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, double nbuckets);
406 : static void hashagg_recompile_expressions(AggState *aggstate, bool minslot,
407 : bool nullcheck);
408 : static void hash_create_memory(AggState *aggstate);
409 : static double hash_choose_num_buckets(double hashentrysize,
410 : double 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 7901824 : select_current_set(AggState *aggstate, int setno, bool is_hash)
457 : {
458 : /*
459 : * When changing this, also adapt ExecAggPlainTransByVal() and
460 : * ExecAggPlainTransByRef().
461 : */
462 7901824 : if (is_hash)
463 7229674 : aggstate->curaggcontext = aggstate->hashcontext;
464 : else
465 672150 : aggstate->curaggcontext = aggstate->aggcontexts[setno];
466 :
467 7901824 : aggstate->current_set = setno;
468 7901824 : }
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 89202 : 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 89202 : if (aggstate->sort_in)
487 : {
488 42 : tuplesort_end(aggstate->sort_in);
489 42 : aggstate->sort_in = NULL;
490 : }
491 :
492 89202 : if (newphase <= 1)
493 : {
494 : /*
495 : * Discard any existing output tuplesort.
496 : */
497 88998 : 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 89202 : 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 89202 : aggstate->current_phase = newphase;
536 89202 : aggstate->phase = &aggstate->phases[newphase];
537 89202 : }
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 28746920 : fetch_input_tuple(AggState *aggstate)
549 : {
550 : TupleTableSlot *slot;
551 :
552 28746920 : 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 28452020 : slot = ExecProcNode(outerPlanState(aggstate));
563 :
564 28746626 : if (!TupIsNull(slot) && aggstate->sort_out)
565 294696 : tuplesort_puttupleslot(aggstate->sort_out, slot);
566 :
567 28746626 : 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 1131886 : 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 1131886 : 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 1131886 : if (pertrans->initValueIsNull)
629 597050 : pergroupstate->transValue = pertrans->initValue;
630 : else
631 : {
632 : MemoryContext oldContext;
633 :
634 534836 : oldContext = MemoryContextSwitchTo(aggstate->curaggcontext->ecxt_per_tuple_memory);
635 1069672 : pergroupstate->transValue = datumCopy(pertrans->initValue,
636 534836 : pertrans->transtypeByVal,
637 534836 : pertrans->transtypeLen);
638 534836 : MemoryContextSwitchTo(oldContext);
639 : }
640 1131886 : 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 1131886 : pergroupstate->noTransValue = pertrans->initValueIsNull;
650 1131886 : }
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 301298 : initialize_aggregates(AggState *aggstate,
667 : AggStatePerGroup *pergroups,
668 : int numReset)
669 : {
670 : int transno;
671 301298 : int numGroupingSets = Max(aggstate->phase->numsets, 1);
672 301298 : int setno = 0;
673 301298 : int numTrans = aggstate->numtrans;
674 301298 : AggStatePerTrans transstates = aggstate->pertrans;
675 :
676 301298 : if (numReset == 0)
677 0 : numReset = numGroupingSets;
678 :
679 616770 : for (setno = 0; setno < numReset; setno++)
680 : {
681 315472 : AggStatePerGroup pergroup = pergroups[setno];
682 :
683 315472 : select_current_set(aggstate, setno, false);
684 :
685 982888 : for (transno = 0; transno < numTrans; transno++)
686 : {
687 667416 : AggStatePerTrans pertrans = &transstates[transno];
688 667416 : AggStatePerGroup pergroupstate = &pergroup[transno];
689 :
690 667416 : initialize_aggregate(aggstate, pertrans, pergroupstate);
691 : }
692 : }
693 301298 : }
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 724314 : advance_transition_function(AggState *aggstate,
708 : AggStatePerTrans pertrans,
709 : AggStatePerGroup pergroupstate)
710 : {
711 724314 : FunctionCallInfo fcinfo = pertrans->transfn_fcinfo;
712 : MemoryContext oldContext;
713 : Datum newVal;
714 :
715 724314 : 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 724314 : oldContext = MemoryContextSwitchTo(aggstate->tmpcontext->ecxt_per_tuple_memory);
763 :
764 : /* set up aggstate->curpertrans for AggGetAggref() */
765 724314 : aggstate->curpertrans = pertrans;
766 :
767 : /*
768 : * OK to call the transition function
769 : */
770 724314 : fcinfo->args[0].value = pergroupstate->transValue;
771 724314 : fcinfo->args[0].isnull = pergroupstate->transValueIsNull;
772 724314 : fcinfo->isnull = false; /* just in case transfn doesn't set it */
773 :
774 724314 : newVal = FunctionCallInvoke(fcinfo);
775 :
776 724314 : 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 724314 : 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 724314 : pergroupstate->transValue = newVal;
800 724314 : pergroupstate->transValueIsNull = fcinfo->isnull;
801 :
802 724314 : 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 29427640 : advance_aggregates(AggState *aggstate)
818 : {
819 29427640 : ExecEvalExprNoReturnSwitchContext(aggstate->phase->evaltrans,
820 : aggstate->tmpcontext);
821 29427562 : }
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 120420 : MemoryContextSwitchTo(oldContext);
899 120420 : continue;
900 : }
901 : else
902 : {
903 724098 : advance_transition_function(aggstate, pertrans, pergroupstate);
904 :
905 724098 : MemoryContextSwitchTo(oldContext);
906 :
907 : /*
908 : * Forget the old value, if any, and remember the new one for
909 : * subsequent equality checks.
910 : */
911 724098 : 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 198810 : oldVal = *newVal;
921 724098 : oldAbbrevVal = newAbbrevVal;
922 724098 : oldIsNull = *isNull;
923 724098 : 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 1118912 : finalize_aggregate(AggState *aggstate,
1045 : AggStatePerAgg peragg,
1046 : AggStatePerGroup pergroupstate,
1047 : Datum *resultVal, bool *resultIsNull)
1048 : {
1049 1118912 : LOCAL_FCINFO(fcinfo, FUNC_MAX_ARGS);
1050 1118912 : bool anynull = false;
1051 : MemoryContext oldContext;
1052 : int i;
1053 : ListCell *lc;
1054 1118912 : AggStatePerTrans pertrans = &aggstate->pertrans[peragg->transno];
1055 :
1056 1118912 : 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 1118912 : i = 1;
1065 1119886 : 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 1118912 : if (OidIsValid(peragg->finalfn_oid))
1080 : {
1081 338914 : int numFinalArgs = peragg->numFinalArgs;
1082 :
1083 : /* set up aggstate->curperagg for AggGetAggref() */
1084 338914 : aggstate->curperagg = peragg;
1085 :
1086 338914 : InitFunctionCallInfoData(*fcinfo, &peragg->finalfn,
1087 : numFinalArgs,
1088 : pertrans->aggCollation,
1089 : (Node *) aggstate, NULL);
1090 :
1091 : /* Fill in the transition state value */
1092 338914 : fcinfo->args[0].value =
1093 338914 : MakeExpandedObjectReadOnly(pergroupstate->transValue,
1094 : pergroupstate->transValueIsNull,
1095 : pertrans->transtypeLen);
1096 338914 : fcinfo->args[0].isnull = pergroupstate->transValueIsNull;
1097 338914 : anynull |= pergroupstate->transValueIsNull;
1098 :
1099 : /* Fill any remaining argument positions with nulls */
1100 491640 : for (; i < numFinalArgs; i++)
1101 : {
1102 152726 : fcinfo->args[i].value = (Datum) 0;
1103 152726 : fcinfo->args[i].isnull = true;
1104 152726 : anynull = true;
1105 : }
1106 :
1107 338914 : 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 338914 : result = FunctionCallInvoke(fcinfo);
1118 338902 : *resultIsNull = fcinfo->isnull;
1119 338902 : *resultVal = MakeExpandedObjectReadOnly(result,
1120 : fcinfo->isnull,
1121 : peragg->resulttypeLen);
1122 : }
1123 338902 : aggstate->curperagg = NULL;
1124 : }
1125 : else
1126 : {
1127 779998 : *resultVal =
1128 779998 : MakeExpandedObjectReadOnly(pergroupstate->transValue,
1129 : pergroupstate->transValueIsNull,
1130 : pertrans->transtypeLen);
1131 779998 : *resultIsNull = pergroupstate->transValueIsNull;
1132 : }
1133 :
1134 1118900 : MemoryContextSwitchTo(oldContext);
1135 1118900 : }
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 16598 : finalize_partialaggregate(AggState *aggstate,
1145 : AggStatePerAgg peragg,
1146 : AggStatePerGroup pergroupstate,
1147 : Datum *resultVal, bool *resultIsNull)
1148 : {
1149 16598 : AggStatePerTrans pertrans = &aggstate->pertrans[peragg->transno];
1150 : MemoryContext oldContext;
1151 :
1152 16598 : 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 16598 : if (OidIsValid(pertrans->serialfn_oid))
1159 : {
1160 : /* Don't call a strict serialization function with NULL input. */
1161 506 : if (pertrans->serialfn.fn_strict && pergroupstate->transValueIsNull)
1162 : {
1163 84 : *resultVal = (Datum) 0;
1164 84 : *resultIsNull = true;
1165 : }
1166 : else
1167 : {
1168 422 : FunctionCallInfo fcinfo = pertrans->serialfn_fcinfo;
1169 : Datum result;
1170 :
1171 422 : fcinfo->args[0].value =
1172 422 : MakeExpandedObjectReadOnly(pergroupstate->transValue,
1173 : pergroupstate->transValueIsNull,
1174 : pertrans->transtypeLen);
1175 422 : fcinfo->args[0].isnull = pergroupstate->transValueIsNull;
1176 422 : fcinfo->isnull = false;
1177 :
1178 422 : result = FunctionCallInvoke(fcinfo);
1179 422 : *resultIsNull = fcinfo->isnull;
1180 422 : *resultVal = MakeExpandedObjectReadOnly(result,
1181 : fcinfo->isnull,
1182 : peragg->resulttypeLen);
1183 : }
1184 : }
1185 : else
1186 : {
1187 16092 : *resultVal =
1188 16092 : MakeExpandedObjectReadOnly(pergroupstate->transValue,
1189 : pergroupstate->transValueIsNull,
1190 : pertrans->transtypeLen);
1191 16092 : *resultIsNull = pergroupstate->transValueIsNull;
1192 : }
1193 :
1194 16598 : MemoryContextSwitchTo(oldContext);
1195 16598 : }
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 8267258 : 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 8267258 : slot_getsomeattrs(inputslot, perhash->largestGrpColIdx);
1210 8267258 : ExecClearTuple(hashslot);
1211 :
1212 20484886 : for (i = 0; i < perhash->numhashGrpCols; i++)
1213 : {
1214 12217628 : int varNumber = perhash->hashGrpColIdxInput[i] - 1;
1215 :
1216 12217628 : hashslot->tts_values[i] = inputslot->tts_values[varNumber];
1217 12217628 : hashslot->tts_isnull[i] = inputslot->tts_isnull[varNumber];
1218 : }
1219 8267258 : ExecStoreVirtualTuple(hashslot);
1220 8267258 : }
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 844536 : prepare_projection_slot(AggState *aggstate, TupleTableSlot *slot, int currentSet)
1248 : {
1249 844536 : if (aggstate->phase->grouped_cols)
1250 : {
1251 558644 : Bitmapset *grouped_cols = aggstate->phase->grouped_cols[currentSet];
1252 :
1253 558644 : aggstate->grouped_cols = grouped_cols;
1254 :
1255 558644 : 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 558584 : else if (aggstate->all_grouped_cols)
1265 : {
1266 : ListCell *lc;
1267 :
1268 : /* all_grouped_cols is arranged in desc order */
1269 558536 : slot_getsomeattrs(slot, linitial_int(aggstate->all_grouped_cols));
1270 :
1271 1525640 : foreach(lc, aggstate->all_grouped_cols)
1272 : {
1273 967104 : int attnum = lfirst_int(lc);
1274 :
1275 967104 : if (!bms_is_member(attnum, grouped_cols))
1276 57832 : slot->tts_isnull[attnum - 1] = true;
1277 : }
1278 : }
1279 : }
1280 844536 : }
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 844536 : finalize_aggregates(AggState *aggstate,
1293 : AggStatePerAgg peraggs,
1294 : AggStatePerGroup pergroup)
1295 : {
1296 844536 : ExprContext *econtext = aggstate->ss.ps.ps_ExprContext;
1297 844536 : Datum *aggvalues = econtext->ecxt_aggvalues;
1298 844536 : 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 1979788 : for (int transno = 0; transno < aggstate->numtrans; transno++)
1306 : {
1307 1135252 : AggStatePerTrans pertrans = &aggstate->pertrans[transno];
1308 : AggStatePerGroup pergroupstate;
1309 :
1310 1135252 : pergroupstate = &pergroup[transno];
1311 :
1312 1135252 : 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 1081410 : 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 1980034 : for (aggno = 0; aggno < aggstate->numaggs; aggno++)
1347 : {
1348 1135510 : AggStatePerAgg peragg = &peraggs[aggno];
1349 1135510 : int transno = peragg->transno;
1350 : AggStatePerGroup pergroupstate;
1351 :
1352 1135510 : pergroupstate = &pergroup[transno];
1353 :
1354 1135510 : if (DO_AGGSPLIT_SKIPFINAL(aggstate->aggsplit))
1355 16598 : finalize_partialaggregate(aggstate, peragg, pergroupstate,
1356 16598 : &aggvalues[aggno], &aggnulls[aggno]);
1357 : else
1358 1118912 : finalize_aggregate(aggstate, peragg, pergroupstate,
1359 1118912 : &aggvalues[aggno], &aggnulls[aggno]);
1360 : }
1361 844524 : }
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 844524 : project_aggregates(AggState *aggstate)
1370 : {
1371 844524 : 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 844524 : 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 738048 : 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 7030 : find_cols(AggState *aggstate, Bitmapset **aggregated, Bitmapset **unaggregated)
1396 : {
1397 7030 : Agg *agg = (Agg *) aggstate->ss.ps.plan;
1398 : FindColsContext context;
1399 :
1400 7030 : context.is_aggref = false;
1401 7030 : context.aggregated = NULL;
1402 7030 : context.unaggregated = NULL;
1403 :
1404 : /* Examine tlist and quals */
1405 7030 : (void) find_cols_walker((Node *) agg->plan.targetlist, &context);
1406 7030 : (void) find_cols_walker((Node *) agg->plan.qual, &context);
1407 :
1408 : /* In some cases, grouping columns will not appear in the tlist */
1409 17908 : for (int i = 0; i < agg->numCols; i++)
1410 10878 : context.unaggregated = bms_add_member(context.unaggregated,
1411 10878 : agg->grpColIdx[i]);
1412 :
1413 7030 : *aggregated = context.aggregated;
1414 7030 : *unaggregated = context.unaggregated;
1415 7030 : }
1416 :
1417 : static bool
1418 83970 : find_cols_walker(Node *node, FindColsContext *context)
1419 : {
1420 83970 : if (node == NULL)
1421 15112 : return false;
1422 68858 : if (IsA(node, Var))
1423 : {
1424 18972 : 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 18972 : if (context->is_aggref)
1430 6098 : context->aggregated = bms_add_member(context->aggregated,
1431 6098 : var->varattno);
1432 : else
1433 12874 : context->unaggregated = bms_add_member(context->unaggregated,
1434 12874 : var->varattno);
1435 18972 : return false;
1436 : }
1437 49886 : if (IsA(node, Aggref))
1438 : {
1439 : Assert(!context->is_aggref);
1440 8582 : context->is_aggref = true;
1441 8582 : expression_tree_walker(node, find_cols_walker, context);
1442 8582 : context->is_aggref = false;
1443 8582 : return false;
1444 : }
1445 41304 : 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 live in the aggstate's hash_tuplescxt
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 17586 : build_hash_tables(AggState *aggstate)
1466 : {
1467 : int setno;
1468 :
1469 35516 : for (setno = 0; setno < aggstate->num_hashes; ++setno)
1470 : {
1471 17930 : AggStatePerHash perhash = &aggstate->perhash[setno];
1472 : double nbuckets;
1473 : Size memory;
1474 :
1475 17930 : if (perhash->hashtable != NULL)
1476 : {
1477 12462 : ResetTupleHashTable(perhash->hashtable);
1478 12462 : continue;
1479 : }
1480 :
1481 5468 : memory = aggstate->hash_mem_limit / aggstate->num_hashes;
1482 :
1483 : /* choose reasonable number of buckets per hashtable */
1484 5468 : nbuckets = hash_choose_num_buckets(aggstate->hashentrysize,
1485 5468 : perhash->aggnode->numGroups,
1486 : memory);
1487 :
1488 : #ifdef USE_INJECTION_POINTS
1489 5468 : if (IS_INJECTION_POINT_ATTACHED("hash-aggregate-oversize-table"))
1490 : {
1491 0 : nbuckets = memory / TupleHashEntrySize();
1492 0 : INJECTION_POINT_CACHED("hash-aggregate-oversize-table", NULL);
1493 : }
1494 : #endif
1495 :
1496 5468 : build_hash_table(aggstate, setno, nbuckets);
1497 : }
1498 :
1499 17586 : aggstate->hash_ngroups_current = 0;
1500 17586 : }
1501 :
1502 : /*
1503 : * Build a single hashtable for this grouping set.
1504 : */
1505 : static void
1506 5468 : build_hash_table(AggState *aggstate, int setno, double nbuckets)
1507 : {
1508 5468 : AggStatePerHash perhash = &aggstate->perhash[setno];
1509 5468 : MemoryContext metacxt = aggstate->hash_metacxt;
1510 5468 : MemoryContext tuplescxt = aggstate->hash_tuplescxt;
1511 5468 : MemoryContext tmpcxt = aggstate->tmpcontext->ecxt_per_tuple_memory;
1512 : Size additionalsize;
1513 :
1514 : Assert(aggstate->aggstrategy == AGG_HASHED ||
1515 : aggstate->aggstrategy == AGG_MIXED);
1516 :
1517 : /*
1518 : * Used to make sure initial hash table allocation does not exceed
1519 : * hash_mem. Note that the estimate does not include space for
1520 : * pass-by-reference transition data values, nor for the representative
1521 : * tuple of each group.
1522 : */
1523 5468 : additionalsize = aggstate->numtrans * sizeof(AggStatePerGroupData);
1524 :
1525 10936 : perhash->hashtable = BuildTupleHashTable(&aggstate->ss.ps,
1526 5468 : perhash->hashslot->tts_tupleDescriptor,
1527 5468 : perhash->hashslot->tts_ops,
1528 : perhash->numCols,
1529 : perhash->hashGrpColIdxHash,
1530 5468 : perhash->eqfuncoids,
1531 : perhash->hashfunctions,
1532 5468 : perhash->aggnode->grpCollations,
1533 : nbuckets,
1534 : additionalsize,
1535 : metacxt,
1536 : tuplescxt,
1537 : tmpcxt,
1538 5468 : DO_AGGSPLIT_SKIPFINAL(aggstate->aggsplit));
1539 5468 : }
1540 :
1541 : /*
1542 : * Compute columns that actually need to be stored in hashtable entries. The
1543 : * incoming tuples from the child plan node will contain grouping columns,
1544 : * other columns referenced in our targetlist and qual, columns used to
1545 : * compute the aggregate functions, and perhaps just junk columns we don't use
1546 : * at all. Only columns of the first two types need to be stored in the
1547 : * hashtable, and getting rid of the others can make the table entries
1548 : * significantly smaller. The hashtable only contains the relevant columns,
1549 : * and is packed/unpacked in lookup_hash_entries() / agg_retrieve_hash_table()
1550 : * into the format of the normal input descriptor.
1551 : *
1552 : * Additional columns, in addition to the columns grouped by, come from two
1553 : * sources: Firstly functionally dependent columns that we don't need to group
1554 : * by themselves, and secondly ctids for row-marks.
1555 : *
1556 : * To eliminate duplicates, we build a bitmapset of the needed columns, and
1557 : * then build an array of the columns included in the hashtable. We might
1558 : * still have duplicates if the passed-in grpColIdx has them, which can happen
1559 : * in edge cases from semijoins/distinct; these can't always be removed,
1560 : * because it's not certain that the duplicate cols will be using the same
1561 : * hash function.
1562 : *
1563 : * Note that the array is preserved over ExecReScanAgg, so we allocate it in
1564 : * the per-query context (unlike the hash table itself).
1565 : */
1566 : static void
1567 7030 : find_hash_columns(AggState *aggstate)
1568 : {
1569 : Bitmapset *base_colnos;
1570 : Bitmapset *aggregated_colnos;
1571 7030 : TupleDesc scanDesc = aggstate->ss.ss_ScanTupleSlot->tts_tupleDescriptor;
1572 7030 : List *outerTlist = outerPlanState(aggstate)->plan->targetlist;
1573 7030 : int numHashes = aggstate->num_hashes;
1574 7030 : EState *estate = aggstate->ss.ps.state;
1575 : int j;
1576 :
1577 : /* Find Vars that will be needed in tlist and qual */
1578 7030 : find_cols(aggstate, &aggregated_colnos, &base_colnos);
1579 7030 : aggstate->colnos_needed = bms_union(base_colnos, aggregated_colnos);
1580 7030 : aggstate->max_colno_needed = 0;
1581 7030 : aggstate->all_cols_needed = true;
1582 :
1583 29642 : for (int i = 0; i < scanDesc->natts; i++)
1584 : {
1585 22612 : int colno = i + 1;
1586 :
1587 22612 : if (bms_is_member(colno, aggstate->colnos_needed))
1588 16290 : aggstate->max_colno_needed = colno;
1589 : else
1590 6322 : aggstate->all_cols_needed = false;
1591 : }
1592 :
1593 14574 : for (j = 0; j < numHashes; ++j)
1594 : {
1595 7544 : AggStatePerHash perhash = &aggstate->perhash[j];
1596 7544 : Bitmapset *colnos = bms_copy(base_colnos);
1597 7544 : AttrNumber *grpColIdx = perhash->aggnode->grpColIdx;
1598 7544 : List *hashTlist = NIL;
1599 : TupleDesc hashDesc;
1600 : int maxCols;
1601 : int i;
1602 :
1603 7544 : perhash->largestGrpColIdx = 0;
1604 :
1605 : /*
1606 : * If we're doing grouping sets, then some Vars might be referenced in
1607 : * tlist/qual for the benefit of other grouping sets, but not needed
1608 : * when hashing; i.e. prepare_projection_slot will null them out, so
1609 : * there'd be no point storing them. Use prepare_projection_slot's
1610 : * logic to determine which.
1611 : */
1612 7544 : if (aggstate->phases[0].grouped_cols)
1613 : {
1614 7544 : Bitmapset *grouped_cols = aggstate->phases[0].grouped_cols[j];
1615 : ListCell *lc;
1616 :
1617 20388 : foreach(lc, aggstate->all_grouped_cols)
1618 : {
1619 12844 : int attnum = lfirst_int(lc);
1620 :
1621 12844 : if (!bms_is_member(attnum, grouped_cols))
1622 1344 : colnos = bms_del_member(colnos, attnum);
1623 : }
1624 : }
1625 :
1626 : /*
1627 : * Compute maximum number of input columns accounting for possible
1628 : * duplications in the grpColIdx array, which can happen in some edge
1629 : * cases where HashAggregate was generated as part of a semijoin or a
1630 : * DISTINCT.
1631 : */
1632 7544 : maxCols = bms_num_members(colnos) + perhash->numCols;
1633 :
1634 7544 : perhash->hashGrpColIdxInput =
1635 7544 : palloc(maxCols * sizeof(AttrNumber));
1636 7544 : perhash->hashGrpColIdxHash =
1637 7544 : palloc(perhash->numCols * sizeof(AttrNumber));
1638 :
1639 : /* Add all the grouping columns to colnos */
1640 19044 : for (i = 0; i < perhash->numCols; i++)
1641 11500 : colnos = bms_add_member(colnos, grpColIdx[i]);
1642 :
1643 : /*
1644 : * First build mapping for columns directly hashed. These are the
1645 : * first, because they'll be accessed when computing hash values and
1646 : * comparing tuples for exact matches. We also build simple mapping
1647 : * for execGrouping, so it knows where to find the to-be-hashed /
1648 : * compared columns in the input.
1649 : */
1650 19044 : for (i = 0; i < perhash->numCols; i++)
1651 : {
1652 11500 : perhash->hashGrpColIdxInput[i] = grpColIdx[i];
1653 11500 : perhash->hashGrpColIdxHash[i] = i + 1;
1654 11500 : perhash->numhashGrpCols++;
1655 : /* delete already mapped columns */
1656 11500 : colnos = bms_del_member(colnos, grpColIdx[i]);
1657 : }
1658 :
1659 : /* and add the remaining columns */
1660 7544 : i = -1;
1661 8818 : while ((i = bms_next_member(colnos, i)) >= 0)
1662 : {
1663 1274 : perhash->hashGrpColIdxInput[perhash->numhashGrpCols] = i;
1664 1274 : perhash->numhashGrpCols++;
1665 : }
1666 :
1667 : /* and build a tuple descriptor for the hashtable */
1668 20318 : for (i = 0; i < perhash->numhashGrpCols; i++)
1669 : {
1670 12774 : int varNumber = perhash->hashGrpColIdxInput[i] - 1;
1671 :
1672 12774 : hashTlist = lappend(hashTlist, list_nth(outerTlist, varNumber));
1673 12774 : perhash->largestGrpColIdx =
1674 12774 : Max(varNumber + 1, perhash->largestGrpColIdx);
1675 : }
1676 :
1677 7544 : hashDesc = ExecTypeFromTL(hashTlist);
1678 :
1679 7544 : execTuplesHashPrepare(perhash->numCols,
1680 7544 : perhash->aggnode->grpOperators,
1681 : &perhash->eqfuncoids,
1682 : &perhash->hashfunctions);
1683 7544 : perhash->hashslot =
1684 7544 : ExecAllocTableSlot(&estate->es_tupleTable, hashDesc,
1685 : &TTSOpsMinimalTuple);
1686 :
1687 7544 : list_free(hashTlist);
1688 7544 : bms_free(colnos);
1689 : }
1690 :
1691 7030 : bms_free(base_colnos);
1692 7030 : }
1693 :
1694 : /*
1695 : * Estimate per-hash-table-entry overhead.
1696 : */
1697 : Size
1698 41704 : hash_agg_entry_size(int numTrans, Size tupleWidth, Size transitionSpace)
1699 : {
1700 : Size tupleChunkSize;
1701 : Size pergroupChunkSize;
1702 : Size transitionChunkSize;
1703 41704 : Size tupleSize = (MAXALIGN(SizeofMinimalTupleHeader) +
1704 : tupleWidth);
1705 41704 : Size pergroupSize = numTrans * sizeof(AggStatePerGroupData);
1706 :
1707 : /*
1708 : * Entries use the Bump allocator, so the chunk sizes are the same as the
1709 : * requested sizes.
1710 : */
1711 41704 : tupleChunkSize = MAXALIGN(tupleSize);
1712 41704 : pergroupChunkSize = pergroupSize;
1713 :
1714 : /*
1715 : * Transition values use AllocSet, which has a chunk header and also uses
1716 : * power-of-two allocations.
1717 : */
1718 41704 : if (transitionSpace > 0)
1719 5398 : transitionChunkSize = CHUNKHDRSZ + pg_nextpower2_size_t(transitionSpace);
1720 : else
1721 36306 : transitionChunkSize = 0;
1722 :
1723 : return
1724 41704 : TupleHashEntrySize() +
1725 41704 : tupleChunkSize +
1726 41704 : pergroupChunkSize +
1727 : transitionChunkSize;
1728 : }
1729 :
1730 : /*
1731 : * hashagg_recompile_expressions()
1732 : *
1733 : * Identifies the right phase, compiles the right expression given the
1734 : * arguments, and then sets phase->evalfunc to that expression.
1735 : *
1736 : * Different versions of the compiled expression are needed depending on
1737 : * whether hash aggregation has spilled or not, and whether it's reading from
1738 : * the outer plan or a tape. Before spilling to disk, the expression reads
1739 : * from the outer plan and does not need to perform a NULL check. After
1740 : * HashAgg begins to spill, new groups will not be created in the hash table,
1741 : * and the AggStatePerGroup array may be NULL; therefore we need to add a null
1742 : * pointer check to the expression. Then, when reading spilled data from a
1743 : * tape, we change the outer slot type to be a fixed minimal tuple slot.
1744 : *
1745 : * It would be wasteful to recompile every time, so cache the compiled
1746 : * expressions in the AggStatePerPhase, and reuse when appropriate.
1747 : */
1748 : static void
1749 65804 : hashagg_recompile_expressions(AggState *aggstate, bool minslot, bool nullcheck)
1750 : {
1751 : AggStatePerPhase phase;
1752 65804 : int i = minslot ? 1 : 0;
1753 65804 : int j = nullcheck ? 1 : 0;
1754 :
1755 : Assert(aggstate->aggstrategy == AGG_HASHED ||
1756 : aggstate->aggstrategy == AGG_MIXED);
1757 :
1758 65804 : if (aggstate->aggstrategy == AGG_HASHED)
1759 13232 : phase = &aggstate->phases[0];
1760 : else /* AGG_MIXED */
1761 52572 : phase = &aggstate->phases[1];
1762 :
1763 65804 : if (phase->evaltrans_cache[i][j] == NULL)
1764 : {
1765 88 : const TupleTableSlotOps *outerops = aggstate->ss.ps.outerops;
1766 88 : bool outerfixed = aggstate->ss.ps.outeropsfixed;
1767 88 : bool dohash = true;
1768 88 : bool dosort = false;
1769 :
1770 : /*
1771 : * If minslot is true, that means we are processing a spilled batch
1772 : * (inside agg_refill_hash_table()), and we must not advance the
1773 : * sorted grouping sets.
1774 : */
1775 88 : if (aggstate->aggstrategy == AGG_MIXED && !minslot)
1776 12 : dosort = true;
1777 :
1778 : /* temporarily change the outerops while compiling the expression */
1779 88 : if (minslot)
1780 : {
1781 44 : aggstate->ss.ps.outerops = &TTSOpsMinimalTuple;
1782 44 : aggstate->ss.ps.outeropsfixed = true;
1783 : }
1784 :
1785 88 : phase->evaltrans_cache[i][j] = ExecBuildAggTrans(aggstate, phase,
1786 : dosort, dohash,
1787 : nullcheck);
1788 :
1789 : /* change back */
1790 88 : aggstate->ss.ps.outerops = outerops;
1791 88 : aggstate->ss.ps.outeropsfixed = outerfixed;
1792 : }
1793 :
1794 65804 : phase->evaltrans = phase->evaltrans_cache[i][j];
1795 65804 : }
1796 :
1797 : /*
1798 : * Set limits that trigger spilling to avoid exceeding hash_mem. Consider the
1799 : * number of partitions we expect to create (if we do spill).
1800 : *
1801 : * There are two limits: a memory limit, and also an ngroups limit. The
1802 : * ngroups limit becomes important when we expect transition values to grow
1803 : * substantially larger than the initial value.
1804 : */
1805 : void
1806 66214 : hash_agg_set_limits(double hashentrysize, double input_groups, int used_bits,
1807 : Size *mem_limit, uint64 *ngroups_limit,
1808 : int *num_partitions)
1809 : {
1810 : int npartitions;
1811 : Size partition_mem;
1812 66214 : Size hash_mem_limit = get_hash_memory_limit();
1813 :
1814 : /* if not expected to spill, use all of hash_mem */
1815 66214 : if (input_groups * hashentrysize <= hash_mem_limit)
1816 : {
1817 63784 : if (num_partitions != NULL)
1818 39174 : *num_partitions = 0;
1819 63784 : *mem_limit = hash_mem_limit;
1820 63784 : *ngroups_limit = hash_mem_limit / hashentrysize;
1821 63784 : return;
1822 : }
1823 :
1824 : /*
1825 : * Calculate expected memory requirements for spilling, which is the size
1826 : * of the buffers needed for all the tapes that need to be open at once.
1827 : * Then, subtract that from the memory available for holding hash tables.
1828 : */
1829 2430 : npartitions = hash_choose_num_partitions(input_groups,
1830 : hashentrysize,
1831 : used_bits,
1832 : NULL);
1833 2430 : if (num_partitions != NULL)
1834 96 : *num_partitions = npartitions;
1835 :
1836 2430 : partition_mem =
1837 2430 : HASHAGG_READ_BUFFER_SIZE +
1838 : HASHAGG_WRITE_BUFFER_SIZE * npartitions;
1839 :
1840 : /*
1841 : * Don't set the limit below 3/4 of hash_mem. In that case, we are at the
1842 : * minimum number of partitions, so we aren't going to dramatically exceed
1843 : * work mem anyway.
1844 : */
1845 2430 : if (hash_mem_limit > 4 * partition_mem)
1846 0 : *mem_limit = hash_mem_limit - partition_mem;
1847 : else
1848 2430 : *mem_limit = hash_mem_limit * 0.75;
1849 :
1850 2430 : if (*mem_limit > hashentrysize)
1851 2430 : *ngroups_limit = *mem_limit / hashentrysize;
1852 : else
1853 0 : *ngroups_limit = 1;
1854 : }
1855 :
1856 : /*
1857 : * hash_agg_check_limits
1858 : *
1859 : * After adding a new group to the hash table, check whether we need to enter
1860 : * spill mode. Allocations may happen without adding new groups (for instance,
1861 : * if the transition state size grows), so this check is imperfect.
1862 : */
1863 : static void
1864 527238 : hash_agg_check_limits(AggState *aggstate)
1865 : {
1866 527238 : uint64 ngroups = aggstate->hash_ngroups_current;
1867 527238 : Size meta_mem = MemoryContextMemAllocated(aggstate->hash_metacxt,
1868 : true);
1869 527238 : Size entry_mem = MemoryContextMemAllocated(aggstate->hash_tuplescxt,
1870 : true);
1871 527238 : Size tval_mem = MemoryContextMemAllocated(aggstate->hashcontext->ecxt_per_tuple_memory,
1872 : true);
1873 527238 : Size total_mem = meta_mem + entry_mem + tval_mem;
1874 527238 : bool do_spill = false;
1875 :
1876 : #ifdef USE_INJECTION_POINTS
1877 527238 : if (ngroups >= 1000)
1878 : {
1879 95650 : if (IS_INJECTION_POINT_ATTACHED("hash-aggregate-spill-1000"))
1880 : {
1881 10 : do_spill = true;
1882 10 : INJECTION_POINT_CACHED("hash-aggregate-spill-1000", NULL);
1883 : }
1884 : }
1885 : #endif
1886 :
1887 : /*
1888 : * Don't spill unless there's at least one group in the hash table so we
1889 : * can be sure to make progress even in edge cases.
1890 : */
1891 527238 : if (aggstate->hash_ngroups_current > 0 &&
1892 527238 : (total_mem > aggstate->hash_mem_limit ||
1893 500826 : ngroups > aggstate->hash_ngroups_limit))
1894 : {
1895 26448 : do_spill = true;
1896 : }
1897 :
1898 527238 : if (do_spill)
1899 26458 : hash_agg_enter_spill_mode(aggstate);
1900 527238 : }
1901 :
1902 : /*
1903 : * Enter "spill mode", meaning that no new groups are added to any of the hash
1904 : * tables. Tuples that would create a new group are instead spilled, and
1905 : * processed later.
1906 : */
1907 : static void
1908 26458 : hash_agg_enter_spill_mode(AggState *aggstate)
1909 : {
1910 26458 : INJECTION_POINT("hash-aggregate-enter-spill-mode", NULL);
1911 26458 : aggstate->hash_spill_mode = true;
1912 26458 : hashagg_recompile_expressions(aggstate, aggstate->table_filled, true);
1913 :
1914 26458 : if (!aggstate->hash_ever_spilled)
1915 : {
1916 : Assert(aggstate->hash_tapeset == NULL);
1917 : Assert(aggstate->hash_spills == NULL);
1918 :
1919 62 : aggstate->hash_ever_spilled = true;
1920 :
1921 62 : aggstate->hash_tapeset = LogicalTapeSetCreate(true, NULL, -1);
1922 :
1923 62 : aggstate->hash_spills = palloc(sizeof(HashAggSpill) * aggstate->num_hashes);
1924 :
1925 184 : for (int setno = 0; setno < aggstate->num_hashes; setno++)
1926 : {
1927 122 : AggStatePerHash perhash = &aggstate->perhash[setno];
1928 122 : HashAggSpill *spill = &aggstate->hash_spills[setno];
1929 :
1930 122 : hashagg_spill_init(spill, aggstate->hash_tapeset, 0,
1931 122 : perhash->aggnode->numGroups,
1932 : aggstate->hashentrysize);
1933 : }
1934 : }
1935 26458 : }
1936 :
1937 : /*
1938 : * Update metrics after filling the hash table.
1939 : *
1940 : * If reading from the outer plan, from_tape should be false; if reading from
1941 : * another tape, from_tape should be true.
1942 : */
1943 : static void
1944 44254 : hash_agg_update_metrics(AggState *aggstate, bool from_tape, int npartitions)
1945 : {
1946 : Size meta_mem;
1947 : Size entry_mem;
1948 : Size hashkey_mem;
1949 : Size buffer_mem;
1950 : Size total_mem;
1951 :
1952 44254 : if (aggstate->aggstrategy != AGG_MIXED &&
1953 17848 : aggstate->aggstrategy != AGG_HASHED)
1954 0 : return;
1955 :
1956 : /* memory for the hash table itself */
1957 44254 : meta_mem = MemoryContextMemAllocated(aggstate->hash_metacxt, true);
1958 :
1959 : /* memory for hash entries */
1960 44254 : entry_mem = MemoryContextMemAllocated(aggstate->hash_tuplescxt, true);
1961 :
1962 : /* memory for byref transition states */
1963 44254 : hashkey_mem = MemoryContextMemAllocated(aggstate->hashcontext->ecxt_per_tuple_memory, true);
1964 :
1965 : /* memory for read/write tape buffers, if spilled */
1966 44254 : buffer_mem = npartitions * HASHAGG_WRITE_BUFFER_SIZE;
1967 44254 : if (from_tape)
1968 26944 : buffer_mem += HASHAGG_READ_BUFFER_SIZE;
1969 :
1970 : /* update peak mem */
1971 44254 : total_mem = meta_mem + entry_mem + hashkey_mem + buffer_mem;
1972 44254 : if (total_mem > aggstate->hash_mem_peak)
1973 4958 : aggstate->hash_mem_peak = total_mem;
1974 :
1975 : /* update disk usage */
1976 44254 : if (aggstate->hash_tapeset != NULL)
1977 : {
1978 27006 : uint64 disk_used = LogicalTapeSetBlocks(aggstate->hash_tapeset) * (BLCKSZ / 1024);
1979 :
1980 27006 : if (aggstate->hash_disk_used < disk_used)
1981 52 : aggstate->hash_disk_used = disk_used;
1982 : }
1983 :
1984 : /* update hashentrysize estimate based on contents */
1985 44254 : if (aggstate->hash_ngroups_current > 0)
1986 : {
1987 43682 : aggstate->hashentrysize =
1988 43682 : TupleHashEntrySize() +
1989 43682 : (hashkey_mem / (double) aggstate->hash_ngroups_current);
1990 : }
1991 : }
1992 :
1993 : /*
1994 : * Create memory contexts used for hash aggregation.
1995 : */
1996 : static void
1997 7030 : hash_create_memory(AggState *aggstate)
1998 : {
1999 7030 : Size maxBlockSize = ALLOCSET_DEFAULT_MAXSIZE;
2000 :
2001 : /*
2002 : * The hashcontext's per-tuple memory will be used for byref transition
2003 : * values and returned by AggCheckCallContext().
2004 : */
2005 7030 : aggstate->hashcontext = CreateWorkExprContext(aggstate->ss.ps.state);
2006 :
2007 : /*
2008 : * The meta context will be used for the bucket array of
2009 : * TupleHashEntryData (or arrays, in the case of grouping sets). As the
2010 : * hash table grows, the bucket array will double in size and the old one
2011 : * will be freed, so an AllocSet is appropriate. For large bucket arrays,
2012 : * the large allocation path will be used, so it's not worth worrying
2013 : * about wasting space due to power-of-two allocations.
2014 : */
2015 7030 : aggstate->hash_metacxt = AllocSetContextCreate(aggstate->ss.ps.state->es_query_cxt,
2016 : "HashAgg meta context",
2017 : ALLOCSET_DEFAULT_SIZES);
2018 :
2019 : /*
2020 : * The hash entries themselves, which include the grouping key
2021 : * (firstTuple) and pergroup data, are stored in the table context. The
2022 : * bump allocator can be used because the entries are not freed until the
2023 : * entire hash table is reset. The bump allocator is faster for
2024 : * allocations and avoids wasting space on the chunk header or
2025 : * power-of-two allocations.
2026 : *
2027 : * Like CreateWorkExprContext(), use smaller sizings for smaller work_mem,
2028 : * to avoid large jumps in memory usage.
2029 : */
2030 :
2031 : /*
2032 : * Like CreateWorkExprContext(), use smaller sizings for smaller work_mem,
2033 : * to avoid large jumps in memory usage.
2034 : */
2035 7030 : maxBlockSize = pg_prevpower2_size_t(work_mem * (Size) 1024 / 16);
2036 :
2037 : /* But no bigger than ALLOCSET_DEFAULT_MAXSIZE */
2038 7030 : maxBlockSize = Min(maxBlockSize, ALLOCSET_DEFAULT_MAXSIZE);
2039 :
2040 : /* and no smaller than ALLOCSET_DEFAULT_INITSIZE */
2041 7030 : maxBlockSize = Max(maxBlockSize, ALLOCSET_DEFAULT_INITSIZE);
2042 :
2043 7030 : aggstate->hash_tuplescxt = BumpContextCreate(aggstate->ss.ps.state->es_query_cxt,
2044 : "HashAgg hashed tuples",
2045 : ALLOCSET_DEFAULT_MINSIZE,
2046 : ALLOCSET_DEFAULT_INITSIZE,
2047 : maxBlockSize);
2048 :
2049 7030 : }
2050 :
2051 : /*
2052 : * Choose a reasonable number of buckets for the initial hash table size.
2053 : */
2054 : static double
2055 5468 : hash_choose_num_buckets(double hashentrysize, double ngroups, Size memory)
2056 : {
2057 : double max_nbuckets;
2058 5468 : double nbuckets = ngroups;
2059 :
2060 5468 : max_nbuckets = memory / hashentrysize;
2061 :
2062 : /*
2063 : * Underestimating is better than overestimating. Too many buckets crowd
2064 : * out space for group keys and transition state values.
2065 : */
2066 5468 : max_nbuckets /= 2;
2067 :
2068 5468 : if (nbuckets > max_nbuckets)
2069 72 : nbuckets = max_nbuckets;
2070 :
2071 : /*
2072 : * BuildTupleHashTable will clamp any obviously-insane result, so we don't
2073 : * need to be too careful here.
2074 : */
2075 5468 : return nbuckets;
2076 : }
2077 :
2078 : /*
2079 : * Determine the number of partitions to create when spilling, which will
2080 : * always be a power of two. If log2_npartitions is non-NULL, set
2081 : * *log2_npartitions to the log2() of the number of partitions.
2082 : */
2083 : static int
2084 15052 : hash_choose_num_partitions(double input_groups, double hashentrysize,
2085 : int used_bits, int *log2_npartitions)
2086 : {
2087 15052 : Size hash_mem_limit = get_hash_memory_limit();
2088 : double partition_limit;
2089 : double mem_wanted;
2090 : double dpartitions;
2091 : int npartitions;
2092 : int partition_bits;
2093 :
2094 : /*
2095 : * Avoid creating so many partitions that the memory requirements of the
2096 : * open partition files are greater than 1/4 of hash_mem.
2097 : */
2098 15052 : partition_limit =
2099 15052 : (hash_mem_limit * 0.25 - HASHAGG_READ_BUFFER_SIZE) /
2100 : HASHAGG_WRITE_BUFFER_SIZE;
2101 :
2102 15052 : mem_wanted = HASHAGG_PARTITION_FACTOR * input_groups * hashentrysize;
2103 :
2104 : /* make enough partitions so that each one is likely to fit in memory */
2105 15052 : dpartitions = 1 + (mem_wanted / hash_mem_limit);
2106 :
2107 15052 : if (dpartitions > partition_limit)
2108 14988 : dpartitions = partition_limit;
2109 :
2110 15052 : if (dpartitions < HASHAGG_MIN_PARTITIONS)
2111 15052 : dpartitions = HASHAGG_MIN_PARTITIONS;
2112 15052 : if (dpartitions > HASHAGG_MAX_PARTITIONS)
2113 0 : dpartitions = HASHAGG_MAX_PARTITIONS;
2114 :
2115 : /* HASHAGG_MAX_PARTITIONS limit makes this safe */
2116 15052 : npartitions = (int) dpartitions;
2117 :
2118 : /* ceil(log2(npartitions)) */
2119 15052 : partition_bits = pg_ceil_log2_32(npartitions);
2120 :
2121 : /* make sure that we don't exhaust the hash bits */
2122 15052 : if (partition_bits + used_bits >= 32)
2123 0 : partition_bits = 32 - used_bits;
2124 :
2125 15052 : if (log2_npartitions != NULL)
2126 12622 : *log2_npartitions = partition_bits;
2127 :
2128 : /* number of partitions will be a power of two */
2129 15052 : npartitions = 1 << partition_bits;
2130 :
2131 15052 : return npartitions;
2132 : }
2133 :
2134 : /*
2135 : * Initialize a freshly-created TupleHashEntry.
2136 : */
2137 : static void
2138 527238 : initialize_hash_entry(AggState *aggstate, TupleHashTable hashtable,
2139 : TupleHashEntry entry)
2140 : {
2141 : AggStatePerGroup pergroup;
2142 : int transno;
2143 :
2144 527238 : aggstate->hash_ngroups_current++;
2145 527238 : hash_agg_check_limits(aggstate);
2146 :
2147 : /* no need to allocate or initialize per-group state */
2148 527238 : if (aggstate->numtrans == 0)
2149 211862 : return;
2150 :
2151 315376 : pergroup = (AggStatePerGroup) TupleHashEntryGetAdditional(hashtable, entry);
2152 :
2153 : /*
2154 : * Initialize aggregates for new tuple group, lookup_hash_entries()
2155 : * already has selected the relevant grouping set.
2156 : */
2157 779846 : for (transno = 0; transno < aggstate->numtrans; transno++)
2158 : {
2159 464470 : AggStatePerTrans pertrans = &aggstate->pertrans[transno];
2160 464470 : AggStatePerGroup pergroupstate = &pergroup[transno];
2161 :
2162 464470 : initialize_aggregate(aggstate, pertrans, pergroupstate);
2163 : }
2164 : }
2165 :
2166 : /*
2167 : * Look up hash entries for the current tuple in all hashed grouping sets.
2168 : *
2169 : * Some entries may be left NULL if we are in "spill mode". The same tuple
2170 : * will belong to different groups for each grouping set, so may match a group
2171 : * already in memory for one set and match a group not in memory for another
2172 : * set. When in "spill mode", the tuple will be spilled for each grouping set
2173 : * where it doesn't match a group in memory.
2174 : *
2175 : * NB: It's possible to spill the same tuple for several different grouping
2176 : * sets. This may seem wasteful, but it's actually a trade-off: if we spill
2177 : * the tuple multiple times for multiple grouping sets, it can be partitioned
2178 : * for each grouping set, making the refilling of the hash table very
2179 : * efficient.
2180 : */
2181 : static void
2182 6916030 : lookup_hash_entries(AggState *aggstate)
2183 : {
2184 6916030 : AggStatePerGroup *pergroup = aggstate->hash_pergroup;
2185 6916030 : TupleTableSlot *outerslot = aggstate->tmpcontext->ecxt_outertuple;
2186 : int setno;
2187 :
2188 13966512 : for (setno = 0; setno < aggstate->num_hashes; setno++)
2189 : {
2190 7050482 : AggStatePerHash perhash = &aggstate->perhash[setno];
2191 7050482 : TupleHashTable hashtable = perhash->hashtable;
2192 7050482 : TupleTableSlot *hashslot = perhash->hashslot;
2193 : TupleHashEntry entry;
2194 : uint32 hash;
2195 7050482 : bool isnew = false;
2196 : bool *p_isnew;
2197 :
2198 : /* if hash table already spilled, don't create new entries */
2199 7050482 : p_isnew = aggstate->hash_spill_mode ? NULL : &isnew;
2200 :
2201 7050482 : select_current_set(aggstate, setno, true);
2202 7050482 : prepare_hash_slot(perhash,
2203 : outerslot,
2204 : hashslot);
2205 :
2206 7050482 : entry = LookupTupleHashEntry(hashtable, hashslot,
2207 : p_isnew, &hash);
2208 :
2209 7050482 : if (entry != NULL)
2210 : {
2211 6283246 : if (isnew)
2212 371558 : initialize_hash_entry(aggstate, hashtable, entry);
2213 6283246 : pergroup[setno] = TupleHashEntryGetAdditional(hashtable, entry);
2214 : }
2215 : else
2216 : {
2217 767236 : HashAggSpill *spill = &aggstate->hash_spills[setno];
2218 767236 : TupleTableSlot *slot = aggstate->tmpcontext->ecxt_outertuple;
2219 :
2220 767236 : if (spill->partitions == NULL)
2221 0 : hashagg_spill_init(spill, aggstate->hash_tapeset, 0,
2222 0 : perhash->aggnode->numGroups,
2223 : aggstate->hashentrysize);
2224 :
2225 767236 : hashagg_spill_tuple(aggstate, spill, slot, hash);
2226 767236 : pergroup[setno] = NULL;
2227 : }
2228 : }
2229 6916030 : }
2230 :
2231 : /*
2232 : * ExecAgg -
2233 : *
2234 : * ExecAgg receives tuples from its outer subplan and aggregates over
2235 : * the appropriate attribute for each aggregate function use (Aggref
2236 : * node) appearing in the targetlist or qual of the node. The number
2237 : * of tuples to aggregate over depends on whether grouped or plain
2238 : * aggregation is selected. In grouped aggregation, we produce a result
2239 : * row for each group; in plain aggregation there's a single result row
2240 : * for the whole query. In either case, the value of each aggregate is
2241 : * stored in the expression context to be used when ExecProject evaluates
2242 : * the result tuple.
2243 : */
2244 : static TupleTableSlot *
2245 820512 : ExecAgg(PlanState *pstate)
2246 : {
2247 820512 : AggState *node = castNode(AggState, pstate);
2248 820512 : TupleTableSlot *result = NULL;
2249 :
2250 820512 : CHECK_FOR_INTERRUPTS();
2251 :
2252 820512 : if (!node->agg_done)
2253 : {
2254 : /* Dispatch based on strategy */
2255 757162 : switch (node->phase->aggstrategy)
2256 : {
2257 484146 : case AGG_HASHED:
2258 484146 : if (!node->table_filled)
2259 17166 : agg_fill_hash_table(node);
2260 : /* FALLTHROUGH */
2261 : case AGG_MIXED:
2262 511508 : result = agg_retrieve_hash_table(node);
2263 511508 : break;
2264 245654 : case AGG_PLAIN:
2265 : case AGG_SORTED:
2266 245654 : result = agg_retrieve_direct(node);
2267 245462 : break;
2268 : }
2269 :
2270 756970 : if (!TupIsNull(result))
2271 738036 : return result;
2272 : }
2273 :
2274 82284 : return NULL;
2275 : }
2276 :
2277 : /*
2278 : * ExecAgg for non-hashed case
2279 : */
2280 : static TupleTableSlot *
2281 245654 : agg_retrieve_direct(AggState *aggstate)
2282 : {
2283 245654 : Agg *node = aggstate->phase->aggnode;
2284 : ExprContext *econtext;
2285 : ExprContext *tmpcontext;
2286 : AggStatePerAgg peragg;
2287 : AggStatePerGroup *pergroups;
2288 : TupleTableSlot *outerslot;
2289 : TupleTableSlot *firstSlot;
2290 : TupleTableSlot *result;
2291 245654 : bool hasGroupingSets = aggstate->phase->numsets > 0;
2292 245654 : int numGroupingSets = Max(aggstate->phase->numsets, 1);
2293 : int currentSet;
2294 : int nextSetSize;
2295 : int numReset;
2296 : int i;
2297 :
2298 : /*
2299 : * get state info from node
2300 : *
2301 : * econtext is the per-output-tuple expression context
2302 : *
2303 : * tmpcontext is the per-input-tuple expression context
2304 : */
2305 245654 : econtext = aggstate->ss.ps.ps_ExprContext;
2306 245654 : tmpcontext = aggstate->tmpcontext;
2307 :
2308 245654 : peragg = aggstate->peragg;
2309 245654 : pergroups = aggstate->pergroups;
2310 245654 : firstSlot = aggstate->ss.ss_ScanTupleSlot;
2311 :
2312 : /*
2313 : * We loop retrieving groups until we find one matching
2314 : * aggstate->ss.ps.qual
2315 : *
2316 : * For grouping sets, we have the invariant that aggstate->projected_set
2317 : * is either -1 (initial call) or the index (starting from 0) in
2318 : * gset_lengths for the group we just completed (either by projecting a
2319 : * row or by discarding it in the qual).
2320 : */
2321 316486 : while (!aggstate->agg_done)
2322 : {
2323 : /*
2324 : * Clear the per-output-tuple context for each group, as well as
2325 : * aggcontext (which contains any pass-by-ref transvalues of the old
2326 : * group). Some aggregate functions store working state in child
2327 : * contexts; those now get reset automatically without us needing to
2328 : * do anything special.
2329 : *
2330 : * We use ReScanExprContext not just ResetExprContext because we want
2331 : * any registered shutdown callbacks to be called. That allows
2332 : * aggregate functions to ensure they've cleaned up any non-memory
2333 : * resources.
2334 : */
2335 316276 : ReScanExprContext(econtext);
2336 :
2337 : /*
2338 : * Determine how many grouping sets need to be reset at this boundary.
2339 : */
2340 316276 : if (aggstate->projected_set >= 0 &&
2341 247052 : aggstate->projected_set < numGroupingSets)
2342 247034 : numReset = aggstate->projected_set + 1;
2343 : else
2344 69242 : numReset = numGroupingSets;
2345 :
2346 : /*
2347 : * numReset can change on a phase boundary, but that's OK; we want to
2348 : * reset the contexts used in _this_ phase, and later, after possibly
2349 : * changing phase, initialize the right number of aggregates for the
2350 : * _new_ phase.
2351 : */
2352 :
2353 654830 : for (i = 0; i < numReset; i++)
2354 : {
2355 338554 : ReScanExprContext(aggstate->aggcontexts[i]);
2356 : }
2357 :
2358 : /*
2359 : * Check if input is complete and there are no more groups to project
2360 : * in this phase; move to next phase or mark as done.
2361 : */
2362 316276 : if (aggstate->input_done == true &&
2363 1614 : aggstate->projected_set >= (numGroupingSets - 1))
2364 : {
2365 798 : if (aggstate->current_phase < aggstate->numphases - 1)
2366 : {
2367 204 : initialize_phase(aggstate, aggstate->current_phase + 1);
2368 204 : aggstate->input_done = false;
2369 204 : aggstate->projected_set = -1;
2370 204 : numGroupingSets = Max(aggstate->phase->numsets, 1);
2371 204 : node = aggstate->phase->aggnode;
2372 204 : numReset = numGroupingSets;
2373 : }
2374 594 : else if (aggstate->aggstrategy == AGG_MIXED)
2375 : {
2376 : /*
2377 : * Mixed mode; we've output all the grouped stuff and have
2378 : * full hashtables, so switch to outputting those.
2379 : */
2380 156 : initialize_phase(aggstate, 0);
2381 156 : aggstate->table_filled = true;
2382 156 : ResetTupleHashIterator(aggstate->perhash[0].hashtable,
2383 : &aggstate->perhash[0].hashiter);
2384 156 : select_current_set(aggstate, 0, true);
2385 156 : return agg_retrieve_hash_table(aggstate);
2386 : }
2387 : else
2388 : {
2389 438 : aggstate->agg_done = true;
2390 438 : break;
2391 : }
2392 : }
2393 :
2394 : /*
2395 : * Get the number of columns in the next grouping set after the last
2396 : * projected one (if any). This is the number of columns to compare to
2397 : * see if we reached the boundary of that set too.
2398 : */
2399 315682 : if (aggstate->projected_set >= 0 &&
2400 246254 : aggstate->projected_set < (numGroupingSets - 1))
2401 27294 : nextSetSize = aggstate->phase->gset_lengths[aggstate->projected_set + 1];
2402 : else
2403 288388 : nextSetSize = 0;
2404 :
2405 : /*----------
2406 : * If a subgroup for the current grouping set is present, project it.
2407 : *
2408 : * We have a new group if:
2409 : * - we're out of input but haven't projected all grouping sets
2410 : * (checked above)
2411 : * OR
2412 : * - we already projected a row that wasn't from the last grouping
2413 : * set
2414 : * AND
2415 : * - the next grouping set has at least one grouping column (since
2416 : * empty grouping sets project only once input is exhausted)
2417 : * AND
2418 : * - the previous and pending rows differ on the grouping columns
2419 : * of the next grouping set
2420 : *----------
2421 : */
2422 315682 : tmpcontext->ecxt_innertuple = econtext->ecxt_outertuple;
2423 315682 : if (aggstate->input_done ||
2424 314866 : (node->aggstrategy != AGG_PLAIN &&
2425 247400 : aggstate->projected_set != -1 &&
2426 245438 : aggstate->projected_set < (numGroupingSets - 1) &&
2427 19946 : nextSetSize > 0 &&
2428 19946 : !ExecQualAndReset(aggstate->phase->eqfunctions[nextSetSize - 1],
2429 : tmpcontext)))
2430 : {
2431 14150 : aggstate->projected_set += 1;
2432 :
2433 : Assert(aggstate->projected_set < numGroupingSets);
2434 14150 : Assert(nextSetSize > 0 || aggstate->input_done);
2435 : }
2436 : else
2437 : {
2438 : /*
2439 : * We no longer care what group we just projected, the next
2440 : * projection will always be the first (or only) grouping set
2441 : * (unless the input proves to be empty).
2442 : */
2443 301532 : aggstate->projected_set = 0;
2444 :
2445 : /*
2446 : * If we don't already have the first tuple of the new group,
2447 : * fetch it from the outer plan.
2448 : */
2449 301532 : if (aggstate->grp_firstTuple == NULL)
2450 : {
2451 69428 : outerslot = fetch_input_tuple(aggstate);
2452 69368 : if (!TupIsNull(outerslot))
2453 : {
2454 : /*
2455 : * Make a copy of the first input tuple; we will use this
2456 : * for comparisons (in group mode) and for projection.
2457 : */
2458 56056 : aggstate->grp_firstTuple = ExecCopySlotHeapTuple(outerslot);
2459 : }
2460 : else
2461 : {
2462 : /* outer plan produced no tuples at all */
2463 13312 : if (hasGroupingSets)
2464 : {
2465 : /*
2466 : * If there was no input at all, we need to project
2467 : * rows only if there are grouping sets of size 0.
2468 : * Note that this implies that there can't be any
2469 : * references to ungrouped Vars, which would otherwise
2470 : * cause issues with the empty output slot.
2471 : *
2472 : * XXX: This is no longer true, we currently deal with
2473 : * this in finalize_aggregates().
2474 : */
2475 78 : aggstate->input_done = true;
2476 :
2477 108 : while (aggstate->phase->gset_lengths[aggstate->projected_set] > 0)
2478 : {
2479 48 : aggstate->projected_set += 1;
2480 48 : if (aggstate->projected_set >= numGroupingSets)
2481 : {
2482 : /*
2483 : * We can't set agg_done here because we might
2484 : * have more phases to do, even though the
2485 : * input is empty. So we need to restart the
2486 : * whole outer loop.
2487 : */
2488 18 : break;
2489 : }
2490 : }
2491 :
2492 78 : if (aggstate->projected_set >= numGroupingSets)
2493 18 : continue;
2494 : }
2495 : else
2496 : {
2497 13234 : aggstate->agg_done = true;
2498 : /* If we are grouping, we should produce no tuples too */
2499 13234 : if (node->aggstrategy != AGG_PLAIN)
2500 156 : return NULL;
2501 : }
2502 : }
2503 : }
2504 :
2505 : /*
2506 : * Initialize working state for a new input tuple group.
2507 : */
2508 301298 : initialize_aggregates(aggstate, pergroups, numReset);
2509 :
2510 301298 : if (aggstate->grp_firstTuple != NULL)
2511 : {
2512 : /*
2513 : * Store the copied first input tuple in the tuple table slot
2514 : * reserved for it. The tuple will be deleted when it is
2515 : * cleared from the slot.
2516 : */
2517 288160 : ExecForceStoreHeapTuple(aggstate->grp_firstTuple,
2518 : firstSlot, true);
2519 288160 : aggstate->grp_firstTuple = NULL; /* don't keep two pointers */
2520 :
2521 : /* set up for first advance_aggregates call */
2522 288160 : tmpcontext->ecxt_outertuple = firstSlot;
2523 :
2524 : /*
2525 : * Process each outer-plan tuple, and then fetch the next one,
2526 : * until we exhaust the outer plan or cross a group boundary.
2527 : */
2528 : for (;;)
2529 : {
2530 : /*
2531 : * During phase 1 only of a mixed agg, we need to update
2532 : * hashtables as well in advance_aggregates.
2533 : */
2534 21782436 : if (aggstate->aggstrategy == AGG_MIXED &&
2535 38062 : aggstate->current_phase == 1)
2536 : {
2537 38062 : lookup_hash_entries(aggstate);
2538 : }
2539 :
2540 : /* Advance the aggregates (or combine functions) */
2541 21782436 : advance_aggregates(aggstate);
2542 :
2543 : /* Reset per-input-tuple context after each tuple */
2544 21782358 : ResetExprContext(tmpcontext);
2545 :
2546 21782358 : outerslot = fetch_input_tuple(aggstate);
2547 21782328 : if (TupIsNull(outerslot))
2548 : {
2549 : /* no more outer-plan tuples available */
2550 :
2551 : /* if we built hash tables, finalize any spills */
2552 55924 : if (aggstate->aggstrategy == AGG_MIXED &&
2553 144 : aggstate->current_phase == 1)
2554 144 : hashagg_finish_initial_spills(aggstate);
2555 :
2556 55924 : if (hasGroupingSets)
2557 : {
2558 720 : aggstate->input_done = true;
2559 720 : break;
2560 : }
2561 : else
2562 : {
2563 55204 : aggstate->agg_done = true;
2564 55204 : break;
2565 : }
2566 : }
2567 : /* set up for next advance_aggregates call */
2568 21726404 : tmpcontext->ecxt_outertuple = outerslot;
2569 :
2570 : /*
2571 : * If we are grouping, check whether we've crossed a group
2572 : * boundary.
2573 : */
2574 21726404 : if (node->aggstrategy != AGG_PLAIN && node->numCols > 0)
2575 : {
2576 2471606 : tmpcontext->ecxt_innertuple = firstSlot;
2577 2471606 : if (!ExecQual(aggstate->phase->eqfunctions[node->numCols - 1],
2578 : tmpcontext))
2579 : {
2580 232128 : aggstate->grp_firstTuple = ExecCopySlotHeapTuple(outerslot);
2581 232128 : break;
2582 : }
2583 : }
2584 : }
2585 : }
2586 :
2587 : /*
2588 : * Use the representative input tuple for any references to
2589 : * non-aggregated input columns in aggregate direct args, the node
2590 : * qual, and the tlist. (If we are not grouping, and there are no
2591 : * input rows at all, we will come here with an empty firstSlot
2592 : * ... but if not grouping, there can't be any references to
2593 : * non-aggregated input columns, so no problem.)
2594 : */
2595 301190 : econtext->ecxt_outertuple = firstSlot;
2596 : }
2597 :
2598 : Assert(aggstate->projected_set >= 0);
2599 :
2600 315340 : currentSet = aggstate->projected_set;
2601 :
2602 315340 : prepare_projection_slot(aggstate, econtext->ecxt_outertuple, currentSet);
2603 :
2604 315340 : select_current_set(aggstate, currentSet, false);
2605 :
2606 315340 : finalize_aggregates(aggstate,
2607 : peragg,
2608 315340 : pergroups[currentSet]);
2609 :
2610 : /*
2611 : * If there's no row to project right now, we must continue rather
2612 : * than returning a null since there might be more groups.
2613 : */
2614 315328 : result = project_aggregates(aggstate);
2615 315316 : if (result)
2616 244502 : return result;
2617 : }
2618 :
2619 : /* No more groups */
2620 648 : return NULL;
2621 : }
2622 :
2623 : /*
2624 : * ExecAgg for hashed case: read input and build hash table
2625 : */
2626 : static void
2627 17166 : agg_fill_hash_table(AggState *aggstate)
2628 : {
2629 : TupleTableSlot *outerslot;
2630 17166 : ExprContext *tmpcontext = aggstate->tmpcontext;
2631 :
2632 : /*
2633 : * Process each outer-plan tuple, and then fetch the next one, until we
2634 : * exhaust the outer plan.
2635 : */
2636 : for (;;)
2637 : {
2638 6895134 : outerslot = fetch_input_tuple(aggstate);
2639 6895134 : if (TupIsNull(outerslot))
2640 : break;
2641 :
2642 : /* set up for lookup_hash_entries and advance_aggregates */
2643 6877968 : tmpcontext->ecxt_outertuple = outerslot;
2644 :
2645 : /* Find or build hashtable entries */
2646 6877968 : lookup_hash_entries(aggstate);
2647 :
2648 : /* Advance the aggregates (or combine functions) */
2649 6877968 : advance_aggregates(aggstate);
2650 :
2651 : /*
2652 : * Reset per-input-tuple context after each tuple, but note that the
2653 : * hash lookups do this too
2654 : */
2655 6877968 : ResetExprContext(aggstate->tmpcontext);
2656 : }
2657 :
2658 : /* finalize spills, if any */
2659 17166 : hashagg_finish_initial_spills(aggstate);
2660 :
2661 17166 : aggstate->table_filled = true;
2662 : /* Initialize to walk the first hash table */
2663 17166 : select_current_set(aggstate, 0, true);
2664 17166 : ResetTupleHashIterator(aggstate->perhash[0].hashtable,
2665 : &aggstate->perhash[0].hashiter);
2666 17166 : }
2667 :
2668 : /*
2669 : * If any data was spilled during hash aggregation, reset the hash table and
2670 : * reprocess one batch of spilled data. After reprocessing a batch, the hash
2671 : * table will again contain data, ready to be consumed by
2672 : * agg_retrieve_hash_table_in_memory().
2673 : *
2674 : * Should only be called after all in memory hash table entries have been
2675 : * finalized and emitted.
2676 : *
2677 : * Return false when input is exhausted and there's no more work to be done;
2678 : * otherwise return true.
2679 : */
2680 : static bool
2681 45074 : agg_refill_hash_table(AggState *aggstate)
2682 : {
2683 : HashAggBatch *batch;
2684 : AggStatePerHash perhash;
2685 : HashAggSpill spill;
2686 45074 : LogicalTapeSet *tapeset = aggstate->hash_tapeset;
2687 45074 : bool spill_initialized = false;
2688 :
2689 45074 : if (aggstate->hash_batches == NIL)
2690 18130 : return false;
2691 :
2692 : /* hash_batches is a stack, with the top item at the end of the list */
2693 26944 : batch = llast(aggstate->hash_batches);
2694 26944 : aggstate->hash_batches = list_delete_last(aggstate->hash_batches);
2695 :
2696 26944 : hash_agg_set_limits(aggstate->hashentrysize, batch->input_card,
2697 : batch->used_bits, &aggstate->hash_mem_limit,
2698 : &aggstate->hash_ngroups_limit, NULL);
2699 :
2700 : /*
2701 : * Each batch only processes one grouping set; set the rest to NULL so
2702 : * that advance_aggregates() knows to ignore them. We don't touch
2703 : * pergroups for sorted grouping sets here, because they will be needed if
2704 : * we rescan later. The expressions for sorted grouping sets will not be
2705 : * evaluated after we recompile anyway.
2706 : */
2707 207428 : MemSet(aggstate->hash_pergroup, 0,
2708 : sizeof(AggStatePerGroup) * aggstate->num_hashes);
2709 :
2710 : /* free memory and reset hash tables */
2711 26944 : ReScanExprContext(aggstate->hashcontext);
2712 207428 : for (int setno = 0; setno < aggstate->num_hashes; setno++)
2713 180484 : ResetTupleHashTable(aggstate->perhash[setno].hashtable);
2714 :
2715 26944 : aggstate->hash_ngroups_current = 0;
2716 :
2717 : /*
2718 : * In AGG_MIXED mode, hash aggregation happens in phase 1 and the output
2719 : * happens in phase 0. So, we switch to phase 1 when processing a batch,
2720 : * and back to phase 0 after the batch is done.
2721 : */
2722 : Assert(aggstate->current_phase == 0);
2723 26944 : if (aggstate->phase->aggstrategy == AGG_MIXED)
2724 : {
2725 26262 : aggstate->current_phase = 1;
2726 26262 : aggstate->phase = &aggstate->phases[aggstate->current_phase];
2727 : }
2728 :
2729 26944 : select_current_set(aggstate, batch->setno, true);
2730 :
2731 26944 : perhash = &aggstate->perhash[aggstate->current_set];
2732 :
2733 : /*
2734 : * Spilled tuples are always read back as MinimalTuples, which may be
2735 : * different from the outer plan, so recompile the aggregate expressions.
2736 : *
2737 : * We still need the NULL check, because we are only processing one
2738 : * grouping set at a time and the rest will be NULL.
2739 : */
2740 26944 : hashagg_recompile_expressions(aggstate, true, true);
2741 :
2742 26944 : INJECTION_POINT("hash-aggregate-process-batch", NULL);
2743 : for (;;)
2744 1216776 : {
2745 1243720 : TupleTableSlot *spillslot = aggstate->hash_spill_rslot;
2746 1243720 : TupleTableSlot *hashslot = perhash->hashslot;
2747 1243720 : TupleHashTable hashtable = perhash->hashtable;
2748 : TupleHashEntry entry;
2749 : MinimalTuple tuple;
2750 : uint32 hash;
2751 1243720 : bool isnew = false;
2752 1243720 : bool *p_isnew = aggstate->hash_spill_mode ? NULL : &isnew;
2753 :
2754 1243720 : CHECK_FOR_INTERRUPTS();
2755 :
2756 1243720 : tuple = hashagg_batch_read(batch, &hash);
2757 1243720 : if (tuple == NULL)
2758 26944 : break;
2759 :
2760 1216776 : ExecStoreMinimalTuple(tuple, spillslot, true);
2761 1216776 : aggstate->tmpcontext->ecxt_outertuple = spillslot;
2762 :
2763 1216776 : prepare_hash_slot(perhash,
2764 1216776 : aggstate->tmpcontext->ecxt_outertuple,
2765 : hashslot);
2766 1216776 : entry = LookupTupleHashEntryHash(hashtable, hashslot,
2767 : p_isnew, hash);
2768 :
2769 1216776 : if (entry != NULL)
2770 : {
2771 767236 : if (isnew)
2772 155680 : initialize_hash_entry(aggstate, hashtable, entry);
2773 767236 : aggstate->hash_pergroup[batch->setno] = TupleHashEntryGetAdditional(hashtable, entry);
2774 767236 : advance_aggregates(aggstate);
2775 : }
2776 : else
2777 : {
2778 449540 : if (!spill_initialized)
2779 : {
2780 : /*
2781 : * Avoid initializing the spill until we actually need it so
2782 : * that we don't assign tapes that will never be used.
2783 : */
2784 12500 : spill_initialized = true;
2785 12500 : hashagg_spill_init(&spill, tapeset, batch->used_bits,
2786 : batch->input_card, aggstate->hashentrysize);
2787 : }
2788 : /* no memory for a new group, spill */
2789 449540 : hashagg_spill_tuple(aggstate, &spill, spillslot, hash);
2790 :
2791 449540 : aggstate->hash_pergroup[batch->setno] = NULL;
2792 : }
2793 :
2794 : /*
2795 : * Reset per-input-tuple context after each tuple, but note that the
2796 : * hash lookups do this too
2797 : */
2798 1216776 : ResetExprContext(aggstate->tmpcontext);
2799 : }
2800 :
2801 26944 : LogicalTapeClose(batch->input_tape);
2802 :
2803 : /* change back to phase 0 */
2804 26944 : aggstate->current_phase = 0;
2805 26944 : aggstate->phase = &aggstate->phases[aggstate->current_phase];
2806 :
2807 26944 : if (spill_initialized)
2808 : {
2809 12500 : hashagg_spill_finish(aggstate, &spill, batch->setno);
2810 12500 : hash_agg_update_metrics(aggstate, true, spill.npartitions);
2811 : }
2812 : else
2813 14444 : hash_agg_update_metrics(aggstate, true, 0);
2814 :
2815 26944 : aggstate->hash_spill_mode = false;
2816 :
2817 : /* prepare to walk the first hash table */
2818 26944 : select_current_set(aggstate, batch->setno, true);
2819 26944 : ResetTupleHashIterator(aggstate->perhash[batch->setno].hashtable,
2820 : &aggstate->perhash[batch->setno].hashiter);
2821 :
2822 26944 : pfree(batch);
2823 :
2824 26944 : return true;
2825 : }
2826 :
2827 : /*
2828 : * ExecAgg for hashed case: retrieving groups from hash table
2829 : *
2830 : * After exhausting in-memory tuples, also try refilling the hash table using
2831 : * previously-spilled tuples. Only returns NULL after all in-memory and
2832 : * spilled tuples are exhausted.
2833 : */
2834 : static TupleTableSlot *
2835 511664 : agg_retrieve_hash_table(AggState *aggstate)
2836 : {
2837 511664 : TupleTableSlot *result = NULL;
2838 :
2839 1032142 : while (result == NULL)
2840 : {
2841 538608 : result = agg_retrieve_hash_table_in_memory(aggstate);
2842 538608 : if (result == NULL)
2843 : {
2844 45074 : if (!agg_refill_hash_table(aggstate))
2845 : {
2846 18130 : aggstate->agg_done = true;
2847 18130 : break;
2848 : }
2849 : }
2850 : }
2851 :
2852 511664 : return result;
2853 : }
2854 :
2855 : /*
2856 : * Retrieve the groups from the in-memory hash tables without considering any
2857 : * spilled tuples.
2858 : */
2859 : static TupleTableSlot *
2860 538608 : agg_retrieve_hash_table_in_memory(AggState *aggstate)
2861 : {
2862 : ExprContext *econtext;
2863 : AggStatePerAgg peragg;
2864 : AggStatePerGroup pergroup;
2865 : TupleHashEntry entry;
2866 : TupleTableSlot *firstSlot;
2867 : TupleTableSlot *result;
2868 : AggStatePerHash perhash;
2869 :
2870 : /*
2871 : * get state info from node.
2872 : *
2873 : * econtext is the per-output-tuple expression context.
2874 : */
2875 538608 : econtext = aggstate->ss.ps.ps_ExprContext;
2876 538608 : peragg = aggstate->peragg;
2877 538608 : firstSlot = aggstate->ss.ss_ScanTupleSlot;
2878 :
2879 : /*
2880 : * Note that perhash (and therefore anything accessed through it) can
2881 : * change inside the loop, as we change between grouping sets.
2882 : */
2883 538608 : perhash = &aggstate->perhash[aggstate->current_set];
2884 :
2885 : /*
2886 : * We loop retrieving groups until we find one satisfying
2887 : * aggstate->ss.ps.qual
2888 : */
2889 : for (;;)
2890 135954 : {
2891 674562 : TupleTableSlot *hashslot = perhash->hashslot;
2892 674562 : TupleHashTable hashtable = perhash->hashtable;
2893 : int i;
2894 :
2895 674562 : CHECK_FOR_INTERRUPTS();
2896 :
2897 : /*
2898 : * Find the next entry in the hash table
2899 : */
2900 674562 : entry = ScanTupleHashTable(hashtable, &perhash->hashiter);
2901 674562 : if (entry == NULL)
2902 : {
2903 145366 : int nextset = aggstate->current_set + 1;
2904 :
2905 145366 : if (nextset < aggstate->num_hashes)
2906 : {
2907 : /*
2908 : * Switch to next grouping set, reinitialize, and restart the
2909 : * loop.
2910 : */
2911 100292 : select_current_set(aggstate, nextset, true);
2912 :
2913 100292 : perhash = &aggstate->perhash[aggstate->current_set];
2914 :
2915 100292 : ResetTupleHashIterator(perhash->hashtable, &perhash->hashiter);
2916 :
2917 100292 : continue;
2918 : }
2919 : else
2920 : {
2921 45074 : return NULL;
2922 : }
2923 : }
2924 :
2925 : /*
2926 : * Clear the per-output-tuple context for each group
2927 : *
2928 : * We intentionally don't use ReScanExprContext here; if any aggs have
2929 : * registered shutdown callbacks, they mustn't be called yet, since we
2930 : * might not be done with that agg.
2931 : */
2932 529196 : ResetExprContext(econtext);
2933 :
2934 : /*
2935 : * Transform representative tuple back into one with the right
2936 : * columns.
2937 : */
2938 529196 : ExecStoreMinimalTuple(TupleHashEntryGetTuple(entry), hashslot, false);
2939 529196 : slot_getallattrs(hashslot);
2940 :
2941 529196 : ExecClearTuple(firstSlot);
2942 529196 : memset(firstSlot->tts_isnull, true,
2943 529196 : firstSlot->tts_tupleDescriptor->natts * sizeof(bool));
2944 :
2945 1392242 : for (i = 0; i < perhash->numhashGrpCols; i++)
2946 : {
2947 863046 : int varNumber = perhash->hashGrpColIdxInput[i] - 1;
2948 :
2949 863046 : firstSlot->tts_values[varNumber] = hashslot->tts_values[i];
2950 863046 : firstSlot->tts_isnull[varNumber] = hashslot->tts_isnull[i];
2951 : }
2952 529196 : ExecStoreVirtualTuple(firstSlot);
2953 :
2954 529196 : pergroup = (AggStatePerGroup) TupleHashEntryGetAdditional(hashtable, entry);
2955 :
2956 : /*
2957 : * Use the representative input tuple for any references to
2958 : * non-aggregated input columns in the qual and tlist.
2959 : */
2960 529196 : econtext->ecxt_outertuple = firstSlot;
2961 :
2962 529196 : prepare_projection_slot(aggstate,
2963 : econtext->ecxt_outertuple,
2964 : aggstate->current_set);
2965 :
2966 529196 : finalize_aggregates(aggstate, peragg, pergroup);
2967 :
2968 529196 : result = project_aggregates(aggstate);
2969 529196 : if (result)
2970 493534 : return result;
2971 : }
2972 :
2973 : /* No more groups */
2974 : return NULL;
2975 : }
2976 :
2977 : /*
2978 : * hashagg_spill_init
2979 : *
2980 : * Called after we determined that spilling is necessary. Chooses the number
2981 : * of partitions to create, and initializes them.
2982 : */
2983 : static void
2984 12622 : hashagg_spill_init(HashAggSpill *spill, LogicalTapeSet *tapeset, int used_bits,
2985 : double input_groups, double hashentrysize)
2986 : {
2987 : int npartitions;
2988 : int partition_bits;
2989 :
2990 12622 : npartitions = hash_choose_num_partitions(input_groups, hashentrysize,
2991 : used_bits, &partition_bits);
2992 :
2993 : #ifdef USE_INJECTION_POINTS
2994 12622 : if (IS_INJECTION_POINT_ATTACHED("hash-aggregate-single-partition"))
2995 : {
2996 10 : npartitions = 1;
2997 10 : partition_bits = 0;
2998 10 : INJECTION_POINT_CACHED("hash-aggregate-single-partition", NULL);
2999 : }
3000 : #endif
3001 :
3002 12622 : spill->partitions = palloc0(sizeof(LogicalTape *) * npartitions);
3003 12622 : spill->ntuples = palloc0(sizeof(int64) * npartitions);
3004 12622 : spill->hll_card = palloc0(sizeof(hyperLogLogState) * npartitions);
3005 :
3006 63080 : for (int i = 0; i < npartitions; i++)
3007 50458 : spill->partitions[i] = LogicalTapeCreate(tapeset);
3008 :
3009 12622 : spill->shift = 32 - used_bits - partition_bits;
3010 12622 : if (spill->shift < 32)
3011 12612 : spill->mask = (npartitions - 1) << spill->shift;
3012 : else
3013 10 : spill->mask = 0;
3014 12622 : spill->npartitions = npartitions;
3015 :
3016 63080 : for (int i = 0; i < npartitions; i++)
3017 50458 : initHyperLogLog(&spill->hll_card[i], HASHAGG_HLL_BIT_WIDTH);
3018 12622 : }
3019 :
3020 : /*
3021 : * hashagg_spill_tuple
3022 : *
3023 : * No room for new groups in the hash table. Save for later in the appropriate
3024 : * partition.
3025 : */
3026 : static Size
3027 1216776 : hashagg_spill_tuple(AggState *aggstate, HashAggSpill *spill,
3028 : TupleTableSlot *inputslot, uint32 hash)
3029 : {
3030 : TupleTableSlot *spillslot;
3031 : int partition;
3032 : MinimalTuple tuple;
3033 : LogicalTape *tape;
3034 1216776 : int total_written = 0;
3035 : bool shouldFree;
3036 :
3037 : Assert(spill->partitions != NULL);
3038 :
3039 : /* spill only attributes that we actually need */
3040 1216776 : if (!aggstate->all_cols_needed)
3041 : {
3042 1572 : spillslot = aggstate->hash_spill_wslot;
3043 1572 : slot_getsomeattrs(inputslot, aggstate->max_colno_needed);
3044 1572 : ExecClearTuple(spillslot);
3045 4716 : for (int i = 0; i < spillslot->tts_tupleDescriptor->natts; i++)
3046 : {
3047 3144 : if (bms_is_member(i + 1, aggstate->colnos_needed))
3048 : {
3049 1572 : spillslot->tts_values[i] = inputslot->tts_values[i];
3050 1572 : spillslot->tts_isnull[i] = inputslot->tts_isnull[i];
3051 : }
3052 : else
3053 1572 : spillslot->tts_isnull[i] = true;
3054 : }
3055 1572 : ExecStoreVirtualTuple(spillslot);
3056 : }
3057 : else
3058 1215204 : spillslot = inputslot;
3059 :
3060 1216776 : tuple = ExecFetchSlotMinimalTuple(spillslot, &shouldFree);
3061 :
3062 1216776 : if (spill->shift < 32)
3063 1195776 : partition = (hash & spill->mask) >> spill->shift;
3064 : else
3065 21000 : partition = 0;
3066 :
3067 1216776 : spill->ntuples[partition]++;
3068 :
3069 : /*
3070 : * All hash values destined for a given partition have some bits in
3071 : * common, which causes bad HLL cardinality estimates. Hash the hash to
3072 : * get a more uniform distribution.
3073 : */
3074 1216776 : addHyperLogLog(&spill->hll_card[partition], hash_bytes_uint32(hash));
3075 :
3076 1216776 : tape = spill->partitions[partition];
3077 :
3078 1216776 : LogicalTapeWrite(tape, &hash, sizeof(uint32));
3079 1216776 : total_written += sizeof(uint32);
3080 :
3081 1216776 : LogicalTapeWrite(tape, tuple, tuple->t_len);
3082 1216776 : total_written += tuple->t_len;
3083 :
3084 1216776 : if (shouldFree)
3085 767236 : pfree(tuple);
3086 :
3087 1216776 : return total_written;
3088 : }
3089 :
3090 : /*
3091 : * hashagg_batch_new
3092 : *
3093 : * Construct a HashAggBatch item, which represents one iteration of HashAgg to
3094 : * be done.
3095 : */
3096 : static HashAggBatch *
3097 26944 : hashagg_batch_new(LogicalTape *input_tape, int setno,
3098 : int64 input_tuples, double input_card, int used_bits)
3099 : {
3100 26944 : HashAggBatch *batch = palloc0(sizeof(HashAggBatch));
3101 :
3102 26944 : batch->setno = setno;
3103 26944 : batch->used_bits = used_bits;
3104 26944 : batch->input_tape = input_tape;
3105 26944 : batch->input_tuples = input_tuples;
3106 26944 : batch->input_card = input_card;
3107 :
3108 26944 : return batch;
3109 : }
3110 :
3111 : /*
3112 : * hashagg_batch_read
3113 : * read the next tuple from a batch's tape. Return NULL if no more.
3114 : */
3115 : static MinimalTuple
3116 1243720 : hashagg_batch_read(HashAggBatch *batch, uint32 *hashp)
3117 : {
3118 1243720 : LogicalTape *tape = batch->input_tape;
3119 : MinimalTuple tuple;
3120 : uint32 t_len;
3121 : size_t nread;
3122 : uint32 hash;
3123 :
3124 1243720 : nread = LogicalTapeRead(tape, &hash, sizeof(uint32));
3125 1243720 : if (nread == 0)
3126 26944 : return NULL;
3127 1216776 : if (nread != sizeof(uint32))
3128 0 : ereport(ERROR,
3129 : (errcode_for_file_access(),
3130 : errmsg_internal("unexpected EOF for tape %p: requested %zu bytes, read %zu bytes",
3131 : tape, sizeof(uint32), nread)));
3132 1216776 : if (hashp != NULL)
3133 1216776 : *hashp = hash;
3134 :
3135 1216776 : nread = LogicalTapeRead(tape, &t_len, sizeof(t_len));
3136 1216776 : if (nread != sizeof(uint32))
3137 0 : ereport(ERROR,
3138 : (errcode_for_file_access(),
3139 : errmsg_internal("unexpected EOF for tape %p: requested %zu bytes, read %zu bytes",
3140 : tape, sizeof(uint32), nread)));
3141 :
3142 1216776 : tuple = (MinimalTuple) palloc(t_len);
3143 1216776 : tuple->t_len = t_len;
3144 :
3145 1216776 : nread = LogicalTapeRead(tape,
3146 : (char *) tuple + sizeof(uint32),
3147 : t_len - sizeof(uint32));
3148 1216776 : if (nread != t_len - sizeof(uint32))
3149 0 : ereport(ERROR,
3150 : (errcode_for_file_access(),
3151 : errmsg_internal("unexpected EOF for tape %p: requested %zu bytes, read %zu bytes",
3152 : tape, t_len - sizeof(uint32), nread)));
3153 :
3154 1216776 : return tuple;
3155 : }
3156 :
3157 : /*
3158 : * hashagg_finish_initial_spills
3159 : *
3160 : * After a HashAggBatch has been processed, it may have spilled tuples to
3161 : * disk. If so, turn the spilled partitions into new batches that must later
3162 : * be executed.
3163 : */
3164 : static void
3165 17310 : hashagg_finish_initial_spills(AggState *aggstate)
3166 : {
3167 : int setno;
3168 17310 : int total_npartitions = 0;
3169 :
3170 17310 : if (aggstate->hash_spills != NULL)
3171 : {
3172 184 : for (setno = 0; setno < aggstate->num_hashes; setno++)
3173 : {
3174 122 : HashAggSpill *spill = &aggstate->hash_spills[setno];
3175 :
3176 122 : total_npartitions += spill->npartitions;
3177 122 : hashagg_spill_finish(aggstate, spill, setno);
3178 : }
3179 :
3180 : /*
3181 : * We're not processing tuples from outer plan any more; only
3182 : * processing batches of spilled tuples. The initial spill structures
3183 : * are no longer needed.
3184 : */
3185 62 : pfree(aggstate->hash_spills);
3186 62 : aggstate->hash_spills = NULL;
3187 : }
3188 :
3189 17310 : hash_agg_update_metrics(aggstate, false, total_npartitions);
3190 17310 : aggstate->hash_spill_mode = false;
3191 17310 : }
3192 :
3193 : /*
3194 : * hashagg_spill_finish
3195 : *
3196 : * Transform spill partitions into new batches.
3197 : */
3198 : static void
3199 12622 : hashagg_spill_finish(AggState *aggstate, HashAggSpill *spill, int setno)
3200 : {
3201 : int i;
3202 12622 : int used_bits = 32 - spill->shift;
3203 :
3204 12622 : if (spill->npartitions == 0)
3205 0 : return; /* didn't spill */
3206 :
3207 63080 : for (i = 0; i < spill->npartitions; i++)
3208 : {
3209 50458 : LogicalTape *tape = spill->partitions[i];
3210 : HashAggBatch *new_batch;
3211 : double cardinality;
3212 :
3213 : /* if the partition is empty, don't create a new batch of work */
3214 50458 : if (spill->ntuples[i] == 0)
3215 23514 : continue;
3216 :
3217 26944 : cardinality = estimateHyperLogLog(&spill->hll_card[i]);
3218 26944 : freeHyperLogLog(&spill->hll_card[i]);
3219 :
3220 : /* rewinding frees the buffer while not in use */
3221 26944 : LogicalTapeRewindForRead(tape, HASHAGG_READ_BUFFER_SIZE);
3222 :
3223 26944 : new_batch = hashagg_batch_new(tape, setno,
3224 26944 : spill->ntuples[i], cardinality,
3225 : used_bits);
3226 26944 : aggstate->hash_batches = lappend(aggstate->hash_batches, new_batch);
3227 26944 : aggstate->hash_batches_used++;
3228 : }
3229 :
3230 12622 : pfree(spill->ntuples);
3231 12622 : pfree(spill->hll_card);
3232 12622 : pfree(spill->partitions);
3233 : }
3234 :
3235 : /*
3236 : * Free resources related to a spilled HashAgg.
3237 : */
3238 : static void
3239 60328 : hashagg_reset_spill_state(AggState *aggstate)
3240 : {
3241 : /* free spills from initial pass */
3242 60328 : if (aggstate->hash_spills != NULL)
3243 : {
3244 : int setno;
3245 :
3246 0 : for (setno = 0; setno < aggstate->num_hashes; setno++)
3247 : {
3248 0 : HashAggSpill *spill = &aggstate->hash_spills[setno];
3249 :
3250 0 : pfree(spill->ntuples);
3251 0 : pfree(spill->partitions);
3252 : }
3253 0 : pfree(aggstate->hash_spills);
3254 0 : aggstate->hash_spills = NULL;
3255 : }
3256 :
3257 : /* free batches */
3258 60328 : list_free_deep(aggstate->hash_batches);
3259 60328 : aggstate->hash_batches = NIL;
3260 :
3261 : /* close tape set */
3262 60328 : if (aggstate->hash_tapeset != NULL)
3263 : {
3264 62 : LogicalTapeSetClose(aggstate->hash_tapeset);
3265 62 : aggstate->hash_tapeset = NULL;
3266 : }
3267 60328 : }
3268 :
3269 :
3270 : /* -----------------
3271 : * ExecInitAgg
3272 : *
3273 : * Creates the run-time information for the agg node produced by the
3274 : * planner and initializes its outer subtree.
3275 : *
3276 : * -----------------
3277 : */
3278 : AggState *
3279 48136 : ExecInitAgg(Agg *node, EState *estate, int eflags)
3280 : {
3281 : AggState *aggstate;
3282 : AggStatePerAgg peraggs;
3283 : AggStatePerTrans pertransstates;
3284 : AggStatePerGroup *pergroups;
3285 : Plan *outerPlan;
3286 : ExprContext *econtext;
3287 : TupleDesc scanDesc;
3288 : int max_aggno;
3289 : int max_transno;
3290 : int numaggrefs;
3291 : int numaggs;
3292 : int numtrans;
3293 : int phase;
3294 : int phaseidx;
3295 : ListCell *l;
3296 48136 : Bitmapset *all_grouped_cols = NULL;
3297 48136 : int numGroupingSets = 1;
3298 : int numPhases;
3299 : int numHashes;
3300 48136 : int i = 0;
3301 48136 : int j = 0;
3302 89474 : bool use_hashing = (node->aggstrategy == AGG_HASHED ||
3303 41338 : node->aggstrategy == AGG_MIXED);
3304 :
3305 : /* check for unsupported flags */
3306 : Assert(!(eflags & (EXEC_FLAG_BACKWARD | EXEC_FLAG_MARK)));
3307 :
3308 : /*
3309 : * create state structure
3310 : */
3311 48136 : aggstate = makeNode(AggState);
3312 48136 : aggstate->ss.ps.plan = (Plan *) node;
3313 48136 : aggstate->ss.ps.state = estate;
3314 48136 : aggstate->ss.ps.ExecProcNode = ExecAgg;
3315 :
3316 48136 : aggstate->aggs = NIL;
3317 48136 : aggstate->numaggs = 0;
3318 48136 : aggstate->numtrans = 0;
3319 48136 : aggstate->aggstrategy = node->aggstrategy;
3320 48136 : aggstate->aggsplit = node->aggsplit;
3321 48136 : aggstate->maxsets = 0;
3322 48136 : aggstate->projected_set = -1;
3323 48136 : aggstate->current_set = 0;
3324 48136 : aggstate->peragg = NULL;
3325 48136 : aggstate->pertrans = NULL;
3326 48136 : aggstate->curperagg = NULL;
3327 48136 : aggstate->curpertrans = NULL;
3328 48136 : aggstate->input_done = false;
3329 48136 : aggstate->agg_done = false;
3330 48136 : aggstate->pergroups = NULL;
3331 48136 : aggstate->grp_firstTuple = NULL;
3332 48136 : aggstate->sort_in = NULL;
3333 48136 : aggstate->sort_out = NULL;
3334 :
3335 : /*
3336 : * phases[0] always exists, but is dummy in sorted/plain mode
3337 : */
3338 48136 : numPhases = (use_hashing ? 1 : 2);
3339 48136 : numHashes = (use_hashing ? 1 : 0);
3340 :
3341 : /*
3342 : * Calculate the maximum number of grouping sets in any phase; this
3343 : * determines the size of some allocations. Also calculate the number of
3344 : * phases, since all hashed/mixed nodes contribute to only a single phase.
3345 : */
3346 48136 : if (node->groupingSets)
3347 : {
3348 914 : numGroupingSets = list_length(node->groupingSets);
3349 :
3350 1912 : foreach(l, node->chain)
3351 : {
3352 998 : Agg *agg = lfirst(l);
3353 :
3354 998 : numGroupingSets = Max(numGroupingSets,
3355 : list_length(agg->groupingSets));
3356 :
3357 : /*
3358 : * additional AGG_HASHED aggs become part of phase 0, but all
3359 : * others add an extra phase.
3360 : */
3361 998 : if (agg->aggstrategy != AGG_HASHED)
3362 484 : ++numPhases;
3363 : else
3364 514 : ++numHashes;
3365 : }
3366 : }
3367 :
3368 48136 : aggstate->maxsets = numGroupingSets;
3369 48136 : aggstate->numphases = numPhases;
3370 :
3371 48136 : aggstate->aggcontexts = (ExprContext **)
3372 48136 : palloc0(sizeof(ExprContext *) * numGroupingSets);
3373 :
3374 : /*
3375 : * Create expression contexts. We need three or more, one for
3376 : * per-input-tuple processing, one for per-output-tuple processing, one
3377 : * for all the hashtables, and one for each grouping set. The per-tuple
3378 : * memory context of the per-grouping-set ExprContexts (aggcontexts)
3379 : * replaces the standalone memory context formerly used to hold transition
3380 : * values. We cheat a little by using ExecAssignExprContext() to build
3381 : * all of them.
3382 : *
3383 : * NOTE: the details of what is stored in aggcontexts and what is stored
3384 : * in the regular per-query memory context are driven by a simple
3385 : * decision: we want to reset the aggcontext at group boundaries (if not
3386 : * hashing) and in ExecReScanAgg to recover no-longer-wanted space.
3387 : */
3388 48136 : ExecAssignExprContext(estate, &aggstate->ss.ps);
3389 48136 : aggstate->tmpcontext = aggstate->ss.ps.ps_ExprContext;
3390 :
3391 97136 : for (i = 0; i < numGroupingSets; ++i)
3392 : {
3393 49000 : ExecAssignExprContext(estate, &aggstate->ss.ps);
3394 49000 : aggstate->aggcontexts[i] = aggstate->ss.ps.ps_ExprContext;
3395 : }
3396 :
3397 48136 : if (use_hashing)
3398 7030 : hash_create_memory(aggstate);
3399 :
3400 48136 : ExecAssignExprContext(estate, &aggstate->ss.ps);
3401 :
3402 : /*
3403 : * Initialize child nodes.
3404 : *
3405 : * If we are doing a hashed aggregation then the child plan does not need
3406 : * to handle REWIND efficiently; see ExecReScanAgg.
3407 : */
3408 48136 : if (node->aggstrategy == AGG_HASHED)
3409 6798 : eflags &= ~EXEC_FLAG_REWIND;
3410 48136 : outerPlan = outerPlan(node);
3411 48136 : outerPlanState(aggstate) = ExecInitNode(outerPlan, estate, eflags);
3412 :
3413 : /*
3414 : * initialize source tuple type.
3415 : */
3416 48136 : aggstate->ss.ps.outerops =
3417 48136 : ExecGetResultSlotOps(outerPlanState(&aggstate->ss),
3418 : &aggstate->ss.ps.outeropsfixed);
3419 48136 : aggstate->ss.ps.outeropsset = true;
3420 :
3421 48136 : ExecCreateScanSlotFromOuterPlan(estate, &aggstate->ss,
3422 : aggstate->ss.ps.outerops);
3423 48136 : scanDesc = aggstate->ss.ss_ScanTupleSlot->tts_tupleDescriptor;
3424 :
3425 : /*
3426 : * If there are more than two phases (including a potential dummy phase
3427 : * 0), input will be resorted using tuplesort. Need a slot for that.
3428 : */
3429 48136 : if (numPhases > 2)
3430 : {
3431 210 : aggstate->sort_slot = ExecInitExtraTupleSlot(estate, scanDesc,
3432 : &TTSOpsMinimalTuple);
3433 :
3434 : /*
3435 : * The output of the tuplesort, and the output from the outer child
3436 : * might not use the same type of slot. In most cases the child will
3437 : * be a Sort, and thus return a TTSOpsMinimalTuple type slot - but the
3438 : * input can also be presorted due an index, in which case it could be
3439 : * a different type of slot.
3440 : *
3441 : * XXX: For efficiency it would be good to instead/additionally
3442 : * generate expressions with corresponding settings of outerops* for
3443 : * the individual phases - deforming is often a bottleneck for
3444 : * aggregations with lots of rows per group. If there's multiple
3445 : * sorts, we know that all but the first use TTSOpsMinimalTuple (via
3446 : * the nodeAgg.c internal tuplesort).
3447 : */
3448 210 : if (aggstate->ss.ps.outeropsfixed &&
3449 210 : aggstate->ss.ps.outerops != &TTSOpsMinimalTuple)
3450 12 : aggstate->ss.ps.outeropsfixed = false;
3451 : }
3452 :
3453 : /*
3454 : * Initialize result type, slot and projection.
3455 : */
3456 48136 : ExecInitResultTupleSlotTL(&aggstate->ss.ps, &TTSOpsVirtual);
3457 48136 : ExecAssignProjectionInfo(&aggstate->ss.ps, NULL);
3458 :
3459 : /*
3460 : * initialize child expressions
3461 : *
3462 : * We expect the parser to have checked that no aggs contain other agg
3463 : * calls in their arguments (and just to be sure, we verify it again while
3464 : * initializing the plan node). This would make no sense under SQL
3465 : * semantics, and it's forbidden by the spec. Because it is true, we
3466 : * don't need to worry about evaluating the aggs in any particular order.
3467 : *
3468 : * Note: execExpr.c finds Aggrefs for us, and adds them to aggstate->aggs.
3469 : * Aggrefs in the qual are found here; Aggrefs in the targetlist are found
3470 : * during ExecAssignProjectionInfo, above.
3471 : */
3472 48136 : aggstate->ss.ps.qual =
3473 48136 : ExecInitQual(node->plan.qual, (PlanState *) aggstate);
3474 :
3475 : /*
3476 : * We should now have found all Aggrefs in the targetlist and quals.
3477 : */
3478 48136 : numaggrefs = list_length(aggstate->aggs);
3479 48136 : max_aggno = -1;
3480 48136 : max_transno = -1;
3481 102250 : foreach(l, aggstate->aggs)
3482 : {
3483 54114 : Aggref *aggref = (Aggref *) lfirst(l);
3484 :
3485 54114 : max_aggno = Max(max_aggno, aggref->aggno);
3486 54114 : max_transno = Max(max_transno, aggref->aggtransno);
3487 : }
3488 48136 : aggstate->numaggs = numaggs = max_aggno + 1;
3489 48136 : aggstate->numtrans = numtrans = max_transno + 1;
3490 :
3491 : /*
3492 : * For each phase, prepare grouping set data and fmgr lookup data for
3493 : * compare functions. Accumulate all_grouped_cols in passing.
3494 : */
3495 48136 : aggstate->phases = palloc0(numPhases * sizeof(AggStatePerPhaseData));
3496 :
3497 48136 : aggstate->num_hashes = numHashes;
3498 48136 : if (numHashes)
3499 : {
3500 7030 : aggstate->perhash = palloc0(sizeof(AggStatePerHashData) * numHashes);
3501 7030 : aggstate->phases[0].numsets = 0;
3502 7030 : aggstate->phases[0].gset_lengths = palloc(numHashes * sizeof(int));
3503 7030 : aggstate->phases[0].grouped_cols = palloc(numHashes * sizeof(Bitmapset *));
3504 : }
3505 :
3506 48136 : phase = 0;
3507 97270 : for (phaseidx = 0; phaseidx <= list_length(node->chain); ++phaseidx)
3508 : {
3509 : Agg *aggnode;
3510 : Sort *sortnode;
3511 :
3512 49134 : if (phaseidx > 0)
3513 : {
3514 998 : aggnode = list_nth_node(Agg, node->chain, phaseidx - 1);
3515 998 : sortnode = castNode(Sort, outerPlan(aggnode));
3516 : }
3517 : else
3518 : {
3519 48136 : aggnode = node;
3520 48136 : sortnode = NULL;
3521 : }
3522 :
3523 : Assert(phase <= 1 || sortnode);
3524 :
3525 49134 : if (aggnode->aggstrategy == AGG_HASHED
3526 41822 : || aggnode->aggstrategy == AGG_MIXED)
3527 7544 : {
3528 7544 : AggStatePerPhase phasedata = &aggstate->phases[0];
3529 : AggStatePerHash perhash;
3530 7544 : Bitmapset *cols = NULL;
3531 :
3532 : Assert(phase == 0);
3533 7544 : i = phasedata->numsets++;
3534 7544 : perhash = &aggstate->perhash[i];
3535 :
3536 : /* phase 0 always points to the "real" Agg in the hash case */
3537 7544 : phasedata->aggnode = node;
3538 7544 : phasedata->aggstrategy = node->aggstrategy;
3539 :
3540 : /* but the actual Agg node representing this hash is saved here */
3541 7544 : perhash->aggnode = aggnode;
3542 :
3543 7544 : phasedata->gset_lengths[i] = perhash->numCols = aggnode->numCols;
3544 :
3545 19044 : for (j = 0; j < aggnode->numCols; ++j)
3546 11500 : cols = bms_add_member(cols, aggnode->grpColIdx[j]);
3547 :
3548 7544 : phasedata->grouped_cols[i] = cols;
3549 :
3550 7544 : all_grouped_cols = bms_add_members(all_grouped_cols, cols);
3551 7544 : continue;
3552 : }
3553 : else
3554 : {
3555 41590 : AggStatePerPhase phasedata = &aggstate->phases[++phase];
3556 : int num_sets;
3557 :
3558 41590 : phasedata->numsets = num_sets = list_length(aggnode->groupingSets);
3559 :
3560 41590 : if (num_sets)
3561 : {
3562 1000 : phasedata->gset_lengths = palloc(num_sets * sizeof(int));
3563 1000 : phasedata->grouped_cols = palloc(num_sets * sizeof(Bitmapset *));
3564 :
3565 1000 : i = 0;
3566 2936 : foreach(l, aggnode->groupingSets)
3567 : {
3568 1936 : int current_length = list_length(lfirst(l));
3569 1936 : Bitmapset *cols = NULL;
3570 :
3571 : /* planner forces this to be correct */
3572 3804 : for (j = 0; j < current_length; ++j)
3573 1868 : cols = bms_add_member(cols, aggnode->grpColIdx[j]);
3574 :
3575 1936 : phasedata->grouped_cols[i] = cols;
3576 1936 : phasedata->gset_lengths[i] = current_length;
3577 :
3578 1936 : ++i;
3579 : }
3580 :
3581 1000 : all_grouped_cols = bms_add_members(all_grouped_cols,
3582 1000 : phasedata->grouped_cols[0]);
3583 : }
3584 : else
3585 : {
3586 : Assert(phaseidx == 0);
3587 :
3588 40590 : phasedata->gset_lengths = NULL;
3589 40590 : phasedata->grouped_cols = NULL;
3590 : }
3591 :
3592 : /*
3593 : * If we are grouping, precompute fmgr lookup data for inner loop.
3594 : */
3595 41590 : if (aggnode->aggstrategy == AGG_SORTED)
3596 : {
3597 : /*
3598 : * Build a separate function for each subset of columns that
3599 : * need to be compared.
3600 : */
3601 2654 : phasedata->eqfunctions =
3602 2654 : (ExprState **) palloc0(aggnode->numCols * sizeof(ExprState *));
3603 :
3604 : /* for each grouping set */
3605 4294 : for (int k = 0; k < phasedata->numsets; k++)
3606 : {
3607 1640 : int length = phasedata->gset_lengths[k];
3608 :
3609 : /* nothing to do for empty grouping set */
3610 1640 : if (length == 0)
3611 338 : continue;
3612 :
3613 : /* if we already had one of this length, it'll do */
3614 1302 : if (phasedata->eqfunctions[length - 1] != NULL)
3615 138 : continue;
3616 :
3617 1164 : phasedata->eqfunctions[length - 1] =
3618 1164 : execTuplesMatchPrepare(scanDesc,
3619 : length,
3620 1164 : aggnode->grpColIdx,
3621 1164 : aggnode->grpOperators,
3622 1164 : aggnode->grpCollations,
3623 : (PlanState *) aggstate);
3624 : }
3625 :
3626 : /* and for all grouped columns, unless already computed */
3627 2654 : if (aggnode->numCols > 0 &&
3628 2560 : phasedata->eqfunctions[aggnode->numCols - 1] == NULL)
3629 : {
3630 1784 : phasedata->eqfunctions[aggnode->numCols - 1] =
3631 1784 : execTuplesMatchPrepare(scanDesc,
3632 : aggnode->numCols,
3633 1784 : aggnode->grpColIdx,
3634 1784 : aggnode->grpOperators,
3635 1784 : aggnode->grpCollations,
3636 : (PlanState *) aggstate);
3637 : }
3638 : }
3639 :
3640 41590 : phasedata->aggnode = aggnode;
3641 41590 : phasedata->aggstrategy = aggnode->aggstrategy;
3642 41590 : phasedata->sortnode = sortnode;
3643 : }
3644 : }
3645 :
3646 : /*
3647 : * Convert all_grouped_cols to a descending-order list.
3648 : */
3649 48136 : i = -1;
3650 60256 : while ((i = bms_next_member(all_grouped_cols, i)) >= 0)
3651 12120 : aggstate->all_grouped_cols = lcons_int(i, aggstate->all_grouped_cols);
3652 :
3653 : /*
3654 : * Set up aggregate-result storage in the output expr context, and also
3655 : * allocate my private per-agg working storage
3656 : */
3657 48136 : econtext = aggstate->ss.ps.ps_ExprContext;
3658 48136 : econtext->ecxt_aggvalues = (Datum *) palloc0(sizeof(Datum) * numaggs);
3659 48136 : econtext->ecxt_aggnulls = (bool *) palloc0(sizeof(bool) * numaggs);
3660 :
3661 48136 : peraggs = (AggStatePerAgg) palloc0(sizeof(AggStatePerAggData) * numaggs);
3662 48136 : pertransstates = (AggStatePerTrans) palloc0(sizeof(AggStatePerTransData) * numtrans);
3663 :
3664 48136 : aggstate->peragg = peraggs;
3665 48136 : aggstate->pertrans = pertransstates;
3666 :
3667 :
3668 48136 : aggstate->all_pergroups =
3669 48136 : (AggStatePerGroup *) palloc0(sizeof(AggStatePerGroup)
3670 48136 : * (numGroupingSets + numHashes));
3671 48136 : pergroups = aggstate->all_pergroups;
3672 :
3673 48136 : if (node->aggstrategy != AGG_HASHED)
3674 : {
3675 83540 : for (i = 0; i < numGroupingSets; i++)
3676 : {
3677 42202 : pergroups[i] = (AggStatePerGroup) palloc0(sizeof(AggStatePerGroupData)
3678 : * numaggs);
3679 : }
3680 :
3681 41338 : aggstate->pergroups = pergroups;
3682 41338 : pergroups += numGroupingSets;
3683 : }
3684 :
3685 : /*
3686 : * Hashing can only appear in the initial phase.
3687 : */
3688 48136 : if (use_hashing)
3689 : {
3690 7030 : Plan *outerplan = outerPlan(node);
3691 7030 : double totalGroups = 0;
3692 :
3693 7030 : aggstate->hash_spill_rslot = ExecInitExtraTupleSlot(estate, scanDesc,
3694 : &TTSOpsMinimalTuple);
3695 7030 : aggstate->hash_spill_wslot = ExecInitExtraTupleSlot(estate, scanDesc,
3696 : &TTSOpsVirtual);
3697 :
3698 : /* this is an array of pointers, not structures */
3699 7030 : aggstate->hash_pergroup = pergroups;
3700 :
3701 14060 : aggstate->hashentrysize = hash_agg_entry_size(aggstate->numtrans,
3702 7030 : outerplan->plan_width,
3703 : node->transitionSpace);
3704 :
3705 : /*
3706 : * Consider all of the grouping sets together when setting the limits
3707 : * and estimating the number of partitions. This can be inaccurate
3708 : * when there is more than one grouping set, but should still be
3709 : * reasonable.
3710 : */
3711 14574 : for (int k = 0; k < aggstate->num_hashes; k++)
3712 7544 : totalGroups += aggstate->perhash[k].aggnode->numGroups;
3713 :
3714 7030 : hash_agg_set_limits(aggstate->hashentrysize, totalGroups, 0,
3715 : &aggstate->hash_mem_limit,
3716 : &aggstate->hash_ngroups_limit,
3717 : &aggstate->hash_planned_partitions);
3718 7030 : find_hash_columns(aggstate);
3719 :
3720 : /* Skip massive memory allocation if we are just doing EXPLAIN */
3721 7030 : if (!(eflags & EXEC_FLAG_EXPLAIN_ONLY))
3722 5184 : build_hash_tables(aggstate);
3723 :
3724 7030 : aggstate->table_filled = false;
3725 :
3726 : /* Initialize this to 1, meaning nothing spilled, yet */
3727 7030 : aggstate->hash_batches_used = 1;
3728 : }
3729 :
3730 : /*
3731 : * Initialize current phase-dependent values to initial phase. The initial
3732 : * phase is 1 (first sort pass) for all strategies that use sorting (if
3733 : * hashing is being done too, then phase 0 is processed last); but if only
3734 : * hashing is being done, then phase 0 is all there is.
3735 : */
3736 48136 : if (node->aggstrategy == AGG_HASHED)
3737 : {
3738 6798 : aggstate->current_phase = 0;
3739 6798 : initialize_phase(aggstate, 0);
3740 6798 : select_current_set(aggstate, 0, true);
3741 : }
3742 : else
3743 : {
3744 41338 : aggstate->current_phase = 1;
3745 41338 : initialize_phase(aggstate, 1);
3746 41338 : select_current_set(aggstate, 0, false);
3747 : }
3748 :
3749 : /*
3750 : * Perform lookups of aggregate function info, and initialize the
3751 : * unchanging fields of the per-agg and per-trans data.
3752 : */
3753 102244 : foreach(l, aggstate->aggs)
3754 : {
3755 54114 : Aggref *aggref = lfirst(l);
3756 : AggStatePerAgg peragg;
3757 : AggStatePerTrans pertrans;
3758 : Oid aggTransFnInputTypes[FUNC_MAX_ARGS];
3759 : int numAggTransFnArgs;
3760 : int numDirectArgs;
3761 : HeapTuple aggTuple;
3762 : Form_pg_aggregate aggform;
3763 : AclResult aclresult;
3764 : Oid finalfn_oid;
3765 : Oid serialfn_oid,
3766 : deserialfn_oid;
3767 : Oid aggOwner;
3768 : Expr *finalfnexpr;
3769 : Oid aggtranstype;
3770 :
3771 : /* Planner should have assigned aggregate to correct level */
3772 : Assert(aggref->agglevelsup == 0);
3773 : /* ... and the split mode should match */
3774 : Assert(aggref->aggsplit == aggstate->aggsplit);
3775 :
3776 54114 : peragg = &peraggs[aggref->aggno];
3777 :
3778 : /* Check if we initialized the state for this aggregate already. */
3779 54114 : if (peragg->aggref != NULL)
3780 484 : continue;
3781 :
3782 53630 : peragg->aggref = aggref;
3783 53630 : peragg->transno = aggref->aggtransno;
3784 :
3785 : /* Fetch the pg_aggregate row */
3786 53630 : aggTuple = SearchSysCache1(AGGFNOID,
3787 : ObjectIdGetDatum(aggref->aggfnoid));
3788 53630 : if (!HeapTupleIsValid(aggTuple))
3789 0 : elog(ERROR, "cache lookup failed for aggregate %u",
3790 : aggref->aggfnoid);
3791 53630 : aggform = (Form_pg_aggregate) GETSTRUCT(aggTuple);
3792 :
3793 : /* Check permission to call aggregate function */
3794 53630 : aclresult = object_aclcheck(ProcedureRelationId, aggref->aggfnoid, GetUserId(),
3795 : ACL_EXECUTE);
3796 53630 : if (aclresult != ACLCHECK_OK)
3797 6 : aclcheck_error(aclresult, OBJECT_AGGREGATE,
3798 6 : get_func_name(aggref->aggfnoid));
3799 53624 : InvokeFunctionExecuteHook(aggref->aggfnoid);
3800 :
3801 : /* planner recorded transition state type in the Aggref itself */
3802 53624 : aggtranstype = aggref->aggtranstype;
3803 : Assert(OidIsValid(aggtranstype));
3804 :
3805 : /* Final function only required if we're finalizing the aggregates */
3806 53624 : if (DO_AGGSPLIT_SKIPFINAL(aggstate->aggsplit))
3807 5388 : peragg->finalfn_oid = finalfn_oid = InvalidOid;
3808 : else
3809 48236 : peragg->finalfn_oid = finalfn_oid = aggform->aggfinalfn;
3810 :
3811 53624 : serialfn_oid = InvalidOid;
3812 53624 : deserialfn_oid = InvalidOid;
3813 :
3814 : /*
3815 : * Check if serialization/deserialization is required. We only do it
3816 : * for aggregates that have transtype INTERNAL.
3817 : */
3818 53624 : if (aggtranstype == INTERNALOID)
3819 : {
3820 : /*
3821 : * The planner should only have generated a serialize agg node if
3822 : * every aggregate with an INTERNAL state has a serialization
3823 : * function. Verify that.
3824 : */
3825 22638 : if (DO_AGGSPLIT_SERIALIZE(aggstate->aggsplit))
3826 : {
3827 : /* serialization only valid when not running finalfn */
3828 : Assert(DO_AGGSPLIT_SKIPFINAL(aggstate->aggsplit));
3829 :
3830 336 : if (!OidIsValid(aggform->aggserialfn))
3831 0 : elog(ERROR, "serialfunc not provided for serialization aggregation");
3832 336 : serialfn_oid = aggform->aggserialfn;
3833 : }
3834 :
3835 : /* Likewise for deserialization functions */
3836 22638 : if (DO_AGGSPLIT_DESERIALIZE(aggstate->aggsplit))
3837 : {
3838 : /* deserialization only valid when combining states */
3839 : Assert(DO_AGGSPLIT_COMBINE(aggstate->aggsplit));
3840 :
3841 120 : if (!OidIsValid(aggform->aggdeserialfn))
3842 0 : elog(ERROR, "deserialfunc not provided for deserialization aggregation");
3843 120 : deserialfn_oid = aggform->aggdeserialfn;
3844 : }
3845 : }
3846 :
3847 : /* Check that aggregate owner has permission to call component fns */
3848 : {
3849 : HeapTuple procTuple;
3850 :
3851 53624 : procTuple = SearchSysCache1(PROCOID,
3852 : ObjectIdGetDatum(aggref->aggfnoid));
3853 53624 : if (!HeapTupleIsValid(procTuple))
3854 0 : elog(ERROR, "cache lookup failed for function %u",
3855 : aggref->aggfnoid);
3856 53624 : aggOwner = ((Form_pg_proc) GETSTRUCT(procTuple))->proowner;
3857 53624 : ReleaseSysCache(procTuple);
3858 :
3859 53624 : if (OidIsValid(finalfn_oid))
3860 : {
3861 24188 : aclresult = object_aclcheck(ProcedureRelationId, finalfn_oid, aggOwner,
3862 : ACL_EXECUTE);
3863 24188 : if (aclresult != ACLCHECK_OK)
3864 0 : aclcheck_error(aclresult, OBJECT_FUNCTION,
3865 0 : get_func_name(finalfn_oid));
3866 24188 : InvokeFunctionExecuteHook(finalfn_oid);
3867 : }
3868 53624 : if (OidIsValid(serialfn_oid))
3869 : {
3870 336 : aclresult = object_aclcheck(ProcedureRelationId, serialfn_oid, aggOwner,
3871 : ACL_EXECUTE);
3872 336 : if (aclresult != ACLCHECK_OK)
3873 0 : aclcheck_error(aclresult, OBJECT_FUNCTION,
3874 0 : get_func_name(serialfn_oid));
3875 336 : InvokeFunctionExecuteHook(serialfn_oid);
3876 : }
3877 53624 : if (OidIsValid(deserialfn_oid))
3878 : {
3879 120 : aclresult = object_aclcheck(ProcedureRelationId, deserialfn_oid, aggOwner,
3880 : ACL_EXECUTE);
3881 120 : if (aclresult != ACLCHECK_OK)
3882 0 : aclcheck_error(aclresult, OBJECT_FUNCTION,
3883 0 : get_func_name(deserialfn_oid));
3884 120 : InvokeFunctionExecuteHook(deserialfn_oid);
3885 : }
3886 : }
3887 :
3888 : /*
3889 : * Get actual datatypes of the (nominal) aggregate inputs. These
3890 : * could be different from the agg's declared input types, when the
3891 : * agg accepts ANY or a polymorphic type.
3892 : */
3893 53624 : numAggTransFnArgs = get_aggregate_argtypes(aggref,
3894 : aggTransFnInputTypes);
3895 :
3896 : /* Count the "direct" arguments, if any */
3897 53624 : numDirectArgs = list_length(aggref->aggdirectargs);
3898 :
3899 : /* Detect how many arguments to pass to the finalfn */
3900 53624 : if (aggform->aggfinalextra)
3901 16244 : peragg->numFinalArgs = numAggTransFnArgs + 1;
3902 : else
3903 37380 : peragg->numFinalArgs = numDirectArgs + 1;
3904 :
3905 : /* Initialize any direct-argument expressions */
3906 53624 : peragg->aggdirectargs = ExecInitExprList(aggref->aggdirectargs,
3907 : (PlanState *) aggstate);
3908 :
3909 : /*
3910 : * build expression trees using actual argument & result types for the
3911 : * finalfn, if it exists and is required.
3912 : */
3913 53624 : if (OidIsValid(finalfn_oid))
3914 : {
3915 24188 : build_aggregate_finalfn_expr(aggTransFnInputTypes,
3916 : peragg->numFinalArgs,
3917 : aggtranstype,
3918 : aggref->aggtype,
3919 : aggref->inputcollid,
3920 : finalfn_oid,
3921 : &finalfnexpr);
3922 24188 : fmgr_info(finalfn_oid, &peragg->finalfn);
3923 24188 : fmgr_info_set_expr((Node *) finalfnexpr, &peragg->finalfn);
3924 : }
3925 :
3926 : /* get info about the output value's datatype */
3927 53624 : get_typlenbyval(aggref->aggtype,
3928 : &peragg->resulttypeLen,
3929 : &peragg->resulttypeByVal);
3930 :
3931 : /*
3932 : * Build working state for invoking the transition function, if we
3933 : * haven't done it already.
3934 : */
3935 53624 : pertrans = &pertransstates[aggref->aggtransno];
3936 53624 : if (pertrans->aggref == NULL)
3937 : {
3938 : Datum textInitVal;
3939 : Datum initValue;
3940 : bool initValueIsNull;
3941 : Oid transfn_oid;
3942 :
3943 : /*
3944 : * If this aggregation is performing state combines, then instead
3945 : * of using the transition function, we'll use the combine
3946 : * function.
3947 : */
3948 53366 : if (DO_AGGSPLIT_COMBINE(aggstate->aggsplit))
3949 : {
3950 2192 : transfn_oid = aggform->aggcombinefn;
3951 :
3952 : /* If not set then the planner messed up */
3953 2192 : if (!OidIsValid(transfn_oid))
3954 0 : elog(ERROR, "combinefn not set for aggregate function");
3955 : }
3956 : else
3957 51174 : transfn_oid = aggform->aggtransfn;
3958 :
3959 53366 : aclresult = object_aclcheck(ProcedureRelationId, transfn_oid, aggOwner, ACL_EXECUTE);
3960 53366 : if (aclresult != ACLCHECK_OK)
3961 0 : aclcheck_error(aclresult, OBJECT_FUNCTION,
3962 0 : get_func_name(transfn_oid));
3963 53366 : InvokeFunctionExecuteHook(transfn_oid);
3964 :
3965 : /*
3966 : * initval is potentially null, so don't try to access it as a
3967 : * struct field. Must do it the hard way with SysCacheGetAttr.
3968 : */
3969 53366 : textInitVal = SysCacheGetAttr(AGGFNOID, aggTuple,
3970 : Anum_pg_aggregate_agginitval,
3971 : &initValueIsNull);
3972 53366 : if (initValueIsNull)
3973 31854 : initValue = (Datum) 0;
3974 : else
3975 21512 : initValue = GetAggInitVal(textInitVal, aggtranstype);
3976 :
3977 53366 : if (DO_AGGSPLIT_COMBINE(aggstate->aggsplit))
3978 : {
3979 2192 : Oid combineFnInputTypes[] = {aggtranstype,
3980 : aggtranstype};
3981 :
3982 : /*
3983 : * When combining there's only one input, the to-be-combined
3984 : * transition value. The transition value is not counted
3985 : * here.
3986 : */
3987 2192 : pertrans->numTransInputs = 1;
3988 :
3989 : /* aggcombinefn always has two arguments of aggtranstype */
3990 2192 : build_pertrans_for_aggref(pertrans, aggstate, estate,
3991 : aggref, transfn_oid, aggtranstype,
3992 : serialfn_oid, deserialfn_oid,
3993 : initValue, initValueIsNull,
3994 : combineFnInputTypes, 2);
3995 :
3996 : /*
3997 : * Ensure that a combine function to combine INTERNAL states
3998 : * is not strict. This should have been checked during CREATE
3999 : * AGGREGATE, but the strict property could have been changed
4000 : * since then.
4001 : */
4002 2192 : if (pertrans->transfn.fn_strict && aggtranstype == INTERNALOID)
4003 0 : ereport(ERROR,
4004 : (errcode(ERRCODE_INVALID_FUNCTION_DEFINITION),
4005 : errmsg("combine function with transition type %s must not be declared STRICT",
4006 : format_type_be(aggtranstype))));
4007 : }
4008 : else
4009 : {
4010 : /* Detect how many arguments to pass to the transfn */
4011 51174 : if (AGGKIND_IS_ORDERED_SET(aggref->aggkind))
4012 252 : pertrans->numTransInputs = list_length(aggref->args);
4013 : else
4014 50922 : pertrans->numTransInputs = numAggTransFnArgs;
4015 :
4016 51174 : build_pertrans_for_aggref(pertrans, aggstate, estate,
4017 : aggref, transfn_oid, aggtranstype,
4018 : serialfn_oid, deserialfn_oid,
4019 : initValue, initValueIsNull,
4020 : aggTransFnInputTypes,
4021 : numAggTransFnArgs);
4022 :
4023 : /*
4024 : * If the transfn is strict and the initval is NULL, make sure
4025 : * input type and transtype are the same (or at least
4026 : * binary-compatible), so that it's OK to use the first
4027 : * aggregated input value as the initial transValue. This
4028 : * should have been checked at agg definition time, but we
4029 : * must check again in case the transfn's strictness property
4030 : * has been changed.
4031 : */
4032 51174 : if (pertrans->transfn.fn_strict && pertrans->initValueIsNull)
4033 : {
4034 5072 : if (numAggTransFnArgs <= numDirectArgs ||
4035 5072 : !IsBinaryCoercible(aggTransFnInputTypes[numDirectArgs],
4036 : aggtranstype))
4037 0 : ereport(ERROR,
4038 : (errcode(ERRCODE_INVALID_FUNCTION_DEFINITION),
4039 : errmsg("aggregate %u needs to have compatible input type and transition type",
4040 : aggref->aggfnoid)));
4041 : }
4042 : }
4043 : }
4044 : else
4045 258 : pertrans->aggshared = true;
4046 53624 : ReleaseSysCache(aggTuple);
4047 : }
4048 :
4049 : /*
4050 : * Last, check whether any more aggregates got added onto the node while
4051 : * we processed the expressions for the aggregate arguments (including not
4052 : * only the regular arguments and FILTER expressions handled immediately
4053 : * above, but any direct arguments we might've handled earlier). If so,
4054 : * we have nested aggregate functions, which is semantically nonsensical,
4055 : * so complain. (This should have been caught by the parser, so we don't
4056 : * need to work hard on a helpful error message; but we defend against it
4057 : * here anyway, just to be sure.)
4058 : */
4059 48130 : if (numaggrefs != list_length(aggstate->aggs))
4060 0 : ereport(ERROR,
4061 : (errcode(ERRCODE_GROUPING_ERROR),
4062 : errmsg("aggregate function calls cannot be nested")));
4063 :
4064 : /*
4065 : * Build expressions doing all the transition work at once. We build a
4066 : * different one for each phase, as the number of transition function
4067 : * invocation can differ between phases. Note this'll work both for
4068 : * transition and combination functions (although there'll only be one
4069 : * phase in the latter case).
4070 : */
4071 137844 : for (phaseidx = 0; phaseidx < aggstate->numphases; phaseidx++)
4072 : {
4073 89714 : AggStatePerPhase phase = &aggstate->phases[phaseidx];
4074 89714 : bool dohash = false;
4075 89714 : bool dosort = false;
4076 :
4077 : /* phase 0 doesn't necessarily exist */
4078 89714 : if (!phase->aggnode)
4079 41100 : continue;
4080 :
4081 48614 : if (aggstate->aggstrategy == AGG_MIXED && phaseidx == 1)
4082 : {
4083 : /*
4084 : * Phase one, and only phase one, in a mixed agg performs both
4085 : * sorting and aggregation.
4086 : */
4087 232 : dohash = true;
4088 232 : dosort = true;
4089 : }
4090 48382 : else if (aggstate->aggstrategy == AGG_MIXED && phaseidx == 0)
4091 : {
4092 : /*
4093 : * No need to compute a transition function for an AGG_MIXED phase
4094 : * 0 - the contents of the hashtables will have been computed
4095 : * during phase 1.
4096 : */
4097 232 : continue;
4098 : }
4099 48150 : else if (phase->aggstrategy == AGG_PLAIN ||
4100 9390 : phase->aggstrategy == AGG_SORTED)
4101 : {
4102 41352 : dohash = false;
4103 41352 : dosort = true;
4104 : }
4105 6798 : else if (phase->aggstrategy == AGG_HASHED)
4106 : {
4107 6798 : dohash = true;
4108 6798 : dosort = false;
4109 : }
4110 : else
4111 : Assert(false);
4112 :
4113 48382 : phase->evaltrans = ExecBuildAggTrans(aggstate, phase, dosort, dohash,
4114 : false);
4115 :
4116 : /* cache compiled expression for outer slot without NULL check */
4117 48382 : phase->evaltrans_cache[0][0] = phase->evaltrans;
4118 : }
4119 :
4120 48130 : return aggstate;
4121 : }
4122 :
4123 : /*
4124 : * Build the state needed to calculate a state value for an aggregate.
4125 : *
4126 : * This initializes all the fields in 'pertrans'. 'aggref' is the aggregate
4127 : * to initialize the state for. 'transfn_oid', 'aggtranstype', and the rest
4128 : * of the arguments could be calculated from 'aggref', but the caller has
4129 : * calculated them already, so might as well pass them.
4130 : *
4131 : * 'transfn_oid' may be either the Oid of the aggtransfn or the aggcombinefn.
4132 : */
4133 : static void
4134 53366 : build_pertrans_for_aggref(AggStatePerTrans pertrans,
4135 : AggState *aggstate, EState *estate,
4136 : Aggref *aggref,
4137 : Oid transfn_oid, Oid aggtranstype,
4138 : Oid aggserialfn, Oid aggdeserialfn,
4139 : Datum initValue, bool initValueIsNull,
4140 : Oid *inputTypes, int numArguments)
4141 : {
4142 53366 : int numGroupingSets = Max(aggstate->maxsets, 1);
4143 : Expr *transfnexpr;
4144 : int numTransArgs;
4145 53366 : Expr *serialfnexpr = NULL;
4146 53366 : Expr *deserialfnexpr = NULL;
4147 : ListCell *lc;
4148 : int numInputs;
4149 : int numDirectArgs;
4150 : List *sortlist;
4151 : int numSortCols;
4152 : int numDistinctCols;
4153 : int i;
4154 :
4155 : /* Begin filling in the pertrans data */
4156 53366 : pertrans->aggref = aggref;
4157 53366 : pertrans->aggshared = false;
4158 53366 : pertrans->aggCollation = aggref->inputcollid;
4159 53366 : pertrans->transfn_oid = transfn_oid;
4160 53366 : pertrans->serialfn_oid = aggserialfn;
4161 53366 : pertrans->deserialfn_oid = aggdeserialfn;
4162 53366 : pertrans->initValue = initValue;
4163 53366 : pertrans->initValueIsNull = initValueIsNull;
4164 :
4165 : /* Count the "direct" arguments, if any */
4166 53366 : numDirectArgs = list_length(aggref->aggdirectargs);
4167 :
4168 : /* Count the number of aggregated input columns */
4169 53366 : pertrans->numInputs = numInputs = list_length(aggref->args);
4170 :
4171 53366 : pertrans->aggtranstype = aggtranstype;
4172 :
4173 : /* account for the current transition state */
4174 53366 : numTransArgs = pertrans->numTransInputs + 1;
4175 :
4176 : /*
4177 : * Set up infrastructure for calling the transfn. Note that invtransfn is
4178 : * not needed here.
4179 : */
4180 53366 : build_aggregate_transfn_expr(inputTypes,
4181 : numArguments,
4182 : numDirectArgs,
4183 53366 : aggref->aggvariadic,
4184 : aggtranstype,
4185 : aggref->inputcollid,
4186 : transfn_oid,
4187 : InvalidOid,
4188 : &transfnexpr,
4189 : NULL);
4190 :
4191 53366 : fmgr_info(transfn_oid, &pertrans->transfn);
4192 53366 : fmgr_info_set_expr((Node *) transfnexpr, &pertrans->transfn);
4193 :
4194 53366 : pertrans->transfn_fcinfo =
4195 53366 : (FunctionCallInfo) palloc(SizeForFunctionCallInfo(numTransArgs));
4196 53366 : InitFunctionCallInfoData(*pertrans->transfn_fcinfo,
4197 : &pertrans->transfn,
4198 : numTransArgs,
4199 : pertrans->aggCollation,
4200 : (Node *) aggstate, NULL);
4201 :
4202 : /* get info about the state value's datatype */
4203 53366 : get_typlenbyval(aggtranstype,
4204 : &pertrans->transtypeLen,
4205 : &pertrans->transtypeByVal);
4206 :
4207 53366 : if (OidIsValid(aggserialfn))
4208 : {
4209 336 : build_aggregate_serialfn_expr(aggserialfn,
4210 : &serialfnexpr);
4211 336 : fmgr_info(aggserialfn, &pertrans->serialfn);
4212 336 : fmgr_info_set_expr((Node *) serialfnexpr, &pertrans->serialfn);
4213 :
4214 336 : pertrans->serialfn_fcinfo =
4215 336 : (FunctionCallInfo) palloc(SizeForFunctionCallInfo(1));
4216 336 : InitFunctionCallInfoData(*pertrans->serialfn_fcinfo,
4217 : &pertrans->serialfn,
4218 : 1,
4219 : InvalidOid,
4220 : (Node *) aggstate, NULL);
4221 : }
4222 :
4223 53366 : if (OidIsValid(aggdeserialfn))
4224 : {
4225 120 : build_aggregate_deserialfn_expr(aggdeserialfn,
4226 : &deserialfnexpr);
4227 120 : fmgr_info(aggdeserialfn, &pertrans->deserialfn);
4228 120 : fmgr_info_set_expr((Node *) deserialfnexpr, &pertrans->deserialfn);
4229 :
4230 120 : pertrans->deserialfn_fcinfo =
4231 120 : (FunctionCallInfo) palloc(SizeForFunctionCallInfo(2));
4232 120 : InitFunctionCallInfoData(*pertrans->deserialfn_fcinfo,
4233 : &pertrans->deserialfn,
4234 : 2,
4235 : InvalidOid,
4236 : (Node *) aggstate, NULL);
4237 : }
4238 :
4239 : /*
4240 : * If we're doing either DISTINCT or ORDER BY for a plain agg, then we
4241 : * have a list of SortGroupClause nodes; fish out the data in them and
4242 : * stick them into arrays. We ignore ORDER BY for an ordered-set agg,
4243 : * however; the agg's transfn and finalfn are responsible for that.
4244 : *
4245 : * When the planner has set the aggpresorted flag, the input to the
4246 : * aggregate is already correctly sorted. For ORDER BY aggregates we can
4247 : * simply treat these as normal aggregates. For presorted DISTINCT
4248 : * aggregates an extra step must be added to remove duplicate consecutive
4249 : * inputs.
4250 : *
4251 : * Note that by construction, if there is a DISTINCT clause then the ORDER
4252 : * BY clause is a prefix of it (see transformDistinctClause).
4253 : */
4254 53366 : if (AGGKIND_IS_ORDERED_SET(aggref->aggkind))
4255 : {
4256 252 : sortlist = NIL;
4257 252 : numSortCols = numDistinctCols = 0;
4258 252 : pertrans->aggsortrequired = false;
4259 : }
4260 53114 : else if (aggref->aggpresorted && aggref->aggdistinct == NIL)
4261 : {
4262 2098 : sortlist = NIL;
4263 2098 : numSortCols = numDistinctCols = 0;
4264 2098 : pertrans->aggsortrequired = false;
4265 : }
4266 51016 : else if (aggref->aggdistinct)
4267 : {
4268 582 : sortlist = aggref->aggdistinct;
4269 582 : numSortCols = numDistinctCols = list_length(sortlist);
4270 : Assert(numSortCols >= list_length(aggref->aggorder));
4271 582 : pertrans->aggsortrequired = !aggref->aggpresorted;
4272 : }
4273 : else
4274 : {
4275 50434 : sortlist = aggref->aggorder;
4276 50434 : numSortCols = list_length(sortlist);
4277 50434 : numDistinctCols = 0;
4278 50434 : pertrans->aggsortrequired = (numSortCols > 0);
4279 : }
4280 :
4281 53366 : pertrans->numSortCols = numSortCols;
4282 53366 : pertrans->numDistinctCols = numDistinctCols;
4283 :
4284 : /*
4285 : * If we have either sorting or filtering to do, create a tupledesc and
4286 : * slot corresponding to the aggregated inputs (including sort
4287 : * expressions) of the agg.
4288 : */
4289 53366 : if (numSortCols > 0 || aggref->aggfilter)
4290 : {
4291 1430 : pertrans->sortdesc = ExecTypeFromTL(aggref->args);
4292 1430 : pertrans->sortslot =
4293 1430 : ExecInitExtraTupleSlot(estate, pertrans->sortdesc,
4294 : &TTSOpsMinimalTuple);
4295 : }
4296 :
4297 53366 : if (numSortCols > 0)
4298 : {
4299 : /*
4300 : * We don't implement DISTINCT or ORDER BY aggs in the HASHED case
4301 : * (yet)
4302 : */
4303 : Assert(aggstate->aggstrategy != AGG_HASHED && aggstate->aggstrategy != AGG_MIXED);
4304 :
4305 : /* ORDER BY aggregates are not supported with partial aggregation */
4306 : Assert(!DO_AGGSPLIT_COMBINE(aggstate->aggsplit));
4307 :
4308 : /* If we have only one input, we need its len/byval info. */
4309 720 : if (numInputs == 1)
4310 : {
4311 570 : get_typlenbyval(inputTypes[numDirectArgs],
4312 : &pertrans->inputtypeLen,
4313 : &pertrans->inputtypeByVal);
4314 : }
4315 150 : else if (numDistinctCols > 0)
4316 : {
4317 : /* we will need an extra slot to store prior values */
4318 108 : pertrans->uniqslot =
4319 108 : ExecInitExtraTupleSlot(estate, pertrans->sortdesc,
4320 : &TTSOpsMinimalTuple);
4321 : }
4322 :
4323 : /* Extract the sort information for use later */
4324 720 : pertrans->sortColIdx =
4325 720 : (AttrNumber *) palloc(numSortCols * sizeof(AttrNumber));
4326 720 : pertrans->sortOperators =
4327 720 : (Oid *) palloc(numSortCols * sizeof(Oid));
4328 720 : pertrans->sortCollations =
4329 720 : (Oid *) palloc(numSortCols * sizeof(Oid));
4330 720 : pertrans->sortNullsFirst =
4331 720 : (bool *) palloc(numSortCols * sizeof(bool));
4332 :
4333 720 : i = 0;
4334 1638 : foreach(lc, sortlist)
4335 : {
4336 918 : SortGroupClause *sortcl = (SortGroupClause *) lfirst(lc);
4337 918 : TargetEntry *tle = get_sortgroupclause_tle(sortcl, aggref->args);
4338 :
4339 : /* the parser should have made sure of this */
4340 : Assert(OidIsValid(sortcl->sortop));
4341 :
4342 918 : pertrans->sortColIdx[i] = tle->resno;
4343 918 : pertrans->sortOperators[i] = sortcl->sortop;
4344 918 : pertrans->sortCollations[i] = exprCollation((Node *) tle->expr);
4345 918 : pertrans->sortNullsFirst[i] = sortcl->nulls_first;
4346 918 : i++;
4347 : }
4348 : Assert(i == numSortCols);
4349 : }
4350 :
4351 53366 : if (aggref->aggdistinct)
4352 : {
4353 : Oid *ops;
4354 :
4355 : Assert(numArguments > 0);
4356 : Assert(list_length(aggref->aggdistinct) == numDistinctCols);
4357 :
4358 582 : ops = palloc(numDistinctCols * sizeof(Oid));
4359 :
4360 582 : i = 0;
4361 1344 : foreach(lc, aggref->aggdistinct)
4362 762 : ops[i++] = ((SortGroupClause *) lfirst(lc))->eqop;
4363 :
4364 : /* lookup / build the necessary comparators */
4365 582 : if (numDistinctCols == 1)
4366 474 : fmgr_info(get_opcode(ops[0]), &pertrans->equalfnOne);
4367 : else
4368 108 : pertrans->equalfnMulti =
4369 108 : execTuplesMatchPrepare(pertrans->sortdesc,
4370 : numDistinctCols,
4371 108 : pertrans->sortColIdx,
4372 : ops,
4373 108 : pertrans->sortCollations,
4374 : &aggstate->ss.ps);
4375 582 : pfree(ops);
4376 : }
4377 :
4378 53366 : pertrans->sortstates = (Tuplesortstate **)
4379 53366 : palloc0(sizeof(Tuplesortstate *) * numGroupingSets);
4380 53366 : }
4381 :
4382 :
4383 : static Datum
4384 21512 : GetAggInitVal(Datum textInitVal, Oid transtype)
4385 : {
4386 : Oid typinput,
4387 : typioparam;
4388 : char *strInitVal;
4389 : Datum initVal;
4390 :
4391 21512 : getTypeInputInfo(transtype, &typinput, &typioparam);
4392 21512 : strInitVal = TextDatumGetCString(textInitVal);
4393 21512 : initVal = OidInputFunctionCall(typinput, strInitVal,
4394 : typioparam, -1);
4395 21512 : pfree(strInitVal);
4396 21512 : return initVal;
4397 : }
4398 :
4399 : void
4400 47926 : ExecEndAgg(AggState *node)
4401 : {
4402 : PlanState *outerPlan;
4403 : int transno;
4404 47926 : int numGroupingSets = Max(node->maxsets, 1);
4405 : int setno;
4406 :
4407 : /*
4408 : * When ending a parallel worker, copy the statistics gathered by the
4409 : * worker back into shared memory so that it can be picked up by the main
4410 : * process to report in EXPLAIN ANALYZE.
4411 : */
4412 47926 : if (node->shared_info && IsParallelWorker())
4413 : {
4414 : AggregateInstrumentation *si;
4415 :
4416 : Assert(ParallelWorkerNumber <= node->shared_info->num_workers);
4417 168 : si = &node->shared_info->sinstrument[ParallelWorkerNumber];
4418 168 : si->hash_batches_used = node->hash_batches_used;
4419 168 : si->hash_disk_used = node->hash_disk_used;
4420 168 : si->hash_mem_peak = node->hash_mem_peak;
4421 : }
4422 :
4423 : /* Make sure we have closed any open tuplesorts */
4424 :
4425 47926 : if (node->sort_in)
4426 162 : tuplesort_end(node->sort_in);
4427 47926 : if (node->sort_out)
4428 48 : tuplesort_end(node->sort_out);
4429 :
4430 47926 : hashagg_reset_spill_state(node);
4431 :
4432 : /* Release hash tables too */
4433 47926 : if (node->hash_metacxt != NULL)
4434 : {
4435 7022 : MemoryContextDelete(node->hash_metacxt);
4436 7022 : node->hash_metacxt = NULL;
4437 : }
4438 47926 : if (node->hash_tuplescxt != NULL)
4439 : {
4440 7022 : MemoryContextDelete(node->hash_tuplescxt);
4441 7022 : node->hash_tuplescxt = NULL;
4442 : }
4443 :
4444 101084 : for (transno = 0; transno < node->numtrans; transno++)
4445 : {
4446 53158 : AggStatePerTrans pertrans = &node->pertrans[transno];
4447 :
4448 107354 : for (setno = 0; setno < numGroupingSets; setno++)
4449 : {
4450 54196 : if (pertrans->sortstates[setno])
4451 0 : tuplesort_end(pertrans->sortstates[setno]);
4452 : }
4453 : }
4454 :
4455 : /* And ensure any agg shutdown callbacks have been called */
4456 96716 : for (setno = 0; setno < numGroupingSets; setno++)
4457 48790 : ReScanExprContext(node->aggcontexts[setno]);
4458 47926 : if (node->hashcontext)
4459 7022 : ReScanExprContext(node->hashcontext);
4460 :
4461 47926 : outerPlan = outerPlanState(node);
4462 47926 : ExecEndNode(outerPlan);
4463 47926 : }
4464 :
4465 : void
4466 54120 : ExecReScanAgg(AggState *node)
4467 : {
4468 54120 : ExprContext *econtext = node->ss.ps.ps_ExprContext;
4469 54120 : PlanState *outerPlan = outerPlanState(node);
4470 54120 : Agg *aggnode = (Agg *) node->ss.ps.plan;
4471 : int transno;
4472 54120 : int numGroupingSets = Max(node->maxsets, 1);
4473 : int setno;
4474 :
4475 54120 : node->agg_done = false;
4476 :
4477 54120 : if (node->aggstrategy == AGG_HASHED)
4478 : {
4479 : /*
4480 : * In the hashed case, if we haven't yet built the hash table then we
4481 : * can just return; nothing done yet, so nothing to undo. If subnode's
4482 : * chgParam is not NULL then it will be re-scanned by ExecProcNode,
4483 : * else no reason to re-scan it at all.
4484 : */
4485 13414 : if (!node->table_filled)
4486 150 : return;
4487 :
4488 : /*
4489 : * If we do have the hash table, and it never spilled, and the subplan
4490 : * does not have any parameter changes, and none of our own parameter
4491 : * changes affect input expressions of the aggregated functions, then
4492 : * we can just rescan the existing hash table; no need to build it
4493 : * again.
4494 : */
4495 13264 : if (outerPlan->chgParam == NULL && !node->hash_ever_spilled &&
4496 916 : !bms_overlap(node->ss.ps.chgParam, aggnode->aggParams))
4497 : {
4498 892 : ResetTupleHashIterator(node->perhash[0].hashtable,
4499 : &node->perhash[0].hashiter);
4500 892 : select_current_set(node, 0, true);
4501 892 : return;
4502 : }
4503 : }
4504 :
4505 : /* Make sure we have closed any open tuplesorts */
4506 120892 : for (transno = 0; transno < node->numtrans; transno++)
4507 : {
4508 135664 : for (setno = 0; setno < numGroupingSets; setno++)
4509 : {
4510 67850 : AggStatePerTrans pertrans = &node->pertrans[transno];
4511 :
4512 67850 : if (pertrans->sortstates[setno])
4513 : {
4514 0 : tuplesort_end(pertrans->sortstates[setno]);
4515 0 : pertrans->sortstates[setno] = NULL;
4516 : }
4517 : }
4518 : }
4519 :
4520 : /*
4521 : * We don't need to ReScanExprContext the output tuple context here;
4522 : * ExecReScan already did it. But we do need to reset our per-grouping-set
4523 : * contexts, which may have transvalues stored in them. (We use rescan
4524 : * rather than just reset because transfns may have registered callbacks
4525 : * that need to be run now.) For the AGG_HASHED case, see below.
4526 : */
4527 :
4528 106192 : for (setno = 0; setno < numGroupingSets; setno++)
4529 : {
4530 53114 : ReScanExprContext(node->aggcontexts[setno]);
4531 : }
4532 :
4533 : /* Release first tuple of group, if we have made a copy */
4534 53078 : if (node->grp_firstTuple != NULL)
4535 : {
4536 0 : heap_freetuple(node->grp_firstTuple);
4537 0 : node->grp_firstTuple = NULL;
4538 : }
4539 53078 : ExecClearTuple(node->ss.ss_ScanTupleSlot);
4540 :
4541 : /* Forget current agg values */
4542 120892 : MemSet(econtext->ecxt_aggvalues, 0, sizeof(Datum) * node->numaggs);
4543 53078 : MemSet(econtext->ecxt_aggnulls, 0, sizeof(bool) * node->numaggs);
4544 :
4545 : /*
4546 : * With AGG_HASHED/MIXED, the hash table is allocated in a sub-context of
4547 : * the hashcontext. This used to be an issue, but now, resetting a context
4548 : * automatically deletes sub-contexts too.
4549 : */
4550 53078 : if (node->aggstrategy == AGG_HASHED || node->aggstrategy == AGG_MIXED)
4551 : {
4552 12402 : hashagg_reset_spill_state(node);
4553 :
4554 12402 : node->hash_ever_spilled = false;
4555 12402 : node->hash_spill_mode = false;
4556 12402 : node->hash_ngroups_current = 0;
4557 :
4558 12402 : ReScanExprContext(node->hashcontext);
4559 : /* Rebuild empty hash table(s) */
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 53078 : if (node->aggstrategy != AGG_HASHED)
4568 : {
4569 : /*
4570 : * Reset the per-group state (in particular, mark transvalues null)
4571 : */
4572 81448 : for (setno = 0; setno < numGroupingSets; setno++)
4573 : {
4574 176346 : MemSet(node->pergroups[setno], 0,
4575 : sizeof(AggStatePerGroupData) * node->numaggs);
4576 : }
4577 :
4578 : /* reset to phase 1 */
4579 40706 : initialize_phase(node, 1);
4580 :
4581 40706 : node->input_done = false;
4582 40706 : node->projected_set = -1;
4583 : }
4584 :
4585 53078 : if (outerPlan->chgParam == NULL)
4586 188 : 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 5439564 : AggCheckCallContext(FunctionCallInfo fcinfo, MemoryContext *aggcontext)
4614 : {
4615 5439564 : if (fcinfo->context && IsA(fcinfo->context, AggState))
4616 : {
4617 5428178 : if (aggcontext)
4618 : {
4619 2751434 : AggState *aggstate = ((AggState *) fcinfo->context);
4620 2751434 : ExprContext *cxt = aggstate->curaggcontext;
4621 :
4622 2751434 : *aggcontext = cxt->ecxt_per_tuple_memory;
4623 : }
4624 5428178 : 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 1560 : ExecAggInitializeWorker(AggState *node, ParallelWorkerContext *pwcxt)
4832 : {
4833 1560 : node->shared_info =
4834 1560 : shm_toc_lookup(pwcxt->toc, node->ss.ps.plan->plan_node_id, true);
4835 1560 : }
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 : }
|
