                                                                                          agg_partial_costs,
6809                                                                                                          dNumPartialPartialGroups));
6810                                 else
6811                                         add_partial_path(partially_grouped_rel, (Path *)
6812                                                                          create_group_path(root,
6813                                                                                                            partially_grouped_rel,
6814                                                                                                            path,
6815                                                                                                            parse->groupClause,
6816                                                                                                            NIL,
6817                                                                                                            dNumPartialPartialGroups));
6818                         }
6819                 }
6820         }
6821
6822         if (can_hash && cheapest_total_path != NULL)
6823         {
6824                 double          hashaggtablesize;
6825
6826                 /* Checked above */
6827                 Assert(parse->hasAggs || parse->groupClause);
6828
6829                 hashaggtablesize =
6830                         estimate_hashagg_tablesize(cheapest_total_path,
6831                                                                            agg_partial_costs,
6832                                                                            dNumPartialGroups);
6833
6834                 /*
6835                  * Tentatively produce a partial HashAgg Path, depending on if it
6836                  * looks as if the hash table will fit in work_mem.
6837                  */
6838                 if (hashaggtablesize < work_mem * 1024L &&
6839                         cheapest_total_path != NULL)
6840                 {
6841                         add_path(partially_grouped_rel, (Path *)
6842                                          create_agg_path(root,
6843                                                                          partially_grouped_rel,
6844                                                                          cheapest_total_path,
6845                                                                          partially_grouped_rel->reltarget,
6846                                                                          AGG_HASHED,
6847                                                                          AGGSPLIT_INITIAL_SERIAL,
6848                                                                          parse->groupClause,
6849                                                                          NIL,
6850                                                                          agg_partial_costs,
6851                                                                          dNumPartialGroups));
6852                 }
6853         }
6854
6855         if (can_hash && cheapest_partial_path != NULL)
6856         {
6857                 double          hashaggtablesize;
6858
6859                 hashaggtablesize =
6860                         estimate_hashagg_tablesize(cheapest_partial_path,
6861                                                                            agg_partial_costs,
6862                                                                            dNumPartialPartialGroups);
6863
6864                 /* Do the same for partial paths. */
6865                 if (hashaggtablesize < work_mem * 1024L &&
6866                         cheapest_partial_path != NULL)
6867                 {
6868                         add_partial_path(partially_grouped_rel, (Path *)
6869                                                          create_agg_path(root,
6870                                                                                          partially_grouped_rel,
6871                                                                                          cheapest_partial_path,
6872                                                                                          partially_grouped_rel->reltarget,
6873                                                                                          AGG_HASHED,
6874                                                                                          AGGSPLIT_INITIAL_SERIAL,
6875                                                                                          parse->groupClause,
6876                                                                                          NIL,
6877                                                                                          agg_partial_costs,
6878                                                                                          dNumPartialPartialGroups));
6879                 }
6880         }
6881
6882         /*
6883          * If there is an FDW that's responsible for all baserels of the query,
6884          * let it consider adding partially grouped ForeignPaths.
6885          */
6886         if (partially_grouped_rel->fdwroutine &&
6887                 partially_grouped_rel->fdwroutine->GetForeignUpperPaths)
6888         {
6889                 FdwRoutine *fdwroutine = partially_grouped_rel->fdwroutine;
6890
6891                 fdwroutine->GetForeignUpperPaths(root,
6892                                                                                  UPPERREL_PARTIAL_GROUP_AGG,
6893                                                                                  input_rel, partially_grouped_rel,
6894                                                                                  extra);
6895         }
6896
6897         return partially_grouped_rel;
6898 }
6899
6900 /*
6901  * Generate Gather and Gather Merge paths for a grouping relation or partial
6902  * grouping relation.
6903  *
6904  * generate_gather_paths does most of the work, but we also consider a special
6905  * case: we could try sorting the data by the group_pathkeys and then applying
6906  * Gather Merge.
6907  *
6908  * NB: This function shouldn't be used for anything other than a grouped or
6909  * partially grouped relation not only because of the fact that it explicitly
6910  * references group_pathkeys but we pass "true" as the third argument to
6911  * generate_gather_paths().
6912  */
6913 static void
6914 gather_grouping_paths(PlannerInfo *root, RelOptInfo *rel)
6915 {
6916         Path       *cheapest_partial_path;
6917
6918         /* Try Gather for unordered paths and Gather Merge for ordered ones. */
6919         generate_gather_paths(root, rel, true);
6920
6921         /* Try cheapest partial path + explicit Sort + Gather Merge. */
6922         cheapest_partial_path = linitial(rel->partial_pathlist);
6923         if (!pathkeys_contained_in(root->group_pathkeys,
6924                                                            cheapest_partial_path->pathkeys))
6925         {
6926                 Path       *path;
6927                 double          total_groups;
6928
6929                 total_groups =
6930                         cheapest_partial_path->rows * cheapest_partial_path->parallel_workers;
6931                 path = (Path *) create_sort_path(root, rel, cheapest_partial_path,
6932                                                                                  root->group_pathkeys,
6933                                                                                  -1.0);
6934                 path = (Path *)
6935                         create_gather_merge_path(root,
6936                                                                          rel,
6937                                                                          path,
6938                                                                          rel->reltarget,
6939                                                                          root->group_pathkeys,
6940                                                                          NULL,
6941                                                                          &total_groups);
6942
6943                 add_path(rel, path);
6944         }
6945 }
6946
6947 /*
6948  * can_partial_agg
6949  *
6950  * Determines whether or not partial grouping and/or aggregation is possible.
6951  * Returns true when possible, false otherwise.
6952  */
6953 static bool
6954 can_partial_agg(PlannerInfo *root, const AggClauseCosts *agg_costs)
6955 {
6956         Query      *parse = root->parse;
6957
6958         if (!parse->hasAggs && parse->groupClause == NIL)
6959         {
6960                 /*
6961                  * We don't know how to do parallel aggregation unless we have either
6962                  * some aggregates or a grouping clause.
6963                  */
6964                 return false;
6965         }
6966         else if (parse->groupingSets)
6967         {
6968                 /* We don't know how to do grouping sets in parallel. */
6969                 return false;
6970         }
6971         else if (agg_costs->hasNonPartial || agg_costs->hasNonSerial)
6972         {
6973                 /* Insufficient support for partial mode. */
6974                 return false;
6975         }
6976
6977         /* Everything looks good. */
6978         return true;
6979 }
6980
6981 /*
6982  * apply_scanjoin_target_to_paths
6983  *
6984  * Adjust the final scan/join relation, and recursively all of its children,
6985  * to generate the final scan/join target.  It would be more correct to model
6986  * this as a separate planning step with a new RelOptInfo at the toplevel and
6987  * for each child relation, but doing it this way is noticeably cheaper.
6988  * Maybe that problem can be solved at some point, but for now we do this.
6989  *
6990  * If tlist_same_exprs is true, then the scan/join target to be applied has
6991  * the same expressions as the existing reltarget, so we need only insert the
6992  * appropriate sortgroupref information.  By avoiding the creation of
6993  * projection paths we save effort both immediately and at plan creation time.
6994  */
6995 static void
6996 apply_scanjoin_target_to_paths(PlannerInfo *root,
6997                                                            RelOptInfo *rel,
6998                                                            List *scanjoin_targets,
6999                                                            List *scanjoin_targets_contain_srfs,
7000                                                            bool scanjoin_target_parallel_safe,
7001                                                            bool tlist_same_exprs)
7002 {
7003         bool            rel_is_partitioned = IS_PARTITIONED_REL(rel);
7004         PathTarget *scanjoin_target;
7005         ListCell   *lc;
7006
7007         /* This recurses, so be paranoid. */
7008         check_stack_depth();
7009
7010         /*
7011          * If the rel is partitioned, we want to drop its existing paths and
7012          * generate new ones.  This function would still be correct if we kept the
7013          * existing paths: we'd modify them to generate the correct target above
7014          * the partitioning Append, and then they'd compete on cost with paths
7015          * generating the target below the Append.  However, in our current cost
7016          * model the latter way is always the same or cheaper cost, so modifying
7017          * the existing paths would just be useless work.  Moreover, when the cost
7018          * is the same, varying roundoff errors might sometimes allow an existing
7019          * path to be picked, resulting in undesirable cross-platform plan
7020          * variations.  So we drop old paths and thereby force the work to be done
7021          * below the Append, except in the case of a non-parallel-safe target.
7022          *
7023          * Some care is needed, because we have to allow generate_gather_paths to
7024          * see the old partial paths in the next stanza.  Hence, zap the main
7025          * pathlist here, then allow generate_gather_paths to add path(s) to the
7026          * main list, and finally zap the partial pathlist.
7027          */
7028         if (rel_is_partitioned)
7029                 rel->pathlist = NIL;
7030
7031         /*
7032          * If the scan/join target is not parallel-safe, partial paths cannot
7033          * generate it.
7034          */
7035         if (!scanjoin_target_parallel_safe)
7036         {
7037                 /*
7038                  * Since we can't generate the final scan/join target in parallel
7039                  * workers, this is our last opportunity to use any partial paths that
7040                  * exist; so build Gather path(s) that use them and emit whatever the
7041                  * current reltarget is.  We don't do this in the case where the
7042                  * target is parallel-safe, since we will be able to generate superior
7043                  * paths by doing it after the final scan/join target has been
7044                  * applied.
7045                  */
7046                 generate_gather_paths(root, rel, false);
7047
7048                 /* Can't use parallel query above this level. */
7049                 rel->partial_pathlist = NIL;
7050                 rel->consider_parallel = false;
7051         }
7052
7053         /* Finish dropping old paths for a partitioned rel, per comment above */
7054         if (rel_is_partitioned)
7055                 rel->partial_pathlist = NIL;
7056
7057         /* Extract SRF-free scan/join target. */
7058         scanjoin_target = linitial_node(PathTarget, scanjoin_targets);
7059
7060         /*
7061          * Apply the SRF-free scan/join target to each existing path.
7062          *
7063          * If the tlist exprs are the same, we can just inject the sortgroupref
7064          * information into the existing pathtargets.  Otherwise, replace each
7065          * path with a projection path that generates the SRF-free scan/join
7066          * target.  This can't change the ordering of paths within rel->pathlist,
7067          * so we just modify the list in place.
7068          */
7069         foreach(lc, rel->pathlist)
7070         {
7071                 Path       *subpath = (Path *) lfirst(lc);
7072
7073                 /* Shouldn't have any parameterized paths anymore */
7074                 Assert(subpath->param_info == NULL);
7075
7076                 if (tlist_same_exprs)
7077                         subpath->pathtarget->sortgrouprefs =
7078                                 scanjoin_target->sortgrouprefs;
7079                 else
7080                 {
7081                         Path       *newpath;
7082
7083                         newpath = (Path *) create_projection_path(root, rel, subpath,
7084                                                                                                           scanjoin_target);
7085                         lfirst(lc) = newpath;
7086                 }
7087         }
7088
7089         /* Likewise adjust the targets for any partial paths. */
7090         foreach(lc, rel->partial_pathlist)
7091         {
7092                 Path       *subpath = (Path *) lfirst(lc);
7093
7094                 /* Shouldn't have any parameterized paths anymore */
7095                 Assert(subpath->param_info == NULL);
7096
7097                 if (tlist_same_exprs)
7098                         subpath->pathtarget->sortgrouprefs =
7099                                 scanjoin_target->sortgrouprefs;
7100                 else
7101                 {
7102                         Path       *newpath;
7103
7104                         newpath = (Path *) create_projection_path(root, rel, subpath,
7105                                                                                                           scanjoin_target);
7106                         lfirst(lc) = newpath;
7107                 }
7108         }
7109
7110         /*
7111          * Now, if final scan/join target contains SRFs, insert ProjectSetPath(s)
7112          * atop each existing path.  (Note that this function doesn't look at the
7113          * cheapest-path fields, which is a good thing because they're bogus right
7114          * now.)
7115          */
7116         if (root->parse->hasTargetSRFs)
7117                 adjust_paths_for_srfs(root, rel,
7118                                                           scanjoin_targets,
7119                                                           scanjoin_targets_contain_srfs);
7120
7121         /*
7122          * Update the rel's target to be the final (with SRFs) scan/join target.
7123          * This now matches the actual output of all the paths, and we might get
7124          * confused in createplan.c if they don't agree.  We must do this now so
7125          * that any append paths made in the next part will use the correct
7126          * pathtarget (cf. create_append_path).
7127          *
7128          * Note that this is also necessary if GetForeignUpperPaths() gets called
7129          * on the final scan/join relation or on any of its children, since the
7130          * FDW might look at the rel's target to create ForeignPaths.
7131          */
7132         rel->reltarget = llast_node(PathTarget, scanjoin_targets);
7133
7134         /*
7135          * If the relation is partitioned, recursively apply the scan/join target
7136          * to all partitions, and generate brand-new Append paths in which the
7137          * scan/join target is computed below the Append rather than above it.
7138          * Since Append is not projection-capable, that might save a separate
7139          * Result node, and it also is important for partitionwise aggregate.
7140          */
7141         if (rel_is_partitioned)
7142         {
7143                 List       *live_children = NIL;
7144                 int                     partition_idx;
7145
7146                 /* Adjust each partition. */
7147                 for (partition_idx = 0; partition_idx < rel->nparts; partition_idx++)
7148                 {
7149                         RelOptInfo *child_rel = rel->part_rels[partition_idx];
7150                         AppendRelInfo **appinfos;
7151                         int                     nappinfos;
7152                         List       *child_scanjoin_targets = NIL;
7153                         ListCell   *lc;
7154
7155                         /* Pruned or dummy children can be ignored. */
7156                         if (child_rel == NULL || IS_DUMMY_REL(child_rel))
7157                                 continue;
7158
7159                         /* Translate scan/join targets for this child. */
7160                         appinfos = find_appinfos_by_relids(root, child_rel->relids,
7161                                                                                            &nappinfos);
7162                         foreach(lc, scanjoin_targets)
7163                         {
7164                                 PathTarget *target = lfirst_node(PathTarget, lc);
7165
7166                                 target = copy_pathtarget(target);
7167                                 target->exprs = (List *)
7168                                         adjust_appendrel_attrs(root,
7169                                                                                    (Node *) target->exprs,
7170                                                                                    nappinfos, appinfos);
7171                                 child_scanjoin_targets = lappend(child_scanjoin_targets,
7172                                                                                                  target);
7173                         }
7174                         pfree(appinfos);
7175
7176                         /* Recursion does the real work. */
7177                         apply_scanjoin_target_to_paths(root, child_rel,
7178                                                                                    child_scanjoin_targets,
7179                                                                                    scanjoin_targets_contain_srfs,
7180                                                                                    scanjoin_target_parallel_safe,
7181                                                                                    tlist_same_exprs);
7182
7183                         /* Save non-dummy children for Append paths. */
7184                         if (!IS_DUMMY_REL(child_rel))
7185                                 live_children = lappend(live_children, child_rel);
7186                 }
7187
7188                 /* Build new paths for this relation by appending child paths. */
7189                 add_paths_to_append_rel(root, rel, live_children);
7190         }
7191
7192         /*
7193          * Consider generating Gather or Gather Merge paths.  We must only do this
7194          * if the relation is parallel safe, and we don't do it for child rels to
7195          * avoid creating multiple Gather nodes within the same plan. We must do
7196          * this after all paths have been generated and before set_cheapest, since
7197          * one of the generated paths may turn out to be the cheapest one.
7198          */
7199         if (rel->consider_parallel && !IS_OTHER_REL(rel))
7200                 generate_gather_paths(root, rel, false);
7201
7202         /*
7203          * Reassess which paths are the cheapest, now that we've potentially added
7204          * new Gather (or Gather Merge) and/or Append (or MergeAppend) paths to
7205          * this relation.
7206          */
7207         set_cheapest(rel);
7208 }
7209
7210 /*
7211  * create_partitionwise_grouping_paths
7212  *
7213  * If the partition keys of input relation are part of the GROUP BY clause, all
7214  * the rows belonging to a given group come from a single partition.  This
7215  * allows aggregation/grouping over a partitioned relation to be broken down
7216  * into aggregation/grouping on each partition.  This should be no worse, and
7217  * often better, than the normal approach.
7218  *
7219  * However, if the GROUP BY clause does not contain all the partition keys,
7220  * rows from a given group may be spread across multiple partitions. In that
7221  * case, we perform partial aggregation for each group, append the results,
7222  * and then finalize aggregation.  This is less certain to win than the
7223  * previous case.  It may win if the PartialAggregate stage greatly reduces
7224  * the number of groups, because fewer rows will pass through the Append node.
7225  * It may lose if we have lots of small groups.
7226  */
7227 static void
7228 create_partitionwise_grouping_paths(PlannerInfo *root,
7229                                                                         RelOptInfo *input_rel,
7230                                                                         RelOptInfo *grouped_rel,
7231                                                                         RelOptInfo *partially_grouped_rel,
7232                                                                         const AggClauseCosts *agg_costs,
7233                                                                         grouping_sets_data *gd,
7234                                                                         PartitionwiseAggregateType patype,
7235                                                                         GroupPathExtraData *extra)
7236 {
7237         int                     nparts = input_rel->nparts;
7238         int                     cnt_parts;
7239         List       *grouped_live_children = NIL;
7240         List       *partially_grouped_live_children = NIL;
7241         PathTarget *target = grouped_rel->reltarget;
7242         bool            partial_grouping_valid = true;
7243
7244         Assert(patype != PARTITIONWISE_AGGREGATE_NONE);
7245         Assert(patype != PARTITIONWISE_AGGREGATE_PARTIAL ||
7246                    partially_grouped_rel != NULL);
7247
7248         /* Add paths for partitionwise aggregation/grouping. */
7249         for (cnt_parts = 0; cnt_parts < nparts; cnt_parts++)
7250         {
7251                 RelOptInfo *child_input_rel = input_rel->part_rels[cnt_parts];
7252                 PathTarget *child_target = copy_pathtarget(target);
7253                 AppendRelInfo **appinfos;
7254                 int                     nappinfos;
7255                 GroupPathExtraData child_extra;
7256                 RelOptInfo *child_grouped_rel;
7257                 RelOptInfo *child_partially_grouped_rel;
7258
7259                 /* Pruned or dummy children can be ignored. */
7260                 if (child_input_rel == NULL || IS_DUMMY_REL(child_input_rel))
7261                         continue;
7262
7263                 /*
7264                  * Copy the given "extra" structure as is and then override the
7265                  * members specific to this child.
7266                  */
7267                 memcpy(&child_extra, extra, sizeof(child_extra));
7268
7269                 appinfos = find_appinfos_by_relids(root, child_input_rel->relids,
7270                                                                                    &nappinfos);
7271
7272                 child_target->exprs = (List *)
7273                         adjust_appendrel_attrs(root,
7274                                                                    (Node *) target->exprs,
7275                                                                    nappinfos, appinfos);
7276
7277                 /* Translate havingQual and targetList. */
7278                 child_extra.havingQual = (Node *)
7279                         adjust_appendrel_attrs(root,
7280                                                                    extra->havingQual,
7281                                                                    nappinfos, appinfos);
7282                 child_extra.targetList = (List *)
7283                         adjust_appendrel_attrs(root,
7284                                                                    (Node *) extra->targetList,
7285                                                                    nappinfos, appinfos);
7286
7287                 /*
7288                  * extra->patype was the value computed for our parent rel; patype is
7289                  * the value for this relation.  For the child, our value is its
7290                  * parent rel's value.
7291                  */
7292                 child_extra.patype = patype;
7293
7294                 /*
7295                  * Create grouping relation to hold fully aggregated grouping and/or
7296                  * aggregation paths for the child.
7297                  */
7298                 child_grouped_rel = make_grouping_rel(root, child_input_rel,
7299                                                                                           child_target,
7300                                                                                           extra->target_parallel_safe,
7301                                                                                           child_extra.havingQual);
7302
7303                 /* Create grouping paths for this child relation. */
7304                 create_ordinary_grouping_paths(root, child_input_rel,
7305                                                                            child_grouped_rel,
7306                                                                            agg_costs, gd, &child_extra,
7307                                                                            &child_partially_grouped_rel);
7308
7309                 if (child_partially_grouped_rel)
7310                 {
7311                         partially_grouped_live_children =
7312                                 lappend(partially_grouped_live_children,
7313                                                 child_partially_grouped_rel);
7314                 }
7315                 else
7316                         partial_grouping_valid = false;
7317
7318                 if (patype == PARTITIONWISE_AGGREGATE_FULL)
7319                 {
7320                         set_cheapest(child_grouped_rel);
7321                         grouped_live_children = lappend(grouped_live_children,
7322                                                                                         child_grouped_rel);
7323                 }
7324
7325                 pfree(appinfos);
7326         }
7327
7328         /*
7329          * Try to create append paths for partially grouped children. For full
7330          * partitionwise aggregation, we might have paths in the partial_pathlist
7331          * if parallel aggregation is possible.  For partial partitionwise
7332          * aggregation, we may have paths in both pathlist and partial_pathlist.
7333          *
7334          * NB: We must have a partially grouped path for every child in order to
7335          * generate a partially grouped path for this relation.
7336          */
7337         if (partially_grouped_rel && partial_grouping_valid)
7338         {
7339                 Assert(partially_grouped_live_children != NIL);
7340
7341                 add_paths_to_append_rel(root, partially_grouped_rel,
7342                                                                 partially_grouped_live_children);
7343
7344                 /*
7345                  * We need call set_cheapest, since the finalization step will use the
7346                  * cheapest path from the rel.
7347                  */
7348                 if (partially_grouped_rel->pathlist)
7349                         set_cheapest(partially_grouped_rel);
7350         }
7351
7352         /* If possible, create append paths for fully grouped children. */
7353         if (patype == PARTITIONWISE_AGGREGATE_FULL)
7354         {
7355                 Assert(grouped_live_children != NIL);
7356
7357                 add_paths_to_append_rel(root, grouped_rel, grouped_live_children);
7358         }
7359 }
7360
7361 /*
7362  * group_by_has_partkey
7363  *
7364  * Returns true, if all the partition keys of the given relation are part of
7365  * the GROUP BY clauses, false otherwise.
7366  */
7367 static bool
7368 group_by_has_partkey(RelOptInfo *input_rel,
7369                                          List *targetList,
7370                                          List *groupClause)
7371 {
7372         List       *groupexprs = get_sortgrouplist_exprs(groupClause, targetList);
7373         int                     cnt = 0;
7374         int                     partnatts;
7375
7376         /* Input relation should be partitioned. */
7377         Assert(input_rel->part_scheme);
7378
7379         /* Rule out early, if there are no partition keys present. */
7380         if (!input_rel->partexprs)
7381                 return false;
7382
7383         partnatts = input_rel->part_scheme->partnatts;
7384
7385         for (cnt = 0; cnt < partnatts; cnt++)
7386         {
7387                 List       *partexprs = input_rel->partexprs[cnt];
7388                 ListCell   *lc;
7389                 bool            found = false;
7390
7391                 foreach(lc, partexprs)
7392                 {
7393                         Expr       *partexpr = lfirst(lc);
7394
7395                         if (list_member(groupexprs, partexpr))
7396                         {
7397                                 found = true;
7398                                 break;
7399                         }
7400                 }
7401
7402                 /*
7403                  * If none of the partition key expressions match with any of the
7404                  * GROUP BY expression, return false.
7405                  */
7406                 if (!found)
7407                         return false;
7408         }
7409
7410         return true;
7411 }