orderedsetaggs.c

Go to the documentation of this file.

/*-------------------------------------------------------------------------
 *
 * orderedsetaggs.c
 *      Ordered-set aggregate functions.
 *
 * Portions Copyright (c) 1996-2020, PostgreSQL Global Development Group
 * Portions Copyright (c) 1994, Regents of the University of California
 *
 *
 * IDENTIFICATION
 *    src/backend/utils/adt/orderedsetaggs.c
 *
 *-------------------------------------------------------------------------
 */
#include "postgres.h"

#include <math.h>

#include "catalog/pg_aggregate.h"
#include "catalog/pg_operator.h"
#include "catalog/pg_type.h"
#include "executor/executor.h"
#include "miscadmin.h"
#include "nodes/nodeFuncs.h"
#include "optimizer/optimizer.h"
#include "utils/array.h"
#include "utils/builtins.h"
#include "utils/lsyscache.h"
#include "utils/memutils.h"
#include "utils/timestamp.h"
#include "utils/tuplesort.h"


/*
 * Generic support for ordered-set aggregates
 *
 * The state for an ordered-set aggregate is divided into a per-group struct
 * (which is the internal-type transition state datum returned to nodeAgg.c)
 * and a per-query struct, which contains data and sub-objects that we can
 * create just once per query because they will not change across groups.
 * The per-query struct and subsidiary data live in the executor's per-query
 * memory context, and go away implicitly at ExecutorEnd().
 *
 * These structs are set up during the first call of the transition function.
 * Because we allow nodeAgg.c to merge ordered-set aggregates (but not
 * hypothetical aggregates) with identical inputs and transition functions,
 * this info must not depend on the particular aggregate (ie, particular
 * final-function), nor on the direct argument(s) of the aggregate.
 */

typedef struct OSAPerQueryState
{
    /* Representative Aggref for this aggregate: */
    Aggref     *aggref;
    /* Memory context containing this struct and other per-query data: */
    MemoryContext qcontext;
    /* Context for expression evaluation */
    ExprContext *econtext;
    /* Do we expect multiple final-function calls within one group? */
    bool        rescan_needed;

    /* These fields are used only when accumulating tuples: */

    /* Tuple descriptor for tuples inserted into sortstate: */
    TupleDesc   tupdesc;
    /* Tuple slot we can use for inserting/extracting tuples: */
    TupleTableSlot *tupslot;
    /* Per-sort-column sorting information */
    int         numSortCols;
    AttrNumber *sortColIdx;
    Oid        *sortOperators;
    Oid        *eqOperators;
    Oid        *sortCollations;
    bool       *sortNullsFirsts;
    /* Equality operator call info, created only if needed: */
    ExprState  *compareTuple;

    /* These fields are used only when accumulating datums: */

    /* Info about datatype of datums being sorted: */
    Oid         sortColType;
    int16       typLen;
    bool        typByVal;
    char        typAlign;
    /* Info about sort ordering: */
    Oid         sortOperator;
    Oid         eqOperator;
    Oid         sortCollation;
    bool        sortNullsFirst;
    /* Equality operator call info, created only if needed: */
    FmgrInfo    equalfn;
} OSAPerQueryState;

typedef struct OSAPerGroupState
{
    /* Link to the per-query state for this aggregate: */
    OSAPerQueryState *qstate;
    /* Memory context containing per-group data: */
    MemoryContext gcontext;
    /* Sort object we're accumulating data in: */
    Tuplesortstate *sortstate;
    /* Number of normal rows inserted into sortstate: */
    int64       number_of_rows;
    /* Have we already done tuplesort_performsort? */
    bool        sort_done;
} OSAPerGroupState;

static void ordered_set_shutdown(Datum arg);


/*
 * Set up working state for an ordered-set aggregate
 */
static OSAPerGroupState *
ordered_set_startup(FunctionCallInfo fcinfo, bool use_tuples)
{
    OSAPerGroupState *osastate;
    OSAPerQueryState *qstate;
    MemoryContext gcontext;
    MemoryContext oldcontext;

    /*
     * Check we're called as aggregate (and not a window function), and get
     * the Agg node's group-lifespan context (which might change from group to
     * group, so we shouldn't cache it in the per-query state).
     */
    if (AggCheckCallContext(fcinfo, &gcontext) != AGG_CONTEXT_AGGREGATE)
        elog(ERROR, "ordered-set aggregate called in non-aggregate context");

    /*
     * We keep a link to the per-query state in fn_extra; if it's not there,
     * create it, and do the per-query setup we need.
     */
    qstate = (OSAPerQueryState *) fcinfo->flinfo->fn_extra;
    if (qstate == NULL)
    {
        Aggref     *aggref;
        MemoryContext qcontext;
        List       *sortlist;
        int         numSortCols;

        /* Get the Aggref so we can examine aggregate's arguments */
        aggref = AggGetAggref(fcinfo);
        if (!aggref)
            elog(ERROR, "ordered-set aggregate called in non-aggregate context");
        if (!AGGKIND_IS_ORDERED_SET(aggref->aggkind))
            elog(ERROR, "ordered-set aggregate support function called for non-ordered-set aggregate");

        /*
         * Prepare per-query structures in the fn_mcxt, which we assume is the
         * executor's per-query context; in any case it's the right place to
         * keep anything found via fn_extra.
         */
        qcontext = fcinfo->flinfo->fn_mcxt;
        oldcontext = MemoryContextSwitchTo(qcontext);

        qstate = (OSAPerQueryState *) palloc0(sizeof(OSAPerQueryState));
        qstate->aggref = aggref;
        qstate->qcontext = qcontext;

        /* We need to support rescans if the trans state is shared */
        qstate->rescan_needed = AggStateIsShared(fcinfo);

        /* Extract the sort information */
        sortlist = aggref->aggorder;
        numSortCols = list_length(sortlist);

        if (use_tuples)
        {
            bool        ishypothetical = (aggref->aggkind == AGGKIND_HYPOTHETICAL);
            ListCell   *lc;
            int         i;

            if (ishypothetical)
                numSortCols++;  /* make space for flag column */
            qstate->numSortCols = numSortCols;
            qstate->sortColIdx = (AttrNumber *) palloc(numSortCols * sizeof(AttrNumber));
            qstate->sortOperators = (Oid *) palloc(numSortCols * sizeof(Oid));
            qstate->eqOperators = (Oid *) palloc(numSortCols * sizeof(Oid));
            qstate->sortCollations = (Oid *) palloc(numSortCols * sizeof(Oid));
            qstate->sortNullsFirsts = (bool *) palloc(numSortCols * sizeof(bool));

            i = 0;
            foreach(lc, sortlist)
            {
                SortGroupClause *sortcl = (SortGroupClause *) lfirst(lc);
                TargetEntry *tle = get_sortgroupclause_tle(sortcl,
                                                           aggref->args);

                /* the parser should have made sure of this */
                Assert(OidIsValid(sortcl->sortop));

                qstate->sortColIdx[i] = tle->resno;
                qstate->sortOperators[i] = sortcl->sortop;
                qstate->eqOperators[i] = sortcl->eqop;
                qstate->sortCollations[i] = exprCollation((Node *) tle->expr);
                qstate->sortNullsFirsts[i] = sortcl->nulls_first;
                i++;
            }

            if (ishypothetical)
            {
                /* Add an integer flag column as the last sort column */
                qstate->sortColIdx[i] = list_length(aggref->args) + 1;
                qstate->sortOperators[i] = Int4LessOperator;
                qstate->eqOperators[i] = Int4EqualOperator;
                qstate->sortCollations[i] = InvalidOid;
                qstate->sortNullsFirsts[i] = false;
                i++;
            }

            Assert(i == numSortCols);

            /*
             * Get a tupledesc corresponding to the aggregated inputs
             * (including sort expressions) of the agg.
             */
            qstate->tupdesc = ExecTypeFromTL(aggref->args);

            /* If we need a flag column, hack the tupledesc to include that */
            if (ishypothetical)
            {
                TupleDesc   newdesc;
                int         natts = qstate->tupdesc->natts;

                newdesc = CreateTemplateTupleDesc(natts + 1);
                for (i = 1; i <= natts; i++)
                    TupleDescCopyEntry(newdesc, i, qstate->tupdesc, i);

                TupleDescInitEntry(newdesc,
                                   (AttrNumber) ++natts,
                                   "flag",
                                   INT4OID,
                                   -1,
                                   0);

                FreeTupleDesc(qstate->tupdesc);
                qstate->tupdesc = newdesc;
            }

            /* Create slot we'll use to store/retrieve rows */
            qstate->tupslot = MakeSingleTupleTableSlot(qstate->tupdesc,
                                                       &TTSOpsMinimalTuple);
        }
        else
        {
            /* Sort single datums */
            SortGroupClause *sortcl;
            TargetEntry *tle;

            if (numSortCols != 1 || aggref->aggkind == AGGKIND_HYPOTHETICAL)
                elog(ERROR, "ordered-set aggregate support function does not support multiple aggregated columns");

            sortcl = (SortGroupClause *) linitial(sortlist);
            tle = get_sortgroupclause_tle(sortcl, aggref->args);

            /* the parser should have made sure of this */
            Assert(OidIsValid(sortcl->sortop));

            /* Save sort ordering info */
            qstate->sortColType = exprType((Node *) tle->expr);
            qstate->sortOperator = sortcl->sortop;
            qstate->eqOperator = sortcl->eqop;
            qstate->sortCollation = exprCollation((Node *) tle->expr);
            qstate->sortNullsFirst = sortcl->nulls_first;

            /* Save datatype info */
            get_typlenbyvalalign(qstate->sortColType,
                                 &qstate->typLen,
                                 &qstate->typByVal,
                                 &qstate->typAlign);
        }

        fcinfo->flinfo->fn_extra = (void *) qstate;

        MemoryContextSwitchTo(oldcontext);
    }

    /* Now build the stuff we need in group-lifespan context */
    oldcontext = MemoryContextSwitchTo(gcontext);

    osastate = (OSAPerGroupState *) palloc(sizeof(OSAPerGroupState));
    osastate->qstate = qstate;
    osastate->gcontext = gcontext;

    /*
     * Initialize tuplesort object.
     */
    if (use_tuples)
        osastate->sortstate = tuplesort_begin_heap(qstate->tupdesc,
                                                   qstate->numSortCols,
                                                   qstate->sortColIdx,
                                                   qstate->sortOperators,
                                                   qstate->sortCollations,
                                                   qstate->sortNullsFirsts,
                                                   work_mem,
                                                   NULL,
                                                   qstate->rescan_needed);
    else
        osastate->sortstate = tuplesort_begin_datum(qstate->sortColType,
                                                    qstate->sortOperator,
                                                    qstate->sortCollation,
                                                    qstate->sortNullsFirst,
                                                    work_mem,
                                                    NULL,
                                                    qstate->rescan_needed);

    osastate->number_of_rows = 0;
    osastate->sort_done = false;

    /* Now register a shutdown callback to clean things up at end of group */
    AggRegisterCallback(fcinfo,
                        ordered_set_shutdown,
                        PointerGetDatum(osastate));

    MemoryContextSwitchTo(oldcontext);

    return osastate;
}

/*
 * Clean up when evaluation of an ordered-set aggregate is complete.
 *
 * We don't need to bother freeing objects in the per-group memory context,
 * since that will get reset anyway by nodeAgg.c; nor should we free anything
 * in the per-query context, which will get cleared (if this was the last
 * group) by ExecutorEnd.  But we must take care to release any potential
 * non-memory resources.
 *
 * In the case where we're not expecting multiple finalfn calls, we could
 * arguably rely on the finalfn to clean up; but it's easier and more testable
 * if we just do it the same way in either case.
 */
static void
ordered_set_shutdown(Datum arg)
{
    OSAPerGroupState *osastate = (OSAPerGroupState *) DatumGetPointer(arg);

    /* Tuplesort object might have temp files. */
    if (osastate->sortstate)
        tuplesort_end(osastate->sortstate);
    osastate->sortstate = NULL;
    /* The tupleslot probably can't be holding a pin, but let's be safe. */
    if (osastate->qstate->tupslot)
        ExecClearTuple(osastate->qstate->tupslot);
}


/*
 * Generic transition function for ordered-set aggregates
 * with a single input column in which we want to suppress nulls
 */
Datum
ordered_set_transition(PG_FUNCTION_ARGS)
{
    OSAPerGroupState *osastate;

    /* If first call, create the transition state workspace */
    if (PG_ARGISNULL(0))
        osastate = ordered_set_startup(fcinfo, false);
    else
        osastate = (OSAPerGroupState *) PG_GETARG_POINTER(0);

    /* Load the datum into the tuplesort object, but only if it's not null */
    if (!PG_ARGISNULL(1))
    {
        tuplesort_putdatum(osastate->sortstate, PG_GETARG_DATUM(1), false);
        osastate->number_of_rows++;
    }

    PG_RETURN_POINTER(osastate);
}

/*
 * Generic transition function for ordered-set aggregates
 * with (potentially) multiple aggregated input columns
 */
Datum
ordered_set_transition_multi(PG_FUNCTION_ARGS)
{
    OSAPerGroupState *osastate;
    TupleTableSlot *slot;
    int         nargs;
    int         i;

    /* If first call, create the transition state workspace */
    if (PG_ARGISNULL(0))
        osastate = ordered_set_startup(fcinfo, true);
    else
        osastate = (OSAPerGroupState *) PG_GETARG_POINTER(0);

    /* Form a tuple from all the other inputs besides the transition value */
    slot = osastate->qstate->tupslot;
    ExecClearTuple(slot);
    nargs = PG_NARGS() - 1;
    for (i = 0; i < nargs; i++)
    {
        slot->tts_values[i] = PG_GETARG_DATUM(i + 1);
        slot->tts_isnull[i] = PG_ARGISNULL(i + 1);
    }
    if (osastate->qstate->aggref->aggkind == AGGKIND_HYPOTHETICAL)
    {
        /* Add a zero flag value to mark this row as a normal input row */
        slot->tts_values[i] = Int32GetDatum(0);
        slot->tts_isnull[i] = false;
        i++;
    }
    Assert(i == slot->tts_tupleDescriptor->natts);
    ExecStoreVirtualTuple(slot);

    /* Load the row into the tuplesort object */
    tuplesort_puttupleslot(osastate->sortstate, slot);
    osastate->number_of_rows++;

    PG_RETURN_POINTER(osastate);
}


/*
 * percentile_disc(float8) within group(anyelement) - discrete percentile
 */
Datum
percentile_disc_final(PG_FUNCTION_ARGS)
{
    OSAPerGroupState *osastate;
    double      percentile;
    Datum       val;
    bool        isnull;
    int64       rownum;

    Assert(AggCheckCallContext(fcinfo, NULL) == AGG_CONTEXT_AGGREGATE);

    /* Get and check the percentile argument */
    if (PG_ARGISNULL(1))
        PG_RETURN_NULL();

    percentile = PG_GETARG_FLOAT8(1);

    if (percentile < 0 || percentile > 1 || isnan(percentile))
        ereport(ERROR,
                (errcode(ERRCODE_NUMERIC_VALUE_OUT_OF_RANGE),
                 errmsg("percentile value %g is not between 0 and 1",
                        percentile)));

    /* If there were no regular rows, the result is NULL */
    if (PG_ARGISNULL(0))
        PG_RETURN_NULL();

    osastate = (OSAPerGroupState *) PG_GETARG_POINTER(0);

    /* number_of_rows could be zero if we only saw NULL input values */
    if (osastate->number_of_rows == 0)
        PG_RETURN_NULL();

    /* Finish the sort, or rescan if we already did */
    if (!osastate->sort_done)
    {
        tuplesort_performsort(osastate->sortstate);
        osastate->sort_done = true;
    }
    else
        tuplesort_rescan(osastate->sortstate);

    /*----------
     * We need the smallest K such that (K/N) >= percentile.
     * N>0, therefore K >= N*percentile, therefore K = ceil(N*percentile).
     * So we skip K-1 rows (if K>0) and return the next row fetched.
     *----------
     */
    rownum = (int64) ceil(percentile * osastate->number_of_rows);
    Assert(rownum <= osastate->number_of_rows);

    if (rownum > 1)
    {
        if (!tuplesort_skiptuples(osastate->sortstate, rownum - 1, true))
            elog(ERROR, "missing row in percentile_disc");
    }

    if (!tuplesort_getdatum(osastate->sortstate, true, &val, &isnull, NULL))
        elog(ERROR, "missing row in percentile_disc");

    /* We shouldn't have stored any nulls, but do the right thing anyway */
    if (isnull)
        PG_RETURN_NULL();
    else
        PG_RETURN_DATUM(val);
}


/*
 * For percentile_cont, we need a way to interpolate between consecutive
 * values. Use a helper function for that, so that we can share the rest
 * of the code between types.
 */
typedef Datum (*LerpFunc) (Datum lo, Datum hi, double pct);

static Datum
float8_lerp(Datum lo, Datum hi, double pct)
{
    double      loval = DatumGetFloat8(lo);
    double      hival = DatumGetFloat8(hi);

    return Float8GetDatum(loval + (pct * (hival - loval)));
}

static Datum
interval_lerp(Datum lo, Datum hi, double pct)
{
    Datum       diff_result = DirectFunctionCall2(interval_mi, hi, lo);
    Datum       mul_result = DirectFunctionCall2(interval_mul,
                                                 diff_result,
                                                 Float8GetDatumFast(pct));

    return DirectFunctionCall2(interval_pl, mul_result, lo);
}

/*
 * Continuous percentile
 */
static Datum
percentile_cont_final_common(FunctionCallInfo fcinfo,
                             Oid expect_type,
                             LerpFunc lerpfunc)
{
    OSAPerGroupState *osastate;
    double      percentile;
    int64       first_row = 0;
    int64       second_row = 0;
    Datum       val;
    Datum       first_val;
    Datum       second_val;
    double      proportion;
    bool        isnull;

    Assert(AggCheckCallContext(fcinfo, NULL) == AGG_CONTEXT_AGGREGATE);

    /* Get and check the percentile argument */
    if (PG_ARGISNULL(1))
        PG_RETURN_NULL();

    percentile = PG_GETARG_FLOAT8(1);

    if (percentile < 0 || percentile > 1 || isnan(percentile))
        ereport(ERROR,
                (errcode(ERRCODE_NUMERIC_VALUE_OUT_OF_RANGE),
                 errmsg("percentile value %g is not between 0 and 1",
                        percentile)));

    /* If there were no regular rows, the result is NULL */
    if (PG_ARGISNULL(0))
        PG_RETURN_NULL();

    osastate = (OSAPerGroupState *) PG_GETARG_POINTER(0);

    /* number_of_rows could be zero if we only saw NULL input values */
    if (osastate->number_of_rows == 0)
        PG_RETURN_NULL();

    Assert(expect_type == osastate->qstate->sortColType);

    /* Finish the sort, or rescan if we already did */
    if (!osastate->sort_done)
    {
        tuplesort_performsort(osastate->sortstate);
        osastate->sort_done = true;
    }
    else
        tuplesort_rescan(osastate->sortstate);

    first_row = floor(percentile * (osastate->number_of_rows - 1));
    second_row = ceil(percentile * (osastate->number_of_rows - 1));

    Assert(first_row < osastate->number_of_rows);

    if (!tuplesort_skiptuples(osastate->sortstate, first_row, true))
        elog(ERROR, "missing row in percentile_cont");

    if (!tuplesort_getdatum(osastate->sortstate, true, &first_val, &isnull, NULL))
        elog(ERROR, "missing row in percentile_cont");
    if (isnull)
        PG_RETURN_NULL();

    if (first_row == second_row)
    {
        val = first_val;
    }
    else
    {
        if (!tuplesort_getdatum(osastate->sortstate, true, &second_val, &isnull, NULL))
            elog(ERROR, "missing row in percentile_cont");

        if (isnull)
            PG_RETURN_NULL();

        proportion = (percentile * (osastate->number_of_rows - 1)) - first_row;
        val = lerpfunc(first_val, second_val, proportion);
    }

    PG_RETURN_DATUM(val);
}

/*
 * percentile_cont(float8) within group (float8)    - continuous percentile
 */
Datum
percentile_cont_float8_final(PG_FUNCTION_ARGS)
{
    return percentile_cont_final_common(fcinfo, FLOAT8OID, float8_lerp);
}

/*
 * percentile_cont(float8) within group (interval)  - continuous percentile
 */
Datum
percentile_cont_interval_final(PG_FUNCTION_ARGS)
{
    return percentile_cont_final_common(fcinfo, INTERVALOID, interval_lerp);
}


/*
 * Support code for handling arrays of percentiles
 *
 * Note: in each pct_info entry, second_row should be equal to or
 * exactly one more than first_row.
 */
struct pct_info
{
    int64       first_row;      /* first row to sample */
    int64       second_row;     /* possible second row to sample */
    double      proportion;     /* interpolation fraction */
    int         idx;            /* index of this item in original array */
};

/*
 * Sort comparator to sort pct_infos by first_row then second_row
 */
static int
pct_info_cmp(const void *pa, const void *pb)
{
    const struct pct_info *a = (const struct pct_info *) pa;
    const struct pct_info *b = (const struct pct_info *) pb;

    if (a->first_row != b->first_row)
        return (a->first_row < b->first_row) ? -1 : 1;
    if (a->second_row != b->second_row)
        return (a->second_row < b->second_row) ? -1 : 1;
    return 0;
}

/*
 * Construct array showing which rows to sample for percentiles.
 */
static struct pct_info *
setup_pct_info(int num_percentiles,
               Datum *percentiles_datum,
               bool *percentiles_null,
               int64 rowcount,
               bool continuous)
{
    struct pct_info *pct_info;
    int         i;

    pct_info = (struct pct_info *) palloc(num_percentiles * sizeof(struct pct_info));

    for (i = 0; i < num_percentiles; i++)
    {
        pct_info[i].idx = i;

        if (percentiles_null[i])
        {
            /* dummy entry for any NULL in array */
            pct_info[i].first_row = 0;
            pct_info[i].second_row = 0;
            pct_info[i].proportion = 0;
        }
        else
        {
            double      p = DatumGetFloat8(percentiles_datum[i]);

            if (p < 0 || p > 1 || isnan(p))
                ereport(ERROR,
                        (errcode(ERRCODE_NUMERIC_VALUE_OUT_OF_RANGE),
                         errmsg("percentile value %g is not between 0 and 1",
                                p)));

            if (continuous)
            {
                pct_info[i].first_row = 1 + floor(p * (rowcount - 1));
                pct_info[i].second_row = 1 + ceil(p * (rowcount - 1));
                pct_info[i].proportion = (p * (rowcount - 1)) - floor(p * (rowcount - 1));
            }
            else
            {
                /*----------
                 * We need the smallest K such that (K/N) >= percentile.
                 * N>0, therefore K >= N*percentile, therefore
                 * K = ceil(N*percentile); but not less than 1.
                 *----------
                 */
                int64       row = (int64) ceil(p * rowcount);

                row = Max(1, row);
                pct_info[i].first_row = row;
                pct_info[i].second_row = row;
                pct_info[i].proportion = 0;
            }
        }
    }

    /*
     * The parameter array wasn't necessarily in sorted order, but we need to
     * visit the rows in order, so sort by first_row/second_row.
     */
    qsort(pct_info, num_percentiles, sizeof(struct pct_info), pct_info_cmp);

    return pct_info;
}

/*
 * percentile_disc(float8[]) within group (anyelement)  - discrete percentiles
 */
Datum
percentile_disc_multi_final(PG_FUNCTION_ARGS)
{
    OSAPerGroupState *osastate;
    ArrayType  *param;
    Datum      *percentiles_datum;
    bool       *percentiles_null;
    int         num_percentiles;
    struct pct_info *pct_info;
    Datum      *result_datum;
    bool       *result_isnull;
    int64       rownum = 0;
    Datum       val = (Datum) 0;
    bool        isnull = true;
    int         i;

    Assert(AggCheckCallContext(fcinfo, NULL) == AGG_CONTEXT_AGGREGATE);

    /* If there were no regular rows, the result is NULL */
    if (PG_ARGISNULL(0))
        PG_RETURN_NULL();

    osastate = (OSAPerGroupState *) PG_GETARG_POINTER(0);

    /* number_of_rows could be zero if we only saw NULL input values */
    if (osastate->number_of_rows == 0)
        PG_RETURN_NULL();

    /* Deconstruct the percentile-array input */
    if (PG_ARGISNULL(1))
        PG_RETURN_NULL();
    param = PG_GETARG_ARRAYTYPE_P(1);

    deconstruct_array(param, FLOAT8OID,
    /* hard-wired info on type float8 */
                      8, FLOAT8PASSBYVAL, 'd',
                      &percentiles_datum,
                      &percentiles_null,
                      &num_percentiles);

    if (num_percentiles == 0)
        PG_RETURN_POINTER(construct_empty_array(osastate->qstate->sortColType));

    pct_info = setup_pct_info(num_percentiles,
                              percentiles_datum,
                              percentiles_null,
                              osastate->number_of_rows,
                              false);

    result_datum = (Datum *) palloc(num_percentiles * sizeof(Datum));
    result_isnull = (bool *) palloc(num_percentiles * sizeof(bool));

    /*
     * Start by dealing with any nulls in the param array - those are sorted
     * to the front on row=0, so set the corresponding result indexes to null
     */
    for (i = 0; i < num_percentiles; i++)
    {
        int         idx = pct_info[i].idx;

        if (pct_info[i].first_row > 0)
            break;

        result_datum[idx] = (Datum) 0;
        result_isnull[idx] = true;
    }

    /*
     * If there's anything left after doing the nulls, then grind the input
     * and extract the needed values
     */
    if (i < num_percentiles)
    {
        /* Finish the sort, or rescan if we already did */
        if (!osastate->sort_done)
        {
            tuplesort_performsort(osastate->sortstate);
            osastate->sort_done = true;
        }
        else
            tuplesort_rescan(osastate->sortstate);

        for (; i < num_percentiles; i++)
        {
            int64       target_row = pct_info[i].first_row;
            int         idx = pct_info[i].idx;

            /* Advance to target row, if not already there */
            if (target_row > rownum)
            {
                if (!tuplesort_skiptuples(osastate->sortstate, target_row - rownum - 1, true))
                    elog(ERROR, "missing row in percentile_disc");

                if (!tuplesort_getdatum(osastate->sortstate, true, &val, &isnull, NULL))
                    elog(ERROR, "missing row in percentile_disc");

                rownum = target_row;
            }

            result_datum[idx] = val;
            result_isnull[idx] = isnull;
        }
    }

    /* We make the output array the same shape as the input */
    PG_RETURN_POINTER(construct_md_array(result_datum, result_isnull,
                                         ARR_NDIM(param),
                                         ARR_DIMS(param),
                                         ARR_LBOUND(param),
                                         osastate->qstate->sortColType,
                                         osastate->qstate->typLen,
                                         osastate->qstate->typByVal,
                                         osastate->qstate->typAlign));
}

/*
 * percentile_cont(float8[]) within group ()    - continuous percentiles
 */
static Datum
percentile_cont_multi_final_common(FunctionCallInfo fcinfo,
                                   Oid expect_type,
                                   int16 typLen, bool typByVal, char typAlign,
                                   LerpFunc lerpfunc)
{
    OSAPerGroupState *osastate;
    ArrayType  *param;
    Datum      *percentiles_datum;
    bool       *percentiles_null;
    int         num_percentiles;
    struct pct_info *pct_info;
    Datum      *result_datum;
    bool       *result_isnull;
    int64       rownum = 0;
    Datum       first_val = (Datum) 0;
    Datum       second_val = (Datum) 0;
    bool        isnull;
    int         i;

    Assert(AggCheckCallContext(fcinfo, NULL) == AGG_CONTEXT_AGGREGATE);

    /* If there were no regular rows, the result is NULL */
    if (PG_ARGISNULL(0))
        PG_RETURN_NULL();

    osastate = (OSAPerGroupState *) PG_GETARG_POINTER(0);

    /* number_of_rows could be zero if we only saw NULL input values */
    if (osastate->number_of_rows == 0)
        PG_RETURN_NULL();

    Assert(expect_type == osastate->qstate->sortColType);

    /* Deconstruct the percentile-array input */
    if (PG_ARGISNULL(1))
        PG_RETURN_NULL();
    param = PG_GETARG_ARRAYTYPE_P(1);

    deconstruct_array(param, FLOAT8OID,
    /* hard-wired info on type float8 */
                      8, FLOAT8PASSBYVAL, 'd',
                      &percentiles_datum,
                      &percentiles_null,
                      &num_percentiles);

    if (num_percentiles == 0)
        PG_RETURN_POINTER(construct_empty_array(osastate->qstate->sortColType));

    pct_info = setup_pct_info(num_percentiles,
                              percentiles_datum,
                              percentiles_null,
                              osastate->number_of_rows,
                              true);

    result_datum = (Datum *) palloc(num_percentiles * sizeof(Datum));
    result_isnull = (bool *) palloc(num_percentiles * sizeof(bool));

    /*
     * Start by dealing with any nulls in the param array - those are sorted
     * to the front on row=0, so set the corresponding result indexes to null
     */
    for (i = 0; i < num_percentiles; i++)
    {
        int         idx = pct_info[i].idx;

        if (pct_info[i].first_row > 0)
            break;

        result_datum[idx] = (Datum) 0;
        result_isnull[idx] = true;
    }

    /*
     * If there's anything left after doing the nulls, then grind the input
     * and extract the needed values
     */
    if (i < num_percentiles)
    {
        /* Finish the sort, or rescan if we already did */
        if (!osastate->sort_done)
        {
            tuplesort_performsort(osastate->sortstate);
            osastate->sort_done = true;
        }
        else
            tuplesort_rescan(osastate->sortstate);

        for (; i < num_percentiles; i++)
        {
            int64       first_row = pct_info[i].first_row;
            int64       second_row = pct_info[i].second_row;
            int         idx = pct_info[i].idx;

            /*
             * Advance to first_row, if not already there.  Note that we might
             * already have rownum beyond first_row, in which case first_val
             * is already correct.  (This occurs when interpolating between
             * the same two input rows as for the previous percentile.)
             */
            if (first_row > rownum)
            {
                if (!tuplesort_skiptuples(osastate->sortstate, first_row - rownum - 1, true))
                    elog(ERROR, "missing row in percentile_cont");

                if (!tuplesort_getdatum(osastate->sortstate, true, &first_val,
                                        &isnull, NULL) || isnull)
                    elog(ERROR, "missing row in percentile_cont");

                rownum = first_row;
                /* Always advance second_val to be latest input value */
                second_val = first_val;
            }
            else if (first_row == rownum)
            {
                /*
                 * We are already at the desired row, so we must previously
                 * have read its value into second_val (and perhaps first_val
                 * as well, but this assignment is harmless in that case).
                 */
                first_val = second_val;
            }

            /* Fetch second_row if needed */
            if (second_row > rownum)
            {
                if (!tuplesort_getdatum(osastate->sortstate, true, &second_val,
                                        &isnull, NULL) || isnull)
                    elog(ERROR, "missing row in percentile_cont");
                rownum++;
            }
            /* We should now certainly be on second_row exactly */
            Assert(second_row == rownum);

            /* Compute appropriate result */
            if (second_row > first_row)
                result_datum[idx] = lerpfunc(first_val, second_val,
                                             pct_info[i].proportion);
            else
                result_datum[idx] = first_val;

            result_isnull[idx] = false;
        }
    }

    /* We make the output array the same shape as the input */
    PG_RETURN_POINTER(construct_md_array(result_datum, result_isnull,
                                         ARR_NDIM(param),
                                         ARR_DIMS(param), ARR_LBOUND(param),
                                         expect_type,
                                         typLen,
                                         typByVal,
                                         typAlign));
}

/*
 * percentile_cont(float8[]) within group (float8)  - continuous percentiles
 */
Datum
percentile_cont_float8_multi_final(PG_FUNCTION_ARGS)
{
    return percentile_cont_multi_final_common(fcinfo,
                                              FLOAT8OID,
    /* hard-wired info on type float8 */
                                              8, FLOAT8PASSBYVAL, 'd',
                                              float8_lerp);
}

/*
 * percentile_cont(float8[]) within group (interval)  - continuous percentiles
 */
Datum
percentile_cont_interval_multi_final(PG_FUNCTION_ARGS)
{
    return percentile_cont_multi_final_common(fcinfo,
                                              INTERVALOID,
    /* hard-wired info on type interval */
                                              16, false, 'd',
                                              interval_lerp);
}


/*
 * mode() within group (anyelement) - most common value
 */
Datum
mode_final(PG_FUNCTION_ARGS)
{
    OSAPerGroupState *osastate;
    Datum       val;
    bool        isnull;
    Datum       mode_val = (Datum) 0;
    int64       mode_freq = 0;
    Datum       last_val = (Datum) 0;
    int64       last_val_freq = 0;
    bool        last_val_is_mode = false;
    FmgrInfo   *equalfn;
    Datum       abbrev_val = (Datum) 0;
    Datum       last_abbrev_val = (Datum) 0;
    bool        shouldfree;

    Assert(AggCheckCallContext(fcinfo, NULL) == AGG_CONTEXT_AGGREGATE);

    /* If there were no regular rows, the result is NULL */
    if (PG_ARGISNULL(0))
        PG_RETURN_NULL();

    osastate = (OSAPerGroupState *) PG_GETARG_POINTER(0);

    /* number_of_rows could be zero if we only saw NULL input values */
    if (osastate->number_of_rows == 0)
        PG_RETURN_NULL();

    /* Look up the equality function for the datatype, if we didn't already */
    equalfn = &(osastate->qstate->equalfn);
    if (!OidIsValid(equalfn->fn_oid))
        fmgr_info_cxt(get_opcode(osastate->qstate->eqOperator), equalfn,
                      osastate->qstate->qcontext);

    shouldfree = !(osastate->qstate->typByVal);

    /* Finish the sort, or rescan if we already did */
    if (!osastate->sort_done)
    {
        tuplesort_performsort(osastate->sortstate);
        osastate->sort_done = true;
    }
    else
        tuplesort_rescan(osastate->sortstate);

    /* Scan tuples and count frequencies */
    while (tuplesort_getdatum(osastate->sortstate, true, &val, &isnull, &abbrev_val))
    {
        /* we don't expect any nulls, but ignore them if found */
        if (isnull)
            continue;

        if (last_val_freq == 0)
        {
            /* first nonnull value - it's the mode for now */
            mode_val = last_val = val;
            mode_freq = last_val_freq = 1;
            last_val_is_mode = true;
            last_abbrev_val = abbrev_val;
        }
        else if (abbrev_val == last_abbrev_val &&
                 DatumGetBool(FunctionCall2Coll(equalfn, PG_GET_COLLATION(), val, last_val)))
        {
            /* value equal to previous value, count it */
            if (last_val_is_mode)
                mode_freq++;    /* needn't maintain last_val_freq */
            else if (++last_val_freq > mode_freq)
            {
                /* last_val becomes new mode */
                if (shouldfree)
                    pfree(DatumGetPointer(mode_val));
                mode_val = last_val;
                mode_freq = last_val_freq;
                last_val_is_mode = true;
            }
            if (shouldfree)
                pfree(DatumGetPointer(val));
        }
        else
        {
            /* val should replace last_val */
            if (shouldfree && !last_val_is_mode)
                pfree(DatumGetPointer(last_val));
            last_val = val;
            /* avoid equality function calls by reusing abbreviated keys */
            last_abbrev_val = abbrev_val;
            last_val_freq = 1;
            last_val_is_mode = false;
        }

        CHECK_FOR_INTERRUPTS();
    }

    if (shouldfree && !last_val_is_mode)
        pfree(DatumGetPointer(last_val));

    if (mode_freq)
        PG_RETURN_DATUM(mode_val);
    else
        PG_RETURN_NULL();
}


/*
 * Common code to sanity-check args for hypothetical-set functions. No need
 * for friendly errors, these can only happen if someone's messing up the
 * aggregate definitions. The checks are needed for security, however.
 */
static void
hypothetical_check_argtypes(FunctionCallInfo fcinfo, int nargs,
                            TupleDesc tupdesc)
{
    int         i;

    /* check that we have an int4 flag column */
    if (!tupdesc ||
        (nargs + 1) != tupdesc->natts ||
        TupleDescAttr(tupdesc, nargs)->atttypid != INT4OID)
        elog(ERROR, "type mismatch in hypothetical-set function");

    /* check that direct args match in type with aggregated args */
    for (i = 0; i < nargs; i++)
    {
        Form_pg_attribute attr = TupleDescAttr(tupdesc, i);

        if (get_fn_expr_argtype(fcinfo->flinfo, i + 1) != attr->atttypid)
            elog(ERROR, "type mismatch in hypothetical-set function");
    }
}

/*
 * compute rank of hypothetical row
 *
 * flag should be -1 to sort hypothetical row ahead of its peers, or +1
 * to sort behind.
 * total number of regular rows is returned into *number_of_rows.
 */
static int64
hypothetical_rank_common(FunctionCallInfo fcinfo, int flag,
                         int64 *number_of_rows)
{
    int         nargs = PG_NARGS() - 1;
    int64       rank = 1;
    OSAPerGroupState *osastate;
    TupleTableSlot *slot;
    int         i;

    Assert(AggCheckCallContext(fcinfo, NULL) == AGG_CONTEXT_AGGREGATE);

    /* If there were no regular rows, the rank is always 1 */
    if (PG_ARGISNULL(0))
    {
        *number_of_rows = 0;
        return 1;
    }

    osastate = (OSAPerGroupState *) PG_GETARG_POINTER(0);
    *number_of_rows = osastate->number_of_rows;

    /* Adjust nargs to be the number of direct (or aggregated) args */
    if (nargs % 2 != 0)
        elog(ERROR, "wrong number of arguments in hypothetical-set function");
    nargs /= 2;

    hypothetical_check_argtypes(fcinfo, nargs, osastate->qstate->tupdesc);

    /* because we need a hypothetical row, we can't share transition state */
    Assert(!osastate->sort_done);

    /* insert the hypothetical row into the sort */
    slot = osastate->qstate->tupslot;
    ExecClearTuple(slot);
    for (i = 0; i < nargs; i++)
    {
        slot->tts_values[i] = PG_GETARG_DATUM(i + 1);
        slot->tts_isnull[i] = PG_ARGISNULL(i + 1);
    }
    slot->tts_values[i] = Int32GetDatum(flag);
    slot->tts_isnull[i] = false;
    ExecStoreVirtualTuple(slot);

    tuplesort_puttupleslot(osastate->sortstate, slot);

    /* finish the sort */
    tuplesort_performsort(osastate->sortstate);
    osastate->sort_done = true;

    /* iterate till we find the hypothetical row */
    while (tuplesort_gettupleslot(osastate->sortstate, true, true, slot, NULL))
    {
        bool        isnull;
        Datum       d = slot_getattr(slot, nargs + 1, &isnull);

        if (!isnull && DatumGetInt32(d) != 0)
            break;

        rank++;

        CHECK_FOR_INTERRUPTS();
    }

    ExecClearTuple(slot);

    return rank;
}


/*
 * rank()  - rank of hypothetical row
 */
Datum
hypothetical_rank_final(PG_FUNCTION_ARGS)
{
    int64       rank;
    int64       rowcount;

    rank = hypothetical_rank_common(fcinfo, -1, &rowcount);

    PG_RETURN_INT64(rank);
}

/*
 * percent_rank()   - percentile rank of hypothetical row
 */
Datum
hypothetical_percent_rank_final(PG_FUNCTION_ARGS)
{
    int64       rank;
    int64       rowcount;
    double      result_val;

    rank = hypothetical_rank_common(fcinfo, -1, &rowcount);

    if (rowcount == 0)
        PG_RETURN_FLOAT8(0);

    result_val = (double) (rank - 1) / (double) (rowcount);

    PG_RETURN_FLOAT8(result_val);
}

/*
 * cume_dist()  - cumulative distribution of hypothetical row
 */
Datum
hypothetical_cume_dist_final(PG_FUNCTION_ARGS)
{
    int64       rank;
    int64       rowcount;
    double      result_val;

    rank = hypothetical_rank_common(fcinfo, 1, &rowcount);

    result_val = (double) (rank) / (double) (rowcount + 1);

    PG_RETURN_FLOAT8(result_val);
}

/*
 * dense_rank() - rank of hypothetical row without gaps in ranking
 */
Datum
hypothetical_dense_rank_final(PG_FUNCTION_ARGS)
{
    ExprContext *econtext;
    ExprState  *compareTuple;
    int         nargs = PG_NARGS() - 1;
    int64       rank = 1;
    int64       duplicate_count = 0;
    OSAPerGroupState *osastate;
    int         numDistinctCols;
    Datum       abbrevVal = (Datum) 0;
    Datum       abbrevOld = (Datum) 0;
    TupleTableSlot *slot;
    TupleTableSlot *extraslot;
    TupleTableSlot *slot2;
    int         i;

    Assert(AggCheckCallContext(fcinfo, NULL) == AGG_CONTEXT_AGGREGATE);

    /* If there were no regular rows, the rank is always 1 */
    if (PG_ARGISNULL(0))
        PG_RETURN_INT64(rank);

    osastate = (OSAPerGroupState *) PG_GETARG_POINTER(0);
    econtext = osastate->qstate->econtext;
    if (!econtext)
    {
        MemoryContext oldcontext;

        /* Make sure to we create econtext under correct parent context. */
        oldcontext = MemoryContextSwitchTo(osastate->qstate->qcontext);
        osastate->qstate->econtext = CreateStandaloneExprContext();
        econtext = osastate->qstate->econtext;
        MemoryContextSwitchTo(oldcontext);
    }

    /* Adjust nargs to be the number of direct (or aggregated) args */
    if (nargs % 2 != 0)
        elog(ERROR, "wrong number of arguments in hypothetical-set function");
    nargs /= 2;

    hypothetical_check_argtypes(fcinfo, nargs, osastate->qstate->tupdesc);

    /*
     * When comparing tuples, we can omit the flag column since we will only
     * compare rows with flag == 0.
     */
    numDistinctCols = osastate->qstate->numSortCols - 1;

    /* Build tuple comparator, if we didn't already */
    compareTuple = osastate->qstate->compareTuple;
    if (compareTuple == NULL)
    {
        AttrNumber *sortColIdx = osastate->qstate->sortColIdx;
        MemoryContext oldContext;

        oldContext = MemoryContextSwitchTo(osastate->qstate->qcontext);
        compareTuple = execTuplesMatchPrepare(osastate->qstate->tupdesc,
                                              numDistinctCols,
                                              sortColIdx,
                                              osastate->qstate->eqOperators,
                                              osastate->qstate->sortCollations,
                                              NULL);
        MemoryContextSwitchTo(oldContext);
        osastate->qstate->compareTuple = compareTuple;
    }

    /* because we need a hypothetical row, we can't share transition state */
    Assert(!osastate->sort_done);

    /* insert the hypothetical row into the sort */
    slot = osastate->qstate->tupslot;
    ExecClearTuple(slot);
    for (i = 0; i < nargs; i++)
    {
        slot->tts_values[i] = PG_GETARG_DATUM(i + 1);
        slot->tts_isnull[i] = PG_ARGISNULL(i + 1);
    }
    slot->tts_values[i] = Int32GetDatum(-1);
    slot->tts_isnull[i] = false;
    ExecStoreVirtualTuple(slot);

    tuplesort_puttupleslot(osastate->sortstate, slot);

    /* finish the sort */
    tuplesort_performsort(osastate->sortstate);
    osastate->sort_done = true;

    /*
     * We alternate fetching into tupslot and extraslot so that we have the
     * previous row available for comparisons.  This is accomplished by
     * swapping the slot pointer variables after each row.
     */
    extraslot = MakeSingleTupleTableSlot(osastate->qstate->tupdesc,
                                         &TTSOpsMinimalTuple);
    slot2 = extraslot;

    /* iterate till we find the hypothetical row */
    while (tuplesort_gettupleslot(osastate->sortstate, true, true, slot,
                                  &abbrevVal))
    {
        bool        isnull;
        Datum       d = slot_getattr(slot, nargs + 1, &isnull);
        TupleTableSlot *tmpslot;

        if (!isnull && DatumGetInt32(d) != 0)
            break;

        /* count non-distinct tuples */
        econtext->ecxt_outertuple = slot;
        econtext->ecxt_innertuple = slot2;

        if (!TupIsNull(slot2) &&
            abbrevVal == abbrevOld &&
            ExecQualAndReset(compareTuple, econtext))
            duplicate_count++;

        tmpslot = slot2;
        slot2 = slot;
        slot = tmpslot;
        /* avoid ExecQual() calls by reusing abbreviated keys */
        abbrevOld = abbrevVal;

        rank++;

        CHECK_FOR_INTERRUPTS();
    }

    ExecClearTuple(slot);
    ExecClearTuple(slot2);

    ExecDropSingleTupleTableSlot(extraslot);

    rank = rank - duplicate_count;

    PG_RETURN_INT64(rank);
}