Move per-agg and per-trans duplicate finding to the planner.
authorHeikki Linnakangas <heikki.linnakangas@iki.fi>
Tue, 24 Nov 2020 08:45:00 +0000 (10:45 +0200)
committerHeikki Linnakangas <heikki.linnakangas@iki.fi>
Tue, 24 Nov 2020 08:45:00 +0000 (10:45 +0200)
This has the advantage that the cost estimates for aggregates can count
the number of calls to transition and final functions correctly.

Bump catalog version, because views can contain Aggrefs.

Reviewed-by: Andres Freund
Discussion: https://www.postgresql.org/message-id/b2e3536b-1dbc-8303-c97e-89cb0b4a9a48%40iki.fi

29 files changed:
contrib/postgres_fdw/postgres_fdw.c
src/backend/executor/execExpr.c
src/backend/executor/execExprInterp.c
src/backend/executor/nodeAgg.c
src/backend/jit/llvm/llvmjit_expr.c
src/backend/nodes/copyfuncs.c
src/backend/nodes/equalfuncs.c
src/backend/nodes/outfuncs.c
src/backend/nodes/readfuncs.c
src/backend/optimizer/path/costsize.c
src/backend/optimizer/plan/planagg.c
src/backend/optimizer/plan/planner.c
src/backend/optimizer/prep/Makefile
src/backend/optimizer/prep/prepagg.c [new file with mode: 0644]
src/backend/optimizer/util/clauses.c
src/backend/parser/parse_func.c
src/backend/utils/adt/selfuncs.c
src/include/catalog/catversion.h
src/include/executor/execExpr.h
src/include/nodes/execnodes.h
src/include/nodes/nodes.h
src/include/nodes/pathnodes.h
src/include/nodes/primnodes.h
src/include/optimizer/clauses.h
src/include/optimizer/prep.h
src/include/utils/selfuncs.h
src/test/regress/expected/partition_aggregate.out
src/test/regress/sql/partition_aggregate.sql
src/tools/pgindent/typedefs.list

index 9c5aaacc515667bc0a687c9e48a47456faad6329..b6c72e1d1e64222717b62bd5af85450ec919be8b 100644 (file)
@@ -32,6 +32,7 @@
 #include "optimizer/pathnode.h"
 #include "optimizer/paths.h"
 #include "optimizer/planmain.h"
+#include "optimizer/prep.h"
 #include "optimizer/restrictinfo.h"
 #include "optimizer/tlist.h"
 #include "parser/parsetree.h"
@@ -2944,16 +2945,7 @@ estimate_path_cost_size(PlannerInfo *root,
            MemSet(&aggcosts, 0, sizeof(AggClauseCosts));
            if (root->parse->hasAggs)
            {
-               get_agg_clause_costs(root, (Node *) fpinfo->grouped_tlist,
-                                    AGGSPLIT_SIMPLE, &aggcosts);
-
-               /*
-                * The cost of aggregates in the HAVING qual will be the same
-                * for each child as it is for the parent, so there's no need
-                * to use a translated version of havingQual.
-                */
-               get_agg_clause_costs(root, (Node *) root->parse->havingQual,
-                                    AGGSPLIT_SIMPLE, &aggcosts);
+               get_agg_clause_costs(root, AGGSPLIT_SIMPLE, &aggcosts);
            }
 
            /* Get number of grouping columns and possible number of groups */
index d76836c09b12ee9a8ba418bbb9232a0dd6e391d9..79b325c7cfb10ad669dbc5d247feb2698f74f930 100644 (file)
@@ -99,8 +99,7 @@ static void ExecBuildAggTransCall(ExprState *state, AggState *aggstate,
  * the same as the per-query context of the associated ExprContext.
  *
  * Any Aggref, WindowFunc, or SubPlan nodes found in the tree are added to
- * the lists of such nodes held by the parent PlanState (or more accurately,
- * the AggrefExprState etc. nodes created for them are added).
+ * the lists of such nodes held by the parent PlanState.
  *
  * Note: there is no ExecEndExpr function; we assume that any resource
  * cleanup needed will be handled by just releasing the memory context
@@ -779,18 +778,15 @@ ExecInitExprRec(Expr *node, ExprState *state,
        case T_Aggref:
            {
                Aggref     *aggref = (Aggref *) node;
-               AggrefExprState *astate = makeNode(AggrefExprState);
 
                scratch.opcode = EEOP_AGGREF;
-               scratch.d.aggref.astate = astate;
-               astate->aggref = aggref;
+               scratch.d.aggref.aggno = aggref->aggno;
 
                if (state->parent && IsA(state->parent, AggState))
                {
                    AggState   *aggstate = (AggState *) state->parent;
 
-                   aggstate->aggs = lappend(aggstate->aggs, astate);
-                   aggstate->numaggs++;
+                   aggstate->aggs = lappend(aggstate->aggs, aggref);
                }
                else
                {
index 26c2b4963215e2a0e46e9db8b0620877ba14e444..c09371ad58f98ba93dfc3d97493ae45bb4d9efaf 100644 (file)
@@ -1494,12 +1494,12 @@ ExecInterpExpr(ExprState *state, ExprContext *econtext, bool *isnull)
             * Returns a Datum whose value is the precomputed aggregate value
             * found in the given expression context.
             */
-           AggrefExprState *aggref = op->d.aggref.astate;
+           int         aggno = op->d.aggref.aggno;
 
            Assert(econtext->ecxt_aggvalues != NULL);
 
-           *op->resvalue = econtext->ecxt_aggvalues[aggref->aggno];
-           *op->resnull = econtext->ecxt_aggnulls[aggref->aggno];
+           *op->resvalue = econtext->ecxt_aggvalues[aggno];
+           *op->resnull = econtext->ecxt_aggnulls[aggno];
 
            EEO_NEXT();
        }
index 5bf9e99bbc86b776ac2488518d3d8626b0fb4e31..dc640feb63114e88795f4070fc70f2f1fca6da68 100644 (file)
@@ -465,14 +465,6 @@ static void build_pertrans_for_aggref(AggStatePerTrans pertrans,
                                      Oid aggserialfn, Oid aggdeserialfn,
                                      Datum initValue, bool initValueIsNull,
                                      Oid *inputTypes, int numArguments);
-static int find_compatible_peragg(Aggref *newagg, AggState *aggstate,
-                                  int lastaggno, List **same_input_transnos);
-static int find_compatible_pertrans(AggState *aggstate, Aggref *newagg,
-                                    bool shareable,
-                                    Oid aggtransfn, Oid aggtranstype,
-                                    Oid aggserialfn, Oid aggdeserialfn,
-                                    Datum initValue, bool initValueIsNull,
-                                    List *transnos);
 
 
 /*
@@ -3244,9 +3236,11 @@ ExecInitAgg(Agg *node, EState *estate, int eflags)
    Plan       *outerPlan;
    ExprContext *econtext;
    TupleDesc   scanDesc;
-   int         numaggs,
-               transno,
-               aggno;
+   int         max_aggno;
+   int         max_transno;
+   int         numaggrefs;
+   int         numaggs;
+   int         numtrans;
    int         phase;
    int         phaseidx;
    ListCell   *l;
@@ -3422,9 +3416,9 @@ ExecInitAgg(Agg *node, EState *estate, int eflags)
     * semantics, and it's forbidden by the spec.  Because it is true, we
     * don't need to worry about evaluating the aggs in any particular order.
     *
-    * Note: execExpr.c finds Aggrefs for us, and adds their AggrefExprState
-    * nodes to aggstate->aggs.  Aggrefs in the qual are found here; Aggrefs
-    * in the targetlist are found during ExecAssignProjectionInfo, below.
+    * Note: execExpr.c finds Aggrefs for us, and adds them to aggstate->aggs.
+    * Aggrefs in the qual are found here; Aggrefs in the targetlist are found
+    * during ExecAssignProjectionInfo, above.
     */
    aggstate->ss.ps.qual =
        ExecInitQual(node->plan.qual, (PlanState *) aggstate);
@@ -3432,8 +3426,18 @@ ExecInitAgg(Agg *node, EState *estate, int eflags)
    /*
     * We should now have found all Aggrefs in the targetlist and quals.
     */
-   numaggs = aggstate->numaggs;
-   Assert(numaggs == list_length(aggstate->aggs));
+   numaggrefs = list_length(aggstate->aggs);
+   max_aggno = -1;
+   max_transno = -1;
+   foreach(l, aggstate->aggs)
+   {
+       Aggref     *aggref = (Aggref *) lfirst(l);
+
+       max_aggno = Max(max_aggno, aggref->aggno);
+       max_transno = Max(max_transno, aggref->aggtransno);
+   }
+   numaggs = max_aggno + 1;
+   numtrans = max_transno + 1;
 
    /*
     * For each phase, prepare grouping set data and fmgr lookup data for
@@ -3604,7 +3608,7 @@ ExecInitAgg(Agg *node, EState *estate, int eflags)
    econtext->ecxt_aggnulls = (bool *) palloc0(sizeof(bool) * numaggs);
 
    peraggs = (AggStatePerAgg) palloc0(sizeof(AggStatePerAggData) * numaggs);
-   pertransstates = (AggStatePerTrans) palloc0(sizeof(AggStatePerTransData) * numaggs);
+   pertransstates = (AggStatePerTrans) palloc0(sizeof(AggStatePerTransData) * numtrans);
 
    aggstate->peragg = peraggs;
    aggstate->pertrans = pertransstates;
@@ -3695,92 +3699,41 @@ ExecInitAgg(Agg *node, EState *estate, int eflags)
        select_current_set(aggstate, 0, false);
    }
 
-   /* -----------------
+   /*
     * Perform lookups of aggregate function info, and initialize the
     * unchanging fields of the per-agg and per-trans data.
-    *
-    * We try to optimize by detecting duplicate aggregate functions so that
-    * their state and final values are re-used, rather than needlessly being
-    * re-calculated independently. We also detect aggregates that are not
-    * the same, but which can share the same transition state.
-    *
-    * Scenarios:
-    *
-    * 1. Identical aggregate function calls appear in the query:
-    *
-    *    SELECT SUM(x) FROM ... HAVING SUM(x) > 0
-    *
-    *    Since these aggregates are identical, we only need to calculate
-    *    the value once. Both aggregates will share the same 'aggno' value.
-    *
-    * 2. Two different aggregate functions appear in the query, but the
-    *    aggregates have the same arguments, transition functions and
-    *    initial values (and, presumably, different final functions):
-    *
-    *    SELECT AVG(x), STDDEV(x) FROM ...
-    *
-    *    In this case we must create a new peragg for the varying aggregate,
-    *    and we need to call the final functions separately, but we need
-    *    only run the transition function once.  (This requires that the
-    *    final functions be nondestructive of the transition state, but
-    *    that's required anyway for other reasons.)
-    *
-    * For either of these optimizations to be valid, all aggregate properties
-    * used in the transition phase must be the same, including any modifiers
-    * such as ORDER BY, DISTINCT and FILTER, and the arguments mustn't
-    * contain any volatile functions.
-    * -----------------
     */
-   aggno = -1;
-   transno = -1;
    foreach(l, aggstate->aggs)
    {
-       AggrefExprState *aggrefstate = (AggrefExprState *) lfirst(l);
-       Aggref     *aggref = aggrefstate->aggref;
+       Aggref     *aggref = lfirst(l);
        AggStatePerAgg peragg;
        AggStatePerTrans pertrans;
-       int         existing_aggno;
-       int         existing_transno;
-       List       *same_input_transnos;
        Oid         inputTypes[FUNC_MAX_ARGS];
        int         numArguments;
        int         numDirectArgs;
        HeapTuple   aggTuple;
        Form_pg_aggregate aggform;
        AclResult   aclresult;
-       Oid         transfn_oid,
-                   finalfn_oid;
-       bool        shareable;
+       Oid         finalfn_oid;
        Oid         serialfn_oid,
                    deserialfn_oid;
+       Oid         aggOwner;
        Expr       *finalfnexpr;
        Oid         aggtranstype;
-       Datum       textInitVal;
-       Datum       initValue;
-       bool        initValueIsNull;
 
        /* Planner should have assigned aggregate to correct level */
        Assert(aggref->agglevelsup == 0);
        /* ... and the split mode should match */
        Assert(aggref->aggsplit == aggstate->aggsplit);
 
-       /* 1. Check for already processed aggs which can be re-used */
-       existing_aggno = find_compatible_peragg(aggref, aggstate, aggno,
-                                               &same_input_transnos);
-       if (existing_aggno != -1)
-       {
-           /*
-            * Existing compatible agg found. so just point the Aggref to the
-            * same per-agg struct.
-            */
-           aggrefstate->aggno = existing_aggno;
+       peragg = &peraggs[aggref->aggno];
+
+       /* Check if we initialized the state for this aggregate already. */
+       if (peragg->aggref != NULL)
            continue;
-       }
 
-       /* Mark Aggref state node with assigned index in the result array */
-       peragg = &peraggs[++aggno];
        peragg->aggref = aggref;
-       aggrefstate->aggno = aggno;
+       peragg->transno = aggref->aggtransno;
 
        /* Fetch the pg_aggregate row */
        aggTuple = SearchSysCache1(AGGFNOID,
@@ -3802,36 +3755,12 @@ ExecInitAgg(Agg *node, EState *estate, int eflags)
        aggtranstype = aggref->aggtranstype;
        Assert(OidIsValid(aggtranstype));
 
-       /*
-        * If this aggregation is performing state combines, then instead of
-        * using the transition function, we'll use the combine function
-        */
-       if (DO_AGGSPLIT_COMBINE(aggstate->aggsplit))
-       {
-           transfn_oid = aggform->aggcombinefn;
-
-           /* If not set then the planner messed up */
-           if (!OidIsValid(transfn_oid))
-               elog(ERROR, "combinefn not set for aggregate function");
-       }
-       else
-           transfn_oid = aggform->aggtransfn;
-
        /* Final function only required if we're finalizing the aggregates */
        if (DO_AGGSPLIT_SKIPFINAL(aggstate->aggsplit))
            peragg->finalfn_oid = finalfn_oid = InvalidOid;
        else
            peragg->finalfn_oid = finalfn_oid = aggform->aggfinalfn;
 
-       /*
-        * If finalfn is marked read-write, we can't share transition states;
-        * but it is okay to share states for AGGMODIFY_SHAREABLE aggs.  Also,
-        * if we're not executing the finalfn here, we can share regardless.
-        */
-       shareable = (aggform->aggfinalmodify != AGGMODIFY_READ_WRITE) ||
-           (finalfn_oid == InvalidOid);
-       peragg->shareable = shareable;
-
        serialfn_oid = InvalidOid;
        deserialfn_oid = InvalidOid;
 
@@ -3871,7 +3800,6 @@ ExecInitAgg(Agg *node, EState *estate, int eflags)
        /* Check that aggregate owner has permission to call component fns */
        {
            HeapTuple   procTuple;
-           Oid         aggOwner;
 
            procTuple = SearchSysCache1(PROCOID,
                                        ObjectIdGetDatum(aggref->aggfnoid));
@@ -3881,12 +3809,6 @@ ExecInitAgg(Agg *node, EState *estate, int eflags)
            aggOwner = ((Form_pg_proc) GETSTRUCT(procTuple))->proowner;
            ReleaseSysCache(procTuple);
 
-           aclresult = pg_proc_aclcheck(transfn_oid, aggOwner,
-                                        ACL_EXECUTE);
-           if (aclresult != ACLCHECK_OK)
-               aclcheck_error(aclresult, OBJECT_FUNCTION,
-                              get_func_name(transfn_oid));
-           InvokeFunctionExecuteHook(transfn_oid);
            if (OidIsValid(finalfn_oid))
            {
                aclresult = pg_proc_aclcheck(finalfn_oid, aggOwner,
@@ -3959,51 +3881,60 @@ ExecInitAgg(Agg *node, EState *estate, int eflags)
                        &peragg->resulttypeByVal);
 
        /*
-        * initval is potentially null, so don't try to access it as a struct
-        * field. Must do it the hard way with SysCacheGetAttr.
+        * Build working state for invoking the transition function, if we
+        * haven't done it already.
         */
-       textInitVal = SysCacheGetAttr(AGGFNOID, aggTuple,
-                                     Anum_pg_aggregate_agginitval,
-                                     &initValueIsNull);
-       if (initValueIsNull)
-           initValue = (Datum) 0;
-       else
-           initValue = GetAggInitVal(textInitVal, aggtranstype);
-
-       /*
-        * 2. Build working state for invoking the transition function, or
-        * look up previously initialized working state, if we can share it.
-        *
-        * find_compatible_peragg() already collected a list of shareable
-        * per-Trans's with the same inputs. Check if any of them have the
-        * same transition function and initial value.
-        */
-       existing_transno = find_compatible_pertrans(aggstate, aggref,
-                                                   shareable,
-                                                   transfn_oid, aggtranstype,
-                                                   serialfn_oid, deserialfn_oid,
-                                                   initValue, initValueIsNull,
-                                                   same_input_transnos);
-       if (existing_transno != -1)
+       pertrans = &pertransstates[aggref->aggtransno];
+       if (pertrans->aggref == NULL)
        {
+           Datum       textInitVal;
+           Datum       initValue;
+           bool        initValueIsNull;
+           Oid         transfn_oid;
+
            /*
-            * Existing compatible trans found, so just point the 'peragg' to
-            * the same per-trans struct, and mark the trans state as shared.
+            * If this aggregation is performing state combines, then instead
+            * of using the transition function, we'll use the combine
+            * function
             */
-           pertrans = &pertransstates[existing_transno];
-           pertrans->aggshared = true;
-           peragg->transno = existing_transno;
-       }
-       else
-       {
-           pertrans = &pertransstates[++transno];
+           if (DO_AGGSPLIT_COMBINE(aggstate->aggsplit))
+           {
+               transfn_oid = aggform->aggcombinefn;
+
+               /* If not set then the planner messed up */
+               if (!OidIsValid(transfn_oid))
+                   elog(ERROR, "combinefn not set for aggregate function");
+           }
+           else
+               transfn_oid = aggform->aggtransfn;
+
+           aclresult = pg_proc_aclcheck(transfn_oid, aggOwner,
+                                        ACL_EXECUTE);
+           if (aclresult != ACLCHECK_OK)
+               aclcheck_error(aclresult, OBJECT_FUNCTION,
+                              get_func_name(transfn_oid));
+           InvokeFunctionExecuteHook(transfn_oid);
+
+           /*
+            * initval is potentially null, so don't try to access it as a
+            * struct field. Must do it the hard way with SysCacheGetAttr.
+            */
+           textInitVal = SysCacheGetAttr(AGGFNOID, aggTuple,
+                                         Anum_pg_aggregate_agginitval,
+                                         &initValueIsNull);
+           if (initValueIsNull)
+               initValue = (Datum) 0;
+           else
+               initValue = GetAggInitVal(textInitVal, aggtranstype);
+
            build_pertrans_for_aggref(pertrans, aggstate, estate,
                                      aggref, transfn_oid, aggtranstype,
                                      serialfn_oid, deserialfn_oid,
                                      initValue, initValueIsNull,
                                      inputTypes, numArguments);
-           peragg->transno = transno;
        }
+       else
+           pertrans->aggshared = true;
        ReleaseSysCache(aggTuple);
    }
 
@@ -4011,8 +3942,8 @@ ExecInitAgg(Agg *node, EState *estate, int eflags)
     * Update aggstate->numaggs to be the number of unique aggregates found.
     * Also set numstates to the number of unique transition states found.
     */
-   aggstate->numaggs = aggno + 1;
-   aggstate->numtrans = transno + 1;
+   aggstate->numaggs = numaggs;
+   aggstate->numtrans = numtrans;
 
    /*
     * Last, check whether any more aggregates got added onto the node while
@@ -4024,7 +3955,7 @@ ExecInitAgg(Agg *node, EState *estate, int eflags)
     * need to work hard on a helpful error message; but we defend against it
     * here anyway, just to be sure.)
     */
-   if (numaggs != list_length(aggstate->aggs))
+   if (numaggrefs != list_length(aggstate->aggs))
        ereport(ERROR,
                (errcode(ERRCODE_GROUPING_ERROR),
                 errmsg("aggregate function calls cannot be nested")));
@@ -4420,147 +4351,6 @@ GetAggInitVal(Datum textInitVal, Oid transtype)
    return initVal;
 }
 
-/*
- * find_compatible_peragg - search for a previously initialized per-Agg struct
- *
- * Searches the previously looked at aggregates to find one which is compatible
- * with this one, with the same input parameters. If no compatible aggregate
- * can be found, returns -1.
- *
- * As a side-effect, this also collects a list of existing, shareable per-Trans
- * structs with matching inputs. If no identical Aggref is found, the list is
- * passed later to find_compatible_pertrans, to see if we can at least reuse
- * the state value of another aggregate.
- */
-static int
-find_compatible_peragg(Aggref *newagg, AggState *aggstate,
-                      int lastaggno, List **same_input_transnos)
-{
-   int         aggno;
-   AggStatePerAgg peraggs;
-
-   *same_input_transnos = NIL;
-
-   /* we mustn't reuse the aggref if it contains volatile function calls */
-   if (contain_volatile_functions((Node *) newagg))
-       return -1;
-
-   peraggs = aggstate->peragg;
-
-   /*
-    * Search through the list of already seen aggregates. If we find an
-    * existing identical aggregate call, then we can re-use that one. While
-    * searching, we'll also collect a list of Aggrefs with the same input
-    * parameters. If no matching Aggref is found, the caller can potentially
-    * still re-use the transition state of one of them.  (At this stage we
-    * just compare the parsetrees; whether different aggregates share the
-    * same transition function will be checked later.)
-    */
-   for (aggno = 0; aggno <= lastaggno; aggno++)
-   {
-       AggStatePerAgg peragg;
-       Aggref     *existingRef;
-
-       peragg = &peraggs[aggno];
-       existingRef = peragg->aggref;
-
-       /* all of the following must be the same or it's no match */
-       if (newagg->inputcollid != existingRef->inputcollid ||
-           newagg->aggtranstype != existingRef->aggtranstype ||
-           newagg->aggstar != existingRef->aggstar ||
-           newagg->aggvariadic != existingRef->aggvariadic ||
-           newagg->aggkind != existingRef->aggkind ||
-           !equal(newagg->args, existingRef->args) ||
-           !equal(newagg->aggorder, existingRef->aggorder) ||
-           !equal(newagg->aggdistinct, existingRef->aggdistinct) ||
-           !equal(newagg->aggfilter, existingRef->aggfilter))
-           continue;
-
-       /* if it's the same aggregate function then report exact match */
-       if (newagg->aggfnoid == existingRef->aggfnoid &&
-           newagg->aggtype == existingRef->aggtype &&
-           newagg->aggcollid == existingRef->aggcollid &&
-           equal(newagg->aggdirectargs, existingRef->aggdirectargs))
-       {
-           list_free(*same_input_transnos);
-           *same_input_transnos = NIL;
-           return aggno;
-       }
-
-       /*
-        * Not identical, but it had the same inputs.  If the final function
-        * permits sharing, return its transno to the caller, in case we can
-        * re-use its per-trans state.  (If there's already sharing going on,
-        * we might report a transno more than once.  find_compatible_pertrans
-        * is cheap enough that it's not worth spending cycles to avoid that.)
-        */
-       if (peragg->shareable)
-           *same_input_transnos = lappend_int(*same_input_transnos,
-                                              peragg->transno);
-   }
-
-   return -1;
-}
-
-/*
- * find_compatible_pertrans - search for a previously initialized per-Trans
- * struct
- *
- * Searches the list of transnos for a per-Trans struct with the same
- * transition function and initial condition. (The inputs have already been
- * verified to match.)
- */
-static int
-find_compatible_pertrans(AggState *aggstate, Aggref *newagg, bool shareable,
-                        Oid aggtransfn, Oid aggtranstype,
-                        Oid aggserialfn, Oid aggdeserialfn,
-                        Datum initValue, bool initValueIsNull,
-                        List *transnos)
-{
-   ListCell   *lc;
-
-   /* If this aggregate can't share transition states, give up */
-   if (!shareable)
-       return -1;
-
-   foreach(lc, transnos)
-   {
-       int         transno = lfirst_int(lc);
-       AggStatePerTrans pertrans = &aggstate->pertrans[transno];
-
-       /*
-        * if the transfns or transition state types are not the same then the
-        * state can't be shared.
-        */
-       if (aggtransfn != pertrans->transfn_oid ||
-           aggtranstype != pertrans->aggtranstype)
-           continue;
-
-       /*
-        * The serialization and deserialization functions must match, if
-        * present, as we're unable to share the trans state for aggregates
-        * which will serialize or deserialize into different formats.
-        * Remember that these will be InvalidOid if they're not required for
-        * this agg node.
-        */
-       if (aggserialfn != pertrans->serialfn_oid ||
-           aggdeserialfn != pertrans->deserialfn_oid)
-           continue;
-
-       /*
-        * Check that the initial condition matches, too.
-        */
-       if (initValueIsNull && pertrans->initValueIsNull)
-           return transno;
-
-       if (!initValueIsNull && !pertrans->initValueIsNull &&
-           datumIsEqual(initValue, pertrans->initValue,
-                        pertrans->transtypeByVal, pertrans->transtypeLen))
-           return transno;
-   }
-   return -1;
-}
-
 void
 ExecEndAgg(AggState *node)
 {
index eb1dea658cb258892739bb8e5c973d7833283d50..f232397cabf86ce142b89dc5a634a6d9d879b4b7 100644 (file)
@@ -1849,20 +1849,11 @@ llvm_compile_expr(ExprState *state)
 
            case EEOP_AGGREF:
                {
-                   AggrefExprState *aggref = op->d.aggref.astate;
-                   LLVMValueRef v_aggnop;
                    LLVMValueRef v_aggno;
                    LLVMValueRef value,
                                isnull;
 
-                   /*
-                    * At this point aggref->aggno is not yet set (it's set up
-                    * in ExecInitAgg() after initializing the expression). So
-                    * load it from memory each time round.
-                    */
-                   v_aggnop = l_ptr_const(&aggref->aggno,
-                                          l_ptr(LLVMInt32Type()));
-                   v_aggno = LLVMBuildLoad(b, v_aggnop, "v_aggno");
+                   v_aggno = l_int32_const(op->d.aggref.aggno);
 
                    /* load agg value / null */
                    value = l_load_gep1(b, v_aggvalues, v_aggno, "aggvalue");
index 5a591d0a7513e3729411ea7a7042ba52442c930b..47b9ffd401793cec776fb3ec8df323cedab49f55 100644 (file)
@@ -1492,6 +1492,8 @@ _copyAggref(const Aggref *from)
    COPY_SCALAR_FIELD(aggkind);
    COPY_SCALAR_FIELD(agglevelsup);
    COPY_SCALAR_FIELD(aggsplit);
+   COPY_SCALAR_FIELD(aggno);
+   COPY_SCALAR_FIELD(aggtransno);
    COPY_LOCATION_FIELD(location);
 
    return newnode;
index e2895a8985d978253835abb86a5ad48b194543af..6cccaea124ef46a1edd932c464620aa00db90f44 100644 (file)
@@ -232,6 +232,8 @@ _equalAggref(const Aggref *a, const Aggref *b)
    COMPARE_SCALAR_FIELD(aggkind);
    COMPARE_SCALAR_FIELD(agglevelsup);
    COMPARE_SCALAR_FIELD(aggsplit);
+   COMPARE_SCALAR_FIELD(aggno);
+   COMPARE_SCALAR_FIELD(aggtransno);
    COMPARE_LOCATION_FIELD(location);
 
    return true;
index f26498cea2d15c6fdac33c74a4bf23a7bd185753..bd5694d88e03fdab6666dee487884ca5e11ed344 100644 (file)
@@ -1153,6 +1153,8 @@ _outAggref(StringInfo str, const Aggref *node)
    WRITE_CHAR_FIELD(aggkind);
    WRITE_UINT_FIELD(agglevelsup);
    WRITE_ENUM_FIELD(aggsplit, AggSplit);
+   WRITE_INT_FIELD(aggno);
+   WRITE_INT_FIELD(aggtransno);
    WRITE_LOCATION_FIELD(location);
 }
 
index ab7b535caaeaa2f9abd61f27467120c70bbb3230..169d5581b91f73f2dc381c92f21c3443252a4a88 100644 (file)
@@ -615,6 +615,8 @@ _readAggref(void)
    READ_CHAR_FIELD(aggkind);
    READ_UINT_FIELD(agglevelsup);
    READ_ENUM_FIELD(aggsplit, AggSplit);
+   READ_INT_FIELD(aggno);
+   READ_INT_FIELD(aggtransno);
    READ_LOCATION_FIELD(location);
 
    READ_DONE();
index f1dfdc1a4a1d66a3c8e02d6459d88eb4acbd9806..22d6935824a48d842d4022b9b13b9f3598f9de59 100644 (file)
@@ -2439,7 +2439,8 @@ cost_agg(Path *path, PlannerInfo *root,
         * than or equal to one, all groups are expected to fit in memory;
         * otherwise we expect to spill.
         */
-       hashentrysize = hash_agg_entry_size(aggcosts->numAggs, input_width,
+       hashentrysize = hash_agg_entry_size(list_length(root->aggtransinfos),
+                                           input_width,
                                            aggcosts->transitionSpace);
        hash_agg_set_limits(hashentrysize, numGroups, 0, &mem_limit,
                            &ngroups_limit, &num_partitions);
index 8634940efc11a350724cc9e6f50c34069c1c19c7..48c4fee8923b8330b051e1fa088c1df17c0cb306 100644 (file)
@@ -47,7 +47,7 @@
 #include "utils/lsyscache.h"
 #include "utils/syscache.h"
 
-static bool find_minmax_aggs_walker(Node *node, List **context);
+static bool can_minmax_aggs(PlannerInfo *root, List **context);
 static bool build_minmax_path(PlannerInfo *root, MinMaxAggInfo *mminfo,
                              Oid eqop, Oid sortop, bool nulls_first);
 static void minmax_qp_callback(PlannerInfo *root, void *extra);
@@ -66,7 +66,8 @@ static Oid    fetch_agg_sort_op(Oid aggfnoid);
  * query_planner(), because we generate indexscan paths by cloning the
  * planner's state and invoking query_planner() on a modified version of
  * the query parsetree.  Thus, all preprocessing needed before query_planner()
- * must already be done.
+ * must already be done.  This relies on the list of aggregates in
+ * root->agginfos, so preprocess_aggrefs() must have been called already, too.
  */
 void
 preprocess_minmax_aggregates(PlannerInfo *root)
@@ -140,9 +141,7 @@ preprocess_minmax_aggregates(PlannerInfo *root)
     * all are MIN/MAX aggregates.  Stop as soon as we find one that isn't.
     */
    aggs_list = NIL;
-   if (find_minmax_aggs_walker((Node *) root->processed_tlist, &aggs_list))
-       return;
-   if (find_minmax_aggs_walker(parse->havingQual, &aggs_list))
+   if (!can_minmax_aggs(root, &aggs_list))
        return;
 
    /*
@@ -227,38 +226,33 @@ preprocess_minmax_aggregates(PlannerInfo *root)
 }
 
 /*
- * find_minmax_aggs_walker
- *     Recursively scan the Aggref nodes in an expression tree, and check
- *     that each one is a MIN/MAX aggregate.  If so, build a list of the
+ * can_minmax_aggs
+ *     Walk through all the aggregates in the query, and check
+ *     if they are all MIN/MAX aggregates.  If so, build a list of the
  *     distinct aggregate calls in the tree.
  *
- * Returns true if a non-MIN/MAX aggregate is found, false otherwise.
- * (This seemingly-backward definition is used because expression_tree_walker
- * aborts the scan on true return, which is what we want.)
- *
- * Found aggregates are added to the list at *context; it's up to the caller
- * to initialize the list to NIL.
+ * Returns false if a non-MIN/MAX aggregate is found, true otherwise.
  *
  * This does not descend into subqueries, and so should be used only after
  * reduction of sublinks to subplans.  There mustn't be outer-aggregate
  * references either.
  */
 static bool
-find_minmax_aggs_walker(Node *node, List **context)
+can_minmax_aggs(PlannerInfo *root, List **context)
 {
-   if (node == NULL)
-       return false;
-   if (IsA(node, Aggref))
+   ListCell   *lc;
+
+   foreach(lc, root->agginfos)
    {
-       Aggref     *aggref = (Aggref *) node;
+       AggInfo    *agginfo = (AggInfo *) lfirst(lc);
+       Aggref     *aggref = agginfo->representative_aggref;
        Oid         aggsortop;
        TargetEntry *curTarget;
        MinMaxAggInfo *mminfo;
-       ListCell   *l;
 
        Assert(aggref->agglevelsup == 0);
        if (list_length(aggref->args) != 1)
-           return true;        /* it couldn't be MIN/MAX */
+           return false;       /* it couldn't be MIN/MAX */
 
        /*
         * ORDER BY is usually irrelevant for MIN/MAX, but it can change the
@@ -274,7 +268,7 @@ find_minmax_aggs_walker(Node *node, List **context)
         * quickly.
         */
        if (aggref->aggorder != NIL)
-           return true;
+           return false;
        /* note: we do not care if DISTINCT is mentioned ... */
 
        /*
@@ -283,30 +277,19 @@ find_minmax_aggs_walker(Node *node, List **context)
         * now, just punt.
         */
        if (aggref->aggfilter != NULL)
-           return true;
+           return false;
 
        aggsortop = fetch_agg_sort_op(aggref->aggfnoid);
        if (!OidIsValid(aggsortop))
-           return true;        /* not a MIN/MAX aggregate */
+           return false;       /* not a MIN/MAX aggregate */
 
        curTarget = (TargetEntry *) linitial(aggref->args);
 
        if (contain_mutable_functions((Node *) curTarget->expr))
-           return true;        /* not potentially indexable */
+           return false;       /* not potentially indexable */
 
        if (type_is_rowtype(exprType((Node *) curTarget->expr)))
-           return true;        /* IS NOT NULL would have weird semantics */
-
-       /*
-        * Check whether it's already in the list, and add it if not.
-        */
-       foreach(l, *context)
-       {
-           mminfo = (MinMaxAggInfo *) lfirst(l);
-           if (mminfo->aggfnoid == aggref->aggfnoid &&
-               equal(mminfo->target, curTarget->expr))
-               return false;
-       }
+           return false;       /* IS NOT NULL would have weird semantics */
 
        mminfo = makeNode(MinMaxAggInfo);
        mminfo->aggfnoid = aggref->aggfnoid;
@@ -318,16 +301,8 @@ find_minmax_aggs_walker(Node *node, List **context)
        mminfo->param = NULL;
 
        *context = lappend(*context, mminfo);
-
-       /*
-        * We need not recurse into the argument, since it can't contain any
-        * aggregates.
-        */
-       return false;
    }
-   Assert(!IsA(node, SubLink));
-   return expression_tree_walker(node, find_minmax_aggs_walker,
-                                 (void *) context);
+   return true;
 }
 
 /*
@@ -368,6 +343,8 @@ build_minmax_path(PlannerInfo *root, MinMaxAggInfo *mminfo,
    subroot->plan_params = NIL;
    subroot->outer_params = NULL;
    subroot->init_plans = NIL;
+   subroot->agginfos = NIL;
+   subroot->aggtransinfos = NIL;
 
    subroot->parse = parse = copyObject(root->parse);
    IncrementVarSublevelsUp((Node *) parse, 1, 1);
index 986d7a52e32ca49841f0574b6565afab97c5ed0b..247f7d462525325998a73e6681bbb912388e9907 100644 (file)
@@ -152,7 +152,6 @@ static RelOptInfo *create_grouping_paths(PlannerInfo *root,
                                         RelOptInfo *input_rel,
                                         PathTarget *target,
                                         bool target_parallel_safe,
-                                        const AggClauseCosts *agg_costs,
                                         grouping_sets_data *gd);
 static bool is_degenerate_grouping(PlannerInfo *root);
 static void create_degenerate_grouping_paths(PlannerInfo *root,
@@ -228,8 +227,7 @@ static RelOptInfo *create_partial_grouping_paths(PlannerInfo *root,
                                                 GroupPathExtraData *extra,
                                                 bool force_rel_creation);
 static void gather_grouping_paths(PlannerInfo *root, RelOptInfo *rel);
-static bool can_partial_agg(PlannerInfo *root,
-                           const AggClauseCosts *agg_costs);
+static bool can_partial_agg(PlannerInfo *root);
 static void apply_scanjoin_target_to_paths(PlannerInfo *root,
                                           RelOptInfo *rel,
                                           List *scanjoin_targets,
@@ -1944,7 +1942,6 @@ grouping_planner(PlannerInfo *root, bool inheritance_update,
        bool        scanjoin_target_parallel_safe;
        bool        scanjoin_target_same_exprs;
        bool        have_grouping;
-       AggClauseCosts agg_costs;
        WindowFuncLists *wflists = NULL;
        List       *activeWindows = NIL;
        grouping_sets_data *gset_data = NULL;
@@ -1975,25 +1972,16 @@ grouping_planner(PlannerInfo *root, bool inheritance_update,
        root->processed_tlist = preprocess_targetlist(root);
 
        /*
-        * Collect statistics about aggregates for estimating costs, and mark
-        * all the aggregates with resolved aggtranstypes.  We must do this
-        * before slicing and dicing the tlist into various pathtargets, else
-        * some copies of the Aggref nodes might escape being marked with the
-        * correct transtypes.
-        *
-        * Note: currently, we do not detect duplicate aggregates here.  This
-        * may result in somewhat-overestimated cost, which is fine for our
-        * purposes since all Paths will get charged the same.  But at some
-        * point we might wish to do that detection in the planner, rather
-        * than during executor startup.
+        * Mark all the aggregates with resolved aggtranstypes, and detect
+        * aggregates that are duplicates or can share transition state.  We
+        * must do this before slicing and dicing the tlist into various
+        * pathtargets, else some copies of the Aggref nodes might escape
+        * being marked.
         */
-       MemSet(&agg_costs, 0, sizeof(AggClauseCosts));
        if (parse->hasAggs)
        {
-           get_agg_clause_costs(root, (Node *) root->processed_tlist,
-                                AGGSPLIT_SIMPLE, &agg_costs);
-           get_agg_clause_costs(root, parse->havingQual, AGGSPLIT_SIMPLE,
-                                &agg_costs);
+           preprocess_aggrefs(root, (Node *) root->processed_tlist);
+           preprocess_aggrefs(root, (Node *) parse->havingQual);
        }
 
        /*
@@ -2198,7 +2186,6 @@ grouping_planner(PlannerInfo *root, bool inheritance_update,
                                                current_rel,
                                                grouping_target,
                                                grouping_target_parallel_safe,
-                                               &agg_costs,
                                                gset_data);
            /* Fix things up if grouping_target contains SRFs */
            if (parse->hasTargetSRFs)
@@ -3790,7 +3777,6 @@ get_number_of_groups(PlannerInfo *root,
  *
  * input_rel: contains the source-data Paths
  * target: the pathtarget for the result Paths to compute
- * agg_costs: cost info about all aggregates in query (in AGGSPLIT_SIMPLE mode)
  * gd: grouping sets data including list of grouping sets and their clauses
  *
  * Note: all Paths in input_rel are expected to return the target computed
@@ -3801,12 +3787,15 @@ create_grouping_paths(PlannerInfo *root,
                      RelOptInfo *input_rel,
                      PathTarget *target,
                      bool target_parallel_safe,
-                     const AggClauseCosts *agg_costs,
                      grouping_sets_data *gd)
 {
    Query      *parse = root->parse;
    RelOptInfo *grouped_rel;
    RelOptInfo *partially_grouped_rel;
+   AggClauseCosts agg_costs;
+
+   MemSet(&agg_costs, 0, sizeof(AggClauseCosts));
+   get_agg_clause_costs(root, AGGSPLIT_SIMPLE, &agg_costs);
 
    /*
     * Create grouping relation to hold fully aggregated grouping and/or
@@ -3862,14 +3851,14 @@ create_grouping_paths(PlannerInfo *root,
         * the other gating conditions, so we want to do it last.
         */
        if ((parse->groupClause != NIL &&
-            agg_costs->numOrderedAggs == 0 &&
+            root->numOrderedAggs == 0 &&
             (gd ? gd->any_hashable : grouping_is_hashable(parse->groupClause))))
            flags |= GROUPING_CAN_USE_HASH;
 
        /*
         * Determine whether partial aggregation is possible.
         */
-       if (can_partial_agg(root, agg_costs))
+       if (can_partial_agg(root))
            flags |= GROUPING_CAN_PARTIAL_AGG;
 
        extra.flags = flags;
@@ -3890,7 +3879,7 @@ create_grouping_paths(PlannerInfo *root,
            extra.patype = PARTITIONWISE_AGGREGATE_NONE;
 
        create_ordinary_grouping_paths(root, input_rel, grouped_rel,
-                                      agg_costs, gd, &extra,
+                                      &agg_costs, gd, &extra,
                                       &partially_grouped_rel);
    }
 
@@ -4248,7 +4237,8 @@ consider_groupingsets_paths(PlannerInfo *root,
            l_start = lnext(gd->rollups, l_start);
        }
 
-       hashsize = estimate_hashagg_tablesize(path,
+       hashsize = estimate_hashagg_tablesize(root,
+                                             path,
                                              agg_costs,
                                              dNumGroups - exclude_groups);
 
@@ -4382,7 +4372,8 @@ consider_groupingsets_paths(PlannerInfo *root,
        /*
         * Account first for space needed for groups we can't sort at all.
         */
-       availspace -= estimate_hashagg_tablesize(path,
+       availspace -= estimate_hashagg_tablesize(root,
+                                                path,
                                                 agg_costs,
                                                 gd->dNumHashGroups);
 
@@ -4433,7 +4424,8 @@ consider_groupingsets_paths(PlannerInfo *root,
 
                if (rollup->hashable)
                {
-                   double      sz = estimate_hashagg_tablesize(path,
+                   double      sz = estimate_hashagg_tablesize(root,
+                                                               path,
                                                                agg_costs,
                                                                rollup->numGroups);
 
@@ -6926,20 +6918,12 @@ create_partial_grouping_paths(PlannerInfo *root,
        MemSet(agg_final_costs, 0, sizeof(AggClauseCosts));
        if (parse->hasAggs)
        {
-           List       *partial_target_exprs;
-
            /* partial phase */
-           partial_target_exprs = partially_grouped_rel->reltarget->exprs;
-           get_agg_clause_costs(root, (Node *) partial_target_exprs,
-                                AGGSPLIT_INITIAL_SERIAL,
+           get_agg_clause_costs(root, AGGSPLIT_INITIAL_SERIAL,
                                 agg_partial_costs);
 
            /* final phase */
-           get_agg_clause_costs(root, (Node *) grouped_rel->reltarget->exprs,
-                                AGGSPLIT_FINAL_DESERIAL,
-                                agg_final_costs);
-           get_agg_clause_costs(root, extra->havingQual,
-                                AGGSPLIT_FINAL_DESERIAL,
+           get_agg_clause_costs(root, AGGSPLIT_FINAL_DESERIAL,
                                 agg_final_costs);
        }
 
@@ -7324,7 +7308,7 @@ gather_grouping_paths(PlannerInfo *root, RelOptInfo *rel)
  * Returns true when possible, false otherwise.
  */
 static bool
-can_partial_agg(PlannerInfo *root, const AggClauseCosts *agg_costs)
+can_partial_agg(PlannerInfo *root)
 {
    Query      *parse = root->parse;
 
@@ -7341,7 +7325,7 @@ can_partial_agg(PlannerInfo *root, const AggClauseCosts *agg_costs)
        /* We don't know how to do grouping sets in parallel. */
        return false;
    }
-   else if (agg_costs->hasNonPartial || agg_costs->hasNonSerial)
+   else if (root->hasNonPartialAggs || root->hasNonSerialAggs)
    {
        /* Insufficient support for partial mode. */
        return false;
index 5733df45737addef79aaf28b65d6d6d8537d3077..6f8c6c8208b3c12e175ad3d8138fdab2901948f1 100644 (file)
@@ -13,6 +13,7 @@ top_builddir = ../../../..
 include $(top_builddir)/src/Makefile.global
 
 OBJS = \
+   prepagg.o \
    prepjointree.o \
    prepqual.o \
    preptlist.o \
diff --git a/src/backend/optimizer/prep/prepagg.c b/src/backend/optimizer/prep/prepagg.c
new file mode 100644 (file)
index 0000000..34ac985
--- /dev/null
@@ -0,0 +1,678 @@
+/*-------------------------------------------------------------------------
+ *
+ * prepagg.c
+ *   Routines to preprocess aggregate function calls
+ *
+ * If there are identical aggregate calls in the query, they only need to
+ * be computed once.  Also, some aggregate functions can share the same
+ * transition state, so that we only need to call the final function for
+ * them separately.  These optimizations are independent of how the
+ * aggregates are executed.
+ *
+ * preprocess_aggrefs() detects those cases, creates AggInfo and
+ * AggTransInfo structs for each aggregate and transition state that needs
+ * to be computed, and sets the 'aggno' and 'transno' fields in the Aggrefs
+ * accordingly.  It also resolves polymorphic transition types, and sets
+ * the 'aggtranstype' fields accordingly.
+ *
+ * XXX: The AggInfo and AggTransInfo structs are thrown away after
+ * planning, so executor startup has to perform some of the same lookups
+ * of transition functions and initial values that we do here.  One day, we
+ * might want to carry that information to the Agg nodes to save the effort
+ * at executor startup.  The Agg nodes are constructed much later in the
+ * planning, however, so it's not trivial.
+ *
+ * Portions Copyright (c) 1996-2019, PostgreSQL Global Development Group
+ * Portions Copyright (c) 1994, Regents of the University of California
+ *
+ *
+ * IDENTIFICATION
+ *   src/backend/optimizer/prep/prepagg.c
+ *
+ *-------------------------------------------------------------------------
+ */
+
+#include "postgres.h"
+
+#include "access/htup_details.h"
+#include "catalog/pg_aggregate.h"
+#include "catalog/pg_type.h"
+#include "nodes/nodeFuncs.h"
+#include "nodes/pathnodes.h"
+#include "optimizer/clauses.h"
+#include "optimizer/cost.h"
+#include "optimizer/optimizer.h"
+#include "optimizer/plancat.h"
+#include "optimizer/prep.h"
+#include "parser/parse_agg.h"
+#include "utils/builtins.h"
+#include "utils/datum.h"
+#include "utils/fmgroids.h"
+#include "utils/lsyscache.h"
+#include "utils/memutils.h"
+#include "utils/syscache.h"
+
+static bool preprocess_aggrefs_walker(Node *node, PlannerInfo *root);
+static int find_compatible_agg(PlannerInfo *root, Aggref *newagg,
+                               List **same_input_transnos);
+static int find_compatible_trans(PlannerInfo *root, Aggref *newagg,
+                                 bool shareable,
+                                 Oid aggtransfn, Oid aggtranstype,
+                                 int transtypeLen, bool transtypeByVal,
+                                 Oid aggcombinefn,
+                                 Oid aggserialfn, Oid aggdeserialfn,
+                                 Datum initValue, bool initValueIsNull,
+                                 List *transnos);
+static Datum GetAggInitVal(Datum textInitVal, Oid transtype);
+
+/* -----------------
+ * Resolve the transition type of all Aggrefs, and determine which Aggrefs
+ * can share aggregate or transition state.
+ *
+ * Information about the aggregates and transition functions are collected
+ * in the root->agginfos and root->aggtransinfos lists.  The 'aggtranstype',
+ * 'aggno', and 'aggtransno' fields in are filled in in each Aggref.
+ *
+ * NOTE: This modifies the Aggrefs in the input expression in-place!
+ *
+ * We try to optimize by detecting duplicate aggregate functions so that
+ * their state and final values are re-used, rather than needlessly being
+ * re-calculated independently.  We also detect aggregates that are not
+ * the same, but which can share the same transition state.
+ *
+ * Scenarios:
+ *
+ * 1. Identical aggregate function calls appear in the query:
+ *
+ *   SELECT SUM(x) FROM ... HAVING SUM(x) > 0
+ *
+ *   Since these aggregates are identical, we only need to calculate
+ *   the value once.  Both aggregates will share the same 'aggno' value.
+ *
+ * 2. Two different aggregate functions appear in the query, but the
+ *   aggregates have the same arguments, transition functions and
+ *   initial values (and, presumably, different final functions):
+ *
+ *   SELECT AVG(x), STDDEV(x) FROM ...
+ *
+ *   In this case we must create a new AggInfo for the varying aggregate,
+ *   and we need to call the final functions separately, but we need
+ *   only run the transition function once.  (This requires that the
+ *   final functions be nondestructive of the transition state, but
+ *   that's required anyway for other reasons.)
+ *
+ * For either of these optimizations to be valid, all aggregate properties
+ * used in the transition phase must be the same, including any modifiers
+ * such as ORDER BY, DISTINCT and FILTER, and the arguments mustn't
+ * contain any volatile functions.
+ * -----------------
+ */
+void
+preprocess_aggrefs(PlannerInfo *root, Node *clause)
+{
+   (void) preprocess_aggrefs_walker(clause, root);
+}
+
+static void
+preprocess_aggref(Aggref *aggref, PlannerInfo *root)
+{
+   HeapTuple   aggTuple;
+   Form_pg_aggregate aggform;
+   Oid         aggtransfn;
+   Oid         aggfinalfn;
+   Oid         aggcombinefn;
+   Oid         aggserialfn;
+   Oid         aggdeserialfn;
+   Oid         aggtranstype;
+   int32       aggtranstypmod;
+   int32       aggtransspace;
+   bool        shareable;
+   int         aggno;
+   int         transno;
+   List       *same_input_transnos;
+   int16       resulttypeLen;
+   bool        resulttypeByVal;
+   Datum       textInitVal;
+   Datum       initValue;
+   bool        initValueIsNull;
+   bool        transtypeByVal;
+   int16       transtypeLen;
+   Oid         inputTypes[FUNC_MAX_ARGS];
+   int         numArguments;
+
+   Assert(aggref->agglevelsup == 0);
+
+   /*
+    * Fetch info about the aggregate from pg_aggregate.  Note it's correct to
+    * ignore the moving-aggregate variant, since what we're concerned with
+    * here is aggregates not window functions.
+    */
+   aggTuple = SearchSysCache1(AGGFNOID,
+                              ObjectIdGetDatum(aggref->aggfnoid));
+   if (!HeapTupleIsValid(aggTuple))
+       elog(ERROR, "cache lookup failed for aggregate %u",
+            aggref->aggfnoid);
+   aggform = (Form_pg_aggregate) GETSTRUCT(aggTuple);
+   aggtransfn = aggform->aggtransfn;
+   aggfinalfn = aggform->aggfinalfn;
+   aggcombinefn = aggform->aggcombinefn;
+   aggserialfn = aggform->aggserialfn;
+   aggdeserialfn = aggform->aggdeserialfn;
+   aggtranstype = aggform->aggtranstype;
+   aggtransspace = aggform->aggtransspace;
+
+   /*
+    * Resolve the possibly-polymorphic aggregate transition type.
+    */
+
+   /* extract argument types (ignoring any ORDER BY expressions) */
+   numArguments = get_aggregate_argtypes(aggref, inputTypes);
+
+   /* resolve actual type of transition state, if polymorphic */
+   aggtranstype = resolve_aggregate_transtype(aggref->aggfnoid,
+                                              aggtranstype,
+                                              inputTypes,
+                                              numArguments);
+   aggref->aggtranstype = aggtranstype;
+
+   /*
+    * If transition state is of same type as first aggregated input, assume
+    * it's the same typmod (same width) as well.  This works for cases like
+    * MAX/MIN and is probably somewhat reasonable otherwise.
+    */
+   aggtranstypmod = -1;
+   if (aggref->args)
+   {
+       TargetEntry *tle = (TargetEntry *) linitial(aggref->args);
+
+       if (aggtranstype == exprType((Node *) tle->expr))
+           aggtranstypmod = exprTypmod((Node *) tle->expr);
+   }
+
+   /*
+    * If finalfn is marked read-write, we can't share transition states; but
+    * it is okay to share states for AGGMODIFY_SHAREABLE aggs.
+    *
+    * In principle, in a partial aggregate, we could share the transition
+    * state even if the final function is marked as read-write, because the
+    * partial aggregate doesn't execute the final function.  But it's too
+    * early to know whether we're going perform a partial aggregate.
+    */
+   shareable = (aggform->aggfinalmodify != AGGMODIFY_READ_WRITE);
+
+   /* get info about the output value's datatype */
+   get_typlenbyval(aggref->aggtype,
+                   &resulttypeLen,
+                   &resulttypeByVal);
+
+   /* get initial value */
+   textInitVal = SysCacheGetAttr(AGGFNOID, aggTuple,
+                                 Anum_pg_aggregate_agginitval,
+                                 &initValueIsNull);
+   if (initValueIsNull)
+       initValue = (Datum) 0;
+   else
+       initValue = GetAggInitVal(textInitVal, aggtranstype);
+
+   ReleaseSysCache(aggTuple);
+
+   /*
+    * 1. See if this is identical to another aggregate function call that
+    * we've seen already.
+    */
+   aggno = find_compatible_agg(root, aggref, &same_input_transnos);
+   if (aggno != -1)
+   {
+       AggInfo    *agginfo = list_nth(root->agginfos, aggno);
+
+       transno = agginfo->transno;
+   }
+   else
+   {
+       AggInfo    *agginfo = palloc(sizeof(AggInfo));
+
+       agginfo->finalfn_oid = aggfinalfn;
+       agginfo->representative_aggref = aggref;
+       agginfo->shareable = shareable;
+
+       aggno = list_length(root->agginfos);
+       root->agginfos = lappend(root->agginfos, agginfo);
+
+       /*
+        * Count it, and check for cases requiring ordered input.  Note that
+        * ordered-set aggs always have nonempty aggorder.  Any ordered-input
+        * case also defeats partial aggregation.
+        */
+       if (aggref->aggorder != NIL || aggref->aggdistinct != NIL)
+       {
+           root->numOrderedAggs++;
+           root->hasNonPartialAggs = true;
+       }
+
+       get_typlenbyval(aggtranstype,
+                       &transtypeLen,
+                       &transtypeByVal);
+
+       /*
+        * 2. See if this aggregate can share transition state with another
+        * aggregate that we've initialized already.
+        */
+       transno = find_compatible_trans(root, aggref, shareable,
+                                       aggtransfn, aggtranstype,
+                                       transtypeLen, transtypeByVal,
+                                       aggcombinefn,
+                                       aggserialfn, aggdeserialfn,
+                                       initValue, initValueIsNull,
+                                       same_input_transnos);
+       if (transno == -1)
+       {
+           AggTransInfo *transinfo = palloc(sizeof(AggTransInfo));
+
+           transinfo->args = aggref->args;
+           transinfo->aggfilter = aggref->aggfilter;
+           transinfo->transfn_oid = aggtransfn;
+           transinfo->combinefn_oid = aggcombinefn;
+           transinfo->serialfn_oid = aggserialfn;
+           transinfo->deserialfn_oid = aggdeserialfn;
+           transinfo->aggtranstype = aggtranstype;
+           transinfo->aggtranstypmod = aggtranstypmod;
+           transinfo->transtypeLen = transtypeLen;
+           transinfo->transtypeByVal = transtypeByVal;
+           transinfo->aggtransspace = aggtransspace;
+           transinfo->initValue = initValue;
+           transinfo->initValueIsNull = initValueIsNull;
+
+           transno = list_length(root->aggtransinfos);
+           root->aggtransinfos = lappend(root->aggtransinfos, transinfo);
+
+           /*
+            * Check whether partial aggregation is feasible, unless we
+            * already found out that we can't do it.
+            */
+           if (!root->hasNonPartialAggs)
+           {
+               /*
+                * If there is no combine function, then partial aggregation
+                * is not possible.
+                */
+               if (!OidIsValid(transinfo->combinefn_oid))
+                   root->hasNonPartialAggs = true;
+
+               /*
+                * If we have any aggs with transtype INTERNAL then we must
+                * check whether they have serialization/deserialization
+                * functions; if not, we can't serialize partial-aggregation
+                * results.
+                */
+               else if (transinfo->aggtranstype == INTERNALOID &&
+                        (!OidIsValid(transinfo->serialfn_oid) ||
+                         !OidIsValid(transinfo->deserialfn_oid)))
+                   root->hasNonSerialAggs = true;
+           }
+       }
+       agginfo->transno = transno;
+   }
+
+   /*
+    * Fill in the fields in the Aggref (aggtranstype was set above already)
+    */
+   aggref->aggno = aggno;
+   aggref->aggtransno = transno;
+}
+
+static bool
+preprocess_aggrefs_walker(Node *node, PlannerInfo *root)
+{
+   if (node == NULL)
+       return false;
+   if (IsA(node, Aggref))
+   {
+       Aggref     *aggref = (Aggref *) node;
+
+       preprocess_aggref(aggref, root);
+
+       /*
+        * We assume that the parser checked that there are no aggregates (of
+        * this level anyway) in the aggregated arguments, direct arguments,
+        * or filter clause.  Hence, we need not recurse into any of them.
+        */
+       return false;
+   }
+   Assert(!IsA(node, SubLink));
+   return expression_tree_walker(node, preprocess_aggrefs_walker,
+                                 (void *) root);
+}
+
+
+/*
+ * find_compatible_agg - search for a previously initialized per-Agg struct
+ *
+ * Searches the previously looked at aggregates to find one which is compatible
+ * with this one, with the same input parameters.  If no compatible aggregate
+ * can be found, returns -1.
+ *
+ * As a side-effect, this also collects a list of existing, shareable per-Trans
+ * structs with matching inputs.  If no identical Aggref is found, the list is
+ * passed later to find_compatible_trans, to see if we can at least reuse
+ * the state value of another aggregate.
+ */
+static int
+find_compatible_agg(PlannerInfo *root, Aggref *newagg,
+                   List **same_input_transnos)
+{
+   ListCell   *lc;
+   int         aggno;
+
+   *same_input_transnos = NIL;
+
+   /* we mustn't reuse the aggref if it contains volatile function calls */
+   if (contain_volatile_functions((Node *) newagg))
+       return -1;
+
+   /*
+    * Search through the list of already seen aggregates.  If we find an
+    * existing identical aggregate call, then we can re-use that one.  While
+    * searching, we'll also collect a list of Aggrefs with the same input
+    * parameters.  If no matching Aggref is found, the caller can potentially
+    * still re-use the transition state of one of them.  (At this stage we
+    * just compare the parsetrees; whether different aggregates share the
+    * same transition function will be checked later.)
+    */
+   aggno = -1;
+   foreach(lc, root->agginfos)
+   {
+       AggInfo    *agginfo = (AggInfo *) lfirst(lc);
+       Aggref     *existingRef;
+
+       aggno++;
+
+       existingRef = agginfo->representative_aggref;
+
+       /* all of the following must be the same or it's no match */
+       if (newagg->inputcollid != existingRef->inputcollid ||
+           newagg->aggtranstype != existingRef->aggtranstype ||
+           newagg->aggstar != existingRef->aggstar ||
+           newagg->aggvariadic != existingRef->aggvariadic ||
+           newagg->aggkind != existingRef->aggkind ||
+           !equal(newagg->args, existingRef->args) ||
+           !equal(newagg->aggorder, existingRef->aggorder) ||
+           !equal(newagg->aggdistinct, existingRef->aggdistinct) ||
+           !equal(newagg->aggfilter, existingRef->aggfilter))
+           continue;
+
+       /* if it's the same aggregate function then report exact match */
+       if (newagg->aggfnoid == existingRef->aggfnoid &&
+           newagg->aggtype == existingRef->aggtype &&
+           newagg->aggcollid == existingRef->aggcollid &&
+           equal(newagg->aggdirectargs, existingRef->aggdirectargs))
+       {
+           list_free(*same_input_transnos);
+           *same_input_transnos = NIL;
+           return aggno;
+       }
+
+       /*
+        * Not identical, but it had the same inputs.  If the final function
+        * permits sharing, return its transno to the caller, in case we can
+        * re-use its per-trans state.  (If there's already sharing going on,
+        * we might report a transno more than once.  find_compatible_trans is
+        * cheap enough that it's not worth spending cycles to avoid that.)
+        */
+       if (agginfo->shareable)
+           *same_input_transnos = lappend_int(*same_input_transnos,
+                                              agginfo->transno);
+   }
+
+   return -1;
+}
+
+/*
+ * find_compatible_trans - search for a previously initialized per-Trans
+ * struct
+ *
+ * Searches the list of transnos for a per-Trans struct with the same
+ * transition function and initial condition. (The inputs have already been
+ * verified to match.)
+ */
+static int
+find_compatible_trans(PlannerInfo *root, Aggref *newagg, bool shareable,
+                     Oid aggtransfn, Oid aggtranstype,
+                     int transtypeLen, bool transtypeByVal,
+                     Oid aggcombinefn,
+                     Oid aggserialfn, Oid aggdeserialfn,
+                     Datum initValue, bool initValueIsNull,
+                     List *transnos)
+{
+   ListCell   *lc;
+
+   /* If this aggregate can't share transition states, give up */
+   if (!shareable)
+       return -1;
+
+   foreach(lc, transnos)
+   {
+       int         transno = lfirst_int(lc);
+       AggTransInfo *pertrans = (AggTransInfo *) list_nth(root->aggtransinfos, transno);
+
+       /*
+        * if the transfns or transition state types are not the same then the
+        * state can't be shared.
+        */
+       if (aggtransfn != pertrans->transfn_oid ||
+           aggtranstype != pertrans->aggtranstype)
+           continue;
+
+       /*
+        * The serialization and deserialization functions must match, if
+        * present, as we're unable to share the trans state for aggregates
+        * which will serialize or deserialize into different formats.
+        * Remember that these will be InvalidOid if they're not required for
+        * this agg node.
+        */
+       if (aggserialfn != pertrans->serialfn_oid ||
+           aggdeserialfn != pertrans->deserialfn_oid)
+           continue;
+
+       /*
+        * Combine function must also match.  We only care about the combine
+        * function with partial aggregates, but it's too early in the
+        * planning to know if we will do partial aggregation, so be
+        * conservative.
+        */
+       if (aggcombinefn != pertrans->combinefn_oid)
+           continue;
+
+       /*
+        * Check that the initial condition matches, too.
+        */
+       if (initValueIsNull && pertrans->initValueIsNull)
+           return transno;
+
+       if (!initValueIsNull && !pertrans->initValueIsNull &&
+           datumIsEqual(initValue, pertrans->initValue,
+                        transtypeByVal, transtypeLen))
+           return transno;
+   }
+   return -1;
+}
+
+static Datum
+GetAggInitVal(Datum textInitVal, Oid transtype)
+{
+   Oid         typinput,
+               typioparam;
+   char       *strInitVal;
+   Datum       initVal;
+
+   getTypeInputInfo(transtype, &typinput, &typioparam);
+   strInitVal = TextDatumGetCString(textInitVal);
+   initVal = OidInputFunctionCall(typinput, strInitVal,
+                                  typioparam, -1);
+   pfree(strInitVal);
+   return initVal;
+}
+
+
+/*
+ * get_agg_clause_costs
+ *   Recursively find the Aggref nodes in an expression tree, and
+ *   accumulate cost information about them.
+ *
+ * 'aggsplit' tells us the expected partial-aggregation mode, which affects
+ * the cost estimates.
+ *
+ * NOTE that the counts/costs are ADDED to those already in *costs ... so
+ * the caller is responsible for zeroing the struct initially.
+ *
+ * We count the nodes, estimate their execution costs, and estimate the total
+ * space needed for their transition state values if all are evaluated in
+ * parallel (as would be done in a HashAgg plan).  Also, we check whether
+ * partial aggregation is feasible.  See AggClauseCosts for the exact set
+ * of statistics collected.
+ *
+ * In addition, we mark Aggref nodes with the correct aggtranstype, so
+ * that that doesn't need to be done repeatedly.  (That makes this function's
+ * name a bit of a misnomer.)
+ *
+ * This does not descend into subqueries, and so should be used only after
+ * reduction of sublinks to subplans, or in contexts where it's known there
+ * are no subqueries.  There mustn't be outer-aggregate references either.
+ */
+void
+get_agg_clause_costs(PlannerInfo *root, AggSplit aggsplit, AggClauseCosts *costs)
+{
+   ListCell   *lc;
+
+   foreach(lc, root->aggtransinfos)
+   {
+       AggTransInfo *transinfo = (AggTransInfo *) lfirst(lc);
+
+       /*
+        * Add the appropriate component function execution costs to
+        * appropriate totals.
+        */
+       if (DO_AGGSPLIT_COMBINE(aggsplit))
+       {
+           /* charge for combining previously aggregated states */
+           add_function_cost(root, transinfo->combinefn_oid, NULL,
+                             &costs->transCost);
+       }
+       else
+           add_function_cost(root, transinfo->transfn_oid, NULL,
+                             &costs->transCost);
+       if (DO_AGGSPLIT_DESERIALIZE(aggsplit) &&
+           OidIsValid(transinfo->deserialfn_oid))
+           add_function_cost(root, transinfo->deserialfn_oid, NULL,
+                             &costs->transCost);
+       if (DO_AGGSPLIT_SERIALIZE(aggsplit) &&
+           OidIsValid(transinfo->serialfn_oid))
+           add_function_cost(root, transinfo->serialfn_oid, NULL,
+                             &costs->finalCost);
+
+       /*
+        * These costs are incurred only by the initial aggregate node, so we
+        * mustn't include them again at upper levels.
+        */
+       if (!DO_AGGSPLIT_COMBINE(aggsplit))
+       {
+           /* add the input expressions' cost to per-input-row costs */
+           QualCost    argcosts;
+
+           cost_qual_eval_node(&argcosts, (Node *) transinfo->args, root);
+           costs->transCost.startup += argcosts.startup;
+           costs->transCost.per_tuple += argcosts.per_tuple;
+
+           /*
+            * Add any filter's cost to per-input-row costs.
+            *
+            * XXX Ideally we should reduce input expression costs according
+            * to filter selectivity, but it's not clear it's worth the
+            * trouble.
+            */
+           if (transinfo->aggfilter)
+           {
+               cost_qual_eval_node(&argcosts, (Node *) transinfo->aggfilter,
+                                   root);
+               costs->transCost.startup += argcosts.startup;
+               costs->transCost.per_tuple += argcosts.per_tuple;
+           }
+       }
+
+       /*
+        * If the transition type is pass-by-value then it doesn't add
+        * anything to the required size of the hashtable.  If it is
+        * pass-by-reference then we have to add the estimated size of the
+        * value itself, plus palloc overhead.
+        */
+       if (!transinfo->transtypeByVal)
+       {
+           int32       avgwidth;
+
+           /* Use average width if aggregate definition gave one */
+           if (transinfo->aggtransspace > 0)
+               avgwidth = transinfo->aggtransspace;
+           else if (transinfo->transfn_oid == F_ARRAY_APPEND)
+           {
+               /*
+                * If the transition function is array_append(), it'll use an
+                * expanded array as transvalue, which will occupy at least
+                * ALLOCSET_SMALL_INITSIZE and possibly more.  Use that as the
+                * estimate for lack of a better idea.
+                */
+               avgwidth = ALLOCSET_SMALL_INITSIZE;
+           }
+           else
+           {
+               avgwidth = get_typavgwidth(transinfo->aggtranstype, transinfo->aggtranstypmod);
+           }
+
+           avgwidth = MAXALIGN(avgwidth);
+           costs->transitionSpace += avgwidth + 2 * sizeof(void *);
+       }
+       else if (transinfo->aggtranstype == INTERNALOID)
+       {
+           /*
+            * INTERNAL transition type is a special case: although INTERNAL
+            * is pass-by-value, it's almost certainly being used as a pointer
+            * to some large data structure.  The aggregate definition can
+            * provide an estimate of the size.  If it doesn't, then we assume
+            * ALLOCSET_DEFAULT_INITSIZE, which is a good guess if the data is
+            * being kept in a private memory context, as is done by
+            * array_agg() for instance.
+            */
+           if (transinfo->aggtransspace > 0)
+               costs->transitionSpace += transinfo->aggtransspace;
+           else
+               costs->transitionSpace += ALLOCSET_DEFAULT_INITSIZE;
+       }
+   }
+
+   foreach(lc, root->agginfos)
+   {
+       AggInfo    *agginfo = (AggInfo *) lfirst(lc);
+       Aggref     *aggref = agginfo->representative_aggref;
+
+       /*
+        * Add the appropriate component function execution costs to
+        * appropriate totals.
+        */
+       if (!DO_AGGSPLIT_SKIPFINAL(aggsplit) &&
+           OidIsValid(agginfo->finalfn_oid))
+           add_function_cost(root, agginfo->finalfn_oid, NULL,
+                             &costs->finalCost);
+
+       /*
+        * If there are direct arguments, treat their evaluation cost like the
+        * cost of the finalfn.
+        */
+       if (aggref->aggdirectargs)
+       {
+           QualCost    argcosts;
+
+           cost_qual_eval_node(&argcosts, (Node *) aggref->aggdirectargs,
+                               root);
+           costs->finalCost.startup += argcosts.startup;
+           costs->finalCost.per_tuple += argcosts.per_tuple;
+       }
+   }
+}
index 85ef873caaf8a9e3e93ff4ddfcc9806518fa1b4e..587d494c34f966f0a46877b921e9d42c4206a4fc 100644 (file)
 #include "utils/syscache.h"
 #include "utils/typcache.h"
 
-
-typedef struct
-{
-   PlannerInfo *root;
-   AggSplit    aggsplit;
-   AggClauseCosts *costs;
-} get_agg_clause_costs_context;
-
 typedef struct
 {
    ParamListInfo boundParams;
@@ -98,8 +90,6 @@ typedef struct
 } max_parallel_hazard_context;
 
 static bool contain_agg_clause_walker(Node *node, void *context);
-static bool get_agg_clause_costs_walker(Node *node,
-                                       get_agg_clause_costs_context *context);
 static bool find_window_functions_walker(Node *node, WindowFuncLists *lists);
 static bool contain_subplans_walker(Node *node, void *context);
 static bool contain_mutable_functions_walker(Node *node, void *context);
@@ -200,284 +190,6 @@ contain_agg_clause_walker(Node *node, void *context)
    return expression_tree_walker(node, contain_agg_clause_walker, context);
 }
 
-/*
- * get_agg_clause_costs
- *   Recursively find the Aggref nodes in an expression tree, and
- *   accumulate cost information about them.
- *
- * 'aggsplit' tells us the expected partial-aggregation mode, which affects
- * the cost estimates.
- *
- * NOTE that the counts/costs are ADDED to those already in *costs ... so
- * the caller is responsible for zeroing the struct initially.
- *
- * We count the nodes, estimate their execution costs, and estimate the total
- * space needed for their transition state values if all are evaluated in
- * parallel (as would be done in a HashAgg plan).  Also, we check whether
- * partial aggregation is feasible.  See AggClauseCosts for the exact set
- * of statistics collected.
- *
- * In addition, we mark Aggref nodes with the correct aggtranstype, so
- * that that doesn't need to be done repeatedly.  (That makes this function's
- * name a bit of a misnomer.)
- *
- * This does not descend into subqueries, and so should be used only after
- * reduction of sublinks to subplans, or in contexts where it's known there
- * are no subqueries.  There mustn't be outer-aggregate references either.
- */
-void
-get_agg_clause_costs(PlannerInfo *root, Node *clause, AggSplit aggsplit,
-                    AggClauseCosts *costs)
-{
-   get_agg_clause_costs_context context;
-
-   context.root = root;
-   context.aggsplit = aggsplit;
-   context.costs = costs;
-   (void) get_agg_clause_costs_walker(clause, &context);
-}
-
-static bool
-get_agg_clause_costs_walker(Node *node, get_agg_clause_costs_context *context)
-{
-   if (node == NULL)
-       return false;
-   if (IsA(node, Aggref))
-   {
-       Aggref     *aggref = (Aggref *) node;
-       AggClauseCosts *costs = context->costs;
-       HeapTuple   aggTuple;
-       Form_pg_aggregate aggform;
-       Oid         aggtransfn;
-       Oid         aggfinalfn;
-       Oid         aggcombinefn;
-       Oid         aggserialfn;
-       Oid         aggdeserialfn;
-       Oid         aggtranstype;
-       int32       aggtransspace;
-       QualCost    argcosts;
-
-       Assert(aggref->agglevelsup == 0);
-
-       /*
-        * Fetch info about aggregate from pg_aggregate.  Note it's correct to
-        * ignore the moving-aggregate variant, since what we're concerned
-        * with here is aggregates not window functions.
-        */
-       aggTuple = SearchSysCache1(AGGFNOID,
-                                  ObjectIdGetDatum(aggref->aggfnoid));
-       if (!HeapTupleIsValid(aggTuple))
-           elog(ERROR, "cache lookup failed for aggregate %u",
-                aggref->aggfnoid);
-       aggform = (Form_pg_aggregate) GETSTRUCT(aggTuple);
-       aggtransfn = aggform->aggtransfn;
-       aggfinalfn = aggform->aggfinalfn;
-       aggcombinefn = aggform->aggcombinefn;
-       aggserialfn = aggform->aggserialfn;
-       aggdeserialfn = aggform->aggdeserialfn;
-       aggtranstype = aggform->aggtranstype;
-       aggtransspace = aggform->aggtransspace;
-       ReleaseSysCache(aggTuple);
-
-       /*
-        * Resolve the possibly-polymorphic aggregate transition type, unless
-        * already done in a previous pass over the expression.
-        */
-       if (OidIsValid(aggref->aggtranstype))
-           aggtranstype = aggref->aggtranstype;
-       else
-       {
-           Oid         inputTypes[FUNC_MAX_ARGS];
-           int         numArguments;
-
-           /* extract argument types (ignoring any ORDER BY expressions) */
-           numArguments = get_aggregate_argtypes(aggref, inputTypes);
-
-           /* resolve actual type of transition state, if polymorphic */
-           aggtranstype = resolve_aggregate_transtype(aggref->aggfnoid,
-                                                      aggtranstype,
-                                                      inputTypes,
-                                                      numArguments);
-           aggref->aggtranstype = aggtranstype;
-       }
-
-       /*
-        * Count it, and check for cases requiring ordered input.  Note that
-        * ordered-set aggs always have nonempty aggorder.  Any ordered-input
-        * case also defeats partial aggregation.
-        */
-       costs->numAggs++;
-       if (aggref->aggorder != NIL || aggref->aggdistinct != NIL)
-       {
-           costs->numOrderedAggs++;
-           costs->hasNonPartial = true;
-       }
-
-       /*
-        * Check whether partial aggregation is feasible, unless we already
-        * found out that we can't do it.
-        */
-       if (!costs->hasNonPartial)
-       {
-           /*
-            * If there is no combine function, then partial aggregation is
-            * not possible.
-            */
-           if (!OidIsValid(aggcombinefn))
-               costs->hasNonPartial = true;
-
-           /*
-            * If we have any aggs with transtype INTERNAL then we must check
-            * whether they have serialization/deserialization functions; if
-            * not, we can't serialize partial-aggregation results.
-            */
-           else if (aggtranstype == INTERNALOID &&
-                    (!OidIsValid(aggserialfn) || !OidIsValid(aggdeserialfn)))
-               costs->hasNonSerial = true;
-       }
-
-       /*
-        * Add the appropriate component function execution costs to
-        * appropriate totals.
-        */
-       if (DO_AGGSPLIT_COMBINE(context->aggsplit))
-       {
-           /* charge for combining previously aggregated states */
-           add_function_cost(context->root, aggcombinefn, NULL,
-                             &costs->transCost);
-       }
-       else
-           add_function_cost(context->root, aggtransfn, NULL,
-                             &costs->transCost);
-       if (DO_AGGSPLIT_DESERIALIZE(context->aggsplit) &&
-           OidIsValid(aggdeserialfn))
-           add_function_cost(context->root, aggdeserialfn, NULL,
-                             &costs->transCost);
-       if (DO_AGGSPLIT_SERIALIZE(context->aggsplit) &&
-           OidIsValid(aggserialfn))
-           add_function_cost(context->root, aggserialfn, NULL,
-                             &costs->finalCost);
-       if (!DO_AGGSPLIT_SKIPFINAL(context->aggsplit) &&
-           OidIsValid(aggfinalfn))
-           add_function_cost(context->root, aggfinalfn, NULL,
-                             &costs->finalCost);
-
-       /*
-        * These costs are incurred only by the initial aggregate node, so we
-        * mustn't include them again at upper levels.
-        */
-       if (!DO_AGGSPLIT_COMBINE(context->aggsplit))
-       {
-           /* add the input expressions' cost to per-input-row costs */
-           cost_qual_eval_node(&argcosts, (Node *) aggref->args, context->root);
-           costs->transCost.startup += argcosts.startup;
-           costs->transCost.per_tuple += argcosts.per_tuple;
-
-           /*
-            * Add any filter's cost to per-input-row costs.
-            *
-            * XXX Ideally we should reduce input expression costs according
-            * to filter selectivity, but it's not clear it's worth the
-            * trouble.
-            */
-           if (aggref->aggfilter)
-           {
-               cost_qual_eval_node(&argcosts, (Node *) aggref->aggfilter,
-                                   context->root);
-               costs->transCost.startup += argcosts.startup;
-               costs->transCost.per_tuple += argcosts.per_tuple;
-           }
-       }
-
-       /*
-        * If there are direct arguments, treat their evaluation cost like the
-        * cost of the finalfn.
-        */
-       if (aggref->aggdirectargs)
-       {
-           cost_qual_eval_node(&argcosts, (Node *) aggref->aggdirectargs,
-                               context->root);
-           costs->finalCost.startup += argcosts.startup;
-           costs->finalCost.per_tuple += argcosts.per_tuple;
-       }
-
-       /*
-        * If the transition type is pass-by-value then it doesn't add
-        * anything to the required size of the hashtable.  If it is
-        * pass-by-reference then we have to add the estimated size of the
-        * value itself, plus palloc overhead.
-        */
-       if (!get_typbyval(aggtranstype))
-       {
-           int32       avgwidth;
-
-           /* Use average width if aggregate definition gave one */
-           if (aggtransspace > 0)
-               avgwidth = aggtransspace;
-           else if (aggtransfn == F_ARRAY_APPEND)
-           {
-               /*
-                * If the transition function is array_append(), it'll use an
-                * expanded array as transvalue, which will occupy at least
-                * ALLOCSET_SMALL_INITSIZE and possibly more.  Use that as the
-                * estimate for lack of a better idea.
-                */
-               avgwidth = ALLOCSET_SMALL_INITSIZE;
-           }
-           else
-           {
-               /*
-                * If transition state is of same type as first aggregated
-                * input, assume it's the same typmod (same width) as well.
-                * This works for cases like MAX/MIN and is probably somewhat
-                * reasonable otherwise.
-                */
-               int32       aggtranstypmod = -1;
-
-               if (aggref->args)
-               {
-                   TargetEntry *tle = (TargetEntry *) linitial(aggref->args);
-
-                   if (aggtranstype == exprType((Node *) tle->expr))
-                       aggtranstypmod = exprTypmod((Node *) tle->expr);
-               }
-
-               avgwidth = get_typavgwidth(aggtranstype, aggtranstypmod);
-           }
-
-           avgwidth = MAXALIGN(avgwidth);
-           costs->transitionSpace += avgwidth + 2 * sizeof(void *);
-       }
-       else if (aggtranstype == INTERNALOID)
-       {
-           /*
-            * INTERNAL transition type is a special case: although INTERNAL
-            * is pass-by-value, it's almost certainly being used as a pointer
-            * to some large data structure.  The aggregate definition can
-            * provide an estimate of the size.  If it doesn't, then we assume
-            * ALLOCSET_DEFAULT_INITSIZE, which is a good guess if the data is
-            * being kept in a private memory context, as is done by
-            * array_agg() for instance.
-            */
-           if (aggtransspace > 0)
-               costs->transitionSpace += aggtransspace;
-           else
-               costs->transitionSpace += ALLOCSET_DEFAULT_INITSIZE;
-       }
-
-       /*
-        * We assume that the parser checked that there are no aggregates (of
-        * this level anyway) in the aggregated arguments, direct arguments,
-        * or filter clause.  Hence, we need not recurse into any of them.
-        */
-       return false;
-   }
-   Assert(!IsA(node, SubLink));
-   return expression_tree_walker(node, get_agg_clause_costs_walker,
-                                 (void *) context);
-}
-
-
 /*****************************************************************************
  *     Window-function clause manipulation
  *****************************************************************************/
index 8b4e3ca5e113e77cbab8850bae088aaabb37d91d..23ac2a2fe65330ae2e647423c30265e3ef755b8b 100644 (file)
@@ -769,6 +769,8 @@ ParseFuncOrColumn(ParseState *pstate, List *funcname, List *fargs,
        aggref->aggkind = aggkind;
        /* agglevelsup will be set by transformAggregateCall */
        aggref->aggsplit = AGGSPLIT_SIMPLE; /* planner might change this */
+       aggref->aggno = -1;     /* planner will set aggno and aggtransno */
+       aggref->aggtransno = -1;
        aggref->location = location;
 
        /*
index bec357fcef042b0fc68596ea2613369631265a58..80bd60f8767b7606f365e3e1e6441893ebb0aaed 100644 (file)
@@ -3839,12 +3839,14 @@ estimate_hash_bucket_stats(PlannerInfo *root, Node *hashkey, double nbuckets,
  * won't store them.  Is this a problem?
  */
 double
-estimate_hashagg_tablesize(Path *path, const AggClauseCosts *agg_costs,
-                          double dNumGroups)
+estimate_hashagg_tablesize(PlannerInfo *root, Path *path,
+                          const AggClauseCosts *agg_costs, double dNumGroups)
 {
-   Size        hashentrysize = hash_agg_entry_size(agg_costs->numAggs,
-                                                   path->pathtarget->width,
-                                                   agg_costs->transitionSpace);
+   Size        hashentrysize;
+
+   hashentrysize = hash_agg_entry_size(list_length(root->aggtransinfos),
+                                       path->pathtarget->width,
+                                       agg_costs->transitionSpace);
 
    /*
     * Note that this disregards the effect of fill-factor and growth policy
index 1276ff8bdac947d4d651b92cf4ad4aeaa7e7dae9..087efddff9b7d4ed1276db5dacdf2e86d7d5185e 100644 (file)
@@ -53,6 +53,6 @@
  */
 
 /*                         yyyymmddN */
-#define CATALOG_VERSION_NO 202011231
+#define CATALOG_VERSION_NO 202011241
 
 #endif
index b792de1bc95e87690eb2c7c13a640420c30f5869..abb489e2062b502b95f6c16f402dba4aae816376 100644 (file)
@@ -564,8 +564,7 @@ typedef struct ExprEvalStep
        /* for EEOP_AGGREF */
        struct
        {
-           /* out-of-line state, modified by nodeAgg.c */
-           AggrefExprState *astate;
+           int         aggno;
        }           aggref;
 
        /* for EEOP_GROUPING_FUNC */
index f6824bf2e1b0dda175d76ee1331e21680a0af15b..61ba4c36662ba470fcd26ed5b15fae33bd5f410c 100644 (file)
@@ -746,17 +746,6 @@ typedef tuplehash_iterator TupleHashIterator;
  * ----------------------------------------------------------------
  */
 
-/* ----------------
- *     AggrefExprState node
- * ----------------
- */
-typedef struct AggrefExprState
-{
-   NodeTag     type;
-   Aggref     *aggref;         /* expression plan node */
-   int         aggno;          /* ID number for agg within its plan node */
-} AggrefExprState;
-
 /* ----------------
  *     WindowFuncExprState node
  * ----------------
index 7ddd8c011bfc329ae67a71f01bc4d71bc5b81935..3684f87a8831db8e1fbec6a8de0263194d5c1ac6 100644 (file)
@@ -206,10 +206,9 @@ typedef enum NodeTag
     * Most Expr-based plan nodes do not have a corresponding expression state
     * node, they're fully handled within execExpr* - but sometimes the state
     * needs to be shared with other parts of the executor, as for example
-    * with AggrefExprState, which nodeAgg.c has to modify.
+    * with SubPlanState, which nodeSubplan.c has to modify.
     */
    T_ExprState,
-   T_AggrefExprState,
    T_WindowFuncExprState,
    T_SetExprState,
    T_SubPlanState,
index 8f62d61702893c3f2f4e9f0c8af47fe51f930f26..abe6f570e3280354109f1a33c93114266c19dbc0 100644 (file)
@@ -55,10 +55,6 @@ typedef struct QualCost
  */
 typedef struct AggClauseCosts
 {
-   int         numAggs;        /* total number of aggregate functions */
-   int         numOrderedAggs; /* number w/ DISTINCT/ORDER BY/WITHIN GROUP */
-   bool        hasNonPartial;  /* does any agg not support partial mode? */
-   bool        hasNonSerial;   /* is any partial agg non-serializable? */
    QualCost    transCost;      /* total per-input-row execution costs */
    QualCost    finalCost;      /* total per-aggregated-row costs */
    Size        transitionSpace;    /* space for pass-by-ref transition data */
@@ -348,6 +344,15 @@ struct PlannerInfo
    bool        hasAlternativeSubPlans; /* true if we've made any of those */
    bool        hasRecursion;   /* true if planning a recursive WITH item */
 
+   /*
+    * Information about aggregates. Filled by preprocess_aggrefs().
+    */
+   List       *agginfos;       /* AggInfo structs */
+   List       *aggtransinfos;  /* AggTransInfo structs */
+   int         numOrderedAggs; /* number w/ DISTINCT/ORDER BY/WITHIN GROUP */
+   bool        hasNonPartialAggs;  /* does any agg not support partial mode? */
+   bool        hasNonSerialAggs;   /* is any partial agg non-serializable? */
+
    /* These fields are used only when hasRecursion is true: */
    int         wt_param_id;    /* PARAM_EXEC ID for the work table */
    struct Path *non_recursive_path;    /* a path for non-recursive term */
@@ -2549,4 +2554,71 @@ typedef struct JoinCostWorkspace
    double      inner_rows_total;
 } JoinCostWorkspace;
 
+/*
+ * AggInfo holds information about an aggregate that needs to be computed.
+ * Multiple Aggrefs in a query can refer to the same AggInfo by having the
+ * same 'aggno' value, so that the aggregate is computed only once.
+ */
+typedef struct AggInfo
+{
+   /*
+    * Link to an Aggref expr this state value is for.
+    *
+    * There can be multiple identical Aggref's sharing the same per-agg. This
+    * points to the first one of them.
+    */
+   Aggref     *representative_aggref;
+
+   int         transno;
+
+   /*
+    * "shareable" is false if this agg cannot share state values with other
+    * aggregates because the final function is read-write.
+    */
+   bool        shareable;
+
+   /* Oid of the final function or InvalidOid */
+   Oid         finalfn_oid;
+
+} AggInfo;
+
+/*
+ * AggTransInfo holds information about transition state that is used by one
+ * or more aggregates in the query.  Multiple aggregates can share the same
+ * transition state, if they have the same inputs and the same transition
+ * function.  Aggrefs that share the same transition info have the same
+ * 'aggtransno' value.
+ */
+typedef struct AggTransInfo
+{
+   List       *args;
+   Expr       *aggfilter;
+
+   /* Oid of the state transition function */
+   Oid         transfn_oid;
+
+   /* Oid of the serialization function or InvalidOid */
+   Oid         serialfn_oid;
+
+   /* Oid of the deserialization function or InvalidOid */
+   Oid         deserialfn_oid;
+
+   /* Oid of the combine function or InvalidOid */
+   Oid         combinefn_oid;
+
+   /* Oid of state value's datatype */
+   Oid         aggtranstype;
+   int32       aggtranstypmod;
+   int         transtypeLen;
+   bool        transtypeByVal;
+   int32       aggtransspace;
+
+   /*
+    * initial value from pg_aggregate entry
+    */
+   Datum       initValue;
+   bool        initValueIsNull;
+
+} AggTransInfo;
+
 #endif                         /* PATHNODES_H */
index 5b190bb99bde04370f031f19df8d325e21a4a492..cdbe781c73479ec1c0f1965428d39ceb853b700b 100644 (file)
@@ -305,6 +305,12 @@ typedef struct Param
  * a crosscheck that the Aggrefs match the plan; but note that when aggsplit
  * indicates a non-final mode, aggtype reflects the transition data type
  * not the SQL-level output type of the aggregate.
+ *
+ * aggno and aggtransno are -1 in the parse stage, and are set in planning.
+ * Aggregates with the same 'aggno' represent the same aggregate expression,
+ * and can share the result.  Aggregates with same 'transno' but different
+ * 'aggno' can share the same transition state, only the final function needs
+ * to be called separately.
  */
 typedef struct Aggref
 {
@@ -326,6 +332,8 @@ typedef struct Aggref
    char        aggkind;        /* aggregate kind (see pg_aggregate.h) */
    Index       agglevelsup;    /* > 0 if agg belongs to outer query */
    AggSplit    aggsplit;       /* expected agg-splitting mode of parent Agg */
+   int         aggno;          /* unique ID within the Agg node */
+   int         aggtransno;     /* unique ID of transition state in the Agg */
    int         location;       /* token location, or -1 if unknown */
 } Aggref;
 
index 2584ffc72fc6dc1c10452bf456a10f545eaa81b2..68855d0cee7aa05fbaf4dc080dad92b1b849ecc3 100644 (file)
@@ -24,8 +24,6 @@ typedef struct
 } WindowFuncLists;
 
 extern bool contain_agg_clause(Node *clause);
-extern void get_agg_clause_costs(PlannerInfo *root, Node *clause,
-                                AggSplit aggsplit, AggClauseCosts *costs);
 
 extern bool contain_window_function(Node *clause);
 extern WindowFuncLists *find_window_functions(Node *clause, Index maxWinRef);
index 19c92302b03261c5b15407552fc04bb331019219..0abe6bec00e782a1fb662adccb20e866e70e93fe 100644 (file)
@@ -38,9 +38,17 @@ extern List *preprocess_targetlist(PlannerInfo *root);
 
 extern PlanRowMark *get_plan_rowmark(List *rowmarks, Index rtindex);
 
+/*
+ * prototypes for prepagg.c
+ */
+extern void get_agg_clause_costs(PlannerInfo *root, AggSplit aggsplit,
+                                AggClauseCosts *agg_costs);
+extern void preprocess_aggrefs(PlannerInfo *root, Node *clause);
+
 /*
  * prototypes for prepunion.c
  */
 extern RelOptInfo *plan_set_operations(PlannerInfo *root);
 
+
 #endif                         /* PREP_H */
index 7ac4a063915968335b3d8cf19093a43d652598b2..3a2cfb7efa665c995c2ad690a7e29c220663624e 100644 (file)
@@ -200,7 +200,7 @@ extern void estimate_hash_bucket_stats(PlannerInfo *root,
                                       Node *hashkey, double nbuckets,
                                       Selectivity *mcv_freq,
                                       Selectivity *bucketsize_frac);
-extern double estimate_hashagg_tablesize(Path *path,
+extern double estimate_hashagg_tablesize(PlannerInfo *root, Path *path,
                                         const AggClauseCosts *agg_costs,
                                         double dNumGroups);
 
index 45c698daf4895092ea4c81f02490d8725c769d67..dfa4b036b52fa32738b79f602019d0265b0aa7c4 100644 (file)
@@ -1412,11 +1412,12 @@ SELECT y, sum(x), avg(x), count(*) FROM pagg_tab_para GROUP BY y HAVING avg(x) <
 (4 rows)
 
 -- Test when parent can produce parallel paths but not any (or some) of its children
+-- (Use one more aggregate to tilt the cost estimates for the plan we want)
 ALTER TABLE pagg_tab_para_p1 SET (parallel_workers = 0);
 ALTER TABLE pagg_tab_para_p3 SET (parallel_workers = 0);
 ANALYZE pagg_tab_para;
 EXPLAIN (COSTS OFF)
-SELECT x, sum(y), avg(y), count(*) FROM pagg_tab_para GROUP BY x HAVING avg(y) < 7 ORDER BY 1, 2, 3;
+SELECT x, sum(y), avg(y), sum(x+y), count(*) FROM pagg_tab_para GROUP BY x HAVING avg(y) < 7 ORDER BY 1, 2, 3;
                                         QUERY PLAN                                         
 -------------------------------------------------------------------------------------------
  Sort
@@ -1436,21 +1437,21 @@ SELECT x, sum(y), avg(y), count(*) FROM pagg_tab_para GROUP BY x HAVING avg(y) <
                                  ->  Parallel Seq Scan on pagg_tab_para_p2 pagg_tab_para_2
 (15 rows)
 
-SELECT x, sum(y), avg(y), count(*) FROM pagg_tab_para GROUP BY x HAVING avg(y) < 7 ORDER BY 1, 2, 3;
- x  | sum  |        avg         | count 
-----+------+--------------------+-------
-  0 | 5000 | 5.0000000000000000 |  1000
-  1 | 6000 | 6.0000000000000000 |  1000
- 10 | 5000 | 5.0000000000000000 |  1000
- 11 | 6000 | 6.0000000000000000 |  1000
- 20 | 5000 | 5.0000000000000000 |  1000
- 21 | 6000 | 6.0000000000000000 |  1000
+SELECT x, sum(y), avg(y), sum(x+y), count(*) FROM pagg_tab_para GROUP BY x HAVING avg(y) < 7 ORDER BY 1, 2, 3;
+ x  | sum  |        avg         |  sum  | count 
+----+------+--------------------+-------+-------
+  0 | 5000 | 5.0000000000000000 |  5000 |  1000
+  1 | 6000 | 6.0000000000000000 |  7000 |  1000
+ 10 | 5000 | 5.0000000000000000 | 15000 |  1000
+ 11 | 6000 | 6.0000000000000000 | 17000 |  1000
+ 20 | 5000 | 5.0000000000000000 | 25000 |  1000
+ 21 | 6000 | 6.0000000000000000 | 27000 |  1000
 (6 rows)
 
 ALTER TABLE pagg_tab_para_p2 SET (parallel_workers = 0);
 ANALYZE pagg_tab_para;
 EXPLAIN (COSTS OFF)
-SELECT x, sum(y), avg(y), count(*) FROM pagg_tab_para GROUP BY x HAVING avg(y) < 7 ORDER BY 1, 2, 3;
+SELECT x, sum(y), avg(y), sum(x+y), count(*) FROM pagg_tab_para GROUP BY x HAVING avg(y) < 7 ORDER BY 1, 2, 3;
                                     QUERY PLAN                                    
 ----------------------------------------------------------------------------------
  Sort
@@ -1470,15 +1471,15 @@ SELECT x, sum(y), avg(y), count(*) FROM pagg_tab_para GROUP BY x HAVING avg(y) <
                                  ->  Seq Scan on pagg_tab_para_p3 pagg_tab_para_3
 (15 rows)
 
-SELECT x, sum(y), avg(y), count(*) FROM pagg_tab_para GROUP BY x HAVING avg(y) < 7 ORDER BY 1, 2, 3;
- x  | sum  |        avg         | count 
-----+------+--------------------+-------
-  0 | 5000 | 5.0000000000000000 |  1000
-  1 | 6000 | 6.0000000000000000 |  1000
- 10 | 5000 | 5.0000000000000000 |  1000
- 11 | 6000 | 6.0000000000000000 |  1000
- 20 | 5000 | 5.0000000000000000 |  1000
- 21 | 6000 | 6.0000000000000000 |  1000
+SELECT x, sum(y), avg(y), sum(x+y), count(*) FROM pagg_tab_para GROUP BY x HAVING avg(y) < 7 ORDER BY 1, 2, 3;
+ x  | sum  |        avg         |  sum  | count 
+----+------+--------------------+-------+-------
+  0 | 5000 | 5.0000000000000000 |  5000 |  1000
+  1 | 6000 | 6.0000000000000000 |  7000 |  1000
+ 10 | 5000 | 5.0000000000000000 | 15000 |  1000
+ 11 | 6000 | 6.0000000000000000 | 17000 |  1000
+ 20 | 5000 | 5.0000000000000000 | 25000 |  1000
+ 21 | 6000 | 6.0000000000000000 | 27000 |  1000
 (6 rows)
 
 -- Reset parallelism parameters to get partitionwise aggregation plan.
index 117f65ecb4f6a16f3eda1d76c9402485547f4c59..c17294b15b01f0130f30b04d217675d64f0c3b84 100644 (file)
@@ -308,20 +308,21 @@ SELECT y, sum(x), avg(x), count(*) FROM pagg_tab_para GROUP BY y HAVING avg(x) <
 SELECT y, sum(x), avg(x), count(*) FROM pagg_tab_para GROUP BY y HAVING avg(x) < 12 ORDER BY 1, 2, 3;
 
 -- Test when parent can produce parallel paths but not any (or some) of its children
+-- (Use one more aggregate to tilt the cost estimates for the plan we want)
 ALTER TABLE pagg_tab_para_p1 SET (parallel_workers = 0);
 ALTER TABLE pagg_tab_para_p3 SET (parallel_workers = 0);
 ANALYZE pagg_tab_para;
 
 EXPLAIN (COSTS OFF)
-SELECT x, sum(y), avg(y), count(*) FROM pagg_tab_para GROUP BY x HAVING avg(y) < 7 ORDER BY 1, 2, 3;
-SELECT x, sum(y), avg(y), count(*) FROM pagg_tab_para GROUP BY x HAVING avg(y) < 7 ORDER BY 1, 2, 3;
+SELECT x, sum(y), avg(y), sum(x+y), count(*) FROM pagg_tab_para GROUP BY x HAVING avg(y) < 7 ORDER BY 1, 2, 3;
+SELECT x, sum(y), avg(y), sum(x+y), count(*) FROM pagg_tab_para GROUP BY x HAVING avg(y) < 7 ORDER BY 1, 2, 3;
 
 ALTER TABLE pagg_tab_para_p2 SET (parallel_workers = 0);
 ANALYZE pagg_tab_para;
 
 EXPLAIN (COSTS OFF)
-SELECT x, sum(y), avg(y), count(*) FROM pagg_tab_para GROUP BY x HAVING avg(y) < 7 ORDER BY 1, 2, 3;
-SELECT x, sum(y), avg(y), count(*) FROM pagg_tab_para GROUP BY x HAVING avg(y) < 7 ORDER BY 1, 2, 3;
+SELECT x, sum(y), avg(y), sum(x+y), count(*) FROM pagg_tab_para GROUP BY x HAVING avg(y) < 7 ORDER BY 1, 2, 3;
+SELECT x, sum(y), avg(y), sum(x+y), count(*) FROM pagg_tab_para GROUP BY x HAVING avg(y) < 7 ORDER BY 1, 2, 3;
 
 -- Reset parallelism parameters to get partitionwise aggregation plan.
 RESET min_parallel_table_scan_size;
index fde701bfd4d16059dc070f0e3809f89b515f7c49..4c40ae37b26a6c37a0057371c1d3707d9bd3f4ef 100644 (file)
@@ -50,7 +50,6 @@ AggStatePerPhase
 AggStatePerTrans
 AggStrategy
 Aggref
-AggrefExprState
 AlenState
 Alias
 AllocBlock