Line data Source code
1 : /*-------------------------------------------------------------------------
2 : *
3 : * nodeAgg.c
4 : * Routines to handle aggregate nodes.
5 : *
6 : * ExecAgg normally evaluates each aggregate in the following steps:
7 : *
8 : * transvalue = initcond
9 : * foreach input_tuple do
10 : * transvalue = transfunc(transvalue, input_value(s))
11 : * result = finalfunc(transvalue, direct_argument(s))
12 : *
13 : * If a finalfunc is not supplied then the result is just the ending
14 : * value of transvalue.
15 : *
16 : * Other behaviors can be selected by the "aggsplit" mode, which exists
17 : * to support partial aggregation. It is possible to:
18 : * * Skip running the finalfunc, so that the output is always the
19 : * final transvalue state.
20 : * * Substitute the combinefunc for the transfunc, so that transvalue
21 : * states (propagated up from a child partial-aggregation step) are merged
22 : * rather than processing raw input rows. (The statements below about
23 : * the transfunc apply equally to the combinefunc, when it's selected.)
24 : * * Apply the serializefunc to the output values (this only makes sense
25 : * when skipping the finalfunc, since the serializefunc works on the
26 : * transvalue data type).
27 : * * Apply the deserializefunc to the input values (this only makes sense
28 : * when using the combinefunc, for similar reasons).
29 : * It is the planner's responsibility to connect up Agg nodes using these
30 : * alternate behaviors in a way that makes sense, with partial aggregation
31 : * results being fed to nodes that expect them.
32 : *
33 : * If a normal aggregate call specifies DISTINCT or ORDER BY, we sort the
34 : * input tuples and eliminate duplicates (if required) before performing
35 : * the above-depicted process. (However, we don't do that for ordered-set
36 : * aggregates; their "ORDER BY" inputs are ordinary aggregate arguments
37 : * so far as this module is concerned.) Note that partial aggregation
38 : * is not supported in these cases, since we couldn't ensure global
39 : * ordering or distinctness of the inputs.
40 : *
41 : * If transfunc is marked "strict" in pg_proc and initcond is NULL,
42 : * then the first non-NULL input_value is assigned directly to transvalue,
43 : * and transfunc isn't applied until the second non-NULL input_value.
44 : * The agg's first input type and transtype must be the same in this case!
45 : *
46 : * If transfunc is marked "strict" then NULL input_values are skipped,
47 : * keeping the previous transvalue. If transfunc is not strict then it
48 : * is called for every input tuple and must deal with NULL initcond
49 : * or NULL input_values for itself.
50 : *
51 : * If finalfunc is marked "strict" then it is not called when the
52 : * ending transvalue is NULL, instead a NULL result is created
53 : * automatically (this is just the usual handling of strict functions,
54 : * of course). A non-strict finalfunc can make its own choice of
55 : * what to return for a NULL ending transvalue.
56 : *
57 : * Ordered-set aggregates are treated specially in one other way: we
58 : * evaluate any "direct" arguments and pass them to the finalfunc along
59 : * with the transition value.
60 : *
61 : * A finalfunc can have additional arguments beyond the transvalue and
62 : * any "direct" arguments, corresponding to the input arguments of the
63 : * aggregate. These are always just passed as NULL. Such arguments may be
64 : * needed to allow resolution of a polymorphic aggregate's result type.
65 : *
66 : * We compute aggregate input expressions and run the transition functions
67 : * in a temporary econtext (aggstate->tmpcontext). This is reset at least
68 : * once per input tuple, so when the transvalue datatype is
69 : * pass-by-reference, we have to be careful to copy it into a longer-lived
70 : * memory context, and free the prior value to avoid memory leakage. We
71 : * store transvalues in another set of econtexts, aggstate->aggcontexts
72 : * (one per grouping set, see below), which are also used for the hashtable
73 : * structures in AGG_HASHED mode. These econtexts are rescanned, not just
74 : * reset, at group boundaries so that aggregate transition functions can
75 : * register shutdown callbacks via AggRegisterCallback.
76 : *
77 : * The node's regular econtext (aggstate->ss.ps.ps_ExprContext) is used to
78 : * run finalize functions and compute the output tuple; this context can be
79 : * reset once per output tuple.
80 : *
81 : * The executor's AggState node is passed as the fmgr "context" value in
82 : * all transfunc and finalfunc calls. It is not recommended that the
83 : * transition functions look at the AggState node directly, but they can
84 : * use AggCheckCallContext() to verify that they are being called by
85 : * nodeAgg.c (and not as ordinary SQL functions). The main reason a
86 : * transition function might want to know this is so that it can avoid
87 : * palloc'ing a fixed-size pass-by-ref transition value on every call:
88 : * it can instead just scribble on and return its left input. Ordinarily
89 : * it is completely forbidden for functions to modify pass-by-ref inputs,
90 : * but in the aggregate case we know the left input is either the initial
91 : * transition value or a previous function result, and in either case its
92 : * value need not be preserved. See int8inc() for an example. Notice that
93 : * the EEOP_AGG_PLAIN_TRANS step is coded to avoid a data copy step when
94 : * the previous transition value pointer is returned. It is also possible
95 : * to avoid repeated data copying when the transition value is an expanded
96 : * object: to do that, the transition function must take care to return
97 : * an expanded object that is in a child context of the memory context
98 : * returned by AggCheckCallContext(). Also, some transition functions want
99 : * to store working state in addition to the nominal transition value; they
100 : * can use the memory context returned by AggCheckCallContext() to do that.
101 : *
102 : * Note: AggCheckCallContext() is available as of PostgreSQL 9.0. The
103 : * AggState is available as context in earlier releases (back to 8.1),
104 : * but direct examination of the node is needed to use it before 9.0.
105 : *
106 : * As of 9.4, aggregate transition functions can also use AggGetAggref()
107 : * to get hold of the Aggref expression node for their aggregate call.
108 : * This is mainly intended for ordered-set aggregates, which are not
109 : * supported as window functions. (A regular aggregate function would
110 : * need some fallback logic to use this, since there's no Aggref node
111 : * for a window function.)
112 : *
113 : * Grouping sets:
114 : *
115 : * A list of grouping sets which is structurally equivalent to a ROLLUP
116 : * clause (e.g. (a,b,c), (a,b), (a)) can be processed in a single pass over
117 : * ordered data. We do this by keeping a separate set of transition values
118 : * for each grouping set being concurrently processed; for each input tuple
119 : * we update them all, and on group boundaries we reset those states
120 : * (starting at the front of the list) whose grouping values have changed
121 : * (the list of grouping sets is ordered from most specific to least
122 : * specific).
123 : *
124 : * Where more complex grouping sets are used, we break them down into
125 : * "phases", where each phase has a different sort order (except phase 0
126 : * which is reserved for hashing). During each phase but the last, the
127 : * input tuples are additionally stored in a tuplesort which is keyed to the
128 : * next phase's sort order; during each phase but the first, the input
129 : * tuples are drawn from the previously sorted data. (The sorting of the
130 : * data for the first phase is handled by the planner, as it might be
131 : * satisfied by underlying nodes.)
132 : *
133 : * Hashing can be mixed with sorted grouping. To do this, we have an
134 : * AGG_MIXED strategy that populates the hashtables during the first sorted
135 : * phase, and switches to reading them out after completing all sort phases.
136 : * We can also support AGG_HASHED with multiple hash tables and no sorting
137 : * at all.
138 : *
139 : * From the perspective of aggregate transition and final functions, the
140 : * only issue regarding grouping sets is this: a single call site (flinfo)
141 : * of an aggregate function may be used for updating several different
142 : * transition values in turn. So the function must not cache in the flinfo
143 : * anything which logically belongs as part of the transition value (most
144 : * importantly, the memory context in which the transition value exists).
145 : * The support API functions (AggCheckCallContext, AggRegisterCallback) are
146 : * sensitive to the grouping set for which the aggregate function is
147 : * currently being called.
148 : *
149 : * Plan structure:
150 : *
151 : * What we get from the planner is actually one "real" Agg node which is
152 : * part of the plan tree proper, but which optionally has an additional list
153 : * of Agg nodes hung off the side via the "chain" field. This is because an
154 : * Agg node happens to be a convenient representation of all the data we
155 : * need for grouping sets.
156 : *
157 : * For many purposes, we treat the "real" node as if it were just the first
158 : * node in the chain. The chain must be ordered such that hashed entries
159 : * come before sorted/plain entries; the real node is marked AGG_MIXED if
160 : * there are both types present (in which case the real node describes one
161 : * of the hashed groupings, other AGG_HASHED nodes may optionally follow in
162 : * the chain, followed in turn by AGG_SORTED or (one) AGG_PLAIN node). If
163 : * the real node is marked AGG_HASHED or AGG_SORTED, then all the chained
164 : * nodes must be of the same type; if it is AGG_PLAIN, there can be no
165 : * chained nodes.
166 : *
167 : * We collect all hashed nodes into a single "phase", numbered 0, and create
168 : * a sorted phase (numbered 1..n) for each AGG_SORTED or AGG_PLAIN node.
169 : * Phase 0 is allocated even if there are no hashes, but remains unused in
170 : * that case.
171 : *
172 : * AGG_HASHED nodes actually refer to only a single grouping set each,
173 : * because for each hashed grouping we need a separate grpColIdx and
174 : * numGroups estimate. AGG_SORTED nodes represent a "rollup", a list of
175 : * grouping sets that share a sort order. Each AGG_SORTED node other than
176 : * the first one has an associated Sort node which describes the sort order
177 : * to be used; the first sorted node takes its input from the outer subtree,
178 : * which the planner has already arranged to provide ordered data.
179 : *
180 : * Memory and ExprContext usage:
181 : *
182 : * Because we're accumulating aggregate values across input rows, we need to
183 : * use more memory contexts than just simple input/output tuple contexts.
184 : * In fact, for a rollup, we need a separate context for each grouping set
185 : * so that we can reset the inner (finer-grained) aggregates on their group
186 : * boundaries while continuing to accumulate values for outer
187 : * (coarser-grained) groupings. On top of this, we might be simultaneously
188 : * populating hashtables; however, we only need one context for all the
189 : * hashtables.
190 : *
191 : * So we create an array, aggcontexts, with an ExprContext for each grouping
192 : * set in the largest rollup that we're going to process, and use the
193 : * per-tuple memory context of those ExprContexts to store the aggregate
194 : * transition values. hashcontext is the single context created to support
195 : * all hash tables.
196 : *
197 : * Spilling To Disk
198 : *
199 : * When performing hash aggregation, if the hash table memory exceeds the
200 : * limit (see hash_agg_check_limits()), we enter "spill mode". In spill
201 : * mode, we advance the transition states only for groups already in the
202 : * hash table. For tuples that would need to create a new hash table
203 : * entries (and initialize new transition states), we instead spill them to
204 : * disk to be processed later. The tuples are spilled in a partitioned
205 : * manner, so that subsequent batches are smaller and less likely to exceed
206 : * hash_mem (if a batch does exceed hash_mem, it must be spilled
207 : * recursively).
208 : *
209 : * Spilled data is written to logical tapes. These provide better control
210 : * over memory usage, disk space, and the number of files than if we were
211 : * to use a BufFile for each spill. We don't know the number of tapes needed
212 : * at the start of the algorithm (because it can recurse), so a tape set is
213 : * allocated at the beginning, and individual tapes are created as needed.
214 : * As a particular tape is read, logtape.c recycles its disk space. When a
215 : * tape is read to completion, it is destroyed entirely.
216 : *
217 : * Tapes' buffers can take up substantial memory when many tapes are open at
218 : * once. We only need one tape open at a time in read mode (using a buffer
219 : * that's a multiple of BLCKSZ); but we need one tape open in write mode (each
220 : * requiring a buffer of size BLCKSZ) for each partition.
221 : *
222 : * Note that it's possible for transition states to start small but then
223 : * grow very large; for instance in the case of ARRAY_AGG. In such cases,
224 : * it's still possible to significantly exceed hash_mem. We try to avoid
225 : * this situation by estimating what will fit in the available memory, and
226 : * imposing a limit on the number of groups separately from the amount of
227 : * memory consumed.
228 : *
229 : * Transition / Combine function invocation:
230 : *
231 : * For performance reasons transition functions, including combine
232 : * functions, aren't invoked one-by-one from nodeAgg.c after computing
233 : * arguments using the expression evaluation engine. Instead
234 : * ExecBuildAggTrans() builds one large expression that does both argument
235 : * evaluation and transition function invocation. That avoids performance
236 : * issues due to repeated uses of expression evaluation, complications due
237 : * to filter expressions having to be evaluated early, and allows to JIT
238 : * the entire expression into one native function.
239 : *
240 : * Portions Copyright (c) 1996-2024, PostgreSQL Global Development Group
241 : * Portions Copyright (c) 1994, Regents of the University of California
242 : *
243 : * IDENTIFICATION
244 : * src/backend/executor/nodeAgg.c
245 : *
246 : *-------------------------------------------------------------------------
247 : */
248 :
249 : #include "postgres.h"
250 :
251 : #include "access/htup_details.h"
252 : #include "access/parallel.h"
253 : #include "catalog/objectaccess.h"
254 : #include "catalog/pg_aggregate.h"
255 : #include "catalog/pg_proc.h"
256 : #include "catalog/pg_type.h"
257 : #include "common/hashfn.h"
258 : #include "executor/execExpr.h"
259 : #include "executor/executor.h"
260 : #include "executor/nodeAgg.h"
261 : #include "lib/hyperloglog.h"
262 : #include "miscadmin.h"
263 : #include "nodes/nodeFuncs.h"
264 : #include "optimizer/optimizer.h"
265 : #include "parser/parse_agg.h"
266 : #include "parser/parse_coerce.h"
267 : #include "utils/acl.h"
268 : #include "utils/builtins.h"
269 : #include "utils/datum.h"
270 : #include "utils/dynahash.h"
271 : #include "utils/expandeddatum.h"
272 : #include "utils/logtape.h"
273 : #include "utils/lsyscache.h"
274 : #include "utils/memutils.h"
275 : #include "utils/syscache.h"
276 : #include "utils/tuplesort.h"
277 :
278 : /*
279 : * Control how many partitions are created when spilling HashAgg to
280 : * disk.
281 : *
282 : * HASHAGG_PARTITION_FACTOR is multiplied by the estimated number of
283 : * partitions needed such that each partition will fit in memory. The factor
284 : * is set higher than one because there's not a high cost to having a few too
285 : * many partitions, and it makes it less likely that a partition will need to
286 : * be spilled recursively. Another benefit of having more, smaller partitions
287 : * is that small hash tables may perform better than large ones due to memory
288 : * caching effects.
289 : *
290 : * We also specify a min and max number of partitions per spill. Too few might
291 : * mean a lot of wasted I/O from repeated spilling of the same tuples. Too
292 : * many will result in lots of memory wasted buffering the spill files (which
293 : * could instead be spent on a larger hash table).
294 : */
295 : #define HASHAGG_PARTITION_FACTOR 1.50
296 : #define HASHAGG_MIN_PARTITIONS 4
297 : #define HASHAGG_MAX_PARTITIONS 1024
298 :
299 : /*
300 : * For reading from tapes, the buffer size must be a multiple of
301 : * BLCKSZ. Larger values help when reading from multiple tapes concurrently,
302 : * but that doesn't happen in HashAgg, so we simply use BLCKSZ. Writing to a
303 : * tape always uses a buffer of size BLCKSZ.
304 : */
305 : #define HASHAGG_READ_BUFFER_SIZE BLCKSZ
306 : #define HASHAGG_WRITE_BUFFER_SIZE BLCKSZ
307 :
308 : /*
309 : * HyperLogLog is used for estimating the cardinality of the spilled tuples in
310 : * a given partition. 5 bits corresponds to a size of about 32 bytes and a
311 : * worst-case error of around 18%. That's effective enough to choose a
312 : * reasonable number of partitions when recursing.
313 : */
314 : #define HASHAGG_HLL_BIT_WIDTH 5
315 :
316 : /*
317 : * Estimate chunk overhead as a constant 16 bytes. XXX: should this be
318 : * improved?
319 : */
320 : #define CHUNKHDRSZ 16
321 :
322 : /*
323 : * Represents partitioned spill data for a single hashtable. Contains the
324 : * necessary information to route tuples to the correct partition, and to
325 : * transform the spilled data into new batches.
326 : *
327 : * The high bits are used for partition selection (when recursing, we ignore
328 : * the bits that have already been used for partition selection at an earlier
329 : * level).
330 : */
331 : typedef struct HashAggSpill
332 : {
333 : int npartitions; /* number of partitions */
334 : LogicalTape **partitions; /* spill partition tapes */
335 : int64 *ntuples; /* number of tuples in each partition */
336 : uint32 mask; /* mask to find partition from hash value */
337 : int shift; /* after masking, shift by this amount */
338 : hyperLogLogState *hll_card; /* cardinality estimate for contents */
339 : } HashAggSpill;
340 :
341 : /*
342 : * Represents work to be done for one pass of hash aggregation (with only one
343 : * grouping set).
344 : *
345 : * Also tracks the bits of the hash already used for partition selection by
346 : * earlier iterations, so that this batch can use new bits. If all bits have
347 : * already been used, no partitioning will be done (any spilled data will go
348 : * to a single output tape).
349 : */
350 : typedef struct HashAggBatch
351 : {
352 : int setno; /* grouping set */
353 : int used_bits; /* number of bits of hash already used */
354 : LogicalTape *input_tape; /* input partition tape */
355 : int64 input_tuples; /* number of tuples in this batch */
356 : double input_card; /* estimated group cardinality */
357 : } HashAggBatch;
358 :
359 : /* used to find referenced colnos */
360 : typedef struct FindColsContext
361 : {
362 : bool is_aggref; /* is under an aggref */
363 : Bitmapset *aggregated; /* column references under an aggref */
364 : Bitmapset *unaggregated; /* other column references */
365 : } FindColsContext;
366 :
367 : static void select_current_set(AggState *aggstate, int setno, bool is_hash);
368 : static void initialize_phase(AggState *aggstate, int newphase);
369 : static TupleTableSlot *fetch_input_tuple(AggState *aggstate);
370 : static void initialize_aggregates(AggState *aggstate,
371 : AggStatePerGroup *pergroups,
372 : int numReset);
373 : static void advance_transition_function(AggState *aggstate,
374 : AggStatePerTrans pertrans,
375 : AggStatePerGroup pergroupstate);
376 : static void advance_aggregates(AggState *aggstate);
377 : static void process_ordered_aggregate_single(AggState *aggstate,
378 : AggStatePerTrans pertrans,
379 : AggStatePerGroup pergroupstate);
380 : static void process_ordered_aggregate_multi(AggState *aggstate,
381 : AggStatePerTrans pertrans,
382 : AggStatePerGroup pergroupstate);
383 : static void finalize_aggregate(AggState *aggstate,
384 : AggStatePerAgg peragg,
385 : AggStatePerGroup pergroupstate,
386 : Datum *resultVal, bool *resultIsNull);
387 : static void finalize_partialaggregate(AggState *aggstate,
388 : AggStatePerAgg peragg,
389 : AggStatePerGroup pergroupstate,
390 : Datum *resultVal, bool *resultIsNull);
391 : static inline void prepare_hash_slot(AggStatePerHash perhash,
392 : TupleTableSlot *inputslot,
393 : TupleTableSlot *hashslot);
394 : static void prepare_projection_slot(AggState *aggstate,
395 : TupleTableSlot *slot,
396 : int currentSet);
397 : static void finalize_aggregates(AggState *aggstate,
398 : AggStatePerAgg peraggs,
399 : AggStatePerGroup pergroup);
400 : static TupleTableSlot *project_aggregates(AggState *aggstate);
401 : static void find_cols(AggState *aggstate, Bitmapset **aggregated,
402 : Bitmapset **unaggregated);
403 : static bool find_cols_walker(Node *node, FindColsContext *context);
404 : static void build_hash_tables(AggState *aggstate);
405 : static void build_hash_table(AggState *aggstate, int setno, long nbuckets);
406 : static void hashagg_recompile_expressions(AggState *aggstate, bool minslot,
407 : bool nullcheck);
408 : static long hash_choose_num_buckets(double hashentrysize,
409 : long ngroups, Size memory);
410 : static int hash_choose_num_partitions(double input_groups,
411 : double hashentrysize,
412 : int used_bits,
413 : int *log2_npartitions);
414 : static void initialize_hash_entry(AggState *aggstate,
415 : TupleHashTable hashtable,
416 : TupleHashEntry entry);
417 : static void lookup_hash_entries(AggState *aggstate);
418 : static TupleTableSlot *agg_retrieve_direct(AggState *aggstate);
419 : static void agg_fill_hash_table(AggState *aggstate);
420 : static bool agg_refill_hash_table(AggState *aggstate);
421 : static TupleTableSlot *agg_retrieve_hash_table(AggState *aggstate);
422 : static TupleTableSlot *agg_retrieve_hash_table_in_memory(AggState *aggstate);
423 : static void hash_agg_check_limits(AggState *aggstate);
424 : static void hash_agg_enter_spill_mode(AggState *aggstate);
425 : static void hash_agg_update_metrics(AggState *aggstate, bool from_tape,
426 : int npartitions);
427 : static void hashagg_finish_initial_spills(AggState *aggstate);
428 : static void hashagg_reset_spill_state(AggState *aggstate);
429 : static HashAggBatch *hashagg_batch_new(LogicalTape *input_tape, int setno,
430 : int64 input_tuples, double input_card,
431 : int used_bits);
432 : static MinimalTuple hashagg_batch_read(HashAggBatch *batch, uint32 *hashp);
433 : static void hashagg_spill_init(HashAggSpill *spill, LogicalTapeSet *tapeset,
434 : int used_bits, double input_groups,
435 : double hashentrysize);
436 : static Size hashagg_spill_tuple(AggState *aggstate, HashAggSpill *spill,
437 : TupleTableSlot *inputslot, uint32 hash);
438 : static void hashagg_spill_finish(AggState *aggstate, HashAggSpill *spill,
439 : int setno);
440 : static Datum GetAggInitVal(Datum textInitVal, Oid transtype);
441 : static void build_pertrans_for_aggref(AggStatePerTrans pertrans,
442 : AggState *aggstate, EState *estate,
443 : Aggref *aggref, Oid transfn_oid,
444 : Oid aggtranstype, Oid aggserialfn,
445 : Oid aggdeserialfn, Datum initValue,
446 : bool initValueIsNull, Oid *inputTypes,
447 : int numArguments);
448 :
449 :
450 : /*
451 : * Select the current grouping set; affects current_set and
452 : * curaggcontext.
453 : */
454 : static void
455 6322870 : select_current_set(AggState *aggstate, int setno, bool is_hash)
456 : {
457 : /*
458 : * When changing this, also adapt ExecAggPlainTransByVal() and
459 : * ExecAggPlainTransByRef().
460 : */
461 6322870 : if (is_hash)
462 5674104 : aggstate->curaggcontext = aggstate->hashcontext;
463 : else
464 648766 : aggstate->curaggcontext = aggstate->aggcontexts[setno];
465 :
466 6322870 : aggstate->current_set = setno;
467 6322870 : }
468 :
469 : /*
470 : * Switch to phase "newphase", which must either be 0 or 1 (to reset) or
471 : * current_phase + 1. Juggle the tuplesorts accordingly.
472 : *
473 : * Phase 0 is for hashing, which we currently handle last in the AGG_MIXED
474 : * case, so when entering phase 0, all we need to do is drop open sorts.
475 : */
476 : static void
477 77458 : initialize_phase(AggState *aggstate, int newphase)
478 : {
479 : Assert(newphase <= 1 || newphase == aggstate->current_phase + 1);
480 :
481 : /*
482 : * Whatever the previous state, we're now done with whatever input
483 : * tuplesort was in use.
484 : */
485 77458 : if (aggstate->sort_in)
486 : {
487 42 : tuplesort_end(aggstate->sort_in);
488 42 : aggstate->sort_in = NULL;
489 : }
490 :
491 77458 : if (newphase <= 1)
492 : {
493 : /*
494 : * Discard any existing output tuplesort.
495 : */
496 77272 : if (aggstate->sort_out)
497 : {
498 6 : tuplesort_end(aggstate->sort_out);
499 6 : aggstate->sort_out = NULL;
500 : }
501 : }
502 : else
503 : {
504 : /*
505 : * The old output tuplesort becomes the new input one, and this is the
506 : * right time to actually sort it.
507 : */
508 186 : aggstate->sort_in = aggstate->sort_out;
509 186 : aggstate->sort_out = NULL;
510 : Assert(aggstate->sort_in);
511 186 : tuplesort_performsort(aggstate->sort_in);
512 : }
513 :
514 : /*
515 : * If this isn't the last phase, we need to sort appropriately for the
516 : * next phase in sequence.
517 : */
518 77458 : if (newphase > 0 && newphase < aggstate->numphases - 1)
519 : {
520 234 : Sort *sortnode = aggstate->phases[newphase + 1].sortnode;
521 234 : PlanState *outerNode = outerPlanState(aggstate);
522 234 : TupleDesc tupDesc = ExecGetResultType(outerNode);
523 :
524 234 : aggstate->sort_out = tuplesort_begin_heap(tupDesc,
525 : sortnode->numCols,
526 : sortnode->sortColIdx,
527 : sortnode->sortOperators,
528 : sortnode->collations,
529 : sortnode->nullsFirst,
530 : work_mem,
531 : NULL, TUPLESORT_NONE);
532 : }
533 :
534 77458 : aggstate->current_phase = newphase;
535 77458 : aggstate->phase = &aggstate->phases[newphase];
536 77458 : }
537 :
538 : /*
539 : * Fetch a tuple from either the outer plan (for phase 1) or from the sorter
540 : * populated by the previous phase. Copy it to the sorter for the next phase
541 : * if any.
542 : *
543 : * Callers cannot rely on memory for tuple in returned slot remaining valid
544 : * past any subsequently fetched tuple.
545 : */
546 : static TupleTableSlot *
547 25074470 : fetch_input_tuple(AggState *aggstate)
548 : {
549 : TupleTableSlot *slot;
550 :
551 25074470 : if (aggstate->sort_in)
552 : {
553 : /* make sure we check for interrupts in either path through here */
554 174882 : CHECK_FOR_INTERRUPTS();
555 174882 : if (!tuplesort_gettupleslot(aggstate->sort_in, true, false,
556 : aggstate->sort_slot, NULL))
557 186 : return NULL;
558 174696 : slot = aggstate->sort_slot;
559 : }
560 : else
561 24899588 : slot = ExecProcNode(outerPlanState(aggstate));
562 :
563 25074266 : if (!TupIsNull(slot) && aggstate->sort_out)
564 174696 : tuplesort_puttupleslot(aggstate->sort_out, slot);
565 :
566 25074266 : return slot;
567 : }
568 :
569 : /*
570 : * (Re)Initialize an individual aggregate.
571 : *
572 : * This function handles only one grouping set, already set in
573 : * aggstate->current_set.
574 : *
575 : * When called, CurrentMemoryContext should be the per-query context.
576 : */
577 : static void
578 1054536 : initialize_aggregate(AggState *aggstate, AggStatePerTrans pertrans,
579 : AggStatePerGroup pergroupstate)
580 : {
581 : /*
582 : * Start a fresh sort operation for each DISTINCT/ORDER BY aggregate.
583 : */
584 1054536 : if (pertrans->aggsortrequired)
585 : {
586 : /*
587 : * In case of rescan, maybe there could be an uncompleted sort
588 : * operation? Clean it up if so.
589 : */
590 53828 : if (pertrans->sortstates[aggstate->current_set])
591 0 : tuplesort_end(pertrans->sortstates[aggstate->current_set]);
592 :
593 :
594 : /*
595 : * We use a plain Datum sorter when there's a single input column;
596 : * otherwise sort the full tuple. (See comments for
597 : * process_ordered_aggregate_single.)
598 : */
599 53828 : if (pertrans->numInputs == 1)
600 : {
601 53756 : Form_pg_attribute attr = TupleDescAttr(pertrans->sortdesc, 0);
602 :
603 53756 : pertrans->sortstates[aggstate->current_set] =
604 53756 : tuplesort_begin_datum(attr->atttypid,
605 53756 : pertrans->sortOperators[0],
606 53756 : pertrans->sortCollations[0],
607 53756 : pertrans->sortNullsFirst[0],
608 : work_mem, NULL, TUPLESORT_NONE);
609 : }
610 : else
611 72 : pertrans->sortstates[aggstate->current_set] =
612 72 : tuplesort_begin_heap(pertrans->sortdesc,
613 : pertrans->numSortCols,
614 : pertrans->sortColIdx,
615 : pertrans->sortOperators,
616 : pertrans->sortCollations,
617 : pertrans->sortNullsFirst,
618 : work_mem, NULL, TUPLESORT_NONE);
619 : }
620 :
621 : /*
622 : * (Re)set transValue to the initial value.
623 : *
624 : * Note that when the initial value is pass-by-ref, we must copy it (into
625 : * the aggcontext) since we will pfree the transValue later.
626 : */
627 1054536 : if (pertrans->initValueIsNull)
628 525594 : pergroupstate->transValue = pertrans->initValue;
629 : else
630 : {
631 : MemoryContext oldContext;
632 :
633 528942 : oldContext = MemoryContextSwitchTo(aggstate->curaggcontext->ecxt_per_tuple_memory);
634 1057884 : pergroupstate->transValue = datumCopy(pertrans->initValue,
635 528942 : pertrans->transtypeByVal,
636 528942 : pertrans->transtypeLen);
637 528942 : MemoryContextSwitchTo(oldContext);
638 : }
639 1054536 : pergroupstate->transValueIsNull = pertrans->initValueIsNull;
640 :
641 : /*
642 : * If the initial value for the transition state doesn't exist in the
643 : * pg_aggregate table then we will let the first non-NULL value returned
644 : * from the outer procNode become the initial value. (This is useful for
645 : * aggregates like max() and min().) The noTransValue flag signals that we
646 : * still need to do this.
647 : */
648 1054536 : pergroupstate->noTransValue = pertrans->initValueIsNull;
649 1054536 : }
650 :
651 : /*
652 : * Initialize all aggregate transition states for a new group of input values.
653 : *
654 : * If there are multiple grouping sets, we initialize only the first numReset
655 : * of them (the grouping sets are ordered so that the most specific one, which
656 : * is reset most often, is first). As a convenience, if numReset is 0, we
657 : * reinitialize all sets.
658 : *
659 : * NB: This cannot be used for hash aggregates, as for those the grouping set
660 : * number has to be specified from further up.
661 : *
662 : * When called, CurrentMemoryContext should be the per-query context.
663 : */
664 : static void
665 292106 : initialize_aggregates(AggState *aggstate,
666 : AggStatePerGroup *pergroups,
667 : int numReset)
668 : {
669 : int transno;
670 292106 : int numGroupingSets = Max(aggstate->phase->numsets, 1);
671 292106 : int setno = 0;
672 292106 : int numTrans = aggstate->numtrans;
673 292106 : AggStatePerTrans transstates = aggstate->pertrans;
674 :
675 292106 : if (numReset == 0)
676 0 : numReset = numGroupingSets;
677 :
678 598368 : for (setno = 0; setno < numReset; setno++)
679 : {
680 306262 : AggStatePerGroup pergroup = pergroups[setno];
681 :
682 306262 : select_current_set(aggstate, setno, false);
683 :
684 964396 : for (transno = 0; transno < numTrans; transno++)
685 : {
686 658134 : AggStatePerTrans pertrans = &transstates[transno];
687 658134 : AggStatePerGroup pergroupstate = &pergroup[transno];
688 :
689 658134 : initialize_aggregate(aggstate, pertrans, pergroupstate);
690 : }
691 : }
692 292106 : }
693 :
694 : /*
695 : * Given new input value(s), advance the transition function of one aggregate
696 : * state within one grouping set only (already set in aggstate->current_set)
697 : *
698 : * The new values (and null flags) have been preloaded into argument positions
699 : * 1 and up in pertrans->transfn_fcinfo, so that we needn't copy them again to
700 : * pass to the transition function. We also expect that the static fields of
701 : * the fcinfo are already initialized; that was done by ExecInitAgg().
702 : *
703 : * It doesn't matter which memory context this is called in.
704 : */
705 : static void
706 724174 : advance_transition_function(AggState *aggstate,
707 : AggStatePerTrans pertrans,
708 : AggStatePerGroup pergroupstate)
709 : {
710 724174 : FunctionCallInfo fcinfo = pertrans->transfn_fcinfo;
711 : MemoryContext oldContext;
712 : Datum newVal;
713 :
714 724174 : if (pertrans->transfn.fn_strict)
715 : {
716 : /*
717 : * For a strict transfn, nothing happens when there's a NULL input; we
718 : * just keep the prior transValue.
719 : */
720 225000 : int numTransInputs = pertrans->numTransInputs;
721 : int i;
722 :
723 450000 : for (i = 1; i <= numTransInputs; i++)
724 : {
725 225000 : if (fcinfo->args[i].isnull)
726 0 : return;
727 : }
728 225000 : if (pergroupstate->noTransValue)
729 : {
730 : /*
731 : * transValue has not been initialized. This is the first non-NULL
732 : * input value. We use it as the initial value for transValue. (We
733 : * already checked that the agg's input type is binary-compatible
734 : * with its transtype, so straight copy here is OK.)
735 : *
736 : * We must copy the datum into aggcontext if it is pass-by-ref. We
737 : * do not need to pfree the old transValue, since it's NULL.
738 : */
739 0 : oldContext = MemoryContextSwitchTo(aggstate->curaggcontext->ecxt_per_tuple_memory);
740 0 : pergroupstate->transValue = datumCopy(fcinfo->args[1].value,
741 0 : pertrans->transtypeByVal,
742 0 : pertrans->transtypeLen);
743 0 : pergroupstate->transValueIsNull = false;
744 0 : pergroupstate->noTransValue = false;
745 0 : MemoryContextSwitchTo(oldContext);
746 0 : return;
747 : }
748 225000 : if (pergroupstate->transValueIsNull)
749 : {
750 : /*
751 : * Don't call a strict function with NULL inputs. Note it is
752 : * possible to get here despite the above tests, if the transfn is
753 : * strict *and* returned a NULL on a prior cycle. If that happens
754 : * we will propagate the NULL all the way to the end.
755 : */
756 0 : return;
757 : }
758 : }
759 :
760 : /* We run the transition functions in per-input-tuple memory context */
761 724174 : oldContext = MemoryContextSwitchTo(aggstate->tmpcontext->ecxt_per_tuple_memory);
762 :
763 : /* set up aggstate->curpertrans for AggGetAggref() */
764 724174 : aggstate->curpertrans = pertrans;
765 :
766 : /*
767 : * OK to call the transition function
768 : */
769 724174 : fcinfo->args[0].value = pergroupstate->transValue;
770 724174 : fcinfo->args[0].isnull = pergroupstate->transValueIsNull;
771 724174 : fcinfo->isnull = false; /* just in case transfn doesn't set it */
772 :
773 724174 : newVal = FunctionCallInvoke(fcinfo);
774 :
775 724174 : aggstate->curpertrans = NULL;
776 :
777 : /*
778 : * If pass-by-ref datatype, must copy the new value into aggcontext and
779 : * free the prior transValue. But if transfn returned a pointer to its
780 : * first input, we don't need to do anything.
781 : *
782 : * It's safe to compare newVal with pergroup->transValue without regard
783 : * for either being NULL, because ExecAggCopyTransValue takes care to set
784 : * transValue to 0 when NULL. Otherwise we could end up accidentally not
785 : * reparenting, when the transValue has the same numerical value as
786 : * newValue, despite being NULL. This is a somewhat hot path, making it
787 : * undesirable to instead solve this with another branch for the common
788 : * case of the transition function returning its (modified) input
789 : * argument.
790 : */
791 724174 : if (!pertrans->transtypeByVal &&
792 0 : DatumGetPointer(newVal) != DatumGetPointer(pergroupstate->transValue))
793 0 : newVal = ExecAggCopyTransValue(aggstate, pertrans,
794 0 : newVal, fcinfo->isnull,
795 : pergroupstate->transValue,
796 0 : pergroupstate->transValueIsNull);
797 :
798 724174 : pergroupstate->transValue = newVal;
799 724174 : pergroupstate->transValueIsNull = fcinfo->isnull;
800 :
801 724174 : MemoryContextSwitchTo(oldContext);
802 : }
803 :
804 : /*
805 : * Advance each aggregate transition state for one input tuple. The input
806 : * tuple has been stored in tmpcontext->ecxt_outertuple, so that it is
807 : * accessible to ExecEvalExpr.
808 : *
809 : * We have two sets of transition states to handle: one for sorted aggregation
810 : * and one for hashed; we do them both here, to avoid multiple evaluation of
811 : * the inputs.
812 : *
813 : * When called, CurrentMemoryContext should be the per-query context.
814 : */
815 : static void
816 25231690 : advance_aggregates(AggState *aggstate)
817 : {
818 : bool dummynull;
819 :
820 25231690 : ExecEvalExprSwitchContext(aggstate->phase->evaltrans,
821 : aggstate->tmpcontext,
822 : &dummynull);
823 25231624 : }
824 :
825 : /*
826 : * Run the transition function for a DISTINCT or ORDER BY aggregate
827 : * with only one input. This is called after we have completed
828 : * entering all the input values into the sort object. We complete the
829 : * sort, read out the values in sorted order, and run the transition
830 : * function on each value (applying DISTINCT if appropriate).
831 : *
832 : * Note that the strictness of the transition function was checked when
833 : * entering the values into the sort, so we don't check it again here;
834 : * we just apply standard SQL DISTINCT logic.
835 : *
836 : * The one-input case is handled separately from the multi-input case
837 : * for performance reasons: for single by-value inputs, such as the
838 : * common case of count(distinct id), the tuplesort_getdatum code path
839 : * is around 300% faster. (The speedup for by-reference types is less
840 : * but still noticeable.)
841 : *
842 : * This function handles only one grouping set (already set in
843 : * aggstate->current_set).
844 : *
845 : * When called, CurrentMemoryContext should be the per-query context.
846 : */
847 : static void
848 53756 : process_ordered_aggregate_single(AggState *aggstate,
849 : AggStatePerTrans pertrans,
850 : AggStatePerGroup pergroupstate)
851 : {
852 53756 : Datum oldVal = (Datum) 0;
853 53756 : bool oldIsNull = true;
854 53756 : bool haveOldVal = false;
855 53756 : MemoryContext workcontext = aggstate->tmpcontext->ecxt_per_tuple_memory;
856 : MemoryContext oldContext;
857 53756 : bool isDistinct = (pertrans->numDistinctCols > 0);
858 53756 : Datum newAbbrevVal = (Datum) 0;
859 53756 : Datum oldAbbrevVal = (Datum) 0;
860 53756 : FunctionCallInfo fcinfo = pertrans->transfn_fcinfo;
861 : Datum *newVal;
862 : bool *isNull;
863 :
864 : Assert(pertrans->numDistinctCols < 2);
865 :
866 53756 : tuplesort_performsort(pertrans->sortstates[aggstate->current_set]);
867 :
868 : /* Load the column into argument 1 (arg 0 will be transition value) */
869 53756 : newVal = &fcinfo->args[1].value;
870 53756 : isNull = &fcinfo->args[1].isnull;
871 :
872 : /*
873 : * Note: if input type is pass-by-ref, the datums returned by the sort are
874 : * freshly palloc'd in the per-query context, so we must be careful to
875 : * pfree them when they are no longer needed.
876 : */
877 :
878 898140 : while (tuplesort_getdatum(pertrans->sortstates[aggstate->current_set],
879 : true, false, newVal, isNull, &newAbbrevVal))
880 : {
881 : /*
882 : * Clear and select the working context for evaluation of the equality
883 : * function and transition function.
884 : */
885 844384 : MemoryContextReset(workcontext);
886 844384 : oldContext = MemoryContextSwitchTo(workcontext);
887 :
888 : /*
889 : * If DISTINCT mode, and not distinct from prior, skip it.
890 : */
891 844384 : if (isDistinct &&
892 310322 : haveOldVal &&
893 0 : ((oldIsNull && *isNull) ||
894 310322 : (!oldIsNull && !*isNull &&
895 605684 : oldAbbrevVal == newAbbrevVal &&
896 295362 : DatumGetBool(FunctionCall2Coll(&pertrans->equalfnOne,
897 : pertrans->aggCollation,
898 : oldVal, *newVal)))))
899 : {
900 120390 : MemoryContextSwitchTo(oldContext);
901 120390 : continue;
902 : }
903 : else
904 : {
905 723994 : advance_transition_function(aggstate, pertrans, pergroupstate);
906 :
907 723994 : MemoryContextSwitchTo(oldContext);
908 :
909 : /*
910 : * Forget the old value, if any, and remember the new one for
911 : * subsequent equality checks.
912 : */
913 723994 : if (!pertrans->inputtypeByVal)
914 : {
915 525288 : if (!oldIsNull)
916 525108 : pfree(DatumGetPointer(oldVal));
917 525288 : if (!*isNull)
918 525228 : oldVal = datumCopy(*newVal, pertrans->inputtypeByVal,
919 525228 : pertrans->inputtypeLen);
920 : }
921 : else
922 198706 : oldVal = *newVal;
923 723994 : oldAbbrevVal = newAbbrevVal;
924 723994 : oldIsNull = *isNull;
925 723994 : haveOldVal = true;
926 : }
927 : }
928 :
929 53756 : if (!oldIsNull && !pertrans->inputtypeByVal)
930 120 : pfree(DatumGetPointer(oldVal));
931 :
932 53756 : tuplesort_end(pertrans->sortstates[aggstate->current_set]);
933 53756 : pertrans->sortstates[aggstate->current_set] = NULL;
934 53756 : }
935 :
936 : /*
937 : * Run the transition function for a DISTINCT or ORDER BY aggregate
938 : * with more than one input. This is called after we have completed
939 : * entering all the input values into the sort object. We complete the
940 : * sort, read out the values in sorted order, and run the transition
941 : * function on each value (applying DISTINCT if appropriate).
942 : *
943 : * This function handles only one grouping set (already set in
944 : * aggstate->current_set).
945 : *
946 : * When called, CurrentMemoryContext should be the per-query context.
947 : */
948 : static void
949 72 : process_ordered_aggregate_multi(AggState *aggstate,
950 : AggStatePerTrans pertrans,
951 : AggStatePerGroup pergroupstate)
952 : {
953 72 : ExprContext *tmpcontext = aggstate->tmpcontext;
954 72 : FunctionCallInfo fcinfo = pertrans->transfn_fcinfo;
955 72 : TupleTableSlot *slot1 = pertrans->sortslot;
956 72 : TupleTableSlot *slot2 = pertrans->uniqslot;
957 72 : int numTransInputs = pertrans->numTransInputs;
958 72 : int numDistinctCols = pertrans->numDistinctCols;
959 72 : Datum newAbbrevVal = (Datum) 0;
960 72 : Datum oldAbbrevVal = (Datum) 0;
961 72 : bool haveOldValue = false;
962 72 : TupleTableSlot *save = aggstate->tmpcontext->ecxt_outertuple;
963 : int i;
964 :
965 72 : tuplesort_performsort(pertrans->sortstates[aggstate->current_set]);
966 :
967 72 : ExecClearTuple(slot1);
968 72 : if (slot2)
969 0 : ExecClearTuple(slot2);
970 :
971 252 : while (tuplesort_gettupleslot(pertrans->sortstates[aggstate->current_set],
972 : true, true, slot1, &newAbbrevVal))
973 : {
974 180 : CHECK_FOR_INTERRUPTS();
975 :
976 180 : tmpcontext->ecxt_outertuple = slot1;
977 180 : tmpcontext->ecxt_innertuple = slot2;
978 :
979 180 : if (numDistinctCols == 0 ||
980 0 : !haveOldValue ||
981 0 : newAbbrevVal != oldAbbrevVal ||
982 0 : !ExecQual(pertrans->equalfnMulti, tmpcontext))
983 : {
984 : /*
985 : * Extract the first numTransInputs columns as datums to pass to
986 : * the transfn.
987 : */
988 180 : slot_getsomeattrs(slot1, numTransInputs);
989 :
990 : /* Load values into fcinfo */
991 : /* Start from 1, since the 0th arg will be the transition value */
992 540 : for (i = 0; i < numTransInputs; i++)
993 : {
994 360 : fcinfo->args[i + 1].value = slot1->tts_values[i];
995 360 : fcinfo->args[i + 1].isnull = slot1->tts_isnull[i];
996 : }
997 :
998 180 : advance_transition_function(aggstate, pertrans, pergroupstate);
999 :
1000 180 : if (numDistinctCols > 0)
1001 : {
1002 : /* swap the slot pointers to retain the current tuple */
1003 0 : TupleTableSlot *tmpslot = slot2;
1004 :
1005 0 : slot2 = slot1;
1006 0 : slot1 = tmpslot;
1007 : /* avoid ExecQual() calls by reusing abbreviated keys */
1008 0 : oldAbbrevVal = newAbbrevVal;
1009 0 : haveOldValue = true;
1010 : }
1011 : }
1012 :
1013 : /* Reset context each time */
1014 180 : ResetExprContext(tmpcontext);
1015 :
1016 180 : ExecClearTuple(slot1);
1017 : }
1018 :
1019 72 : if (slot2)
1020 0 : ExecClearTuple(slot2);
1021 :
1022 72 : tuplesort_end(pertrans->sortstates[aggstate->current_set]);
1023 72 : pertrans->sortstates[aggstate->current_set] = NULL;
1024 :
1025 : /* restore previous slot, potentially in use for grouping sets */
1026 72 : tmpcontext->ecxt_outertuple = save;
1027 72 : }
1028 :
1029 : /*
1030 : * Compute the final value of one aggregate.
1031 : *
1032 : * This function handles only one grouping set (already set in
1033 : * aggstate->current_set).
1034 : *
1035 : * The finalfn will be run, and the result delivered, in the
1036 : * output-tuple context; caller's CurrentMemoryContext does not matter.
1037 : * (But note that in some cases, such as when there is no finalfn, the
1038 : * result might be a pointer to or into the agg's transition value.)
1039 : *
1040 : * The finalfn uses the state as set in the transno. This also might be
1041 : * being used by another aggregate function, so it's important that we do
1042 : * nothing destructive here. Moreover, the aggregate's final value might
1043 : * get used in multiple places, so we mustn't return a R/W expanded datum.
1044 : */
1045 : static void
1046 1046476 : finalize_aggregate(AggState *aggstate,
1047 : AggStatePerAgg peragg,
1048 : AggStatePerGroup pergroupstate,
1049 : Datum *resultVal, bool *resultIsNull)
1050 : {
1051 1046476 : LOCAL_FCINFO(fcinfo, FUNC_MAX_ARGS);
1052 1046476 : bool anynull = false;
1053 : MemoryContext oldContext;
1054 : int i;
1055 : ListCell *lc;
1056 1046476 : AggStatePerTrans pertrans = &aggstate->pertrans[peragg->transno];
1057 :
1058 1046476 : oldContext = MemoryContextSwitchTo(aggstate->ss.ps.ps_ExprContext->ecxt_per_tuple_memory);
1059 :
1060 : /*
1061 : * Evaluate any direct arguments. We do this even if there's no finalfn
1062 : * (which is unlikely anyway), so that side-effects happen as expected.
1063 : * The direct arguments go into arg positions 1 and up, leaving position 0
1064 : * for the transition state value.
1065 : */
1066 1046476 : i = 1;
1067 1047450 : foreach(lc, peragg->aggdirectargs)
1068 : {
1069 974 : ExprState *expr = (ExprState *) lfirst(lc);
1070 :
1071 974 : fcinfo->args[i].value = ExecEvalExpr(expr,
1072 : aggstate->ss.ps.ps_ExprContext,
1073 : &fcinfo->args[i].isnull);
1074 974 : anynull |= fcinfo->args[i].isnull;
1075 974 : i++;
1076 : }
1077 :
1078 : /*
1079 : * Apply the agg's finalfn if one is provided, else return transValue.
1080 : */
1081 1046476 : if (OidIsValid(peragg->finalfn_oid))
1082 : {
1083 270624 : int numFinalArgs = peragg->numFinalArgs;
1084 :
1085 : /* set up aggstate->curperagg for AggGetAggref() */
1086 270624 : aggstate->curperagg = peragg;
1087 :
1088 270624 : InitFunctionCallInfoData(*fcinfo, &peragg->finalfn,
1089 : numFinalArgs,
1090 : pertrans->aggCollation,
1091 : (void *) aggstate, NULL);
1092 :
1093 : /* Fill in the transition state value */
1094 270624 : fcinfo->args[0].value =
1095 270624 : MakeExpandedObjectReadOnly(pergroupstate->transValue,
1096 : pergroupstate->transValueIsNull,
1097 : pertrans->transtypeLen);
1098 270624 : fcinfo->args[0].isnull = pergroupstate->transValueIsNull;
1099 270624 : anynull |= pergroupstate->transValueIsNull;
1100 :
1101 : /* Fill any remaining argument positions with nulls */
1102 356010 : for (; i < numFinalArgs; i++)
1103 : {
1104 85386 : fcinfo->args[i].value = (Datum) 0;
1105 85386 : fcinfo->args[i].isnull = true;
1106 85386 : anynull = true;
1107 : }
1108 :
1109 270624 : if (fcinfo->flinfo->fn_strict && anynull)
1110 : {
1111 : /* don't call a strict function with NULL inputs */
1112 0 : *resultVal = (Datum) 0;
1113 0 : *resultIsNull = true;
1114 : }
1115 : else
1116 : {
1117 : Datum result;
1118 :
1119 270624 : result = FunctionCallInvoke(fcinfo);
1120 270612 : *resultIsNull = fcinfo->isnull;
1121 270612 : *resultVal = MakeExpandedObjectReadOnly(result,
1122 : fcinfo->isnull,
1123 : peragg->resulttypeLen);
1124 : }
1125 270612 : aggstate->curperagg = NULL;
1126 : }
1127 : else
1128 : {
1129 775852 : *resultVal =
1130 775852 : MakeExpandedObjectReadOnly(pergroupstate->transValue,
1131 : pergroupstate->transValueIsNull,
1132 : pertrans->transtypeLen);
1133 775852 : *resultIsNull = pergroupstate->transValueIsNull;
1134 : }
1135 :
1136 1046464 : MemoryContextSwitchTo(oldContext);
1137 1046464 : }
1138 :
1139 : /*
1140 : * Compute the output value of one partial aggregate.
1141 : *
1142 : * The serialization function will be run, and the result delivered, in the
1143 : * output-tuple context; caller's CurrentMemoryContext does not matter.
1144 : */
1145 : static void
1146 11676 : finalize_partialaggregate(AggState *aggstate,
1147 : AggStatePerAgg peragg,
1148 : AggStatePerGroup pergroupstate,
1149 : Datum *resultVal, bool *resultIsNull)
1150 : {
1151 11676 : AggStatePerTrans pertrans = &aggstate->pertrans[peragg->transno];
1152 : MemoryContext oldContext;
1153 :
1154 11676 : oldContext = MemoryContextSwitchTo(aggstate->ss.ps.ps_ExprContext->ecxt_per_tuple_memory);
1155 :
1156 : /*
1157 : * serialfn_oid will be set if we must serialize the transvalue before
1158 : * returning it
1159 : */
1160 11676 : if (OidIsValid(pertrans->serialfn_oid))
1161 : {
1162 : /* Don't call a strict serialization function with NULL input. */
1163 426 : if (pertrans->serialfn.fn_strict && pergroupstate->transValueIsNull)
1164 : {
1165 124 : *resultVal = (Datum) 0;
1166 124 : *resultIsNull = true;
1167 : }
1168 : else
1169 : {
1170 302 : FunctionCallInfo fcinfo = pertrans->serialfn_fcinfo;
1171 : Datum result;
1172 :
1173 302 : fcinfo->args[0].value =
1174 302 : MakeExpandedObjectReadOnly(pergroupstate->transValue,
1175 : pergroupstate->transValueIsNull,
1176 : pertrans->transtypeLen);
1177 302 : fcinfo->args[0].isnull = pergroupstate->transValueIsNull;
1178 302 : fcinfo->isnull = false;
1179 :
1180 302 : result = FunctionCallInvoke(fcinfo);
1181 302 : *resultIsNull = fcinfo->isnull;
1182 302 : *resultVal = MakeExpandedObjectReadOnly(result,
1183 : fcinfo->isnull,
1184 : peragg->resulttypeLen);
1185 : }
1186 : }
1187 : else
1188 : {
1189 11250 : *resultVal =
1190 11250 : MakeExpandedObjectReadOnly(pergroupstate->transValue,
1191 : pergroupstate->transValueIsNull,
1192 : pertrans->transtypeLen);
1193 11250 : *resultIsNull = pergroupstate->transValueIsNull;
1194 : }
1195 :
1196 11676 : MemoryContextSwitchTo(oldContext);
1197 11676 : }
1198 :
1199 : /*
1200 : * Extract the attributes that make up the grouping key into the
1201 : * hashslot. This is necessary to compute the hash or perform a lookup.
1202 : */
1203 : static inline void
1204 6166166 : prepare_hash_slot(AggStatePerHash perhash,
1205 : TupleTableSlot *inputslot,
1206 : TupleTableSlot *hashslot)
1207 : {
1208 : int i;
1209 :
1210 : /* transfer just the needed columns into hashslot */
1211 6166166 : slot_getsomeattrs(inputslot, perhash->largestGrpColIdx);
1212 6166166 : ExecClearTuple(hashslot);
1213 :
1214 15307718 : for (i = 0; i < perhash->numhashGrpCols; i++)
1215 : {
1216 9141552 : int varNumber = perhash->hashGrpColIdxInput[i] - 1;
1217 :
1218 9141552 : hashslot->tts_values[i] = inputslot->tts_values[varNumber];
1219 9141552 : hashslot->tts_isnull[i] = inputslot->tts_isnull[varNumber];
1220 : }
1221 6166166 : ExecStoreVirtualTuple(hashslot);
1222 6166166 : }
1223 :
1224 : /*
1225 : * Prepare to finalize and project based on the specified representative tuple
1226 : * slot and grouping set.
1227 : *
1228 : * In the specified tuple slot, force to null all attributes that should be
1229 : * read as null in the context of the current grouping set. Also stash the
1230 : * current group bitmap where GroupingExpr can get at it.
1231 : *
1232 : * This relies on three conditions:
1233 : *
1234 : * 1) Nothing is ever going to try and extract the whole tuple from this slot,
1235 : * only reference it in evaluations, which will only access individual
1236 : * attributes.
1237 : *
1238 : * 2) No system columns are going to need to be nulled. (If a system column is
1239 : * referenced in a group clause, it is actually projected in the outer plan
1240 : * tlist.)
1241 : *
1242 : * 3) Within a given phase, we never need to recover the value of an attribute
1243 : * once it has been set to null.
1244 : *
1245 : * Poking into the slot this way is a bit ugly, but the consensus is that the
1246 : * alternative was worse.
1247 : */
1248 : static void
1249 855284 : prepare_projection_slot(AggState *aggstate, TupleTableSlot *slot, int currentSet)
1250 : {
1251 855284 : if (aggstate->phase->grouped_cols)
1252 : {
1253 578398 : Bitmapset *grouped_cols = aggstate->phase->grouped_cols[currentSet];
1254 :
1255 578398 : aggstate->grouped_cols = grouped_cols;
1256 :
1257 578398 : if (TTS_EMPTY(slot))
1258 : {
1259 : /*
1260 : * Force all values to be NULL if working on an empty input tuple
1261 : * (i.e. an empty grouping set for which no input rows were
1262 : * supplied).
1263 : */
1264 48 : ExecStoreAllNullTuple(slot);
1265 : }
1266 578350 : else if (aggstate->all_grouped_cols)
1267 : {
1268 : ListCell *lc;
1269 :
1270 : /* all_grouped_cols is arranged in desc order */
1271 578302 : slot_getsomeattrs(slot, linitial_int(aggstate->all_grouped_cols));
1272 :
1273 1699366 : foreach(lc, aggstate->all_grouped_cols)
1274 : {
1275 1121064 : int attnum = lfirst_int(lc);
1276 :
1277 1121064 : if (!bms_is_member(attnum, grouped_cols))
1278 57346 : slot->tts_isnull[attnum - 1] = true;
1279 : }
1280 : }
1281 : }
1282 855284 : }
1283 :
1284 : /*
1285 : * Compute the final value of all aggregates for one group.
1286 : *
1287 : * This function handles only one grouping set at a time, which the caller must
1288 : * have selected. It's also the caller's responsibility to adjust the supplied
1289 : * pergroup parameter to point to the current set's transvalues.
1290 : *
1291 : * Results are stored in the output econtext aggvalues/aggnulls.
1292 : */
1293 : static void
1294 855284 : finalize_aggregates(AggState *aggstate,
1295 : AggStatePerAgg peraggs,
1296 : AggStatePerGroup pergroup)
1297 : {
1298 855284 : ExprContext *econtext = aggstate->ss.ps.ps_ExprContext;
1299 855284 : Datum *aggvalues = econtext->ecxt_aggvalues;
1300 855284 : bool *aggnulls = econtext->ecxt_aggnulls;
1301 : int aggno;
1302 :
1303 : /*
1304 : * If there were any DISTINCT and/or ORDER BY aggregates, sort their
1305 : * inputs and run the transition functions.
1306 : */
1307 1913178 : for (int transno = 0; transno < aggstate->numtrans; transno++)
1308 : {
1309 1057894 : AggStatePerTrans pertrans = &aggstate->pertrans[transno];
1310 : AggStatePerGroup pergroupstate;
1311 :
1312 1057894 : pergroupstate = &pergroup[transno];
1313 :
1314 1057894 : if (pertrans->aggsortrequired)
1315 : {
1316 : Assert(aggstate->aggstrategy != AGG_HASHED &&
1317 : aggstate->aggstrategy != AGG_MIXED);
1318 :
1319 53828 : if (pertrans->numInputs == 1)
1320 53756 : process_ordered_aggregate_single(aggstate,
1321 : pertrans,
1322 : pergroupstate);
1323 : else
1324 72 : process_ordered_aggregate_multi(aggstate,
1325 : pertrans,
1326 : pergroupstate);
1327 : }
1328 1004066 : else if (pertrans->numDistinctCols > 0 && pertrans->haslast)
1329 : {
1330 18348 : pertrans->haslast = false;
1331 :
1332 18348 : if (pertrans->numDistinctCols == 1)
1333 : {
1334 18252 : if (!pertrans->inputtypeByVal && !pertrans->lastisnull)
1335 250 : pfree(DatumGetPointer(pertrans->lastdatum));
1336 :
1337 18252 : pertrans->lastisnull = false;
1338 18252 : pertrans->lastdatum = (Datum) 0;
1339 : }
1340 : else
1341 96 : ExecClearTuple(pertrans->uniqslot);
1342 : }
1343 : }
1344 :
1345 : /*
1346 : * Run the final functions.
1347 : */
1348 1913424 : for (aggno = 0; aggno < aggstate->numaggs; aggno++)
1349 : {
1350 1058152 : AggStatePerAgg peragg = &peraggs[aggno];
1351 1058152 : int transno = peragg->transno;
1352 : AggStatePerGroup pergroupstate;
1353 :
1354 1058152 : pergroupstate = &pergroup[transno];
1355 :
1356 1058152 : if (DO_AGGSPLIT_SKIPFINAL(aggstate->aggsplit))
1357 11676 : finalize_partialaggregate(aggstate, peragg, pergroupstate,
1358 11676 : &aggvalues[aggno], &aggnulls[aggno]);
1359 : else
1360 1046476 : finalize_aggregate(aggstate, peragg, pergroupstate,
1361 1046476 : &aggvalues[aggno], &aggnulls[aggno]);
1362 : }
1363 855272 : }
1364 :
1365 : /*
1366 : * Project the result of a group (whose aggs have already been calculated by
1367 : * finalize_aggregates). Returns the result slot, or NULL if no row is
1368 : * projected (suppressed by qual).
1369 : */
1370 : static TupleTableSlot *
1371 855272 : project_aggregates(AggState *aggstate)
1372 : {
1373 855272 : ExprContext *econtext = aggstate->ss.ps.ps_ExprContext;
1374 :
1375 : /*
1376 : * Check the qual (HAVING clause); if the group does not match, ignore it.
1377 : */
1378 855272 : if (ExecQual(aggstate->ss.ps.qual, econtext))
1379 : {
1380 : /*
1381 : * Form and return projection tuple using the aggregate results and
1382 : * the representative input tuple.
1383 : */
1384 748704 : return ExecProject(aggstate->ss.ps.ps_ProjInfo);
1385 : }
1386 : else
1387 106568 : InstrCountFiltered1(aggstate, 1);
1388 :
1389 106568 : return NULL;
1390 : }
1391 :
1392 : /*
1393 : * Find input-tuple columns that are needed, dividing them into
1394 : * aggregated and unaggregated sets.
1395 : */
1396 : static void
1397 5108 : find_cols(AggState *aggstate, Bitmapset **aggregated, Bitmapset **unaggregated)
1398 : {
1399 5108 : Agg *agg = (Agg *) aggstate->ss.ps.plan;
1400 : FindColsContext context;
1401 :
1402 5108 : context.is_aggref = false;
1403 5108 : context.aggregated = NULL;
1404 5108 : context.unaggregated = NULL;
1405 :
1406 : /* Examine tlist and quals */
1407 5108 : (void) find_cols_walker((Node *) agg->plan.targetlist, &context);
1408 5108 : (void) find_cols_walker((Node *) agg->plan.qual, &context);
1409 :
1410 : /* In some cases, grouping columns will not appear in the tlist */
1411 13148 : for (int i = 0; i < agg->numCols; i++)
1412 8040 : context.unaggregated = bms_add_member(context.unaggregated,
1413 8040 : agg->grpColIdx[i]);
1414 :
1415 5108 : *aggregated = context.aggregated;
1416 5108 : *unaggregated = context.unaggregated;
1417 5108 : }
1418 :
1419 : static bool
1420 63046 : find_cols_walker(Node *node, FindColsContext *context)
1421 : {
1422 63046 : if (node == NULL)
1423 11142 : return false;
1424 51904 : if (IsA(node, Var))
1425 : {
1426 13524 : Var *var = (Var *) node;
1427 :
1428 : /* setrefs.c should have set the varno to OUTER_VAR */
1429 : Assert(var->varno == OUTER_VAR);
1430 : Assert(var->varlevelsup == 0);
1431 13524 : if (context->is_aggref)
1432 4370 : context->aggregated = bms_add_member(context->aggregated,
1433 4370 : var->varattno);
1434 : else
1435 9154 : context->unaggregated = bms_add_member(context->unaggregated,
1436 9154 : var->varattno);
1437 13524 : return false;
1438 : }
1439 38380 : if (IsA(node, Aggref))
1440 : {
1441 : Assert(!context->is_aggref);
1442 6558 : context->is_aggref = true;
1443 6558 : expression_tree_walker(node, find_cols_walker, (void *) context);
1444 6558 : context->is_aggref = false;
1445 6558 : return false;
1446 : }
1447 31822 : return expression_tree_walker(node, find_cols_walker,
1448 : (void *) context);
1449 : }
1450 :
1451 : /*
1452 : * (Re-)initialize the hash table(s) to empty.
1453 : *
1454 : * To implement hashed aggregation, we need a hashtable that stores a
1455 : * representative tuple and an array of AggStatePerGroup structs for each
1456 : * distinct set of GROUP BY column values. We compute the hash key from the
1457 : * GROUP BY columns. The per-group data is allocated in lookup_hash_entry(),
1458 : * for each entry.
1459 : *
1460 : * We have a separate hashtable and associated perhash data structure for each
1461 : * grouping set for which we're doing hashing.
1462 : *
1463 : * The contents of the hash tables always live in the hashcontext's per-tuple
1464 : * memory context (there is only one of these for all tables together, since
1465 : * they are all reset at the same time).
1466 : */
1467 : static void
1468 4060 : build_hash_tables(AggState *aggstate)
1469 : {
1470 : int setno;
1471 :
1472 8404 : for (setno = 0; setno < aggstate->num_hashes; ++setno)
1473 : {
1474 4344 : AggStatePerHash perhash = &aggstate->perhash[setno];
1475 : long nbuckets;
1476 : Size memory;
1477 :
1478 4344 : if (perhash->hashtable != NULL)
1479 : {
1480 138 : ResetTupleHashTable(perhash->hashtable);
1481 138 : continue;
1482 : }
1483 :
1484 : Assert(perhash->aggnode->numGroups > 0);
1485 :
1486 4206 : memory = aggstate->hash_mem_limit / aggstate->num_hashes;
1487 :
1488 : /* choose reasonable number of buckets per hashtable */
1489 4206 : nbuckets = hash_choose_num_buckets(aggstate->hashentrysize,
1490 4206 : perhash->aggnode->numGroups,
1491 : memory);
1492 :
1493 4206 : build_hash_table(aggstate, setno, nbuckets);
1494 : }
1495 :
1496 4060 : aggstate->hash_ngroups_current = 0;
1497 4060 : }
1498 :
1499 : /*
1500 : * Build a single hashtable for this grouping set.
1501 : */
1502 : static void
1503 4206 : build_hash_table(AggState *aggstate, int setno, long nbuckets)
1504 : {
1505 4206 : AggStatePerHash perhash = &aggstate->perhash[setno];
1506 4206 : MemoryContext metacxt = aggstate->hash_metacxt;
1507 4206 : MemoryContext hashcxt = aggstate->hashcontext->ecxt_per_tuple_memory;
1508 4206 : MemoryContext tmpcxt = aggstate->tmpcontext->ecxt_per_tuple_memory;
1509 : Size additionalsize;
1510 :
1511 : Assert(aggstate->aggstrategy == AGG_HASHED ||
1512 : aggstate->aggstrategy == AGG_MIXED);
1513 :
1514 : /*
1515 : * Used to make sure initial hash table allocation does not exceed
1516 : * hash_mem. Note that the estimate does not include space for
1517 : * pass-by-reference transition data values, nor for the representative
1518 : * tuple of each group.
1519 : */
1520 4206 : additionalsize = aggstate->numtrans * sizeof(AggStatePerGroupData);
1521 :
1522 8412 : perhash->hashtable = BuildTupleHashTableExt(&aggstate->ss.ps,
1523 4206 : perhash->hashslot->tts_tupleDescriptor,
1524 : perhash->numCols,
1525 : perhash->hashGrpColIdxHash,
1526 4206 : perhash->eqfuncoids,
1527 : perhash->hashfunctions,
1528 4206 : perhash->aggnode->grpCollations,
1529 : nbuckets,
1530 : additionalsize,
1531 : metacxt,
1532 : hashcxt,
1533 : tmpcxt,
1534 4206 : DO_AGGSPLIT_SKIPFINAL(aggstate->aggsplit));
1535 4206 : }
1536 :
1537 : /*
1538 : * Compute columns that actually need to be stored in hashtable entries. The
1539 : * incoming tuples from the child plan node will contain grouping columns,
1540 : * other columns referenced in our targetlist and qual, columns used to
1541 : * compute the aggregate functions, and perhaps just junk columns we don't use
1542 : * at all. Only columns of the first two types need to be stored in the
1543 : * hashtable, and getting rid of the others can make the table entries
1544 : * significantly smaller. The hashtable only contains the relevant columns,
1545 : * and is packed/unpacked in lookup_hash_entry() / agg_retrieve_hash_table()
1546 : * into the format of the normal input descriptor.
1547 : *
1548 : * Additional columns, in addition to the columns grouped by, come from two
1549 : * sources: Firstly functionally dependent columns that we don't need to group
1550 : * by themselves, and secondly ctids for row-marks.
1551 : *
1552 : * To eliminate duplicates, we build a bitmapset of the needed columns, and
1553 : * then build an array of the columns included in the hashtable. We might
1554 : * still have duplicates if the passed-in grpColIdx has them, which can happen
1555 : * in edge cases from semijoins/distinct; these can't always be removed,
1556 : * because it's not certain that the duplicate cols will be using the same
1557 : * hash function.
1558 : *
1559 : * Note that the array is preserved over ExecReScanAgg, so we allocate it in
1560 : * the per-query context (unlike the hash table itself).
1561 : */
1562 : static void
1563 5108 : find_hash_columns(AggState *aggstate)
1564 : {
1565 : Bitmapset *base_colnos;
1566 : Bitmapset *aggregated_colnos;
1567 5108 : TupleDesc scanDesc = aggstate->ss.ss_ScanTupleSlot->tts_tupleDescriptor;
1568 5108 : List *outerTlist = outerPlanState(aggstate)->plan->targetlist;
1569 5108 : int numHashes = aggstate->num_hashes;
1570 5108 : EState *estate = aggstate->ss.ps.state;
1571 : int j;
1572 :
1573 : /* Find Vars that will be needed in tlist and qual */
1574 5108 : find_cols(aggstate, &aggregated_colnos, &base_colnos);
1575 5108 : aggstate->colnos_needed = bms_union(base_colnos, aggregated_colnos);
1576 5108 : aggstate->max_colno_needed = 0;
1577 5108 : aggstate->all_cols_needed = true;
1578 :
1579 21506 : for (int i = 0; i < scanDesc->natts; i++)
1580 : {
1581 16398 : int colno = i + 1;
1582 :
1583 16398 : if (bms_is_member(colno, aggstate->colnos_needed))
1584 11258 : aggstate->max_colno_needed = colno;
1585 : else
1586 5140 : aggstate->all_cols_needed = false;
1587 : }
1588 :
1589 10610 : for (j = 0; j < numHashes; ++j)
1590 : {
1591 5502 : AggStatePerHash perhash = &aggstate->perhash[j];
1592 5502 : Bitmapset *colnos = bms_copy(base_colnos);
1593 5502 : AttrNumber *grpColIdx = perhash->aggnode->grpColIdx;
1594 5502 : List *hashTlist = NIL;
1595 : TupleDesc hashDesc;
1596 : int maxCols;
1597 : int i;
1598 :
1599 5502 : perhash->largestGrpColIdx = 0;
1600 :
1601 : /*
1602 : * If we're doing grouping sets, then some Vars might be referenced in
1603 : * tlist/qual for the benefit of other grouping sets, but not needed
1604 : * when hashing; i.e. prepare_projection_slot will null them out, so
1605 : * there'd be no point storing them. Use prepare_projection_slot's
1606 : * logic to determine which.
1607 : */
1608 5502 : if (aggstate->phases[0].grouped_cols)
1609 : {
1610 5502 : Bitmapset *grouped_cols = aggstate->phases[0].grouped_cols[j];
1611 : ListCell *lc;
1612 :
1613 15046 : foreach(lc, aggstate->all_grouped_cols)
1614 : {
1615 9544 : int attnum = lfirst_int(lc);
1616 :
1617 9544 : if (!bms_is_member(attnum, grouped_cols))
1618 1008 : colnos = bms_del_member(colnos, attnum);
1619 : }
1620 : }
1621 :
1622 : /*
1623 : * Compute maximum number of input columns accounting for possible
1624 : * duplications in the grpColIdx array, which can happen in some edge
1625 : * cases where HashAggregate was generated as part of a semijoin or a
1626 : * DISTINCT.
1627 : */
1628 5502 : maxCols = bms_num_members(colnos) + perhash->numCols;
1629 :
1630 5502 : perhash->hashGrpColIdxInput =
1631 5502 : palloc(maxCols * sizeof(AttrNumber));
1632 5502 : perhash->hashGrpColIdxHash =
1633 5502 : palloc(perhash->numCols * sizeof(AttrNumber));
1634 :
1635 : /* Add all the grouping columns to colnos */
1636 14044 : for (i = 0; i < perhash->numCols; i++)
1637 8542 : colnos = bms_add_member(colnos, grpColIdx[i]);
1638 :
1639 : /*
1640 : * First build mapping for columns directly hashed. These are the
1641 : * first, because they'll be accessed when computing hash values and
1642 : * comparing tuples for exact matches. We also build simple mapping
1643 : * for execGrouping, so it knows where to find the to-be-hashed /
1644 : * compared columns in the input.
1645 : */
1646 14044 : for (i = 0; i < perhash->numCols; i++)
1647 : {
1648 8542 : perhash->hashGrpColIdxInput[i] = grpColIdx[i];
1649 8542 : perhash->hashGrpColIdxHash[i] = i + 1;
1650 8542 : perhash->numhashGrpCols++;
1651 : /* delete already mapped columns */
1652 8542 : colnos = bms_del_member(colnos, grpColIdx[i]);
1653 : }
1654 :
1655 : /* and add the remaining columns */
1656 5502 : i = -1;
1657 6142 : while ((i = bms_next_member(colnos, i)) >= 0)
1658 : {
1659 640 : perhash->hashGrpColIdxInput[perhash->numhashGrpCols] = i;
1660 640 : perhash->numhashGrpCols++;
1661 : }
1662 :
1663 : /* and build a tuple descriptor for the hashtable */
1664 14684 : for (i = 0; i < perhash->numhashGrpCols; i++)
1665 : {
1666 9182 : int varNumber = perhash->hashGrpColIdxInput[i] - 1;
1667 :
1668 9182 : hashTlist = lappend(hashTlist, list_nth(outerTlist, varNumber));
1669 9182 : perhash->largestGrpColIdx =
1670 9182 : Max(varNumber + 1, perhash->largestGrpColIdx);
1671 : }
1672 :
1673 5502 : hashDesc = ExecTypeFromTL(hashTlist);
1674 :
1675 5502 : execTuplesHashPrepare(perhash->numCols,
1676 5502 : perhash->aggnode->grpOperators,
1677 : &perhash->eqfuncoids,
1678 : &perhash->hashfunctions);
1679 5502 : perhash->hashslot =
1680 5502 : ExecAllocTableSlot(&estate->es_tupleTable, hashDesc,
1681 : &TTSOpsMinimalTuple);
1682 :
1683 5502 : list_free(hashTlist);
1684 5502 : bms_free(colnos);
1685 : }
1686 :
1687 5108 : bms_free(base_colnos);
1688 5108 : }
1689 :
1690 : /*
1691 : * Estimate per-hash-table-entry overhead.
1692 : */
1693 : Size
1694 25082 : hash_agg_entry_size(int numTrans, Size tupleWidth, Size transitionSpace)
1695 : {
1696 : Size tupleChunkSize;
1697 : Size pergroupChunkSize;
1698 : Size transitionChunkSize;
1699 25082 : Size tupleSize = (MAXALIGN(SizeofMinimalTupleHeader) +
1700 : tupleWidth);
1701 25082 : Size pergroupSize = numTrans * sizeof(AggStatePerGroupData);
1702 :
1703 25082 : tupleChunkSize = CHUNKHDRSZ + tupleSize;
1704 :
1705 25082 : if (pergroupSize > 0)
1706 10322 : pergroupChunkSize = CHUNKHDRSZ + pergroupSize;
1707 : else
1708 14760 : pergroupChunkSize = 0;
1709 :
1710 25082 : if (transitionSpace > 0)
1711 4732 : transitionChunkSize = CHUNKHDRSZ + transitionSpace;
1712 : else
1713 20350 : transitionChunkSize = 0;
1714 :
1715 : return
1716 : sizeof(TupleHashEntryData) +
1717 25082 : tupleChunkSize +
1718 25082 : pergroupChunkSize +
1719 : transitionChunkSize;
1720 : }
1721 :
1722 : /*
1723 : * hashagg_recompile_expressions()
1724 : *
1725 : * Identifies the right phase, compiles the right expression given the
1726 : * arguments, and then sets phase->evalfunc to that expression.
1727 : *
1728 : * Different versions of the compiled expression are needed depending on
1729 : * whether hash aggregation has spilled or not, and whether it's reading from
1730 : * the outer plan or a tape. Before spilling to disk, the expression reads
1731 : * from the outer plan and does not need to perform a NULL check. After
1732 : * HashAgg begins to spill, new groups will not be created in the hash table,
1733 : * and the AggStatePerGroup array may be NULL; therefore we need to add a null
1734 : * pointer check to the expression. Then, when reading spilled data from a
1735 : * tape, we change the outer slot type to be a fixed minimal tuple slot.
1736 : *
1737 : * It would be wasteful to recompile every time, so cache the compiled
1738 : * expressions in the AggStatePerPhase, and reuse when appropriate.
1739 : */
1740 : static void
1741 53430 : hashagg_recompile_expressions(AggState *aggstate, bool minslot, bool nullcheck)
1742 : {
1743 : AggStatePerPhase phase;
1744 53430 : int i = minslot ? 1 : 0;
1745 53430 : int j = nullcheck ? 1 : 0;
1746 :
1747 : Assert(aggstate->aggstrategy == AGG_HASHED ||
1748 : aggstate->aggstrategy == AGG_MIXED);
1749 :
1750 53430 : if (aggstate->aggstrategy == AGG_HASHED)
1751 858 : phase = &aggstate->phases[0];
1752 : else /* AGG_MIXED */
1753 52572 : phase = &aggstate->phases[1];
1754 :
1755 53430 : if (phase->evaltrans_cache[i][j] == NULL)
1756 : {
1757 72 : const TupleTableSlotOps *outerops = aggstate->ss.ps.outerops;
1758 72 : bool outerfixed = aggstate->ss.ps.outeropsfixed;
1759 72 : bool dohash = true;
1760 72 : bool dosort = false;
1761 :
1762 : /*
1763 : * If minslot is true, that means we are processing a spilled batch
1764 : * (inside agg_refill_hash_table()), and we must not advance the
1765 : * sorted grouping sets.
1766 : */
1767 72 : if (aggstate->aggstrategy == AGG_MIXED && !minslot)
1768 12 : dosort = true;
1769 :
1770 : /* temporarily change the outerops while compiling the expression */
1771 72 : if (minslot)
1772 : {
1773 36 : aggstate->ss.ps.outerops = &TTSOpsMinimalTuple;
1774 36 : aggstate->ss.ps.outeropsfixed = true;
1775 : }
1776 :
1777 72 : phase->evaltrans_cache[i][j] = ExecBuildAggTrans(aggstate, phase,
1778 : dosort, dohash,
1779 : nullcheck);
1780 :
1781 : /* change back */
1782 72 : aggstate->ss.ps.outerops = outerops;
1783 72 : aggstate->ss.ps.outeropsfixed = outerfixed;
1784 : }
1785 :
1786 53430 : phase->evaltrans = phase->evaltrans_cache[i][j];
1787 53430 : }
1788 :
1789 : /*
1790 : * Set limits that trigger spilling to avoid exceeding hash_mem. Consider the
1791 : * number of partitions we expect to create (if we do spill).
1792 : *
1793 : * There are two limits: a memory limit, and also an ngroups limit. The
1794 : * ngroups limit becomes important when we expect transition values to grow
1795 : * substantially larger than the initial value.
1796 : */
1797 : void
1798 50014 : hash_agg_set_limits(double hashentrysize, double input_groups, int used_bits,
1799 : Size *mem_limit, uint64 *ngroups_limit,
1800 : int *num_partitions)
1801 : {
1802 : int npartitions;
1803 : Size partition_mem;
1804 50014 : Size hash_mem_limit = get_hash_memory_limit();
1805 :
1806 : /* if not expected to spill, use all of hash_mem */
1807 50014 : if (input_groups * hashentrysize <= hash_mem_limit)
1808 : {
1809 47590 : if (num_partitions != NULL)
1810 23014 : *num_partitions = 0;
1811 47590 : *mem_limit = hash_mem_limit;
1812 47590 : *ngroups_limit = hash_mem_limit / hashentrysize;
1813 47590 : return;
1814 : }
1815 :
1816 : /*
1817 : * Calculate expected memory requirements for spilling, which is the size
1818 : * of the buffers needed for all the tapes that need to be open at once.
1819 : * Then, subtract that from the memory available for holding hash tables.
1820 : */
1821 2424 : npartitions = hash_choose_num_partitions(input_groups,
1822 : hashentrysize,
1823 : used_bits,
1824 : NULL);
1825 2424 : if (num_partitions != NULL)
1826 90 : *num_partitions = npartitions;
1827 :
1828 2424 : partition_mem =
1829 2424 : HASHAGG_READ_BUFFER_SIZE +
1830 : HASHAGG_WRITE_BUFFER_SIZE * npartitions;
1831 :
1832 : /*
1833 : * Don't set the limit below 3/4 of hash_mem. In that case, we are at the
1834 : * minimum number of partitions, so we aren't going to dramatically exceed
1835 : * work mem anyway.
1836 : */
1837 2424 : if (hash_mem_limit > 4 * partition_mem)
1838 0 : *mem_limit = hash_mem_limit - partition_mem;
1839 : else
1840 2424 : *mem_limit = hash_mem_limit * 0.75;
1841 :
1842 2424 : if (*mem_limit > hashentrysize)
1843 2424 : *ngroups_limit = *mem_limit / hashentrysize;
1844 : else
1845 0 : *ngroups_limit = 1;
1846 : }
1847 :
1848 : /*
1849 : * hash_agg_check_limits
1850 : *
1851 : * After adding a new group to the hash table, check whether we need to enter
1852 : * spill mode. Allocations may happen without adding new groups (for instance,
1853 : * if the transition state size grows), so this check is imperfect.
1854 : */
1855 : static void
1856 423438 : hash_agg_check_limits(AggState *aggstate)
1857 : {
1858 423438 : uint64 ngroups = aggstate->hash_ngroups_current;
1859 423438 : Size meta_mem = MemoryContextMemAllocated(aggstate->hash_metacxt,
1860 : true);
1861 423438 : Size hashkey_mem = MemoryContextMemAllocated(aggstate->hashcontext->ecxt_per_tuple_memory,
1862 : true);
1863 :
1864 : /*
1865 : * Don't spill unless there's at least one group in the hash table so we
1866 : * can be sure to make progress even in edge cases.
1867 : */
1868 423438 : if (aggstate->hash_ngroups_current > 0 &&
1869 423438 : (meta_mem + hashkey_mem > aggstate->hash_mem_limit ||
1870 397026 : ngroups > aggstate->hash_ngroups_limit))
1871 : {
1872 26442 : hash_agg_enter_spill_mode(aggstate);
1873 : }
1874 423438 : }
1875 :
1876 : /*
1877 : * Enter "spill mode", meaning that no new groups are added to any of the hash
1878 : * tables. Tuples that would create a new group are instead spilled, and
1879 : * processed later.
1880 : */
1881 : static void
1882 26442 : hash_agg_enter_spill_mode(AggState *aggstate)
1883 : {
1884 26442 : aggstate->hash_spill_mode = true;
1885 26442 : hashagg_recompile_expressions(aggstate, aggstate->table_filled, true);
1886 :
1887 26442 : if (!aggstate->hash_ever_spilled)
1888 : {
1889 : Assert(aggstate->hash_tapeset == NULL);
1890 : Assert(aggstate->hash_spills == NULL);
1891 :
1892 54 : aggstate->hash_ever_spilled = true;
1893 :
1894 54 : aggstate->hash_tapeset = LogicalTapeSetCreate(true, NULL, -1);
1895 :
1896 54 : aggstate->hash_spills = palloc(sizeof(HashAggSpill) * aggstate->num_hashes);
1897 :
1898 168 : for (int setno = 0; setno < aggstate->num_hashes; setno++)
1899 : {
1900 114 : AggStatePerHash perhash = &aggstate->perhash[setno];
1901 114 : HashAggSpill *spill = &aggstate->hash_spills[setno];
1902 :
1903 114 : hashagg_spill_init(spill, aggstate->hash_tapeset, 0,
1904 114 : perhash->aggnode->numGroups,
1905 : aggstate->hashentrysize);
1906 : }
1907 : }
1908 26442 : }
1909 :
1910 : /*
1911 : * Update metrics after filling the hash table.
1912 : *
1913 : * If reading from the outer plan, from_tape should be false; if reading from
1914 : * another tape, from_tape should be true.
1915 : */
1916 : static void
1917 30694 : hash_agg_update_metrics(AggState *aggstate, bool from_tape, int npartitions)
1918 : {
1919 : Size meta_mem;
1920 : Size hashkey_mem;
1921 : Size buffer_mem;
1922 : Size total_mem;
1923 :
1924 30694 : if (aggstate->aggstrategy != AGG_MIXED &&
1925 4306 : aggstate->aggstrategy != AGG_HASHED)
1926 0 : return;
1927 :
1928 : /* memory for the hash table itself */
1929 30694 : meta_mem = MemoryContextMemAllocated(aggstate->hash_metacxt, true);
1930 :
1931 : /* memory for the group keys and transition states */
1932 30694 : hashkey_mem = MemoryContextMemAllocated(aggstate->hashcontext->ecxt_per_tuple_memory, true);
1933 :
1934 : /* memory for read/write tape buffers, if spilled */
1935 30694 : buffer_mem = npartitions * HASHAGG_WRITE_BUFFER_SIZE;
1936 30694 : if (from_tape)
1937 26910 : buffer_mem += HASHAGG_READ_BUFFER_SIZE;
1938 :
1939 : /* update peak mem */
1940 30694 : total_mem = meta_mem + hashkey_mem + buffer_mem;
1941 30694 : if (total_mem > aggstate->hash_mem_peak)
1942 3748 : aggstate->hash_mem_peak = total_mem;
1943 :
1944 : /* update disk usage */
1945 30694 : if (aggstate->hash_tapeset != NULL)
1946 : {
1947 26964 : uint64 disk_used = LogicalTapeSetBlocks(aggstate->hash_tapeset) * (BLCKSZ / 1024);
1948 :
1949 26964 : if (aggstate->hash_disk_used < disk_used)
1950 42 : aggstate->hash_disk_used = disk_used;
1951 : }
1952 :
1953 : /* update hashentrysize estimate based on contents */
1954 30694 : if (aggstate->hash_ngroups_current > 0)
1955 : {
1956 30358 : aggstate->hashentrysize =
1957 30358 : sizeof(TupleHashEntryData) +
1958 30358 : (hashkey_mem / (double) aggstate->hash_ngroups_current);
1959 : }
1960 : }
1961 :
1962 : /*
1963 : * Choose a reasonable number of buckets for the initial hash table size.
1964 : */
1965 : static long
1966 4206 : hash_choose_num_buckets(double hashentrysize, long ngroups, Size memory)
1967 : {
1968 : long max_nbuckets;
1969 4206 : long nbuckets = ngroups;
1970 :
1971 4206 : max_nbuckets = memory / hashentrysize;
1972 :
1973 : /*
1974 : * Underestimating is better than overestimating. Too many buckets crowd
1975 : * out space for group keys and transition state values.
1976 : */
1977 4206 : max_nbuckets >>= 1;
1978 :
1979 4206 : if (nbuckets > max_nbuckets)
1980 72 : nbuckets = max_nbuckets;
1981 :
1982 4206 : return Max(nbuckets, 1);
1983 : }
1984 :
1985 : /*
1986 : * Determine the number of partitions to create when spilling, which will
1987 : * always be a power of two. If log2_npartitions is non-NULL, set
1988 : * *log2_npartitions to the log2() of the number of partitions.
1989 : */
1990 : static int
1991 15030 : hash_choose_num_partitions(double input_groups, double hashentrysize,
1992 : int used_bits, int *log2_npartitions)
1993 : {
1994 15030 : Size hash_mem_limit = get_hash_memory_limit();
1995 : double partition_limit;
1996 : double mem_wanted;
1997 : double dpartitions;
1998 : int npartitions;
1999 : int partition_bits;
2000 :
2001 : /*
2002 : * Avoid creating so many partitions that the memory requirements of the
2003 : * open partition files are greater than 1/4 of hash_mem.
2004 : */
2005 15030 : partition_limit =
2006 15030 : (hash_mem_limit * 0.25 - HASHAGG_READ_BUFFER_SIZE) /
2007 : HASHAGG_WRITE_BUFFER_SIZE;
2008 :
2009 15030 : mem_wanted = HASHAGG_PARTITION_FACTOR * input_groups * hashentrysize;
2010 :
2011 : /* make enough partitions so that each one is likely to fit in memory */
2012 15030 : dpartitions = 1 + (mem_wanted / hash_mem_limit);
2013 :
2014 15030 : if (dpartitions > partition_limit)
2015 15012 : dpartitions = partition_limit;
2016 :
2017 15030 : if (dpartitions < HASHAGG_MIN_PARTITIONS)
2018 15030 : dpartitions = HASHAGG_MIN_PARTITIONS;
2019 15030 : if (dpartitions > HASHAGG_MAX_PARTITIONS)
2020 0 : dpartitions = HASHAGG_MAX_PARTITIONS;
2021 :
2022 : /* HASHAGG_MAX_PARTITIONS limit makes this safe */
2023 15030 : npartitions = (int) dpartitions;
2024 :
2025 : /* ceil(log2(npartitions)) */
2026 15030 : partition_bits = my_log2(npartitions);
2027 :
2028 : /* make sure that we don't exhaust the hash bits */
2029 15030 : if (partition_bits + used_bits >= 32)
2030 0 : partition_bits = 32 - used_bits;
2031 :
2032 15030 : if (log2_npartitions != NULL)
2033 12606 : *log2_npartitions = partition_bits;
2034 :
2035 : /* number of partitions will be a power of two */
2036 15030 : npartitions = 1 << partition_bits;
2037 :
2038 15030 : return npartitions;
2039 : }
2040 :
2041 : /*
2042 : * Initialize a freshly-created TupleHashEntry.
2043 : */
2044 : static void
2045 423438 : initialize_hash_entry(AggState *aggstate, TupleHashTable hashtable,
2046 : TupleHashEntry entry)
2047 : {
2048 : AggStatePerGroup pergroup;
2049 : int transno;
2050 :
2051 423438 : aggstate->hash_ngroups_current++;
2052 423438 : hash_agg_check_limits(aggstate);
2053 :
2054 : /* no need to allocate or initialize per-group state */
2055 423438 : if (aggstate->numtrans == 0)
2056 172566 : return;
2057 :
2058 : pergroup = (AggStatePerGroup)
2059 250872 : MemoryContextAlloc(hashtable->tablecxt,
2060 250872 : sizeof(AggStatePerGroupData) * aggstate->numtrans);
2061 :
2062 250872 : entry->additional = pergroup;
2063 :
2064 : /*
2065 : * Initialize aggregates for new tuple group, lookup_hash_entries()
2066 : * already has selected the relevant grouping set.
2067 : */
2068 647274 : for (transno = 0; transno < aggstate->numtrans; transno++)
2069 : {
2070 396402 : AggStatePerTrans pertrans = &aggstate->pertrans[transno];
2071 396402 : AggStatePerGroup pergroupstate = &pergroup[transno];
2072 :
2073 396402 : initialize_aggregate(aggstate, pertrans, pergroupstate);
2074 : }
2075 : }
2076 :
2077 : /*
2078 : * Look up hash entries for the current tuple in all hashed grouping sets.
2079 : *
2080 : * Be aware that lookup_hash_entry can reset the tmpcontext.
2081 : *
2082 : * Some entries may be left NULL if we are in "spill mode". The same tuple
2083 : * will belong to different groups for each grouping set, so may match a group
2084 : * already in memory for one set and match a group not in memory for another
2085 : * set. When in "spill mode", the tuple will be spilled for each grouping set
2086 : * where it doesn't match a group in memory.
2087 : *
2088 : * NB: It's possible to spill the same tuple for several different grouping
2089 : * sets. This may seem wasteful, but it's actually a trade-off: if we spill
2090 : * the tuple multiple times for multiple grouping sets, it can be partitioned
2091 : * for each grouping set, making the refilling of the hash table very
2092 : * efficient.
2093 : */
2094 : static void
2095 5376046 : lookup_hash_entries(AggState *aggstate)
2096 : {
2097 5376046 : AggStatePerGroup *pergroup = aggstate->hash_pergroup;
2098 5376046 : TupleTableSlot *outerslot = aggstate->tmpcontext->ecxt_outertuple;
2099 : int setno;
2100 :
2101 10886388 : for (setno = 0; setno < aggstate->num_hashes; setno++)
2102 : {
2103 5510342 : AggStatePerHash perhash = &aggstate->perhash[setno];
2104 5510342 : TupleHashTable hashtable = perhash->hashtable;
2105 5510342 : TupleTableSlot *hashslot = perhash->hashslot;
2106 : TupleHashEntry entry;
2107 : uint32 hash;
2108 5510342 : bool isnew = false;
2109 : bool *p_isnew;
2110 :
2111 : /* if hash table already spilled, don't create new entries */
2112 5510342 : p_isnew = aggstate->hash_spill_mode ? NULL : &isnew;
2113 :
2114 5510342 : select_current_set(aggstate, setno, true);
2115 5510342 : prepare_hash_slot(perhash,
2116 : outerslot,
2117 : hashslot);
2118 :
2119 5510342 : entry = LookupTupleHashEntry(hashtable, hashslot,
2120 : p_isnew, &hash);
2121 :
2122 5510342 : if (entry != NULL)
2123 : {
2124 5288462 : if (isnew)
2125 328944 : initialize_hash_entry(aggstate, hashtable, entry);
2126 5288462 : pergroup[setno] = entry->additional;
2127 : }
2128 : else
2129 : {
2130 221880 : HashAggSpill *spill = &aggstate->hash_spills[setno];
2131 221880 : TupleTableSlot *slot = aggstate->tmpcontext->ecxt_outertuple;
2132 :
2133 221880 : if (spill->partitions == NULL)
2134 0 : hashagg_spill_init(spill, aggstate->hash_tapeset, 0,
2135 0 : perhash->aggnode->numGroups,
2136 : aggstate->hashentrysize);
2137 :
2138 221880 : hashagg_spill_tuple(aggstate, spill, slot, hash);
2139 221880 : pergroup[setno] = NULL;
2140 : }
2141 : }
2142 5376046 : }
2143 :
2144 : /*
2145 : * ExecAgg -
2146 : *
2147 : * ExecAgg receives tuples from its outer subplan and aggregates over
2148 : * the appropriate attribute for each aggregate function use (Aggref
2149 : * node) appearing in the targetlist or qual of the node. The number
2150 : * of tuples to aggregate over depends on whether grouped or plain
2151 : * aggregation is selected. In grouped aggregation, we produce a result
2152 : * row for each group; in plain aggregation there's a single result row
2153 : * for the whole query. In either case, the value of each aggregate is
2154 : * stored in the expression context to be used when ExecProject evaluates
2155 : * the result tuple.
2156 : */
2157 : static TupleTableSlot *
2158 811500 : ExecAgg(PlanState *pstate)
2159 : {
2160 811500 : AggState *node = castNode(AggState, pstate);
2161 811500 : TupleTableSlot *result = NULL;
2162 :
2163 811500 : CHECK_FOR_INTERRUPTS();
2164 :
2165 811500 : if (!node->agg_done)
2166 : {
2167 : /* Dispatch based on strategy */
2168 754476 : switch (node->phase->aggstrategy)
2169 : {
2170 490722 : case AGG_HASHED:
2171 490722 : if (!node->table_filled)
2172 3658 : agg_fill_hash_table(node);
2173 : /* FALLTHROUGH */
2174 : case AGG_MIXED:
2175 518066 : result = agg_retrieve_hash_table(node);
2176 518066 : break;
2177 236410 : case AGG_PLAIN:
2178 : case AGG_SORTED:
2179 236410 : result = agg_retrieve_direct(node);
2180 236314 : break;
2181 : }
2182 :
2183 754380 : if (!TupIsNull(result))
2184 748704 : return result;
2185 : }
2186 :
2187 62700 : return NULL;
2188 : }
2189 :
2190 : /*
2191 : * ExecAgg for non-hashed case
2192 : */
2193 : static TupleTableSlot *
2194 236410 : agg_retrieve_direct(AggState *aggstate)
2195 : {
2196 236410 : Agg *node = aggstate->phase->aggnode;
2197 : ExprContext *econtext;
2198 : ExprContext *tmpcontext;
2199 : AggStatePerAgg peragg;
2200 : AggStatePerGroup *pergroups;
2201 : TupleTableSlot *outerslot;
2202 : TupleTableSlot *firstSlot;
2203 : TupleTableSlot *result;
2204 236410 : bool hasGroupingSets = aggstate->phase->numsets > 0;
2205 236410 : int numGroupingSets = Max(aggstate->phase->numsets, 1);
2206 : int currentSet;
2207 : int nextSetSize;
2208 : int numReset;
2209 : int i;
2210 :
2211 : /*
2212 : * get state info from node
2213 : *
2214 : * econtext is the per-output-tuple expression context
2215 : *
2216 : * tmpcontext is the per-input-tuple expression context
2217 : */
2218 236410 : econtext = aggstate->ss.ps.ps_ExprContext;
2219 236410 : tmpcontext = aggstate->tmpcontext;
2220 :
2221 236410 : peragg = aggstate->peragg;
2222 236410 : pergroups = aggstate->pergroups;
2223 236410 : firstSlot = aggstate->ss.ss_ScanTupleSlot;
2224 :
2225 : /*
2226 : * We loop retrieving groups until we find one matching
2227 : * aggstate->ss.ps.qual
2228 : *
2229 : * For grouping sets, we have the invariant that aggstate->projected_set
2230 : * is either -1 (initial call) or the index (starting from 0) in
2231 : * gset_lengths for the group we just completed (either by projecting a
2232 : * row or by discarding it in the qual).
2233 : */
2234 307364 : while (!aggstate->agg_done)
2235 : {
2236 : /*
2237 : * Clear the per-output-tuple context for each group, as well as
2238 : * aggcontext (which contains any pass-by-ref transvalues of the old
2239 : * group). Some aggregate functions store working state in child
2240 : * contexts; those now get reset automatically without us needing to
2241 : * do anything special.
2242 : *
2243 : * We use ReScanExprContext not just ResetExprContext because we want
2244 : * any registered shutdown callbacks to be called. That allows
2245 : * aggregate functions to ensure they've cleaned up any non-memory
2246 : * resources.
2247 : */
2248 306966 : ReScanExprContext(econtext);
2249 :
2250 : /*
2251 : * Determine how many grouping sets need to be reset at this boundary.
2252 : */
2253 306966 : if (aggstate->projected_set >= 0 &&
2254 246084 : aggstate->projected_set < numGroupingSets)
2255 246078 : numReset = aggstate->projected_set + 1;
2256 : else
2257 60888 : numReset = numGroupingSets;
2258 :
2259 : /*
2260 : * numReset can change on a phase boundary, but that's OK; we want to
2261 : * reset the contexts used in _this_ phase, and later, after possibly
2262 : * changing phase, initialize the right number of aggregates for the
2263 : * _new_ phase.
2264 : */
2265 :
2266 636174 : for (i = 0; i < numReset; i++)
2267 : {
2268 329208 : ReScanExprContext(aggstate->aggcontexts[i]);
2269 : }
2270 :
2271 : /*
2272 : * Check if input is complete and there are no more groups to project
2273 : * in this phase; move to next phase or mark as done.
2274 : */
2275 306966 : if (aggstate->input_done == true &&
2276 1524 : aggstate->projected_set >= (numGroupingSets - 1))
2277 : {
2278 720 : if (aggstate->current_phase < aggstate->numphases - 1)
2279 : {
2280 186 : initialize_phase(aggstate, aggstate->current_phase + 1);
2281 186 : aggstate->input_done = false;
2282 186 : aggstate->projected_set = -1;
2283 186 : numGroupingSets = Max(aggstate->phase->numsets, 1);
2284 186 : node = aggstate->phase->aggnode;
2285 186 : numReset = numGroupingSets;
2286 : }
2287 534 : else if (aggstate->aggstrategy == AGG_MIXED)
2288 : {
2289 : /*
2290 : * Mixed mode; we've output all the grouped stuff and have
2291 : * full hashtables, so switch to outputting those.
2292 : */
2293 138 : initialize_phase(aggstate, 0);
2294 138 : aggstate->table_filled = true;
2295 138 : ResetTupleHashIterator(aggstate->perhash[0].hashtable,
2296 : &aggstate->perhash[0].hashiter);
2297 138 : select_current_set(aggstate, 0, true);
2298 138 : return agg_retrieve_hash_table(aggstate);
2299 : }
2300 : else
2301 : {
2302 396 : aggstate->agg_done = true;
2303 396 : break;
2304 : }
2305 : }
2306 :
2307 : /*
2308 : * Get the number of columns in the next grouping set after the last
2309 : * projected one (if any). This is the number of columns to compare to
2310 : * see if we reached the boundary of that set too.
2311 : */
2312 306432 : if (aggstate->projected_set >= 0 &&
2313 245364 : aggstate->projected_set < (numGroupingSets - 1))
2314 27276 : nextSetSize = aggstate->phase->gset_lengths[aggstate->projected_set + 1];
2315 : else
2316 279156 : nextSetSize = 0;
2317 :
2318 : /*----------
2319 : * If a subgroup for the current grouping set is present, project it.
2320 : *
2321 : * We have a new group if:
2322 : * - we're out of input but haven't projected all grouping sets
2323 : * (checked above)
2324 : * OR
2325 : * - we already projected a row that wasn't from the last grouping
2326 : * set
2327 : * AND
2328 : * - the next grouping set has at least one grouping column (since
2329 : * empty grouping sets project only once input is exhausted)
2330 : * AND
2331 : * - the previous and pending rows differ on the grouping columns
2332 : * of the next grouping set
2333 : *----------
2334 : */
2335 306432 : tmpcontext->ecxt_innertuple = econtext->ecxt_outertuple;
2336 306432 : if (aggstate->input_done ||
2337 305628 : (node->aggstrategy != AGG_PLAIN &&
2338 246226 : aggstate->projected_set != -1 &&
2339 244560 : aggstate->projected_set < (numGroupingSets - 1) &&
2340 19940 : nextSetSize > 0 &&
2341 19940 : !ExecQualAndReset(aggstate->phase->eqfunctions[nextSetSize - 1],
2342 : tmpcontext)))
2343 : {
2344 14138 : aggstate->projected_set += 1;
2345 :
2346 : Assert(aggstate->projected_set < numGroupingSets);
2347 14138 : Assert(nextSetSize > 0 || aggstate->input_done);
2348 : }
2349 : else
2350 : {
2351 : /*
2352 : * We no longer care what group we just projected, the next
2353 : * projection will always be the first (or only) grouping set
2354 : * (unless the input proves to be empty).
2355 : */
2356 292294 : aggstate->projected_set = 0;
2357 :
2358 : /*
2359 : * If we don't already have the first tuple of the new group,
2360 : * fetch it from the outer plan.
2361 : */
2362 292294 : if (aggstate->grp_firstTuple == NULL)
2363 : {
2364 61068 : outerslot = fetch_input_tuple(aggstate);
2365 61050 : if (!TupIsNull(outerslot))
2366 : {
2367 : /*
2368 : * Make a copy of the first input tuple; we will use this
2369 : * for comparisons (in group mode) and for projection.
2370 : */
2371 49470 : aggstate->grp_firstTuple = ExecCopySlotHeapTuple(outerslot);
2372 : }
2373 : else
2374 : {
2375 : /* outer plan produced no tuples at all */
2376 11580 : if (hasGroupingSets)
2377 : {
2378 : /*
2379 : * If there was no input at all, we need to project
2380 : * rows only if there are grouping sets of size 0.
2381 : * Note that this implies that there can't be any
2382 : * references to ungrouped Vars, which would otherwise
2383 : * cause issues with the empty output slot.
2384 : *
2385 : * XXX: This is no longer true, we currently deal with
2386 : * this in finalize_aggregates().
2387 : */
2388 54 : aggstate->input_done = true;
2389 :
2390 78 : while (aggstate->phase->gset_lengths[aggstate->projected_set] > 0)
2391 : {
2392 30 : aggstate->projected_set += 1;
2393 30 : if (aggstate->projected_set >= numGroupingSets)
2394 : {
2395 : /*
2396 : * We can't set agg_done here because we might
2397 : * have more phases to do, even though the
2398 : * input is empty. So we need to restart the
2399 : * whole outer loop.
2400 : */
2401 6 : break;
2402 : }
2403 : }
2404 :
2405 54 : if (aggstate->projected_set >= numGroupingSets)
2406 6 : continue;
2407 : }
2408 : else
2409 : {
2410 11526 : aggstate->agg_done = true;
2411 : /* If we are grouping, we should produce no tuples too */
2412 11526 : if (node->aggstrategy != AGG_PLAIN)
2413 164 : return NULL;
2414 : }
2415 : }
2416 : }
2417 :
2418 : /*
2419 : * Initialize working state for a new input tuple group.
2420 : */
2421 292106 : initialize_aggregates(aggstate, pergroups, numReset);
2422 :
2423 292106 : if (aggstate->grp_firstTuple != NULL)
2424 : {
2425 : /*
2426 : * Store the copied first input tuple in the tuple table slot
2427 : * reserved for it. The tuple will be deleted when it is
2428 : * cleared from the slot.
2429 : */
2430 280696 : ExecForceStoreHeapTuple(aggstate->grp_firstTuple,
2431 : firstSlot, true);
2432 280696 : aggstate->grp_firstTuple = NULL; /* don't keep two pointers */
2433 :
2434 : /* set up for first advance_aggregates call */
2435 280696 : tmpcontext->ecxt_outertuple = firstSlot;
2436 :
2437 : /*
2438 : * Process each outer-plan tuple, and then fetch the next one,
2439 : * until we exhaust the outer plan or cross a group boundary.
2440 : */
2441 : for (;;)
2442 : {
2443 : /*
2444 : * During phase 1 only of a mixed agg, we need to update
2445 : * hashtables as well in advance_aggregates.
2446 : */
2447 19671808 : if (aggstate->aggstrategy == AGG_MIXED &&
2448 38044 : aggstate->current_phase == 1)
2449 : {
2450 38044 : lookup_hash_entries(aggstate);
2451 : }
2452 :
2453 : /* Advance the aggregates (or combine functions) */
2454 19671808 : advance_aggregates(aggstate);
2455 :
2456 : /* Reset per-input-tuple context after each tuple */
2457 19671742 : ResetExprContext(tmpcontext);
2458 :
2459 19671742 : outerslot = fetch_input_tuple(aggstate);
2460 19671742 : if (TupIsNull(outerslot))
2461 : {
2462 : /* no more outer-plan tuples available */
2463 :
2464 : /* if we built hash tables, finalize any spills */
2465 49398 : if (aggstate->aggstrategy == AGG_MIXED &&
2466 126 : aggstate->current_phase == 1)
2467 126 : hashagg_finish_initial_spills(aggstate);
2468 :
2469 49398 : if (hasGroupingSets)
2470 : {
2471 666 : aggstate->input_done = true;
2472 666 : break;
2473 : }
2474 : else
2475 : {
2476 48732 : aggstate->agg_done = true;
2477 48732 : break;
2478 : }
2479 : }
2480 : /* set up for next advance_aggregates call */
2481 19622344 : tmpcontext->ecxt_outertuple = outerslot;
2482 :
2483 : /*
2484 : * If we are grouping, check whether we've crossed a group
2485 : * boundary.
2486 : */
2487 19622344 : if (node->aggstrategy != AGG_PLAIN && node->numCols > 0)
2488 : {
2489 1811040 : tmpcontext->ecxt_innertuple = firstSlot;
2490 1811040 : if (!ExecQual(aggstate->phase->eqfunctions[node->numCols - 1],
2491 : tmpcontext))
2492 : {
2493 231232 : aggstate->grp_firstTuple = ExecCopySlotHeapTuple(outerslot);
2494 231232 : break;
2495 : }
2496 : }
2497 : }
2498 : }
2499 :
2500 : /*
2501 : * Use the representative input tuple for any references to
2502 : * non-aggregated input columns in aggregate direct args, the node
2503 : * qual, and the tlist. (If we are not grouping, and there are no
2504 : * input rows at all, we will come here with an empty firstSlot
2505 : * ... but if not grouping, there can't be any references to
2506 : * non-aggregated input columns, so no problem.)
2507 : */
2508 292040 : econtext->ecxt_outertuple = firstSlot;
2509 : }
2510 :
2511 : Assert(aggstate->projected_set >= 0);
2512 :
2513 306178 : currentSet = aggstate->projected_set;
2514 :
2515 306178 : prepare_projection_slot(aggstate, econtext->ecxt_outertuple, currentSet);
2516 :
2517 306178 : select_current_set(aggstate, currentSet, false);
2518 :
2519 306178 : finalize_aggregates(aggstate,
2520 : peragg,
2521 306178 : pergroups[currentSet]);
2522 :
2523 : /*
2524 : * If there's no row to project right now, we must continue rather
2525 : * than returning a null since there might be more groups.
2526 : */
2527 306166 : result = project_aggregates(aggstate);
2528 306166 : if (result)
2529 235218 : return result;
2530 : }
2531 :
2532 : /* No more groups */
2533 794 : return NULL;
2534 : }
2535 :
2536 : /*
2537 : * ExecAgg for hashed case: read input and build hash table
2538 : */
2539 : static void
2540 3658 : agg_fill_hash_table(AggState *aggstate)
2541 : {
2542 : TupleTableSlot *outerslot;
2543 3658 : ExprContext *tmpcontext = aggstate->tmpcontext;
2544 :
2545 : /*
2546 : * Process each outer-plan tuple, and then fetch the next one, until we
2547 : * exhaust the outer plan.
2548 : */
2549 : for (;;)
2550 : {
2551 5341660 : outerslot = fetch_input_tuple(aggstate);
2552 5341660 : if (TupIsNull(outerslot))
2553 : break;
2554 :
2555 : /* set up for lookup_hash_entries and advance_aggregates */
2556 5338002 : tmpcontext->ecxt_outertuple = outerslot;
2557 :
2558 : /* Find or build hashtable entries */
2559 5338002 : lookup_hash_entries(aggstate);
2560 :
2561 : /* Advance the aggregates (or combine functions) */
2562 5338002 : advance_aggregates(aggstate);
2563 :
2564 : /*
2565 : * Reset per-input-tuple context after each tuple, but note that the
2566 : * hash lookups do this too
2567 : */
2568 5338002 : ResetExprContext(aggstate->tmpcontext);
2569 : }
2570 :
2571 : /* finalize spills, if any */
2572 3658 : hashagg_finish_initial_spills(aggstate);
2573 :
2574 3658 : aggstate->table_filled = true;
2575 : /* Initialize to walk the first hash table */
2576 3658 : select_current_set(aggstate, 0, true);
2577 3658 : ResetTupleHashIterator(aggstate->perhash[0].hashtable,
2578 : &aggstate->perhash[0].hashiter);
2579 3658 : }
2580 :
2581 : /*
2582 : * If any data was spilled during hash aggregation, reset the hash table and
2583 : * reprocess one batch of spilled data. After reprocessing a batch, the hash
2584 : * table will again contain data, ready to be consumed by
2585 : * agg_retrieve_hash_table_in_memory().
2586 : *
2587 : * Should only be called after all in memory hash table entries have been
2588 : * finalized and emitted.
2589 : *
2590 : * Return false when input is exhausted and there's no more work to be done;
2591 : * otherwise return true.
2592 : */
2593 : static bool
2594 31628 : agg_refill_hash_table(AggState *aggstate)
2595 : {
2596 : HashAggBatch *batch;
2597 : AggStatePerHash perhash;
2598 : HashAggSpill spill;
2599 31628 : LogicalTapeSet *tapeset = aggstate->hash_tapeset;
2600 31628 : bool spill_initialized = false;
2601 :
2602 31628 : if (aggstate->hash_batches == NIL)
2603 4718 : return false;
2604 :
2605 : /* hash_batches is a stack, with the top item at the end of the list */
2606 26910 : batch = llast(aggstate->hash_batches);
2607 26910 : aggstate->hash_batches = list_delete_last(aggstate->hash_batches);
2608 :
2609 26910 : hash_agg_set_limits(aggstate->hashentrysize, batch->input_card,
2610 : batch->used_bits, &aggstate->hash_mem_limit,
2611 : &aggstate->hash_ngroups_limit, NULL);
2612 :
2613 : /*
2614 : * Each batch only processes one grouping set; set the rest to NULL so
2615 : * that advance_aggregates() knows to ignore them. We don't touch
2616 : * pergroups for sorted grouping sets here, because they will be needed if
2617 : * we rescan later. The expressions for sorted grouping sets will not be
2618 : * evaluated after we recompile anyway.
2619 : */
2620 207360 : MemSet(aggstate->hash_pergroup, 0,
2621 : sizeof(AggStatePerGroup) * aggstate->num_hashes);
2622 :
2623 : /* free memory and reset hash tables */
2624 26910 : ReScanExprContext(aggstate->hashcontext);
2625 207360 : for (int setno = 0; setno < aggstate->num_hashes; setno++)
2626 180450 : ResetTupleHashTable(aggstate->perhash[setno].hashtable);
2627 :
2628 26910 : aggstate->hash_ngroups_current = 0;
2629 :
2630 : /*
2631 : * In AGG_MIXED mode, hash aggregation happens in phase 1 and the output
2632 : * happens in phase 0. So, we switch to phase 1 when processing a batch,
2633 : * and back to phase 0 after the batch is done.
2634 : */
2635 : Assert(aggstate->current_phase == 0);
2636 26910 : if (aggstate->phase->aggstrategy == AGG_MIXED)
2637 : {
2638 26262 : aggstate->current_phase = 1;
2639 26262 : aggstate->phase = &aggstate->phases[aggstate->current_phase];
2640 : }
2641 :
2642 26910 : select_current_set(aggstate, batch->setno, true);
2643 :
2644 26910 : perhash = &aggstate->perhash[aggstate->current_set];
2645 :
2646 : /*
2647 : * Spilled tuples are always read back as MinimalTuples, which may be
2648 : * different from the outer plan, so recompile the aggregate expressions.
2649 : *
2650 : * We still need the NULL check, because we are only processing one
2651 : * grouping set at a time and the rest will be NULL.
2652 : */
2653 26910 : hashagg_recompile_expressions(aggstate, true, true);
2654 :
2655 : for (;;)
2656 655824 : {
2657 682734 : TupleTableSlot *spillslot = aggstate->hash_spill_rslot;
2658 682734 : TupleTableSlot *hashslot = perhash->hashslot;
2659 : TupleHashEntry entry;
2660 : MinimalTuple tuple;
2661 : uint32 hash;
2662 682734 : bool isnew = false;
2663 682734 : bool *p_isnew = aggstate->hash_spill_mode ? NULL : &isnew;
2664 :
2665 682734 : CHECK_FOR_INTERRUPTS();
2666 :
2667 682734 : tuple = hashagg_batch_read(batch, &hash);
2668 682734 : if (tuple == NULL)
2669 26910 : break;
2670 :
2671 655824 : ExecStoreMinimalTuple(tuple, spillslot, true);
2672 655824 : aggstate->tmpcontext->ecxt_outertuple = spillslot;
2673 :
2674 655824 : prepare_hash_slot(perhash,
2675 655824 : aggstate->tmpcontext->ecxt_outertuple,
2676 : hashslot);
2677 655824 : entry = LookupTupleHashEntryHash(perhash->hashtable, hashslot,
2678 : p_isnew, hash);
2679 :
2680 655824 : if (entry != NULL)
2681 : {
2682 221880 : if (isnew)
2683 94494 : initialize_hash_entry(aggstate, perhash->hashtable, entry);
2684 221880 : aggstate->hash_pergroup[batch->setno] = entry->additional;
2685 221880 : advance_aggregates(aggstate);
2686 : }
2687 : else
2688 : {
2689 433944 : if (!spill_initialized)
2690 : {
2691 : /*
2692 : * Avoid initializing the spill until we actually need it so
2693 : * that we don't assign tapes that will never be used.
2694 : */
2695 12492 : spill_initialized = true;
2696 12492 : hashagg_spill_init(&spill, tapeset, batch->used_bits,
2697 : batch->input_card, aggstate->hashentrysize);
2698 : }
2699 : /* no memory for a new group, spill */
2700 433944 : hashagg_spill_tuple(aggstate, &spill, spillslot, hash);
2701 :
2702 433944 : aggstate->hash_pergroup[batch->setno] = NULL;
2703 : }
2704 :
2705 : /*
2706 : * Reset per-input-tuple context after each tuple, but note that the
2707 : * hash lookups do this too
2708 : */
2709 655824 : ResetExprContext(aggstate->tmpcontext);
2710 : }
2711 :
2712 26910 : LogicalTapeClose(batch->input_tape);
2713 :
2714 : /* change back to phase 0 */
2715 26910 : aggstate->current_phase = 0;
2716 26910 : aggstate->phase = &aggstate->phases[aggstate->current_phase];
2717 :
2718 26910 : if (spill_initialized)
2719 : {
2720 12492 : hashagg_spill_finish(aggstate, &spill, batch->setno);
2721 12492 : hash_agg_update_metrics(aggstate, true, spill.npartitions);
2722 : }
2723 : else
2724 14418 : hash_agg_update_metrics(aggstate, true, 0);
2725 :
2726 26910 : aggstate->hash_spill_mode = false;
2727 :
2728 : /* prepare to walk the first hash table */
2729 26910 : select_current_set(aggstate, batch->setno, true);
2730 26910 : ResetTupleHashIterator(aggstate->perhash[batch->setno].hashtable,
2731 : &aggstate->perhash[batch->setno].hashiter);
2732 :
2733 26910 : pfree(batch);
2734 :
2735 26910 : return true;
2736 : }
2737 :
2738 : /*
2739 : * ExecAgg for hashed case: retrieving groups from hash table
2740 : *
2741 : * After exhausting in-memory tuples, also try refilling the hash table using
2742 : * previously-spilled tuples. Only returns NULL after all in-memory and
2743 : * spilled tuples are exhausted.
2744 : */
2745 : static TupleTableSlot *
2746 518204 : agg_retrieve_hash_table(AggState *aggstate)
2747 : {
2748 518204 : TupleTableSlot *result = NULL;
2749 :
2750 1058600 : while (result == NULL)
2751 : {
2752 545114 : result = agg_retrieve_hash_table_in_memory(aggstate);
2753 545114 : if (result == NULL)
2754 : {
2755 31628 : if (!agg_refill_hash_table(aggstate))
2756 : {
2757 4718 : aggstate->agg_done = true;
2758 4718 : break;
2759 : }
2760 : }
2761 : }
2762 :
2763 518204 : return result;
2764 : }
2765 :
2766 : /*
2767 : * Retrieve the groups from the in-memory hash tables without considering any
2768 : * spilled tuples.
2769 : */
2770 : static TupleTableSlot *
2771 545114 : agg_retrieve_hash_table_in_memory(AggState *aggstate)
2772 : {
2773 : ExprContext *econtext;
2774 : AggStatePerAgg peragg;
2775 : AggStatePerGroup pergroup;
2776 : TupleHashEntryData *entry;
2777 : TupleTableSlot *firstSlot;
2778 : TupleTableSlot *result;
2779 : AggStatePerHash perhash;
2780 :
2781 : /*
2782 : * get state info from node.
2783 : *
2784 : * econtext is the per-output-tuple expression context.
2785 : */
2786 545114 : econtext = aggstate->ss.ps.ps_ExprContext;
2787 545114 : peragg = aggstate->peragg;
2788 545114 : firstSlot = aggstate->ss.ss_ScanTupleSlot;
2789 :
2790 : /*
2791 : * Note that perhash (and therefore anything accessed through it) can
2792 : * change inside the loop, as we change between grouping sets.
2793 : */
2794 545114 : perhash = &aggstate->perhash[aggstate->current_set];
2795 :
2796 : /*
2797 : * We loop retrieving groups until we find one satisfying
2798 : * aggstate->ss.ps.qual
2799 : */
2800 : for (;;)
2801 135852 : {
2802 680966 : TupleTableSlot *hashslot = perhash->hashslot;
2803 : int i;
2804 :
2805 680966 : CHECK_FOR_INTERRUPTS();
2806 :
2807 : /*
2808 : * Find the next entry in the hash table
2809 : */
2810 680966 : entry = ScanTupleHashTable(perhash->hashtable, &perhash->hashiter);
2811 680966 : if (entry == NULL)
2812 : {
2813 131860 : int nextset = aggstate->current_set + 1;
2814 :
2815 131860 : if (nextset < aggstate->num_hashes)
2816 : {
2817 : /*
2818 : * Switch to next grouping set, reinitialize, and restart the
2819 : * loop.
2820 : */
2821 100232 : select_current_set(aggstate, nextset, true);
2822 :
2823 100232 : perhash = &aggstate->perhash[aggstate->current_set];
2824 :
2825 100232 : ResetTupleHashIterator(perhash->hashtable, &perhash->hashiter);
2826 :
2827 100232 : continue;
2828 : }
2829 : else
2830 : {
2831 31628 : return NULL;
2832 : }
2833 : }
2834 :
2835 : /*
2836 : * Clear the per-output-tuple context for each group
2837 : *
2838 : * We intentionally don't use ReScanExprContext here; if any aggs have
2839 : * registered shutdown callbacks, they mustn't be called yet, since we
2840 : * might not be done with that agg.
2841 : */
2842 549106 : ResetExprContext(econtext);
2843 :
2844 : /*
2845 : * Transform representative tuple back into one with the right
2846 : * columns.
2847 : */
2848 549106 : ExecStoreMinimalTuple(entry->firstTuple, hashslot, false);
2849 549106 : slot_getallattrs(hashslot);
2850 :
2851 549106 : ExecClearTuple(firstSlot);
2852 549106 : memset(firstSlot->tts_isnull, true,
2853 549106 : firstSlot->tts_tupleDescriptor->natts * sizeof(bool));
2854 :
2855 1701796 : for (i = 0; i < perhash->numhashGrpCols; i++)
2856 : {
2857 1152690 : int varNumber = perhash->hashGrpColIdxInput[i] - 1;
2858 :
2859 1152690 : firstSlot->tts_values[varNumber] = hashslot->tts_values[i];
2860 1152690 : firstSlot->tts_isnull[varNumber] = hashslot->tts_isnull[i];
2861 : }
2862 549106 : ExecStoreVirtualTuple(firstSlot);
2863 :
2864 549106 : pergroup = (AggStatePerGroup) entry->additional;
2865 :
2866 : /*
2867 : * Use the representative input tuple for any references to
2868 : * non-aggregated input columns in the qual and tlist.
2869 : */
2870 549106 : econtext->ecxt_outertuple = firstSlot;
2871 :
2872 549106 : prepare_projection_slot(aggstate,
2873 : econtext->ecxt_outertuple,
2874 : aggstate->current_set);
2875 :
2876 549106 : finalize_aggregates(aggstate, peragg, pergroup);
2877 :
2878 549106 : result = project_aggregates(aggstate);
2879 549106 : if (result)
2880 513486 : return result;
2881 : }
2882 :
2883 : /* No more groups */
2884 : return NULL;
2885 : }
2886 :
2887 : /*
2888 : * hashagg_spill_init
2889 : *
2890 : * Called after we determined that spilling is necessary. Chooses the number
2891 : * of partitions to create, and initializes them.
2892 : */
2893 : static void
2894 12606 : hashagg_spill_init(HashAggSpill *spill, LogicalTapeSet *tapeset, int used_bits,
2895 : double input_groups, double hashentrysize)
2896 : {
2897 : int npartitions;
2898 : int partition_bits;
2899 :
2900 12606 : npartitions = hash_choose_num_partitions(input_groups, hashentrysize,
2901 : used_bits, &partition_bits);
2902 :
2903 12606 : spill->partitions = palloc0(sizeof(LogicalTape *) * npartitions);
2904 12606 : spill->ntuples = palloc0(sizeof(int64) * npartitions);
2905 12606 : spill->hll_card = palloc0(sizeof(hyperLogLogState) * npartitions);
2906 :
2907 63030 : for (int i = 0; i < npartitions; i++)
2908 50424 : spill->partitions[i] = LogicalTapeCreate(tapeset);
2909 :
2910 12606 : spill->shift = 32 - used_bits - partition_bits;
2911 12606 : spill->mask = (npartitions - 1) << spill->shift;
2912 12606 : spill->npartitions = npartitions;
2913 :
2914 63030 : for (int i = 0; i < npartitions; i++)
2915 50424 : initHyperLogLog(&spill->hll_card[i], HASHAGG_HLL_BIT_WIDTH);
2916 12606 : }
2917 :
2918 : /*
2919 : * hashagg_spill_tuple
2920 : *
2921 : * No room for new groups in the hash table. Save for later in the appropriate
2922 : * partition.
2923 : */
2924 : static Size
2925 655824 : hashagg_spill_tuple(AggState *aggstate, HashAggSpill *spill,
2926 : TupleTableSlot *inputslot, uint32 hash)
2927 : {
2928 : TupleTableSlot *spillslot;
2929 : int partition;
2930 : MinimalTuple tuple;
2931 : LogicalTape *tape;
2932 655824 : int total_written = 0;
2933 : bool shouldFree;
2934 :
2935 : Assert(spill->partitions != NULL);
2936 :
2937 : /* spill only attributes that we actually need */
2938 655824 : if (!aggstate->all_cols_needed)
2939 : {
2940 3492 : spillslot = aggstate->hash_spill_wslot;
2941 3492 : slot_getsomeattrs(inputslot, aggstate->max_colno_needed);
2942 3492 : ExecClearTuple(spillslot);
2943 10476 : for (int i = 0; i < spillslot->tts_tupleDescriptor->natts; i++)
2944 : {
2945 6984 : if (bms_is_member(i + 1, aggstate->colnos_needed))
2946 : {
2947 3492 : spillslot->tts_values[i] = inputslot->tts_values[i];
2948 3492 : spillslot->tts_isnull[i] = inputslot->tts_isnull[i];
2949 : }
2950 : else
2951 3492 : spillslot->tts_isnull[i] = true;
2952 : }
2953 3492 : ExecStoreVirtualTuple(spillslot);
2954 : }
2955 : else
2956 652332 : spillslot = inputslot;
2957 :
2958 655824 : tuple = ExecFetchSlotMinimalTuple(spillslot, &shouldFree);
2959 :
2960 655824 : partition = (hash & spill->mask) >> spill->shift;
2961 655824 : spill->ntuples[partition]++;
2962 :
2963 : /*
2964 : * All hash values destined for a given partition have some bits in
2965 : * common, which causes bad HLL cardinality estimates. Hash the hash to
2966 : * get a more uniform distribution.
2967 : */
2968 655824 : addHyperLogLog(&spill->hll_card[partition], hash_bytes_uint32(hash));
2969 :
2970 655824 : tape = spill->partitions[partition];
2971 :
2972 655824 : LogicalTapeWrite(tape, &hash, sizeof(uint32));
2973 655824 : total_written += sizeof(uint32);
2974 :
2975 655824 : LogicalTapeWrite(tape, tuple, tuple->t_len);
2976 655824 : total_written += tuple->t_len;
2977 :
2978 655824 : if (shouldFree)
2979 221880 : pfree(tuple);
2980 :
2981 655824 : return total_written;
2982 : }
2983 :
2984 : /*
2985 : * hashagg_batch_new
2986 : *
2987 : * Construct a HashAggBatch item, which represents one iteration of HashAgg to
2988 : * be done.
2989 : */
2990 : static HashAggBatch *
2991 26910 : hashagg_batch_new(LogicalTape *input_tape, int setno,
2992 : int64 input_tuples, double input_card, int used_bits)
2993 : {
2994 26910 : HashAggBatch *batch = palloc0(sizeof(HashAggBatch));
2995 :
2996 26910 : batch->setno = setno;
2997 26910 : batch->used_bits = used_bits;
2998 26910 : batch->input_tape = input_tape;
2999 26910 : batch->input_tuples = input_tuples;
3000 26910 : batch->input_card = input_card;
3001 :
3002 26910 : return batch;
3003 : }
3004 :
3005 : /*
3006 : * hashagg_batch_read
3007 : * read the next tuple from a batch's tape. Return NULL if no more.
3008 : */
3009 : static MinimalTuple
3010 682734 : hashagg_batch_read(HashAggBatch *batch, uint32 *hashp)
3011 : {
3012 682734 : LogicalTape *tape = batch->input_tape;
3013 : MinimalTuple tuple;
3014 : uint32 t_len;
3015 : size_t nread;
3016 : uint32 hash;
3017 :
3018 682734 : nread = LogicalTapeRead(tape, &hash, sizeof(uint32));
3019 682734 : if (nread == 0)
3020 26910 : return NULL;
3021 655824 : if (nread != sizeof(uint32))
3022 0 : ereport(ERROR,
3023 : (errcode_for_file_access(),
3024 : errmsg_internal("unexpected EOF for tape %p: requested %zu bytes, read %zu bytes",
3025 : tape, sizeof(uint32), nread)));
3026 655824 : if (hashp != NULL)
3027 655824 : *hashp = hash;
3028 :
3029 655824 : nread = LogicalTapeRead(tape, &t_len, sizeof(t_len));
3030 655824 : if (nread != sizeof(uint32))
3031 0 : ereport(ERROR,
3032 : (errcode_for_file_access(),
3033 : errmsg_internal("unexpected EOF for tape %p: requested %zu bytes, read %zu bytes",
3034 : tape, sizeof(uint32), nread)));
3035 :
3036 655824 : tuple = (MinimalTuple) palloc(t_len);
3037 655824 : tuple->t_len = t_len;
3038 :
3039 655824 : nread = LogicalTapeRead(tape,
3040 : (char *) tuple + sizeof(uint32),
3041 : t_len - sizeof(uint32));
3042 655824 : if (nread != t_len - sizeof(uint32))
3043 0 : ereport(ERROR,
3044 : (errcode_for_file_access(),
3045 : errmsg_internal("unexpected EOF for tape %p: requested %zu bytes, read %zu bytes",
3046 : tape, t_len - sizeof(uint32), nread)));
3047 :
3048 655824 : return tuple;
3049 : }
3050 :
3051 : /*
3052 : * hashagg_finish_initial_spills
3053 : *
3054 : * After a HashAggBatch has been processed, it may have spilled tuples to
3055 : * disk. If so, turn the spilled partitions into new batches that must later
3056 : * be executed.
3057 : */
3058 : static void
3059 3784 : hashagg_finish_initial_spills(AggState *aggstate)
3060 : {
3061 : int setno;
3062 3784 : int total_npartitions = 0;
3063 :
3064 3784 : if (aggstate->hash_spills != NULL)
3065 : {
3066 168 : for (setno = 0; setno < aggstate->num_hashes; setno++)
3067 : {
3068 114 : HashAggSpill *spill = &aggstate->hash_spills[setno];
3069 :
3070 114 : total_npartitions += spill->npartitions;
3071 114 : hashagg_spill_finish(aggstate, spill, setno);
3072 : }
3073 :
3074 : /*
3075 : * We're not processing tuples from outer plan any more; only
3076 : * processing batches of spilled tuples. The initial spill structures
3077 : * are no longer needed.
3078 : */
3079 54 : pfree(aggstate->hash_spills);
3080 54 : aggstate->hash_spills = NULL;
3081 : }
3082 :
3083 3784 : hash_agg_update_metrics(aggstate, false, total_npartitions);
3084 3784 : aggstate->hash_spill_mode = false;
3085 3784 : }
3086 :
3087 : /*
3088 : * hashagg_spill_finish
3089 : *
3090 : * Transform spill partitions into new batches.
3091 : */
3092 : static void
3093 12606 : hashagg_spill_finish(AggState *aggstate, HashAggSpill *spill, int setno)
3094 : {
3095 : int i;
3096 12606 : int used_bits = 32 - spill->shift;
3097 :
3098 12606 : if (spill->npartitions == 0)
3099 0 : return; /* didn't spill */
3100 :
3101 63030 : for (i = 0; i < spill->npartitions; i++)
3102 : {
3103 50424 : LogicalTape *tape = spill->partitions[i];
3104 : HashAggBatch *new_batch;
3105 : double cardinality;
3106 :
3107 : /* if the partition is empty, don't create a new batch of work */
3108 50424 : if (spill->ntuples[i] == 0)
3109 23514 : continue;
3110 :
3111 26910 : cardinality = estimateHyperLogLog(&spill->hll_card[i]);
3112 26910 : freeHyperLogLog(&spill->hll_card[i]);
3113 :
3114 : /* rewinding frees the buffer while not in use */
3115 26910 : LogicalTapeRewindForRead(tape, HASHAGG_READ_BUFFER_SIZE);
3116 :
3117 26910 : new_batch = hashagg_batch_new(tape, setno,
3118 26910 : spill->ntuples[i], cardinality,
3119 : used_bits);
3120 26910 : aggstate->hash_batches = lappend(aggstate->hash_batches, new_batch);
3121 26910 : aggstate->hash_batches_used++;
3122 : }
3123 :
3124 12606 : pfree(spill->ntuples);
3125 12606 : pfree(spill->hll_card);
3126 12606 : pfree(spill->partitions);
3127 : }
3128 :
3129 : /*
3130 : * Free resources related to a spilled HashAgg.
3131 : */
3132 : static void
3133 41204 : hashagg_reset_spill_state(AggState *aggstate)
3134 : {
3135 : /* free spills from initial pass */
3136 41204 : if (aggstate->hash_spills != NULL)
3137 : {
3138 : int setno;
3139 :
3140 0 : for (setno = 0; setno < aggstate->num_hashes; setno++)
3141 : {
3142 0 : HashAggSpill *spill = &aggstate->hash_spills[setno];
3143 :
3144 0 : pfree(spill->ntuples);
3145 0 : pfree(spill->partitions);
3146 : }
3147 0 : pfree(aggstate->hash_spills);
3148 0 : aggstate->hash_spills = NULL;
3149 : }
3150 :
3151 : /* free batches */
3152 41204 : list_free_deep(aggstate->hash_batches);
3153 41204 : aggstate->hash_batches = NIL;
3154 :
3155 : /* close tape set */
3156 41204 : if (aggstate->hash_tapeset != NULL)
3157 : {
3158 54 : LogicalTapeSetClose(aggstate->hash_tapeset);
3159 54 : aggstate->hash_tapeset = NULL;
3160 : }
3161 41204 : }
3162 :
3163 :
3164 : /* -----------------
3165 : * ExecInitAgg
3166 : *
3167 : * Creates the run-time information for the agg node produced by the
3168 : * planner and initializes its outer subtree.
3169 : *
3170 : * -----------------
3171 : */
3172 : AggState *
3173 41238 : ExecInitAgg(Agg *node, EState *estate, int eflags)
3174 : {
3175 : AggState *aggstate;
3176 : AggStatePerAgg peraggs;
3177 : AggStatePerTrans pertransstates;
3178 : AggStatePerGroup *pergroups;
3179 : Plan *outerPlan;
3180 : ExprContext *econtext;
3181 : TupleDesc scanDesc;
3182 : int max_aggno;
3183 : int max_transno;
3184 : int numaggrefs;
3185 : int numaggs;
3186 : int numtrans;
3187 : int phase;
3188 : int phaseidx;
3189 : ListCell *l;
3190 41238 : Bitmapset *all_grouped_cols = NULL;
3191 41238 : int numGroupingSets = 1;
3192 : int numPhases;
3193 : int numHashes;
3194 41238 : int i = 0;
3195 41238 : int j = 0;
3196 77564 : bool use_hashing = (node->aggstrategy == AGG_HASHED ||
3197 36326 : node->aggstrategy == AGG_MIXED);
3198 :
3199 : /* check for unsupported flags */
3200 : Assert(!(eflags & (EXEC_FLAG_BACKWARD | EXEC_FLAG_MARK)));
3201 :
3202 : /*
3203 : * create state structure
3204 : */
3205 41238 : aggstate = makeNode(AggState);
3206 41238 : aggstate->ss.ps.plan = (Plan *) node;
3207 41238 : aggstate->ss.ps.state = estate;
3208 41238 : aggstate->ss.ps.ExecProcNode = ExecAgg;
3209 :
3210 41238 : aggstate->aggs = NIL;
3211 41238 : aggstate->numaggs = 0;
3212 41238 : aggstate->numtrans = 0;
3213 41238 : aggstate->aggstrategy = node->aggstrategy;
3214 41238 : aggstate->aggsplit = node->aggsplit;
3215 41238 : aggstate->maxsets = 0;
3216 41238 : aggstate->projected_set = -1;
3217 41238 : aggstate->current_set = 0;
3218 41238 : aggstate->peragg = NULL;
3219 41238 : aggstate->pertrans = NULL;
3220 41238 : aggstate->curperagg = NULL;
3221 41238 : aggstate->curpertrans = NULL;
3222 41238 : aggstate->input_done = false;
3223 41238 : aggstate->agg_done = false;
3224 41238 : aggstate->pergroups = NULL;
3225 41238 : aggstate->grp_firstTuple = NULL;
3226 41238 : aggstate->sort_in = NULL;
3227 41238 : aggstate->sort_out = NULL;
3228 :
3229 : /*
3230 : * phases[0] always exists, but is dummy in sorted/plain mode
3231 : */
3232 41238 : numPhases = (use_hashing ? 1 : 2);
3233 41238 : numHashes = (use_hashing ? 1 : 0);
3234 :
3235 : /*
3236 : * Calculate the maximum number of grouping sets in any phase; this
3237 : * determines the size of some allocations. Also calculate the number of
3238 : * phases, since all hashed/mixed nodes contribute to only a single phase.
3239 : */
3240 41238 : if (node->groupingSets)
3241 : {
3242 716 : numGroupingSets = list_length(node->groupingSets);
3243 :
3244 1534 : foreach(l, node->chain)
3245 : {
3246 818 : Agg *agg = lfirst(l);
3247 :
3248 818 : numGroupingSets = Max(numGroupingSets,
3249 : list_length(agg->groupingSets));
3250 :
3251 : /*
3252 : * additional AGG_HASHED aggs become part of phase 0, but all
3253 : * others add an extra phase.
3254 : */
3255 818 : if (agg->aggstrategy != AGG_HASHED)
3256 424 : ++numPhases;
3257 : else
3258 394 : ++numHashes;
3259 : }
3260 : }
3261 :
3262 41238 : aggstate->maxsets = numGroupingSets;
3263 41238 : aggstate->numphases = numPhases;
3264 :
3265 41238 : aggstate->aggcontexts = (ExprContext **)
3266 41238 : palloc0(sizeof(ExprContext *) * numGroupingSets);
3267 :
3268 : /*
3269 : * Create expression contexts. We need three or more, one for
3270 : * per-input-tuple processing, one for per-output-tuple processing, one
3271 : * for all the hashtables, and one for each grouping set. The per-tuple
3272 : * memory context of the per-grouping-set ExprContexts (aggcontexts)
3273 : * replaces the standalone memory context formerly used to hold transition
3274 : * values. We cheat a little by using ExecAssignExprContext() to build
3275 : * all of them.
3276 : *
3277 : * NOTE: the details of what is stored in aggcontexts and what is stored
3278 : * in the regular per-query memory context are driven by a simple
3279 : * decision: we want to reset the aggcontext at group boundaries (if not
3280 : * hashing) and in ExecReScanAgg to recover no-longer-wanted space.
3281 : */
3282 41238 : ExecAssignExprContext(estate, &aggstate->ss.ps);
3283 41238 : aggstate->tmpcontext = aggstate->ss.ps.ps_ExprContext;
3284 :
3285 83304 : for (i = 0; i < numGroupingSets; ++i)
3286 : {
3287 42066 : ExecAssignExprContext(estate, &aggstate->ss.ps);
3288 42066 : aggstate->aggcontexts[i] = aggstate->ss.ps.ps_ExprContext;
3289 : }
3290 :
3291 41238 : if (use_hashing)
3292 5108 : aggstate->hashcontext = CreateWorkExprContext(estate);
3293 :
3294 41238 : ExecAssignExprContext(estate, &aggstate->ss.ps);
3295 :
3296 : /*
3297 : * Initialize child nodes.
3298 : *
3299 : * If we are doing a hashed aggregation then the child plan does not need
3300 : * to handle REWIND efficiently; see ExecReScanAgg.
3301 : */
3302 41238 : if (node->aggstrategy == AGG_HASHED)
3303 4912 : eflags &= ~EXEC_FLAG_REWIND;
3304 41238 : outerPlan = outerPlan(node);
3305 41238 : outerPlanState(aggstate) = ExecInitNode(outerPlan, estate, eflags);
3306 :
3307 : /*
3308 : * initialize source tuple type.
3309 : */
3310 41238 : aggstate->ss.ps.outerops =
3311 41238 : ExecGetResultSlotOps(outerPlanState(&aggstate->ss),
3312 : &aggstate->ss.ps.outeropsfixed);
3313 41238 : aggstate->ss.ps.outeropsset = true;
3314 :
3315 41238 : ExecCreateScanSlotFromOuterPlan(estate, &aggstate->ss,
3316 : aggstate->ss.ps.outerops);
3317 41238 : scanDesc = aggstate->ss.ss_ScanTupleSlot->tts_tupleDescriptor;
3318 :
3319 : /*
3320 : * If there are more than two phases (including a potential dummy phase
3321 : * 0), input will be resorted using tuplesort. Need a slot for that.
3322 : */
3323 41238 : if (numPhases > 2)
3324 : {
3325 186 : aggstate->sort_slot = ExecInitExtraTupleSlot(estate, scanDesc,
3326 : &TTSOpsMinimalTuple);
3327 :
3328 : /*
3329 : * The output of the tuplesort, and the output from the outer child
3330 : * might not use the same type of slot. In most cases the child will
3331 : * be a Sort, and thus return a TTSOpsMinimalTuple type slot - but the
3332 : * input can also be presorted due an index, in which case it could be
3333 : * a different type of slot.
3334 : *
3335 : * XXX: For efficiency it would be good to instead/additionally
3336 : * generate expressions with corresponding settings of outerops* for
3337 : * the individual phases - deforming is often a bottleneck for
3338 : * aggregations with lots of rows per group. If there's multiple
3339 : * sorts, we know that all but the first use TTSOpsMinimalTuple (via
3340 : * the nodeAgg.c internal tuplesort).
3341 : */
3342 186 : if (aggstate->ss.ps.outeropsfixed &&
3343 186 : aggstate->ss.ps.outerops != &TTSOpsMinimalTuple)
3344 24 : aggstate->ss.ps.outeropsfixed = false;
3345 : }
3346 :
3347 : /*
3348 : * Initialize result type, slot and projection.
3349 : */
3350 41238 : ExecInitResultTupleSlotTL(&aggstate->ss.ps, &TTSOpsVirtual);
3351 41238 : ExecAssignProjectionInfo(&aggstate->ss.ps, NULL);
3352 :
3353 : /*
3354 : * initialize child expressions
3355 : *
3356 : * We expect the parser to have checked that no aggs contain other agg
3357 : * calls in their arguments (and just to be sure, we verify it again while
3358 : * initializing the plan node). This would make no sense under SQL
3359 : * semantics, and it's forbidden by the spec. Because it is true, we
3360 : * don't need to worry about evaluating the aggs in any particular order.
3361 : *
3362 : * Note: execExpr.c finds Aggrefs for us, and adds them to aggstate->aggs.
3363 : * Aggrefs in the qual are found here; Aggrefs in the targetlist are found
3364 : * during ExecAssignProjectionInfo, above.
3365 : */
3366 41238 : aggstate->ss.ps.qual =
3367 41238 : ExecInitQual(node->plan.qual, (PlanState *) aggstate);
3368 :
3369 : /*
3370 : * We should now have found all Aggrefs in the targetlist and quals.
3371 : */
3372 41238 : numaggrefs = list_length(aggstate->aggs);
3373 41238 : max_aggno = -1;
3374 41238 : max_transno = -1;
3375 88094 : foreach(l, aggstate->aggs)
3376 : {
3377 46856 : Aggref *aggref = (Aggref *) lfirst(l);
3378 :
3379 46856 : max_aggno = Max(max_aggno, aggref->aggno);
3380 46856 : max_transno = Max(max_transno, aggref->aggtransno);
3381 : }
3382 41238 : numaggs = max_aggno + 1;
3383 41238 : numtrans = max_transno + 1;
3384 :
3385 : /*
3386 : * For each phase, prepare grouping set data and fmgr lookup data for
3387 : * compare functions. Accumulate all_grouped_cols in passing.
3388 : */
3389 41238 : aggstate->phases = palloc0(numPhases * sizeof(AggStatePerPhaseData));
3390 :
3391 41238 : aggstate->num_hashes = numHashes;
3392 41238 : if (numHashes)
3393 : {
3394 5108 : aggstate->perhash = palloc0(sizeof(AggStatePerHashData) * numHashes);
3395 5108 : aggstate->phases[0].numsets = 0;
3396 5108 : aggstate->phases[0].gset_lengths = palloc(numHashes * sizeof(int));
3397 5108 : aggstate->phases[0].grouped_cols = palloc(numHashes * sizeof(Bitmapset *));
3398 : }
3399 :
3400 41238 : phase = 0;
3401 83294 : for (phaseidx = 0; phaseidx <= list_length(node->chain); ++phaseidx)
3402 : {
3403 : Agg *aggnode;
3404 : Sort *sortnode;
3405 :
3406 42056 : if (phaseidx > 0)
3407 : {
3408 818 : aggnode = list_nth_node(Agg, node->chain, phaseidx - 1);
3409 818 : sortnode = castNode(Sort, outerPlan(aggnode));
3410 : }
3411 : else
3412 : {
3413 41238 : aggnode = node;
3414 41238 : sortnode = NULL;
3415 : }
3416 :
3417 : Assert(phase <= 1 || sortnode);
3418 :
3419 42056 : if (aggnode->aggstrategy == AGG_HASHED
3420 36750 : || aggnode->aggstrategy == AGG_MIXED)
3421 : {
3422 5502 : AggStatePerPhase phasedata = &aggstate->phases[0];
3423 : AggStatePerHash perhash;
3424 5502 : Bitmapset *cols = NULL;
3425 :
3426 : Assert(phase == 0);
3427 5502 : i = phasedata->numsets++;
3428 5502 : perhash = &aggstate->perhash[i];
3429 :
3430 : /* phase 0 always points to the "real" Agg in the hash case */
3431 5502 : phasedata->aggnode = node;
3432 5502 : phasedata->aggstrategy = node->aggstrategy;
3433 :
3434 : /* but the actual Agg node representing this hash is saved here */
3435 5502 : perhash->aggnode = aggnode;
3436 :
3437 5502 : phasedata->gset_lengths[i] = perhash->numCols = aggnode->numCols;
3438 :
3439 14044 : for (j = 0; j < aggnode->numCols; ++j)
3440 8542 : cols = bms_add_member(cols, aggnode->grpColIdx[j]);
3441 :
3442 5502 : phasedata->grouped_cols[i] = cols;
3443 :
3444 5502 : all_grouped_cols = bms_add_members(all_grouped_cols, cols);
3445 5502 : continue;
3446 : }
3447 : else
3448 : {
3449 36554 : AggStatePerPhase phasedata = &aggstate->phases[++phase];
3450 : int num_sets;
3451 :
3452 36554 : phasedata->numsets = num_sets = list_length(aggnode->groupingSets);
3453 :
3454 36554 : if (num_sets)
3455 : {
3456 862 : phasedata->gset_lengths = palloc(num_sets * sizeof(int));
3457 862 : phasedata->grouped_cols = palloc(num_sets * sizeof(Bitmapset *));
3458 :
3459 862 : i = 0;
3460 2624 : foreach(l, aggnode->groupingSets)
3461 : {
3462 1762 : int current_length = list_length(lfirst(l));
3463 1762 : Bitmapset *cols = NULL;
3464 :
3465 : /* planner forces this to be correct */
3466 3498 : for (j = 0; j < current_length; ++j)
3467 1736 : cols = bms_add_member(cols, aggnode->grpColIdx[j]);
3468 :
3469 1762 : phasedata->grouped_cols[i] = cols;
3470 1762 : phasedata->gset_lengths[i] = current_length;
3471 :
3472 1762 : ++i;
3473 : }
3474 :
3475 862 : all_grouped_cols = bms_add_members(all_grouped_cols,
3476 862 : phasedata->grouped_cols[0]);
3477 : }
3478 : else
3479 : {
3480 : Assert(phaseidx == 0);
3481 :
3482 35692 : phasedata->gset_lengths = NULL;
3483 35692 : phasedata->grouped_cols = NULL;
3484 : }
3485 :
3486 : /*
3487 : * If we are grouping, precompute fmgr lookup data for inner loop.
3488 : */
3489 36554 : if (aggnode->aggstrategy == AGG_SORTED)
3490 : {
3491 : /*
3492 : * Build a separate function for each subset of columns that
3493 : * need to be compared.
3494 : */
3495 2218 : phasedata->eqfunctions =
3496 2218 : (ExprState **) palloc0(aggnode->numCols * sizeof(ExprState *));
3497 :
3498 : /* for each grouping set */
3499 3738 : for (int k = 0; k < phasedata->numsets; k++)
3500 : {
3501 1520 : int length = phasedata->gset_lengths[k];
3502 :
3503 : /* nothing to do for empty grouping set */
3504 1520 : if (length == 0)
3505 326 : continue;
3506 :
3507 : /* if we already had one of this length, it'll do */
3508 1194 : if (phasedata->eqfunctions[length - 1] != NULL)
3509 138 : continue;
3510 :
3511 1056 : phasedata->eqfunctions[length - 1] =
3512 1056 : execTuplesMatchPrepare(scanDesc,
3513 : length,
3514 1056 : aggnode->grpColIdx,
3515 1056 : aggnode->grpOperators,
3516 1056 : aggnode->grpCollations,
3517 : (PlanState *) aggstate);
3518 : }
3519 :
3520 : /* and for all grouped columns, unless already computed */
3521 2218 : if (aggnode->numCols > 0 &&
3522 2124 : phasedata->eqfunctions[aggnode->numCols - 1] == NULL)
3523 : {
3524 1432 : phasedata->eqfunctions[aggnode->numCols - 1] =
3525 1432 : execTuplesMatchPrepare(scanDesc,
3526 : aggnode->numCols,
3527 1432 : aggnode->grpColIdx,
3528 1432 : aggnode->grpOperators,
3529 1432 : aggnode->grpCollations,
3530 : (PlanState *) aggstate);
3531 : }
3532 : }
3533 :
3534 36554 : phasedata->aggnode = aggnode;
3535 36554 : phasedata->aggstrategy = aggnode->aggstrategy;
3536 36554 : phasedata->sortnode = sortnode;
3537 : }
3538 : }
3539 :
3540 : /*
3541 : * Convert all_grouped_cols to a descending-order list.
3542 : */
3543 41238 : i = -1;
3544 50334 : while ((i = bms_next_member(all_grouped_cols, i)) >= 0)
3545 9096 : aggstate->all_grouped_cols = lcons_int(i, aggstate->all_grouped_cols);
3546 :
3547 : /*
3548 : * Set up aggregate-result storage in the output expr context, and also
3549 : * allocate my private per-agg working storage
3550 : */
3551 41238 : econtext = aggstate->ss.ps.ps_ExprContext;
3552 41238 : econtext->ecxt_aggvalues = (Datum *) palloc0(sizeof(Datum) * numaggs);
3553 41238 : econtext->ecxt_aggnulls = (bool *) palloc0(sizeof(bool) * numaggs);
3554 :
3555 41238 : peraggs = (AggStatePerAgg) palloc0(sizeof(AggStatePerAggData) * numaggs);
3556 41238 : pertransstates = (AggStatePerTrans) palloc0(sizeof(AggStatePerTransData) * numtrans);
3557 :
3558 41238 : aggstate->peragg = peraggs;
3559 41238 : aggstate->pertrans = pertransstates;
3560 :
3561 :
3562 41238 : aggstate->all_pergroups =
3563 41238 : (AggStatePerGroup *) palloc0(sizeof(AggStatePerGroup)
3564 41238 : * (numGroupingSets + numHashes));
3565 41238 : pergroups = aggstate->all_pergroups;
3566 :
3567 41238 : if (node->aggstrategy != AGG_HASHED)
3568 : {
3569 73480 : for (i = 0; i < numGroupingSets; i++)
3570 : {
3571 37154 : pergroups[i] = (AggStatePerGroup) palloc0(sizeof(AggStatePerGroupData)
3572 : * numaggs);
3573 : }
3574 :
3575 36326 : aggstate->pergroups = pergroups;
3576 36326 : pergroups += numGroupingSets;
3577 : }
3578 :
3579 : /*
3580 : * Hashing can only appear in the initial phase.
3581 : */
3582 41238 : if (use_hashing)
3583 : {
3584 5108 : Plan *outerplan = outerPlan(node);
3585 5108 : uint64 totalGroups = 0;
3586 :
3587 5108 : aggstate->hash_metacxt = AllocSetContextCreate(aggstate->ss.ps.state->es_query_cxt,
3588 : "HashAgg meta context",
3589 : ALLOCSET_DEFAULT_SIZES);
3590 5108 : aggstate->hash_spill_rslot = ExecInitExtraTupleSlot(estate, scanDesc,
3591 : &TTSOpsMinimalTuple);
3592 5108 : aggstate->hash_spill_wslot = ExecInitExtraTupleSlot(estate, scanDesc,
3593 : &TTSOpsVirtual);
3594 :
3595 : /* this is an array of pointers, not structures */
3596 5108 : aggstate->hash_pergroup = pergroups;
3597 :
3598 10216 : aggstate->hashentrysize = hash_agg_entry_size(aggstate->numtrans,
3599 5108 : outerplan->plan_width,
3600 : node->transitionSpace);
3601 :
3602 : /*
3603 : * Consider all of the grouping sets together when setting the limits
3604 : * and estimating the number of partitions. This can be inaccurate
3605 : * when there is more than one grouping set, but should still be
3606 : * reasonable.
3607 : */
3608 10610 : for (int k = 0; k < aggstate->num_hashes; k++)
3609 5502 : totalGroups += aggstate->perhash[k].aggnode->numGroups;
3610 :
3611 5108 : hash_agg_set_limits(aggstate->hashentrysize, totalGroups, 0,
3612 : &aggstate->hash_mem_limit,
3613 : &aggstate->hash_ngroups_limit,
3614 : &aggstate->hash_planned_partitions);
3615 5108 : find_hash_columns(aggstate);
3616 :
3617 : /* Skip massive memory allocation if we are just doing EXPLAIN */
3618 5108 : if (!(eflags & EXEC_FLAG_EXPLAIN_ONLY))
3619 3982 : build_hash_tables(aggstate);
3620 :
3621 5108 : aggstate->table_filled = false;
3622 :
3623 : /* Initialize this to 1, meaning nothing spilled, yet */
3624 5108 : aggstate->hash_batches_used = 1;
3625 : }
3626 :
3627 : /*
3628 : * Initialize current phase-dependent values to initial phase. The initial
3629 : * phase is 1 (first sort pass) for all strategies that use sorting (if
3630 : * hashing is being done too, then phase 0 is processed last); but if only
3631 : * hashing is being done, then phase 0 is all there is.
3632 : */
3633 41238 : if (node->aggstrategy == AGG_HASHED)
3634 : {
3635 4912 : aggstate->current_phase = 0;
3636 4912 : initialize_phase(aggstate, 0);
3637 4912 : select_current_set(aggstate, 0, true);
3638 : }
3639 : else
3640 : {
3641 36326 : aggstate->current_phase = 1;
3642 36326 : initialize_phase(aggstate, 1);
3643 36326 : select_current_set(aggstate, 0, false);
3644 : }
3645 :
3646 : /*
3647 : * Perform lookups of aggregate function info, and initialize the
3648 : * unchanging fields of the per-agg and per-trans data.
3649 : */
3650 88088 : foreach(l, aggstate->aggs)
3651 : {
3652 46856 : Aggref *aggref = lfirst(l);
3653 : AggStatePerAgg peragg;
3654 : AggStatePerTrans pertrans;
3655 : Oid aggTransFnInputTypes[FUNC_MAX_ARGS];
3656 : int numAggTransFnArgs;
3657 : int numDirectArgs;
3658 : HeapTuple aggTuple;
3659 : Form_pg_aggregate aggform;
3660 : AclResult aclresult;
3661 : Oid finalfn_oid;
3662 : Oid serialfn_oid,
3663 : deserialfn_oid;
3664 : Oid aggOwner;
3665 : Expr *finalfnexpr;
3666 : Oid aggtranstype;
3667 :
3668 : /* Planner should have assigned aggregate to correct level */
3669 : Assert(aggref->agglevelsup == 0);
3670 : /* ... and the split mode should match */
3671 : Assert(aggref->aggsplit == aggstate->aggsplit);
3672 :
3673 46856 : peragg = &peraggs[aggref->aggno];
3674 :
3675 : /* Check if we initialized the state for this aggregate already. */
3676 46856 : if (peragg->aggref != NULL)
3677 448 : continue;
3678 :
3679 46408 : peragg->aggref = aggref;
3680 46408 : peragg->transno = aggref->aggtransno;
3681 :
3682 : /* Fetch the pg_aggregate row */
3683 46408 : aggTuple = SearchSysCache1(AGGFNOID,
3684 : ObjectIdGetDatum(aggref->aggfnoid));
3685 46408 : if (!HeapTupleIsValid(aggTuple))
3686 0 : elog(ERROR, "cache lookup failed for aggregate %u",
3687 : aggref->aggfnoid);
3688 46408 : aggform = (Form_pg_aggregate) GETSTRUCT(aggTuple);
3689 :
3690 : /* Check permission to call aggregate function */
3691 46408 : aclresult = object_aclcheck(ProcedureRelationId, aggref->aggfnoid, GetUserId(),
3692 : ACL_EXECUTE);
3693 46408 : if (aclresult != ACLCHECK_OK)
3694 6 : aclcheck_error(aclresult, OBJECT_AGGREGATE,
3695 6 : get_func_name(aggref->aggfnoid));
3696 46402 : InvokeFunctionExecuteHook(aggref->aggfnoid);
3697 :
3698 : /* planner recorded transition state type in the Aggref itself */
3699 46402 : aggtranstype = aggref->aggtranstype;
3700 : Assert(OidIsValid(aggtranstype));
3701 :
3702 : /* Final function only required if we're finalizing the aggregates */
3703 46402 : if (DO_AGGSPLIT_SKIPFINAL(aggstate->aggsplit))
3704 4250 : peragg->finalfn_oid = finalfn_oid = InvalidOid;
3705 : else
3706 42152 : peragg->finalfn_oid = finalfn_oid = aggform->aggfinalfn;
3707 :
3708 46402 : serialfn_oid = InvalidOid;
3709 46402 : deserialfn_oid = InvalidOid;
3710 :
3711 : /*
3712 : * Check if serialization/deserialization is required. We only do it
3713 : * for aggregates that have transtype INTERNAL.
3714 : */
3715 46402 : if (aggtranstype == INTERNALOID)
3716 : {
3717 : /*
3718 : * The planner should only have generated a serialize agg node if
3719 : * every aggregate with an INTERNAL state has a serialization
3720 : * function. Verify that.
3721 : */
3722 18978 : if (DO_AGGSPLIT_SERIALIZE(aggstate->aggsplit))
3723 : {
3724 : /* serialization only valid when not running finalfn */
3725 : Assert(DO_AGGSPLIT_SKIPFINAL(aggstate->aggsplit));
3726 :
3727 336 : if (!OidIsValid(aggform->aggserialfn))
3728 0 : elog(ERROR, "serialfunc not provided for serialization aggregation");
3729 336 : serialfn_oid = aggform->aggserialfn;
3730 : }
3731 :
3732 : /* Likewise for deserialization functions */
3733 18978 : if (DO_AGGSPLIT_DESERIALIZE(aggstate->aggsplit))
3734 : {
3735 : /* deserialization only valid when combining states */
3736 : Assert(DO_AGGSPLIT_COMBINE(aggstate->aggsplit));
3737 :
3738 120 : if (!OidIsValid(aggform->aggdeserialfn))
3739 0 : elog(ERROR, "deserialfunc not provided for deserialization aggregation");
3740 120 : deserialfn_oid = aggform->aggdeserialfn;
3741 : }
3742 : }
3743 :
3744 : /* Check that aggregate owner has permission to call component fns */
3745 : {
3746 : HeapTuple procTuple;
3747 :
3748 46402 : procTuple = SearchSysCache1(PROCOID,
3749 : ObjectIdGetDatum(aggref->aggfnoid));
3750 46402 : if (!HeapTupleIsValid(procTuple))
3751 0 : elog(ERROR, "cache lookup failed for function %u",
3752 : aggref->aggfnoid);
3753 46402 : aggOwner = ((Form_pg_proc) GETSTRUCT(procTuple))->proowner;
3754 46402 : ReleaseSysCache(procTuple);
3755 :
3756 46402 : if (OidIsValid(finalfn_oid))
3757 : {
3758 20396 : aclresult = object_aclcheck(ProcedureRelationId, finalfn_oid, aggOwner,
3759 : ACL_EXECUTE);
3760 20396 : if (aclresult != ACLCHECK_OK)
3761 0 : aclcheck_error(aclresult, OBJECT_FUNCTION,
3762 0 : get_func_name(finalfn_oid));
3763 20396 : InvokeFunctionExecuteHook(finalfn_oid);
3764 : }
3765 46402 : if (OidIsValid(serialfn_oid))
3766 : {
3767 336 : aclresult = object_aclcheck(ProcedureRelationId, serialfn_oid, aggOwner,
3768 : ACL_EXECUTE);
3769 336 : if (aclresult != ACLCHECK_OK)
3770 0 : aclcheck_error(aclresult, OBJECT_FUNCTION,
3771 0 : get_func_name(serialfn_oid));
3772 336 : InvokeFunctionExecuteHook(serialfn_oid);
3773 : }
3774 46402 : if (OidIsValid(deserialfn_oid))
3775 : {
3776 120 : aclresult = object_aclcheck(ProcedureRelationId, deserialfn_oid, aggOwner,
3777 : ACL_EXECUTE);
3778 120 : if (aclresult != ACLCHECK_OK)
3779 0 : aclcheck_error(aclresult, OBJECT_FUNCTION,
3780 0 : get_func_name(deserialfn_oid));
3781 120 : InvokeFunctionExecuteHook(deserialfn_oid);
3782 : }
3783 : }
3784 :
3785 : /*
3786 : * Get actual datatypes of the (nominal) aggregate inputs. These
3787 : * could be different from the agg's declared input types, when the
3788 : * agg accepts ANY or a polymorphic type.
3789 : */
3790 46402 : numAggTransFnArgs = get_aggregate_argtypes(aggref,
3791 : aggTransFnInputTypes);
3792 :
3793 : /* Count the "direct" arguments, if any */
3794 46402 : numDirectArgs = list_length(aggref->aggdirectargs);
3795 :
3796 : /* Detect how many arguments to pass to the finalfn */
3797 46402 : if (aggform->aggfinalextra)
3798 13400 : peragg->numFinalArgs = numAggTransFnArgs + 1;
3799 : else
3800 33002 : peragg->numFinalArgs = numDirectArgs + 1;
3801 :
3802 : /* Initialize any direct-argument expressions */
3803 46402 : peragg->aggdirectargs = ExecInitExprList(aggref->aggdirectargs,
3804 : (PlanState *) aggstate);
3805 :
3806 : /*
3807 : * build expression trees using actual argument & result types for the
3808 : * finalfn, if it exists and is required.
3809 : */
3810 46402 : if (OidIsValid(finalfn_oid))
3811 : {
3812 20396 : build_aggregate_finalfn_expr(aggTransFnInputTypes,
3813 : peragg->numFinalArgs,
3814 : aggtranstype,
3815 : aggref->aggtype,
3816 : aggref->inputcollid,
3817 : finalfn_oid,
3818 : &finalfnexpr);
3819 20396 : fmgr_info(finalfn_oid, &peragg->finalfn);
3820 20396 : fmgr_info_set_expr((Node *) finalfnexpr, &peragg->finalfn);
3821 : }
3822 :
3823 : /* get info about the output value's datatype */
3824 46402 : get_typlenbyval(aggref->aggtype,
3825 : &peragg->resulttypeLen,
3826 : &peragg->resulttypeByVal);
3827 :
3828 : /*
3829 : * Build working state for invoking the transition function, if we
3830 : * haven't done it already.
3831 : */
3832 46402 : pertrans = &pertransstates[aggref->aggtransno];
3833 46402 : if (pertrans->aggref == NULL)
3834 : {
3835 : Datum textInitVal;
3836 : Datum initValue;
3837 : bool initValueIsNull;
3838 : Oid transfn_oid;
3839 :
3840 : /*
3841 : * If this aggregation is performing state combines, then instead
3842 : * of using the transition function, we'll use the combine
3843 : * function.
3844 : */
3845 46144 : if (DO_AGGSPLIT_COMBINE(aggstate->aggsplit))
3846 : {
3847 1358 : transfn_oid = aggform->aggcombinefn;
3848 :
3849 : /* If not set then the planner messed up */
3850 1358 : if (!OidIsValid(transfn_oid))
3851 0 : elog(ERROR, "combinefn not set for aggregate function");
3852 : }
3853 : else
3854 44786 : transfn_oid = aggform->aggtransfn;
3855 :
3856 46144 : aclresult = object_aclcheck(ProcedureRelationId, transfn_oid, aggOwner, ACL_EXECUTE);
3857 46144 : if (aclresult != ACLCHECK_OK)
3858 0 : aclcheck_error(aclresult, OBJECT_FUNCTION,
3859 0 : get_func_name(transfn_oid));
3860 46144 : InvokeFunctionExecuteHook(transfn_oid);
3861 :
3862 : /*
3863 : * initval is potentially null, so don't try to access it as a
3864 : * struct field. Must do it the hard way with SysCacheGetAttr.
3865 : */
3866 46144 : textInitVal = SysCacheGetAttr(AGGFNOID, aggTuple,
3867 : Anum_pg_aggregate_agginitval,
3868 : &initValueIsNull);
3869 46144 : if (initValueIsNull)
3870 27016 : initValue = (Datum) 0;
3871 : else
3872 19128 : initValue = GetAggInitVal(textInitVal, aggtranstype);
3873 :
3874 46144 : if (DO_AGGSPLIT_COMBINE(aggstate->aggsplit))
3875 : {
3876 1358 : Oid combineFnInputTypes[] = {aggtranstype,
3877 : aggtranstype};
3878 :
3879 : /*
3880 : * When combining there's only one input, the to-be-combined
3881 : * transition value. The transition value is not counted
3882 : * here.
3883 : */
3884 1358 : pertrans->numTransInputs = 1;
3885 :
3886 : /* aggcombinefn always has two arguments of aggtranstype */
3887 1358 : build_pertrans_for_aggref(pertrans, aggstate, estate,
3888 : aggref, transfn_oid, aggtranstype,
3889 : serialfn_oid, deserialfn_oid,
3890 : initValue, initValueIsNull,
3891 : combineFnInputTypes, 2);
3892 :
3893 : /*
3894 : * Ensure that a combine function to combine INTERNAL states
3895 : * is not strict. This should have been checked during CREATE
3896 : * AGGREGATE, but the strict property could have been changed
3897 : * since then.
3898 : */
3899 1358 : if (pertrans->transfn.fn_strict && aggtranstype == INTERNALOID)
3900 0 : ereport(ERROR,
3901 : (errcode(ERRCODE_INVALID_FUNCTION_DEFINITION),
3902 : errmsg("combine function with transition type %s must not be declared STRICT",
3903 : format_type_be(aggtranstype))));
3904 : }
3905 : else
3906 : {
3907 : /* Detect how many arguments to pass to the transfn */
3908 44786 : if (AGGKIND_IS_ORDERED_SET(aggref->aggkind))
3909 252 : pertrans->numTransInputs = list_length(aggref->args);
3910 : else
3911 44534 : pertrans->numTransInputs = numAggTransFnArgs;
3912 :
3913 44786 : build_pertrans_for_aggref(pertrans, aggstate, estate,
3914 : aggref, transfn_oid, aggtranstype,
3915 : serialfn_oid, deserialfn_oid,
3916 : initValue, initValueIsNull,
3917 : aggTransFnInputTypes,
3918 : numAggTransFnArgs);
3919 :
3920 : /*
3921 : * If the transfn is strict and the initval is NULL, make sure
3922 : * input type and transtype are the same (or at least
3923 : * binary-compatible), so that it's OK to use the first
3924 : * aggregated input value as the initial transValue. This
3925 : * should have been checked at agg definition time, but we
3926 : * must check again in case the transfn's strictness property
3927 : * has been changed.
3928 : */
3929 44786 : if (pertrans->transfn.fn_strict && pertrans->initValueIsNull)
3930 : {
3931 4834 : if (numAggTransFnArgs <= numDirectArgs ||
3932 4834 : !IsBinaryCoercible(aggTransFnInputTypes[numDirectArgs],
3933 : aggtranstype))
3934 0 : ereport(ERROR,
3935 : (errcode(ERRCODE_INVALID_FUNCTION_DEFINITION),
3936 : errmsg("aggregate %u needs to have compatible input type and transition type",
3937 : aggref->aggfnoid)));
3938 : }
3939 : }
3940 : }
3941 : else
3942 258 : pertrans->aggshared = true;
3943 46402 : ReleaseSysCache(aggTuple);
3944 : }
3945 :
3946 : /*
3947 : * Update aggstate->numaggs to be the number of unique aggregates found.
3948 : * Also set numstates to the number of unique transition states found.
3949 : */
3950 41232 : aggstate->numaggs = numaggs;
3951 41232 : aggstate->numtrans = numtrans;
3952 :
3953 : /*
3954 : * Last, check whether any more aggregates got added onto the node while
3955 : * we processed the expressions for the aggregate arguments (including not
3956 : * only the regular arguments and FILTER expressions handled immediately
3957 : * above, but any direct arguments we might've handled earlier). If so,
3958 : * we have nested aggregate functions, which is semantically nonsensical,
3959 : * so complain. (This should have been caught by the parser, so we don't
3960 : * need to work hard on a helpful error message; but we defend against it
3961 : * here anyway, just to be sure.)
3962 : */
3963 41232 : if (numaggrefs != list_length(aggstate->aggs))
3964 0 : ereport(ERROR,
3965 : (errcode(ERRCODE_GROUPING_ERROR),
3966 : errmsg("aggregate function calls cannot be nested")));
3967 :
3968 : /*
3969 : * Build expressions doing all the transition work at once. We build a
3970 : * different one for each phase, as the number of transition function
3971 : * invocation can differ between phases. Note this'll work both for
3972 : * transition and combination functions (although there'll only be one
3973 : * phase in the latter case).
3974 : */
3975 119012 : for (phaseidx = 0; phaseidx < aggstate->numphases; phaseidx++)
3976 : {
3977 77780 : AggStatePerPhase phase = &aggstate->phases[phaseidx];
3978 77780 : bool dohash = false;
3979 77780 : bool dosort = false;
3980 :
3981 : /* phase 0 doesn't necessarily exist */
3982 77780 : if (!phase->aggnode)
3983 36124 : continue;
3984 :
3985 41656 : if (aggstate->aggstrategy == AGG_MIXED && phaseidx == 1)
3986 : {
3987 : /*
3988 : * Phase one, and only phase one, in a mixed agg performs both
3989 : * sorting and aggregation.
3990 : */
3991 196 : dohash = true;
3992 196 : dosort = true;
3993 : }
3994 41460 : else if (aggstate->aggstrategy == AGG_MIXED && phaseidx == 0)
3995 : {
3996 : /*
3997 : * No need to compute a transition function for an AGG_MIXED phase
3998 : * 0 - the contents of the hashtables will have been computed
3999 : * during phase 1.
4000 : */
4001 196 : continue;
4002 : }
4003 41264 : else if (phase->aggstrategy == AGG_PLAIN ||
4004 7068 : phase->aggstrategy == AGG_SORTED)
4005 : {
4006 36352 : dohash = false;
4007 36352 : dosort = true;
4008 : }
4009 4912 : else if (phase->aggstrategy == AGG_HASHED)
4010 : {
4011 4912 : dohash = true;
4012 4912 : dosort = false;
4013 : }
4014 : else
4015 : Assert(false);
4016 :
4017 41460 : phase->evaltrans = ExecBuildAggTrans(aggstate, phase, dosort, dohash,
4018 : false);
4019 :
4020 : /* cache compiled expression for outer slot without NULL check */
4021 41460 : phase->evaltrans_cache[0][0] = phase->evaltrans;
4022 : }
4023 :
4024 41232 : return aggstate;
4025 : }
4026 :
4027 : /*
4028 : * Build the state needed to calculate a state value for an aggregate.
4029 : *
4030 : * This initializes all the fields in 'pertrans'. 'aggref' is the aggregate
4031 : * to initialize the state for. 'transfn_oid', 'aggtranstype', and the rest
4032 : * of the arguments could be calculated from 'aggref', but the caller has
4033 : * calculated them already, so might as well pass them.
4034 : *
4035 : * 'transfn_oid' may be either the Oid of the aggtransfn or the aggcombinefn.
4036 : */
4037 : static void
4038 46144 : build_pertrans_for_aggref(AggStatePerTrans pertrans,
4039 : AggState *aggstate, EState *estate,
4040 : Aggref *aggref,
4041 : Oid transfn_oid, Oid aggtranstype,
4042 : Oid aggserialfn, Oid aggdeserialfn,
4043 : Datum initValue, bool initValueIsNull,
4044 : Oid *inputTypes, int numArguments)
4045 : {
4046 46144 : int numGroupingSets = Max(aggstate->maxsets, 1);
4047 : Expr *transfnexpr;
4048 : int numTransArgs;
4049 46144 : Expr *serialfnexpr = NULL;
4050 46144 : Expr *deserialfnexpr = NULL;
4051 : ListCell *lc;
4052 : int numInputs;
4053 : int numDirectArgs;
4054 : List *sortlist;
4055 : int numSortCols;
4056 : int numDistinctCols;
4057 : int i;
4058 :
4059 : /* Begin filling in the pertrans data */
4060 46144 : pertrans->aggref = aggref;
4061 46144 : pertrans->aggshared = false;
4062 46144 : pertrans->aggCollation = aggref->inputcollid;
4063 46144 : pertrans->transfn_oid = transfn_oid;
4064 46144 : pertrans->serialfn_oid = aggserialfn;
4065 46144 : pertrans->deserialfn_oid = aggdeserialfn;
4066 46144 : pertrans->initValue = initValue;
4067 46144 : pertrans->initValueIsNull = initValueIsNull;
4068 :
4069 : /* Count the "direct" arguments, if any */
4070 46144 : numDirectArgs = list_length(aggref->aggdirectargs);
4071 :
4072 : /* Count the number of aggregated input columns */
4073 46144 : pertrans->numInputs = numInputs = list_length(aggref->args);
4074 :
4075 46144 : pertrans->aggtranstype = aggtranstype;
4076 :
4077 : /* account for the current transition state */
4078 46144 : numTransArgs = pertrans->numTransInputs + 1;
4079 :
4080 : /*
4081 : * Set up infrastructure for calling the transfn. Note that invtransfn is
4082 : * not needed here.
4083 : */
4084 46144 : build_aggregate_transfn_expr(inputTypes,
4085 : numArguments,
4086 : numDirectArgs,
4087 46144 : aggref->aggvariadic,
4088 : aggtranstype,
4089 : aggref->inputcollid,
4090 : transfn_oid,
4091 : InvalidOid,
4092 : &transfnexpr,
4093 : NULL);
4094 :
4095 46144 : fmgr_info(transfn_oid, &pertrans->transfn);
4096 46144 : fmgr_info_set_expr((Node *) transfnexpr, &pertrans->transfn);
4097 :
4098 46144 : pertrans->transfn_fcinfo =
4099 46144 : (FunctionCallInfo) palloc(SizeForFunctionCallInfo(numTransArgs));
4100 46144 : InitFunctionCallInfoData(*pertrans->transfn_fcinfo,
4101 : &pertrans->transfn,
4102 : numTransArgs,
4103 : pertrans->aggCollation,
4104 : (void *) aggstate, NULL);
4105 :
4106 : /* get info about the state value's datatype */
4107 46144 : get_typlenbyval(aggtranstype,
4108 : &pertrans->transtypeLen,
4109 : &pertrans->transtypeByVal);
4110 :
4111 46144 : if (OidIsValid(aggserialfn))
4112 : {
4113 336 : build_aggregate_serialfn_expr(aggserialfn,
4114 : &serialfnexpr);
4115 336 : fmgr_info(aggserialfn, &pertrans->serialfn);
4116 336 : fmgr_info_set_expr((Node *) serialfnexpr, &pertrans->serialfn);
4117 :
4118 336 : pertrans->serialfn_fcinfo =
4119 336 : (FunctionCallInfo) palloc(SizeForFunctionCallInfo(1));
4120 336 : InitFunctionCallInfoData(*pertrans->serialfn_fcinfo,
4121 : &pertrans->serialfn,
4122 : 1,
4123 : InvalidOid,
4124 : (void *) aggstate, NULL);
4125 : }
4126 :
4127 46144 : if (OidIsValid(aggdeserialfn))
4128 : {
4129 120 : build_aggregate_deserialfn_expr(aggdeserialfn,
4130 : &deserialfnexpr);
4131 120 : fmgr_info(aggdeserialfn, &pertrans->deserialfn);
4132 120 : fmgr_info_set_expr((Node *) deserialfnexpr, &pertrans->deserialfn);
4133 :
4134 120 : pertrans->deserialfn_fcinfo =
4135 120 : (FunctionCallInfo) palloc(SizeForFunctionCallInfo(2));
4136 120 : InitFunctionCallInfoData(*pertrans->deserialfn_fcinfo,
4137 : &pertrans->deserialfn,
4138 : 2,
4139 : InvalidOid,
4140 : (void *) aggstate, NULL);
4141 : }
4142 :
4143 : /*
4144 : * If we're doing either DISTINCT or ORDER BY for a plain agg, then we
4145 : * have a list of SortGroupClause nodes; fish out the data in them and
4146 : * stick them into arrays. We ignore ORDER BY for an ordered-set agg,
4147 : * however; the agg's transfn and finalfn are responsible for that.
4148 : *
4149 : * When the planner has set the aggpresorted flag, the input to the
4150 : * aggregate is already correctly sorted. For ORDER BY aggregates we can
4151 : * simply treat these as normal aggregates. For presorted DISTINCT
4152 : * aggregates an extra step must be added to remove duplicate consecutive
4153 : * inputs.
4154 : *
4155 : * Note that by construction, if there is a DISTINCT clause then the ORDER
4156 : * BY clause is a prefix of it (see transformDistinctClause).
4157 : */
4158 46144 : if (AGGKIND_IS_ORDERED_SET(aggref->aggkind))
4159 : {
4160 252 : sortlist = NIL;
4161 252 : numSortCols = numDistinctCols = 0;
4162 252 : pertrans->aggsortrequired = false;
4163 : }
4164 45892 : else if (aggref->aggpresorted && aggref->aggdistinct == NIL)
4165 : {
4166 1510 : sortlist = NIL;
4167 1510 : numSortCols = numDistinctCols = 0;
4168 1510 : pertrans->aggsortrequired = false;
4169 : }
4170 44382 : else if (aggref->aggdistinct)
4171 : {
4172 558 : sortlist = aggref->aggdistinct;
4173 558 : numSortCols = numDistinctCols = list_length(sortlist);
4174 : Assert(numSortCols >= list_length(aggref->aggorder));
4175 558 : pertrans->aggsortrequired = !aggref->aggpresorted;
4176 : }
4177 : else
4178 : {
4179 43824 : sortlist = aggref->aggorder;
4180 43824 : numSortCols = list_length(sortlist);
4181 43824 : numDistinctCols = 0;
4182 43824 : pertrans->aggsortrequired = (numSortCols > 0);
4183 : }
4184 :
4185 46144 : pertrans->numSortCols = numSortCols;
4186 46144 : pertrans->numDistinctCols = numDistinctCols;
4187 :
4188 : /*
4189 : * If we have either sorting or filtering to do, create a tupledesc and
4190 : * slot corresponding to the aggregated inputs (including sort
4191 : * expressions) of the agg.
4192 : */
4193 46144 : if (numSortCols > 0 || aggref->aggfilter)
4194 : {
4195 1392 : pertrans->sortdesc = ExecTypeFromTL(aggref->args);
4196 1392 : pertrans->sortslot =
4197 1392 : ExecInitExtraTupleSlot(estate, pertrans->sortdesc,
4198 : &TTSOpsMinimalTuple);
4199 : }
4200 :
4201 46144 : if (numSortCols > 0)
4202 : {
4203 : /*
4204 : * We don't implement DISTINCT or ORDER BY aggs in the HASHED case
4205 : * (yet)
4206 : */
4207 : Assert(aggstate->aggstrategy != AGG_HASHED && aggstate->aggstrategy != AGG_MIXED);
4208 :
4209 : /* ORDER BY aggregates are not supported with partial aggregation */
4210 : Assert(!DO_AGGSPLIT_COMBINE(aggstate->aggsplit));
4211 :
4212 : /* If we have only one input, we need its len/byval info. */
4213 684 : if (numInputs == 1)
4214 : {
4215 558 : get_typlenbyval(inputTypes[numDirectArgs],
4216 : &pertrans->inputtypeLen,
4217 : &pertrans->inputtypeByVal);
4218 : }
4219 126 : else if (numDistinctCols > 0)
4220 : {
4221 : /* we will need an extra slot to store prior values */
4222 96 : pertrans->uniqslot =
4223 96 : ExecInitExtraTupleSlot(estate, pertrans->sortdesc,
4224 : &TTSOpsMinimalTuple);
4225 : }
4226 :
4227 : /* Extract the sort information for use later */
4228 684 : pertrans->sortColIdx =
4229 684 : (AttrNumber *) palloc(numSortCols * sizeof(AttrNumber));
4230 684 : pertrans->sortOperators =
4231 684 : (Oid *) palloc(numSortCols * sizeof(Oid));
4232 684 : pertrans->sortCollations =
4233 684 : (Oid *) palloc(numSortCols * sizeof(Oid));
4234 684 : pertrans->sortNullsFirst =
4235 684 : (bool *) palloc(numSortCols * sizeof(bool));
4236 :
4237 684 : i = 0;
4238 1554 : foreach(lc, sortlist)
4239 : {
4240 870 : SortGroupClause *sortcl = (SortGroupClause *) lfirst(lc);
4241 870 : TargetEntry *tle = get_sortgroupclause_tle(sortcl, aggref->args);
4242 :
4243 : /* the parser should have made sure of this */
4244 : Assert(OidIsValid(sortcl->sortop));
4245 :
4246 870 : pertrans->sortColIdx[i] = tle->resno;
4247 870 : pertrans->sortOperators[i] = sortcl->sortop;
4248 870 : pertrans->sortCollations[i] = exprCollation((Node *) tle->expr);
4249 870 : pertrans->sortNullsFirst[i] = sortcl->nulls_first;
4250 870 : i++;
4251 : }
4252 : Assert(i == numSortCols);
4253 : }
4254 :
4255 46144 : if (aggref->aggdistinct)
4256 : {
4257 : Oid *ops;
4258 :
4259 : Assert(numArguments > 0);
4260 : Assert(list_length(aggref->aggdistinct) == numDistinctCols);
4261 :
4262 558 : ops = palloc(numDistinctCols * sizeof(Oid));
4263 :
4264 558 : i = 0;
4265 1284 : foreach(lc, aggref->aggdistinct)
4266 726 : ops[i++] = ((SortGroupClause *) lfirst(lc))->eqop;
4267 :
4268 : /* lookup / build the necessary comparators */
4269 558 : if (numDistinctCols == 1)
4270 462 : fmgr_info(get_opcode(ops[0]), &pertrans->equalfnOne);
4271 : else
4272 96 : pertrans->equalfnMulti =
4273 96 : execTuplesMatchPrepare(pertrans->sortdesc,
4274 : numDistinctCols,
4275 96 : pertrans->sortColIdx,
4276 : ops,
4277 96 : pertrans->sortCollations,
4278 : &aggstate->ss.ps);
4279 558 : pfree(ops);
4280 : }
4281 :
4282 46144 : pertrans->sortstates = (Tuplesortstate **)
4283 46144 : palloc0(sizeof(Tuplesortstate *) * numGroupingSets);
4284 46144 : }
4285 :
4286 :
4287 : static Datum
4288 19128 : GetAggInitVal(Datum textInitVal, Oid transtype)
4289 : {
4290 : Oid typinput,
4291 : typioparam;
4292 : char *strInitVal;
4293 : Datum initVal;
4294 :
4295 19128 : getTypeInputInfo(transtype, &typinput, &typioparam);
4296 19128 : strInitVal = TextDatumGetCString(textInitVal);
4297 19128 : initVal = OidInputFunctionCall(typinput, strInitVal,
4298 : typioparam, -1);
4299 19128 : pfree(strInitVal);
4300 19128 : return initVal;
4301 : }
4302 :
4303 : void
4304 41126 : ExecEndAgg(AggState *node)
4305 : {
4306 : PlanState *outerPlan;
4307 : int transno;
4308 41126 : int numGroupingSets = Max(node->maxsets, 1);
4309 : int setno;
4310 :
4311 : /*
4312 : * When ending a parallel worker, copy the statistics gathered by the
4313 : * worker back into shared memory so that it can be picked up by the main
4314 : * process to report in EXPLAIN ANALYZE.
4315 : */
4316 41126 : if (node->shared_info && IsParallelWorker())
4317 : {
4318 : AggregateInstrumentation *si;
4319 :
4320 : Assert(ParallelWorkerNumber <= node->shared_info->num_workers);
4321 174 : si = &node->shared_info->sinstrument[ParallelWorkerNumber];
4322 174 : si->hash_batches_used = node->hash_batches_used;
4323 174 : si->hash_disk_used = node->hash_disk_used;
4324 174 : si->hash_mem_peak = node->hash_mem_peak;
4325 : }
4326 :
4327 : /* Make sure we have closed any open tuplesorts */
4328 :
4329 41126 : if (node->sort_in)
4330 144 : tuplesort_end(node->sort_in);
4331 41126 : if (node->sort_out)
4332 42 : tuplesort_end(node->sort_out);
4333 :
4334 41126 : hashagg_reset_spill_state(node);
4335 :
4336 41126 : if (node->hash_metacxt != NULL)
4337 : {
4338 5100 : MemoryContextDelete(node->hash_metacxt);
4339 5100 : node->hash_metacxt = NULL;
4340 : }
4341 :
4342 87160 : for (transno = 0; transno < node->numtrans; transno++)
4343 : {
4344 46034 : AggStatePerTrans pertrans = &node->pertrans[transno];
4345 :
4346 93070 : for (setno = 0; setno < numGroupingSets; setno++)
4347 : {
4348 47036 : if (pertrans->sortstates[setno])
4349 0 : tuplesort_end(pertrans->sortstates[setno]);
4350 : }
4351 : }
4352 :
4353 : /* And ensure any agg shutdown callbacks have been called */
4354 83080 : for (setno = 0; setno < numGroupingSets; setno++)
4355 41954 : ReScanExprContext(node->aggcontexts[setno]);
4356 41126 : if (node->hashcontext)
4357 5100 : ReScanExprContext(node->hashcontext);
4358 :
4359 41126 : outerPlan = outerPlanState(node);
4360 41126 : ExecEndNode(outerPlan);
4361 41126 : }
4362 :
4363 : void
4364 36984 : ExecReScanAgg(AggState *node)
4365 : {
4366 36984 : ExprContext *econtext = node->ss.ps.ps_ExprContext;
4367 36984 : PlanState *outerPlan = outerPlanState(node);
4368 36984 : Agg *aggnode = (Agg *) node->ss.ps.plan;
4369 : int transno;
4370 36984 : int numGroupingSets = Max(node->maxsets, 1);
4371 : int setno;
4372 :
4373 36984 : node->agg_done = false;
4374 :
4375 36984 : if (node->aggstrategy == AGG_HASHED)
4376 : {
4377 : /*
4378 : * In the hashed case, if we haven't yet built the hash table then we
4379 : * can just return; nothing done yet, so nothing to undo. If subnode's
4380 : * chgParam is not NULL then it will be re-scanned by ExecProcNode,
4381 : * else no reason to re-scan it at all.
4382 : */
4383 1088 : if (!node->table_filled)
4384 38 : return;
4385 :
4386 : /*
4387 : * If we do have the hash table, and it never spilled, and the subplan
4388 : * does not have any parameter changes, and none of our own parameter
4389 : * changes affect input expressions of the aggregated functions, then
4390 : * we can just rescan the existing hash table; no need to build it
4391 : * again.
4392 : */
4393 1050 : if (outerPlan->chgParam == NULL && !node->hash_ever_spilled &&
4394 1026 : !bms_overlap(node->ss.ps.chgParam, aggnode->aggParams))
4395 : {
4396 1002 : ResetTupleHashIterator(node->perhash[0].hashtable,
4397 : &node->perhash[0].hashiter);
4398 1002 : select_current_set(node, 0, true);
4399 1002 : return;
4400 : }
4401 : }
4402 :
4403 : /* Make sure we have closed any open tuplesorts */
4404 98912 : for (transno = 0; transno < node->numtrans; transno++)
4405 : {
4406 125972 : for (setno = 0; setno < numGroupingSets; setno++)
4407 : {
4408 63004 : AggStatePerTrans pertrans = &node->pertrans[transno];
4409 :
4410 63004 : if (pertrans->sortstates[setno])
4411 : {
4412 0 : tuplesort_end(pertrans->sortstates[setno]);
4413 0 : pertrans->sortstates[setno] = NULL;
4414 : }
4415 : }
4416 : }
4417 :
4418 : /*
4419 : * We don't need to ReScanExprContext the output tuple context here;
4420 : * ExecReScan already did it. But we do need to reset our per-grouping-set
4421 : * contexts, which may have transvalues stored in them. (We use rescan
4422 : * rather than just reset because transfns may have registered callbacks
4423 : * that need to be run now.) For the AGG_HASHED case, see below.
4424 : */
4425 :
4426 71924 : for (setno = 0; setno < numGroupingSets; setno++)
4427 : {
4428 35980 : ReScanExprContext(node->aggcontexts[setno]);
4429 : }
4430 :
4431 : /* Release first tuple of group, if we have made a copy */
4432 35944 : if (node->grp_firstTuple != NULL)
4433 : {
4434 0 : heap_freetuple(node->grp_firstTuple);
4435 0 : node->grp_firstTuple = NULL;
4436 : }
4437 35944 : ExecClearTuple(node->ss.ss_ScanTupleSlot);
4438 :
4439 : /* Forget current agg values */
4440 98912 : MemSet(econtext->ecxt_aggvalues, 0, sizeof(Datum) * node->numaggs);
4441 35944 : MemSet(econtext->ecxt_aggnulls, 0, sizeof(bool) * node->numaggs);
4442 :
4443 : /*
4444 : * With AGG_HASHED/MIXED, the hash table is allocated in a sub-context of
4445 : * the hashcontext. This used to be an issue, but now, resetting a context
4446 : * automatically deletes sub-contexts too.
4447 : */
4448 35944 : if (node->aggstrategy == AGG_HASHED || node->aggstrategy == AGG_MIXED)
4449 : {
4450 78 : hashagg_reset_spill_state(node);
4451 :
4452 78 : node->hash_ever_spilled = false;
4453 78 : node->hash_spill_mode = false;
4454 78 : node->hash_ngroups_current = 0;
4455 :
4456 78 : ReScanExprContext(node->hashcontext);
4457 : /* Rebuild an empty hash table */
4458 78 : build_hash_tables(node);
4459 78 : node->table_filled = false;
4460 : /* iterator will be reset when the table is filled */
4461 :
4462 78 : hashagg_recompile_expressions(node, false, false);
4463 : }
4464 :
4465 35944 : if (node->aggstrategy != AGG_HASHED)
4466 : {
4467 : /*
4468 : * Reset the per-group state (in particular, mark transvalues null)
4469 : */
4470 71828 : for (setno = 0; setno < numGroupingSets; setno++)
4471 : {
4472 161844 : MemSet(node->pergroups[setno], 0,
4473 : sizeof(AggStatePerGroupData) * node->numaggs);
4474 : }
4475 :
4476 : /* reset to phase 1 */
4477 35896 : initialize_phase(node, 1);
4478 :
4479 35896 : node->input_done = false;
4480 35896 : node->projected_set = -1;
4481 : }
4482 :
4483 35944 : if (outerPlan->chgParam == NULL)
4484 160 : ExecReScan(outerPlan);
4485 : }
4486 :
4487 :
4488 : /***********************************************************************
4489 : * API exposed to aggregate functions
4490 : ***********************************************************************/
4491 :
4492 :
4493 : /*
4494 : * AggCheckCallContext - test if a SQL function is being called as an aggregate
4495 : *
4496 : * The transition and/or final functions of an aggregate may want to verify
4497 : * that they are being called as aggregates, rather than as plain SQL
4498 : * functions. They should use this function to do so. The return value
4499 : * is nonzero if being called as an aggregate, or zero if not. (Specific
4500 : * nonzero values are AGG_CONTEXT_AGGREGATE or AGG_CONTEXT_WINDOW, but more
4501 : * values could conceivably appear in future.)
4502 : *
4503 : * If aggcontext isn't NULL, the function also stores at *aggcontext the
4504 : * identity of the memory context that aggregate transition values are being
4505 : * stored in. Note that the same aggregate call site (flinfo) may be called
4506 : * interleaved on different transition values in different contexts, so it's
4507 : * not kosher to cache aggcontext under fn_extra. It is, however, kosher to
4508 : * cache it in the transvalue itself (for internal-type transvalues).
4509 : */
4510 : int
4511 4438458 : AggCheckCallContext(FunctionCallInfo fcinfo, MemoryContext *aggcontext)
4512 : {
4513 4438458 : if (fcinfo->context && IsA(fcinfo->context, AggState))
4514 : {
4515 4427078 : if (aggcontext)
4516 : {
4517 1839450 : AggState *aggstate = ((AggState *) fcinfo->context);
4518 1839450 : ExprContext *cxt = aggstate->curaggcontext;
4519 :
4520 1839450 : *aggcontext = cxt->ecxt_per_tuple_memory;
4521 : }
4522 4427078 : return AGG_CONTEXT_AGGREGATE;
4523 : }
4524 11380 : if (fcinfo->context && IsA(fcinfo->context, WindowAggState))
4525 : {
4526 9512 : if (aggcontext)
4527 710 : *aggcontext = ((WindowAggState *) fcinfo->context)->curaggcontext;
4528 9512 : return AGG_CONTEXT_WINDOW;
4529 : }
4530 :
4531 : /* this is just to prevent "uninitialized variable" warnings */
4532 1868 : if (aggcontext)
4533 1820 : *aggcontext = NULL;
4534 1868 : return 0;
4535 : }
4536 :
4537 : /*
4538 : * AggGetAggref - allow an aggregate support function to get its Aggref
4539 : *
4540 : * If the function is being called as an aggregate support function,
4541 : * return the Aggref node for the aggregate call. Otherwise, return NULL.
4542 : *
4543 : * Aggregates sharing the same inputs and transition functions can get
4544 : * merged into a single transition calculation. If the transition function
4545 : * calls AggGetAggref, it will get some one of the Aggrefs for which it is
4546 : * executing. It must therefore not pay attention to the Aggref fields that
4547 : * relate to the final function, as those are indeterminate. But if a final
4548 : * function calls AggGetAggref, it will get a precise result.
4549 : *
4550 : * Note that if an aggregate is being used as a window function, this will
4551 : * return NULL. We could provide a similar function to return the relevant
4552 : * WindowFunc node in such cases, but it's not needed yet.
4553 : */
4554 : Aggref *
4555 246 : AggGetAggref(FunctionCallInfo fcinfo)
4556 : {
4557 246 : if (fcinfo->context && IsA(fcinfo->context, AggState))
4558 : {
4559 246 : AggState *aggstate = (AggState *) fcinfo->context;
4560 : AggStatePerAgg curperagg;
4561 : AggStatePerTrans curpertrans;
4562 :
4563 : /* check curperagg (valid when in a final function) */
4564 246 : curperagg = aggstate->curperagg;
4565 :
4566 246 : if (curperagg)
4567 0 : return curperagg->aggref;
4568 :
4569 : /* check curpertrans (valid when in a transition function) */
4570 246 : curpertrans = aggstate->curpertrans;
4571 :
4572 246 : if (curpertrans)
4573 246 : return curpertrans->aggref;
4574 : }
4575 0 : return NULL;
4576 : }
4577 :
4578 : /*
4579 : * AggGetTempMemoryContext - fetch short-term memory context for aggregates
4580 : *
4581 : * This is useful in agg final functions; the context returned is one that
4582 : * the final function can safely reset as desired. This isn't useful for
4583 : * transition functions, since the context returned MAY (we don't promise)
4584 : * be the same as the context those are called in.
4585 : *
4586 : * As above, this is currently not useful for aggs called as window functions.
4587 : */
4588 : MemoryContext
4589 0 : AggGetTempMemoryContext(FunctionCallInfo fcinfo)
4590 : {
4591 0 : if (fcinfo->context && IsA(fcinfo->context, AggState))
4592 : {
4593 0 : AggState *aggstate = (AggState *) fcinfo->context;
4594 :
4595 0 : return aggstate->tmpcontext->ecxt_per_tuple_memory;
4596 : }
4597 0 : return NULL;
4598 : }
4599 :
4600 : /*
4601 : * AggStateIsShared - find out whether transition state is shared
4602 : *
4603 : * If the function is being called as an aggregate support function,
4604 : * return true if the aggregate's transition state is shared across
4605 : * multiple aggregates, false if it is not.
4606 : *
4607 : * Returns true if not called as an aggregate support function.
4608 : * This is intended as a conservative answer, ie "no you'd better not
4609 : * scribble on your input". In particular, will return true if the
4610 : * aggregate is being used as a window function, which is a scenario
4611 : * in which changing the transition state is a bad idea. We might
4612 : * want to refine the behavior for the window case in future.
4613 : */
4614 : bool
4615 246 : AggStateIsShared(FunctionCallInfo fcinfo)
4616 : {
4617 246 : if (fcinfo->context && IsA(fcinfo->context, AggState))
4618 : {
4619 246 : AggState *aggstate = (AggState *) fcinfo->context;
4620 : AggStatePerAgg curperagg;
4621 : AggStatePerTrans curpertrans;
4622 :
4623 : /* check curperagg (valid when in a final function) */
4624 246 : curperagg = aggstate->curperagg;
4625 :
4626 246 : if (curperagg)
4627 0 : return aggstate->pertrans[curperagg->transno].aggshared;
4628 :
4629 : /* check curpertrans (valid when in a transition function) */
4630 246 : curpertrans = aggstate->curpertrans;
4631 :
4632 246 : if (curpertrans)
4633 246 : return curpertrans->aggshared;
4634 : }
4635 0 : return true;
4636 : }
4637 :
4638 : /*
4639 : * AggRegisterCallback - register a cleanup callback for an aggregate
4640 : *
4641 : * This is useful for aggs to register shutdown callbacks, which will ensure
4642 : * that non-memory resources are freed. The callback will occur just before
4643 : * the associated aggcontext (as returned by AggCheckCallContext) is reset,
4644 : * either between groups or as a result of rescanning the query. The callback
4645 : * will NOT be called on error paths. The typical use-case is for freeing of
4646 : * tuplestores or tuplesorts maintained in aggcontext, or pins held by slots
4647 : * created by the agg functions. (The callback will not be called until after
4648 : * the result of the finalfn is no longer needed, so it's safe for the finalfn
4649 : * to return data that will be freed by the callback.)
4650 : *
4651 : * As above, this is currently not useful for aggs called as window functions.
4652 : */
4653 : void
4654 660 : AggRegisterCallback(FunctionCallInfo fcinfo,
4655 : ExprContextCallbackFunction func,
4656 : Datum arg)
4657 : {
4658 660 : if (fcinfo->context && IsA(fcinfo->context, AggState))
4659 : {
4660 660 : AggState *aggstate = (AggState *) fcinfo->context;
4661 660 : ExprContext *cxt = aggstate->curaggcontext;
4662 :
4663 660 : RegisterExprContextCallback(cxt, func, arg);
4664 :
4665 660 : return;
4666 : }
4667 0 : elog(ERROR, "aggregate function cannot register a callback in this context");
4668 : }
4669 :
4670 :
4671 : /* ----------------------------------------------------------------
4672 : * Parallel Query Support
4673 : * ----------------------------------------------------------------
4674 : */
4675 :
4676 : /* ----------------------------------------------------------------
4677 : * ExecAggEstimate
4678 : *
4679 : * Estimate space required to propagate aggregate statistics.
4680 : * ----------------------------------------------------------------
4681 : */
4682 : void
4683 564 : ExecAggEstimate(AggState *node, ParallelContext *pcxt)
4684 : {
4685 : Size size;
4686 :
4687 : /* don't need this if not instrumenting or no workers */
4688 564 : if (!node->ss.ps.instrument || pcxt->nworkers == 0)
4689 456 : return;
4690 :
4691 108 : size = mul_size(pcxt->nworkers, sizeof(AggregateInstrumentation));
4692 108 : size = add_size(size, offsetof(SharedAggInfo, sinstrument));
4693 108 : shm_toc_estimate_chunk(&pcxt->estimator, size);
4694 108 : shm_toc_estimate_keys(&pcxt->estimator, 1);
4695 : }
4696 :
4697 : /* ----------------------------------------------------------------
4698 : * ExecAggInitializeDSM
4699 : *
4700 : * Initialize DSM space for aggregate statistics.
4701 : * ----------------------------------------------------------------
4702 : */
4703 : void
4704 564 : ExecAggInitializeDSM(AggState *node, ParallelContext *pcxt)
4705 : {
4706 : Size size;
4707 :
4708 : /* don't need this if not instrumenting or no workers */
4709 564 : if (!node->ss.ps.instrument || pcxt->nworkers == 0)
4710 456 : return;
4711 :
4712 108 : size = offsetof(SharedAggInfo, sinstrument)
4713 108 : + pcxt->nworkers * sizeof(AggregateInstrumentation);
4714 108 : node->shared_info = shm_toc_allocate(pcxt->toc, size);
4715 : /* ensure any unfilled slots will contain zeroes */
4716 108 : memset(node->shared_info, 0, size);
4717 108 : node->shared_info->num_workers = pcxt->nworkers;
4718 108 : shm_toc_insert(pcxt->toc, node->ss.ps.plan->plan_node_id,
4719 108 : node->shared_info);
4720 : }
4721 :
4722 : /* ----------------------------------------------------------------
4723 : * ExecAggInitializeWorker
4724 : *
4725 : * Attach worker to DSM space for aggregate statistics.
4726 : * ----------------------------------------------------------------
4727 : */
4728 : void
4729 1556 : ExecAggInitializeWorker(AggState *node, ParallelWorkerContext *pwcxt)
4730 : {
4731 1556 : node->shared_info =
4732 1556 : shm_toc_lookup(pwcxt->toc, node->ss.ps.plan->plan_node_id, true);
4733 1556 : }
4734 :
4735 : /* ----------------------------------------------------------------
4736 : * ExecAggRetrieveInstrumentation
4737 : *
4738 : * Transfer aggregate statistics from DSM to private memory.
4739 : * ----------------------------------------------------------------
4740 : */
4741 : void
4742 108 : ExecAggRetrieveInstrumentation(AggState *node)
4743 : {
4744 : Size size;
4745 : SharedAggInfo *si;
4746 :
4747 108 : if (node->shared_info == NULL)
4748 0 : return;
4749 :
4750 108 : size = offsetof(SharedAggInfo, sinstrument)
4751 108 : + node->shared_info->num_workers * sizeof(AggregateInstrumentation);
4752 108 : si = palloc(size);
4753 108 : memcpy(si, node->shared_info, size);
4754 108 : node->shared_info = si;
4755 : }
|
