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) 1996-2025, 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/expandeddatum.h"
271 : #include "utils/injection_point.h"
272 : #include "utils/logtape.h"
273 : #include "utils/lsyscache.h"
274 : #include "utils/memutils.h"
275 : #include "utils/memutils_memorychunk.h"
276 : #include "utils/syscache.h"
277 : #include "utils/tuplesort.h"
278 :
279 : /*
280 : * Control how many partitions are created when spilling HashAgg to
281 : * disk.
282 : *
283 : * HASHAGG_PARTITION_FACTOR is multiplied by the estimated number of
284 : * partitions needed such that each partition will fit in memory. The factor
285 : * is set higher than one because there's not a high cost to having a few too
286 : * many partitions, and it makes it less likely that a partition will need to
287 : * be spilled recursively. Another benefit of having more, smaller partitions
288 : * is that small hash tables may perform better than large ones due to memory
289 : * caching effects.
290 : *
291 : * We also specify a min and max number of partitions per spill. Too few might
292 : * mean a lot of wasted I/O from repeated spilling of the same tuples. Too
293 : * many will result in lots of memory wasted buffering the spill files (which
294 : * could instead be spent on a larger hash table).
295 : */
296 : #define HASHAGG_PARTITION_FACTOR 1.50
297 : #define HASHAGG_MIN_PARTITIONS 4
298 : #define HASHAGG_MAX_PARTITIONS 1024
299 :
300 : /*
301 : * For reading from tapes, the buffer size must be a multiple of
302 : * BLCKSZ. Larger values help when reading from multiple tapes concurrently,
303 : * but that doesn't happen in HashAgg, so we simply use BLCKSZ. Writing to a
304 : * tape always uses a buffer of size BLCKSZ.
305 : */
306 : #define HASHAGG_READ_BUFFER_SIZE BLCKSZ
307 : #define HASHAGG_WRITE_BUFFER_SIZE BLCKSZ
308 :
309 : /*
310 : * HyperLogLog is used for estimating the cardinality of the spilled tuples in
311 : * a given partition. 5 bits corresponds to a size of about 32 bytes and a
312 : * worst-case error of around 18%. That's effective enough to choose a
313 : * reasonable number of partitions when recursing.
314 : */
315 : #define HASHAGG_HLL_BIT_WIDTH 5
316 :
317 : /*
318 : * Assume the palloc overhead always uses sizeof(MemoryChunk) bytes.
319 : */
320 : #define CHUNKHDRSZ sizeof(MemoryChunk)
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 void hash_create_memory(AggState *aggstate);
409 : static long hash_choose_num_buckets(double hashentrysize,
410 : long ngroups, Size memory);
411 : static int hash_choose_num_partitions(double input_groups,
412 : double hashentrysize,
413 : int used_bits,
414 : int *log2_npartitions);
415 : static void initialize_hash_entry(AggState *aggstate,
416 : TupleHashTable hashtable,
417 : TupleHashEntry entry);
418 : static void lookup_hash_entries(AggState *aggstate);
419 : static TupleTableSlot *agg_retrieve_direct(AggState *aggstate);
420 : static void agg_fill_hash_table(AggState *aggstate);
421 : static bool agg_refill_hash_table(AggState *aggstate);
422 : static TupleTableSlot *agg_retrieve_hash_table(AggState *aggstate);
423 : static TupleTableSlot *agg_retrieve_hash_table_in_memory(AggState *aggstate);
424 : static void hash_agg_check_limits(AggState *aggstate);
425 : static void hash_agg_enter_spill_mode(AggState *aggstate);
426 : static void hash_agg_update_metrics(AggState *aggstate, bool from_tape,
427 : int npartitions);
428 : static void hashagg_finish_initial_spills(AggState *aggstate);
429 : static void hashagg_reset_spill_state(AggState *aggstate);
430 : static HashAggBatch *hashagg_batch_new(LogicalTape *input_tape, int setno,
431 : int64 input_tuples, double input_card,
432 : int used_bits);
433 : static MinimalTuple hashagg_batch_read(HashAggBatch *batch, uint32 *hashp);
434 : static void hashagg_spill_init(HashAggSpill *spill, LogicalTapeSet *tapeset,
435 : int used_bits, double input_groups,
436 : double hashentrysize);
437 : static Size hashagg_spill_tuple(AggState *aggstate, HashAggSpill *spill,
438 : TupleTableSlot *inputslot, uint32 hash);
439 : static void hashagg_spill_finish(AggState *aggstate, HashAggSpill *spill,
440 : int setno);
441 : static Datum GetAggInitVal(Datum textInitVal, Oid transtype);
442 : static void build_pertrans_for_aggref(AggStatePerTrans pertrans,
443 : AggState *aggstate, EState *estate,
444 : Aggref *aggref, Oid transfn_oid,
445 : Oid aggtranstype, Oid aggserialfn,
446 : Oid aggdeserialfn, Datum initValue,
447 : bool initValueIsNull, Oid *inputTypes,
448 : int numArguments);
449 :
450 :
451 : /*
452 : * Select the current grouping set; affects current_set and
453 : * curaggcontext.
454 : */
455 : static void
456 7904028 : select_current_set(AggState *aggstate, int setno, bool is_hash)
457 : {
458 : /*
459 : * When changing this, also adapt ExecAggPlainTransByVal() and
460 : * ExecAggPlainTransByRef().
461 : */
462 7904028 : if (is_hash)
463 7232672 : aggstate->curaggcontext = aggstate->hashcontext;
464 : else
465 671356 : aggstate->curaggcontext = aggstate->aggcontexts[setno];
466 :
467 7904028 : aggstate->current_set = setno;
468 7904028 : }
469 :
470 : /*
471 : * Switch to phase "newphase", which must either be 0 or 1 (to reset) or
472 : * current_phase + 1. Juggle the tuplesorts accordingly.
473 : *
474 : * Phase 0 is for hashing, which we currently handle last in the AGG_MIXED
475 : * case, so when entering phase 0, all we need to do is drop open sorts.
476 : */
477 : static void
478 88910 : initialize_phase(AggState *aggstate, int newphase)
479 : {
480 : Assert(newphase <= 1 || newphase == aggstate->current_phase + 1);
481 :
482 : /*
483 : * Whatever the previous state, we're now done with whatever input
484 : * tuplesort was in use.
485 : */
486 88910 : if (aggstate->sort_in)
487 : {
488 42 : tuplesort_end(aggstate->sort_in);
489 42 : aggstate->sort_in = NULL;
490 : }
491 :
492 88910 : if (newphase <= 1)
493 : {
494 : /*
495 : * Discard any existing output tuplesort.
496 : */
497 88706 : if (aggstate->sort_out)
498 : {
499 6 : tuplesort_end(aggstate->sort_out);
500 6 : aggstate->sort_out = NULL;
501 : }
502 : }
503 : else
504 : {
505 : /*
506 : * The old output tuplesort becomes the new input one, and this is the
507 : * right time to actually sort it.
508 : */
509 204 : aggstate->sort_in = aggstate->sort_out;
510 204 : aggstate->sort_out = NULL;
511 : Assert(aggstate->sort_in);
512 204 : tuplesort_performsort(aggstate->sort_in);
513 : }
514 :
515 : /*
516 : * If this isn't the last phase, we need to sort appropriately for the
517 : * next phase in sequence.
518 : */
519 88910 : if (newphase > 0 && newphase < aggstate->numphases - 1)
520 : {
521 258 : Sort *sortnode = aggstate->phases[newphase + 1].sortnode;
522 258 : PlanState *outerNode = outerPlanState(aggstate);
523 258 : TupleDesc tupDesc = ExecGetResultType(outerNode);
524 :
525 258 : aggstate->sort_out = tuplesort_begin_heap(tupDesc,
526 : sortnode->numCols,
527 : sortnode->sortColIdx,
528 : sortnode->sortOperators,
529 : sortnode->collations,
530 : sortnode->nullsFirst,
531 : work_mem,
532 : NULL, TUPLESORT_NONE);
533 : }
534 :
535 88910 : aggstate->current_phase = newphase;
536 88910 : aggstate->phase = &aggstate->phases[newphase];
537 88910 : }
538 :
539 : /*
540 : * Fetch a tuple from either the outer plan (for phase 1) or from the sorter
541 : * populated by the previous phase. Copy it to the sorter for the next phase
542 : * if any.
543 : *
544 : * Callers cannot rely on memory for tuple in returned slot remaining valid
545 : * past any subsequently fetched tuple.
546 : */
547 : static TupleTableSlot *
548 28764648 : fetch_input_tuple(AggState *aggstate)
549 : {
550 : TupleTableSlot *slot;
551 :
552 28764648 : if (aggstate->sort_in)
553 : {
554 : /* make sure we check for interrupts in either path through here */
555 294900 : CHECK_FOR_INTERRUPTS();
556 294900 : if (!tuplesort_gettupleslot(aggstate->sort_in, true, false,
557 : aggstate->sort_slot, NULL))
558 204 : return NULL;
559 294696 : slot = aggstate->sort_slot;
560 : }
561 : else
562 28469748 : slot = ExecProcNode(outerPlanState(aggstate));
563 :
564 28764354 : if (!TupIsNull(slot) && aggstate->sort_out)
565 294696 : tuplesort_puttupleslot(aggstate->sort_out, slot);
566 :
567 28764354 : return slot;
568 : }
569 :
570 : /*
571 : * (Re)Initialize an individual aggregate.
572 : *
573 : * This function handles only one grouping set, already set in
574 : * aggstate->current_set.
575 : *
576 : * When called, CurrentMemoryContext should be the per-query context.
577 : */
578 : static void
579 1132176 : initialize_aggregate(AggState *aggstate, AggStatePerTrans pertrans,
580 : AggStatePerGroup pergroupstate)
581 : {
582 : /*
583 : * Start a fresh sort operation for each DISTINCT/ORDER BY aggregate.
584 : */
585 1132176 : if (pertrans->aggsortrequired)
586 : {
587 : /*
588 : * In case of rescan, maybe there could be an uncompleted sort
589 : * operation? Clean it up if so.
590 : */
591 53842 : if (pertrans->sortstates[aggstate->current_set])
592 0 : tuplesort_end(pertrans->sortstates[aggstate->current_set]);
593 :
594 :
595 : /*
596 : * We use a plain Datum sorter when there's a single input column;
597 : * otherwise sort the full tuple. (See comments for
598 : * process_ordered_aggregate_single.)
599 : */
600 53842 : if (pertrans->numInputs == 1)
601 : {
602 53758 : Form_pg_attribute attr = TupleDescAttr(pertrans->sortdesc, 0);
603 :
604 53758 : pertrans->sortstates[aggstate->current_set] =
605 53758 : tuplesort_begin_datum(attr->atttypid,
606 53758 : pertrans->sortOperators[0],
607 53758 : pertrans->sortCollations[0],
608 53758 : pertrans->sortNullsFirst[0],
609 : work_mem, NULL, TUPLESORT_NONE);
610 : }
611 : else
612 84 : pertrans->sortstates[aggstate->current_set] =
613 84 : tuplesort_begin_heap(pertrans->sortdesc,
614 : pertrans->numSortCols,
615 : pertrans->sortColIdx,
616 : pertrans->sortOperators,
617 : pertrans->sortCollations,
618 : pertrans->sortNullsFirst,
619 : work_mem, NULL, TUPLESORT_NONE);
620 : }
621 :
622 : /*
623 : * (Re)set transValue to the initial value.
624 : *
625 : * Note that when the initial value is pass-by-ref, we must copy it (into
626 : * the aggcontext) since we will pfree the transValue later.
627 : */
628 1132176 : if (pertrans->initValueIsNull)
629 597282 : pergroupstate->transValue = pertrans->initValue;
630 : else
631 : {
632 : MemoryContext oldContext;
633 :
634 534894 : oldContext = MemoryContextSwitchTo(aggstate->curaggcontext->ecxt_per_tuple_memory);
635 1069788 : pergroupstate->transValue = datumCopy(pertrans->initValue,
636 534894 : pertrans->transtypeByVal,
637 534894 : pertrans->transtypeLen);
638 534894 : MemoryContextSwitchTo(oldContext);
639 : }
640 1132176 : pergroupstate->transValueIsNull = pertrans->initValueIsNull;
641 :
642 : /*
643 : * If the initial value for the transition state doesn't exist in the
644 : * pg_aggregate table then we will let the first non-NULL value returned
645 : * from the outer procNode become the initial value. (This is useful for
646 : * aggregates like max() and min().) The noTransValue flag signals that we
647 : * still need to do this.
648 : */
649 1132176 : pergroupstate->noTransValue = pertrans->initValueIsNull;
650 1132176 : }
651 :
652 : /*
653 : * Initialize all aggregate transition states for a new group of input values.
654 : *
655 : * If there are multiple grouping sets, we initialize only the first numReset
656 : * of them (the grouping sets are ordered so that the most specific one, which
657 : * is reset most often, is first). As a convenience, if numReset is 0, we
658 : * reinitialize all sets.
659 : *
660 : * NB: This cannot be used for hash aggregates, as for those the grouping set
661 : * number has to be specified from further up.
662 : *
663 : * When called, CurrentMemoryContext should be the per-query context.
664 : */
665 : static void
666 301046 : initialize_aggregates(AggState *aggstate,
667 : AggStatePerGroup *pergroups,
668 : int numReset)
669 : {
670 : int transno;
671 301046 : int numGroupingSets = Max(aggstate->phase->numsets, 1);
672 301046 : int setno = 0;
673 301046 : int numTrans = aggstate->numtrans;
674 301046 : AggStatePerTrans transstates = aggstate->pertrans;
675 :
676 301046 : if (numReset == 0)
677 0 : numReset = numGroupingSets;
678 :
679 616266 : for (setno = 0; setno < numReset; setno++)
680 : {
681 315220 : AggStatePerGroup pergroup = pergroups[setno];
682 :
683 315220 : select_current_set(aggstate, setno, false);
684 :
685 982484 : for (transno = 0; transno < numTrans; transno++)
686 : {
687 667264 : AggStatePerTrans pertrans = &transstates[transno];
688 667264 : AggStatePerGroup pergroupstate = &pergroup[transno];
689 :
690 667264 : initialize_aggregate(aggstate, pertrans, pergroupstate);
691 : }
692 : }
693 301046 : }
694 :
695 : /*
696 : * Given new input value(s), advance the transition function of one aggregate
697 : * state within one grouping set only (already set in aggstate->current_set)
698 : *
699 : * The new values (and null flags) have been preloaded into argument positions
700 : * 1 and up in pertrans->transfn_fcinfo, so that we needn't copy them again to
701 : * pass to the transition function. We also expect that the static fields of
702 : * the fcinfo are already initialized; that was done by ExecInitAgg().
703 : *
704 : * It doesn't matter which memory context this is called in.
705 : */
706 : static void
707 724230 : advance_transition_function(AggState *aggstate,
708 : AggStatePerTrans pertrans,
709 : AggStatePerGroup pergroupstate)
710 : {
711 724230 : FunctionCallInfo fcinfo = pertrans->transfn_fcinfo;
712 : MemoryContext oldContext;
713 : Datum newVal;
714 :
715 724230 : if (pertrans->transfn.fn_strict)
716 : {
717 : /*
718 : * For a strict transfn, nothing happens when there's a NULL input; we
719 : * just keep the prior transValue.
720 : */
721 225000 : int numTransInputs = pertrans->numTransInputs;
722 : int i;
723 :
724 450000 : for (i = 1; i <= numTransInputs; i++)
725 : {
726 225000 : if (fcinfo->args[i].isnull)
727 0 : return;
728 : }
729 225000 : if (pergroupstate->noTransValue)
730 : {
731 : /*
732 : * transValue has not been initialized. This is the first non-NULL
733 : * input value. We use it as the initial value for transValue. (We
734 : * already checked that the agg's input type is binary-compatible
735 : * with its transtype, so straight copy here is OK.)
736 : *
737 : * We must copy the datum into aggcontext if it is pass-by-ref. We
738 : * do not need to pfree the old transValue, since it's NULL.
739 : */
740 0 : oldContext = MemoryContextSwitchTo(aggstate->curaggcontext->ecxt_per_tuple_memory);
741 0 : pergroupstate->transValue = datumCopy(fcinfo->args[1].value,
742 0 : pertrans->transtypeByVal,
743 0 : pertrans->transtypeLen);
744 0 : pergroupstate->transValueIsNull = false;
745 0 : pergroupstate->noTransValue = false;
746 0 : MemoryContextSwitchTo(oldContext);
747 0 : return;
748 : }
749 225000 : if (pergroupstate->transValueIsNull)
750 : {
751 : /*
752 : * Don't call a strict function with NULL inputs. Note it is
753 : * possible to get here despite the above tests, if the transfn is
754 : * strict *and* returned a NULL on a prior cycle. If that happens
755 : * we will propagate the NULL all the way to the end.
756 : */
757 0 : return;
758 : }
759 : }
760 :
761 : /* We run the transition functions in per-input-tuple memory context */
762 724230 : oldContext = MemoryContextSwitchTo(aggstate->tmpcontext->ecxt_per_tuple_memory);
763 :
764 : /* set up aggstate->curpertrans for AggGetAggref() */
765 724230 : aggstate->curpertrans = pertrans;
766 :
767 : /*
768 : * OK to call the transition function
769 : */
770 724230 : fcinfo->args[0].value = pergroupstate->transValue;
771 724230 : fcinfo->args[0].isnull = pergroupstate->transValueIsNull;
772 724230 : fcinfo->isnull = false; /* just in case transfn doesn't set it */
773 :
774 724230 : newVal = FunctionCallInvoke(fcinfo);
775 :
776 724230 : aggstate->curpertrans = NULL;
777 :
778 : /*
779 : * If pass-by-ref datatype, must copy the new value into aggcontext and
780 : * free the prior transValue. But if transfn returned a pointer to its
781 : * first input, we don't need to do anything.
782 : *
783 : * It's safe to compare newVal with pergroup->transValue without regard
784 : * for either being NULL, because ExecAggCopyTransValue takes care to set
785 : * transValue to 0 when NULL. Otherwise we could end up accidentally not
786 : * reparenting, when the transValue has the same numerical value as
787 : * newValue, despite being NULL. This is a somewhat hot path, making it
788 : * undesirable to instead solve this with another branch for the common
789 : * case of the transition function returning its (modified) input
790 : * argument.
791 : */
792 724230 : if (!pertrans->transtypeByVal &&
793 0 : DatumGetPointer(newVal) != DatumGetPointer(pergroupstate->transValue))
794 0 : newVal = ExecAggCopyTransValue(aggstate, pertrans,
795 0 : newVal, fcinfo->isnull,
796 : pergroupstate->transValue,
797 0 : pergroupstate->transValueIsNull);
798 :
799 724230 : pergroupstate->transValue = newVal;
800 724230 : pergroupstate->transValueIsNull = fcinfo->isnull;
801 :
802 724230 : MemoryContextSwitchTo(oldContext);
803 : }
804 :
805 : /*
806 : * Advance each aggregate transition state for one input tuple. The input
807 : * tuple has been stored in tmpcontext->ecxt_outertuple, so that it is
808 : * accessible to ExecEvalExpr.
809 : *
810 : * We have two sets of transition states to handle: one for sorted aggregation
811 : * and one for hashed; we do them both here, to avoid multiple evaluation of
812 : * the inputs.
813 : *
814 : * When called, CurrentMemoryContext should be the per-query context.
815 : */
816 : static void
817 29445628 : advance_aggregates(AggState *aggstate)
818 : {
819 29445628 : ExecEvalExprNoReturnSwitchContext(aggstate->phase->evaltrans,
820 : aggstate->tmpcontext);
821 29445550 : }
822 :
823 : /*
824 : * Run the transition function for a DISTINCT or ORDER BY aggregate
825 : * with only one input. This is called after we have completed
826 : * entering all the input values into the sort object. We complete the
827 : * sort, read out the values in sorted order, and run the transition
828 : * function on each value (applying DISTINCT if appropriate).
829 : *
830 : * Note that the strictness of the transition function was checked when
831 : * entering the values into the sort, so we don't check it again here;
832 : * we just apply standard SQL DISTINCT logic.
833 : *
834 : * The one-input case is handled separately from the multi-input case
835 : * for performance reasons: for single by-value inputs, such as the
836 : * common case of count(distinct id), the tuplesort_getdatum code path
837 : * is around 300% faster. (The speedup for by-reference types is less
838 : * but still noticeable.)
839 : *
840 : * This function handles only one grouping set (already set in
841 : * aggstate->current_set).
842 : *
843 : * When called, CurrentMemoryContext should be the per-query context.
844 : */
845 : static void
846 53758 : process_ordered_aggregate_single(AggState *aggstate,
847 : AggStatePerTrans pertrans,
848 : AggStatePerGroup pergroupstate)
849 : {
850 53758 : Datum oldVal = (Datum) 0;
851 53758 : bool oldIsNull = true;
852 53758 : bool haveOldVal = false;
853 53758 : MemoryContext workcontext = aggstate->tmpcontext->ecxt_per_tuple_memory;
854 : MemoryContext oldContext;
855 53758 : bool isDistinct = (pertrans->numDistinctCols > 0);
856 53758 : Datum newAbbrevVal = (Datum) 0;
857 53758 : Datum oldAbbrevVal = (Datum) 0;
858 53758 : FunctionCallInfo fcinfo = pertrans->transfn_fcinfo;
859 : Datum *newVal;
860 : bool *isNull;
861 :
862 : Assert(pertrans->numDistinctCols < 2);
863 :
864 53758 : tuplesort_performsort(pertrans->sortstates[aggstate->current_set]);
865 :
866 : /* Load the column into argument 1 (arg 0 will be transition value) */
867 53758 : newVal = &fcinfo->args[1].value;
868 53758 : isNull = &fcinfo->args[1].isnull;
869 :
870 : /*
871 : * Note: if input type is pass-by-ref, the datums returned by the sort are
872 : * freshly palloc'd in the per-query context, so we must be careful to
873 : * pfree them when they are no longer needed.
874 : */
875 :
876 898276 : while (tuplesort_getdatum(pertrans->sortstates[aggstate->current_set],
877 : true, false, newVal, isNull, &newAbbrevVal))
878 : {
879 : /*
880 : * Clear and select the working context for evaluation of the equality
881 : * function and transition function.
882 : */
883 844518 : MemoryContextReset(workcontext);
884 844518 : oldContext = MemoryContextSwitchTo(workcontext);
885 :
886 : /*
887 : * If DISTINCT mode, and not distinct from prior, skip it.
888 : */
889 844518 : if (isDistinct &&
890 310454 : haveOldVal &&
891 0 : ((oldIsNull && *isNull) ||
892 310454 : (!oldIsNull && !*isNull &&
893 605948 : oldAbbrevVal == newAbbrevVal &&
894 295494 : DatumGetBool(FunctionCall2Coll(&pertrans->equalfnOne,
895 : pertrans->aggCollation,
896 : oldVal, *newVal)))))
897 : {
898 120504 : MemoryContextSwitchTo(oldContext);
899 120504 : continue;
900 : }
901 : else
902 : {
903 724014 : advance_transition_function(aggstate, pertrans, pergroupstate);
904 :
905 724014 : MemoryContextSwitchTo(oldContext);
906 :
907 : /*
908 : * Forget the old value, if any, and remember the new one for
909 : * subsequent equality checks.
910 : */
911 724014 : if (!pertrans->inputtypeByVal)
912 : {
913 525288 : if (!oldIsNull)
914 525108 : pfree(DatumGetPointer(oldVal));
915 525288 : if (!*isNull)
916 525228 : oldVal = datumCopy(*newVal, pertrans->inputtypeByVal,
917 525228 : pertrans->inputtypeLen);
918 : }
919 : else
920 198726 : oldVal = *newVal;
921 724014 : oldAbbrevVal = newAbbrevVal;
922 724014 : oldIsNull = *isNull;
923 724014 : haveOldVal = true;
924 : }
925 : }
926 :
927 53758 : if (!oldIsNull && !pertrans->inputtypeByVal)
928 120 : pfree(DatumGetPointer(oldVal));
929 :
930 53758 : tuplesort_end(pertrans->sortstates[aggstate->current_set]);
931 53758 : pertrans->sortstates[aggstate->current_set] = NULL;
932 53758 : }
933 :
934 : /*
935 : * Run the transition function for a DISTINCT or ORDER BY aggregate
936 : * with more than one input. This is called after we have completed
937 : * entering all the input values into the sort object. We complete the
938 : * sort, read out the values in sorted order, and run the transition
939 : * function on each value (applying DISTINCT if appropriate).
940 : *
941 : * This function handles only one grouping set (already set in
942 : * aggstate->current_set).
943 : *
944 : * When called, CurrentMemoryContext should be the per-query context.
945 : */
946 : static void
947 84 : process_ordered_aggregate_multi(AggState *aggstate,
948 : AggStatePerTrans pertrans,
949 : AggStatePerGroup pergroupstate)
950 : {
951 84 : ExprContext *tmpcontext = aggstate->tmpcontext;
952 84 : FunctionCallInfo fcinfo = pertrans->transfn_fcinfo;
953 84 : TupleTableSlot *slot1 = pertrans->sortslot;
954 84 : TupleTableSlot *slot2 = pertrans->uniqslot;
955 84 : int numTransInputs = pertrans->numTransInputs;
956 84 : int numDistinctCols = pertrans->numDistinctCols;
957 84 : Datum newAbbrevVal = (Datum) 0;
958 84 : Datum oldAbbrevVal = (Datum) 0;
959 84 : bool haveOldValue = false;
960 84 : TupleTableSlot *save = aggstate->tmpcontext->ecxt_outertuple;
961 : int i;
962 :
963 84 : tuplesort_performsort(pertrans->sortstates[aggstate->current_set]);
964 :
965 84 : ExecClearTuple(slot1);
966 84 : if (slot2)
967 0 : ExecClearTuple(slot2);
968 :
969 300 : while (tuplesort_gettupleslot(pertrans->sortstates[aggstate->current_set],
970 : true, true, slot1, &newAbbrevVal))
971 : {
972 216 : CHECK_FOR_INTERRUPTS();
973 :
974 216 : tmpcontext->ecxt_outertuple = slot1;
975 216 : tmpcontext->ecxt_innertuple = slot2;
976 :
977 216 : if (numDistinctCols == 0 ||
978 0 : !haveOldValue ||
979 0 : newAbbrevVal != oldAbbrevVal ||
980 0 : !ExecQual(pertrans->equalfnMulti, tmpcontext))
981 : {
982 : /*
983 : * Extract the first numTransInputs columns as datums to pass to
984 : * the transfn.
985 : */
986 216 : slot_getsomeattrs(slot1, numTransInputs);
987 :
988 : /* Load values into fcinfo */
989 : /* Start from 1, since the 0th arg will be the transition value */
990 612 : for (i = 0; i < numTransInputs; i++)
991 : {
992 396 : fcinfo->args[i + 1].value = slot1->tts_values[i];
993 396 : fcinfo->args[i + 1].isnull = slot1->tts_isnull[i];
994 : }
995 :
996 216 : advance_transition_function(aggstate, pertrans, pergroupstate);
997 :
998 216 : if (numDistinctCols > 0)
999 : {
1000 : /* swap the slot pointers to retain the current tuple */
1001 0 : TupleTableSlot *tmpslot = slot2;
1002 :
1003 0 : slot2 = slot1;
1004 0 : slot1 = tmpslot;
1005 : /* avoid ExecQual() calls by reusing abbreviated keys */
1006 0 : oldAbbrevVal = newAbbrevVal;
1007 0 : haveOldValue = true;
1008 : }
1009 : }
1010 :
1011 : /* Reset context each time */
1012 216 : ResetExprContext(tmpcontext);
1013 :
1014 216 : ExecClearTuple(slot1);
1015 : }
1016 :
1017 84 : if (slot2)
1018 0 : ExecClearTuple(slot2);
1019 :
1020 84 : tuplesort_end(pertrans->sortstates[aggstate->current_set]);
1021 84 : pertrans->sortstates[aggstate->current_set] = NULL;
1022 :
1023 : /* restore previous slot, potentially in use for grouping sets */
1024 84 : tmpcontext->ecxt_outertuple = save;
1025 84 : }
1026 :
1027 : /*
1028 : * Compute the final value of one aggregate.
1029 : *
1030 : * This function handles only one grouping set (already set in
1031 : * aggstate->current_set).
1032 : *
1033 : * The finalfn will be run, and the result delivered, in the
1034 : * output-tuple context; caller's CurrentMemoryContext does not matter.
1035 : * (But note that in some cases, such as when there is no finalfn, the
1036 : * result might be a pointer to or into the agg's transition value.)
1037 : *
1038 : * The finalfn uses the state as set in the transno. This also might be
1039 : * being used by another aggregate function, so it's important that we do
1040 : * nothing destructive here. Moreover, the aggregate's final value might
1041 : * get used in multiple places, so we mustn't return a R/W expanded datum.
1042 : */
1043 : static void
1044 1118736 : finalize_aggregate(AggState *aggstate,
1045 : AggStatePerAgg peragg,
1046 : AggStatePerGroup pergroupstate,
1047 : Datum *resultVal, bool *resultIsNull)
1048 : {
1049 1118736 : LOCAL_FCINFO(fcinfo, FUNC_MAX_ARGS);
1050 1118736 : bool anynull = false;
1051 : MemoryContext oldContext;
1052 : int i;
1053 : ListCell *lc;
1054 1118736 : AggStatePerTrans pertrans = &aggstate->pertrans[peragg->transno];
1055 :
1056 1118736 : oldContext = MemoryContextSwitchTo(aggstate->ss.ps.ps_ExprContext->ecxt_per_tuple_memory);
1057 :
1058 : /*
1059 : * Evaluate any direct arguments. We do this even if there's no finalfn
1060 : * (which is unlikely anyway), so that side-effects happen as expected.
1061 : * The direct arguments go into arg positions 1 and up, leaving position 0
1062 : * for the transition state value.
1063 : */
1064 1118736 : i = 1;
1065 1119710 : foreach(lc, peragg->aggdirectargs)
1066 : {
1067 974 : ExprState *expr = (ExprState *) lfirst(lc);
1068 :
1069 974 : fcinfo->args[i].value = ExecEvalExpr(expr,
1070 : aggstate->ss.ps.ps_ExprContext,
1071 : &fcinfo->args[i].isnull);
1072 974 : anynull |= fcinfo->args[i].isnull;
1073 974 : i++;
1074 : }
1075 :
1076 : /*
1077 : * Apply the agg's finalfn if one is provided, else return transValue.
1078 : */
1079 1118736 : if (OidIsValid(peragg->finalfn_oid))
1080 : {
1081 338908 : int numFinalArgs = peragg->numFinalArgs;
1082 :
1083 : /* set up aggstate->curperagg for AggGetAggref() */
1084 338908 : aggstate->curperagg = peragg;
1085 :
1086 338908 : InitFunctionCallInfoData(*fcinfo, &peragg->finalfn,
1087 : numFinalArgs,
1088 : pertrans->aggCollation,
1089 : (Node *) aggstate, NULL);
1090 :
1091 : /* Fill in the transition state value */
1092 338908 : fcinfo->args[0].value =
1093 338908 : MakeExpandedObjectReadOnly(pergroupstate->transValue,
1094 : pergroupstate->transValueIsNull,
1095 : pertrans->transtypeLen);
1096 338908 : fcinfo->args[0].isnull = pergroupstate->transValueIsNull;
1097 338908 : anynull |= pergroupstate->transValueIsNull;
1098 :
1099 : /* Fill any remaining argument positions with nulls */
1100 491628 : for (; i < numFinalArgs; i++)
1101 : {
1102 152720 : fcinfo->args[i].value = (Datum) 0;
1103 152720 : fcinfo->args[i].isnull = true;
1104 152720 : anynull = true;
1105 : }
1106 :
1107 338908 : if (fcinfo->flinfo->fn_strict && anynull)
1108 : {
1109 : /* don't call a strict function with NULL inputs */
1110 0 : *resultVal = (Datum) 0;
1111 0 : *resultIsNull = true;
1112 : }
1113 : else
1114 : {
1115 : Datum result;
1116 :
1117 338908 : result = FunctionCallInvoke(fcinfo);
1118 338896 : *resultIsNull = fcinfo->isnull;
1119 338896 : *resultVal = MakeExpandedObjectReadOnly(result,
1120 : fcinfo->isnull,
1121 : peragg->resulttypeLen);
1122 : }
1123 338896 : aggstate->curperagg = NULL;
1124 : }
1125 : else
1126 : {
1127 779828 : *resultVal =
1128 779828 : MakeExpandedObjectReadOnly(pergroupstate->transValue,
1129 : pergroupstate->transValueIsNull,
1130 : pertrans->transtypeLen);
1131 779828 : *resultIsNull = pergroupstate->transValueIsNull;
1132 : }
1133 :
1134 1118724 : MemoryContextSwitchTo(oldContext);
1135 1118724 : }
1136 :
1137 : /*
1138 : * Compute the output value of one partial aggregate.
1139 : *
1140 : * The serialization function will be run, and the result delivered, in the
1141 : * output-tuple context; caller's CurrentMemoryContext does not matter.
1142 : */
1143 : static void
1144 17040 : finalize_partialaggregate(AggState *aggstate,
1145 : AggStatePerAgg peragg,
1146 : AggStatePerGroup pergroupstate,
1147 : Datum *resultVal, bool *resultIsNull)
1148 : {
1149 17040 : AggStatePerTrans pertrans = &aggstate->pertrans[peragg->transno];
1150 : MemoryContext oldContext;
1151 :
1152 17040 : oldContext = MemoryContextSwitchTo(aggstate->ss.ps.ps_ExprContext->ecxt_per_tuple_memory);
1153 :
1154 : /*
1155 : * serialfn_oid will be set if we must serialize the transvalue before
1156 : * returning it
1157 : */
1158 17040 : if (OidIsValid(pertrans->serialfn_oid))
1159 : {
1160 : /* Don't call a strict serialization function with NULL input. */
1161 666 : if (pertrans->serialfn.fn_strict && pergroupstate->transValueIsNull)
1162 : {
1163 76 : *resultVal = (Datum) 0;
1164 76 : *resultIsNull = true;
1165 : }
1166 : else
1167 : {
1168 590 : FunctionCallInfo fcinfo = pertrans->serialfn_fcinfo;
1169 : Datum result;
1170 :
1171 590 : fcinfo->args[0].value =
1172 590 : MakeExpandedObjectReadOnly(pergroupstate->transValue,
1173 : pergroupstate->transValueIsNull,
1174 : pertrans->transtypeLen);
1175 590 : fcinfo->args[0].isnull = pergroupstate->transValueIsNull;
1176 590 : fcinfo->isnull = false;
1177 :
1178 590 : result = FunctionCallInvoke(fcinfo);
1179 590 : *resultIsNull = fcinfo->isnull;
1180 590 : *resultVal = MakeExpandedObjectReadOnly(result,
1181 : fcinfo->isnull,
1182 : peragg->resulttypeLen);
1183 : }
1184 : }
1185 : else
1186 : {
1187 16374 : *resultVal =
1188 16374 : MakeExpandedObjectReadOnly(pergroupstate->transValue,
1189 : pergroupstate->transValueIsNull,
1190 : pertrans->transtypeLen);
1191 16374 : *resultIsNull = pergroupstate->transValueIsNull;
1192 : }
1193 :
1194 17040 : MemoryContextSwitchTo(oldContext);
1195 17040 : }
1196 :
1197 : /*
1198 : * Extract the attributes that make up the grouping key into the
1199 : * hashslot. This is necessary to compute the hash or perform a lookup.
1200 : */
1201 : static inline void
1202 8270288 : prepare_hash_slot(AggStatePerHash perhash,
1203 : TupleTableSlot *inputslot,
1204 : TupleTableSlot *hashslot)
1205 : {
1206 : int i;
1207 :
1208 : /* transfer just the needed columns into hashslot */
1209 8270288 : slot_getsomeattrs(inputslot, perhash->largestGrpColIdx);
1210 8270288 : ExecClearTuple(hashslot);
1211 :
1212 20490618 : for (i = 0; i < perhash->numhashGrpCols; i++)
1213 : {
1214 12220330 : int varNumber = perhash->hashGrpColIdxInput[i] - 1;
1215 :
1216 12220330 : hashslot->tts_values[i] = inputslot->tts_values[varNumber];
1217 12220330 : hashslot->tts_isnull[i] = inputslot->tts_isnull[varNumber];
1218 : }
1219 8270288 : ExecStoreVirtualTuple(hashslot);
1220 8270288 : }
1221 :
1222 : /*
1223 : * Prepare to finalize and project based on the specified representative tuple
1224 : * slot and grouping set.
1225 : *
1226 : * In the specified tuple slot, force to null all attributes that should be
1227 : * read as null in the context of the current grouping set. Also stash the
1228 : * current group bitmap where GroupingExpr can get at it.
1229 : *
1230 : * This relies on three conditions:
1231 : *
1232 : * 1) Nothing is ever going to try and extract the whole tuple from this slot,
1233 : * only reference it in evaluations, which will only access individual
1234 : * attributes.
1235 : *
1236 : * 2) No system columns are going to need to be nulled. (If a system column is
1237 : * referenced in a group clause, it is actually projected in the outer plan
1238 : * tlist.)
1239 : *
1240 : * 3) Within a given phase, we never need to recover the value of an attribute
1241 : * once it has been set to null.
1242 : *
1243 : * Poking into the slot this way is a bit ugly, but the consensus is that the
1244 : * alternative was worse.
1245 : */
1246 : static void
1247 847904 : prepare_projection_slot(AggState *aggstate, TupleTableSlot *slot, int currentSet)
1248 : {
1249 847904 : if (aggstate->phase->grouped_cols)
1250 : {
1251 562264 : Bitmapset *grouped_cols = aggstate->phase->grouped_cols[currentSet];
1252 :
1253 562264 : aggstate->grouped_cols = grouped_cols;
1254 :
1255 562264 : if (TTS_EMPTY(slot))
1256 : {
1257 : /*
1258 : * Force all values to be NULL if working on an empty input tuple
1259 : * (i.e. an empty grouping set for which no input rows were
1260 : * supplied).
1261 : */
1262 60 : ExecStoreAllNullTuple(slot);
1263 : }
1264 562204 : else if (aggstate->all_grouped_cols)
1265 : {
1266 : ListCell *lc;
1267 :
1268 : /* all_grouped_cols is arranged in desc order */
1269 562156 : slot_getsomeattrs(slot, linitial_int(aggstate->all_grouped_cols));
1270 :
1271 1532738 : foreach(lc, aggstate->all_grouped_cols)
1272 : {
1273 970582 : int attnum = lfirst_int(lc);
1274 :
1275 970582 : if (!bms_is_member(attnum, grouped_cols))
1276 57832 : slot->tts_isnull[attnum - 1] = true;
1277 : }
1278 : }
1279 : }
1280 847904 : }
1281 :
1282 : /*
1283 : * Compute the final value of all aggregates for one group.
1284 : *
1285 : * This function handles only one grouping set at a time, which the caller must
1286 : * have selected. It's also the caller's responsibility to adjust the supplied
1287 : * pergroup parameter to point to the current set's transvalues.
1288 : *
1289 : * Results are stored in the output econtext aggvalues/aggnulls.
1290 : */
1291 : static void
1292 847904 : finalize_aggregates(AggState *aggstate,
1293 : AggStatePerAgg peraggs,
1294 : AggStatePerGroup pergroup)
1295 : {
1296 847904 : ExprContext *econtext = aggstate->ss.ps.ps_ExprContext;
1297 847904 : Datum *aggvalues = econtext->ecxt_aggvalues;
1298 847904 : bool *aggnulls = econtext->ecxt_aggnulls;
1299 : int aggno;
1300 :
1301 : /*
1302 : * If there were any DISTINCT and/or ORDER BY aggregates, sort their
1303 : * inputs and run the transition functions.
1304 : */
1305 1983422 : for (int transno = 0; transno < aggstate->numtrans; transno++)
1306 : {
1307 1135518 : AggStatePerTrans pertrans = &aggstate->pertrans[transno];
1308 : AggStatePerGroup pergroupstate;
1309 :
1310 1135518 : pergroupstate = &pergroup[transno];
1311 :
1312 1135518 : if (pertrans->aggsortrequired)
1313 : {
1314 : Assert(aggstate->aggstrategy != AGG_HASHED &&
1315 : aggstate->aggstrategy != AGG_MIXED);
1316 :
1317 53842 : if (pertrans->numInputs == 1)
1318 53758 : process_ordered_aggregate_single(aggstate,
1319 : pertrans,
1320 : pergroupstate);
1321 : else
1322 84 : process_ordered_aggregate_multi(aggstate,
1323 : pertrans,
1324 : pergroupstate);
1325 : }
1326 1081676 : else if (pertrans->numDistinctCols > 0 && pertrans->haslast)
1327 : {
1328 18358 : pertrans->haslast = false;
1329 :
1330 18358 : if (pertrans->numDistinctCols == 1)
1331 : {
1332 18262 : if (!pertrans->inputtypeByVal && !pertrans->lastisnull)
1333 262 : pfree(DatumGetPointer(pertrans->lastdatum));
1334 :
1335 18262 : pertrans->lastisnull = false;
1336 18262 : pertrans->lastdatum = (Datum) 0;
1337 : }
1338 : else
1339 96 : ExecClearTuple(pertrans->uniqslot);
1340 : }
1341 : }
1342 :
1343 : /*
1344 : * Run the final functions.
1345 : */
1346 1983668 : for (aggno = 0; aggno < aggstate->numaggs; aggno++)
1347 : {
1348 1135776 : AggStatePerAgg peragg = &peraggs[aggno];
1349 1135776 : int transno = peragg->transno;
1350 : AggStatePerGroup pergroupstate;
1351 :
1352 1135776 : pergroupstate = &pergroup[transno];
1353 :
1354 1135776 : if (DO_AGGSPLIT_SKIPFINAL(aggstate->aggsplit))
1355 17040 : finalize_partialaggregate(aggstate, peragg, pergroupstate,
1356 17040 : &aggvalues[aggno], &aggnulls[aggno]);
1357 : else
1358 1118736 : finalize_aggregate(aggstate, peragg, pergroupstate,
1359 1118736 : &aggvalues[aggno], &aggnulls[aggno]);
1360 : }
1361 847892 : }
1362 :
1363 : /*
1364 : * Project the result of a group (whose aggs have already been calculated by
1365 : * finalize_aggregates). Returns the result slot, or NULL if no row is
1366 : * projected (suppressed by qual).
1367 : */
1368 : static TupleTableSlot *
1369 847892 : project_aggregates(AggState *aggstate)
1370 : {
1371 847892 : ExprContext *econtext = aggstate->ss.ps.ps_ExprContext;
1372 :
1373 : /*
1374 : * Check the qual (HAVING clause); if the group does not match, ignore it.
1375 : */
1376 847892 : if (ExecQual(aggstate->ss.ps.qual, econtext))
1377 : {
1378 : /*
1379 : * Form and return projection tuple using the aggregate results and
1380 : * the representative input tuple.
1381 : */
1382 741416 : return ExecProject(aggstate->ss.ps.ps_ProjInfo);
1383 : }
1384 : else
1385 106476 : InstrCountFiltered1(aggstate, 1);
1386 :
1387 106476 : return NULL;
1388 : }
1389 :
1390 : /*
1391 : * Find input-tuple columns that are needed, dividing them into
1392 : * aggregated and unaggregated sets.
1393 : */
1394 : static void
1395 7008 : find_cols(AggState *aggstate, Bitmapset **aggregated, Bitmapset **unaggregated)
1396 : {
1397 7008 : Agg *agg = (Agg *) aggstate->ss.ps.plan;
1398 : FindColsContext context;
1399 :
1400 7008 : context.is_aggref = false;
1401 7008 : context.aggregated = NULL;
1402 7008 : context.unaggregated = NULL;
1403 :
1404 : /* Examine tlist and quals */
1405 7008 : (void) find_cols_walker((Node *) agg->plan.targetlist, &context);
1406 7008 : (void) find_cols_walker((Node *) agg->plan.qual, &context);
1407 :
1408 : /* In some cases, grouping columns will not appear in the tlist */
1409 17840 : for (int i = 0; i < agg->numCols; i++)
1410 10832 : context.unaggregated = bms_add_member(context.unaggregated,
1411 10832 : agg->grpColIdx[i]);
1412 :
1413 7008 : *aggregated = context.aggregated;
1414 7008 : *unaggregated = context.unaggregated;
1415 7008 : }
1416 :
1417 : static bool
1418 83758 : find_cols_walker(Node *node, FindColsContext *context)
1419 : {
1420 83758 : if (node == NULL)
1421 15078 : return false;
1422 68680 : if (IsA(node, Var))
1423 : {
1424 18906 : Var *var = (Var *) node;
1425 :
1426 : /* setrefs.c should have set the varno to OUTER_VAR */
1427 : Assert(var->varno == OUTER_VAR);
1428 : Assert(var->varlevelsup == 0);
1429 18906 : if (context->is_aggref)
1430 6090 : context->aggregated = bms_add_member(context->aggregated,
1431 6090 : var->varattno);
1432 : else
1433 12816 : context->unaggregated = bms_add_member(context->unaggregated,
1434 12816 : var->varattno);
1435 18906 : return false;
1436 : }
1437 49774 : if (IsA(node, Aggref))
1438 : {
1439 : Assert(!context->is_aggref);
1440 8570 : context->is_aggref = true;
1441 8570 : expression_tree_walker(node, find_cols_walker, context);
1442 8570 : context->is_aggref = false;
1443 8570 : return false;
1444 : }
1445 41204 : return expression_tree_walker(node, find_cols_walker, context);
1446 : }
1447 :
1448 : /*
1449 : * (Re-)initialize the hash table(s) to empty.
1450 : *
1451 : * To implement hashed aggregation, we need a hashtable that stores a
1452 : * representative tuple and an array of AggStatePerGroup structs for each
1453 : * distinct set of GROUP BY column values. We compute the hash key from the
1454 : * GROUP BY columns. The per-group data is allocated in initialize_hash_entry(),
1455 : * for each entry.
1456 : *
1457 : * We have a separate hashtable and associated perhash data structure for each
1458 : * grouping set for which we're doing hashing.
1459 : *
1460 : * The contents of the hash tables always live in the hashcontext's per-tuple
1461 : * memory context (there is only one of these for all tables together, since
1462 : * they are all reset at the same time).
1463 : */
1464 : static void
1465 17564 : build_hash_tables(AggState *aggstate)
1466 : {
1467 : int setno;
1468 :
1469 35472 : for (setno = 0; setno < aggstate->num_hashes; ++setno)
1470 : {
1471 17908 : AggStatePerHash perhash = &aggstate->perhash[setno];
1472 : long nbuckets;
1473 : Size memory;
1474 :
1475 17908 : if (perhash->hashtable != NULL)
1476 : {
1477 12462 : ResetTupleHashTable(perhash->hashtable);
1478 12462 : continue;
1479 : }
1480 :
1481 : Assert(perhash->aggnode->numGroups > 0);
1482 :
1483 5446 : memory = aggstate->hash_mem_limit / aggstate->num_hashes;
1484 :
1485 : /* choose reasonable number of buckets per hashtable */
1486 5446 : nbuckets = hash_choose_num_buckets(aggstate->hashentrysize,
1487 5446 : perhash->aggnode->numGroups,
1488 : memory);
1489 :
1490 : #ifdef USE_INJECTION_POINTS
1491 5446 : if (IS_INJECTION_POINT_ATTACHED("hash-aggregate-oversize-table"))
1492 : {
1493 0 : nbuckets = memory / TupleHashEntrySize();
1494 0 : INJECTION_POINT_CACHED("hash-aggregate-oversize-table", NULL);
1495 : }
1496 : #endif
1497 :
1498 5446 : build_hash_table(aggstate, setno, nbuckets);
1499 : }
1500 :
1501 17564 : aggstate->hash_ngroups_current = 0;
1502 17564 : }
1503 :
1504 : /*
1505 : * Build a single hashtable for this grouping set.
1506 : */
1507 : static void
1508 5446 : build_hash_table(AggState *aggstate, int setno, long nbuckets)
1509 : {
1510 5446 : AggStatePerHash perhash = &aggstate->perhash[setno];
1511 5446 : MemoryContext metacxt = aggstate->hash_metacxt;
1512 5446 : MemoryContext tablecxt = aggstate->hash_tablecxt;
1513 5446 : MemoryContext tmpcxt = aggstate->tmpcontext->ecxt_per_tuple_memory;
1514 : Size additionalsize;
1515 :
1516 : Assert(aggstate->aggstrategy == AGG_HASHED ||
1517 : aggstate->aggstrategy == AGG_MIXED);
1518 :
1519 : /*
1520 : * Used to make sure initial hash table allocation does not exceed
1521 : * hash_mem. Note that the estimate does not include space for
1522 : * pass-by-reference transition data values, nor for the representative
1523 : * tuple of each group.
1524 : */
1525 5446 : additionalsize = aggstate->numtrans * sizeof(AggStatePerGroupData);
1526 :
1527 10892 : perhash->hashtable = BuildTupleHashTable(&aggstate->ss.ps,
1528 5446 : perhash->hashslot->tts_tupleDescriptor,
1529 5446 : perhash->hashslot->tts_ops,
1530 : perhash->numCols,
1531 : perhash->hashGrpColIdxHash,
1532 5446 : perhash->eqfuncoids,
1533 : perhash->hashfunctions,
1534 5446 : perhash->aggnode->grpCollations,
1535 : nbuckets,
1536 : additionalsize,
1537 : metacxt,
1538 : tablecxt,
1539 : tmpcxt,
1540 5446 : DO_AGGSPLIT_SKIPFINAL(aggstate->aggsplit));
1541 5446 : }
1542 :
1543 : /*
1544 : * Compute columns that actually need to be stored in hashtable entries. The
1545 : * incoming tuples from the child plan node will contain grouping columns,
1546 : * other columns referenced in our targetlist and qual, columns used to
1547 : * compute the aggregate functions, and perhaps just junk columns we don't use
1548 : * at all. Only columns of the first two types need to be stored in the
1549 : * hashtable, and getting rid of the others can make the table entries
1550 : * significantly smaller. The hashtable only contains the relevant columns,
1551 : * and is packed/unpacked in lookup_hash_entries() / agg_retrieve_hash_table()
1552 : * into the format of the normal input descriptor.
1553 : *
1554 : * Additional columns, in addition to the columns grouped by, come from two
1555 : * sources: Firstly functionally dependent columns that we don't need to group
1556 : * by themselves, and secondly ctids for row-marks.
1557 : *
1558 : * To eliminate duplicates, we build a bitmapset of the needed columns, and
1559 : * then build an array of the columns included in the hashtable. We might
1560 : * still have duplicates if the passed-in grpColIdx has them, which can happen
1561 : * in edge cases from semijoins/distinct; these can't always be removed,
1562 : * because it's not certain that the duplicate cols will be using the same
1563 : * hash function.
1564 : *
1565 : * Note that the array is preserved over ExecReScanAgg, so we allocate it in
1566 : * the per-query context (unlike the hash table itself).
1567 : */
1568 : static void
1569 7008 : find_hash_columns(AggState *aggstate)
1570 : {
1571 : Bitmapset *base_colnos;
1572 : Bitmapset *aggregated_colnos;
1573 7008 : TupleDesc scanDesc = aggstate->ss.ss_ScanTupleSlot->tts_tupleDescriptor;
1574 7008 : List *outerTlist = outerPlanState(aggstate)->plan->targetlist;
1575 7008 : int numHashes = aggstate->num_hashes;
1576 7008 : EState *estate = aggstate->ss.ps.state;
1577 : int j;
1578 :
1579 : /* Find Vars that will be needed in tlist and qual */
1580 7008 : find_cols(aggstate, &aggregated_colnos, &base_colnos);
1581 7008 : aggstate->colnos_needed = bms_union(base_colnos, aggregated_colnos);
1582 7008 : aggstate->max_colno_needed = 0;
1583 7008 : aggstate->all_cols_needed = true;
1584 :
1585 29318 : for (int i = 0; i < scanDesc->natts; i++)
1586 : {
1587 22310 : int colno = i + 1;
1588 :
1589 22310 : if (bms_is_member(colno, aggstate->colnos_needed))
1590 16228 : aggstate->max_colno_needed = colno;
1591 : else
1592 6082 : aggstate->all_cols_needed = false;
1593 : }
1594 :
1595 14530 : for (j = 0; j < numHashes; ++j)
1596 : {
1597 7522 : AggStatePerHash perhash = &aggstate->perhash[j];
1598 7522 : Bitmapset *colnos = bms_copy(base_colnos);
1599 7522 : AttrNumber *grpColIdx = perhash->aggnode->grpColIdx;
1600 7522 : List *hashTlist = NIL;
1601 : TupleDesc hashDesc;
1602 : int maxCols;
1603 : int i;
1604 :
1605 7522 : perhash->largestGrpColIdx = 0;
1606 :
1607 : /*
1608 : * If we're doing grouping sets, then some Vars might be referenced in
1609 : * tlist/qual for the benefit of other grouping sets, but not needed
1610 : * when hashing; i.e. prepare_projection_slot will null them out, so
1611 : * there'd be no point storing them. Use prepare_projection_slot's
1612 : * logic to determine which.
1613 : */
1614 7522 : if (aggstate->phases[0].grouped_cols)
1615 : {
1616 7522 : Bitmapset *grouped_cols = aggstate->phases[0].grouped_cols[j];
1617 : ListCell *lc;
1618 :
1619 20320 : foreach(lc, aggstate->all_grouped_cols)
1620 : {
1621 12798 : int attnum = lfirst_int(lc);
1622 :
1623 12798 : if (!bms_is_member(attnum, grouped_cols))
1624 1344 : colnos = bms_del_member(colnos, attnum);
1625 : }
1626 : }
1627 :
1628 : /*
1629 : * Compute maximum number of input columns accounting for possible
1630 : * duplications in the grpColIdx array, which can happen in some edge
1631 : * cases where HashAggregate was generated as part of a semijoin or a
1632 : * DISTINCT.
1633 : */
1634 7522 : maxCols = bms_num_members(colnos) + perhash->numCols;
1635 :
1636 7522 : perhash->hashGrpColIdxInput =
1637 7522 : palloc(maxCols * sizeof(AttrNumber));
1638 7522 : perhash->hashGrpColIdxHash =
1639 7522 : palloc(perhash->numCols * sizeof(AttrNumber));
1640 :
1641 : /* Add all the grouping columns to colnos */
1642 18976 : for (i = 0; i < perhash->numCols; i++)
1643 11454 : colnos = bms_add_member(colnos, grpColIdx[i]);
1644 :
1645 : /*
1646 : * First build mapping for columns directly hashed. These are the
1647 : * first, because they'll be accessed when computing hash values and
1648 : * comparing tuples for exact matches. We also build simple mapping
1649 : * for execGrouping, so it knows where to find the to-be-hashed /
1650 : * compared columns in the input.
1651 : */
1652 18976 : for (i = 0; i < perhash->numCols; i++)
1653 : {
1654 11454 : perhash->hashGrpColIdxInput[i] = grpColIdx[i];
1655 11454 : perhash->hashGrpColIdxHash[i] = i + 1;
1656 11454 : perhash->numhashGrpCols++;
1657 : /* delete already mapped columns */
1658 11454 : colnos = bms_del_member(colnos, grpColIdx[i]);
1659 : }
1660 :
1661 : /* and add the remaining columns */
1662 7522 : i = -1;
1663 8784 : while ((i = bms_next_member(colnos, i)) >= 0)
1664 : {
1665 1262 : perhash->hashGrpColIdxInput[perhash->numhashGrpCols] = i;
1666 1262 : perhash->numhashGrpCols++;
1667 : }
1668 :
1669 : /* and build a tuple descriptor for the hashtable */
1670 20238 : for (i = 0; i < perhash->numhashGrpCols; i++)
1671 : {
1672 12716 : int varNumber = perhash->hashGrpColIdxInput[i] - 1;
1673 :
1674 12716 : hashTlist = lappend(hashTlist, list_nth(outerTlist, varNumber));
1675 12716 : perhash->largestGrpColIdx =
1676 12716 : Max(varNumber + 1, perhash->largestGrpColIdx);
1677 : }
1678 :
1679 7522 : hashDesc = ExecTypeFromTL(hashTlist);
1680 :
1681 7522 : execTuplesHashPrepare(perhash->numCols,
1682 7522 : perhash->aggnode->grpOperators,
1683 : &perhash->eqfuncoids,
1684 : &perhash->hashfunctions);
1685 7522 : perhash->hashslot =
1686 7522 : ExecAllocTableSlot(&estate->es_tupleTable, hashDesc,
1687 : &TTSOpsMinimalTuple);
1688 :
1689 7522 : list_free(hashTlist);
1690 7522 : bms_free(colnos);
1691 : }
1692 :
1693 7008 : bms_free(base_colnos);
1694 7008 : }
1695 :
1696 : /*
1697 : * Estimate per-hash-table-entry overhead.
1698 : */
1699 : Size
1700 41506 : hash_agg_entry_size(int numTrans, Size tupleWidth, Size transitionSpace)
1701 : {
1702 : Size tupleChunkSize;
1703 : Size pergroupChunkSize;
1704 : Size transitionChunkSize;
1705 41506 : Size tupleSize = (MAXALIGN(SizeofMinimalTupleHeader) +
1706 : tupleWidth);
1707 41506 : Size pergroupSize = numTrans * sizeof(AggStatePerGroupData);
1708 :
1709 : /*
1710 : * Entries use the Bump allocator, so the chunk sizes are the same as the
1711 : * requested sizes.
1712 : */
1713 41506 : tupleChunkSize = MAXALIGN(tupleSize);
1714 41506 : pergroupChunkSize = pergroupSize;
1715 :
1716 : /*
1717 : * Transition values use AllocSet, which has a chunk header and also uses
1718 : * power-of-two allocations.
1719 : */
1720 41506 : if (transitionSpace > 0)
1721 5394 : transitionChunkSize = CHUNKHDRSZ + pg_nextpower2_size_t(transitionSpace);
1722 : else
1723 36112 : transitionChunkSize = 0;
1724 :
1725 : return
1726 41506 : TupleHashEntrySize() +
1727 41506 : tupleChunkSize +
1728 41506 : pergroupChunkSize +
1729 : transitionChunkSize;
1730 : }
1731 :
1732 : /*
1733 : * hashagg_recompile_expressions()
1734 : *
1735 : * Identifies the right phase, compiles the right expression given the
1736 : * arguments, and then sets phase->evalfunc to that expression.
1737 : *
1738 : * Different versions of the compiled expression are needed depending on
1739 : * whether hash aggregation has spilled or not, and whether it's reading from
1740 : * the outer plan or a tape. Before spilling to disk, the expression reads
1741 : * from the outer plan and does not need to perform a NULL check. After
1742 : * HashAgg begins to spill, new groups will not be created in the hash table,
1743 : * and the AggStatePerGroup array may be NULL; therefore we need to add a null
1744 : * pointer check to the expression. Then, when reading spilled data from a
1745 : * tape, we change the outer slot type to be a fixed minimal tuple slot.
1746 : *
1747 : * It would be wasteful to recompile every time, so cache the compiled
1748 : * expressions in the AggStatePerPhase, and reuse when appropriate.
1749 : */
1750 : static void
1751 65804 : hashagg_recompile_expressions(AggState *aggstate, bool minslot, bool nullcheck)
1752 : {
1753 : AggStatePerPhase phase;
1754 65804 : int i = minslot ? 1 : 0;
1755 65804 : int j = nullcheck ? 1 : 0;
1756 :
1757 : Assert(aggstate->aggstrategy == AGG_HASHED ||
1758 : aggstate->aggstrategy == AGG_MIXED);
1759 :
1760 65804 : if (aggstate->aggstrategy == AGG_HASHED)
1761 13232 : phase = &aggstate->phases[0];
1762 : else /* AGG_MIXED */
1763 52572 : phase = &aggstate->phases[1];
1764 :
1765 65804 : if (phase->evaltrans_cache[i][j] == NULL)
1766 : {
1767 88 : const TupleTableSlotOps *outerops = aggstate->ss.ps.outerops;
1768 88 : bool outerfixed = aggstate->ss.ps.outeropsfixed;
1769 88 : bool dohash = true;
1770 88 : bool dosort = false;
1771 :
1772 : /*
1773 : * If minslot is true, that means we are processing a spilled batch
1774 : * (inside agg_refill_hash_table()), and we must not advance the
1775 : * sorted grouping sets.
1776 : */
1777 88 : if (aggstate->aggstrategy == AGG_MIXED && !minslot)
1778 12 : dosort = true;
1779 :
1780 : /* temporarily change the outerops while compiling the expression */
1781 88 : if (minslot)
1782 : {
1783 44 : aggstate->ss.ps.outerops = &TTSOpsMinimalTuple;
1784 44 : aggstate->ss.ps.outeropsfixed = true;
1785 : }
1786 :
1787 88 : phase->evaltrans_cache[i][j] = ExecBuildAggTrans(aggstate, phase,
1788 : dosort, dohash,
1789 : nullcheck);
1790 :
1791 : /* change back */
1792 88 : aggstate->ss.ps.outerops = outerops;
1793 88 : aggstate->ss.ps.outeropsfixed = outerfixed;
1794 : }
1795 :
1796 65804 : phase->evaltrans = phase->evaltrans_cache[i][j];
1797 65804 : }
1798 :
1799 : /*
1800 : * Set limits that trigger spilling to avoid exceeding hash_mem. Consider the
1801 : * number of partitions we expect to create (if we do spill).
1802 : *
1803 : * There are two limits: a memory limit, and also an ngroups limit. The
1804 : * ngroups limit becomes important when we expect transition values to grow
1805 : * substantially larger than the initial value.
1806 : */
1807 : void
1808 66016 : hash_agg_set_limits(double hashentrysize, double input_groups, int used_bits,
1809 : Size *mem_limit, uint64 *ngroups_limit,
1810 : int *num_partitions)
1811 : {
1812 : int npartitions;
1813 : Size partition_mem;
1814 66016 : Size hash_mem_limit = get_hash_memory_limit();
1815 :
1816 : /* if not expected to spill, use all of hash_mem */
1817 66016 : if (input_groups * hashentrysize <= hash_mem_limit)
1818 : {
1819 63586 : if (num_partitions != NULL)
1820 38976 : *num_partitions = 0;
1821 63586 : *mem_limit = hash_mem_limit;
1822 63586 : *ngroups_limit = hash_mem_limit / hashentrysize;
1823 63586 : return;
1824 : }
1825 :
1826 : /*
1827 : * Calculate expected memory requirements for spilling, which is the size
1828 : * of the buffers needed for all the tapes that need to be open at once.
1829 : * Then, subtract that from the memory available for holding hash tables.
1830 : */
1831 2430 : npartitions = hash_choose_num_partitions(input_groups,
1832 : hashentrysize,
1833 : used_bits,
1834 : NULL);
1835 2430 : if (num_partitions != NULL)
1836 96 : *num_partitions = npartitions;
1837 :
1838 2430 : partition_mem =
1839 2430 : HASHAGG_READ_BUFFER_SIZE +
1840 : HASHAGG_WRITE_BUFFER_SIZE * npartitions;
1841 :
1842 : /*
1843 : * Don't set the limit below 3/4 of hash_mem. In that case, we are at the
1844 : * minimum number of partitions, so we aren't going to dramatically exceed
1845 : * work mem anyway.
1846 : */
1847 2430 : if (hash_mem_limit > 4 * partition_mem)
1848 0 : *mem_limit = hash_mem_limit - partition_mem;
1849 : else
1850 2430 : *mem_limit = hash_mem_limit * 0.75;
1851 :
1852 2430 : if (*mem_limit > hashentrysize)
1853 2430 : *ngroups_limit = *mem_limit / hashentrysize;
1854 : else
1855 0 : *ngroups_limit = 1;
1856 : }
1857 :
1858 : /*
1859 : * hash_agg_check_limits
1860 : *
1861 : * After adding a new group to the hash table, check whether we need to enter
1862 : * spill mode. Allocations may happen without adding new groups (for instance,
1863 : * if the transition state size grows), so this check is imperfect.
1864 : */
1865 : static void
1866 530882 : hash_agg_check_limits(AggState *aggstate)
1867 : {
1868 530882 : uint64 ngroups = aggstate->hash_ngroups_current;
1869 530882 : Size meta_mem = MemoryContextMemAllocated(aggstate->hash_metacxt,
1870 : true);
1871 530882 : Size entry_mem = MemoryContextMemAllocated(aggstate->hash_tablecxt,
1872 : true);
1873 530882 : Size tval_mem = MemoryContextMemAllocated(aggstate->hashcontext->ecxt_per_tuple_memory,
1874 : true);
1875 530882 : Size total_mem = meta_mem + entry_mem + tval_mem;
1876 530882 : bool do_spill = false;
1877 :
1878 : #ifdef USE_INJECTION_POINTS
1879 530882 : if (ngroups >= 1000)
1880 : {
1881 95650 : if (IS_INJECTION_POINT_ATTACHED("hash-aggregate-spill-1000"))
1882 : {
1883 10 : do_spill = true;
1884 10 : INJECTION_POINT_CACHED("hash-aggregate-spill-1000", NULL);
1885 : }
1886 : }
1887 : #endif
1888 :
1889 : /*
1890 : * Don't spill unless there's at least one group in the hash table so we
1891 : * can be sure to make progress even in edge cases.
1892 : */
1893 530882 : if (aggstate->hash_ngroups_current > 0 &&
1894 530882 : (total_mem > aggstate->hash_mem_limit ||
1895 504470 : ngroups > aggstate->hash_ngroups_limit))
1896 : {
1897 26448 : do_spill = true;
1898 : }
1899 :
1900 530882 : if (do_spill)
1901 26458 : hash_agg_enter_spill_mode(aggstate);
1902 530882 : }
1903 :
1904 : /*
1905 : * Enter "spill mode", meaning that no new groups are added to any of the hash
1906 : * tables. Tuples that would create a new group are instead spilled, and
1907 : * processed later.
1908 : */
1909 : static void
1910 26458 : hash_agg_enter_spill_mode(AggState *aggstate)
1911 : {
1912 26458 : INJECTION_POINT("hash-aggregate-enter-spill-mode", NULL);
1913 26458 : aggstate->hash_spill_mode = true;
1914 26458 : hashagg_recompile_expressions(aggstate, aggstate->table_filled, true);
1915 :
1916 26458 : if (!aggstate->hash_ever_spilled)
1917 : {
1918 : Assert(aggstate->hash_tapeset == NULL);
1919 : Assert(aggstate->hash_spills == NULL);
1920 :
1921 62 : aggstate->hash_ever_spilled = true;
1922 :
1923 62 : aggstate->hash_tapeset = LogicalTapeSetCreate(true, NULL, -1);
1924 :
1925 62 : aggstate->hash_spills = palloc(sizeof(HashAggSpill) * aggstate->num_hashes);
1926 :
1927 184 : for (int setno = 0; setno < aggstate->num_hashes; setno++)
1928 : {
1929 122 : AggStatePerHash perhash = &aggstate->perhash[setno];
1930 122 : HashAggSpill *spill = &aggstate->hash_spills[setno];
1931 :
1932 122 : hashagg_spill_init(spill, aggstate->hash_tapeset, 0,
1933 122 : perhash->aggnode->numGroups,
1934 : aggstate->hashentrysize);
1935 : }
1936 : }
1937 26458 : }
1938 :
1939 : /*
1940 : * Update metrics after filling the hash table.
1941 : *
1942 : * If reading from the outer plan, from_tape should be false; if reading from
1943 : * another tape, from_tape should be true.
1944 : */
1945 : static void
1946 44246 : hash_agg_update_metrics(AggState *aggstate, bool from_tape, int npartitions)
1947 : {
1948 : Size meta_mem;
1949 : Size entry_mem;
1950 : Size hashkey_mem;
1951 : Size buffer_mem;
1952 : Size total_mem;
1953 :
1954 44246 : if (aggstate->aggstrategy != AGG_MIXED &&
1955 17840 : aggstate->aggstrategy != AGG_HASHED)
1956 0 : return;
1957 :
1958 : /* memory for the hash table itself */
1959 44246 : meta_mem = MemoryContextMemAllocated(aggstate->hash_metacxt, true);
1960 :
1961 : /* memory for hash entries */
1962 44246 : entry_mem = MemoryContextMemAllocated(aggstate->hash_tablecxt, true);
1963 :
1964 : /* memory for byref transition states */
1965 44246 : hashkey_mem = MemoryContextMemAllocated(aggstate->hashcontext->ecxt_per_tuple_memory, true);
1966 :
1967 : /* memory for read/write tape buffers, if spilled */
1968 44246 : buffer_mem = npartitions * HASHAGG_WRITE_BUFFER_SIZE;
1969 44246 : if (from_tape)
1970 26944 : buffer_mem += HASHAGG_READ_BUFFER_SIZE;
1971 :
1972 : /* update peak mem */
1973 44246 : total_mem = meta_mem + entry_mem + hashkey_mem + buffer_mem;
1974 44246 : if (total_mem > aggstate->hash_mem_peak)
1975 4950 : aggstate->hash_mem_peak = total_mem;
1976 :
1977 : /* update disk usage */
1978 44246 : if (aggstate->hash_tapeset != NULL)
1979 : {
1980 27006 : uint64 disk_used = LogicalTapeSetBlocks(aggstate->hash_tapeset) * (BLCKSZ / 1024);
1981 :
1982 27006 : if (aggstate->hash_disk_used < disk_used)
1983 52 : aggstate->hash_disk_used = disk_used;
1984 : }
1985 :
1986 : /* update hashentrysize estimate based on contents */
1987 44246 : if (aggstate->hash_ngroups_current > 0)
1988 : {
1989 43688 : aggstate->hashentrysize =
1990 43688 : TupleHashEntrySize() +
1991 43688 : (hashkey_mem / (double) aggstate->hash_ngroups_current);
1992 : }
1993 : }
1994 :
1995 : /*
1996 : * Create memory contexts used for hash aggregation.
1997 : */
1998 : static void
1999 7008 : hash_create_memory(AggState *aggstate)
2000 : {
2001 7008 : Size maxBlockSize = ALLOCSET_DEFAULT_MAXSIZE;
2002 :
2003 : /*
2004 : * The hashcontext's per-tuple memory will be used for byref transition
2005 : * values and returned by AggCheckCallContext().
2006 : */
2007 7008 : aggstate->hashcontext = CreateWorkExprContext(aggstate->ss.ps.state);
2008 :
2009 : /*
2010 : * The meta context will be used for the bucket array of
2011 : * TupleHashEntryData (or arrays, in the case of grouping sets). As the
2012 : * hash table grows, the bucket array will double in size and the old one
2013 : * will be freed, so an AllocSet is appropriate. For large bucket arrays,
2014 : * the large allocation path will be used, so it's not worth worrying
2015 : * about wasting space due to power-of-two allocations.
2016 : */
2017 7008 : aggstate->hash_metacxt = AllocSetContextCreate(aggstate->ss.ps.state->es_query_cxt,
2018 : "HashAgg meta context",
2019 : ALLOCSET_DEFAULT_SIZES);
2020 :
2021 : /*
2022 : * The hash entries themselves, which include the grouping key
2023 : * (firstTuple) and pergroup data, are stored in the table context. The
2024 : * bump allocator can be used because the entries are not freed until the
2025 : * entire hash table is reset. The bump allocator is faster for
2026 : * allocations and avoids wasting space on the chunk header or
2027 : * power-of-two allocations.
2028 : *
2029 : * Like CreateWorkExprContext(), use smaller sizings for smaller work_mem,
2030 : * to avoid large jumps in memory usage.
2031 : */
2032 :
2033 : /*
2034 : * Like CreateWorkExprContext(), use smaller sizings for smaller work_mem,
2035 : * to avoid large jumps in memory usage.
2036 : */
2037 7008 : maxBlockSize = pg_prevpower2_size_t(work_mem * (Size) 1024 / 16);
2038 :
2039 : /* But no bigger than ALLOCSET_DEFAULT_MAXSIZE */
2040 7008 : maxBlockSize = Min(maxBlockSize, ALLOCSET_DEFAULT_MAXSIZE);
2041 :
2042 : /* and no smaller than ALLOCSET_DEFAULT_INITSIZE */
2043 7008 : maxBlockSize = Max(maxBlockSize, ALLOCSET_DEFAULT_INITSIZE);
2044 :
2045 7008 : aggstate->hash_tablecxt = BumpContextCreate(aggstate->ss.ps.state->es_query_cxt,
2046 : "HashAgg table context",
2047 : ALLOCSET_DEFAULT_MINSIZE,
2048 : ALLOCSET_DEFAULT_INITSIZE,
2049 : maxBlockSize);
2050 :
2051 7008 : }
2052 :
2053 : /*
2054 : * Choose a reasonable number of buckets for the initial hash table size.
2055 : */
2056 : static long
2057 5446 : hash_choose_num_buckets(double hashentrysize, long ngroups, Size memory)
2058 : {
2059 : long max_nbuckets;
2060 5446 : long nbuckets = ngroups;
2061 :
2062 5446 : max_nbuckets = memory / hashentrysize;
2063 :
2064 : /*
2065 : * Underestimating is better than overestimating. Too many buckets crowd
2066 : * out space for group keys and transition state values.
2067 : */
2068 5446 : max_nbuckets >>= 1;
2069 :
2070 5446 : if (nbuckets > max_nbuckets)
2071 72 : nbuckets = max_nbuckets;
2072 :
2073 5446 : return Max(nbuckets, 1);
2074 : }
2075 :
2076 : /*
2077 : * Determine the number of partitions to create when spilling, which will
2078 : * always be a power of two. If log2_npartitions is non-NULL, set
2079 : * *log2_npartitions to the log2() of the number of partitions.
2080 : */
2081 : static int
2082 15052 : hash_choose_num_partitions(double input_groups, double hashentrysize,
2083 : int used_bits, int *log2_npartitions)
2084 : {
2085 15052 : Size hash_mem_limit = get_hash_memory_limit();
2086 : double partition_limit;
2087 : double mem_wanted;
2088 : double dpartitions;
2089 : int npartitions;
2090 : int partition_bits;
2091 :
2092 : /*
2093 : * Avoid creating so many partitions that the memory requirements of the
2094 : * open partition files are greater than 1/4 of hash_mem.
2095 : */
2096 15052 : partition_limit =
2097 15052 : (hash_mem_limit * 0.25 - HASHAGG_READ_BUFFER_SIZE) /
2098 : HASHAGG_WRITE_BUFFER_SIZE;
2099 :
2100 15052 : mem_wanted = HASHAGG_PARTITION_FACTOR * input_groups * hashentrysize;
2101 :
2102 : /* make enough partitions so that each one is likely to fit in memory */
2103 15052 : dpartitions = 1 + (mem_wanted / hash_mem_limit);
2104 :
2105 15052 : if (dpartitions > partition_limit)
2106 14988 : dpartitions = partition_limit;
2107 :
2108 15052 : if (dpartitions < HASHAGG_MIN_PARTITIONS)
2109 15052 : dpartitions = HASHAGG_MIN_PARTITIONS;
2110 15052 : if (dpartitions > HASHAGG_MAX_PARTITIONS)
2111 0 : dpartitions = HASHAGG_MAX_PARTITIONS;
2112 :
2113 : /* HASHAGG_MAX_PARTITIONS limit makes this safe */
2114 15052 : npartitions = (int) dpartitions;
2115 :
2116 : /* ceil(log2(npartitions)) */
2117 15052 : partition_bits = pg_ceil_log2_32(npartitions);
2118 :
2119 : /* make sure that we don't exhaust the hash bits */
2120 15052 : if (partition_bits + used_bits >= 32)
2121 0 : partition_bits = 32 - used_bits;
2122 :
2123 15052 : if (log2_npartitions != NULL)
2124 12622 : *log2_npartitions = partition_bits;
2125 :
2126 : /* number of partitions will be a power of two */
2127 15052 : npartitions = 1 << partition_bits;
2128 :
2129 15052 : return npartitions;
2130 : }
2131 :
2132 : /*
2133 : * Initialize a freshly-created TupleHashEntry.
2134 : */
2135 : static void
2136 530882 : initialize_hash_entry(AggState *aggstate, TupleHashTable hashtable,
2137 : TupleHashEntry entry)
2138 : {
2139 : AggStatePerGroup pergroup;
2140 : int transno;
2141 :
2142 530882 : aggstate->hash_ngroups_current++;
2143 530882 : hash_agg_check_limits(aggstate);
2144 :
2145 : /* no need to allocate or initialize per-group state */
2146 530882 : if (aggstate->numtrans == 0)
2147 215372 : return;
2148 :
2149 315510 : pergroup = (AggStatePerGroup) TupleHashEntryGetAdditional(hashtable, entry);
2150 :
2151 : /*
2152 : * Initialize aggregates for new tuple group, lookup_hash_entries()
2153 : * already has selected the relevant grouping set.
2154 : */
2155 780422 : for (transno = 0; transno < aggstate->numtrans; transno++)
2156 : {
2157 464912 : AggStatePerTrans pertrans = &aggstate->pertrans[transno];
2158 464912 : AggStatePerGroup pergroupstate = &pergroup[transno];
2159 :
2160 464912 : initialize_aggregate(aggstate, pertrans, pergroupstate);
2161 : }
2162 : }
2163 :
2164 : /*
2165 : * Look up hash entries for the current tuple in all hashed grouping sets.
2166 : *
2167 : * Some entries may be left NULL if we are in "spill mode". The same tuple
2168 : * will belong to different groups for each grouping set, so may match a group
2169 : * already in memory for one set and match a group not in memory for another
2170 : * set. When in "spill mode", the tuple will be spilled for each grouping set
2171 : * where it doesn't match a group in memory.
2172 : *
2173 : * NB: It's possible to spill the same tuple for several different grouping
2174 : * sets. This may seem wasteful, but it's actually a trade-off: if we spill
2175 : * the tuple multiple times for multiple grouping sets, it can be partitioned
2176 : * for each grouping set, making the refilling of the hash table very
2177 : * efficient.
2178 : */
2179 : static void
2180 6919060 : lookup_hash_entries(AggState *aggstate)
2181 : {
2182 6919060 : AggStatePerGroup *pergroup = aggstate->hash_pergroup;
2183 6919060 : TupleTableSlot *outerslot = aggstate->tmpcontext->ecxt_outertuple;
2184 : int setno;
2185 :
2186 13972572 : for (setno = 0; setno < aggstate->num_hashes; setno++)
2187 : {
2188 7053512 : AggStatePerHash perhash = &aggstate->perhash[setno];
2189 7053512 : TupleHashTable hashtable = perhash->hashtable;
2190 7053512 : TupleTableSlot *hashslot = perhash->hashslot;
2191 : TupleHashEntry entry;
2192 : uint32 hash;
2193 7053512 : bool isnew = false;
2194 : bool *p_isnew;
2195 :
2196 : /* if hash table already spilled, don't create new entries */
2197 7053512 : p_isnew = aggstate->hash_spill_mode ? NULL : &isnew;
2198 :
2199 7053512 : select_current_set(aggstate, setno, true);
2200 7053512 : prepare_hash_slot(perhash,
2201 : outerslot,
2202 : hashslot);
2203 :
2204 7053512 : entry = LookupTupleHashEntry(hashtable, hashslot,
2205 : p_isnew, &hash);
2206 :
2207 7053512 : if (entry != NULL)
2208 : {
2209 6286276 : if (isnew)
2210 375202 : initialize_hash_entry(aggstate, hashtable, entry);
2211 6286276 : pergroup[setno] = TupleHashEntryGetAdditional(hashtable, entry);
2212 : }
2213 : else
2214 : {
2215 767236 : HashAggSpill *spill = &aggstate->hash_spills[setno];
2216 767236 : TupleTableSlot *slot = aggstate->tmpcontext->ecxt_outertuple;
2217 :
2218 767236 : if (spill->partitions == NULL)
2219 0 : hashagg_spill_init(spill, aggstate->hash_tapeset, 0,
2220 0 : perhash->aggnode->numGroups,
2221 : aggstate->hashentrysize);
2222 :
2223 767236 : hashagg_spill_tuple(aggstate, spill, slot, hash);
2224 767236 : pergroup[setno] = NULL;
2225 : }
2226 : }
2227 6919060 : }
2228 :
2229 : /*
2230 : * ExecAgg -
2231 : *
2232 : * ExecAgg receives tuples from its outer subplan and aggregates over
2233 : * the appropriate attribute for each aggregate function use (Aggref
2234 : * node) appearing in the targetlist or qual of the node. The number
2235 : * of tuples to aggregate over depends on whether grouped or plain
2236 : * aggregation is selected. In grouped aggregation, we produce a result
2237 : * row for each group; in plain aggregation there's a single result row
2238 : * for the whole query. In either case, the value of each aggregate is
2239 : * stored in the expression context to be used when ExecProject evaluates
2240 : * the result tuple.
2241 : */
2242 : static TupleTableSlot *
2243 823646 : ExecAgg(PlanState *pstate)
2244 : {
2245 823646 : AggState *node = castNode(AggState, pstate);
2246 823646 : TupleTableSlot *result = NULL;
2247 :
2248 823646 : CHECK_FOR_INTERRUPTS();
2249 :
2250 823646 : if (!node->agg_done)
2251 : {
2252 : /* Dispatch based on strategy */
2253 760520 : switch (node->phase->aggstrategy)
2254 : {
2255 487756 : case AGG_HASHED:
2256 487756 : if (!node->table_filled)
2257 17158 : agg_fill_hash_table(node);
2258 : /* FALLTHROUGH */
2259 : case AGG_MIXED:
2260 515118 : result = agg_retrieve_hash_table(node);
2261 515118 : break;
2262 245402 : case AGG_PLAIN:
2263 : case AGG_SORTED:
2264 245402 : result = agg_retrieve_direct(node);
2265 245210 : break;
2266 : }
2267 :
2268 760328 : if (!TupIsNull(result))
2269 741404 : return result;
2270 : }
2271 :
2272 82050 : return NULL;
2273 : }
2274 :
2275 : /*
2276 : * ExecAgg for non-hashed case
2277 : */
2278 : static TupleTableSlot *
2279 245402 : agg_retrieve_direct(AggState *aggstate)
2280 : {
2281 245402 : Agg *node = aggstate->phase->aggnode;
2282 : ExprContext *econtext;
2283 : ExprContext *tmpcontext;
2284 : AggStatePerAgg peragg;
2285 : AggStatePerGroup *pergroups;
2286 : TupleTableSlot *outerslot;
2287 : TupleTableSlot *firstSlot;
2288 : TupleTableSlot *result;
2289 245402 : bool hasGroupingSets = aggstate->phase->numsets > 0;
2290 245402 : int numGroupingSets = Max(aggstate->phase->numsets, 1);
2291 : int currentSet;
2292 : int nextSetSize;
2293 : int numReset;
2294 : int i;
2295 :
2296 : /*
2297 : * get state info from node
2298 : *
2299 : * econtext is the per-output-tuple expression context
2300 : *
2301 : * tmpcontext is the per-input-tuple expression context
2302 : */
2303 245402 : econtext = aggstate->ss.ps.ps_ExprContext;
2304 245402 : tmpcontext = aggstate->tmpcontext;
2305 :
2306 245402 : peragg = aggstate->peragg;
2307 245402 : pergroups = aggstate->pergroups;
2308 245402 : firstSlot = aggstate->ss.ss_ScanTupleSlot;
2309 :
2310 : /*
2311 : * We loop retrieving groups until we find one matching
2312 : * aggstate->ss.ps.qual
2313 : *
2314 : * For grouping sets, we have the invariant that aggstate->projected_set
2315 : * is either -1 (initial call) or the index (starting from 0) in
2316 : * gset_lengths for the group we just completed (either by projecting a
2317 : * row or by discarding it in the qual).
2318 : */
2319 316234 : while (!aggstate->agg_done)
2320 : {
2321 : /*
2322 : * Clear the per-output-tuple context for each group, as well as
2323 : * aggcontext (which contains any pass-by-ref transvalues of the old
2324 : * group). Some aggregate functions store working state in child
2325 : * contexts; those now get reset automatically without us needing to
2326 : * do anything special.
2327 : *
2328 : * We use ReScanExprContext not just ResetExprContext because we want
2329 : * any registered shutdown callbacks to be called. That allows
2330 : * aggregate functions to ensure they've cleaned up any non-memory
2331 : * resources.
2332 : */
2333 316024 : ReScanExprContext(econtext);
2334 :
2335 : /*
2336 : * Determine how many grouping sets need to be reset at this boundary.
2337 : */
2338 316024 : if (aggstate->projected_set >= 0 &&
2339 247052 : aggstate->projected_set < numGroupingSets)
2340 247034 : numReset = aggstate->projected_set + 1;
2341 : else
2342 68990 : numReset = numGroupingSets;
2343 :
2344 : /*
2345 : * numReset can change on a phase boundary, but that's OK; we want to
2346 : * reset the contexts used in _this_ phase, and later, after possibly
2347 : * changing phase, initialize the right number of aggregates for the
2348 : * _new_ phase.
2349 : */
2350 :
2351 654326 : for (i = 0; i < numReset; i++)
2352 : {
2353 338302 : ReScanExprContext(aggstate->aggcontexts[i]);
2354 : }
2355 :
2356 : /*
2357 : * Check if input is complete and there are no more groups to project
2358 : * in this phase; move to next phase or mark as done.
2359 : */
2360 316024 : if (aggstate->input_done == true &&
2361 1614 : aggstate->projected_set >= (numGroupingSets - 1))
2362 : {
2363 798 : if (aggstate->current_phase < aggstate->numphases - 1)
2364 : {
2365 204 : initialize_phase(aggstate, aggstate->current_phase + 1);
2366 204 : aggstate->input_done = false;
2367 204 : aggstate->projected_set = -1;
2368 204 : numGroupingSets = Max(aggstate->phase->numsets, 1);
2369 204 : node = aggstate->phase->aggnode;
2370 204 : numReset = numGroupingSets;
2371 : }
2372 594 : else if (aggstate->aggstrategy == AGG_MIXED)
2373 : {
2374 : /*
2375 : * Mixed mode; we've output all the grouped stuff and have
2376 : * full hashtables, so switch to outputting those.
2377 : */
2378 156 : initialize_phase(aggstate, 0);
2379 156 : aggstate->table_filled = true;
2380 156 : ResetTupleHashIterator(aggstate->perhash[0].hashtable,
2381 : &aggstate->perhash[0].hashiter);
2382 156 : select_current_set(aggstate, 0, true);
2383 156 : return agg_retrieve_hash_table(aggstate);
2384 : }
2385 : else
2386 : {
2387 438 : aggstate->agg_done = true;
2388 438 : break;
2389 : }
2390 : }
2391 :
2392 : /*
2393 : * Get the number of columns in the next grouping set after the last
2394 : * projected one (if any). This is the number of columns to compare to
2395 : * see if we reached the boundary of that set too.
2396 : */
2397 315430 : if (aggstate->projected_set >= 0 &&
2398 246254 : aggstate->projected_set < (numGroupingSets - 1))
2399 27294 : nextSetSize = aggstate->phase->gset_lengths[aggstate->projected_set + 1];
2400 : else
2401 288136 : nextSetSize = 0;
2402 :
2403 : /*----------
2404 : * If a subgroup for the current grouping set is present, project it.
2405 : *
2406 : * We have a new group if:
2407 : * - we're out of input but haven't projected all grouping sets
2408 : * (checked above)
2409 : * OR
2410 : * - we already projected a row that wasn't from the last grouping
2411 : * set
2412 : * AND
2413 : * - the next grouping set has at least one grouping column (since
2414 : * empty grouping sets project only once input is exhausted)
2415 : * AND
2416 : * - the previous and pending rows differ on the grouping columns
2417 : * of the next grouping set
2418 : *----------
2419 : */
2420 315430 : tmpcontext->ecxt_innertuple = econtext->ecxt_outertuple;
2421 315430 : if (aggstate->input_done ||
2422 314614 : (node->aggstrategy != AGG_PLAIN &&
2423 247280 : aggstate->projected_set != -1 &&
2424 245438 : aggstate->projected_set < (numGroupingSets - 1) &&
2425 19946 : nextSetSize > 0 &&
2426 19946 : !ExecQualAndReset(aggstate->phase->eqfunctions[nextSetSize - 1],
2427 : tmpcontext)))
2428 : {
2429 14150 : aggstate->projected_set += 1;
2430 :
2431 : Assert(aggstate->projected_set < numGroupingSets);
2432 14150 : Assert(nextSetSize > 0 || aggstate->input_done);
2433 : }
2434 : else
2435 : {
2436 : /*
2437 : * We no longer care what group we just projected, the next
2438 : * projection will always be the first (or only) grouping set
2439 : * (unless the input proves to be empty).
2440 : */
2441 301280 : aggstate->projected_set = 0;
2442 :
2443 : /*
2444 : * If we don't already have the first tuple of the new group,
2445 : * fetch it from the outer plan.
2446 : */
2447 301280 : if (aggstate->grp_firstTuple == NULL)
2448 : {
2449 69176 : outerslot = fetch_input_tuple(aggstate);
2450 69116 : if (!TupIsNull(outerslot))
2451 : {
2452 : /*
2453 : * Make a copy of the first input tuple; we will use this
2454 : * for comparisons (in group mode) and for projection.
2455 : */
2456 55932 : aggstate->grp_firstTuple = ExecCopySlotHeapTuple(outerslot);
2457 : }
2458 : else
2459 : {
2460 : /* outer plan produced no tuples at all */
2461 13184 : if (hasGroupingSets)
2462 : {
2463 : /*
2464 : * If there was no input at all, we need to project
2465 : * rows only if there are grouping sets of size 0.
2466 : * Note that this implies that there can't be any
2467 : * references to ungrouped Vars, which would otherwise
2468 : * cause issues with the empty output slot.
2469 : *
2470 : * XXX: This is no longer true, we currently deal with
2471 : * this in finalize_aggregates().
2472 : */
2473 78 : aggstate->input_done = true;
2474 :
2475 108 : while (aggstate->phase->gset_lengths[aggstate->projected_set] > 0)
2476 : {
2477 48 : aggstate->projected_set += 1;
2478 48 : if (aggstate->projected_set >= numGroupingSets)
2479 : {
2480 : /*
2481 : * We can't set agg_done here because we might
2482 : * have more phases to do, even though the
2483 : * input is empty. So we need to restart the
2484 : * whole outer loop.
2485 : */
2486 18 : break;
2487 : }
2488 : }
2489 :
2490 78 : if (aggstate->projected_set >= numGroupingSets)
2491 18 : continue;
2492 : }
2493 : else
2494 : {
2495 13106 : aggstate->agg_done = true;
2496 : /* If we are grouping, we should produce no tuples too */
2497 13106 : if (node->aggstrategy != AGG_PLAIN)
2498 156 : return NULL;
2499 : }
2500 : }
2501 : }
2502 :
2503 : /*
2504 : * Initialize working state for a new input tuple group.
2505 : */
2506 301046 : initialize_aggregates(aggstate, pergroups, numReset);
2507 :
2508 301046 : if (aggstate->grp_firstTuple != NULL)
2509 : {
2510 : /*
2511 : * Store the copied first input tuple in the tuple table slot
2512 : * reserved for it. The tuple will be deleted when it is
2513 : * cleared from the slot.
2514 : */
2515 288036 : ExecForceStoreHeapTuple(aggstate->grp_firstTuple,
2516 : firstSlot, true);
2517 288036 : aggstate->grp_firstTuple = NULL; /* don't keep two pointers */
2518 :
2519 : /* set up for first advance_aggregates call */
2520 288036 : tmpcontext->ecxt_outertuple = firstSlot;
2521 :
2522 : /*
2523 : * Process each outer-plan tuple, and then fetch the next one,
2524 : * until we exhaust the outer plan or cross a group boundary.
2525 : */
2526 : for (;;)
2527 : {
2528 : /*
2529 : * During phase 1 only of a mixed agg, we need to update
2530 : * hashtables as well in advance_aggregates.
2531 : */
2532 21797394 : if (aggstate->aggstrategy == AGG_MIXED &&
2533 38062 : aggstate->current_phase == 1)
2534 : {
2535 38062 : lookup_hash_entries(aggstate);
2536 : }
2537 :
2538 : /* Advance the aggregates (or combine functions) */
2539 21797394 : advance_aggregates(aggstate);
2540 :
2541 : /* Reset per-input-tuple context after each tuple */
2542 21797316 : ResetExprContext(tmpcontext);
2543 :
2544 21797316 : outerslot = fetch_input_tuple(aggstate);
2545 21797286 : if (TupIsNull(outerslot))
2546 : {
2547 : /* no more outer-plan tuples available */
2548 :
2549 : /* if we built hash tables, finalize any spills */
2550 55818 : if (aggstate->aggstrategy == AGG_MIXED &&
2551 144 : aggstate->current_phase == 1)
2552 144 : hashagg_finish_initial_spills(aggstate);
2553 :
2554 55818 : if (hasGroupingSets)
2555 : {
2556 720 : aggstate->input_done = true;
2557 720 : break;
2558 : }
2559 : else
2560 : {
2561 55098 : aggstate->agg_done = true;
2562 55098 : break;
2563 : }
2564 : }
2565 : /* set up for next advance_aggregates call */
2566 21741468 : tmpcontext->ecxt_outertuple = outerslot;
2567 :
2568 : /*
2569 : * If we are grouping, check whether we've crossed a group
2570 : * boundary.
2571 : */
2572 21741468 : if (node->aggstrategy != AGG_PLAIN && node->numCols > 0)
2573 : {
2574 2467116 : tmpcontext->ecxt_innertuple = firstSlot;
2575 2467116 : if (!ExecQual(aggstate->phase->eqfunctions[node->numCols - 1],
2576 : tmpcontext))
2577 : {
2578 232110 : aggstate->grp_firstTuple = ExecCopySlotHeapTuple(outerslot);
2579 232110 : break;
2580 : }
2581 : }
2582 : }
2583 : }
2584 :
2585 : /*
2586 : * Use the representative input tuple for any references to
2587 : * non-aggregated input columns in aggregate direct args, the node
2588 : * qual, and the tlist. (If we are not grouping, and there are no
2589 : * input rows at all, we will come here with an empty firstSlot
2590 : * ... but if not grouping, there can't be any references to
2591 : * non-aggregated input columns, so no problem.)
2592 : */
2593 300938 : econtext->ecxt_outertuple = firstSlot;
2594 : }
2595 :
2596 : Assert(aggstate->projected_set >= 0);
2597 :
2598 315088 : currentSet = aggstate->projected_set;
2599 :
2600 315088 : prepare_projection_slot(aggstate, econtext->ecxt_outertuple, currentSet);
2601 :
2602 315088 : select_current_set(aggstate, currentSet, false);
2603 :
2604 315088 : finalize_aggregates(aggstate,
2605 : peragg,
2606 315088 : pergroups[currentSet]);
2607 :
2608 : /*
2609 : * If there's no row to project right now, we must continue rather
2610 : * than returning a null since there might be more groups.
2611 : */
2612 315076 : result = project_aggregates(aggstate);
2613 315064 : if (result)
2614 244250 : return result;
2615 : }
2616 :
2617 : /* No more groups */
2618 648 : return NULL;
2619 : }
2620 :
2621 : /*
2622 : * ExecAgg for hashed case: read input and build hash table
2623 : */
2624 : static void
2625 17158 : agg_fill_hash_table(AggState *aggstate)
2626 : {
2627 : TupleTableSlot *outerslot;
2628 17158 : ExprContext *tmpcontext = aggstate->tmpcontext;
2629 :
2630 : /*
2631 : * Process each outer-plan tuple, and then fetch the next one, until we
2632 : * exhaust the outer plan.
2633 : */
2634 : for (;;)
2635 : {
2636 6898156 : outerslot = fetch_input_tuple(aggstate);
2637 6898156 : if (TupIsNull(outerslot))
2638 : break;
2639 :
2640 : /* set up for lookup_hash_entries and advance_aggregates */
2641 6880998 : tmpcontext->ecxt_outertuple = outerslot;
2642 :
2643 : /* Find or build hashtable entries */
2644 6880998 : lookup_hash_entries(aggstate);
2645 :
2646 : /* Advance the aggregates (or combine functions) */
2647 6880998 : advance_aggregates(aggstate);
2648 :
2649 : /*
2650 : * Reset per-input-tuple context after each tuple, but note that the
2651 : * hash lookups do this too
2652 : */
2653 6880998 : ResetExprContext(aggstate->tmpcontext);
2654 : }
2655 :
2656 : /* finalize spills, if any */
2657 17158 : hashagg_finish_initial_spills(aggstate);
2658 :
2659 17158 : aggstate->table_filled = true;
2660 : /* Initialize to walk the first hash table */
2661 17158 : select_current_set(aggstate, 0, true);
2662 17158 : ResetTupleHashIterator(aggstate->perhash[0].hashtable,
2663 : &aggstate->perhash[0].hashiter);
2664 17158 : }
2665 :
2666 : /*
2667 : * If any data was spilled during hash aggregation, reset the hash table and
2668 : * reprocess one batch of spilled data. After reprocessing a batch, the hash
2669 : * table will again contain data, ready to be consumed by
2670 : * agg_retrieve_hash_table_in_memory().
2671 : *
2672 : * Should only be called after all in memory hash table entries have been
2673 : * finalized and emitted.
2674 : *
2675 : * Return false when input is exhausted and there's no more work to be done;
2676 : * otherwise return true.
2677 : */
2678 : static bool
2679 45064 : agg_refill_hash_table(AggState *aggstate)
2680 : {
2681 : HashAggBatch *batch;
2682 : AggStatePerHash perhash;
2683 : HashAggSpill spill;
2684 45064 : LogicalTapeSet *tapeset = aggstate->hash_tapeset;
2685 45064 : bool spill_initialized = false;
2686 :
2687 45064 : if (aggstate->hash_batches == NIL)
2688 18120 : return false;
2689 :
2690 : /* hash_batches is a stack, with the top item at the end of the list */
2691 26944 : batch = llast(aggstate->hash_batches);
2692 26944 : aggstate->hash_batches = list_delete_last(aggstate->hash_batches);
2693 :
2694 26944 : hash_agg_set_limits(aggstate->hashentrysize, batch->input_card,
2695 : batch->used_bits, &aggstate->hash_mem_limit,
2696 : &aggstate->hash_ngroups_limit, NULL);
2697 :
2698 : /*
2699 : * Each batch only processes one grouping set; set the rest to NULL so
2700 : * that advance_aggregates() knows to ignore them. We don't touch
2701 : * pergroups for sorted grouping sets here, because they will be needed if
2702 : * we rescan later. The expressions for sorted grouping sets will not be
2703 : * evaluated after we recompile anyway.
2704 : */
2705 207428 : MemSet(aggstate->hash_pergroup, 0,
2706 : sizeof(AggStatePerGroup) * aggstate->num_hashes);
2707 :
2708 : /* free memory and reset hash tables */
2709 26944 : ReScanExprContext(aggstate->hashcontext);
2710 26944 : MemoryContextReset(aggstate->hash_tablecxt);
2711 207428 : for (int setno = 0; setno < aggstate->num_hashes; setno++)
2712 180484 : ResetTupleHashTable(aggstate->perhash[setno].hashtable);
2713 :
2714 26944 : aggstate->hash_ngroups_current = 0;
2715 :
2716 : /*
2717 : * In AGG_MIXED mode, hash aggregation happens in phase 1 and the output
2718 : * happens in phase 0. So, we switch to phase 1 when processing a batch,
2719 : * and back to phase 0 after the batch is done.
2720 : */
2721 : Assert(aggstate->current_phase == 0);
2722 26944 : if (aggstate->phase->aggstrategy == AGG_MIXED)
2723 : {
2724 26262 : aggstate->current_phase = 1;
2725 26262 : aggstate->phase = &aggstate->phases[aggstate->current_phase];
2726 : }
2727 :
2728 26944 : select_current_set(aggstate, batch->setno, true);
2729 :
2730 26944 : perhash = &aggstate->perhash[aggstate->current_set];
2731 :
2732 : /*
2733 : * Spilled tuples are always read back as MinimalTuples, which may be
2734 : * different from the outer plan, so recompile the aggregate expressions.
2735 : *
2736 : * We still need the NULL check, because we are only processing one
2737 : * grouping set at a time and the rest will be NULL.
2738 : */
2739 26944 : hashagg_recompile_expressions(aggstate, true, true);
2740 :
2741 26944 : INJECTION_POINT("hash-aggregate-process-batch", NULL);
2742 : for (;;)
2743 1216776 : {
2744 1243720 : TupleTableSlot *spillslot = aggstate->hash_spill_rslot;
2745 1243720 : TupleTableSlot *hashslot = perhash->hashslot;
2746 1243720 : TupleHashTable hashtable = perhash->hashtable;
2747 : TupleHashEntry entry;
2748 : MinimalTuple tuple;
2749 : uint32 hash;
2750 1243720 : bool isnew = false;
2751 1243720 : bool *p_isnew = aggstate->hash_spill_mode ? NULL : &isnew;
2752 :
2753 1243720 : CHECK_FOR_INTERRUPTS();
2754 :
2755 1243720 : tuple = hashagg_batch_read(batch, &hash);
2756 1243720 : if (tuple == NULL)
2757 26944 : break;
2758 :
2759 1216776 : ExecStoreMinimalTuple(tuple, spillslot, true);
2760 1216776 : aggstate->tmpcontext->ecxt_outertuple = spillslot;
2761 :
2762 1216776 : prepare_hash_slot(perhash,
2763 1216776 : aggstate->tmpcontext->ecxt_outertuple,
2764 : hashslot);
2765 1216776 : entry = LookupTupleHashEntryHash(hashtable, hashslot,
2766 : p_isnew, hash);
2767 :
2768 1216776 : if (entry != NULL)
2769 : {
2770 767236 : if (isnew)
2771 155680 : initialize_hash_entry(aggstate, hashtable, entry);
2772 767236 : aggstate->hash_pergroup[batch->setno] = TupleHashEntryGetAdditional(hashtable, entry);
2773 767236 : advance_aggregates(aggstate);
2774 : }
2775 : else
2776 : {
2777 449540 : if (!spill_initialized)
2778 : {
2779 : /*
2780 : * Avoid initializing the spill until we actually need it so
2781 : * that we don't assign tapes that will never be used.
2782 : */
2783 12500 : spill_initialized = true;
2784 12500 : hashagg_spill_init(&spill, tapeset, batch->used_bits,
2785 : batch->input_card, aggstate->hashentrysize);
2786 : }
2787 : /* no memory for a new group, spill */
2788 449540 : hashagg_spill_tuple(aggstate, &spill, spillslot, hash);
2789 :
2790 449540 : aggstate->hash_pergroup[batch->setno] = NULL;
2791 : }
2792 :
2793 : /*
2794 : * Reset per-input-tuple context after each tuple, but note that the
2795 : * hash lookups do this too
2796 : */
2797 1216776 : ResetExprContext(aggstate->tmpcontext);
2798 : }
2799 :
2800 26944 : LogicalTapeClose(batch->input_tape);
2801 :
2802 : /* change back to phase 0 */
2803 26944 : aggstate->current_phase = 0;
2804 26944 : aggstate->phase = &aggstate->phases[aggstate->current_phase];
2805 :
2806 26944 : if (spill_initialized)
2807 : {
2808 12500 : hashagg_spill_finish(aggstate, &spill, batch->setno);
2809 12500 : hash_agg_update_metrics(aggstate, true, spill.npartitions);
2810 : }
2811 : else
2812 14444 : hash_agg_update_metrics(aggstate, true, 0);
2813 :
2814 26944 : aggstate->hash_spill_mode = false;
2815 :
2816 : /* prepare to walk the first hash table */
2817 26944 : select_current_set(aggstate, batch->setno, true);
2818 26944 : ResetTupleHashIterator(aggstate->perhash[batch->setno].hashtable,
2819 : &aggstate->perhash[batch->setno].hashiter);
2820 :
2821 26944 : pfree(batch);
2822 :
2823 26944 : return true;
2824 : }
2825 :
2826 : /*
2827 : * ExecAgg for hashed case: retrieving groups from hash table
2828 : *
2829 : * After exhausting in-memory tuples, also try refilling the hash table using
2830 : * previously-spilled tuples. Only returns NULL after all in-memory and
2831 : * spilled tuples are exhausted.
2832 : */
2833 : static TupleTableSlot *
2834 515274 : agg_retrieve_hash_table(AggState *aggstate)
2835 : {
2836 515274 : TupleTableSlot *result = NULL;
2837 :
2838 1039372 : while (result == NULL)
2839 : {
2840 542218 : result = agg_retrieve_hash_table_in_memory(aggstate);
2841 542218 : if (result == NULL)
2842 : {
2843 45064 : if (!agg_refill_hash_table(aggstate))
2844 : {
2845 18120 : aggstate->agg_done = true;
2846 18120 : break;
2847 : }
2848 : }
2849 : }
2850 :
2851 515274 : return result;
2852 : }
2853 :
2854 : /*
2855 : * Retrieve the groups from the in-memory hash tables without considering any
2856 : * spilled tuples.
2857 : */
2858 : static TupleTableSlot *
2859 542218 : agg_retrieve_hash_table_in_memory(AggState *aggstate)
2860 : {
2861 : ExprContext *econtext;
2862 : AggStatePerAgg peragg;
2863 : AggStatePerGroup pergroup;
2864 : TupleHashEntry entry;
2865 : TupleTableSlot *firstSlot;
2866 : TupleTableSlot *result;
2867 : AggStatePerHash perhash;
2868 :
2869 : /*
2870 : * get state info from node.
2871 : *
2872 : * econtext is the per-output-tuple expression context.
2873 : */
2874 542218 : econtext = aggstate->ss.ps.ps_ExprContext;
2875 542218 : peragg = aggstate->peragg;
2876 542218 : firstSlot = aggstate->ss.ss_ScanTupleSlot;
2877 :
2878 : /*
2879 : * Note that perhash (and therefore anything accessed through it) can
2880 : * change inside the loop, as we change between grouping sets.
2881 : */
2882 542218 : perhash = &aggstate->perhash[aggstate->current_set];
2883 :
2884 : /*
2885 : * We loop retrieving groups until we find one satisfying
2886 : * aggstate->ss.ps.qual
2887 : */
2888 : for (;;)
2889 135954 : {
2890 678172 : TupleTableSlot *hashslot = perhash->hashslot;
2891 678172 : TupleHashTable hashtable = perhash->hashtable;
2892 : int i;
2893 :
2894 678172 : CHECK_FOR_INTERRUPTS();
2895 :
2896 : /*
2897 : * Find the next entry in the hash table
2898 : */
2899 678172 : entry = ScanTupleHashTable(hashtable, &perhash->hashiter);
2900 678172 : if (entry == NULL)
2901 : {
2902 145356 : int nextset = aggstate->current_set + 1;
2903 :
2904 145356 : if (nextset < aggstate->num_hashes)
2905 : {
2906 : /*
2907 : * Switch to next grouping set, reinitialize, and restart the
2908 : * loop.
2909 : */
2910 100292 : select_current_set(aggstate, nextset, true);
2911 :
2912 100292 : perhash = &aggstate->perhash[aggstate->current_set];
2913 :
2914 100292 : ResetTupleHashIterator(perhash->hashtable, &perhash->hashiter);
2915 :
2916 100292 : continue;
2917 : }
2918 : else
2919 : {
2920 45064 : return NULL;
2921 : }
2922 : }
2923 :
2924 : /*
2925 : * Clear the per-output-tuple context for each group
2926 : *
2927 : * We intentionally don't use ReScanExprContext here; if any aggs have
2928 : * registered shutdown callbacks, they mustn't be called yet, since we
2929 : * might not be done with that agg.
2930 : */
2931 532816 : ResetExprContext(econtext);
2932 :
2933 : /*
2934 : * Transform representative tuple back into one with the right
2935 : * columns.
2936 : */
2937 532816 : ExecStoreMinimalTuple(TupleHashEntryGetTuple(entry), hashslot, false);
2938 532816 : slot_getallattrs(hashslot);
2939 :
2940 532816 : ExecClearTuple(firstSlot);
2941 532816 : memset(firstSlot->tts_isnull, true,
2942 532816 : firstSlot->tts_tupleDescriptor->natts * sizeof(bool));
2943 :
2944 1399166 : for (i = 0; i < perhash->numhashGrpCols; i++)
2945 : {
2946 866350 : int varNumber = perhash->hashGrpColIdxInput[i] - 1;
2947 :
2948 866350 : firstSlot->tts_values[varNumber] = hashslot->tts_values[i];
2949 866350 : firstSlot->tts_isnull[varNumber] = hashslot->tts_isnull[i];
2950 : }
2951 532816 : ExecStoreVirtualTuple(firstSlot);
2952 :
2953 532816 : pergroup = (AggStatePerGroup) TupleHashEntryGetAdditional(hashtable, entry);
2954 :
2955 : /*
2956 : * Use the representative input tuple for any references to
2957 : * non-aggregated input columns in the qual and tlist.
2958 : */
2959 532816 : econtext->ecxt_outertuple = firstSlot;
2960 :
2961 532816 : prepare_projection_slot(aggstate,
2962 : econtext->ecxt_outertuple,
2963 : aggstate->current_set);
2964 :
2965 532816 : finalize_aggregates(aggstate, peragg, pergroup);
2966 :
2967 532816 : result = project_aggregates(aggstate);
2968 532816 : if (result)
2969 497154 : return result;
2970 : }
2971 :
2972 : /* No more groups */
2973 : return NULL;
2974 : }
2975 :
2976 : /*
2977 : * hashagg_spill_init
2978 : *
2979 : * Called after we determined that spilling is necessary. Chooses the number
2980 : * of partitions to create, and initializes them.
2981 : */
2982 : static void
2983 12622 : hashagg_spill_init(HashAggSpill *spill, LogicalTapeSet *tapeset, int used_bits,
2984 : double input_groups, double hashentrysize)
2985 : {
2986 : int npartitions;
2987 : int partition_bits;
2988 :
2989 12622 : npartitions = hash_choose_num_partitions(input_groups, hashentrysize,
2990 : used_bits, &partition_bits);
2991 :
2992 : #ifdef USE_INJECTION_POINTS
2993 12622 : if (IS_INJECTION_POINT_ATTACHED("hash-aggregate-single-partition"))
2994 : {
2995 10 : npartitions = 1;
2996 10 : partition_bits = 0;
2997 10 : INJECTION_POINT_CACHED("hash-aggregate-single-partition", NULL);
2998 : }
2999 : #endif
3000 :
3001 12622 : spill->partitions = palloc0(sizeof(LogicalTape *) * npartitions);
3002 12622 : spill->ntuples = palloc0(sizeof(int64) * npartitions);
3003 12622 : spill->hll_card = palloc0(sizeof(hyperLogLogState) * npartitions);
3004 :
3005 63080 : for (int i = 0; i < npartitions; i++)
3006 50458 : spill->partitions[i] = LogicalTapeCreate(tapeset);
3007 :
3008 12622 : spill->shift = 32 - used_bits - partition_bits;
3009 12622 : if (spill->shift < 32)
3010 12612 : spill->mask = (npartitions - 1) << spill->shift;
3011 : else
3012 10 : spill->mask = 0;
3013 12622 : spill->npartitions = npartitions;
3014 :
3015 63080 : for (int i = 0; i < npartitions; i++)
3016 50458 : initHyperLogLog(&spill->hll_card[i], HASHAGG_HLL_BIT_WIDTH);
3017 12622 : }
3018 :
3019 : /*
3020 : * hashagg_spill_tuple
3021 : *
3022 : * No room for new groups in the hash table. Save for later in the appropriate
3023 : * partition.
3024 : */
3025 : static Size
3026 1216776 : hashagg_spill_tuple(AggState *aggstate, HashAggSpill *spill,
3027 : TupleTableSlot *inputslot, uint32 hash)
3028 : {
3029 : TupleTableSlot *spillslot;
3030 : int partition;
3031 : MinimalTuple tuple;
3032 : LogicalTape *tape;
3033 1216776 : int total_written = 0;
3034 : bool shouldFree;
3035 :
3036 : Assert(spill->partitions != NULL);
3037 :
3038 : /* spill only attributes that we actually need */
3039 1216776 : if (!aggstate->all_cols_needed)
3040 : {
3041 1572 : spillslot = aggstate->hash_spill_wslot;
3042 1572 : slot_getsomeattrs(inputslot, aggstate->max_colno_needed);
3043 1572 : ExecClearTuple(spillslot);
3044 4716 : for (int i = 0; i < spillslot->tts_tupleDescriptor->natts; i++)
3045 : {
3046 3144 : if (bms_is_member(i + 1, aggstate->colnos_needed))
3047 : {
3048 1572 : spillslot->tts_values[i] = inputslot->tts_values[i];
3049 1572 : spillslot->tts_isnull[i] = inputslot->tts_isnull[i];
3050 : }
3051 : else
3052 1572 : spillslot->tts_isnull[i] = true;
3053 : }
3054 1572 : ExecStoreVirtualTuple(spillslot);
3055 : }
3056 : else
3057 1215204 : spillslot = inputslot;
3058 :
3059 1216776 : tuple = ExecFetchSlotMinimalTuple(spillslot, &shouldFree);
3060 :
3061 1216776 : if (spill->shift < 32)
3062 1195776 : partition = (hash & spill->mask) >> spill->shift;
3063 : else
3064 21000 : partition = 0;
3065 :
3066 1216776 : spill->ntuples[partition]++;
3067 :
3068 : /*
3069 : * All hash values destined for a given partition have some bits in
3070 : * common, which causes bad HLL cardinality estimates. Hash the hash to
3071 : * get a more uniform distribution.
3072 : */
3073 1216776 : addHyperLogLog(&spill->hll_card[partition], hash_bytes_uint32(hash));
3074 :
3075 1216776 : tape = spill->partitions[partition];
3076 :
3077 1216776 : LogicalTapeWrite(tape, &hash, sizeof(uint32));
3078 1216776 : total_written += sizeof(uint32);
3079 :
3080 1216776 : LogicalTapeWrite(tape, tuple, tuple->t_len);
3081 1216776 : total_written += tuple->t_len;
3082 :
3083 1216776 : if (shouldFree)
3084 767236 : pfree(tuple);
3085 :
3086 1216776 : return total_written;
3087 : }
3088 :
3089 : /*
3090 : * hashagg_batch_new
3091 : *
3092 : * Construct a HashAggBatch item, which represents one iteration of HashAgg to
3093 : * be done.
3094 : */
3095 : static HashAggBatch *
3096 26944 : hashagg_batch_new(LogicalTape *input_tape, int setno,
3097 : int64 input_tuples, double input_card, int used_bits)
3098 : {
3099 26944 : HashAggBatch *batch = palloc0(sizeof(HashAggBatch));
3100 :
3101 26944 : batch->setno = setno;
3102 26944 : batch->used_bits = used_bits;
3103 26944 : batch->input_tape = input_tape;
3104 26944 : batch->input_tuples = input_tuples;
3105 26944 : batch->input_card = input_card;
3106 :
3107 26944 : return batch;
3108 : }
3109 :
3110 : /*
3111 : * hashagg_batch_read
3112 : * read the next tuple from a batch's tape. Return NULL if no more.
3113 : */
3114 : static MinimalTuple
3115 1243720 : hashagg_batch_read(HashAggBatch *batch, uint32 *hashp)
3116 : {
3117 1243720 : LogicalTape *tape = batch->input_tape;
3118 : MinimalTuple tuple;
3119 : uint32 t_len;
3120 : size_t nread;
3121 : uint32 hash;
3122 :
3123 1243720 : nread = LogicalTapeRead(tape, &hash, sizeof(uint32));
3124 1243720 : if (nread == 0)
3125 26944 : return NULL;
3126 1216776 : if (nread != sizeof(uint32))
3127 0 : ereport(ERROR,
3128 : (errcode_for_file_access(),
3129 : errmsg_internal("unexpected EOF for tape %p: requested %zu bytes, read %zu bytes",
3130 : tape, sizeof(uint32), nread)));
3131 1216776 : if (hashp != NULL)
3132 1216776 : *hashp = hash;
3133 :
3134 1216776 : nread = LogicalTapeRead(tape, &t_len, sizeof(t_len));
3135 1216776 : if (nread != sizeof(uint32))
3136 0 : ereport(ERROR,
3137 : (errcode_for_file_access(),
3138 : errmsg_internal("unexpected EOF for tape %p: requested %zu bytes, read %zu bytes",
3139 : tape, sizeof(uint32), nread)));
3140 :
3141 1216776 : tuple = (MinimalTuple) palloc(t_len);
3142 1216776 : tuple->t_len = t_len;
3143 :
3144 1216776 : nread = LogicalTapeRead(tape,
3145 : (char *) tuple + sizeof(uint32),
3146 : t_len - sizeof(uint32));
3147 1216776 : if (nread != t_len - sizeof(uint32))
3148 0 : ereport(ERROR,
3149 : (errcode_for_file_access(),
3150 : errmsg_internal("unexpected EOF for tape %p: requested %zu bytes, read %zu bytes",
3151 : tape, t_len - sizeof(uint32), nread)));
3152 :
3153 1216776 : return tuple;
3154 : }
3155 :
3156 : /*
3157 : * hashagg_finish_initial_spills
3158 : *
3159 : * After a HashAggBatch has been processed, it may have spilled tuples to
3160 : * disk. If so, turn the spilled partitions into new batches that must later
3161 : * be executed.
3162 : */
3163 : static void
3164 17302 : hashagg_finish_initial_spills(AggState *aggstate)
3165 : {
3166 : int setno;
3167 17302 : int total_npartitions = 0;
3168 :
3169 17302 : if (aggstate->hash_spills != NULL)
3170 : {
3171 184 : for (setno = 0; setno < aggstate->num_hashes; setno++)
3172 : {
3173 122 : HashAggSpill *spill = &aggstate->hash_spills[setno];
3174 :
3175 122 : total_npartitions += spill->npartitions;
3176 122 : hashagg_spill_finish(aggstate, spill, setno);
3177 : }
3178 :
3179 : /*
3180 : * We're not processing tuples from outer plan any more; only
3181 : * processing batches of spilled tuples. The initial spill structures
3182 : * are no longer needed.
3183 : */
3184 62 : pfree(aggstate->hash_spills);
3185 62 : aggstate->hash_spills = NULL;
3186 : }
3187 :
3188 17302 : hash_agg_update_metrics(aggstate, false, total_npartitions);
3189 17302 : aggstate->hash_spill_mode = false;
3190 17302 : }
3191 :
3192 : /*
3193 : * hashagg_spill_finish
3194 : *
3195 : * Transform spill partitions into new batches.
3196 : */
3197 : static void
3198 12622 : hashagg_spill_finish(AggState *aggstate, HashAggSpill *spill, int setno)
3199 : {
3200 : int i;
3201 12622 : int used_bits = 32 - spill->shift;
3202 :
3203 12622 : if (spill->npartitions == 0)
3204 0 : return; /* didn't spill */
3205 :
3206 63080 : for (i = 0; i < spill->npartitions; i++)
3207 : {
3208 50458 : LogicalTape *tape = spill->partitions[i];
3209 : HashAggBatch *new_batch;
3210 : double cardinality;
3211 :
3212 : /* if the partition is empty, don't create a new batch of work */
3213 50458 : if (spill->ntuples[i] == 0)
3214 23514 : continue;
3215 :
3216 26944 : cardinality = estimateHyperLogLog(&spill->hll_card[i]);
3217 26944 : freeHyperLogLog(&spill->hll_card[i]);
3218 :
3219 : /* rewinding frees the buffer while not in use */
3220 26944 : LogicalTapeRewindForRead(tape, HASHAGG_READ_BUFFER_SIZE);
3221 :
3222 26944 : new_batch = hashagg_batch_new(tape, setno,
3223 26944 : spill->ntuples[i], cardinality,
3224 : used_bits);
3225 26944 : aggstate->hash_batches = lappend(aggstate->hash_batches, new_batch);
3226 26944 : aggstate->hash_batches_used++;
3227 : }
3228 :
3229 12622 : pfree(spill->ntuples);
3230 12622 : pfree(spill->hll_card);
3231 12622 : pfree(spill->partitions);
3232 : }
3233 :
3234 : /*
3235 : * Free resources related to a spilled HashAgg.
3236 : */
3237 : static void
3238 60016 : hashagg_reset_spill_state(AggState *aggstate)
3239 : {
3240 : /* free spills from initial pass */
3241 60016 : if (aggstate->hash_spills != NULL)
3242 : {
3243 : int setno;
3244 :
3245 0 : for (setno = 0; setno < aggstate->num_hashes; setno++)
3246 : {
3247 0 : HashAggSpill *spill = &aggstate->hash_spills[setno];
3248 :
3249 0 : pfree(spill->ntuples);
3250 0 : pfree(spill->partitions);
3251 : }
3252 0 : pfree(aggstate->hash_spills);
3253 0 : aggstate->hash_spills = NULL;
3254 : }
3255 :
3256 : /* free batches */
3257 60016 : list_free_deep(aggstate->hash_batches);
3258 60016 : aggstate->hash_batches = NIL;
3259 :
3260 : /* close tape set */
3261 60016 : if (aggstate->hash_tapeset != NULL)
3262 : {
3263 62 : LogicalTapeSetClose(aggstate->hash_tapeset);
3264 62 : aggstate->hash_tapeset = NULL;
3265 : }
3266 60016 : }
3267 :
3268 :
3269 : /* -----------------
3270 : * ExecInitAgg
3271 : *
3272 : * Creates the run-time information for the agg node produced by the
3273 : * planner and initializes its outer subtree.
3274 : *
3275 : * -----------------
3276 : */
3277 : AggState *
3278 47824 : ExecInitAgg(Agg *node, EState *estate, int eflags)
3279 : {
3280 : AggState *aggstate;
3281 : AggStatePerAgg peraggs;
3282 : AggStatePerTrans pertransstates;
3283 : AggStatePerGroup *pergroups;
3284 : Plan *outerPlan;
3285 : ExprContext *econtext;
3286 : TupleDesc scanDesc;
3287 : int max_aggno;
3288 : int max_transno;
3289 : int numaggrefs;
3290 : int numaggs;
3291 : int numtrans;
3292 : int phase;
3293 : int phaseidx;
3294 : ListCell *l;
3295 47824 : Bitmapset *all_grouped_cols = NULL;
3296 47824 : int numGroupingSets = 1;
3297 : int numPhases;
3298 : int numHashes;
3299 47824 : int i = 0;
3300 47824 : int j = 0;
3301 88872 : bool use_hashing = (node->aggstrategy == AGG_HASHED ||
3302 41048 : node->aggstrategy == AGG_MIXED);
3303 :
3304 : /* check for unsupported flags */
3305 : Assert(!(eflags & (EXEC_FLAG_BACKWARD | EXEC_FLAG_MARK)));
3306 :
3307 : /*
3308 : * create state structure
3309 : */
3310 47824 : aggstate = makeNode(AggState);
3311 47824 : aggstate->ss.ps.plan = (Plan *) node;
3312 47824 : aggstate->ss.ps.state = estate;
3313 47824 : aggstate->ss.ps.ExecProcNode = ExecAgg;
3314 :
3315 47824 : aggstate->aggs = NIL;
3316 47824 : aggstate->numaggs = 0;
3317 47824 : aggstate->numtrans = 0;
3318 47824 : aggstate->aggstrategy = node->aggstrategy;
3319 47824 : aggstate->aggsplit = node->aggsplit;
3320 47824 : aggstate->maxsets = 0;
3321 47824 : aggstate->projected_set = -1;
3322 47824 : aggstate->current_set = 0;
3323 47824 : aggstate->peragg = NULL;
3324 47824 : aggstate->pertrans = NULL;
3325 47824 : aggstate->curperagg = NULL;
3326 47824 : aggstate->curpertrans = NULL;
3327 47824 : aggstate->input_done = false;
3328 47824 : aggstate->agg_done = false;
3329 47824 : aggstate->pergroups = NULL;
3330 47824 : aggstate->grp_firstTuple = NULL;
3331 47824 : aggstate->sort_in = NULL;
3332 47824 : aggstate->sort_out = NULL;
3333 :
3334 : /*
3335 : * phases[0] always exists, but is dummy in sorted/plain mode
3336 : */
3337 47824 : numPhases = (use_hashing ? 1 : 2);
3338 47824 : numHashes = (use_hashing ? 1 : 0);
3339 :
3340 : /*
3341 : * Calculate the maximum number of grouping sets in any phase; this
3342 : * determines the size of some allocations. Also calculate the number of
3343 : * phases, since all hashed/mixed nodes contribute to only a single phase.
3344 : */
3345 47824 : if (node->groupingSets)
3346 : {
3347 914 : numGroupingSets = list_length(node->groupingSets);
3348 :
3349 1912 : foreach(l, node->chain)
3350 : {
3351 998 : Agg *agg = lfirst(l);
3352 :
3353 998 : numGroupingSets = Max(numGroupingSets,
3354 : list_length(agg->groupingSets));
3355 :
3356 : /*
3357 : * additional AGG_HASHED aggs become part of phase 0, but all
3358 : * others add an extra phase.
3359 : */
3360 998 : if (agg->aggstrategy != AGG_HASHED)
3361 484 : ++numPhases;
3362 : else
3363 514 : ++numHashes;
3364 : }
3365 : }
3366 :
3367 47824 : aggstate->maxsets = numGroupingSets;
3368 47824 : aggstate->numphases = numPhases;
3369 :
3370 47824 : aggstate->aggcontexts = (ExprContext **)
3371 47824 : palloc0(sizeof(ExprContext *) * numGroupingSets);
3372 :
3373 : /*
3374 : * Create expression contexts. We need three or more, one for
3375 : * per-input-tuple processing, one for per-output-tuple processing, one
3376 : * for all the hashtables, and one for each grouping set. The per-tuple
3377 : * memory context of the per-grouping-set ExprContexts (aggcontexts)
3378 : * replaces the standalone memory context formerly used to hold transition
3379 : * values. We cheat a little by using ExecAssignExprContext() to build
3380 : * all of them.
3381 : *
3382 : * NOTE: the details of what is stored in aggcontexts and what is stored
3383 : * in the regular per-query memory context are driven by a simple
3384 : * decision: we want to reset the aggcontext at group boundaries (if not
3385 : * hashing) and in ExecReScanAgg to recover no-longer-wanted space.
3386 : */
3387 47824 : ExecAssignExprContext(estate, &aggstate->ss.ps);
3388 47824 : aggstate->tmpcontext = aggstate->ss.ps.ps_ExprContext;
3389 :
3390 96512 : for (i = 0; i < numGroupingSets; ++i)
3391 : {
3392 48688 : ExecAssignExprContext(estate, &aggstate->ss.ps);
3393 48688 : aggstate->aggcontexts[i] = aggstate->ss.ps.ps_ExprContext;
3394 : }
3395 :
3396 47824 : if (use_hashing)
3397 7008 : hash_create_memory(aggstate);
3398 :
3399 47824 : ExecAssignExprContext(estate, &aggstate->ss.ps);
3400 :
3401 : /*
3402 : * Initialize child nodes.
3403 : *
3404 : * If we are doing a hashed aggregation then the child plan does not need
3405 : * to handle REWIND efficiently; see ExecReScanAgg.
3406 : */
3407 47824 : if (node->aggstrategy == AGG_HASHED)
3408 6776 : eflags &= ~EXEC_FLAG_REWIND;
3409 47824 : outerPlan = outerPlan(node);
3410 47824 : outerPlanState(aggstate) = ExecInitNode(outerPlan, estate, eflags);
3411 :
3412 : /*
3413 : * initialize source tuple type.
3414 : */
3415 47824 : aggstate->ss.ps.outerops =
3416 47824 : ExecGetResultSlotOps(outerPlanState(&aggstate->ss),
3417 : &aggstate->ss.ps.outeropsfixed);
3418 47824 : aggstate->ss.ps.outeropsset = true;
3419 :
3420 47824 : ExecCreateScanSlotFromOuterPlan(estate, &aggstate->ss,
3421 : aggstate->ss.ps.outerops);
3422 47824 : scanDesc = aggstate->ss.ss_ScanTupleSlot->tts_tupleDescriptor;
3423 :
3424 : /*
3425 : * If there are more than two phases (including a potential dummy phase
3426 : * 0), input will be resorted using tuplesort. Need a slot for that.
3427 : */
3428 47824 : if (numPhases > 2)
3429 : {
3430 210 : aggstate->sort_slot = ExecInitExtraTupleSlot(estate, scanDesc,
3431 : &TTSOpsMinimalTuple);
3432 :
3433 : /*
3434 : * The output of the tuplesort, and the output from the outer child
3435 : * might not use the same type of slot. In most cases the child will
3436 : * be a Sort, and thus return a TTSOpsMinimalTuple type slot - but the
3437 : * input can also be presorted due an index, in which case it could be
3438 : * a different type of slot.
3439 : *
3440 : * XXX: For efficiency it would be good to instead/additionally
3441 : * generate expressions with corresponding settings of outerops* for
3442 : * the individual phases - deforming is often a bottleneck for
3443 : * aggregations with lots of rows per group. If there's multiple
3444 : * sorts, we know that all but the first use TTSOpsMinimalTuple (via
3445 : * the nodeAgg.c internal tuplesort).
3446 : */
3447 210 : if (aggstate->ss.ps.outeropsfixed &&
3448 210 : aggstate->ss.ps.outerops != &TTSOpsMinimalTuple)
3449 12 : aggstate->ss.ps.outeropsfixed = false;
3450 : }
3451 :
3452 : /*
3453 : * Initialize result type, slot and projection.
3454 : */
3455 47824 : ExecInitResultTupleSlotTL(&aggstate->ss.ps, &TTSOpsVirtual);
3456 47824 : ExecAssignProjectionInfo(&aggstate->ss.ps, NULL);
3457 :
3458 : /*
3459 : * initialize child expressions
3460 : *
3461 : * We expect the parser to have checked that no aggs contain other agg
3462 : * calls in their arguments (and just to be sure, we verify it again while
3463 : * initializing the plan node). This would make no sense under SQL
3464 : * semantics, and it's forbidden by the spec. Because it is true, we
3465 : * don't need to worry about evaluating the aggs in any particular order.
3466 : *
3467 : * Note: execExpr.c finds Aggrefs for us, and adds them to aggstate->aggs.
3468 : * Aggrefs in the qual are found here; Aggrefs in the targetlist are found
3469 : * during ExecAssignProjectionInfo, above.
3470 : */
3471 47824 : aggstate->ss.ps.qual =
3472 47824 : ExecInitQual(node->plan.qual, (PlanState *) aggstate);
3473 :
3474 : /*
3475 : * We should now have found all Aggrefs in the targetlist and quals.
3476 : */
3477 47824 : numaggrefs = list_length(aggstate->aggs);
3478 47824 : max_aggno = -1;
3479 47824 : max_transno = -1;
3480 101736 : foreach(l, aggstate->aggs)
3481 : {
3482 53912 : Aggref *aggref = (Aggref *) lfirst(l);
3483 :
3484 53912 : max_aggno = Max(max_aggno, aggref->aggno);
3485 53912 : max_transno = Max(max_transno, aggref->aggtransno);
3486 : }
3487 47824 : aggstate->numaggs = numaggs = max_aggno + 1;
3488 47824 : aggstate->numtrans = numtrans = max_transno + 1;
3489 :
3490 : /*
3491 : * For each phase, prepare grouping set data and fmgr lookup data for
3492 : * compare functions. Accumulate all_grouped_cols in passing.
3493 : */
3494 47824 : aggstate->phases = palloc0(numPhases * sizeof(AggStatePerPhaseData));
3495 :
3496 47824 : aggstate->num_hashes = numHashes;
3497 47824 : if (numHashes)
3498 : {
3499 7008 : aggstate->perhash = palloc0(sizeof(AggStatePerHashData) * numHashes);
3500 7008 : aggstate->phases[0].numsets = 0;
3501 7008 : aggstate->phases[0].gset_lengths = palloc(numHashes * sizeof(int));
3502 7008 : aggstate->phases[0].grouped_cols = palloc(numHashes * sizeof(Bitmapset *));
3503 : }
3504 :
3505 47824 : phase = 0;
3506 96646 : for (phaseidx = 0; phaseidx <= list_length(node->chain); ++phaseidx)
3507 : {
3508 : Agg *aggnode;
3509 : Sort *sortnode;
3510 :
3511 48822 : if (phaseidx > 0)
3512 : {
3513 998 : aggnode = list_nth_node(Agg, node->chain, phaseidx - 1);
3514 998 : sortnode = castNode(Sort, outerPlan(aggnode));
3515 : }
3516 : else
3517 : {
3518 47824 : aggnode = node;
3519 47824 : sortnode = NULL;
3520 : }
3521 :
3522 : Assert(phase <= 1 || sortnode);
3523 :
3524 48822 : if (aggnode->aggstrategy == AGG_HASHED
3525 41532 : || aggnode->aggstrategy == AGG_MIXED)
3526 7522 : {
3527 7522 : AggStatePerPhase phasedata = &aggstate->phases[0];
3528 : AggStatePerHash perhash;
3529 7522 : Bitmapset *cols = NULL;
3530 :
3531 : Assert(phase == 0);
3532 7522 : i = phasedata->numsets++;
3533 7522 : perhash = &aggstate->perhash[i];
3534 :
3535 : /* phase 0 always points to the "real" Agg in the hash case */
3536 7522 : phasedata->aggnode = node;
3537 7522 : phasedata->aggstrategy = node->aggstrategy;
3538 :
3539 : /* but the actual Agg node representing this hash is saved here */
3540 7522 : perhash->aggnode = aggnode;
3541 :
3542 7522 : phasedata->gset_lengths[i] = perhash->numCols = aggnode->numCols;
3543 :
3544 18976 : for (j = 0; j < aggnode->numCols; ++j)
3545 11454 : cols = bms_add_member(cols, aggnode->grpColIdx[j]);
3546 :
3547 7522 : phasedata->grouped_cols[i] = cols;
3548 :
3549 7522 : all_grouped_cols = bms_add_members(all_grouped_cols, cols);
3550 7522 : continue;
3551 : }
3552 : else
3553 : {
3554 41300 : AggStatePerPhase phasedata = &aggstate->phases[++phase];
3555 : int num_sets;
3556 :
3557 41300 : phasedata->numsets = num_sets = list_length(aggnode->groupingSets);
3558 :
3559 41300 : if (num_sets)
3560 : {
3561 1000 : phasedata->gset_lengths = palloc(num_sets * sizeof(int));
3562 1000 : phasedata->grouped_cols = palloc(num_sets * sizeof(Bitmapset *));
3563 :
3564 1000 : i = 0;
3565 2936 : foreach(l, aggnode->groupingSets)
3566 : {
3567 1936 : int current_length = list_length(lfirst(l));
3568 1936 : Bitmapset *cols = NULL;
3569 :
3570 : /* planner forces this to be correct */
3571 3804 : for (j = 0; j < current_length; ++j)
3572 1868 : cols = bms_add_member(cols, aggnode->grpColIdx[j]);
3573 :
3574 1936 : phasedata->grouped_cols[i] = cols;
3575 1936 : phasedata->gset_lengths[i] = current_length;
3576 :
3577 1936 : ++i;
3578 : }
3579 :
3580 1000 : all_grouped_cols = bms_add_members(all_grouped_cols,
3581 1000 : phasedata->grouped_cols[0]);
3582 : }
3583 : else
3584 : {
3585 : Assert(phaseidx == 0);
3586 :
3587 40300 : phasedata->gset_lengths = NULL;
3588 40300 : phasedata->grouped_cols = NULL;
3589 : }
3590 :
3591 : /*
3592 : * If we are grouping, precompute fmgr lookup data for inner loop.
3593 : */
3594 41300 : if (aggnode->aggstrategy == AGG_SORTED)
3595 : {
3596 : /*
3597 : * Build a separate function for each subset of columns that
3598 : * need to be compared.
3599 : */
3600 2516 : phasedata->eqfunctions =
3601 2516 : (ExprState **) palloc0(aggnode->numCols * sizeof(ExprState *));
3602 :
3603 : /* for each grouping set */
3604 4156 : for (int k = 0; k < phasedata->numsets; k++)
3605 : {
3606 1640 : int length = phasedata->gset_lengths[k];
3607 :
3608 : /* nothing to do for empty grouping set */
3609 1640 : if (length == 0)
3610 338 : continue;
3611 :
3612 : /* if we already had one of this length, it'll do */
3613 1302 : if (phasedata->eqfunctions[length - 1] != NULL)
3614 138 : continue;
3615 :
3616 1164 : phasedata->eqfunctions[length - 1] =
3617 1164 : execTuplesMatchPrepare(scanDesc,
3618 : length,
3619 1164 : aggnode->grpColIdx,
3620 1164 : aggnode->grpOperators,
3621 1164 : aggnode->grpCollations,
3622 : (PlanState *) aggstate);
3623 : }
3624 :
3625 : /* and for all grouped columns, unless already computed */
3626 2516 : if (aggnode->numCols > 0 &&
3627 2422 : phasedata->eqfunctions[aggnode->numCols - 1] == NULL)
3628 : {
3629 1646 : phasedata->eqfunctions[aggnode->numCols - 1] =
3630 1646 : execTuplesMatchPrepare(scanDesc,
3631 : aggnode->numCols,
3632 1646 : aggnode->grpColIdx,
3633 1646 : aggnode->grpOperators,
3634 1646 : aggnode->grpCollations,
3635 : (PlanState *) aggstate);
3636 : }
3637 : }
3638 :
3639 41300 : phasedata->aggnode = aggnode;
3640 41300 : phasedata->aggstrategy = aggnode->aggstrategy;
3641 41300 : phasedata->sortnode = sortnode;
3642 : }
3643 : }
3644 :
3645 : /*
3646 : * Convert all_grouped_cols to a descending-order list.
3647 : */
3648 47824 : i = -1;
3649 59898 : while ((i = bms_next_member(all_grouped_cols, i)) >= 0)
3650 12074 : aggstate->all_grouped_cols = lcons_int(i, aggstate->all_grouped_cols);
3651 :
3652 : /*
3653 : * Set up aggregate-result storage in the output expr context, and also
3654 : * allocate my private per-agg working storage
3655 : */
3656 47824 : econtext = aggstate->ss.ps.ps_ExprContext;
3657 47824 : econtext->ecxt_aggvalues = (Datum *) palloc0(sizeof(Datum) * numaggs);
3658 47824 : econtext->ecxt_aggnulls = (bool *) palloc0(sizeof(bool) * numaggs);
3659 :
3660 47824 : peraggs = (AggStatePerAgg) palloc0(sizeof(AggStatePerAggData) * numaggs);
3661 47824 : pertransstates = (AggStatePerTrans) palloc0(sizeof(AggStatePerTransData) * numtrans);
3662 :
3663 47824 : aggstate->peragg = peraggs;
3664 47824 : aggstate->pertrans = pertransstates;
3665 :
3666 :
3667 47824 : aggstate->all_pergroups =
3668 47824 : (AggStatePerGroup *) palloc0(sizeof(AggStatePerGroup)
3669 47824 : * (numGroupingSets + numHashes));
3670 47824 : pergroups = aggstate->all_pergroups;
3671 :
3672 47824 : if (node->aggstrategy != AGG_HASHED)
3673 : {
3674 82960 : for (i = 0; i < numGroupingSets; i++)
3675 : {
3676 41912 : pergroups[i] = (AggStatePerGroup) palloc0(sizeof(AggStatePerGroupData)
3677 : * numaggs);
3678 : }
3679 :
3680 41048 : aggstate->pergroups = pergroups;
3681 41048 : pergroups += numGroupingSets;
3682 : }
3683 :
3684 : /*
3685 : * Hashing can only appear in the initial phase.
3686 : */
3687 47824 : if (use_hashing)
3688 : {
3689 7008 : Plan *outerplan = outerPlan(node);
3690 7008 : uint64 totalGroups = 0;
3691 :
3692 7008 : aggstate->hash_spill_rslot = ExecInitExtraTupleSlot(estate, scanDesc,
3693 : &TTSOpsMinimalTuple);
3694 7008 : aggstate->hash_spill_wslot = ExecInitExtraTupleSlot(estate, scanDesc,
3695 : &TTSOpsVirtual);
3696 :
3697 : /* this is an array of pointers, not structures */
3698 7008 : aggstate->hash_pergroup = pergroups;
3699 :
3700 14016 : aggstate->hashentrysize = hash_agg_entry_size(aggstate->numtrans,
3701 7008 : outerplan->plan_width,
3702 : node->transitionSpace);
3703 :
3704 : /*
3705 : * Consider all of the grouping sets together when setting the limits
3706 : * and estimating the number of partitions. This can be inaccurate
3707 : * when there is more than one grouping set, but should still be
3708 : * reasonable.
3709 : */
3710 14530 : for (int k = 0; k < aggstate->num_hashes; k++)
3711 7522 : totalGroups += aggstate->perhash[k].aggnode->numGroups;
3712 :
3713 7008 : hash_agg_set_limits(aggstate->hashentrysize, totalGroups, 0,
3714 : &aggstate->hash_mem_limit,
3715 : &aggstate->hash_ngroups_limit,
3716 : &aggstate->hash_planned_partitions);
3717 7008 : find_hash_columns(aggstate);
3718 :
3719 : /* Skip massive memory allocation if we are just doing EXPLAIN */
3720 7008 : if (!(eflags & EXEC_FLAG_EXPLAIN_ONLY))
3721 5162 : build_hash_tables(aggstate);
3722 :
3723 7008 : aggstate->table_filled = false;
3724 :
3725 : /* Initialize this to 1, meaning nothing spilled, yet */
3726 7008 : aggstate->hash_batches_used = 1;
3727 : }
3728 :
3729 : /*
3730 : * Initialize current phase-dependent values to initial phase. The initial
3731 : * phase is 1 (first sort pass) for all strategies that use sorting (if
3732 : * hashing is being done too, then phase 0 is processed last); but if only
3733 : * hashing is being done, then phase 0 is all there is.
3734 : */
3735 47824 : if (node->aggstrategy == AGG_HASHED)
3736 : {
3737 6776 : aggstate->current_phase = 0;
3738 6776 : initialize_phase(aggstate, 0);
3739 6776 : select_current_set(aggstate, 0, true);
3740 : }
3741 : else
3742 : {
3743 41048 : aggstate->current_phase = 1;
3744 41048 : initialize_phase(aggstate, 1);
3745 41048 : select_current_set(aggstate, 0, false);
3746 : }
3747 :
3748 : /*
3749 : * Perform lookups of aggregate function info, and initialize the
3750 : * unchanging fields of the per-agg and per-trans data.
3751 : */
3752 101730 : foreach(l, aggstate->aggs)
3753 : {
3754 53912 : Aggref *aggref = lfirst(l);
3755 : AggStatePerAgg peragg;
3756 : AggStatePerTrans pertrans;
3757 : Oid aggTransFnInputTypes[FUNC_MAX_ARGS];
3758 : int numAggTransFnArgs;
3759 : int numDirectArgs;
3760 : HeapTuple aggTuple;
3761 : Form_pg_aggregate aggform;
3762 : AclResult aclresult;
3763 : Oid finalfn_oid;
3764 : Oid serialfn_oid,
3765 : deserialfn_oid;
3766 : Oid aggOwner;
3767 : Expr *finalfnexpr;
3768 : Oid aggtranstype;
3769 :
3770 : /* Planner should have assigned aggregate to correct level */
3771 : Assert(aggref->agglevelsup == 0);
3772 : /* ... and the split mode should match */
3773 : Assert(aggref->aggsplit == aggstate->aggsplit);
3774 :
3775 53912 : peragg = &peraggs[aggref->aggno];
3776 :
3777 : /* Check if we initialized the state for this aggregate already. */
3778 53912 : if (peragg->aggref != NULL)
3779 484 : continue;
3780 :
3781 53428 : peragg->aggref = aggref;
3782 53428 : peragg->transno = aggref->aggtransno;
3783 :
3784 : /* Fetch the pg_aggregate row */
3785 53428 : aggTuple = SearchSysCache1(AGGFNOID,
3786 : ObjectIdGetDatum(aggref->aggfnoid));
3787 53428 : if (!HeapTupleIsValid(aggTuple))
3788 0 : elog(ERROR, "cache lookup failed for aggregate %u",
3789 : aggref->aggfnoid);
3790 53428 : aggform = (Form_pg_aggregate) GETSTRUCT(aggTuple);
3791 :
3792 : /* Check permission to call aggregate function */
3793 53428 : aclresult = object_aclcheck(ProcedureRelationId, aggref->aggfnoid, GetUserId(),
3794 : ACL_EXECUTE);
3795 53428 : if (aclresult != ACLCHECK_OK)
3796 6 : aclcheck_error(aclresult, OBJECT_AGGREGATE,
3797 6 : get_func_name(aggref->aggfnoid));
3798 53422 : InvokeFunctionExecuteHook(aggref->aggfnoid);
3799 :
3800 : /* planner recorded transition state type in the Aggref itself */
3801 53422 : aggtranstype = aggref->aggtranstype;
3802 : Assert(OidIsValid(aggtranstype));
3803 :
3804 : /* Final function only required if we're finalizing the aggregates */
3805 53422 : if (DO_AGGSPLIT_SKIPFINAL(aggstate->aggsplit))
3806 5376 : peragg->finalfn_oid = finalfn_oid = InvalidOid;
3807 : else
3808 48046 : peragg->finalfn_oid = finalfn_oid = aggform->aggfinalfn;
3809 :
3810 53422 : serialfn_oid = InvalidOid;
3811 53422 : deserialfn_oid = InvalidOid;
3812 :
3813 : /*
3814 : * Check if serialization/deserialization is required. We only do it
3815 : * for aggregates that have transtype INTERNAL.
3816 : */
3817 53422 : if (aggtranstype == INTERNALOID)
3818 : {
3819 : /*
3820 : * The planner should only have generated a serialize agg node if
3821 : * every aggregate with an INTERNAL state has a serialization
3822 : * function. Verify that.
3823 : */
3824 22632 : if (DO_AGGSPLIT_SERIALIZE(aggstate->aggsplit))
3825 : {
3826 : /* serialization only valid when not running finalfn */
3827 : Assert(DO_AGGSPLIT_SKIPFINAL(aggstate->aggsplit));
3828 :
3829 336 : if (!OidIsValid(aggform->aggserialfn))
3830 0 : elog(ERROR, "serialfunc not provided for serialization aggregation");
3831 336 : serialfn_oid = aggform->aggserialfn;
3832 : }
3833 :
3834 : /* Likewise for deserialization functions */
3835 22632 : if (DO_AGGSPLIT_DESERIALIZE(aggstate->aggsplit))
3836 : {
3837 : /* deserialization only valid when combining states */
3838 : Assert(DO_AGGSPLIT_COMBINE(aggstate->aggsplit));
3839 :
3840 120 : if (!OidIsValid(aggform->aggdeserialfn))
3841 0 : elog(ERROR, "deserialfunc not provided for deserialization aggregation");
3842 120 : deserialfn_oid = aggform->aggdeserialfn;
3843 : }
3844 : }
3845 :
3846 : /* Check that aggregate owner has permission to call component fns */
3847 : {
3848 : HeapTuple procTuple;
3849 :
3850 53422 : procTuple = SearchSysCache1(PROCOID,
3851 : ObjectIdGetDatum(aggref->aggfnoid));
3852 53422 : if (!HeapTupleIsValid(procTuple))
3853 0 : elog(ERROR, "cache lookup failed for function %u",
3854 : aggref->aggfnoid);
3855 53422 : aggOwner = ((Form_pg_proc) GETSTRUCT(procTuple))->proowner;
3856 53422 : ReleaseSysCache(procTuple);
3857 :
3858 53422 : if (OidIsValid(finalfn_oid))
3859 : {
3860 24182 : aclresult = object_aclcheck(ProcedureRelationId, finalfn_oid, aggOwner,
3861 : ACL_EXECUTE);
3862 24182 : if (aclresult != ACLCHECK_OK)
3863 0 : aclcheck_error(aclresult, OBJECT_FUNCTION,
3864 0 : get_func_name(finalfn_oid));
3865 24182 : InvokeFunctionExecuteHook(finalfn_oid);
3866 : }
3867 53422 : if (OidIsValid(serialfn_oid))
3868 : {
3869 336 : aclresult = object_aclcheck(ProcedureRelationId, serialfn_oid, aggOwner,
3870 : ACL_EXECUTE);
3871 336 : if (aclresult != ACLCHECK_OK)
3872 0 : aclcheck_error(aclresult, OBJECT_FUNCTION,
3873 0 : get_func_name(serialfn_oid));
3874 336 : InvokeFunctionExecuteHook(serialfn_oid);
3875 : }
3876 53422 : if (OidIsValid(deserialfn_oid))
3877 : {
3878 120 : aclresult = object_aclcheck(ProcedureRelationId, deserialfn_oid, aggOwner,
3879 : ACL_EXECUTE);
3880 120 : if (aclresult != ACLCHECK_OK)
3881 0 : aclcheck_error(aclresult, OBJECT_FUNCTION,
3882 0 : get_func_name(deserialfn_oid));
3883 120 : InvokeFunctionExecuteHook(deserialfn_oid);
3884 : }
3885 : }
3886 :
3887 : /*
3888 : * Get actual datatypes of the (nominal) aggregate inputs. These
3889 : * could be different from the agg's declared input types, when the
3890 : * agg accepts ANY or a polymorphic type.
3891 : */
3892 53422 : numAggTransFnArgs = get_aggregate_argtypes(aggref,
3893 : aggTransFnInputTypes);
3894 :
3895 : /* Count the "direct" arguments, if any */
3896 53422 : numDirectArgs = list_length(aggref->aggdirectargs);
3897 :
3898 : /* Detect how many arguments to pass to the finalfn */
3899 53422 : if (aggform->aggfinalextra)
3900 16238 : peragg->numFinalArgs = numAggTransFnArgs + 1;
3901 : else
3902 37184 : peragg->numFinalArgs = numDirectArgs + 1;
3903 :
3904 : /* Initialize any direct-argument expressions */
3905 53422 : peragg->aggdirectargs = ExecInitExprList(aggref->aggdirectargs,
3906 : (PlanState *) aggstate);
3907 :
3908 : /*
3909 : * build expression trees using actual argument & result types for the
3910 : * finalfn, if it exists and is required.
3911 : */
3912 53422 : if (OidIsValid(finalfn_oid))
3913 : {
3914 24182 : build_aggregate_finalfn_expr(aggTransFnInputTypes,
3915 : peragg->numFinalArgs,
3916 : aggtranstype,
3917 : aggref->aggtype,
3918 : aggref->inputcollid,
3919 : finalfn_oid,
3920 : &finalfnexpr);
3921 24182 : fmgr_info(finalfn_oid, &peragg->finalfn);
3922 24182 : fmgr_info_set_expr((Node *) finalfnexpr, &peragg->finalfn);
3923 : }
3924 :
3925 : /* get info about the output value's datatype */
3926 53422 : get_typlenbyval(aggref->aggtype,
3927 : &peragg->resulttypeLen,
3928 : &peragg->resulttypeByVal);
3929 :
3930 : /*
3931 : * Build working state for invoking the transition function, if we
3932 : * haven't done it already.
3933 : */
3934 53422 : pertrans = &pertransstates[aggref->aggtransno];
3935 53422 : if (pertrans->aggref == NULL)
3936 : {
3937 : Datum textInitVal;
3938 : Datum initValue;
3939 : bool initValueIsNull;
3940 : Oid transfn_oid;
3941 :
3942 : /*
3943 : * If this aggregation is performing state combines, then instead
3944 : * of using the transition function, we'll use the combine
3945 : * function.
3946 : */
3947 53164 : if (DO_AGGSPLIT_COMBINE(aggstate->aggsplit))
3948 : {
3949 2192 : transfn_oid = aggform->aggcombinefn;
3950 :
3951 : /* If not set then the planner messed up */
3952 2192 : if (!OidIsValid(transfn_oid))
3953 0 : elog(ERROR, "combinefn not set for aggregate function");
3954 : }
3955 : else
3956 50972 : transfn_oid = aggform->aggtransfn;
3957 :
3958 53164 : aclresult = object_aclcheck(ProcedureRelationId, transfn_oid, aggOwner, ACL_EXECUTE);
3959 53164 : if (aclresult != ACLCHECK_OK)
3960 0 : aclcheck_error(aclresult, OBJECT_FUNCTION,
3961 0 : get_func_name(transfn_oid));
3962 53164 : InvokeFunctionExecuteHook(transfn_oid);
3963 :
3964 : /*
3965 : * initval is potentially null, so don't try to access it as a
3966 : * struct field. Must do it the hard way with SysCacheGetAttr.
3967 : */
3968 53164 : textInitVal = SysCacheGetAttr(AGGFNOID, aggTuple,
3969 : Anum_pg_aggregate_agginitval,
3970 : &initValueIsNull);
3971 53164 : if (initValueIsNull)
3972 31816 : initValue = (Datum) 0;
3973 : else
3974 21348 : initValue = GetAggInitVal(textInitVal, aggtranstype);
3975 :
3976 53164 : if (DO_AGGSPLIT_COMBINE(aggstate->aggsplit))
3977 : {
3978 2192 : Oid combineFnInputTypes[] = {aggtranstype,
3979 : aggtranstype};
3980 :
3981 : /*
3982 : * When combining there's only one input, the to-be-combined
3983 : * transition value. The transition value is not counted
3984 : * here.
3985 : */
3986 2192 : pertrans->numTransInputs = 1;
3987 :
3988 : /* aggcombinefn always has two arguments of aggtranstype */
3989 2192 : build_pertrans_for_aggref(pertrans, aggstate, estate,
3990 : aggref, transfn_oid, aggtranstype,
3991 : serialfn_oid, deserialfn_oid,
3992 : initValue, initValueIsNull,
3993 : combineFnInputTypes, 2);
3994 :
3995 : /*
3996 : * Ensure that a combine function to combine INTERNAL states
3997 : * is not strict. This should have been checked during CREATE
3998 : * AGGREGATE, but the strict property could have been changed
3999 : * since then.
4000 : */
4001 2192 : if (pertrans->transfn.fn_strict && aggtranstype == INTERNALOID)
4002 0 : ereport(ERROR,
4003 : (errcode(ERRCODE_INVALID_FUNCTION_DEFINITION),
4004 : errmsg("combine function with transition type %s must not be declared STRICT",
4005 : format_type_be(aggtranstype))));
4006 : }
4007 : else
4008 : {
4009 : /* Detect how many arguments to pass to the transfn */
4010 50972 : if (AGGKIND_IS_ORDERED_SET(aggref->aggkind))
4011 252 : pertrans->numTransInputs = list_length(aggref->args);
4012 : else
4013 50720 : pertrans->numTransInputs = numAggTransFnArgs;
4014 :
4015 50972 : build_pertrans_for_aggref(pertrans, aggstate, estate,
4016 : aggref, transfn_oid, aggtranstype,
4017 : serialfn_oid, deserialfn_oid,
4018 : initValue, initValueIsNull,
4019 : aggTransFnInputTypes,
4020 : numAggTransFnArgs);
4021 :
4022 : /*
4023 : * If the transfn is strict and the initval is NULL, make sure
4024 : * input type and transtype are the same (or at least
4025 : * binary-compatible), so that it's OK to use the first
4026 : * aggregated input value as the initial transValue. This
4027 : * should have been checked at agg definition time, but we
4028 : * must check again in case the transfn's strictness property
4029 : * has been changed.
4030 : */
4031 50972 : if (pertrans->transfn.fn_strict && pertrans->initValueIsNull)
4032 : {
4033 5056 : if (numAggTransFnArgs <= numDirectArgs ||
4034 5056 : !IsBinaryCoercible(aggTransFnInputTypes[numDirectArgs],
4035 : aggtranstype))
4036 0 : ereport(ERROR,
4037 : (errcode(ERRCODE_INVALID_FUNCTION_DEFINITION),
4038 : errmsg("aggregate %u needs to have compatible input type and transition type",
4039 : aggref->aggfnoid)));
4040 : }
4041 : }
4042 : }
4043 : else
4044 258 : pertrans->aggshared = true;
4045 53422 : ReleaseSysCache(aggTuple);
4046 : }
4047 :
4048 : /*
4049 : * Last, check whether any more aggregates got added onto the node while
4050 : * we processed the expressions for the aggregate arguments (including not
4051 : * only the regular arguments and FILTER expressions handled immediately
4052 : * above, but any direct arguments we might've handled earlier). If so,
4053 : * we have nested aggregate functions, which is semantically nonsensical,
4054 : * so complain. (This should have been caught by the parser, so we don't
4055 : * need to work hard on a helpful error message; but we defend against it
4056 : * here anyway, just to be sure.)
4057 : */
4058 47818 : if (numaggrefs != list_length(aggstate->aggs))
4059 0 : ereport(ERROR,
4060 : (errcode(ERRCODE_GROUPING_ERROR),
4061 : errmsg("aggregate function calls cannot be nested")));
4062 :
4063 : /*
4064 : * Build expressions doing all the transition work at once. We build a
4065 : * different one for each phase, as the number of transition function
4066 : * invocation can differ between phases. Note this'll work both for
4067 : * transition and combination functions (although there'll only be one
4068 : * phase in the latter case).
4069 : */
4070 136930 : for (phaseidx = 0; phaseidx < aggstate->numphases; phaseidx++)
4071 : {
4072 89112 : AggStatePerPhase phase = &aggstate->phases[phaseidx];
4073 89112 : bool dohash = false;
4074 89112 : bool dosort = false;
4075 :
4076 : /* phase 0 doesn't necessarily exist */
4077 89112 : if (!phase->aggnode)
4078 40810 : continue;
4079 :
4080 48302 : if (aggstate->aggstrategy == AGG_MIXED && phaseidx == 1)
4081 : {
4082 : /*
4083 : * Phase one, and only phase one, in a mixed agg performs both
4084 : * sorting and aggregation.
4085 : */
4086 232 : dohash = true;
4087 232 : dosort = true;
4088 : }
4089 48070 : else if (aggstate->aggstrategy == AGG_MIXED && phaseidx == 0)
4090 : {
4091 : /*
4092 : * No need to compute a transition function for an AGG_MIXED phase
4093 : * 0 - the contents of the hashtables will have been computed
4094 : * during phase 1.
4095 : */
4096 232 : continue;
4097 : }
4098 47838 : else if (phase->aggstrategy == AGG_PLAIN ||
4099 9230 : phase->aggstrategy == AGG_SORTED)
4100 : {
4101 41062 : dohash = false;
4102 41062 : dosort = true;
4103 : }
4104 6776 : else if (phase->aggstrategy == AGG_HASHED)
4105 : {
4106 6776 : dohash = true;
4107 6776 : dosort = false;
4108 : }
4109 : else
4110 : Assert(false);
4111 :
4112 48070 : phase->evaltrans = ExecBuildAggTrans(aggstate, phase, dosort, dohash,
4113 : false);
4114 :
4115 : /* cache compiled expression for outer slot without NULL check */
4116 48070 : phase->evaltrans_cache[0][0] = phase->evaltrans;
4117 : }
4118 :
4119 47818 : return aggstate;
4120 : }
4121 :
4122 : /*
4123 : * Build the state needed to calculate a state value for an aggregate.
4124 : *
4125 : * This initializes all the fields in 'pertrans'. 'aggref' is the aggregate
4126 : * to initialize the state for. 'transfn_oid', 'aggtranstype', and the rest
4127 : * of the arguments could be calculated from 'aggref', but the caller has
4128 : * calculated them already, so might as well pass them.
4129 : *
4130 : * 'transfn_oid' may be either the Oid of the aggtransfn or the aggcombinefn.
4131 : */
4132 : static void
4133 53164 : build_pertrans_for_aggref(AggStatePerTrans pertrans,
4134 : AggState *aggstate, EState *estate,
4135 : Aggref *aggref,
4136 : Oid transfn_oid, Oid aggtranstype,
4137 : Oid aggserialfn, Oid aggdeserialfn,
4138 : Datum initValue, bool initValueIsNull,
4139 : Oid *inputTypes, int numArguments)
4140 : {
4141 53164 : int numGroupingSets = Max(aggstate->maxsets, 1);
4142 : Expr *transfnexpr;
4143 : int numTransArgs;
4144 53164 : Expr *serialfnexpr = NULL;
4145 53164 : Expr *deserialfnexpr = NULL;
4146 : ListCell *lc;
4147 : int numInputs;
4148 : int numDirectArgs;
4149 : List *sortlist;
4150 : int numSortCols;
4151 : int numDistinctCols;
4152 : int i;
4153 :
4154 : /* Begin filling in the pertrans data */
4155 53164 : pertrans->aggref = aggref;
4156 53164 : pertrans->aggshared = false;
4157 53164 : pertrans->aggCollation = aggref->inputcollid;
4158 53164 : pertrans->transfn_oid = transfn_oid;
4159 53164 : pertrans->serialfn_oid = aggserialfn;
4160 53164 : pertrans->deserialfn_oid = aggdeserialfn;
4161 53164 : pertrans->initValue = initValue;
4162 53164 : pertrans->initValueIsNull = initValueIsNull;
4163 :
4164 : /* Count the "direct" arguments, if any */
4165 53164 : numDirectArgs = list_length(aggref->aggdirectargs);
4166 :
4167 : /* Count the number of aggregated input columns */
4168 53164 : pertrans->numInputs = numInputs = list_length(aggref->args);
4169 :
4170 53164 : pertrans->aggtranstype = aggtranstype;
4171 :
4172 : /* account for the current transition state */
4173 53164 : numTransArgs = pertrans->numTransInputs + 1;
4174 :
4175 : /*
4176 : * Set up infrastructure for calling the transfn. Note that invtransfn is
4177 : * not needed here.
4178 : */
4179 53164 : build_aggregate_transfn_expr(inputTypes,
4180 : numArguments,
4181 : numDirectArgs,
4182 53164 : aggref->aggvariadic,
4183 : aggtranstype,
4184 : aggref->inputcollid,
4185 : transfn_oid,
4186 : InvalidOid,
4187 : &transfnexpr,
4188 : NULL);
4189 :
4190 53164 : fmgr_info(transfn_oid, &pertrans->transfn);
4191 53164 : fmgr_info_set_expr((Node *) transfnexpr, &pertrans->transfn);
4192 :
4193 53164 : pertrans->transfn_fcinfo =
4194 53164 : (FunctionCallInfo) palloc(SizeForFunctionCallInfo(numTransArgs));
4195 53164 : InitFunctionCallInfoData(*pertrans->transfn_fcinfo,
4196 : &pertrans->transfn,
4197 : numTransArgs,
4198 : pertrans->aggCollation,
4199 : (Node *) aggstate, NULL);
4200 :
4201 : /* get info about the state value's datatype */
4202 53164 : get_typlenbyval(aggtranstype,
4203 : &pertrans->transtypeLen,
4204 : &pertrans->transtypeByVal);
4205 :
4206 53164 : if (OidIsValid(aggserialfn))
4207 : {
4208 336 : build_aggregate_serialfn_expr(aggserialfn,
4209 : &serialfnexpr);
4210 336 : fmgr_info(aggserialfn, &pertrans->serialfn);
4211 336 : fmgr_info_set_expr((Node *) serialfnexpr, &pertrans->serialfn);
4212 :
4213 336 : pertrans->serialfn_fcinfo =
4214 336 : (FunctionCallInfo) palloc(SizeForFunctionCallInfo(1));
4215 336 : InitFunctionCallInfoData(*pertrans->serialfn_fcinfo,
4216 : &pertrans->serialfn,
4217 : 1,
4218 : InvalidOid,
4219 : (Node *) aggstate, NULL);
4220 : }
4221 :
4222 53164 : if (OidIsValid(aggdeserialfn))
4223 : {
4224 120 : build_aggregate_deserialfn_expr(aggdeserialfn,
4225 : &deserialfnexpr);
4226 120 : fmgr_info(aggdeserialfn, &pertrans->deserialfn);
4227 120 : fmgr_info_set_expr((Node *) deserialfnexpr, &pertrans->deserialfn);
4228 :
4229 120 : pertrans->deserialfn_fcinfo =
4230 120 : (FunctionCallInfo) palloc(SizeForFunctionCallInfo(2));
4231 120 : InitFunctionCallInfoData(*pertrans->deserialfn_fcinfo,
4232 : &pertrans->deserialfn,
4233 : 2,
4234 : InvalidOid,
4235 : (Node *) aggstate, NULL);
4236 : }
4237 :
4238 : /*
4239 : * If we're doing either DISTINCT or ORDER BY for a plain agg, then we
4240 : * have a list of SortGroupClause nodes; fish out the data in them and
4241 : * stick them into arrays. We ignore ORDER BY for an ordered-set agg,
4242 : * however; the agg's transfn and finalfn are responsible for that.
4243 : *
4244 : * When the planner has set the aggpresorted flag, the input to the
4245 : * aggregate is already correctly sorted. For ORDER BY aggregates we can
4246 : * simply treat these as normal aggregates. For presorted DISTINCT
4247 : * aggregates an extra step must be added to remove duplicate consecutive
4248 : * inputs.
4249 : *
4250 : * Note that by construction, if there is a DISTINCT clause then the ORDER
4251 : * BY clause is a prefix of it (see transformDistinctClause).
4252 : */
4253 53164 : if (AGGKIND_IS_ORDERED_SET(aggref->aggkind))
4254 : {
4255 252 : sortlist = NIL;
4256 252 : numSortCols = numDistinctCols = 0;
4257 252 : pertrans->aggsortrequired = false;
4258 : }
4259 52912 : else if (aggref->aggpresorted && aggref->aggdistinct == NIL)
4260 : {
4261 2098 : sortlist = NIL;
4262 2098 : numSortCols = numDistinctCols = 0;
4263 2098 : pertrans->aggsortrequired = false;
4264 : }
4265 50814 : else if (aggref->aggdistinct)
4266 : {
4267 582 : sortlist = aggref->aggdistinct;
4268 582 : numSortCols = numDistinctCols = list_length(sortlist);
4269 : Assert(numSortCols >= list_length(aggref->aggorder));
4270 582 : pertrans->aggsortrequired = !aggref->aggpresorted;
4271 : }
4272 : else
4273 : {
4274 50232 : sortlist = aggref->aggorder;
4275 50232 : numSortCols = list_length(sortlist);
4276 50232 : numDistinctCols = 0;
4277 50232 : pertrans->aggsortrequired = (numSortCols > 0);
4278 : }
4279 :
4280 53164 : pertrans->numSortCols = numSortCols;
4281 53164 : pertrans->numDistinctCols = numDistinctCols;
4282 :
4283 : /*
4284 : * If we have either sorting or filtering to do, create a tupledesc and
4285 : * slot corresponding to the aggregated inputs (including sort
4286 : * expressions) of the agg.
4287 : */
4288 53164 : if (numSortCols > 0 || aggref->aggfilter)
4289 : {
4290 1430 : pertrans->sortdesc = ExecTypeFromTL(aggref->args);
4291 1430 : pertrans->sortslot =
4292 1430 : ExecInitExtraTupleSlot(estate, pertrans->sortdesc,
4293 : &TTSOpsMinimalTuple);
4294 : }
4295 :
4296 53164 : if (numSortCols > 0)
4297 : {
4298 : /*
4299 : * We don't implement DISTINCT or ORDER BY aggs in the HASHED case
4300 : * (yet)
4301 : */
4302 : Assert(aggstate->aggstrategy != AGG_HASHED && aggstate->aggstrategy != AGG_MIXED);
4303 :
4304 : /* ORDER BY aggregates are not supported with partial aggregation */
4305 : Assert(!DO_AGGSPLIT_COMBINE(aggstate->aggsplit));
4306 :
4307 : /* If we have only one input, we need its len/byval info. */
4308 720 : if (numInputs == 1)
4309 : {
4310 570 : get_typlenbyval(inputTypes[numDirectArgs],
4311 : &pertrans->inputtypeLen,
4312 : &pertrans->inputtypeByVal);
4313 : }
4314 150 : else if (numDistinctCols > 0)
4315 : {
4316 : /* we will need an extra slot to store prior values */
4317 108 : pertrans->uniqslot =
4318 108 : ExecInitExtraTupleSlot(estate, pertrans->sortdesc,
4319 : &TTSOpsMinimalTuple);
4320 : }
4321 :
4322 : /* Extract the sort information for use later */
4323 720 : pertrans->sortColIdx =
4324 720 : (AttrNumber *) palloc(numSortCols * sizeof(AttrNumber));
4325 720 : pertrans->sortOperators =
4326 720 : (Oid *) palloc(numSortCols * sizeof(Oid));
4327 720 : pertrans->sortCollations =
4328 720 : (Oid *) palloc(numSortCols * sizeof(Oid));
4329 720 : pertrans->sortNullsFirst =
4330 720 : (bool *) palloc(numSortCols * sizeof(bool));
4331 :
4332 720 : i = 0;
4333 1638 : foreach(lc, sortlist)
4334 : {
4335 918 : SortGroupClause *sortcl = (SortGroupClause *) lfirst(lc);
4336 918 : TargetEntry *tle = get_sortgroupclause_tle(sortcl, aggref->args);
4337 :
4338 : /* the parser should have made sure of this */
4339 : Assert(OidIsValid(sortcl->sortop));
4340 :
4341 918 : pertrans->sortColIdx[i] = tle->resno;
4342 918 : pertrans->sortOperators[i] = sortcl->sortop;
4343 918 : pertrans->sortCollations[i] = exprCollation((Node *) tle->expr);
4344 918 : pertrans->sortNullsFirst[i] = sortcl->nulls_first;
4345 918 : i++;
4346 : }
4347 : Assert(i == numSortCols);
4348 : }
4349 :
4350 53164 : if (aggref->aggdistinct)
4351 : {
4352 : Oid *ops;
4353 :
4354 : Assert(numArguments > 0);
4355 : Assert(list_length(aggref->aggdistinct) == numDistinctCols);
4356 :
4357 582 : ops = palloc(numDistinctCols * sizeof(Oid));
4358 :
4359 582 : i = 0;
4360 1344 : foreach(lc, aggref->aggdistinct)
4361 762 : ops[i++] = ((SortGroupClause *) lfirst(lc))->eqop;
4362 :
4363 : /* lookup / build the necessary comparators */
4364 582 : if (numDistinctCols == 1)
4365 474 : fmgr_info(get_opcode(ops[0]), &pertrans->equalfnOne);
4366 : else
4367 108 : pertrans->equalfnMulti =
4368 108 : execTuplesMatchPrepare(pertrans->sortdesc,
4369 : numDistinctCols,
4370 108 : pertrans->sortColIdx,
4371 : ops,
4372 108 : pertrans->sortCollations,
4373 : &aggstate->ss.ps);
4374 582 : pfree(ops);
4375 : }
4376 :
4377 53164 : pertrans->sortstates = (Tuplesortstate **)
4378 53164 : palloc0(sizeof(Tuplesortstate *) * numGroupingSets);
4379 53164 : }
4380 :
4381 :
4382 : static Datum
4383 21348 : GetAggInitVal(Datum textInitVal, Oid transtype)
4384 : {
4385 : Oid typinput,
4386 : typioparam;
4387 : char *strInitVal;
4388 : Datum initVal;
4389 :
4390 21348 : getTypeInputInfo(transtype, &typinput, &typioparam);
4391 21348 : strInitVal = TextDatumGetCString(textInitVal);
4392 21348 : initVal = OidInputFunctionCall(typinput, strInitVal,
4393 : typioparam, -1);
4394 21348 : pfree(strInitVal);
4395 21348 : return initVal;
4396 : }
4397 :
4398 : void
4399 47614 : ExecEndAgg(AggState *node)
4400 : {
4401 : PlanState *outerPlan;
4402 : int transno;
4403 47614 : int numGroupingSets = Max(node->maxsets, 1);
4404 : int setno;
4405 :
4406 : /*
4407 : * When ending a parallel worker, copy the statistics gathered by the
4408 : * worker back into shared memory so that it can be picked up by the main
4409 : * process to report in EXPLAIN ANALYZE.
4410 : */
4411 47614 : if (node->shared_info && IsParallelWorker())
4412 : {
4413 : AggregateInstrumentation *si;
4414 :
4415 : Assert(ParallelWorkerNumber <= node->shared_info->num_workers);
4416 168 : si = &node->shared_info->sinstrument[ParallelWorkerNumber];
4417 168 : si->hash_batches_used = node->hash_batches_used;
4418 168 : si->hash_disk_used = node->hash_disk_used;
4419 168 : si->hash_mem_peak = node->hash_mem_peak;
4420 : }
4421 :
4422 : /* Make sure we have closed any open tuplesorts */
4423 :
4424 47614 : if (node->sort_in)
4425 162 : tuplesort_end(node->sort_in);
4426 47614 : if (node->sort_out)
4427 48 : tuplesort_end(node->sort_out);
4428 :
4429 47614 : hashagg_reset_spill_state(node);
4430 :
4431 47614 : if (node->hash_metacxt != NULL)
4432 : {
4433 7000 : MemoryContextDelete(node->hash_metacxt);
4434 7000 : node->hash_metacxt = NULL;
4435 : }
4436 47614 : if (node->hash_tablecxt != NULL)
4437 : {
4438 7000 : MemoryContextDelete(node->hash_tablecxt);
4439 7000 : node->hash_tablecxt = NULL;
4440 : }
4441 :
4442 :
4443 100570 : for (transno = 0; transno < node->numtrans; transno++)
4444 : {
4445 52956 : AggStatePerTrans pertrans = &node->pertrans[transno];
4446 :
4447 106950 : for (setno = 0; setno < numGroupingSets; setno++)
4448 : {
4449 53994 : if (pertrans->sortstates[setno])
4450 0 : tuplesort_end(pertrans->sortstates[setno]);
4451 : }
4452 : }
4453 :
4454 : /* And ensure any agg shutdown callbacks have been called */
4455 96092 : for (setno = 0; setno < numGroupingSets; setno++)
4456 48478 : ReScanExprContext(node->aggcontexts[setno]);
4457 47614 : if (node->hashcontext)
4458 7000 : ReScanExprContext(node->hashcontext);
4459 :
4460 47614 : outerPlan = outerPlanState(node);
4461 47614 : ExecEndNode(outerPlan);
4462 47614 : }
4463 :
4464 : void
4465 54138 : ExecReScanAgg(AggState *node)
4466 : {
4467 54138 : ExprContext *econtext = node->ss.ps.ps_ExprContext;
4468 54138 : PlanState *outerPlan = outerPlanState(node);
4469 54138 : Agg *aggnode = (Agg *) node->ss.ps.plan;
4470 : int transno;
4471 54138 : int numGroupingSets = Max(node->maxsets, 1);
4472 : int setno;
4473 :
4474 54138 : node->agg_done = false;
4475 :
4476 54138 : if (node->aggstrategy == AGG_HASHED)
4477 : {
4478 : /*
4479 : * In the hashed case, if we haven't yet built the hash table then we
4480 : * can just return; nothing done yet, so nothing to undo. If subnode's
4481 : * chgParam is not NULL then it will be re-scanned by ExecProcNode,
4482 : * else no reason to re-scan it at all.
4483 : */
4484 13412 : if (!node->table_filled)
4485 150 : return;
4486 :
4487 : /*
4488 : * If we do have the hash table, and it never spilled, and the subplan
4489 : * does not have any parameter changes, and none of our own parameter
4490 : * changes affect input expressions of the aggregated functions, then
4491 : * we can just rescan the existing hash table; no need to build it
4492 : * again.
4493 : */
4494 13262 : if (outerPlan->chgParam == NULL && !node->hash_ever_spilled &&
4495 914 : !bms_overlap(node->ss.ps.chgParam, aggnode->aggParams))
4496 : {
4497 890 : ResetTupleHashIterator(node->perhash[0].hashtable,
4498 : &node->perhash[0].hashiter);
4499 890 : select_current_set(node, 0, true);
4500 890 : return;
4501 : }
4502 : }
4503 :
4504 : /* Make sure we have closed any open tuplesorts */
4505 120932 : for (transno = 0; transno < node->numtrans; transno++)
4506 : {
4507 135704 : for (setno = 0; setno < numGroupingSets; setno++)
4508 : {
4509 67870 : AggStatePerTrans pertrans = &node->pertrans[transno];
4510 :
4511 67870 : if (pertrans->sortstates[setno])
4512 : {
4513 0 : tuplesort_end(pertrans->sortstates[setno]);
4514 0 : pertrans->sortstates[setno] = NULL;
4515 : }
4516 : }
4517 : }
4518 :
4519 : /*
4520 : * We don't need to ReScanExprContext the output tuple context here;
4521 : * ExecReScan already did it. But we do need to reset our per-grouping-set
4522 : * contexts, which may have transvalues stored in them. (We use rescan
4523 : * rather than just reset because transfns may have registered callbacks
4524 : * that need to be run now.) For the AGG_HASHED case, see below.
4525 : */
4526 :
4527 106232 : for (setno = 0; setno < numGroupingSets; setno++)
4528 : {
4529 53134 : ReScanExprContext(node->aggcontexts[setno]);
4530 : }
4531 :
4532 : /* Release first tuple of group, if we have made a copy */
4533 53098 : if (node->grp_firstTuple != NULL)
4534 : {
4535 0 : heap_freetuple(node->grp_firstTuple);
4536 0 : node->grp_firstTuple = NULL;
4537 : }
4538 53098 : ExecClearTuple(node->ss.ss_ScanTupleSlot);
4539 :
4540 : /* Forget current agg values */
4541 120932 : MemSet(econtext->ecxt_aggvalues, 0, sizeof(Datum) * node->numaggs);
4542 53098 : MemSet(econtext->ecxt_aggnulls, 0, sizeof(bool) * node->numaggs);
4543 :
4544 : /*
4545 : * With AGG_HASHED/MIXED, the hash table is allocated in a sub-context of
4546 : * the hashcontext. This used to be an issue, but now, resetting a context
4547 : * automatically deletes sub-contexts too.
4548 : */
4549 53098 : if (node->aggstrategy == AGG_HASHED || node->aggstrategy == AGG_MIXED)
4550 : {
4551 12402 : hashagg_reset_spill_state(node);
4552 :
4553 12402 : node->hash_ever_spilled = false;
4554 12402 : node->hash_spill_mode = false;
4555 12402 : node->hash_ngroups_current = 0;
4556 :
4557 12402 : ReScanExprContext(node->hashcontext);
4558 12402 : MemoryContextReset(node->hash_tablecxt);
4559 : /* Rebuild an empty hash table */
4560 12402 : build_hash_tables(node);
4561 12402 : node->table_filled = false;
4562 : /* iterator will be reset when the table is filled */
4563 :
4564 12402 : hashagg_recompile_expressions(node, false, false);
4565 : }
4566 :
4567 53098 : if (node->aggstrategy != AGG_HASHED)
4568 : {
4569 : /*
4570 : * Reset the per-group state (in particular, mark transvalues null)
4571 : */
4572 81488 : for (setno = 0; setno < numGroupingSets; setno++)
4573 : {
4574 176406 : MemSet(node->pergroups[setno], 0,
4575 : sizeof(AggStatePerGroupData) * node->numaggs);
4576 : }
4577 :
4578 : /* reset to phase 1 */
4579 40726 : initialize_phase(node, 1);
4580 :
4581 40726 : node->input_done = false;
4582 40726 : node->projected_set = -1;
4583 : }
4584 :
4585 53098 : if (outerPlan->chgParam == NULL)
4586 218 : ExecReScan(outerPlan);
4587 : }
4588 :
4589 :
4590 : /***********************************************************************
4591 : * API exposed to aggregate functions
4592 : ***********************************************************************/
4593 :
4594 :
4595 : /*
4596 : * AggCheckCallContext - test if a SQL function is being called as an aggregate
4597 : *
4598 : * The transition and/or final functions of an aggregate may want to verify
4599 : * that they are being called as aggregates, rather than as plain SQL
4600 : * functions. They should use this function to do so. The return value
4601 : * is nonzero if being called as an aggregate, or zero if not. (Specific
4602 : * nonzero values are AGG_CONTEXT_AGGREGATE or AGG_CONTEXT_WINDOW, but more
4603 : * values could conceivably appear in future.)
4604 : *
4605 : * If aggcontext isn't NULL, the function also stores at *aggcontext the
4606 : * identity of the memory context that aggregate transition values are being
4607 : * stored in. Note that the same aggregate call site (flinfo) may be called
4608 : * interleaved on different transition values in different contexts, so it's
4609 : * not kosher to cache aggcontext under fn_extra. It is, however, kosher to
4610 : * cache it in the transvalue itself (for internal-type transvalues).
4611 : */
4612 : int
4613 5439956 : AggCheckCallContext(FunctionCallInfo fcinfo, MemoryContext *aggcontext)
4614 : {
4615 5439956 : if (fcinfo->context && IsA(fcinfo->context, AggState))
4616 : {
4617 5428570 : if (aggcontext)
4618 : {
4619 2751676 : AggState *aggstate = ((AggState *) fcinfo->context);
4620 2751676 : ExprContext *cxt = aggstate->curaggcontext;
4621 :
4622 2751676 : *aggcontext = cxt->ecxt_per_tuple_memory;
4623 : }
4624 5428570 : return AGG_CONTEXT_AGGREGATE;
4625 : }
4626 11386 : if (fcinfo->context && IsA(fcinfo->context, WindowAggState))
4627 : {
4628 9512 : if (aggcontext)
4629 710 : *aggcontext = ((WindowAggState *) fcinfo->context)->curaggcontext;
4630 9512 : return AGG_CONTEXT_WINDOW;
4631 : }
4632 :
4633 : /* this is just to prevent "uninitialized variable" warnings */
4634 1874 : if (aggcontext)
4635 1826 : *aggcontext = NULL;
4636 1874 : return 0;
4637 : }
4638 :
4639 : /*
4640 : * AggGetAggref - allow an aggregate support function to get its Aggref
4641 : *
4642 : * If the function is being called as an aggregate support function,
4643 : * return the Aggref node for the aggregate call. Otherwise, return NULL.
4644 : *
4645 : * Aggregates sharing the same inputs and transition functions can get
4646 : * merged into a single transition calculation. If the transition function
4647 : * calls AggGetAggref, it will get some one of the Aggrefs for which it is
4648 : * executing. It must therefore not pay attention to the Aggref fields that
4649 : * relate to the final function, as those are indeterminate. But if a final
4650 : * function calls AggGetAggref, it will get a precise result.
4651 : *
4652 : * Note that if an aggregate is being used as a window function, this will
4653 : * return NULL. We could provide a similar function to return the relevant
4654 : * WindowFunc node in such cases, but it's not needed yet.
4655 : */
4656 : Aggref *
4657 246 : AggGetAggref(FunctionCallInfo fcinfo)
4658 : {
4659 246 : if (fcinfo->context && IsA(fcinfo->context, AggState))
4660 : {
4661 246 : AggState *aggstate = (AggState *) fcinfo->context;
4662 : AggStatePerAgg curperagg;
4663 : AggStatePerTrans curpertrans;
4664 :
4665 : /* check curperagg (valid when in a final function) */
4666 246 : curperagg = aggstate->curperagg;
4667 :
4668 246 : if (curperagg)
4669 0 : return curperagg->aggref;
4670 :
4671 : /* check curpertrans (valid when in a transition function) */
4672 246 : curpertrans = aggstate->curpertrans;
4673 :
4674 246 : if (curpertrans)
4675 246 : return curpertrans->aggref;
4676 : }
4677 0 : return NULL;
4678 : }
4679 :
4680 : /*
4681 : * AggGetTempMemoryContext - fetch short-term memory context for aggregates
4682 : *
4683 : * This is useful in agg final functions; the context returned is one that
4684 : * the final function can safely reset as desired. This isn't useful for
4685 : * transition functions, since the context returned MAY (we don't promise)
4686 : * be the same as the context those are called in.
4687 : *
4688 : * As above, this is currently not useful for aggs called as window functions.
4689 : */
4690 : MemoryContext
4691 0 : AggGetTempMemoryContext(FunctionCallInfo fcinfo)
4692 : {
4693 0 : if (fcinfo->context && IsA(fcinfo->context, AggState))
4694 : {
4695 0 : AggState *aggstate = (AggState *) fcinfo->context;
4696 :
4697 0 : return aggstate->tmpcontext->ecxt_per_tuple_memory;
4698 : }
4699 0 : return NULL;
4700 : }
4701 :
4702 : /*
4703 : * AggStateIsShared - find out whether transition state is shared
4704 : *
4705 : * If the function is being called as an aggregate support function,
4706 : * return true if the aggregate's transition state is shared across
4707 : * multiple aggregates, false if it is not.
4708 : *
4709 : * Returns true if not called as an aggregate support function.
4710 : * This is intended as a conservative answer, ie "no you'd better not
4711 : * scribble on your input". In particular, will return true if the
4712 : * aggregate is being used as a window function, which is a scenario
4713 : * in which changing the transition state is a bad idea. We might
4714 : * want to refine the behavior for the window case in future.
4715 : */
4716 : bool
4717 246 : AggStateIsShared(FunctionCallInfo fcinfo)
4718 : {
4719 246 : if (fcinfo->context && IsA(fcinfo->context, AggState))
4720 : {
4721 246 : AggState *aggstate = (AggState *) fcinfo->context;
4722 : AggStatePerAgg curperagg;
4723 : AggStatePerTrans curpertrans;
4724 :
4725 : /* check curperagg (valid when in a final function) */
4726 246 : curperagg = aggstate->curperagg;
4727 :
4728 246 : if (curperagg)
4729 0 : return aggstate->pertrans[curperagg->transno].aggshared;
4730 :
4731 : /* check curpertrans (valid when in a transition function) */
4732 246 : curpertrans = aggstate->curpertrans;
4733 :
4734 246 : if (curpertrans)
4735 246 : return curpertrans->aggshared;
4736 : }
4737 0 : return true;
4738 : }
4739 :
4740 : /*
4741 : * AggRegisterCallback - register a cleanup callback for an aggregate
4742 : *
4743 : * This is useful for aggs to register shutdown callbacks, which will ensure
4744 : * that non-memory resources are freed. The callback will occur just before
4745 : * the associated aggcontext (as returned by AggCheckCallContext) is reset,
4746 : * either between groups or as a result of rescanning the query. The callback
4747 : * will NOT be called on error paths. The typical use-case is for freeing of
4748 : * tuplestores or tuplesorts maintained in aggcontext, or pins held by slots
4749 : * created by the agg functions. (The callback will not be called until after
4750 : * the result of the finalfn is no longer needed, so it's safe for the finalfn
4751 : * to return data that will be freed by the callback.)
4752 : *
4753 : * As above, this is currently not useful for aggs called as window functions.
4754 : */
4755 : void
4756 660 : AggRegisterCallback(FunctionCallInfo fcinfo,
4757 : ExprContextCallbackFunction func,
4758 : Datum arg)
4759 : {
4760 660 : if (fcinfo->context && IsA(fcinfo->context, AggState))
4761 : {
4762 660 : AggState *aggstate = (AggState *) fcinfo->context;
4763 660 : ExprContext *cxt = aggstate->curaggcontext;
4764 :
4765 660 : RegisterExprContextCallback(cxt, func, arg);
4766 :
4767 660 : return;
4768 : }
4769 0 : elog(ERROR, "aggregate function cannot register a callback in this context");
4770 : }
4771 :
4772 :
4773 : /* ----------------------------------------------------------------
4774 : * Parallel Query Support
4775 : * ----------------------------------------------------------------
4776 : */
4777 :
4778 : /* ----------------------------------------------------------------
4779 : * ExecAggEstimate
4780 : *
4781 : * Estimate space required to propagate aggregate statistics.
4782 : * ----------------------------------------------------------------
4783 : */
4784 : void
4785 560 : ExecAggEstimate(AggState *node, ParallelContext *pcxt)
4786 : {
4787 : Size size;
4788 :
4789 : /* don't need this if not instrumenting or no workers */
4790 560 : if (!node->ss.ps.instrument || pcxt->nworkers == 0)
4791 458 : return;
4792 :
4793 102 : size = mul_size(pcxt->nworkers, sizeof(AggregateInstrumentation));
4794 102 : size = add_size(size, offsetof(SharedAggInfo, sinstrument));
4795 102 : shm_toc_estimate_chunk(&pcxt->estimator, size);
4796 102 : shm_toc_estimate_keys(&pcxt->estimator, 1);
4797 : }
4798 :
4799 : /* ----------------------------------------------------------------
4800 : * ExecAggInitializeDSM
4801 : *
4802 : * Initialize DSM space for aggregate statistics.
4803 : * ----------------------------------------------------------------
4804 : */
4805 : void
4806 560 : ExecAggInitializeDSM(AggState *node, ParallelContext *pcxt)
4807 : {
4808 : Size size;
4809 :
4810 : /* don't need this if not instrumenting or no workers */
4811 560 : if (!node->ss.ps.instrument || pcxt->nworkers == 0)
4812 458 : return;
4813 :
4814 102 : size = offsetof(SharedAggInfo, sinstrument)
4815 102 : + pcxt->nworkers * sizeof(AggregateInstrumentation);
4816 102 : node->shared_info = shm_toc_allocate(pcxt->toc, size);
4817 : /* ensure any unfilled slots will contain zeroes */
4818 102 : memset(node->shared_info, 0, size);
4819 102 : node->shared_info->num_workers = pcxt->nworkers;
4820 102 : shm_toc_insert(pcxt->toc, node->ss.ps.plan->plan_node_id,
4821 102 : node->shared_info);
4822 : }
4823 :
4824 : /* ----------------------------------------------------------------
4825 : * ExecAggInitializeWorker
4826 : *
4827 : * Attach worker to DSM space for aggregate statistics.
4828 : * ----------------------------------------------------------------
4829 : */
4830 : void
4831 1556 : ExecAggInitializeWorker(AggState *node, ParallelWorkerContext *pwcxt)
4832 : {
4833 1556 : node->shared_info =
4834 1556 : shm_toc_lookup(pwcxt->toc, node->ss.ps.plan->plan_node_id, true);
4835 1556 : }
4836 :
4837 : /* ----------------------------------------------------------------
4838 : * ExecAggRetrieveInstrumentation
4839 : *
4840 : * Transfer aggregate statistics from DSM to private memory.
4841 : * ----------------------------------------------------------------
4842 : */
4843 : void
4844 102 : ExecAggRetrieveInstrumentation(AggState *node)
4845 : {
4846 : Size size;
4847 : SharedAggInfo *si;
4848 :
4849 102 : if (node->shared_info == NULL)
4850 0 : return;
4851 :
4852 102 : size = offsetof(SharedAggInfo, sinstrument)
4853 102 : + node->shared_info->num_workers * sizeof(AggregateInstrumentation);
4854 102 : si = palloc(size);
4855 102 : memcpy(si, node->shared_info, size);
4856 102 : node->shared_info = si;
4857 : }
|
