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) 1996-2023, 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/makefuncs.h"
264 : #include "nodes/nodeFuncs.h"
265 : #include "optimizer/optimizer.h"
266 : #include "parser/parse_agg.h"
267 : #include "parser/parse_coerce.h"
268 : #include "utils/acl.h"
269 : #include "utils/builtins.h"
270 : #include "utils/datum.h"
271 : #include "utils/dynahash.h"
272 : #include "utils/expandeddatum.h"
273 : #include "utils/logtape.h"
274 : #include "utils/lsyscache.h"
275 : #include "utils/memutils.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 : * Estimate chunk overhead as a constant 16 bytes. XXX: should this be
319 : * improved?
320 : */
321 : #define CHUNKHDRSZ 16
322 :
323 : /*
324 : * Represents partitioned spill data for a single hashtable. Contains the
325 : * necessary information to route tuples to the correct partition, and to
326 : * transform the spilled data into new batches.
327 : *
328 : * The high bits are used for partition selection (when recursing, we ignore
329 : * the bits that have already been used for partition selection at an earlier
330 : * level).
331 : */
332 : typedef struct HashAggSpill
333 : {
334 : int npartitions; /* number of partitions */
335 : LogicalTape **partitions; /* spill partition tapes */
336 : int64 *ntuples; /* number of tuples in each partition */
337 : uint32 mask; /* mask to find partition from hash value */
338 : int shift; /* after masking, shift by this amount */
339 : hyperLogLogState *hll_card; /* cardinality estimate for contents */
340 : } HashAggSpill;
341 :
342 : /*
343 : * Represents work to be done for one pass of hash aggregation (with only one
344 : * grouping set).
345 : *
346 : * Also tracks the bits of the hash already used for partition selection by
347 : * earlier iterations, so that this batch can use new bits. If all bits have
348 : * already been used, no partitioning will be done (any spilled data will go
349 : * to a single output tape).
350 : */
351 : typedef struct HashAggBatch
352 : {
353 : int setno; /* grouping set */
354 : int used_bits; /* number of bits of hash already used */
355 : LogicalTape *input_tape; /* input partition tape */
356 : int64 input_tuples; /* number of tuples in this batch */
357 : double input_card; /* estimated group cardinality */
358 : } HashAggBatch;
359 :
360 : /* used to find referenced colnos */
361 : typedef struct FindColsContext
362 : {
363 : bool is_aggref; /* is under an aggref */
364 : Bitmapset *aggregated; /* column references under an aggref */
365 : Bitmapset *unaggregated; /* other column references */
366 : } FindColsContext;
367 :
368 : static void select_current_set(AggState *aggstate, int setno, bool is_hash);
369 : static void initialize_phase(AggState *aggstate, int newphase);
370 : static TupleTableSlot *fetch_input_tuple(AggState *aggstate);
371 : static void initialize_aggregates(AggState *aggstate,
372 : AggStatePerGroup *pergroups,
373 : int numReset);
374 : static void advance_transition_function(AggState *aggstate,
375 : AggStatePerTrans pertrans,
376 : AggStatePerGroup pergroupstate);
377 : static void advance_aggregates(AggState *aggstate);
378 : static void process_ordered_aggregate_single(AggState *aggstate,
379 : AggStatePerTrans pertrans,
380 : AggStatePerGroup pergroupstate);
381 : static void process_ordered_aggregate_multi(AggState *aggstate,
382 : AggStatePerTrans pertrans,
383 : AggStatePerGroup pergroupstate);
384 : static void finalize_aggregate(AggState *aggstate,
385 : AggStatePerAgg peragg,
386 : AggStatePerGroup pergroupstate,
387 : Datum *resultVal, bool *resultIsNull);
388 : static void finalize_partialaggregate(AggState *aggstate,
389 : AggStatePerAgg peragg,
390 : AggStatePerGroup pergroupstate,
391 : Datum *resultVal, bool *resultIsNull);
392 : static inline void prepare_hash_slot(AggStatePerHash perhash,
393 : TupleTableSlot *inputslot,
394 : TupleTableSlot *hashslot);
395 : static void prepare_projection_slot(AggState *aggstate,
396 : TupleTableSlot *slot,
397 : int currentSet);
398 : static void finalize_aggregates(AggState *aggstate,
399 : AggStatePerAgg peraggs,
400 : AggStatePerGroup pergroup);
401 : static TupleTableSlot *project_aggregates(AggState *aggstate);
402 : static void find_cols(AggState *aggstate, Bitmapset **aggregated,
403 : Bitmapset **unaggregated);
404 : static bool find_cols_walker(Node *node, FindColsContext *context);
405 : static void build_hash_tables(AggState *aggstate);
406 : static void build_hash_table(AggState *aggstate, int setno, long nbuckets);
407 : static void hashagg_recompile_expressions(AggState *aggstate, bool minslot,
408 : bool nullcheck);
409 : static long hash_choose_num_buckets(double hashentrysize,
410 : long ngroups, Size memory);
411 : static int hash_choose_num_partitions(double input_groups,
412 : double hashentrysize,
413 : int used_bits,
414 : int *log2_npartitions);
415 : static void initialize_hash_entry(AggState *aggstate,
416 : TupleHashTable hashtable,
417 : TupleHashEntry entry);
418 : static void lookup_hash_entries(AggState *aggstate);
419 : static TupleTableSlot *agg_retrieve_direct(AggState *aggstate);
420 : static void agg_fill_hash_table(AggState *aggstate);
421 : static bool agg_refill_hash_table(AggState *aggstate);
422 : static TupleTableSlot *agg_retrieve_hash_table(AggState *aggstate);
423 : static TupleTableSlot *agg_retrieve_hash_table_in_memory(AggState *aggstate);
424 : static void hash_agg_check_limits(AggState *aggstate);
425 : static void hash_agg_enter_spill_mode(AggState *aggstate);
426 : static void hash_agg_update_metrics(AggState *aggstate, bool from_tape,
427 : int npartitions);
428 : static void hashagg_finish_initial_spills(AggState *aggstate);
429 : static void hashagg_reset_spill_state(AggState *aggstate);
430 : static HashAggBatch *hashagg_batch_new(LogicalTape *input_tape, int setno,
431 : int64 input_tuples, double input_card,
432 : int used_bits);
433 : static MinimalTuple hashagg_batch_read(HashAggBatch *batch, uint32 *hashp);
434 : static void hashagg_spill_init(HashAggSpill *spill, LogicalTapeSet *tapeset,
435 : int used_bits, double input_groups,
436 : double hashentrysize);
437 : static Size hashagg_spill_tuple(AggState *aggstate, HashAggSpill *spill,
438 : TupleTableSlot *inputslot, uint32 hash);
439 : static void hashagg_spill_finish(AggState *aggstate, HashAggSpill *spill,
440 : int setno);
441 : static Datum GetAggInitVal(Datum textInitVal, Oid transtype);
442 : static void build_pertrans_for_aggref(AggStatePerTrans pertrans,
443 : AggState *aggstate, EState *estate,
444 : Aggref *aggref, Oid transfn_oid,
445 : Oid aggtranstype, Oid aggserialfn,
446 : Oid aggdeserialfn, Datum initValue,
447 : bool initValueIsNull, Oid *inputTypes,
448 : int numArguments);
449 :
450 :
451 : /*
452 : * Select the current grouping set; affects current_set and
453 : * curaggcontext.
454 : */
455 : static void
456 6263562 : select_current_set(AggState *aggstate, int setno, bool is_hash)
457 : {
458 : /*
459 : * When changing this, also adapt ExecAggPlainTransByVal() and
460 : * ExecAggPlainTransByRef().
461 : */
462 6263562 : if (is_hash)
463 5624176 : aggstate->curaggcontext = aggstate->hashcontext;
464 : else
465 639386 : aggstate->curaggcontext = aggstate->aggcontexts[setno];
466 :
467 6263562 : aggstate->current_set = setno;
468 6263562 : }
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 75694 : 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 75694 : if (aggstate->sort_in)
487 : {
488 42 : tuplesort_end(aggstate->sort_in);
489 42 : aggstate->sort_in = NULL;
490 : }
491 :
492 75694 : if (newphase <= 1)
493 : {
494 : /*
495 : * Discard any existing output tuplesort.
496 : */
497 75508 : 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 : */
509 186 : aggstate->sort_in = aggstate->sort_out;
510 186 : aggstate->sort_out = NULL;
511 : Assert(aggstate->sort_in);
512 186 : 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 75694 : if (newphase > 0 && newphase < aggstate->numphases - 1)
520 : {
521 234 : Sort *sortnode = aggstate->phases[newphase + 1].sortnode;
522 234 : PlanState *outerNode = outerPlanState(aggstate);
523 234 : TupleDesc tupDesc = ExecGetResultType(outerNode);
524 :
525 234 : 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 75694 : aggstate->current_phase = newphase;
536 75694 : aggstate->phase = &aggstate->phases[newphase];
537 75694 : }
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 24895084 : fetch_input_tuple(AggState *aggstate)
549 : {
550 : TupleTableSlot *slot;
551 :
552 24895084 : if (aggstate->sort_in)
553 : {
554 : /* make sure we check for interrupts in either path through here */
555 174882 : CHECK_FOR_INTERRUPTS();
556 174882 : if (!tuplesort_gettupleslot(aggstate->sort_in, true, false,
557 : aggstate->sort_slot, NULL))
558 186 : return NULL;
559 174696 : slot = aggstate->sort_slot;
560 : }
561 : else
562 24720202 : slot = ExecProcNode(outerPlanState(aggstate));
563 :
564 24894880 : if (!TupIsNull(slot) && aggstate->sort_out)
565 174696 : tuplesort_puttupleslot(aggstate->sort_out, slot);
566 :
567 24894880 : 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 1032562 : 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 1032562 : 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 53622 : 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 53622 : if (pertrans->numInputs == 1)
601 : {
602 53550 : Form_pg_attribute attr = TupleDescAttr(pertrans->sortdesc, 0);
603 :
604 53550 : pertrans->sortstates[aggstate->current_set] =
605 53550 : tuplesort_begin_datum(attr->atttypid,
606 53550 : pertrans->sortOperators[0],
607 53550 : pertrans->sortCollations[0],
608 53550 : pertrans->sortNullsFirst[0],
609 : work_mem, NULL, TUPLESORT_NONE);
610 : }
611 : else
612 72 : pertrans->sortstates[aggstate->current_set] =
613 72 : 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 1032562 : if (pertrans->initValueIsNull)
629 522258 : pergroupstate->transValue = pertrans->initValue;
630 : else
631 : {
632 : MemoryContext oldContext;
633 :
634 510304 : oldContext = MemoryContextSwitchTo(aggstate->curaggcontext->ecxt_per_tuple_memory);
635 1020608 : pergroupstate->transValue = datumCopy(pertrans->initValue,
636 510304 : pertrans->transtypeByVal,
637 510304 : pertrans->transtypeLen);
638 510304 : MemoryContextSwitchTo(oldContext);
639 : }
640 1032562 : 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 1032562 : pergroupstate->noTransValue = pertrans->initValueIsNull;
650 1032562 : }
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 288202 : initialize_aggregates(AggState *aggstate,
667 : AggStatePerGroup *pergroups,
668 : int numReset)
669 : {
670 : int transno;
671 288202 : int numGroupingSets = Max(aggstate->phase->numsets, 1);
672 288202 : int setno = 0;
673 288202 : int numTrans = aggstate->numtrans;
674 288202 : AggStatePerTrans transstates = aggstate->pertrans;
675 :
676 288202 : if (numReset == 0)
677 0 : numReset = numGroupingSets;
678 :
679 590560 : for (setno = 0; setno < numReset; setno++)
680 : {
681 302358 : AggStatePerGroup pergroup = pergroups[setno];
682 :
683 302358 : select_current_set(aggstate, setno, false);
684 :
685 956468 : for (transno = 0; transno < numTrans; transno++)
686 : {
687 654110 : AggStatePerTrans pertrans = &transstates[transno];
688 654110 : AggStatePerGroup pergroupstate = &pergroup[transno];
689 :
690 654110 : initialize_aggregate(aggstate, pertrans, pergroupstate);
691 : }
692 : }
693 288202 : }
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 705888 : advance_transition_function(AggState *aggstate,
708 : AggStatePerTrans pertrans,
709 : AggStatePerGroup pergroupstate)
710 : {
711 705888 : FunctionCallInfo fcinfo = pertrans->transfn_fcinfo;
712 : MemoryContext oldContext;
713 : Datum newVal;
714 :
715 705888 : 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 705888 : oldContext = MemoryContextSwitchTo(aggstate->tmpcontext->ecxt_per_tuple_memory);
763 :
764 : /* set up aggstate->curpertrans for AggGetAggref() */
765 705888 : aggstate->curpertrans = pertrans;
766 :
767 : /*
768 : * OK to call the transition function
769 : */
770 705888 : fcinfo->args[0].value = pergroupstate->transValue;
771 705888 : fcinfo->args[0].isnull = pergroupstate->transValueIsNull;
772 705888 : fcinfo->isnull = false; /* just in case transfn doesn't set it */
773 :
774 705888 : newVal = FunctionCallInvoke(fcinfo);
775 :
776 705888 : 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 705888 : 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 705888 : pergroupstate->transValue = newVal;
800 705888 : pergroupstate->transValueIsNull = fcinfo->isnull;
801 :
802 705888 : 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 25045902 : advance_aggregates(AggState *aggstate)
818 : {
819 : bool dummynull;
820 :
821 25045902 : ExecEvalExprSwitchContext(aggstate->phase->evaltrans,
822 : aggstate->tmpcontext,
823 : &dummynull);
824 25045836 : }
825 :
826 : /*
827 : * Run the transition function for a DISTINCT or ORDER BY aggregate
828 : * with only one input. This is called after we have completed
829 : * entering all the input values into the sort object. We complete the
830 : * sort, read out the values in sorted order, and run the transition
831 : * function on each value (applying DISTINCT if appropriate).
832 : *
833 : * Note that the strictness of the transition function was checked when
834 : * entering the values into the sort, so we don't check it again here;
835 : * we just apply standard SQL DISTINCT logic.
836 : *
837 : * The one-input case is handled separately from the multi-input case
838 : * for performance reasons: for single by-value inputs, such as the
839 : * common case of count(distinct id), the tuplesort_getdatum code path
840 : * is around 300% faster. (The speedup for by-reference types is less
841 : * but still noticeable.)
842 : *
843 : * This function handles only one grouping set (already set in
844 : * aggstate->current_set).
845 : *
846 : * When called, CurrentMemoryContext should be the per-query context.
847 : */
848 : static void
849 53550 : process_ordered_aggregate_single(AggState *aggstate,
850 : AggStatePerTrans pertrans,
851 : AggStatePerGroup pergroupstate)
852 : {
853 53550 : Datum oldVal = (Datum) 0;
854 53550 : bool oldIsNull = true;
855 53550 : bool haveOldVal = false;
856 53550 : MemoryContext workcontext = aggstate->tmpcontext->ecxt_per_tuple_memory;
857 : MemoryContext oldContext;
858 53550 : bool isDistinct = (pertrans->numDistinctCols > 0);
859 53550 : Datum newAbbrevVal = (Datum) 0;
860 53550 : Datum oldAbbrevVal = (Datum) 0;
861 53550 : FunctionCallInfo fcinfo = pertrans->transfn_fcinfo;
862 : Datum *newVal;
863 : bool *isNull;
864 :
865 : Assert(pertrans->numDistinctCols < 2);
866 :
867 53550 : tuplesort_performsort(pertrans->sortstates[aggstate->current_set]);
868 :
869 : /* Load the column into argument 1 (arg 0 will be transition value) */
870 53550 : newVal = &fcinfo->args[1].value;
871 53550 : isNull = &fcinfo->args[1].isnull;
872 :
873 : /*
874 : * Note: if input type is pass-by-ref, the datums returned by the sort are
875 : * freshly palloc'd in the per-query context, so we must be careful to
876 : * pfree them when they are no longer needed.
877 : */
878 :
879 877722 : while (tuplesort_getdatum(pertrans->sortstates[aggstate->current_set],
880 : true, false, newVal, isNull, &newAbbrevVal))
881 : {
882 : /*
883 : * Clear and select the working context for evaluation of the equality
884 : * function and transition function.
885 : */
886 824172 : MemoryContextReset(workcontext);
887 824172 : oldContext = MemoryContextSwitchTo(workcontext);
888 :
889 : /*
890 : * If DISTINCT mode, and not distinct from prior, skip it.
891 : */
892 824172 : if (isDistinct &&
893 290316 : haveOldVal &&
894 0 : ((oldIsNull && *isNull) ||
895 290316 : (!oldIsNull && !*isNull &&
896 535752 : oldAbbrevVal == newAbbrevVal &&
897 245436 : DatumGetBool(FunctionCall2Coll(&pertrans->equalfnOne,
898 : pertrans->aggCollation,
899 : oldVal, *newVal)))))
900 : {
901 118464 : MemoryContextSwitchTo(oldContext);
902 118464 : continue;
903 : }
904 : else
905 : {
906 705708 : advance_transition_function(aggstate, pertrans, pergroupstate);
907 :
908 705708 : MemoryContextSwitchTo(oldContext);
909 :
910 : /*
911 : * Forget the old value, if any, and remember the new one for
912 : * subsequent equality checks.
913 : */
914 705708 : if (!pertrans->inputtypeByVal)
915 : {
916 525288 : if (!oldIsNull)
917 525108 : pfree(DatumGetPointer(oldVal));
918 525288 : if (!*isNull)
919 525228 : oldVal = datumCopy(*newVal, pertrans->inputtypeByVal,
920 525228 : pertrans->inputtypeLen);
921 : }
922 : else
923 180420 : oldVal = *newVal;
924 705708 : oldAbbrevVal = newAbbrevVal;
925 705708 : oldIsNull = *isNull;
926 705708 : haveOldVal = true;
927 : }
928 : }
929 :
930 53550 : if (!oldIsNull && !pertrans->inputtypeByVal)
931 120 : pfree(DatumGetPointer(oldVal));
932 :
933 53550 : tuplesort_end(pertrans->sortstates[aggstate->current_set]);
934 53550 : pertrans->sortstates[aggstate->current_set] = NULL;
935 53550 : }
936 :
937 : /*
938 : * Run the transition function for a DISTINCT or ORDER BY aggregate
939 : * with more than one input. This is called after we have completed
940 : * entering all the input values into the sort object. We complete the
941 : * sort, read out the values in sorted order, and run the transition
942 : * function on each value (applying DISTINCT if appropriate).
943 : *
944 : * This function handles only one grouping set (already set in
945 : * aggstate->current_set).
946 : *
947 : * When called, CurrentMemoryContext should be the per-query context.
948 : */
949 : static void
950 72 : process_ordered_aggregate_multi(AggState *aggstate,
951 : AggStatePerTrans pertrans,
952 : AggStatePerGroup pergroupstate)
953 : {
954 72 : ExprContext *tmpcontext = aggstate->tmpcontext;
955 72 : FunctionCallInfo fcinfo = pertrans->transfn_fcinfo;
956 72 : TupleTableSlot *slot1 = pertrans->sortslot;
957 72 : TupleTableSlot *slot2 = pertrans->uniqslot;
958 72 : int numTransInputs = pertrans->numTransInputs;
959 72 : int numDistinctCols = pertrans->numDistinctCols;
960 72 : Datum newAbbrevVal = (Datum) 0;
961 72 : Datum oldAbbrevVal = (Datum) 0;
962 72 : bool haveOldValue = false;
963 72 : TupleTableSlot *save = aggstate->tmpcontext->ecxt_outertuple;
964 : int i;
965 :
966 72 : tuplesort_performsort(pertrans->sortstates[aggstate->current_set]);
967 :
968 72 : ExecClearTuple(slot1);
969 72 : if (slot2)
970 0 : ExecClearTuple(slot2);
971 :
972 252 : while (tuplesort_gettupleslot(pertrans->sortstates[aggstate->current_set],
973 : true, true, slot1, &newAbbrevVal))
974 : {
975 180 : CHECK_FOR_INTERRUPTS();
976 :
977 180 : tmpcontext->ecxt_outertuple = slot1;
978 180 : tmpcontext->ecxt_innertuple = slot2;
979 :
980 180 : if (numDistinctCols == 0 ||
981 0 : !haveOldValue ||
982 0 : newAbbrevVal != oldAbbrevVal ||
983 0 : !ExecQual(pertrans->equalfnMulti, tmpcontext))
984 : {
985 : /*
986 : * Extract the first numTransInputs columns as datums to pass to
987 : * the transfn.
988 : */
989 180 : slot_getsomeattrs(slot1, numTransInputs);
990 :
991 : /* Load values into fcinfo */
992 : /* Start from 1, since the 0th arg will be the transition value */
993 540 : for (i = 0; i < numTransInputs; i++)
994 : {
995 360 : fcinfo->args[i + 1].value = slot1->tts_values[i];
996 360 : fcinfo->args[i + 1].isnull = slot1->tts_isnull[i];
997 : }
998 :
999 180 : advance_transition_function(aggstate, pertrans, pergroupstate);
1000 :
1001 180 : if (numDistinctCols > 0)
1002 : {
1003 : /* swap the slot pointers to retain the current tuple */
1004 0 : TupleTableSlot *tmpslot = slot2;
1005 :
1006 0 : slot2 = slot1;
1007 0 : slot1 = tmpslot;
1008 : /* avoid ExecQual() calls by reusing abbreviated keys */
1009 0 : oldAbbrevVal = newAbbrevVal;
1010 0 : haveOldValue = true;
1011 : }
1012 : }
1013 :
1014 : /* Reset context each time */
1015 180 : ResetExprContext(tmpcontext);
1016 :
1017 180 : ExecClearTuple(slot1);
1018 : }
1019 :
1020 72 : if (slot2)
1021 0 : ExecClearTuple(slot2);
1022 :
1023 72 : tuplesort_end(pertrans->sortstates[aggstate->current_set]);
1024 72 : pertrans->sortstates[aggstate->current_set] = NULL;
1025 :
1026 : /* restore previous slot, potentially in use for grouping sets */
1027 72 : tmpcontext->ecxt_outertuple = save;
1028 72 : }
1029 :
1030 : /*
1031 : * Compute the final value of one aggregate.
1032 : *
1033 : * This function handles only one grouping set (already set in
1034 : * aggstate->current_set).
1035 : *
1036 : * The finalfn will be run, and the result delivered, in the
1037 : * output-tuple context; caller's CurrentMemoryContext does not matter.
1038 : * (But note that in some cases, such as when there is no finalfn, the
1039 : * result might be a pointer to or into the agg's transition value.)
1040 : *
1041 : * The finalfn uses the state as set in the transno. This also might be
1042 : * being used by another aggregate function, so it's important that we do
1043 : * nothing destructive here. Moreover, the aggregate's final value might
1044 : * get used in multiple places, so we mustn't return a R/W expanded datum.
1045 : */
1046 : static void
1047 1024600 : finalize_aggregate(AggState *aggstate,
1048 : AggStatePerAgg peragg,
1049 : AggStatePerGroup pergroupstate,
1050 : Datum *resultVal, bool *resultIsNull)
1051 : {
1052 1024600 : LOCAL_FCINFO(fcinfo, FUNC_MAX_ARGS);
1053 1024600 : bool anynull = false;
1054 : MemoryContext oldContext;
1055 : int i;
1056 : ListCell *lc;
1057 1024600 : AggStatePerTrans pertrans = &aggstate->pertrans[peragg->transno];
1058 :
1059 1024600 : oldContext = MemoryContextSwitchTo(aggstate->ss.ps.ps_ExprContext->ecxt_per_tuple_memory);
1060 :
1061 : /*
1062 : * Evaluate any direct arguments. We do this even if there's no finalfn
1063 : * (which is unlikely anyway), so that side-effects happen as expected.
1064 : * The direct arguments go into arg positions 1 and up, leaving position 0
1065 : * for the transition state value.
1066 : */
1067 1024600 : i = 1;
1068 1025574 : foreach(lc, peragg->aggdirectargs)
1069 : {
1070 974 : ExprState *expr = (ExprState *) lfirst(lc);
1071 :
1072 974 : fcinfo->args[i].value = ExecEvalExpr(expr,
1073 : aggstate->ss.ps.ps_ExprContext,
1074 : &fcinfo->args[i].isnull);
1075 974 : anynull |= fcinfo->args[i].isnull;
1076 974 : i++;
1077 : }
1078 :
1079 : /*
1080 : * Apply the agg's finalfn if one is provided, else return transValue.
1081 : */
1082 1024600 : if (OidIsValid(peragg->finalfn_oid))
1083 : {
1084 267198 : int numFinalArgs = peragg->numFinalArgs;
1085 :
1086 : /* set up aggstate->curperagg for AggGetAggref() */
1087 267198 : aggstate->curperagg = peragg;
1088 :
1089 267198 : InitFunctionCallInfoData(*fcinfo, &peragg->finalfn,
1090 : numFinalArgs,
1091 : pertrans->aggCollation,
1092 : (void *) aggstate, NULL);
1093 :
1094 : /* Fill in the transition state value */
1095 267198 : fcinfo->args[0].value =
1096 267198 : MakeExpandedObjectReadOnly(pergroupstate->transValue,
1097 : pergroupstate->transValueIsNull,
1098 : pertrans->transtypeLen);
1099 267198 : fcinfo->args[0].isnull = pergroupstate->transValueIsNull;
1100 267198 : anynull |= pergroupstate->transValueIsNull;
1101 :
1102 : /* Fill any remaining argument positions with nulls */
1103 349198 : for (; i < numFinalArgs; i++)
1104 : {
1105 82000 : fcinfo->args[i].value = (Datum) 0;
1106 82000 : fcinfo->args[i].isnull = true;
1107 82000 : anynull = true;
1108 : }
1109 :
1110 267198 : if (fcinfo->flinfo->fn_strict && anynull)
1111 : {
1112 : /* don't call a strict function with NULL inputs */
1113 0 : *resultVal = (Datum) 0;
1114 0 : *resultIsNull = true;
1115 : }
1116 : else
1117 : {
1118 : Datum result;
1119 :
1120 267198 : result = FunctionCallInvoke(fcinfo);
1121 267186 : *resultIsNull = fcinfo->isnull;
1122 267186 : *resultVal = MakeExpandedObjectReadOnly(result,
1123 : fcinfo->isnull,
1124 : peragg->resulttypeLen);
1125 : }
1126 267186 : aggstate->curperagg = NULL;
1127 : }
1128 : else
1129 : {
1130 757402 : *resultVal =
1131 757402 : MakeExpandedObjectReadOnly(pergroupstate->transValue,
1132 : pergroupstate->transValueIsNull,
1133 : pertrans->transtypeLen);
1134 757402 : *resultIsNull = pergroupstate->transValueIsNull;
1135 : }
1136 :
1137 1024588 : MemoryContextSwitchTo(oldContext);
1138 1024588 : }
1139 :
1140 : /*
1141 : * Compute the output value of one partial aggregate.
1142 : *
1143 : * The serialization function will be run, and the result delivered, in the
1144 : * output-tuple context; caller's CurrentMemoryContext does not matter.
1145 : */
1146 : static void
1147 11746 : finalize_partialaggregate(AggState *aggstate,
1148 : AggStatePerAgg peragg,
1149 : AggStatePerGroup pergroupstate,
1150 : Datum *resultVal, bool *resultIsNull)
1151 : {
1152 11746 : AggStatePerTrans pertrans = &aggstate->pertrans[peragg->transno];
1153 : MemoryContext oldContext;
1154 :
1155 11746 : oldContext = MemoryContextSwitchTo(aggstate->ss.ps.ps_ExprContext->ecxt_per_tuple_memory);
1156 :
1157 : /*
1158 : * serialfn_oid will be set if we must serialize the transvalue before
1159 : * returning it
1160 : */
1161 11746 : if (OidIsValid(pertrans->serialfn_oid))
1162 : {
1163 : /* Don't call a strict serialization function with NULL input. */
1164 586 : if (pertrans->serialfn.fn_strict && pergroupstate->transValueIsNull)
1165 : {
1166 124 : *resultVal = (Datum) 0;
1167 124 : *resultIsNull = true;
1168 : }
1169 : else
1170 : {
1171 462 : FunctionCallInfo fcinfo = pertrans->serialfn_fcinfo;
1172 : Datum result;
1173 :
1174 462 : fcinfo->args[0].value =
1175 462 : MakeExpandedObjectReadOnly(pergroupstate->transValue,
1176 : pergroupstate->transValueIsNull,
1177 : pertrans->transtypeLen);
1178 462 : fcinfo->args[0].isnull = pergroupstate->transValueIsNull;
1179 462 : fcinfo->isnull = false;
1180 :
1181 462 : result = FunctionCallInvoke(fcinfo);
1182 462 : *resultIsNull = fcinfo->isnull;
1183 462 : *resultVal = MakeExpandedObjectReadOnly(result,
1184 : fcinfo->isnull,
1185 : peragg->resulttypeLen);
1186 : }
1187 : }
1188 : else
1189 : {
1190 11160 : *resultVal =
1191 11160 : MakeExpandedObjectReadOnly(pergroupstate->transValue,
1192 : pergroupstate->transValueIsNull,
1193 : pertrans->transtypeLen);
1194 11160 : *resultIsNull = pergroupstate->transValueIsNull;
1195 : }
1196 :
1197 11746 : MemoryContextSwitchTo(oldContext);
1198 11746 : }
1199 :
1200 : /*
1201 : * Extract the attributes that make up the grouping key into the
1202 : * hashslot. This is necessary to compute the hash or perform a lookup.
1203 : */
1204 : static inline void
1205 6103762 : prepare_hash_slot(AggStatePerHash perhash,
1206 : TupleTableSlot *inputslot,
1207 : TupleTableSlot *hashslot)
1208 : {
1209 : int i;
1210 :
1211 : /* transfer just the needed columns into hashslot */
1212 6103762 : slot_getsomeattrs(inputslot, perhash->largestGrpColIdx);
1213 6103762 : ExecClearTuple(hashslot);
1214 :
1215 15182796 : for (i = 0; i < perhash->numhashGrpCols; i++)
1216 : {
1217 9079034 : int varNumber = perhash->hashGrpColIdxInput[i] - 1;
1218 :
1219 9079034 : hashslot->tts_values[i] = inputslot->tts_values[varNumber];
1220 9079034 : hashslot->tts_isnull[i] = inputslot->tts_isnull[varNumber];
1221 : }
1222 6103762 : ExecStoreVirtualTuple(hashslot);
1223 6103762 : }
1224 :
1225 : /*
1226 : * Prepare to finalize and project based on the specified representative tuple
1227 : * slot and grouping set.
1228 : *
1229 : * In the specified tuple slot, force to null all attributes that should be
1230 : * read as null in the context of the current grouping set. Also stash the
1231 : * current group bitmap where GroupingExpr can get at it.
1232 : *
1233 : * This relies on three conditions:
1234 : *
1235 : * 1) Nothing is ever going to try and extract the whole tuple from this slot,
1236 : * only reference it in evaluations, which will only access individual
1237 : * attributes.
1238 : *
1239 : * 2) No system columns are going to need to be nulled. (If a system column is
1240 : * referenced in a group clause, it is actually projected in the outer plan
1241 : * tlist.)
1242 : *
1243 : * 3) Within a given phase, we never need to recover the value of an attribute
1244 : * once it has been set to null.
1245 : *
1246 : * Poking into the slot this way is a bit ugly, but the consensus is that the
1247 : * alternative was worse.
1248 : */
1249 : static void
1250 711706 : prepare_projection_slot(AggState *aggstate, TupleTableSlot *slot, int currentSet)
1251 : {
1252 711706 : if (aggstate->phase->grouped_cols)
1253 : {
1254 438724 : Bitmapset *grouped_cols = aggstate->phase->grouped_cols[currentSet];
1255 :
1256 438724 : aggstate->grouped_cols = grouped_cols;
1257 :
1258 438724 : if (TTS_EMPTY(slot))
1259 : {
1260 : /*
1261 : * Force all values to be NULL if working on an empty input tuple
1262 : * (i.e. an empty grouping set for which no input rows were
1263 : * supplied).
1264 : */
1265 48 : ExecStoreAllNullTuple(slot);
1266 : }
1267 438676 : else if (aggstate->all_grouped_cols)
1268 : {
1269 : ListCell *lc;
1270 :
1271 : /* all_grouped_cols is arranged in desc order */
1272 438622 : slot_getsomeattrs(slot, linitial_int(aggstate->all_grouped_cols));
1273 :
1274 1308696 : foreach(lc, aggstate->all_grouped_cols)
1275 : {
1276 870074 : int attnum = lfirst_int(lc);
1277 :
1278 870074 : if (!bms_is_member(attnum, grouped_cols))
1279 57346 : slot->tts_isnull[attnum - 1] = true;
1280 : }
1281 : }
1282 : }
1283 711706 : }
1284 :
1285 : /*
1286 : * Compute the final value of all aggregates for one group.
1287 : *
1288 : * This function handles only one grouping set at a time, which the caller must
1289 : * have selected. It's also the caller's responsibility to adjust the supplied
1290 : * pergroup parameter to point to the current set's transvalues.
1291 : *
1292 : * Results are stored in the output econtext aggvalues/aggnulls.
1293 : */
1294 : static void
1295 711706 : finalize_aggregates(AggState *aggstate,
1296 : AggStatePerAgg peraggs,
1297 : AggStatePerGroup pergroup)
1298 : {
1299 711706 : ExprContext *econtext = aggstate->ss.ps.ps_ExprContext;
1300 711706 : Datum *aggvalues = econtext->ecxt_aggvalues;
1301 711706 : bool *aggnulls = econtext->ecxt_aggnulls;
1302 : int aggno;
1303 :
1304 : /*
1305 : * If there were any DISTINCT and/or ORDER BY aggregates, sort their
1306 : * inputs and run the transition functions.
1307 : */
1308 1747794 : for (int transno = 0; transno < aggstate->numtrans; transno++)
1309 : {
1310 1036088 : AggStatePerTrans pertrans = &aggstate->pertrans[transno];
1311 : AggStatePerGroup pergroupstate;
1312 :
1313 1036088 : pergroupstate = &pergroup[transno];
1314 :
1315 1036088 : if (pertrans->aggsortrequired)
1316 : {
1317 : Assert(aggstate->aggstrategy != AGG_HASHED &&
1318 : aggstate->aggstrategy != AGG_MIXED);
1319 :
1320 53622 : if (pertrans->numInputs == 1)
1321 53550 : process_ordered_aggregate_single(aggstate,
1322 : pertrans,
1323 : pergroupstate);
1324 : else
1325 72 : process_ordered_aggregate_multi(aggstate,
1326 : pertrans,
1327 : pergroupstate);
1328 : }
1329 982466 : else if (pertrans->numDistinctCols > 0 && pertrans->haslast)
1330 : {
1331 18330 : pertrans->haslast = false;
1332 :
1333 18330 : if (pertrans->numDistinctCols == 1)
1334 : {
1335 18240 : if (!pertrans->inputtypeByVal && !pertrans->lastisnull)
1336 238 : pfree(DatumGetPointer(pertrans->lastdatum));
1337 :
1338 18240 : pertrans->lastisnull = false;
1339 18240 : pertrans->lastdatum = (Datum) 0;
1340 : }
1341 : else
1342 90 : ExecClearTuple(pertrans->uniqslot);
1343 : }
1344 : }
1345 :
1346 : /*
1347 : * Run the final functions.
1348 : */
1349 1748040 : for (aggno = 0; aggno < aggstate->numaggs; aggno++)
1350 : {
1351 1036346 : AggStatePerAgg peragg = &peraggs[aggno];
1352 1036346 : int transno = peragg->transno;
1353 : AggStatePerGroup pergroupstate;
1354 :
1355 1036346 : pergroupstate = &pergroup[transno];
1356 :
1357 1036346 : if (DO_AGGSPLIT_SKIPFINAL(aggstate->aggsplit))
1358 11746 : finalize_partialaggregate(aggstate, peragg, pergroupstate,
1359 11746 : &aggvalues[aggno], &aggnulls[aggno]);
1360 : else
1361 1024600 : finalize_aggregate(aggstate, peragg, pergroupstate,
1362 1024600 : &aggvalues[aggno], &aggnulls[aggno]);
1363 : }
1364 711694 : }
1365 :
1366 : /*
1367 : * Project the result of a group (whose aggs have already been calculated by
1368 : * finalize_aggregates). Returns the result slot, or NULL if no row is
1369 : * projected (suppressed by qual).
1370 : */
1371 : static TupleTableSlot *
1372 711694 : project_aggregates(AggState *aggstate)
1373 : {
1374 711694 : ExprContext *econtext = aggstate->ss.ps.ps_ExprContext;
1375 :
1376 : /*
1377 : * Check the qual (HAVING clause); if the group does not match, ignore it.
1378 : */
1379 711694 : if (ExecQual(aggstate->ss.ps.qual, econtext))
1380 : {
1381 : /*
1382 : * Form and return projection tuple using the aggregate results and
1383 : * the representative input tuple.
1384 : */
1385 623396 : return ExecProject(aggstate->ss.ps.ps_ProjInfo);
1386 : }
1387 : else
1388 88298 : InstrCountFiltered1(aggstate, 1);
1389 :
1390 88298 : return NULL;
1391 : }
1392 :
1393 : /*
1394 : * Find input-tuple columns that are needed, dividing them into
1395 : * aggregated and unaggregated sets.
1396 : */
1397 : static void
1398 7564 : find_cols(AggState *aggstate, Bitmapset **aggregated, Bitmapset **unaggregated)
1399 : {
1400 7564 : Agg *agg = (Agg *) aggstate->ss.ps.plan;
1401 : FindColsContext context;
1402 :
1403 7564 : context.is_aggref = false;
1404 7564 : context.aggregated = NULL;
1405 7564 : context.unaggregated = NULL;
1406 :
1407 : /* Examine tlist and quals */
1408 7564 : (void) find_cols_walker((Node *) agg->plan.targetlist, &context);
1409 7564 : (void) find_cols_walker((Node *) agg->plan.qual, &context);
1410 :
1411 : /* In some cases, grouping columns will not appear in the tlist */
1412 23032 : for (int i = 0; i < agg->numCols; i++)
1413 15468 : context.unaggregated = bms_add_member(context.unaggregated,
1414 15468 : agg->grpColIdx[i]);
1415 :
1416 7564 : *aggregated = context.aggregated;
1417 7564 : *unaggregated = context.unaggregated;
1418 7564 : }
1419 :
1420 : static bool
1421 82014 : find_cols_walker(Node *node, FindColsContext *context)
1422 : {
1423 82014 : if (node == NULL)
1424 13572 : return false;
1425 68442 : if (IsA(node, Var))
1426 : {
1427 20912 : Var *var = (Var *) node;
1428 :
1429 : /* setrefs.c should have set the varno to OUTER_VAR */
1430 : Assert(var->varno == OUTER_VAR);
1431 : Assert(var->varlevelsup == 0);
1432 20912 : if (context->is_aggref)
1433 4368 : context->aggregated = bms_add_member(context->aggregated,
1434 4368 : var->varattno);
1435 : else
1436 16544 : context->unaggregated = bms_add_member(context->unaggregated,
1437 16544 : var->varattno);
1438 20912 : return false;
1439 : }
1440 47530 : if (IsA(node, Aggref))
1441 : {
1442 : Assert(!context->is_aggref);
1443 6502 : context->is_aggref = true;
1444 6502 : expression_tree_walker(node, find_cols_walker, (void *) context);
1445 6502 : context->is_aggref = false;
1446 6502 : return false;
1447 : }
1448 41028 : return expression_tree_walker(node, find_cols_walker,
1449 : (void *) context);
1450 : }
1451 :
1452 : /*
1453 : * (Re-)initialize the hash table(s) to empty.
1454 : *
1455 : * To implement hashed aggregation, we need a hashtable that stores a
1456 : * representative tuple and an array of AggStatePerGroup structs for each
1457 : * distinct set of GROUP BY column values. We compute the hash key from the
1458 : * GROUP BY columns. The per-group data is allocated in lookup_hash_entry(),
1459 : * for each entry.
1460 : *
1461 : * We have a separate hashtable and associated perhash data structure for each
1462 : * grouping set for which we're doing hashing.
1463 : *
1464 : * The contents of the hash tables always live in the hashcontext's per-tuple
1465 : * memory context (there is only one of these for all tables together, since
1466 : * they are all reset at the same time).
1467 : */
1468 : static void
1469 14162 : build_hash_tables(AggState *aggstate)
1470 : {
1471 : int setno;
1472 :
1473 28608 : for (setno = 0; setno < aggstate->num_hashes; ++setno)
1474 : {
1475 14446 : AggStatePerHash perhash = &aggstate->perhash[setno];
1476 : long nbuckets;
1477 : Size memory;
1478 :
1479 14446 : if (perhash->hashtable != NULL)
1480 : {
1481 7758 : ResetTupleHashTable(perhash->hashtable);
1482 7758 : continue;
1483 : }
1484 :
1485 : Assert(perhash->aggnode->numGroups > 0);
1486 :
1487 6688 : memory = aggstate->hash_mem_limit / aggstate->num_hashes;
1488 :
1489 : /* choose reasonable number of buckets per hashtable */
1490 6688 : nbuckets = hash_choose_num_buckets(aggstate->hashentrysize,
1491 6688 : perhash->aggnode->numGroups,
1492 : memory);
1493 :
1494 6688 : build_hash_table(aggstate, setno, nbuckets);
1495 : }
1496 :
1497 14162 : aggstate->hash_ngroups_current = 0;
1498 14162 : }
1499 :
1500 : /*
1501 : * Build a single hashtable for this grouping set.
1502 : */
1503 : static void
1504 6688 : build_hash_table(AggState *aggstate, int setno, long nbuckets)
1505 : {
1506 6688 : AggStatePerHash perhash = &aggstate->perhash[setno];
1507 6688 : MemoryContext metacxt = aggstate->hash_metacxt;
1508 6688 : MemoryContext hashcxt = aggstate->hashcontext->ecxt_per_tuple_memory;
1509 6688 : MemoryContext tmpcxt = aggstate->tmpcontext->ecxt_per_tuple_memory;
1510 : Size additionalsize;
1511 :
1512 : Assert(aggstate->aggstrategy == AGG_HASHED ||
1513 : aggstate->aggstrategy == AGG_MIXED);
1514 :
1515 : /*
1516 : * Used to make sure initial hash table allocation does not exceed
1517 : * hash_mem. Note that the estimate does not include space for
1518 : * pass-by-reference transition data values, nor for the representative
1519 : * tuple of each group.
1520 : */
1521 6688 : additionalsize = aggstate->numtrans * sizeof(AggStatePerGroupData);
1522 :
1523 13376 : perhash->hashtable = BuildTupleHashTableExt(&aggstate->ss.ps,
1524 6688 : perhash->hashslot->tts_tupleDescriptor,
1525 : perhash->numCols,
1526 : perhash->hashGrpColIdxHash,
1527 6688 : perhash->eqfuncoids,
1528 : perhash->hashfunctions,
1529 6688 : perhash->aggnode->grpCollations,
1530 : nbuckets,
1531 : additionalsize,
1532 : metacxt,
1533 : hashcxt,
1534 : tmpcxt,
1535 6688 : DO_AGGSPLIT_SKIPFINAL(aggstate->aggsplit));
1536 6688 : }
1537 :
1538 : /*
1539 : * Compute columns that actually need to be stored in hashtable entries. The
1540 : * incoming tuples from the child plan node will contain grouping columns,
1541 : * other columns referenced in our targetlist and qual, columns used to
1542 : * compute the aggregate functions, and perhaps just junk columns we don't use
1543 : * at all. Only columns of the first two types need to be stored in the
1544 : * hashtable, and getting rid of the others can make the table entries
1545 : * significantly smaller. The hashtable only contains the relevant columns,
1546 : * and is packed/unpacked in lookup_hash_entry() / agg_retrieve_hash_table()
1547 : * into the format of the normal input descriptor.
1548 : *
1549 : * Additional columns, in addition to the columns grouped by, come from two
1550 : * sources: Firstly functionally dependent columns that we don't need to group
1551 : * by themselves, and secondly ctids for row-marks.
1552 : *
1553 : * To eliminate duplicates, we build a bitmapset of the needed columns, and
1554 : * then build an array of the columns included in the hashtable. We might
1555 : * still have duplicates if the passed-in grpColIdx has them, which can happen
1556 : * in edge cases from semijoins/distinct; these can't always be removed,
1557 : * because it's not certain that the duplicate cols will be using the same
1558 : * hash function.
1559 : *
1560 : * Note that the array is preserved over ExecReScanAgg, so we allocate it in
1561 : * the per-query context (unlike the hash table itself).
1562 : */
1563 : static void
1564 7564 : find_hash_columns(AggState *aggstate)
1565 : {
1566 : Bitmapset *base_colnos;
1567 : Bitmapset *aggregated_colnos;
1568 7564 : TupleDesc scanDesc = aggstate->ss.ss_ScanTupleSlot->tts_tupleDescriptor;
1569 7564 : List *outerTlist = outerPlanState(aggstate)->plan->targetlist;
1570 7564 : int numHashes = aggstate->num_hashes;
1571 7564 : EState *estate = aggstate->ss.ps.state;
1572 : int j;
1573 :
1574 : /* Find Vars that will be needed in tlist and qual */
1575 7564 : find_cols(aggstate, &aggregated_colnos, &base_colnos);
1576 7564 : aggstate->colnos_needed = bms_union(base_colnos, aggregated_colnos);
1577 7564 : aggstate->max_colno_needed = 0;
1578 7564 : aggstate->all_cols_needed = true;
1579 :
1580 31410 : for (int i = 0; i < scanDesc->natts; i++)
1581 : {
1582 23846 : int colno = i + 1;
1583 :
1584 23846 : if (bms_is_member(colno, aggstate->colnos_needed))
1585 18672 : aggstate->max_colno_needed = colno;
1586 : else
1587 5174 : aggstate->all_cols_needed = false;
1588 : }
1589 :
1590 15522 : for (j = 0; j < numHashes; ++j)
1591 : {
1592 7958 : AggStatePerHash perhash = &aggstate->perhash[j];
1593 7958 : Bitmapset *colnos = bms_copy(base_colnos);
1594 7958 : AttrNumber *grpColIdx = perhash->aggnode->grpColIdx;
1595 7958 : List *hashTlist = NIL;
1596 : TupleDesc hashDesc;
1597 : int maxCols;
1598 : int i;
1599 :
1600 7958 : perhash->largestGrpColIdx = 0;
1601 :
1602 : /*
1603 : * If we're doing grouping sets, then some Vars might be referenced in
1604 : * tlist/qual for the benefit of other grouping sets, but not needed
1605 : * when hashing; i.e. prepare_projection_slot will null them out, so
1606 : * there'd be no point storing them. Use prepare_projection_slot's
1607 : * logic to determine which.
1608 : */
1609 7958 : if (aggstate->phases[0].grouped_cols)
1610 : {
1611 7958 : Bitmapset *grouped_cols = aggstate->phases[0].grouped_cols[j];
1612 : ListCell *lc;
1613 :
1614 24930 : foreach(lc, aggstate->all_grouped_cols)
1615 : {
1616 16972 : int attnum = lfirst_int(lc);
1617 :
1618 16972 : if (!bms_is_member(attnum, grouped_cols))
1619 1008 : colnos = bms_del_member(colnos, attnum);
1620 : }
1621 : }
1622 :
1623 : /*
1624 : * Compute maximum number of input columns accounting for possible
1625 : * duplications in the grpColIdx array, which can happen in some edge
1626 : * cases where HashAggregate was generated as part of a semijoin or a
1627 : * DISTINCT.
1628 : */
1629 7958 : maxCols = bms_num_members(colnos) + perhash->numCols;
1630 :
1631 7958 : perhash->hashGrpColIdxInput =
1632 7958 : palloc(maxCols * sizeof(AttrNumber));
1633 7958 : perhash->hashGrpColIdxHash =
1634 7958 : palloc(perhash->numCols * sizeof(AttrNumber));
1635 :
1636 : /* Add all the grouping columns to colnos */
1637 23928 : for (i = 0; i < perhash->numCols; i++)
1638 15970 : colnos = bms_add_member(colnos, grpColIdx[i]);
1639 :
1640 : /*
1641 : * First build mapping for columns directly hashed. These are the
1642 : * first, because they'll be accessed when computing hash values and
1643 : * comparing tuples for exact matches. We also build simple mapping
1644 : * for execGrouping, so it knows where to find the to-be-hashed /
1645 : * compared columns in the input.
1646 : */
1647 23928 : for (i = 0; i < perhash->numCols; i++)
1648 : {
1649 15970 : perhash->hashGrpColIdxInput[i] = grpColIdx[i];
1650 15970 : perhash->hashGrpColIdxHash[i] = i + 1;
1651 15970 : perhash->numhashGrpCols++;
1652 : /* delete already mapped columns */
1653 15970 : colnos = bms_del_member(colnos, grpColIdx[i]);
1654 : }
1655 :
1656 : /* and add the remaining columns */
1657 7958 : i = -1;
1658 8586 : while ((i = bms_next_member(colnos, i)) >= 0)
1659 : {
1660 628 : perhash->hashGrpColIdxInput[perhash->numhashGrpCols] = i;
1661 628 : perhash->numhashGrpCols++;
1662 : }
1663 :
1664 : /* and build a tuple descriptor for the hashtable */
1665 24556 : for (i = 0; i < perhash->numhashGrpCols; i++)
1666 : {
1667 16598 : int varNumber = perhash->hashGrpColIdxInput[i] - 1;
1668 :
1669 16598 : hashTlist = lappend(hashTlist, list_nth(outerTlist, varNumber));
1670 16598 : perhash->largestGrpColIdx =
1671 16598 : Max(varNumber + 1, perhash->largestGrpColIdx);
1672 : }
1673 :
1674 7958 : hashDesc = ExecTypeFromTL(hashTlist);
1675 :
1676 7958 : execTuplesHashPrepare(perhash->numCols,
1677 7958 : perhash->aggnode->grpOperators,
1678 : &perhash->eqfuncoids,
1679 : &perhash->hashfunctions);
1680 7958 : perhash->hashslot =
1681 7958 : ExecAllocTableSlot(&estate->es_tupleTable, hashDesc,
1682 : &TTSOpsMinimalTuple);
1683 :
1684 7958 : list_free(hashTlist);
1685 7958 : bms_free(colnos);
1686 : }
1687 :
1688 7564 : bms_free(base_colnos);
1689 7564 : }
1690 :
1691 : /*
1692 : * Estimate per-hash-table-entry overhead.
1693 : */
1694 : Size
1695 29388 : hash_agg_entry_size(int numTrans, Size tupleWidth, Size transitionSpace)
1696 : {
1697 : Size tupleChunkSize;
1698 : Size pergroupChunkSize;
1699 : Size transitionChunkSize;
1700 29388 : Size tupleSize = (MAXALIGN(SizeofMinimalTupleHeader) +
1701 : tupleWidth);
1702 29388 : Size pergroupSize = numTrans * sizeof(AggStatePerGroupData);
1703 :
1704 29388 : tupleChunkSize = CHUNKHDRSZ + tupleSize;
1705 :
1706 29388 : if (pergroupSize > 0)
1707 10224 : pergroupChunkSize = CHUNKHDRSZ + pergroupSize;
1708 : else
1709 19164 : pergroupChunkSize = 0;
1710 :
1711 29388 : if (transitionSpace > 0)
1712 4728 : transitionChunkSize = CHUNKHDRSZ + transitionSpace;
1713 : else
1714 24660 : transitionChunkSize = 0;
1715 :
1716 : return
1717 : sizeof(TupleHashEntryData) +
1718 29388 : tupleChunkSize +
1719 29388 : pergroupChunkSize +
1720 : transitionChunkSize;
1721 : }
1722 :
1723 : /*
1724 : * hashagg_recompile_expressions()
1725 : *
1726 : * Identifies the right phase, compiles the right expression given the
1727 : * arguments, and then sets phase->evalfunc to that expression.
1728 : *
1729 : * Different versions of the compiled expression are needed depending on
1730 : * whether hash aggregation has spilled or not, and whether it's reading from
1731 : * the outer plan or a tape. Before spilling to disk, the expression reads
1732 : * from the outer plan and does not need to perform a NULL check. After
1733 : * HashAgg begins to spill, new groups will not be created in the hash table,
1734 : * and the AggStatePerGroup array may be NULL; therefore we need to add a null
1735 : * pointer check to the expression. Then, when reading spilled data from a
1736 : * tape, we change the outer slot type to be a fixed minimal tuple slot.
1737 : *
1738 : * It would be wasteful to recompile every time, so cache the compiled
1739 : * expressions in the AggStatePerPhase, and reuse when appropriate.
1740 : */
1741 : static void
1742 61050 : hashagg_recompile_expressions(AggState *aggstate, bool minslot, bool nullcheck)
1743 : {
1744 : AggStatePerPhase phase;
1745 61050 : int i = minslot ? 1 : 0;
1746 61050 : int j = nullcheck ? 1 : 0;
1747 :
1748 : Assert(aggstate->aggstrategy == AGG_HASHED ||
1749 : aggstate->aggstrategy == AGG_MIXED);
1750 :
1751 61050 : if (aggstate->aggstrategy == AGG_HASHED)
1752 8478 : phase = &aggstate->phases[0];
1753 : else /* AGG_MIXED */
1754 52572 : phase = &aggstate->phases[1];
1755 :
1756 61050 : if (phase->evaltrans_cache[i][j] == NULL)
1757 : {
1758 72 : const TupleTableSlotOps *outerops = aggstate->ss.ps.outerops;
1759 72 : bool outerfixed = aggstate->ss.ps.outeropsfixed;
1760 72 : bool dohash = true;
1761 72 : bool dosort = false;
1762 :
1763 : /*
1764 : * If minslot is true, that means we are processing a spilled batch
1765 : * (inside agg_refill_hash_table()), and we must not advance the
1766 : * sorted grouping sets.
1767 : */
1768 72 : if (aggstate->aggstrategy == AGG_MIXED && !minslot)
1769 12 : dosort = true;
1770 :
1771 : /* temporarily change the outerops while compiling the expression */
1772 72 : if (minslot)
1773 : {
1774 36 : aggstate->ss.ps.outerops = &TTSOpsMinimalTuple;
1775 36 : aggstate->ss.ps.outeropsfixed = true;
1776 : }
1777 :
1778 72 : phase->evaltrans_cache[i][j] = ExecBuildAggTrans(aggstate, phase,
1779 : dosort, dohash,
1780 : nullcheck);
1781 :
1782 : /* change back */
1783 72 : aggstate->ss.ps.outerops = outerops;
1784 72 : aggstate->ss.ps.outeropsfixed = outerfixed;
1785 : }
1786 :
1787 61050 : phase->evaltrans = phase->evaltrans_cache[i][j];
1788 61050 : }
1789 :
1790 : /*
1791 : * Set limits that trigger spilling to avoid exceeding hash_mem. Consider the
1792 : * number of partitions we expect to create (if we do spill).
1793 : *
1794 : * There are two limits: a memory limit, and also an ngroups limit. The
1795 : * ngroups limit becomes important when we expect transition values to grow
1796 : * substantially larger than the initial value.
1797 : */
1798 : void
1799 54320 : hash_agg_set_limits(double hashentrysize, double input_groups, int used_bits,
1800 : Size *mem_limit, uint64 *ngroups_limit,
1801 : int *num_partitions)
1802 : {
1803 : int npartitions;
1804 : Size partition_mem;
1805 54320 : Size hash_mem_limit = get_hash_memory_limit();
1806 :
1807 : /* if not expected to spill, use all of hash_mem */
1808 54320 : if (input_groups * hashentrysize <= hash_mem_limit)
1809 : {
1810 51896 : if (num_partitions != NULL)
1811 27320 : *num_partitions = 0;
1812 51896 : *mem_limit = hash_mem_limit;
1813 51896 : *ngroups_limit = hash_mem_limit / hashentrysize;
1814 51896 : return;
1815 : }
1816 :
1817 : /*
1818 : * Calculate expected memory requirements for spilling, which is the size
1819 : * of the buffers needed for all the tapes that need to be open at once.
1820 : * Then, subtract that from the memory available for holding hash tables.
1821 : */
1822 2424 : npartitions = hash_choose_num_partitions(input_groups,
1823 : hashentrysize,
1824 : used_bits,
1825 : NULL);
1826 2424 : if (num_partitions != NULL)
1827 90 : *num_partitions = npartitions;
1828 :
1829 2424 : partition_mem =
1830 2424 : HASHAGG_READ_BUFFER_SIZE +
1831 : HASHAGG_WRITE_BUFFER_SIZE * npartitions;
1832 :
1833 : /*
1834 : * Don't set the limit below 3/4 of hash_mem. In that case, we are at the
1835 : * minimum number of partitions, so we aren't going to dramatically exceed
1836 : * work mem anyway.
1837 : */
1838 2424 : if (hash_mem_limit > 4 * partition_mem)
1839 0 : *mem_limit = hash_mem_limit - partition_mem;
1840 : else
1841 2424 : *mem_limit = hash_mem_limit * 0.75;
1842 :
1843 2424 : if (*mem_limit > hashentrysize)
1844 2424 : *ngroups_limit = *mem_limit / hashentrysize;
1845 : else
1846 0 : *ngroups_limit = 1;
1847 : }
1848 :
1849 : /*
1850 : * hash_agg_check_limits
1851 : *
1852 : * After adding a new group to the hash table, check whether we need to enter
1853 : * spill mode. Allocations may happen without adding new groups (for instance,
1854 : * if the transition state size grows), so this check is imperfect.
1855 : */
1856 : static void
1857 407356 : hash_agg_check_limits(AggState *aggstate)
1858 : {
1859 407356 : uint64 ngroups = aggstate->hash_ngroups_current;
1860 407356 : Size meta_mem = MemoryContextMemAllocated(aggstate->hash_metacxt,
1861 : true);
1862 407356 : Size hashkey_mem = MemoryContextMemAllocated(aggstate->hashcontext->ecxt_per_tuple_memory,
1863 : true);
1864 :
1865 : /*
1866 : * Don't spill unless there's at least one group in the hash table so we
1867 : * can be sure to make progress even in edge cases.
1868 : */
1869 407356 : if (aggstate->hash_ngroups_current > 0 &&
1870 407356 : (meta_mem + hashkey_mem > aggstate->hash_mem_limit ||
1871 380944 : ngroups > aggstate->hash_ngroups_limit))
1872 : {
1873 26442 : hash_agg_enter_spill_mode(aggstate);
1874 : }
1875 407356 : }
1876 :
1877 : /*
1878 : * Enter "spill mode", meaning that no new groups are added to any of the hash
1879 : * tables. Tuples that would create a new group are instead spilled, and
1880 : * processed later.
1881 : */
1882 : static void
1883 26442 : hash_agg_enter_spill_mode(AggState *aggstate)
1884 : {
1885 26442 : aggstate->hash_spill_mode = true;
1886 26442 : hashagg_recompile_expressions(aggstate, aggstate->table_filled, true);
1887 :
1888 26442 : if (!aggstate->hash_ever_spilled)
1889 : {
1890 : Assert(aggstate->hash_tapeset == NULL);
1891 : Assert(aggstate->hash_spills == NULL);
1892 :
1893 54 : aggstate->hash_ever_spilled = true;
1894 :
1895 54 : aggstate->hash_tapeset = LogicalTapeSetCreate(true, NULL, -1);
1896 :
1897 54 : aggstate->hash_spills = palloc(sizeof(HashAggSpill) * aggstate->num_hashes);
1898 :
1899 168 : for (int setno = 0; setno < aggstate->num_hashes; setno++)
1900 : {
1901 114 : AggStatePerHash perhash = &aggstate->perhash[setno];
1902 114 : HashAggSpill *spill = &aggstate->hash_spills[setno];
1903 :
1904 114 : hashagg_spill_init(spill, aggstate->hash_tapeset, 0,
1905 114 : perhash->aggnode->numGroups,
1906 : aggstate->hashentrysize);
1907 : }
1908 : }
1909 26442 : }
1910 :
1911 : /*
1912 : * Update metrics after filling the hash table.
1913 : *
1914 : * If reading from the outer plan, from_tape should be false; if reading from
1915 : * another tape, from_tape should be true.
1916 : */
1917 : static void
1918 40794 : hash_agg_update_metrics(AggState *aggstate, bool from_tape, int npartitions)
1919 : {
1920 : Size meta_mem;
1921 : Size hashkey_mem;
1922 : Size buffer_mem;
1923 : Size total_mem;
1924 :
1925 40794 : if (aggstate->aggstrategy != AGG_MIXED &&
1926 14406 : aggstate->aggstrategy != AGG_HASHED)
1927 0 : return;
1928 :
1929 : /* memory for the hash table itself */
1930 40794 : meta_mem = MemoryContextMemAllocated(aggstate->hash_metacxt, true);
1931 :
1932 : /* memory for the group keys and transition states */
1933 40794 : hashkey_mem = MemoryContextMemAllocated(aggstate->hashcontext->ecxt_per_tuple_memory, true);
1934 :
1935 : /* memory for read/write tape buffers, if spilled */
1936 40794 : buffer_mem = npartitions * HASHAGG_WRITE_BUFFER_SIZE;
1937 40794 : if (from_tape)
1938 26910 : buffer_mem += HASHAGG_READ_BUFFER_SIZE;
1939 :
1940 : /* update peak mem */
1941 40794 : total_mem = meta_mem + hashkey_mem + buffer_mem;
1942 40794 : if (total_mem > aggstate->hash_mem_peak)
1943 6228 : aggstate->hash_mem_peak = total_mem;
1944 :
1945 : /* update disk usage */
1946 40794 : if (aggstate->hash_tapeset != NULL)
1947 : {
1948 26964 : uint64 disk_used = LogicalTapeSetBlocks(aggstate->hash_tapeset) * (BLCKSZ / 1024);
1949 :
1950 26964 : if (aggstate->hash_disk_used < disk_used)
1951 42 : aggstate->hash_disk_used = disk_used;
1952 : }
1953 :
1954 : /* update hashentrysize estimate based on contents */
1955 40794 : if (aggstate->hash_ngroups_current > 0)
1956 : {
1957 38158 : aggstate->hashentrysize =
1958 38158 : sizeof(TupleHashEntryData) +
1959 38158 : (hashkey_mem / (double) aggstate->hash_ngroups_current);
1960 : }
1961 : }
1962 :
1963 : /*
1964 : * Choose a reasonable number of buckets for the initial hash table size.
1965 : */
1966 : static long
1967 6688 : hash_choose_num_buckets(double hashentrysize, long ngroups, Size memory)
1968 : {
1969 : long max_nbuckets;
1970 6688 : long nbuckets = ngroups;
1971 :
1972 6688 : max_nbuckets = memory / hashentrysize;
1973 :
1974 : /*
1975 : * Underestimating is better than overestimating. Too many buckets crowd
1976 : * out space for group keys and transition state values.
1977 : */
1978 6688 : max_nbuckets >>= 1;
1979 :
1980 6688 : if (nbuckets > max_nbuckets)
1981 72 : nbuckets = max_nbuckets;
1982 :
1983 6688 : return Max(nbuckets, 1);
1984 : }
1985 :
1986 : /*
1987 : * Determine the number of partitions to create when spilling, which will
1988 : * always be a power of two. If log2_npartitions is non-NULL, set
1989 : * *log2_npartitions to the log2() of the number of partitions.
1990 : */
1991 : static int
1992 15030 : hash_choose_num_partitions(double input_groups, double hashentrysize,
1993 : int used_bits, int *log2_npartitions)
1994 : {
1995 15030 : Size hash_mem_limit = get_hash_memory_limit();
1996 : double partition_limit;
1997 : double mem_wanted;
1998 : double dpartitions;
1999 : int npartitions;
2000 : int partition_bits;
2001 :
2002 : /*
2003 : * Avoid creating so many partitions that the memory requirements of the
2004 : * open partition files are greater than 1/4 of hash_mem.
2005 : */
2006 15030 : partition_limit =
2007 15030 : (hash_mem_limit * 0.25 - HASHAGG_READ_BUFFER_SIZE) /
2008 : HASHAGG_WRITE_BUFFER_SIZE;
2009 :
2010 15030 : mem_wanted = HASHAGG_PARTITION_FACTOR * input_groups * hashentrysize;
2011 :
2012 : /* make enough partitions so that each one is likely to fit in memory */
2013 15030 : dpartitions = 1 + (mem_wanted / hash_mem_limit);
2014 :
2015 15030 : if (dpartitions > partition_limit)
2016 15012 : dpartitions = partition_limit;
2017 :
2018 15030 : if (dpartitions < HASHAGG_MIN_PARTITIONS)
2019 15030 : dpartitions = HASHAGG_MIN_PARTITIONS;
2020 15030 : if (dpartitions > HASHAGG_MAX_PARTITIONS)
2021 0 : dpartitions = HASHAGG_MAX_PARTITIONS;
2022 :
2023 : /* HASHAGG_MAX_PARTITIONS limit makes this safe */
2024 15030 : npartitions = (int) dpartitions;
2025 :
2026 : /* ceil(log2(npartitions)) */
2027 15030 : partition_bits = my_log2(npartitions);
2028 :
2029 : /* make sure that we don't exhaust the hash bits */
2030 15030 : if (partition_bits + used_bits >= 32)
2031 0 : partition_bits = 32 - used_bits;
2032 :
2033 15030 : if (log2_npartitions != NULL)
2034 12606 : *log2_npartitions = partition_bits;
2035 :
2036 : /* number of partitions will be a power of two */
2037 15030 : npartitions = 1 << partition_bits;
2038 :
2039 15030 : return npartitions;
2040 : }
2041 :
2042 : /*
2043 : * Initialize a freshly-created TupleHashEntry.
2044 : */
2045 : static void
2046 407356 : initialize_hash_entry(AggState *aggstate, TupleHashTable hashtable,
2047 : TupleHashEntry entry)
2048 : {
2049 : AggStatePerGroup pergroup;
2050 : int transno;
2051 :
2052 407356 : aggstate->hash_ngroups_current++;
2053 407356 : hash_agg_check_limits(aggstate);
2054 :
2055 : /* no need to allocate or initialize per-group state */
2056 407356 : if (aggstate->numtrans == 0)
2057 174514 : return;
2058 :
2059 : pergroup = (AggStatePerGroup)
2060 232842 : MemoryContextAlloc(hashtable->tablecxt,
2061 232842 : sizeof(AggStatePerGroupData) * aggstate->numtrans);
2062 :
2063 232842 : entry->additional = pergroup;
2064 :
2065 : /*
2066 : * Initialize aggregates for new tuple group, lookup_hash_entries()
2067 : * already has selected the relevant grouping set.
2068 : */
2069 611294 : for (transno = 0; transno < aggstate->numtrans; transno++)
2070 : {
2071 378452 : AggStatePerTrans pertrans = &aggstate->pertrans[transno];
2072 378452 : AggStatePerGroup pergroupstate = &pergroup[transno];
2073 :
2074 378452 : initialize_aggregate(aggstate, pertrans, pergroupstate);
2075 : }
2076 : }
2077 :
2078 : /*
2079 : * Look up hash entries for the current tuple in all hashed grouping sets.
2080 : *
2081 : * Be aware that lookup_hash_entry can reset the tmpcontext.
2082 : *
2083 : * Some entries may be left NULL if we are in "spill mode". The same tuple
2084 : * will belong to different groups for each grouping set, so may match a group
2085 : * already in memory for one set and match a group not in memory for another
2086 : * set. When in "spill mode", the tuple will be spilled for each grouping set
2087 : * where it doesn't match a group in memory.
2088 : *
2089 : * NB: It's possible to spill the same tuple for several different grouping
2090 : * sets. This may seem wasteful, but it's actually a trade-off: if we spill
2091 : * the tuple multiple times for multiple grouping sets, it can be partitioned
2092 : * for each grouping set, making the refilling of the hash table very
2093 : * efficient.
2094 : */
2095 : static void
2096 5313642 : lookup_hash_entries(AggState *aggstate)
2097 : {
2098 5313642 : AggStatePerGroup *pergroup = aggstate->hash_pergroup;
2099 5313642 : TupleTableSlot *outerslot = aggstate->tmpcontext->ecxt_outertuple;
2100 : int setno;
2101 :
2102 10761580 : for (setno = 0; setno < aggstate->num_hashes; setno++)
2103 : {
2104 5447938 : AggStatePerHash perhash = &aggstate->perhash[setno];
2105 5447938 : TupleHashTable hashtable = perhash->hashtable;
2106 5447938 : TupleTableSlot *hashslot = perhash->hashslot;
2107 : TupleHashEntry entry;
2108 : uint32 hash;
2109 5447938 : bool isnew = false;
2110 : bool *p_isnew;
2111 :
2112 : /* if hash table already spilled, don't create new entries */
2113 5447938 : p_isnew = aggstate->hash_spill_mode ? NULL : &isnew;
2114 :
2115 5447938 : select_current_set(aggstate, setno, true);
2116 5447938 : prepare_hash_slot(perhash,
2117 : outerslot,
2118 : hashslot);
2119 :
2120 5447938 : entry = LookupTupleHashEntry(hashtable, hashslot,
2121 : p_isnew, &hash);
2122 :
2123 5447938 : if (entry != NULL)
2124 : {
2125 5226058 : if (isnew)
2126 312862 : initialize_hash_entry(aggstate, hashtable, entry);
2127 5226058 : pergroup[setno] = entry->additional;
2128 : }
2129 : else
2130 : {
2131 221880 : HashAggSpill *spill = &aggstate->hash_spills[setno];
2132 221880 : TupleTableSlot *slot = aggstate->tmpcontext->ecxt_outertuple;
2133 :
2134 221880 : if (spill->partitions == NULL)
2135 0 : hashagg_spill_init(spill, aggstate->hash_tapeset, 0,
2136 0 : perhash->aggnode->numGroups,
2137 : aggstate->hashentrysize);
2138 :
2139 221880 : hashagg_spill_tuple(aggstate, spill, slot, hash);
2140 221880 : pergroup[setno] = NULL;
2141 : }
2142 : }
2143 5313642 : }
2144 :
2145 : /*
2146 : * ExecAgg -
2147 : *
2148 : * ExecAgg receives tuples from its outer subplan and aggregates over
2149 : * the appropriate attribute for each aggregate function use (Aggref
2150 : * node) appearing in the targetlist or qual of the node. The number
2151 : * of tuples to aggregate over depends on whether grouped or plain
2152 : * aggregation is selected. In grouped aggregation, we produce a result
2153 : * row for each group; in plain aggregation there's a single result row
2154 : * for the whole query. In either case, the value of each aggregate is
2155 : * stored in the expression context to be used when ExecProject evaluates
2156 : * the result tuple.
2157 : */
2158 : static TupleTableSlot *
2159 692722 : ExecAgg(PlanState *pstate)
2160 : {
2161 692722 : AggState *node = castNode(AggState, pstate);
2162 692722 : TupleTableSlot *result = NULL;
2163 :
2164 692722 : CHECK_FOR_INTERRUPTS();
2165 :
2166 692722 : if (!node->agg_done)
2167 : {
2168 : /* Dispatch based on strategy */
2169 638918 : switch (node->phase->aggstrategy)
2170 : {
2171 379068 : case AGG_HASHED:
2172 379068 : if (!node->table_filled)
2173 13758 : agg_fill_hash_table(node);
2174 : /* FALLTHROUGH */
2175 : case AGG_MIXED:
2176 406412 : result = agg_retrieve_hash_table(node);
2177 406412 : break;
2178 232506 : case AGG_PLAIN:
2179 : case AGG_SORTED:
2180 232506 : result = agg_retrieve_direct(node);
2181 232410 : break;
2182 : }
2183 :
2184 638822 : if (!TupIsNull(result))
2185 623396 : return result;
2186 : }
2187 :
2188 69230 : return NULL;
2189 : }
2190 :
2191 : /*
2192 : * ExecAgg for non-hashed case
2193 : */
2194 : static TupleTableSlot *
2195 232506 : agg_retrieve_direct(AggState *aggstate)
2196 : {
2197 232506 : Agg *node = aggstate->phase->aggnode;
2198 : ExprContext *econtext;
2199 : ExprContext *tmpcontext;
2200 : AggStatePerAgg peragg;
2201 : AggStatePerGroup *pergroups;
2202 : TupleTableSlot *outerslot;
2203 : TupleTableSlot *firstSlot;
2204 : TupleTableSlot *result;
2205 232506 : bool hasGroupingSets = aggstate->phase->numsets > 0;
2206 232506 : int numGroupingSets = Max(aggstate->phase->numsets, 1);
2207 : int currentSet;
2208 : int nextSetSize;
2209 : int numReset;
2210 : int i;
2211 :
2212 : /*
2213 : * get state info from node
2214 : *
2215 : * econtext is the per-output-tuple expression context
2216 : *
2217 : * tmpcontext is the per-input-tuple expression context
2218 : */
2219 232506 : econtext = aggstate->ss.ps.ps_ExprContext;
2220 232506 : tmpcontext = aggstate->tmpcontext;
2221 :
2222 232506 : peragg = aggstate->peragg;
2223 232506 : pergroups = aggstate->pergroups;
2224 232506 : firstSlot = aggstate->ss.ss_ScanTupleSlot;
2225 :
2226 : /*
2227 : * We loop retrieving groups until we find one matching
2228 : * aggstate->ss.ps.qual
2229 : *
2230 : * For grouping sets, we have the invariant that aggstate->projected_set
2231 : * is either -1 (initial call) or the index (starting from 0) in
2232 : * gset_lengths for the group we just completed (either by projecting a
2233 : * row or by discarding it in the qual).
2234 : */
2235 303190 : while (!aggstate->agg_done)
2236 : {
2237 : /*
2238 : * Clear the per-output-tuple context for each group, as well as
2239 : * aggcontext (which contains any pass-by-ref transvalues of the old
2240 : * group). Some aggregate functions store working state in child
2241 : * contexts; those now get reset automatically without us needing to
2242 : * do anything special.
2243 : *
2244 : * We use ReScanExprContext not just ResetExprContext because we want
2245 : * any registered shutdown callbacks to be called. That allows
2246 : * aggregate functions to ensure they've cleaned up any non-memory
2247 : * resources.
2248 : */
2249 303062 : ReScanExprContext(econtext);
2250 :
2251 : /*
2252 : * Determine how many grouping sets need to be reset at this boundary.
2253 : */
2254 303062 : if (aggstate->projected_set >= 0 &&
2255 245878 : aggstate->projected_set < numGroupingSets)
2256 245872 : numReset = aggstate->projected_set + 1;
2257 : else
2258 57190 : numReset = numGroupingSets;
2259 :
2260 : /*
2261 : * numReset can change on a phase boundary, but that's OK; we want to
2262 : * reset the contexts used in _this_ phase, and later, after possibly
2263 : * changing phase, initialize the right number of aggregates for the
2264 : * _new_ phase.
2265 : */
2266 :
2267 628366 : for (i = 0; i < numReset; i++)
2268 : {
2269 325304 : ReScanExprContext(aggstate->aggcontexts[i]);
2270 : }
2271 :
2272 : /*
2273 : * Check if input is complete and there are no more groups to project
2274 : * in this phase; move to next phase or mark as done.
2275 : */
2276 303062 : if (aggstate->input_done == true &&
2277 1524 : aggstate->projected_set >= (numGroupingSets - 1))
2278 : {
2279 720 : if (aggstate->current_phase < aggstate->numphases - 1)
2280 : {
2281 186 : initialize_phase(aggstate, aggstate->current_phase + 1);
2282 186 : aggstate->input_done = false;
2283 186 : aggstate->projected_set = -1;
2284 186 : numGroupingSets = Max(aggstate->phase->numsets, 1);
2285 186 : node = aggstate->phase->aggnode;
2286 186 : numReset = numGroupingSets;
2287 : }
2288 534 : else if (aggstate->aggstrategy == AGG_MIXED)
2289 : {
2290 : /*
2291 : * Mixed mode; we've output all the grouped stuff and have
2292 : * full hashtables, so switch to outputting those.
2293 : */
2294 138 : initialize_phase(aggstate, 0);
2295 138 : aggstate->table_filled = true;
2296 138 : ResetTupleHashIterator(aggstate->perhash[0].hashtable,
2297 : &aggstate->perhash[0].hashiter);
2298 138 : select_current_set(aggstate, 0, true);
2299 138 : return agg_retrieve_hash_table(aggstate);
2300 : }
2301 : else
2302 : {
2303 396 : aggstate->agg_done = true;
2304 396 : break;
2305 : }
2306 : }
2307 :
2308 : /*
2309 : * Get the number of columns in the next grouping set after the last
2310 : * projected one (if any). This is the number of columns to compare to
2311 : * see if we reached the boundary of that set too.
2312 : */
2313 302528 : if (aggstate->projected_set >= 0 &&
2314 245158 : aggstate->projected_set < (numGroupingSets - 1))
2315 27276 : nextSetSize = aggstate->phase->gset_lengths[aggstate->projected_set + 1];
2316 : else
2317 275252 : nextSetSize = 0;
2318 :
2319 : /*----------
2320 : * If a subgroup for the current grouping set is present, project it.
2321 : *
2322 : * We have a new group if:
2323 : * - we're out of input but haven't projected all grouping sets
2324 : * (checked above)
2325 : * OR
2326 : * - we already projected a row that wasn't from the last grouping
2327 : * set
2328 : * AND
2329 : * - the next grouping set has at least one grouping column (since
2330 : * empty grouping sets project only once input is exhausted)
2331 : * AND
2332 : * - the previous and pending rows differ on the grouping columns
2333 : * of the next grouping set
2334 : *----------
2335 : */
2336 302528 : tmpcontext->ecxt_innertuple = econtext->ecxt_outertuple;
2337 302528 : if (aggstate->input_done ||
2338 301724 : (node->aggstrategy != AGG_PLAIN &&
2339 246012 : aggstate->projected_set != -1 &&
2340 244354 : aggstate->projected_set < (numGroupingSets - 1) &&
2341 19940 : nextSetSize > 0 &&
2342 19940 : !ExecQualAndReset(aggstate->phase->eqfunctions[nextSetSize - 1],
2343 : tmpcontext)))
2344 : {
2345 14138 : aggstate->projected_set += 1;
2346 :
2347 : Assert(aggstate->projected_set < numGroupingSets);
2348 14138 : Assert(nextSetSize > 0 || aggstate->input_done);
2349 : }
2350 : else
2351 : {
2352 : /*
2353 : * We no longer care what group we just projected, the next
2354 : * projection will always be the first (or only) grouping set
2355 : * (unless the input proves to be empty).
2356 : */
2357 288390 : aggstate->projected_set = 0;
2358 :
2359 : /*
2360 : * If we don't already have the first tuple of the new group,
2361 : * fetch it from the outer plan.
2362 : */
2363 288390 : if (aggstate->grp_firstTuple == NULL)
2364 : {
2365 57370 : outerslot = fetch_input_tuple(aggstate);
2366 57352 : if (!TupIsNull(outerslot))
2367 : {
2368 : /*
2369 : * Make a copy of the first input tuple; we will use this
2370 : * for comparisons (in group mode) and for projection.
2371 : */
2372 46780 : aggstate->grp_firstTuple = ExecCopySlotHeapTuple(outerslot);
2373 : }
2374 : else
2375 : {
2376 : /* outer plan produced no tuples at all */
2377 10572 : if (hasGroupingSets)
2378 : {
2379 : /*
2380 : * If there was no input at all, we need to project
2381 : * rows only if there are grouping sets of size 0.
2382 : * Note that this implies that there can't be any
2383 : * references to ungrouped Vars, which would otherwise
2384 : * cause issues with the empty output slot.
2385 : *
2386 : * XXX: This is no longer true, we currently deal with
2387 : * this in finalize_aggregates().
2388 : */
2389 54 : aggstate->input_done = true;
2390 :
2391 78 : while (aggstate->phase->gset_lengths[aggstate->projected_set] > 0)
2392 : {
2393 30 : aggstate->projected_set += 1;
2394 30 : if (aggstate->projected_set >= numGroupingSets)
2395 : {
2396 : /*
2397 : * We can't set agg_done here because we might
2398 : * have more phases to do, even though the
2399 : * input is empty. So we need to restart the
2400 : * whole outer loop.
2401 : */
2402 6 : break;
2403 : }
2404 : }
2405 :
2406 54 : if (aggstate->projected_set >= numGroupingSets)
2407 6 : continue;
2408 : }
2409 : else
2410 : {
2411 10518 : aggstate->agg_done = true;
2412 : /* If we are grouping, we should produce no tuples too */
2413 10518 : if (node->aggstrategy != AGG_PLAIN)
2414 164 : return NULL;
2415 : }
2416 : }
2417 : }
2418 :
2419 : /*
2420 : * Initialize working state for a new input tuple group.
2421 : */
2422 288202 : initialize_aggregates(aggstate, pergroups, numReset);
2423 :
2424 288202 : if (aggstate->grp_firstTuple != NULL)
2425 : {
2426 : /*
2427 : * Store the copied first input tuple in the tuple table slot
2428 : * reserved for it. The tuple will be deleted when it is
2429 : * cleared from the slot.
2430 : */
2431 277800 : ExecForceStoreHeapTuple(aggstate->grp_firstTuple,
2432 : firstSlot, true);
2433 277800 : aggstate->grp_firstTuple = NULL; /* don't keep two pointers */
2434 :
2435 : /* set up for first advance_aggregates call */
2436 277800 : tmpcontext->ecxt_outertuple = firstSlot;
2437 :
2438 : /*
2439 : * Process each outer-plan tuple, and then fetch the next one,
2440 : * until we exhaust the outer plan or cross a group boundary.
2441 : */
2442 : for (;;)
2443 : {
2444 : /*
2445 : * During phase 1 only of a mixed agg, we need to update
2446 : * hashtables as well in advance_aggregates.
2447 : */
2448 19548424 : if (aggstate->aggstrategy == AGG_MIXED &&
2449 38044 : aggstate->current_phase == 1)
2450 : {
2451 38044 : lookup_hash_entries(aggstate);
2452 : }
2453 :
2454 : /* Advance the aggregates (or combine functions) */
2455 19548424 : advance_aggregates(aggstate);
2456 :
2457 : /* Reset per-input-tuple context after each tuple */
2458 19548358 : ResetExprContext(tmpcontext);
2459 :
2460 19548358 : outerslot = fetch_input_tuple(aggstate);
2461 19548358 : if (TupIsNull(outerslot))
2462 : {
2463 : /* no more outer-plan tuples available */
2464 :
2465 : /* if we built hash tables, finalize any spills */
2466 46708 : if (aggstate->aggstrategy == AGG_MIXED &&
2467 126 : aggstate->current_phase == 1)
2468 126 : hashagg_finish_initial_spills(aggstate);
2469 :
2470 46708 : if (hasGroupingSets)
2471 : {
2472 666 : aggstate->input_done = true;
2473 666 : break;
2474 : }
2475 : else
2476 : {
2477 46042 : aggstate->agg_done = true;
2478 46042 : break;
2479 : }
2480 : }
2481 : /* set up for next advance_aggregates call */
2482 19501650 : tmpcontext->ecxt_outertuple = outerslot;
2483 :
2484 : /*
2485 : * If we are grouping, check whether we've crossed a group
2486 : * boundary.
2487 : */
2488 19501650 : if (node->aggstrategy != AGG_PLAIN && node->numCols > 0)
2489 : {
2490 1782694 : tmpcontext->ecxt_innertuple = firstSlot;
2491 1782694 : if (!ExecQual(aggstate->phase->eqfunctions[node->numCols - 1],
2492 : tmpcontext))
2493 : {
2494 231026 : aggstate->grp_firstTuple = ExecCopySlotHeapTuple(outerslot);
2495 231026 : break;
2496 : }
2497 : }
2498 : }
2499 : }
2500 :
2501 : /*
2502 : * Use the representative input tuple for any references to
2503 : * non-aggregated input columns in aggregate direct args, the node
2504 : * qual, and the tlist. (If we are not grouping, and there are no
2505 : * input rows at all, we will come here with an empty firstSlot
2506 : * ... but if not grouping, there can't be any references to
2507 : * non-aggregated input columns, so no problem.)
2508 : */
2509 288136 : econtext->ecxt_outertuple = firstSlot;
2510 : }
2511 :
2512 : Assert(aggstate->projected_set >= 0);
2513 :
2514 302274 : currentSet = aggstate->projected_set;
2515 :
2516 302274 : prepare_projection_slot(aggstate, econtext->ecxt_outertuple, currentSet);
2517 :
2518 302274 : select_current_set(aggstate, currentSet, false);
2519 :
2520 302274 : finalize_aggregates(aggstate,
2521 : peragg,
2522 302274 : pergroups[currentSet]);
2523 :
2524 : /*
2525 : * If there's no row to project right now, we must continue rather
2526 : * than returning a null since there might be more groups.
2527 : */
2528 302262 : result = project_aggregates(aggstate);
2529 302262 : if (result)
2530 231584 : return result;
2531 : }
2532 :
2533 : /* No more groups */
2534 524 : return NULL;
2535 : }
2536 :
2537 : /*
2538 : * ExecAgg for hashed case: read input and build hash table
2539 : */
2540 : static void
2541 13758 : agg_fill_hash_table(AggState *aggstate)
2542 : {
2543 : TupleTableSlot *outerslot;
2544 13758 : ExprContext *tmpcontext = aggstate->tmpcontext;
2545 :
2546 : /*
2547 : * Process each outer-plan tuple, and then fetch the next one, until we
2548 : * exhaust the outer plan.
2549 : */
2550 : for (;;)
2551 : {
2552 5289356 : outerslot = fetch_input_tuple(aggstate);
2553 5289356 : if (TupIsNull(outerslot))
2554 : break;
2555 :
2556 : /* set up for lookup_hash_entries and advance_aggregates */
2557 5275598 : tmpcontext->ecxt_outertuple = outerslot;
2558 :
2559 : /* Find or build hashtable entries */
2560 5275598 : lookup_hash_entries(aggstate);
2561 :
2562 : /* Advance the aggregates (or combine functions) */
2563 5275598 : advance_aggregates(aggstate);
2564 :
2565 : /*
2566 : * Reset per-input-tuple context after each tuple, but note that the
2567 : * hash lookups do this too
2568 : */
2569 5275598 : ResetExprContext(aggstate->tmpcontext);
2570 : }
2571 :
2572 : /* finalize spills, if any */
2573 13758 : hashagg_finish_initial_spills(aggstate);
2574 :
2575 13758 : aggstate->table_filled = true;
2576 : /* Initialize to walk the first hash table */
2577 13758 : select_current_set(aggstate, 0, true);
2578 13758 : ResetTupleHashIterator(aggstate->perhash[0].hashtable,
2579 : &aggstate->perhash[0].hashiter);
2580 13758 : }
2581 :
2582 : /*
2583 : * If any data was spilled during hash aggregation, reset the hash table and
2584 : * reprocess one batch of spilled data. After reprocessing a batch, the hash
2585 : * table will again contain data, ready to be consumed by
2586 : * agg_retrieve_hash_table_in_memory().
2587 : *
2588 : * Should only be called after all in memory hash table entries have been
2589 : * finalized and emitted.
2590 : *
2591 : * Return false when input is exhausted and there's no more work to be done;
2592 : * otherwise return true.
2593 : */
2594 : static bool
2595 41648 : agg_refill_hash_table(AggState *aggstate)
2596 : {
2597 : HashAggBatch *batch;
2598 : AggStatePerHash perhash;
2599 : HashAggSpill spill;
2600 41648 : LogicalTapeSet *tapeset = aggstate->hash_tapeset;
2601 41648 : bool spill_initialized = false;
2602 :
2603 41648 : if (aggstate->hash_batches == NIL)
2604 14738 : return false;
2605 :
2606 : /* hash_batches is a stack, with the top item at the end of the list */
2607 26910 : batch = llast(aggstate->hash_batches);
2608 26910 : aggstate->hash_batches = list_delete_last(aggstate->hash_batches);
2609 :
2610 26910 : hash_agg_set_limits(aggstate->hashentrysize, batch->input_card,
2611 : batch->used_bits, &aggstate->hash_mem_limit,
2612 : &aggstate->hash_ngroups_limit, NULL);
2613 :
2614 : /*
2615 : * Each batch only processes one grouping set; set the rest to NULL so
2616 : * that advance_aggregates() knows to ignore them. We don't touch
2617 : * pergroups for sorted grouping sets here, because they will be needed if
2618 : * we rescan later. The expressions for sorted grouping sets will not be
2619 : * evaluated after we recompile anyway.
2620 : */
2621 207360 : MemSet(aggstate->hash_pergroup, 0,
2622 : sizeof(AggStatePerGroup) * aggstate->num_hashes);
2623 :
2624 : /* free memory and reset hash tables */
2625 26910 : ReScanExprContext(aggstate->hashcontext);
2626 207360 : for (int setno = 0; setno < aggstate->num_hashes; setno++)
2627 180450 : ResetTupleHashTable(aggstate->perhash[setno].hashtable);
2628 :
2629 26910 : aggstate->hash_ngroups_current = 0;
2630 :
2631 : /*
2632 : * In AGG_MIXED mode, hash aggregation happens in phase 1 and the output
2633 : * happens in phase 0. So, we switch to phase 1 when processing a batch,
2634 : * and back to phase 0 after the batch is done.
2635 : */
2636 : Assert(aggstate->current_phase == 0);
2637 26910 : if (aggstate->phase->aggstrategy == AGG_MIXED)
2638 : {
2639 26262 : aggstate->current_phase = 1;
2640 26262 : aggstate->phase = &aggstate->phases[aggstate->current_phase];
2641 : }
2642 :
2643 26910 : select_current_set(aggstate, batch->setno, true);
2644 :
2645 26910 : perhash = &aggstate->perhash[aggstate->current_set];
2646 :
2647 : /*
2648 : * Spilled tuples are always read back as MinimalTuples, which may be
2649 : * different from the outer plan, so recompile the aggregate expressions.
2650 : *
2651 : * We still need the NULL check, because we are only processing one
2652 : * grouping set at a time and the rest will be NULL.
2653 : */
2654 26910 : hashagg_recompile_expressions(aggstate, true, true);
2655 :
2656 : for (;;)
2657 655824 : {
2658 682734 : TupleTableSlot *spillslot = aggstate->hash_spill_rslot;
2659 682734 : TupleTableSlot *hashslot = perhash->hashslot;
2660 : TupleHashEntry entry;
2661 : MinimalTuple tuple;
2662 : uint32 hash;
2663 682734 : bool isnew = false;
2664 682734 : bool *p_isnew = aggstate->hash_spill_mode ? NULL : &isnew;
2665 :
2666 682734 : CHECK_FOR_INTERRUPTS();
2667 :
2668 682734 : tuple = hashagg_batch_read(batch, &hash);
2669 682734 : if (tuple == NULL)
2670 26910 : break;
2671 :
2672 655824 : ExecStoreMinimalTuple(tuple, spillslot, true);
2673 655824 : aggstate->tmpcontext->ecxt_outertuple = spillslot;
2674 :
2675 655824 : prepare_hash_slot(perhash,
2676 655824 : aggstate->tmpcontext->ecxt_outertuple,
2677 : hashslot);
2678 655824 : entry = LookupTupleHashEntryHash(perhash->hashtable, hashslot,
2679 : p_isnew, hash);
2680 :
2681 655824 : if (entry != NULL)
2682 : {
2683 221880 : if (isnew)
2684 94494 : initialize_hash_entry(aggstate, perhash->hashtable, entry);
2685 221880 : aggstate->hash_pergroup[batch->setno] = entry->additional;
2686 221880 : advance_aggregates(aggstate);
2687 : }
2688 : else
2689 : {
2690 433944 : if (!spill_initialized)
2691 : {
2692 : /*
2693 : * Avoid initializing the spill until we actually need it so
2694 : * that we don't assign tapes that will never be used.
2695 : */
2696 12492 : spill_initialized = true;
2697 12492 : hashagg_spill_init(&spill, tapeset, batch->used_bits,
2698 : batch->input_card, aggstate->hashentrysize);
2699 : }
2700 : /* no memory for a new group, spill */
2701 433944 : hashagg_spill_tuple(aggstate, &spill, spillslot, hash);
2702 :
2703 433944 : aggstate->hash_pergroup[batch->setno] = NULL;
2704 : }
2705 :
2706 : /*
2707 : * Reset per-input-tuple context after each tuple, but note that the
2708 : * hash lookups do this too
2709 : */
2710 655824 : ResetExprContext(aggstate->tmpcontext);
2711 : }
2712 :
2713 26910 : LogicalTapeClose(batch->input_tape);
2714 :
2715 : /* change back to phase 0 */
2716 26910 : aggstate->current_phase = 0;
2717 26910 : aggstate->phase = &aggstate->phases[aggstate->current_phase];
2718 :
2719 26910 : if (spill_initialized)
2720 : {
2721 12492 : hashagg_spill_finish(aggstate, &spill, batch->setno);
2722 12492 : hash_agg_update_metrics(aggstate, true, spill.npartitions);
2723 : }
2724 : else
2725 14418 : hash_agg_update_metrics(aggstate, true, 0);
2726 :
2727 26910 : aggstate->hash_spill_mode = false;
2728 :
2729 : /* prepare to walk the first hash table */
2730 26910 : select_current_set(aggstate, batch->setno, true);
2731 26910 : ResetTupleHashIterator(aggstate->perhash[batch->setno].hashtable,
2732 : &aggstate->perhash[batch->setno].hashiter);
2733 :
2734 26910 : pfree(batch);
2735 :
2736 26910 : return true;
2737 : }
2738 :
2739 : /*
2740 : * ExecAgg for hashed case: retrieving groups from hash table
2741 : *
2742 : * After exhausting in-memory tuples, also try refilling the hash table using
2743 : * previously-spilled tuples. Only returns NULL after all in-memory and
2744 : * spilled tuples are exhausted.
2745 : */
2746 : static TupleTableSlot *
2747 406550 : agg_retrieve_hash_table(AggState *aggstate)
2748 : {
2749 406550 : TupleTableSlot *result = NULL;
2750 :
2751 825272 : while (result == NULL)
2752 : {
2753 433460 : result = agg_retrieve_hash_table_in_memory(aggstate);
2754 433460 : if (result == NULL)
2755 : {
2756 41648 : if (!agg_refill_hash_table(aggstate))
2757 : {
2758 14738 : aggstate->agg_done = true;
2759 14738 : break;
2760 : }
2761 : }
2762 : }
2763 :
2764 406550 : return result;
2765 : }
2766 :
2767 : /*
2768 : * Retrieve the groups from the in-memory hash tables without considering any
2769 : * spilled tuples.
2770 : */
2771 : static TupleTableSlot *
2772 433460 : agg_retrieve_hash_table_in_memory(AggState *aggstate)
2773 : {
2774 : ExprContext *econtext;
2775 : AggStatePerAgg peragg;
2776 : AggStatePerGroup pergroup;
2777 : TupleHashEntryData *entry;
2778 : TupleTableSlot *firstSlot;
2779 : TupleTableSlot *result;
2780 : AggStatePerHash perhash;
2781 :
2782 : /*
2783 : * get state info from node.
2784 : *
2785 : * econtext is the per-output-tuple expression context.
2786 : */
2787 433460 : econtext = aggstate->ss.ps.ps_ExprContext;
2788 433460 : peragg = aggstate->peragg;
2789 433460 : firstSlot = aggstate->ss.ss_ScanTupleSlot;
2790 :
2791 : /*
2792 : * Note that perhash (and therefore anything accessed through it) can
2793 : * change inside the loop, as we change between grouping sets.
2794 : */
2795 433460 : perhash = &aggstate->perhash[aggstate->current_set];
2796 :
2797 : /*
2798 : * We loop retrieving groups until we find one satisfying
2799 : * aggstate->ss.ps.qual
2800 : */
2801 : for (;;)
2802 117852 : {
2803 551312 : TupleTableSlot *hashslot = perhash->hashslot;
2804 : int i;
2805 :
2806 551312 : CHECK_FOR_INTERRUPTS();
2807 :
2808 : /*
2809 : * Find the next entry in the hash table
2810 : */
2811 551312 : entry = ScanTupleHashTable(perhash->hashtable, &perhash->hashiter);
2812 551312 : if (entry == NULL)
2813 : {
2814 141880 : int nextset = aggstate->current_set + 1;
2815 :
2816 141880 : if (nextset < aggstate->num_hashes)
2817 : {
2818 : /*
2819 : * Switch to next grouping set, reinitialize, and restart the
2820 : * loop.
2821 : */
2822 100232 : select_current_set(aggstate, nextset, true);
2823 :
2824 100232 : perhash = &aggstate->perhash[aggstate->current_set];
2825 :
2826 100232 : ResetTupleHashIterator(perhash->hashtable, &perhash->hashiter);
2827 :
2828 100232 : continue;
2829 : }
2830 : else
2831 : {
2832 41648 : return NULL;
2833 : }
2834 : }
2835 :
2836 : /*
2837 : * Clear the per-output-tuple context for each group
2838 : *
2839 : * We intentionally don't use ReScanExprContext here; if any aggs have
2840 : * registered shutdown callbacks, they mustn't be called yet, since we
2841 : * might not be done with that agg.
2842 : */
2843 409432 : ResetExprContext(econtext);
2844 :
2845 : /*
2846 : * Transform representative tuple back into one with the right
2847 : * columns.
2848 : */
2849 409432 : ExecStoreMinimalTuple(entry->firstTuple, hashslot, false);
2850 409432 : slot_getallattrs(hashslot);
2851 :
2852 409432 : ExecClearTuple(firstSlot);
2853 409432 : memset(firstSlot->tts_isnull, true,
2854 409432 : firstSlot->tts_tupleDescriptor->natts * sizeof(bool));
2855 :
2856 1175082 : for (i = 0; i < perhash->numhashGrpCols; i++)
2857 : {
2858 765650 : int varNumber = perhash->hashGrpColIdxInput[i] - 1;
2859 :
2860 765650 : firstSlot->tts_values[varNumber] = hashslot->tts_values[i];
2861 765650 : firstSlot->tts_isnull[varNumber] = hashslot->tts_isnull[i];
2862 : }
2863 409432 : ExecStoreVirtualTuple(firstSlot);
2864 :
2865 409432 : pergroup = (AggStatePerGroup) entry->additional;
2866 :
2867 : /*
2868 : * Use the representative input tuple for any references to
2869 : * non-aggregated input columns in the qual and tlist.
2870 : */
2871 409432 : econtext->ecxt_outertuple = firstSlot;
2872 :
2873 409432 : prepare_projection_slot(aggstate,
2874 : econtext->ecxt_outertuple,
2875 : aggstate->current_set);
2876 :
2877 409432 : finalize_aggregates(aggstate, peragg, pergroup);
2878 :
2879 409432 : result = project_aggregates(aggstate);
2880 409432 : if (result)
2881 391812 : return result;
2882 : }
2883 :
2884 : /* No more groups */
2885 : return NULL;
2886 : }
2887 :
2888 : /*
2889 : * hashagg_spill_init
2890 : *
2891 : * Called after we determined that spilling is necessary. Chooses the number
2892 : * of partitions to create, and initializes them.
2893 : */
2894 : static void
2895 12606 : hashagg_spill_init(HashAggSpill *spill, LogicalTapeSet *tapeset, int used_bits,
2896 : double input_groups, double hashentrysize)
2897 : {
2898 : int npartitions;
2899 : int partition_bits;
2900 :
2901 12606 : npartitions = hash_choose_num_partitions(input_groups, hashentrysize,
2902 : used_bits, &partition_bits);
2903 :
2904 12606 : spill->partitions = palloc0(sizeof(LogicalTape *) * npartitions);
2905 12606 : spill->ntuples = palloc0(sizeof(int64) * npartitions);
2906 12606 : spill->hll_card = palloc0(sizeof(hyperLogLogState) * npartitions);
2907 :
2908 63030 : for (int i = 0; i < npartitions; i++)
2909 50424 : spill->partitions[i] = LogicalTapeCreate(tapeset);
2910 :
2911 12606 : spill->shift = 32 - used_bits - partition_bits;
2912 12606 : spill->mask = (npartitions - 1) << spill->shift;
2913 12606 : spill->npartitions = npartitions;
2914 :
2915 63030 : for (int i = 0; i < npartitions; i++)
2916 50424 : initHyperLogLog(&spill->hll_card[i], HASHAGG_HLL_BIT_WIDTH);
2917 12606 : }
2918 :
2919 : /*
2920 : * hashagg_spill_tuple
2921 : *
2922 : * No room for new groups in the hash table. Save for later in the appropriate
2923 : * partition.
2924 : */
2925 : static Size
2926 655824 : hashagg_spill_tuple(AggState *aggstate, HashAggSpill *spill,
2927 : TupleTableSlot *inputslot, uint32 hash)
2928 : {
2929 : TupleTableSlot *spillslot;
2930 : int partition;
2931 : MinimalTuple tuple;
2932 : LogicalTape *tape;
2933 655824 : int total_written = 0;
2934 : bool shouldFree;
2935 :
2936 : Assert(spill->partitions != NULL);
2937 :
2938 : /* spill only attributes that we actually need */
2939 655824 : if (!aggstate->all_cols_needed)
2940 : {
2941 3492 : spillslot = aggstate->hash_spill_wslot;
2942 3492 : slot_getsomeattrs(inputslot, aggstate->max_colno_needed);
2943 3492 : ExecClearTuple(spillslot);
2944 10476 : for (int i = 0; i < spillslot->tts_tupleDescriptor->natts; i++)
2945 : {
2946 6984 : if (bms_is_member(i + 1, aggstate->colnos_needed))
2947 : {
2948 3492 : spillslot->tts_values[i] = inputslot->tts_values[i];
2949 3492 : spillslot->tts_isnull[i] = inputslot->tts_isnull[i];
2950 : }
2951 : else
2952 3492 : spillslot->tts_isnull[i] = true;
2953 : }
2954 3492 : ExecStoreVirtualTuple(spillslot);
2955 : }
2956 : else
2957 652332 : spillslot = inputslot;
2958 :
2959 655824 : tuple = ExecFetchSlotMinimalTuple(spillslot, &shouldFree);
2960 :
2961 655824 : partition = (hash & spill->mask) >> spill->shift;
2962 655824 : spill->ntuples[partition]++;
2963 :
2964 : /*
2965 : * All hash values destined for a given partition have some bits in
2966 : * common, which causes bad HLL cardinality estimates. Hash the hash to
2967 : * get a more uniform distribution.
2968 : */
2969 655824 : addHyperLogLog(&spill->hll_card[partition], hash_bytes_uint32(hash));
2970 :
2971 655824 : tape = spill->partitions[partition];
2972 :
2973 655824 : LogicalTapeWrite(tape, &hash, sizeof(uint32));
2974 655824 : total_written += sizeof(uint32);
2975 :
2976 655824 : LogicalTapeWrite(tape, tuple, tuple->t_len);
2977 655824 : total_written += tuple->t_len;
2978 :
2979 655824 : if (shouldFree)
2980 221880 : pfree(tuple);
2981 :
2982 655824 : return total_written;
2983 : }
2984 :
2985 : /*
2986 : * hashagg_batch_new
2987 : *
2988 : * Construct a HashAggBatch item, which represents one iteration of HashAgg to
2989 : * be done.
2990 : */
2991 : static HashAggBatch *
2992 26910 : hashagg_batch_new(LogicalTape *input_tape, int setno,
2993 : int64 input_tuples, double input_card, int used_bits)
2994 : {
2995 26910 : HashAggBatch *batch = palloc0(sizeof(HashAggBatch));
2996 :
2997 26910 : batch->setno = setno;
2998 26910 : batch->used_bits = used_bits;
2999 26910 : batch->input_tape = input_tape;
3000 26910 : batch->input_tuples = input_tuples;
3001 26910 : batch->input_card = input_card;
3002 :
3003 26910 : return batch;
3004 : }
3005 :
3006 : /*
3007 : * hashagg_batch_read
3008 : * read the next tuple from a batch's tape. Return NULL if no more.
3009 : */
3010 : static MinimalTuple
3011 682734 : hashagg_batch_read(HashAggBatch *batch, uint32 *hashp)
3012 : {
3013 682734 : LogicalTape *tape = batch->input_tape;
3014 : MinimalTuple tuple;
3015 : uint32 t_len;
3016 : size_t nread;
3017 : uint32 hash;
3018 :
3019 682734 : nread = LogicalTapeRead(tape, &hash, sizeof(uint32));
3020 682734 : if (nread == 0)
3021 26910 : return NULL;
3022 655824 : if (nread != sizeof(uint32))
3023 0 : ereport(ERROR,
3024 : (errcode_for_file_access(),
3025 : errmsg_internal("unexpected EOF for tape %p: requested %zu bytes, read %zu bytes",
3026 : tape, sizeof(uint32), nread)));
3027 655824 : if (hashp != NULL)
3028 655824 : *hashp = hash;
3029 :
3030 655824 : nread = LogicalTapeRead(tape, &t_len, sizeof(t_len));
3031 655824 : if (nread != sizeof(uint32))
3032 0 : ereport(ERROR,
3033 : (errcode_for_file_access(),
3034 : errmsg_internal("unexpected EOF for tape %p: requested %zu bytes, read %zu bytes",
3035 : tape, sizeof(uint32), nread)));
3036 :
3037 655824 : tuple = (MinimalTuple) palloc(t_len);
3038 655824 : tuple->t_len = t_len;
3039 :
3040 655824 : nread = LogicalTapeRead(tape,
3041 : (char *) tuple + sizeof(uint32),
3042 : t_len - sizeof(uint32));
3043 655824 : if (nread != t_len - sizeof(uint32))
3044 0 : ereport(ERROR,
3045 : (errcode_for_file_access(),
3046 : errmsg_internal("unexpected EOF for tape %p: requested %zu bytes, read %zu bytes",
3047 : tape, t_len - sizeof(uint32), nread)));
3048 :
3049 655824 : return tuple;
3050 : }
3051 :
3052 : /*
3053 : * hashagg_finish_initial_spills
3054 : *
3055 : * After a HashAggBatch has been processed, it may have spilled tuples to
3056 : * disk. If so, turn the spilled partitions into new batches that must later
3057 : * be executed.
3058 : */
3059 : static void
3060 13884 : hashagg_finish_initial_spills(AggState *aggstate)
3061 : {
3062 : int setno;
3063 13884 : int total_npartitions = 0;
3064 :
3065 13884 : if (aggstate->hash_spills != NULL)
3066 : {
3067 168 : for (setno = 0; setno < aggstate->num_hashes; setno++)
3068 : {
3069 114 : HashAggSpill *spill = &aggstate->hash_spills[setno];
3070 :
3071 114 : total_npartitions += spill->npartitions;
3072 114 : hashagg_spill_finish(aggstate, spill, setno);
3073 : }
3074 :
3075 : /*
3076 : * We're not processing tuples from outer plan any more; only
3077 : * processing batches of spilled tuples. The initial spill structures
3078 : * are no longer needed.
3079 : */
3080 54 : pfree(aggstate->hash_spills);
3081 54 : aggstate->hash_spills = NULL;
3082 : }
3083 :
3084 13884 : hash_agg_update_metrics(aggstate, false, total_npartitions);
3085 13884 : aggstate->hash_spill_mode = false;
3086 13884 : }
3087 :
3088 : /*
3089 : * hashagg_spill_finish
3090 : *
3091 : * Transform spill partitions into new batches.
3092 : */
3093 : static void
3094 12606 : hashagg_spill_finish(AggState *aggstate, HashAggSpill *spill, int setno)
3095 : {
3096 : int i;
3097 12606 : int used_bits = 32 - spill->shift;
3098 :
3099 12606 : if (spill->npartitions == 0)
3100 0 : return; /* didn't spill */
3101 :
3102 63030 : for (i = 0; i < spill->npartitions; i++)
3103 : {
3104 50424 : LogicalTape *tape = spill->partitions[i];
3105 : HashAggBatch *new_batch;
3106 : double cardinality;
3107 :
3108 : /* if the partition is empty, don't create a new batch of work */
3109 50424 : if (spill->ntuples[i] == 0)
3110 23514 : continue;
3111 :
3112 26910 : cardinality = estimateHyperLogLog(&spill->hll_card[i]);
3113 26910 : freeHyperLogLog(&spill->hll_card[i]);
3114 :
3115 : /* rewinding frees the buffer while not in use */
3116 26910 : LogicalTapeRewindForRead(tape, HASHAGG_READ_BUFFER_SIZE);
3117 :
3118 26910 : new_batch = hashagg_batch_new(tape, setno,
3119 26910 : spill->ntuples[i], cardinality,
3120 : used_bits);
3121 26910 : aggstate->hash_batches = lappend(aggstate->hash_batches, new_batch);
3122 26910 : aggstate->hash_batches_used++;
3123 : }
3124 :
3125 12606 : pfree(spill->ntuples);
3126 12606 : pfree(spill->hll_card);
3127 12606 : pfree(spill->partitions);
3128 : }
3129 :
3130 : /*
3131 : * Free resources related to a spilled HashAgg.
3132 : */
3133 : static void
3134 49708 : hashagg_reset_spill_state(AggState *aggstate)
3135 : {
3136 : /* free spills from initial pass */
3137 49708 : if (aggstate->hash_spills != NULL)
3138 : {
3139 : int setno;
3140 :
3141 0 : for (setno = 0; setno < aggstate->num_hashes; setno++)
3142 : {
3143 0 : HashAggSpill *spill = &aggstate->hash_spills[setno];
3144 :
3145 0 : pfree(spill->ntuples);
3146 0 : pfree(spill->partitions);
3147 : }
3148 0 : pfree(aggstate->hash_spills);
3149 0 : aggstate->hash_spills = NULL;
3150 : }
3151 :
3152 : /* free batches */
3153 49708 : list_free_deep(aggstate->hash_batches);
3154 49708 : aggstate->hash_batches = NIL;
3155 :
3156 : /* close tape set */
3157 49708 : if (aggstate->hash_tapeset != NULL)
3158 : {
3159 54 : LogicalTapeSetClose(aggstate->hash_tapeset);
3160 54 : aggstate->hash_tapeset = NULL;
3161 : }
3162 49708 : }
3163 :
3164 :
3165 : /* -----------------
3166 : * ExecInitAgg
3167 : *
3168 : * Creates the run-time information for the agg node produced by the
3169 : * planner and initializes its outer subtree.
3170 : *
3171 : * -----------------
3172 : */
3173 : AggState *
3174 42122 : ExecInitAgg(Agg *node, EState *estate, int eflags)
3175 : {
3176 : AggState *aggstate;
3177 : AggStatePerAgg peraggs;
3178 : AggStatePerTrans pertransstates;
3179 : AggStatePerGroup *pergroups;
3180 : Plan *outerPlan;
3181 : ExprContext *econtext;
3182 : TupleDesc scanDesc;
3183 : int max_aggno;
3184 : int max_transno;
3185 : int numaggrefs;
3186 : int numaggs;
3187 : int numtrans;
3188 : int phase;
3189 : int phaseidx;
3190 : ListCell *l;
3191 42122 : Bitmapset *all_grouped_cols = NULL;
3192 42122 : int numGroupingSets = 1;
3193 : int numPhases;
3194 : int numHashes;
3195 42122 : int i = 0;
3196 42122 : int j = 0;
3197 76876 : bool use_hashing = (node->aggstrategy == AGG_HASHED ||
3198 34754 : node->aggstrategy == AGG_MIXED);
3199 :
3200 : /* check for unsupported flags */
3201 : Assert(!(eflags & (EXEC_FLAG_BACKWARD | EXEC_FLAG_MARK)));
3202 :
3203 : /*
3204 : * create state structure
3205 : */
3206 42122 : aggstate = makeNode(AggState);
3207 42122 : aggstate->ss.ps.plan = (Plan *) node;
3208 42122 : aggstate->ss.ps.state = estate;
3209 42122 : aggstate->ss.ps.ExecProcNode = ExecAgg;
3210 :
3211 42122 : aggstate->aggs = NIL;
3212 42122 : aggstate->numaggs = 0;
3213 42122 : aggstate->numtrans = 0;
3214 42122 : aggstate->aggstrategy = node->aggstrategy;
3215 42122 : aggstate->aggsplit = node->aggsplit;
3216 42122 : aggstate->maxsets = 0;
3217 42122 : aggstate->projected_set = -1;
3218 42122 : aggstate->current_set = 0;
3219 42122 : aggstate->peragg = NULL;
3220 42122 : aggstate->pertrans = NULL;
3221 42122 : aggstate->curperagg = NULL;
3222 42122 : aggstate->curpertrans = NULL;
3223 42122 : aggstate->input_done = false;
3224 42122 : aggstate->agg_done = false;
3225 42122 : aggstate->pergroups = NULL;
3226 42122 : aggstate->grp_firstTuple = NULL;
3227 42122 : aggstate->sort_in = NULL;
3228 42122 : aggstate->sort_out = NULL;
3229 :
3230 : /*
3231 : * phases[0] always exists, but is dummy in sorted/plain mode
3232 : */
3233 42122 : numPhases = (use_hashing ? 1 : 2);
3234 42122 : numHashes = (use_hashing ? 1 : 0);
3235 :
3236 : /*
3237 : * Calculate the maximum number of grouping sets in any phase; this
3238 : * determines the size of some allocations. Also calculate the number of
3239 : * phases, since all hashed/mixed nodes contribute to only a single phase.
3240 : */
3241 42122 : if (node->groupingSets)
3242 : {
3243 716 : numGroupingSets = list_length(node->groupingSets);
3244 :
3245 1534 : foreach(l, node->chain)
3246 : {
3247 818 : Agg *agg = lfirst(l);
3248 :
3249 818 : numGroupingSets = Max(numGroupingSets,
3250 : list_length(agg->groupingSets));
3251 :
3252 : /*
3253 : * additional AGG_HASHED aggs become part of phase 0, but all
3254 : * others add an extra phase.
3255 : */
3256 818 : if (agg->aggstrategy != AGG_HASHED)
3257 424 : ++numPhases;
3258 : else
3259 394 : ++numHashes;
3260 : }
3261 : }
3262 :
3263 42122 : aggstate->maxsets = numGroupingSets;
3264 42122 : aggstate->numphases = numPhases;
3265 :
3266 42122 : aggstate->aggcontexts = (ExprContext **)
3267 42122 : palloc0(sizeof(ExprContext *) * numGroupingSets);
3268 :
3269 : /*
3270 : * Create expression contexts. We need three or more, one for
3271 : * per-input-tuple processing, one for per-output-tuple processing, one
3272 : * for all the hashtables, and one for each grouping set. The per-tuple
3273 : * memory context of the per-grouping-set ExprContexts (aggcontexts)
3274 : * replaces the standalone memory context formerly used to hold transition
3275 : * values. We cheat a little by using ExecAssignExprContext() to build
3276 : * all of them.
3277 : *
3278 : * NOTE: the details of what is stored in aggcontexts and what is stored
3279 : * in the regular per-query memory context are driven by a simple
3280 : * decision: we want to reset the aggcontext at group boundaries (if not
3281 : * hashing) and in ExecReScanAgg to recover no-longer-wanted space.
3282 : */
3283 42122 : ExecAssignExprContext(estate, &aggstate->ss.ps);
3284 42122 : aggstate->tmpcontext = aggstate->ss.ps.ps_ExprContext;
3285 :
3286 85072 : for (i = 0; i < numGroupingSets; ++i)
3287 : {
3288 42950 : ExecAssignExprContext(estate, &aggstate->ss.ps);
3289 42950 : aggstate->aggcontexts[i] = aggstate->ss.ps.ps_ExprContext;
3290 : }
3291 :
3292 42122 : if (use_hashing)
3293 7564 : aggstate->hashcontext = CreateWorkExprContext(estate);
3294 :
3295 42122 : ExecAssignExprContext(estate, &aggstate->ss.ps);
3296 :
3297 : /*
3298 : * Initialize child nodes.
3299 : *
3300 : * If we are doing a hashed aggregation then the child plan does not need
3301 : * to handle REWIND efficiently; see ExecReScanAgg.
3302 : */
3303 42122 : if (node->aggstrategy == AGG_HASHED)
3304 7368 : eflags &= ~EXEC_FLAG_REWIND;
3305 42122 : outerPlan = outerPlan(node);
3306 42122 : outerPlanState(aggstate) = ExecInitNode(outerPlan, estate, eflags);
3307 :
3308 : /*
3309 : * initialize source tuple type.
3310 : */
3311 42122 : aggstate->ss.ps.outerops =
3312 42122 : ExecGetResultSlotOps(outerPlanState(&aggstate->ss),
3313 : &aggstate->ss.ps.outeropsfixed);
3314 42122 : aggstate->ss.ps.outeropsset = true;
3315 :
3316 42122 : ExecCreateScanSlotFromOuterPlan(estate, &aggstate->ss,
3317 : aggstate->ss.ps.outerops);
3318 42122 : scanDesc = aggstate->ss.ss_ScanTupleSlot->tts_tupleDescriptor;
3319 :
3320 : /*
3321 : * If there are more than two phases (including a potential dummy phase
3322 : * 0), input will be resorted using tuplesort. Need a slot for that.
3323 : */
3324 42122 : if (numPhases > 2)
3325 : {
3326 186 : aggstate->sort_slot = ExecInitExtraTupleSlot(estate, scanDesc,
3327 : &TTSOpsMinimalTuple);
3328 :
3329 : /*
3330 : * The output of the tuplesort, and the output from the outer child
3331 : * might not use the same type of slot. In most cases the child will
3332 : * be a Sort, and thus return a TTSOpsMinimalTuple type slot - but the
3333 : * input can also be presorted due an index, in which case it could be
3334 : * a different type of slot.
3335 : *
3336 : * XXX: For efficiency it would be good to instead/additionally
3337 : * generate expressions with corresponding settings of outerops* for
3338 : * the individual phases - deforming is often a bottleneck for
3339 : * aggregations with lots of rows per group. If there's multiple
3340 : * sorts, we know that all but the first use TTSOpsMinimalTuple (via
3341 : * the nodeAgg.c internal tuplesort).
3342 : */
3343 186 : if (aggstate->ss.ps.outeropsfixed &&
3344 186 : aggstate->ss.ps.outerops != &TTSOpsMinimalTuple)
3345 24 : aggstate->ss.ps.outeropsfixed = false;
3346 : }
3347 :
3348 : /*
3349 : * Initialize result type, slot and projection.
3350 : */
3351 42122 : ExecInitResultTupleSlotTL(&aggstate->ss.ps, &TTSOpsVirtual);
3352 42122 : ExecAssignProjectionInfo(&aggstate->ss.ps, NULL);
3353 :
3354 : /*
3355 : * initialize child expressions
3356 : *
3357 : * We expect the parser to have checked that no aggs contain other agg
3358 : * calls in their arguments (and just to be sure, we verify it again while
3359 : * initializing the plan node). This would make no sense under SQL
3360 : * semantics, and it's forbidden by the spec. Because it is true, we
3361 : * don't need to worry about evaluating the aggs in any particular order.
3362 : *
3363 : * Note: execExpr.c finds Aggrefs for us, and adds them to aggstate->aggs.
3364 : * Aggrefs in the qual are found here; Aggrefs in the targetlist are found
3365 : * during ExecAssignProjectionInfo, above.
3366 : */
3367 42122 : aggstate->ss.ps.qual =
3368 42122 : ExecInitQual(node->plan.qual, (PlanState *) aggstate);
3369 :
3370 : /*
3371 : * We should now have found all Aggrefs in the targetlist and quals.
3372 : */
3373 42122 : numaggrefs = list_length(aggstate->aggs);
3374 42122 : max_aggno = -1;
3375 42122 : max_transno = -1;
3376 87234 : foreach(l, aggstate->aggs)
3377 : {
3378 45112 : Aggref *aggref = (Aggref *) lfirst(l);
3379 :
3380 45112 : max_aggno = Max(max_aggno, aggref->aggno);
3381 45112 : max_transno = Max(max_transno, aggref->aggtransno);
3382 : }
3383 42122 : numaggs = max_aggno + 1;
3384 42122 : numtrans = max_transno + 1;
3385 :
3386 : /*
3387 : * For each phase, prepare grouping set data and fmgr lookup data for
3388 : * compare functions. Accumulate all_grouped_cols in passing.
3389 : */
3390 42122 : aggstate->phases = palloc0(numPhases * sizeof(AggStatePerPhaseData));
3391 :
3392 42122 : aggstate->num_hashes = numHashes;
3393 42122 : if (numHashes)
3394 : {
3395 7564 : aggstate->perhash = palloc0(sizeof(AggStatePerHashData) * numHashes);
3396 7564 : aggstate->phases[0].numsets = 0;
3397 7564 : aggstate->phases[0].gset_lengths = palloc(numHashes * sizeof(int));
3398 7564 : aggstate->phases[0].grouped_cols = palloc(numHashes * sizeof(Bitmapset *));
3399 : }
3400 :
3401 42122 : phase = 0;
3402 85062 : for (phaseidx = 0; phaseidx <= list_length(node->chain); ++phaseidx)
3403 : {
3404 : Agg *aggnode;
3405 : Sort *sortnode;
3406 :
3407 42940 : if (phaseidx > 0)
3408 : {
3409 818 : aggnode = list_nth_node(Agg, node->chain, phaseidx - 1);
3410 818 : sortnode = castNode(Sort, outerPlan(aggnode));
3411 : }
3412 : else
3413 : {
3414 42122 : aggnode = node;
3415 42122 : sortnode = NULL;
3416 : }
3417 :
3418 : Assert(phase <= 1 || sortnode);
3419 :
3420 42940 : if (aggnode->aggstrategy == AGG_HASHED
3421 35178 : || aggnode->aggstrategy == AGG_MIXED)
3422 : {
3423 7958 : AggStatePerPhase phasedata = &aggstate->phases[0];
3424 : AggStatePerHash perhash;
3425 7958 : Bitmapset *cols = NULL;
3426 :
3427 : Assert(phase == 0);
3428 7958 : i = phasedata->numsets++;
3429 7958 : perhash = &aggstate->perhash[i];
3430 :
3431 : /* phase 0 always points to the "real" Agg in the hash case */
3432 7958 : phasedata->aggnode = node;
3433 7958 : phasedata->aggstrategy = node->aggstrategy;
3434 :
3435 : /* but the actual Agg node representing this hash is saved here */
3436 7958 : perhash->aggnode = aggnode;
3437 :
3438 7958 : phasedata->gset_lengths[i] = perhash->numCols = aggnode->numCols;
3439 :
3440 23928 : for (j = 0; j < aggnode->numCols; ++j)
3441 15970 : cols = bms_add_member(cols, aggnode->grpColIdx[j]);
3442 :
3443 7958 : phasedata->grouped_cols[i] = cols;
3444 :
3445 7958 : all_grouped_cols = bms_add_members(all_grouped_cols, cols);
3446 7958 : continue;
3447 : }
3448 : else
3449 : {
3450 34982 : AggStatePerPhase phasedata = &aggstate->phases[++phase];
3451 : int num_sets;
3452 :
3453 34982 : phasedata->numsets = num_sets = list_length(aggnode->groupingSets);
3454 :
3455 34982 : if (num_sets)
3456 : {
3457 862 : phasedata->gset_lengths = palloc(num_sets * sizeof(int));
3458 862 : phasedata->grouped_cols = palloc(num_sets * sizeof(Bitmapset *));
3459 :
3460 862 : i = 0;
3461 2624 : foreach(l, aggnode->groupingSets)
3462 : {
3463 1762 : int current_length = list_length(lfirst(l));
3464 1762 : Bitmapset *cols = NULL;
3465 :
3466 : /* planner forces this to be correct */
3467 3498 : for (j = 0; j < current_length; ++j)
3468 1736 : cols = bms_add_member(cols, aggnode->grpColIdx[j]);
3469 :
3470 1762 : phasedata->grouped_cols[i] = cols;
3471 1762 : phasedata->gset_lengths[i] = current_length;
3472 :
3473 1762 : ++i;
3474 : }
3475 :
3476 862 : all_grouped_cols = bms_add_members(all_grouped_cols,
3477 862 : phasedata->grouped_cols[0]);
3478 : }
3479 : else
3480 : {
3481 : Assert(phaseidx == 0);
3482 :
3483 34120 : phasedata->gset_lengths = NULL;
3484 34120 : phasedata->grouped_cols = NULL;
3485 : }
3486 :
3487 : /*
3488 : * If we are grouping, precompute fmgr lookup data for inner loop.
3489 : */
3490 34982 : if (aggnode->aggstrategy == AGG_SORTED)
3491 : {
3492 : /*
3493 : * Build a separate function for each subset of columns that
3494 : * need to be compared.
3495 : */
3496 2150 : phasedata->eqfunctions =
3497 2150 : (ExprState **) palloc0(aggnode->numCols * sizeof(ExprState *));
3498 :
3499 : /* for each grouping set */
3500 3670 : for (int k = 0; k < phasedata->numsets; k++)
3501 : {
3502 1520 : int length = phasedata->gset_lengths[k];
3503 :
3504 : /* nothing to do for empty grouping set */
3505 1520 : if (length == 0)
3506 326 : continue;
3507 :
3508 : /* if we already had one of this length, it'll do */
3509 1194 : if (phasedata->eqfunctions[length - 1] != NULL)
3510 138 : continue;
3511 :
3512 1056 : phasedata->eqfunctions[length - 1] =
3513 1056 : execTuplesMatchPrepare(scanDesc,
3514 : length,
3515 1056 : aggnode->grpColIdx,
3516 1056 : aggnode->grpOperators,
3517 1056 : aggnode->grpCollations,
3518 : (PlanState *) aggstate);
3519 : }
3520 :
3521 : /* and for all grouped columns, unless already computed */
3522 2150 : if (aggnode->numCols > 0 &&
3523 2056 : phasedata->eqfunctions[aggnode->numCols - 1] == NULL)
3524 : {
3525 1364 : phasedata->eqfunctions[aggnode->numCols - 1] =
3526 1364 : execTuplesMatchPrepare(scanDesc,
3527 : aggnode->numCols,
3528 1364 : aggnode->grpColIdx,
3529 1364 : aggnode->grpOperators,
3530 1364 : aggnode->grpCollations,
3531 : (PlanState *) aggstate);
3532 : }
3533 : }
3534 :
3535 34982 : phasedata->aggnode = aggnode;
3536 34982 : phasedata->aggstrategy = aggnode->aggstrategy;
3537 34982 : phasedata->sortnode = sortnode;
3538 : }
3539 : }
3540 :
3541 : /*
3542 : * Convert all_grouped_cols to a descending-order list.
3543 : */
3544 42122 : i = -1;
3545 58646 : while ((i = bms_next_member(all_grouped_cols, i)) >= 0)
3546 16524 : aggstate->all_grouped_cols = lcons_int(i, aggstate->all_grouped_cols);
3547 :
3548 : /*
3549 : * Set up aggregate-result storage in the output expr context, and also
3550 : * allocate my private per-agg working storage
3551 : */
3552 42122 : econtext = aggstate->ss.ps.ps_ExprContext;
3553 42122 : econtext->ecxt_aggvalues = (Datum *) palloc0(sizeof(Datum) * numaggs);
3554 42122 : econtext->ecxt_aggnulls = (bool *) palloc0(sizeof(bool) * numaggs);
3555 :
3556 42122 : peraggs = (AggStatePerAgg) palloc0(sizeof(AggStatePerAggData) * numaggs);
3557 42122 : pertransstates = (AggStatePerTrans) palloc0(sizeof(AggStatePerTransData) * numtrans);
3558 :
3559 42122 : aggstate->peragg = peraggs;
3560 42122 : aggstate->pertrans = pertransstates;
3561 :
3562 :
3563 42122 : aggstate->all_pergroups =
3564 42122 : (AggStatePerGroup *) palloc0(sizeof(AggStatePerGroup)
3565 42122 : * (numGroupingSets + numHashes));
3566 42122 : pergroups = aggstate->all_pergroups;
3567 :
3568 42122 : if (node->aggstrategy != AGG_HASHED)
3569 : {
3570 70336 : for (i = 0; i < numGroupingSets; i++)
3571 : {
3572 35582 : pergroups[i] = (AggStatePerGroup) palloc0(sizeof(AggStatePerGroupData)
3573 : * numaggs);
3574 : }
3575 :
3576 34754 : aggstate->pergroups = pergroups;
3577 34754 : pergroups += numGroupingSets;
3578 : }
3579 :
3580 : /*
3581 : * Hashing can only appear in the initial phase.
3582 : */
3583 42122 : if (use_hashing)
3584 : {
3585 7564 : Plan *outerplan = outerPlan(node);
3586 7564 : uint64 totalGroups = 0;
3587 :
3588 7564 : aggstate->hash_metacxt = AllocSetContextCreate(aggstate->ss.ps.state->es_query_cxt,
3589 : "HashAgg meta context",
3590 : ALLOCSET_DEFAULT_SIZES);
3591 7564 : aggstate->hash_spill_rslot = ExecInitExtraTupleSlot(estate, scanDesc,
3592 : &TTSOpsMinimalTuple);
3593 7564 : aggstate->hash_spill_wslot = ExecInitExtraTupleSlot(estate, scanDesc,
3594 : &TTSOpsVirtual);
3595 :
3596 : /* this is an array of pointers, not structures */
3597 7564 : aggstate->hash_pergroup = pergroups;
3598 :
3599 15128 : aggstate->hashentrysize = hash_agg_entry_size(aggstate->numtrans,
3600 7564 : outerplan->plan_width,
3601 : node->transitionSpace);
3602 :
3603 : /*
3604 : * Consider all of the grouping sets together when setting the limits
3605 : * and estimating the number of partitions. This can be inaccurate
3606 : * when there is more than one grouping set, but should still be
3607 : * reasonable.
3608 : */
3609 15522 : for (int k = 0; k < aggstate->num_hashes; k++)
3610 7958 : totalGroups += aggstate->perhash[k].aggnode->numGroups;
3611 :
3612 7564 : hash_agg_set_limits(aggstate->hashentrysize, totalGroups, 0,
3613 : &aggstate->hash_mem_limit,
3614 : &aggstate->hash_ngroups_limit,
3615 : &aggstate->hash_planned_partitions);
3616 7564 : find_hash_columns(aggstate);
3617 :
3618 : /* Skip massive memory allocation if we are just doing EXPLAIN */
3619 7564 : if (!(eflags & EXEC_FLAG_EXPLAIN_ONLY))
3620 6464 : build_hash_tables(aggstate);
3621 :
3622 7564 : aggstate->table_filled = false;
3623 :
3624 : /* Initialize this to 1, meaning nothing spilled, yet */
3625 7564 : aggstate->hash_batches_used = 1;
3626 : }
3627 :
3628 : /*
3629 : * Initialize current phase-dependent values to initial phase. The initial
3630 : * phase is 1 (first sort pass) for all strategies that use sorting (if
3631 : * hashing is being done too, then phase 0 is processed last); but if only
3632 : * hashing is being done, then phase 0 is all there is.
3633 : */
3634 42122 : if (node->aggstrategy == AGG_HASHED)
3635 : {
3636 7368 : aggstate->current_phase = 0;
3637 7368 : initialize_phase(aggstate, 0);
3638 7368 : select_current_set(aggstate, 0, true);
3639 : }
3640 : else
3641 : {
3642 34754 : aggstate->current_phase = 1;
3643 34754 : initialize_phase(aggstate, 1);
3644 34754 : select_current_set(aggstate, 0, false);
3645 : }
3646 :
3647 : /*
3648 : * Perform lookups of aggregate function info, and initialize the
3649 : * unchanging fields of the per-agg and per-trans data.
3650 : */
3651 87228 : foreach(l, aggstate->aggs)
3652 : {
3653 45112 : Aggref *aggref = lfirst(l);
3654 : AggStatePerAgg peragg;
3655 : AggStatePerTrans pertrans;
3656 : Oid aggTransFnInputTypes[FUNC_MAX_ARGS];
3657 : int numAggTransFnArgs;
3658 : int numDirectArgs;
3659 : HeapTuple aggTuple;
3660 : Form_pg_aggregate aggform;
3661 : AclResult aclresult;
3662 : Oid finalfn_oid;
3663 : Oid serialfn_oid,
3664 : deserialfn_oid;
3665 : Oid aggOwner;
3666 : Expr *finalfnexpr;
3667 : Oid aggtranstype;
3668 :
3669 : /* Planner should have assigned aggregate to correct level */
3670 : Assert(aggref->agglevelsup == 0);
3671 : /* ... and the split mode should match */
3672 : Assert(aggref->aggsplit == aggstate->aggsplit);
3673 :
3674 45112 : peragg = &peraggs[aggref->aggno];
3675 :
3676 : /* Check if we initialized the state for this aggregate already. */
3677 45112 : if (peragg->aggref != NULL)
3678 430 : continue;
3679 :
3680 44682 : peragg->aggref = aggref;
3681 44682 : peragg->transno = aggref->aggtransno;
3682 :
3683 : /* Fetch the pg_aggregate row */
3684 44682 : aggTuple = SearchSysCache1(AGGFNOID,
3685 : ObjectIdGetDatum(aggref->aggfnoid));
3686 44682 : if (!HeapTupleIsValid(aggTuple))
3687 0 : elog(ERROR, "cache lookup failed for aggregate %u",
3688 : aggref->aggfnoid);
3689 44682 : aggform = (Form_pg_aggregate) GETSTRUCT(aggTuple);
3690 :
3691 : /* Check permission to call aggregate function */
3692 44682 : aclresult = object_aclcheck(ProcedureRelationId, aggref->aggfnoid, GetUserId(),
3693 : ACL_EXECUTE);
3694 44682 : if (aclresult != ACLCHECK_OK)
3695 6 : aclcheck_error(aclresult, OBJECT_AGGREGATE,
3696 6 : get_func_name(aggref->aggfnoid));
3697 44676 : InvokeFunctionExecuteHook(aggref->aggfnoid);
3698 :
3699 : /* planner recorded transition state type in the Aggref itself */
3700 44676 : aggtranstype = aggref->aggtranstype;
3701 : Assert(OidIsValid(aggtranstype));
3702 :
3703 : /* Final function only required if we're finalizing the aggregates */
3704 44676 : if (DO_AGGSPLIT_SKIPFINAL(aggstate->aggsplit))
3705 4208 : peragg->finalfn_oid = finalfn_oid = InvalidOid;
3706 : else
3707 40468 : peragg->finalfn_oid = finalfn_oid = aggform->aggfinalfn;
3708 :
3709 44676 : serialfn_oid = InvalidOid;
3710 44676 : deserialfn_oid = InvalidOid;
3711 :
3712 : /*
3713 : * Check if serialization/deserialization is required. We only do it
3714 : * for aggregates that have transtype INTERNAL.
3715 : */
3716 44676 : if (aggtranstype == INTERNALOID)
3717 : {
3718 : /*
3719 : * The planner should only have generated a serialize agg node if
3720 : * every aggregate with an INTERNAL state has a serialization
3721 : * function. Verify that.
3722 : */
3723 18018 : if (DO_AGGSPLIT_SERIALIZE(aggstate->aggsplit))
3724 : {
3725 : /* serialization only valid when not running finalfn */
3726 : Assert(DO_AGGSPLIT_SKIPFINAL(aggstate->aggsplit));
3727 :
3728 336 : if (!OidIsValid(aggform->aggserialfn))
3729 0 : elog(ERROR, "serialfunc not provided for serialization aggregation");
3730 336 : serialfn_oid = aggform->aggserialfn;
3731 : }
3732 :
3733 : /* Likewise for deserialization functions */
3734 18018 : if (DO_AGGSPLIT_DESERIALIZE(aggstate->aggsplit))
3735 : {
3736 : /* deserialization only valid when combining states */
3737 : Assert(DO_AGGSPLIT_COMBINE(aggstate->aggsplit));
3738 :
3739 120 : if (!OidIsValid(aggform->aggdeserialfn))
3740 0 : elog(ERROR, "deserialfunc not provided for deserialization aggregation");
3741 120 : deserialfn_oid = aggform->aggdeserialfn;
3742 : }
3743 : }
3744 :
3745 : /* Check that aggregate owner has permission to call component fns */
3746 : {
3747 : HeapTuple procTuple;
3748 :
3749 44676 : procTuple = SearchSysCache1(PROCOID,
3750 : ObjectIdGetDatum(aggref->aggfnoid));
3751 44676 : if (!HeapTupleIsValid(procTuple))
3752 0 : elog(ERROR, "cache lookup failed for function %u",
3753 : aggref->aggfnoid);
3754 44676 : aggOwner = ((Form_pg_proc) GETSTRUCT(procTuple))->proowner;
3755 44676 : ReleaseSysCache(procTuple);
3756 :
3757 44676 : if (OidIsValid(finalfn_oid))
3758 : {
3759 19412 : aclresult = object_aclcheck(ProcedureRelationId, finalfn_oid, aggOwner,
3760 : ACL_EXECUTE);
3761 19412 : if (aclresult != ACLCHECK_OK)
3762 0 : aclcheck_error(aclresult, OBJECT_FUNCTION,
3763 0 : get_func_name(finalfn_oid));
3764 19412 : InvokeFunctionExecuteHook(finalfn_oid);
3765 : }
3766 44676 : if (OidIsValid(serialfn_oid))
3767 : {
3768 336 : aclresult = object_aclcheck(ProcedureRelationId, serialfn_oid, aggOwner,
3769 : ACL_EXECUTE);
3770 336 : if (aclresult != ACLCHECK_OK)
3771 0 : aclcheck_error(aclresult, OBJECT_FUNCTION,
3772 0 : get_func_name(serialfn_oid));
3773 336 : InvokeFunctionExecuteHook(serialfn_oid);
3774 : }
3775 44676 : if (OidIsValid(deserialfn_oid))
3776 : {
3777 120 : aclresult = object_aclcheck(ProcedureRelationId, deserialfn_oid, aggOwner,
3778 : ACL_EXECUTE);
3779 120 : if (aclresult != ACLCHECK_OK)
3780 0 : aclcheck_error(aclresult, OBJECT_FUNCTION,
3781 0 : get_func_name(deserialfn_oid));
3782 120 : InvokeFunctionExecuteHook(deserialfn_oid);
3783 : }
3784 : }
3785 :
3786 : /*
3787 : * Get actual datatypes of the (nominal) aggregate inputs. These
3788 : * could be different from the agg's declared input types, when the
3789 : * agg accepts ANY or a polymorphic type.
3790 : */
3791 44676 : numAggTransFnArgs = get_aggregate_argtypes(aggref,
3792 : aggTransFnInputTypes);
3793 :
3794 : /* Count the "direct" arguments, if any */
3795 44676 : numDirectArgs = list_length(aggref->aggdirectargs);
3796 :
3797 : /* Detect how many arguments to pass to the finalfn */
3798 44676 : if (aggform->aggfinalextra)
3799 12748 : peragg->numFinalArgs = numAggTransFnArgs + 1;
3800 : else
3801 31928 : peragg->numFinalArgs = numDirectArgs + 1;
3802 :
3803 : /* Initialize any direct-argument expressions */
3804 44676 : peragg->aggdirectargs = ExecInitExprList(aggref->aggdirectargs,
3805 : (PlanState *) aggstate);
3806 :
3807 : /*
3808 : * build expression trees using actual argument & result types for the
3809 : * finalfn, if it exists and is required.
3810 : */
3811 44676 : if (OidIsValid(finalfn_oid))
3812 : {
3813 19412 : build_aggregate_finalfn_expr(aggTransFnInputTypes,
3814 : peragg->numFinalArgs,
3815 : aggtranstype,
3816 : aggref->aggtype,
3817 : aggref->inputcollid,
3818 : finalfn_oid,
3819 : &finalfnexpr);
3820 19412 : fmgr_info(finalfn_oid, &peragg->finalfn);
3821 19412 : fmgr_info_set_expr((Node *) finalfnexpr, &peragg->finalfn);
3822 : }
3823 :
3824 : /* get info about the output value's datatype */
3825 44676 : get_typlenbyval(aggref->aggtype,
3826 : &peragg->resulttypeLen,
3827 : &peragg->resulttypeByVal);
3828 :
3829 : /*
3830 : * Build working state for invoking the transition function, if we
3831 : * haven't done it already.
3832 : */
3833 44676 : pertrans = &pertransstates[aggref->aggtransno];
3834 44676 : if (pertrans->aggref == NULL)
3835 : {
3836 : Datum textInitVal;
3837 : Datum initValue;
3838 : bool initValueIsNull;
3839 : Oid transfn_oid;
3840 :
3841 : /*
3842 : * If this aggregation is performing state combines, then instead
3843 : * of using the transition function, we'll use the combine
3844 : * function.
3845 : */
3846 44418 : if (DO_AGGSPLIT_COMBINE(aggstate->aggsplit))
3847 : {
3848 1340 : transfn_oid = aggform->aggcombinefn;
3849 :
3850 : /* If not set then the planner messed up */
3851 1340 : if (!OidIsValid(transfn_oid))
3852 0 : elog(ERROR, "combinefn not set for aggregate function");
3853 : }
3854 : else
3855 43078 : transfn_oid = aggform->aggtransfn;
3856 :
3857 44418 : aclresult = object_aclcheck(ProcedureRelationId, transfn_oid, aggOwner, ACL_EXECUTE);
3858 44418 : if (aclresult != ACLCHECK_OK)
3859 0 : aclcheck_error(aclresult, OBJECT_FUNCTION,
3860 0 : get_func_name(transfn_oid));
3861 44418 : InvokeFunctionExecuteHook(transfn_oid);
3862 :
3863 : /*
3864 : * initval is potentially null, so don't try to access it as a
3865 : * struct field. Must do it the hard way with SysCacheGetAttr.
3866 : */
3867 44418 : textInitVal = SysCacheGetAttr(AGGFNOID, aggTuple,
3868 : Anum_pg_aggregate_agginitval,
3869 : &initValueIsNull);
3870 44418 : if (initValueIsNull)
3871 25982 : initValue = (Datum) 0;
3872 : else
3873 18436 : initValue = GetAggInitVal(textInitVal, aggtranstype);
3874 :
3875 44418 : if (DO_AGGSPLIT_COMBINE(aggstate->aggsplit))
3876 : {
3877 1340 : Oid combineFnInputTypes[] = {aggtranstype,
3878 : aggtranstype};
3879 :
3880 : /*
3881 : * When combining there's only one input, the to-be-combined
3882 : * transition value. The transition value is not counted
3883 : * here.
3884 : */
3885 1340 : pertrans->numTransInputs = 1;
3886 :
3887 : /* aggcombinefn always has two arguments of aggtranstype */
3888 1340 : build_pertrans_for_aggref(pertrans, aggstate, estate,
3889 : aggref, transfn_oid, aggtranstype,
3890 : serialfn_oid, deserialfn_oid,
3891 : initValue, initValueIsNull,
3892 : combineFnInputTypes, 2);
3893 :
3894 : /*
3895 : * Ensure that a combine function to combine INTERNAL states
3896 : * is not strict. This should have been checked during CREATE
3897 : * AGGREGATE, but the strict property could have been changed
3898 : * since then.
3899 : */
3900 1340 : if (pertrans->transfn.fn_strict && aggtranstype == INTERNALOID)
3901 0 : ereport(ERROR,
3902 : (errcode(ERRCODE_INVALID_FUNCTION_DEFINITION),
3903 : errmsg("combine function with transition type %s must not be declared STRICT",
3904 : format_type_be(aggtranstype))));
3905 : }
3906 : else
3907 : {
3908 : /* Detect how many arguments to pass to the transfn */
3909 43078 : if (AGGKIND_IS_ORDERED_SET(aggref->aggkind))
3910 252 : pertrans->numTransInputs = list_length(aggref->args);
3911 : else
3912 42826 : pertrans->numTransInputs = numAggTransFnArgs;
3913 :
3914 43078 : build_pertrans_for_aggref(pertrans, aggstate, estate,
3915 : aggref, transfn_oid, aggtranstype,
3916 : serialfn_oid, deserialfn_oid,
3917 : initValue, initValueIsNull,
3918 : aggTransFnInputTypes,
3919 : numAggTransFnArgs);
3920 :
3921 : /*
3922 : * If the transfn is strict and the initval is NULL, make sure
3923 : * input type and transtype are the same (or at least
3924 : * binary-compatible), so that it's OK to use the first
3925 : * aggregated input value as the initial transValue. This
3926 : * should have been checked at agg definition time, but we
3927 : * must check again in case the transfn's strictness property
3928 : * has been changed.
3929 : */
3930 43078 : if (pertrans->transfn.fn_strict && pertrans->initValueIsNull)
3931 : {
3932 4766 : if (numAggTransFnArgs <= numDirectArgs ||
3933 4766 : !IsBinaryCoercible(aggTransFnInputTypes[numDirectArgs],
3934 : aggtranstype))
3935 0 : ereport(ERROR,
3936 : (errcode(ERRCODE_INVALID_FUNCTION_DEFINITION),
3937 : errmsg("aggregate %u needs to have compatible input type and transition type",
3938 : aggref->aggfnoid)));
3939 : }
3940 : }
3941 : }
3942 : else
3943 258 : pertrans->aggshared = true;
3944 44676 : ReleaseSysCache(aggTuple);
3945 : }
3946 :
3947 : /*
3948 : * Update aggstate->numaggs to be the number of unique aggregates found.
3949 : * Also set numstates to the number of unique transition states found.
3950 : */
3951 42116 : aggstate->numaggs = numaggs;
3952 42116 : aggstate->numtrans = numtrans;
3953 :
3954 : /*
3955 : * Last, check whether any more aggregates got added onto the node while
3956 : * we processed the expressions for the aggregate arguments (including not
3957 : * only the regular arguments and FILTER expressions handled immediately
3958 : * above, but any direct arguments we might've handled earlier). If so,
3959 : * we have nested aggregate functions, which is semantically nonsensical,
3960 : * so complain. (This should have been caught by the parser, so we don't
3961 : * need to work hard on a helpful error message; but we defend against it
3962 : * here anyway, just to be sure.)
3963 : */
3964 42116 : if (numaggrefs != list_length(aggstate->aggs))
3965 0 : ereport(ERROR,
3966 : (errcode(ERRCODE_GROUPING_ERROR),
3967 : errmsg("aggregate function calls cannot be nested")));
3968 :
3969 : /*
3970 : * Build expressions doing all the transition work at once. We build a
3971 : * different one for each phase, as the number of transition function
3972 : * invocation can differ between phases. Note this'll work both for
3973 : * transition and combination functions (although there'll only be one
3974 : * phase in the latter case).
3975 : */
3976 119208 : for (phaseidx = 0; phaseidx < aggstate->numphases; phaseidx++)
3977 : {
3978 77092 : AggStatePerPhase phase = &aggstate->phases[phaseidx];
3979 77092 : bool dohash = false;
3980 77092 : bool dosort = false;
3981 :
3982 : /* phase 0 doesn't necessarily exist */
3983 77092 : if (!phase->aggnode)
3984 34552 : continue;
3985 :
3986 42540 : if (aggstate->aggstrategy == AGG_MIXED && phaseidx == 1)
3987 : {
3988 : /*
3989 : * Phase one, and only phase one, in a mixed agg performs both
3990 : * sorting and aggregation.
3991 : */
3992 196 : dohash = true;
3993 196 : dosort = true;
3994 : }
3995 42344 : else if (aggstate->aggstrategy == AGG_MIXED && phaseidx == 0)
3996 : {
3997 : /*
3998 : * No need to compute a transition function for an AGG_MIXED phase
3999 : * 0 - the contents of the hashtables will have been computed
4000 : * during phase 1.
4001 : */
4002 196 : continue;
4003 : }
4004 42148 : else if (phase->aggstrategy == AGG_PLAIN ||
4005 9456 : phase->aggstrategy == AGG_SORTED)
4006 : {
4007 34780 : dohash = false;
4008 34780 : dosort = true;
4009 : }
4010 7368 : else if (phase->aggstrategy == AGG_HASHED)
4011 : {
4012 7368 : dohash = true;
4013 7368 : dosort = false;
4014 : }
4015 : else
4016 : Assert(false);
4017 :
4018 42344 : phase->evaltrans = ExecBuildAggTrans(aggstate, phase, dosort, dohash,
4019 : false);
4020 :
4021 : /* cache compiled expression for outer slot without NULL check */
4022 42344 : phase->evaltrans_cache[0][0] = phase->evaltrans;
4023 : }
4024 :
4025 42116 : return aggstate;
4026 : }
4027 :
4028 : /*
4029 : * Build the state needed to calculate a state value for an aggregate.
4030 : *
4031 : * This initializes all the fields in 'pertrans'. 'aggref' is the aggregate
4032 : * to initialize the state for. 'transfn_oid', 'aggtranstype', and the rest
4033 : * of the arguments could be calculated from 'aggref', but the caller has
4034 : * calculated them already, so might as well pass them.
4035 : *
4036 : * 'transfn_oid' may be either the Oid of the aggtransfn or the aggcombinefn.
4037 : */
4038 : static void
4039 44418 : build_pertrans_for_aggref(AggStatePerTrans pertrans,
4040 : AggState *aggstate, EState *estate,
4041 : Aggref *aggref,
4042 : Oid transfn_oid, Oid aggtranstype,
4043 : Oid aggserialfn, Oid aggdeserialfn,
4044 : Datum initValue, bool initValueIsNull,
4045 : Oid *inputTypes, int numArguments)
4046 : {
4047 44418 : int numGroupingSets = Max(aggstate->maxsets, 1);
4048 : Expr *transfnexpr;
4049 : int numTransArgs;
4050 44418 : Expr *serialfnexpr = NULL;
4051 44418 : Expr *deserialfnexpr = NULL;
4052 : ListCell *lc;
4053 : int numInputs;
4054 : int numDirectArgs;
4055 : List *sortlist;
4056 : int numSortCols;
4057 : int numDistinctCols;
4058 : int i;
4059 :
4060 : /* Begin filling in the pertrans data */
4061 44418 : pertrans->aggref = aggref;
4062 44418 : pertrans->aggshared = false;
4063 44418 : pertrans->aggCollation = aggref->inputcollid;
4064 44418 : pertrans->transfn_oid = transfn_oid;
4065 44418 : pertrans->serialfn_oid = aggserialfn;
4066 44418 : pertrans->deserialfn_oid = aggdeserialfn;
4067 44418 : pertrans->initValue = initValue;
4068 44418 : pertrans->initValueIsNull = initValueIsNull;
4069 :
4070 : /* Count the "direct" arguments, if any */
4071 44418 : numDirectArgs = list_length(aggref->aggdirectargs);
4072 :
4073 : /* Count the number of aggregated input columns */
4074 44418 : pertrans->numInputs = numInputs = list_length(aggref->args);
4075 :
4076 44418 : pertrans->aggtranstype = aggtranstype;
4077 :
4078 : /* account for the current transition state */
4079 44418 : numTransArgs = pertrans->numTransInputs + 1;
4080 :
4081 : /*
4082 : * Set up infrastructure for calling the transfn. Note that invtransfn is
4083 : * not needed here.
4084 : */
4085 44418 : build_aggregate_transfn_expr(inputTypes,
4086 : numArguments,
4087 : numDirectArgs,
4088 44418 : aggref->aggvariadic,
4089 : aggtranstype,
4090 : aggref->inputcollid,
4091 : transfn_oid,
4092 : InvalidOid,
4093 : &transfnexpr,
4094 : NULL);
4095 :
4096 44418 : fmgr_info(transfn_oid, &pertrans->transfn);
4097 44418 : fmgr_info_set_expr((Node *) transfnexpr, &pertrans->transfn);
4098 :
4099 44418 : pertrans->transfn_fcinfo =
4100 44418 : (FunctionCallInfo) palloc(SizeForFunctionCallInfo(numTransArgs));
4101 44418 : InitFunctionCallInfoData(*pertrans->transfn_fcinfo,
4102 : &pertrans->transfn,
4103 : numTransArgs,
4104 : pertrans->aggCollation,
4105 : (void *) aggstate, NULL);
4106 :
4107 : /* get info about the state value's datatype */
4108 44418 : get_typlenbyval(aggtranstype,
4109 : &pertrans->transtypeLen,
4110 : &pertrans->transtypeByVal);
4111 :
4112 44418 : if (OidIsValid(aggserialfn))
4113 : {
4114 336 : build_aggregate_serialfn_expr(aggserialfn,
4115 : &serialfnexpr);
4116 336 : fmgr_info(aggserialfn, &pertrans->serialfn);
4117 336 : fmgr_info_set_expr((Node *) serialfnexpr, &pertrans->serialfn);
4118 :
4119 336 : pertrans->serialfn_fcinfo =
4120 336 : (FunctionCallInfo) palloc(SizeForFunctionCallInfo(1));
4121 336 : InitFunctionCallInfoData(*pertrans->serialfn_fcinfo,
4122 : &pertrans->serialfn,
4123 : 1,
4124 : InvalidOid,
4125 : (void *) aggstate, NULL);
4126 : }
4127 :
4128 44418 : if (OidIsValid(aggdeserialfn))
4129 : {
4130 120 : build_aggregate_deserialfn_expr(aggdeserialfn,
4131 : &deserialfnexpr);
4132 120 : fmgr_info(aggdeserialfn, &pertrans->deserialfn);
4133 120 : fmgr_info_set_expr((Node *) deserialfnexpr, &pertrans->deserialfn);
4134 :
4135 120 : pertrans->deserialfn_fcinfo =
4136 120 : (FunctionCallInfo) palloc(SizeForFunctionCallInfo(2));
4137 120 : InitFunctionCallInfoData(*pertrans->deserialfn_fcinfo,
4138 : &pertrans->deserialfn,
4139 : 2,
4140 : InvalidOid,
4141 : (void *) aggstate, NULL);
4142 : }
4143 :
4144 : /*
4145 : * If we're doing either DISTINCT or ORDER BY for a plain agg, then we
4146 : * have a list of SortGroupClause nodes; fish out the data in them and
4147 : * stick them into arrays. We ignore ORDER BY for an ordered-set agg,
4148 : * however; the agg's transfn and finalfn are responsible for that.
4149 : *
4150 : * When the planner has set the aggpresorted flag, the input to the
4151 : * aggregate is already correctly sorted. For ORDER BY aggregates we can
4152 : * simply treat these as normal aggregates. For presorted DISTINCT
4153 : * aggregates an extra step must be added to remove duplicate consecutive
4154 : * inputs.
4155 : *
4156 : * Note that by construction, if there is a DISTINCT clause then the ORDER
4157 : * BY clause is a prefix of it (see transformDistinctClause).
4158 : */
4159 44418 : if (AGGKIND_IS_ORDERED_SET(aggref->aggkind))
4160 : {
4161 252 : sortlist = NIL;
4162 252 : numSortCols = numDistinctCols = 0;
4163 252 : pertrans->aggsortrequired = false;
4164 : }
4165 44166 : else if (aggref->aggpresorted && aggref->aggdistinct == NIL)
4166 : {
4167 1432 : sortlist = NIL;
4168 1432 : numSortCols = numDistinctCols = 0;
4169 1432 : pertrans->aggsortrequired = false;
4170 : }
4171 42734 : else if (aggref->aggdistinct)
4172 : {
4173 528 : sortlist = aggref->aggdistinct;
4174 528 : numSortCols = numDistinctCols = list_length(sortlist);
4175 : Assert(numSortCols >= list_length(aggref->aggorder));
4176 528 : pertrans->aggsortrequired = !aggref->aggpresorted;
4177 : }
4178 : else
4179 : {
4180 42206 : sortlist = aggref->aggorder;
4181 42206 : numSortCols = list_length(sortlist);
4182 42206 : numDistinctCols = 0;
4183 42206 : pertrans->aggsortrequired = (numSortCols > 0);
4184 : }
4185 :
4186 44418 : pertrans->numSortCols = numSortCols;
4187 44418 : pertrans->numDistinctCols = numDistinctCols;
4188 :
4189 : /*
4190 : * If we have either sorting or filtering to do, create a tupledesc and
4191 : * slot corresponding to the aggregated inputs (including sort
4192 : * expressions) of the agg.
4193 : */
4194 44418 : if (numSortCols > 0 || aggref->aggfilter)
4195 : {
4196 1284 : pertrans->sortdesc = ExecTypeFromTL(aggref->args);
4197 1284 : pertrans->sortslot =
4198 1284 : ExecInitExtraTupleSlot(estate, pertrans->sortdesc,
4199 : &TTSOpsMinimalTuple);
4200 : }
4201 :
4202 44418 : if (numSortCols > 0)
4203 : {
4204 : /*
4205 : * We don't implement DISTINCT or ORDER BY aggs in the HASHED case
4206 : * (yet)
4207 : */
4208 : Assert(aggstate->aggstrategy != AGG_HASHED && aggstate->aggstrategy != AGG_MIXED);
4209 :
4210 : /* ORDER BY aggregates are not supported with partial aggregation */
4211 : Assert(!DO_AGGSPLIT_COMBINE(aggstate->aggsplit));
4212 :
4213 : /* If we have only one input, we need its len/byval info. */
4214 654 : if (numInputs == 1)
4215 : {
4216 534 : get_typlenbyval(inputTypes[numDirectArgs],
4217 : &pertrans->inputtypeLen,
4218 : &pertrans->inputtypeByVal);
4219 : }
4220 120 : else if (numDistinctCols > 0)
4221 : {
4222 : /* we will need an extra slot to store prior values */
4223 90 : pertrans->uniqslot =
4224 90 : ExecInitExtraTupleSlot(estate, pertrans->sortdesc,
4225 : &TTSOpsMinimalTuple);
4226 : }
4227 :
4228 : /* Extract the sort information for use later */
4229 654 : pertrans->sortColIdx =
4230 654 : (AttrNumber *) palloc(numSortCols * sizeof(AttrNumber));
4231 654 : pertrans->sortOperators =
4232 654 : (Oid *) palloc(numSortCols * sizeof(Oid));
4233 654 : pertrans->sortCollations =
4234 654 : (Oid *) palloc(numSortCols * sizeof(Oid));
4235 654 : pertrans->sortNullsFirst =
4236 654 : (bool *) palloc(numSortCols * sizeof(bool));
4237 :
4238 654 : i = 0;
4239 1488 : foreach(lc, sortlist)
4240 : {
4241 834 : SortGroupClause *sortcl = (SortGroupClause *) lfirst(lc);
4242 834 : TargetEntry *tle = get_sortgroupclause_tle(sortcl, aggref->args);
4243 :
4244 : /* the parser should have made sure of this */
4245 : Assert(OidIsValid(sortcl->sortop));
4246 :
4247 834 : pertrans->sortColIdx[i] = tle->resno;
4248 834 : pertrans->sortOperators[i] = sortcl->sortop;
4249 834 : pertrans->sortCollations[i] = exprCollation((Node *) tle->expr);
4250 834 : pertrans->sortNullsFirst[i] = sortcl->nulls_first;
4251 834 : i++;
4252 : }
4253 : Assert(i == numSortCols);
4254 : }
4255 :
4256 44418 : if (aggref->aggdistinct)
4257 : {
4258 : Oid *ops;
4259 :
4260 : Assert(numArguments > 0);
4261 : Assert(list_length(aggref->aggdistinct) == numDistinctCols);
4262 :
4263 528 : ops = palloc(numDistinctCols * sizeof(Oid));
4264 :
4265 528 : i = 0;
4266 1218 : foreach(lc, aggref->aggdistinct)
4267 690 : ops[i++] = ((SortGroupClause *) lfirst(lc))->eqop;
4268 :
4269 : /* lookup / build the necessary comparators */
4270 528 : if (numDistinctCols == 1)
4271 438 : fmgr_info(get_opcode(ops[0]), &pertrans->equalfnOne);
4272 : else
4273 90 : pertrans->equalfnMulti =
4274 90 : execTuplesMatchPrepare(pertrans->sortdesc,
4275 : numDistinctCols,
4276 90 : pertrans->sortColIdx,
4277 : ops,
4278 90 : pertrans->sortCollations,
4279 : &aggstate->ss.ps);
4280 528 : pfree(ops);
4281 : }
4282 :
4283 44418 : pertrans->sortstates = (Tuplesortstate **)
4284 44418 : palloc0(sizeof(Tuplesortstate *) * numGroupingSets);
4285 44418 : }
4286 :
4287 :
4288 : static Datum
4289 18436 : GetAggInitVal(Datum textInitVal, Oid transtype)
4290 : {
4291 : Oid typinput,
4292 : typioparam;
4293 : char *strInitVal;
4294 : Datum initVal;
4295 :
4296 18436 : getTypeInputInfo(transtype, &typinput, &typioparam);
4297 18436 : strInitVal = TextDatumGetCString(textInitVal);
4298 18436 : initVal = OidInputFunctionCall(typinput, strInitVal,
4299 : typioparam, -1);
4300 18436 : pfree(strInitVal);
4301 18436 : return initVal;
4302 : }
4303 :
4304 : void
4305 42010 : ExecEndAgg(AggState *node)
4306 : {
4307 : PlanState *outerPlan;
4308 : int transno;
4309 42010 : int numGroupingSets = Max(node->maxsets, 1);
4310 : int setno;
4311 :
4312 : /*
4313 : * When ending a parallel worker, copy the statistics gathered by the
4314 : * worker back into shared memory so that it can be picked up by the main
4315 : * process to report in EXPLAIN ANALYZE.
4316 : */
4317 42010 : if (node->shared_info && IsParallelWorker())
4318 : {
4319 : AggregateInstrumentation *si;
4320 :
4321 : Assert(ParallelWorkerNumber <= node->shared_info->num_workers);
4322 174 : si = &node->shared_info->sinstrument[ParallelWorkerNumber];
4323 174 : si->hash_batches_used = node->hash_batches_used;
4324 174 : si->hash_disk_used = node->hash_disk_used;
4325 174 : si->hash_mem_peak = node->hash_mem_peak;
4326 : }
4327 :
4328 : /* Make sure we have closed any open tuplesorts */
4329 :
4330 42010 : if (node->sort_in)
4331 144 : tuplesort_end(node->sort_in);
4332 42010 : if (node->sort_out)
4333 42 : tuplesort_end(node->sort_out);
4334 :
4335 42010 : hashagg_reset_spill_state(node);
4336 :
4337 42010 : if (node->hash_metacxt != NULL)
4338 : {
4339 7556 : MemoryContextDelete(node->hash_metacxt);
4340 7556 : node->hash_metacxt = NULL;
4341 : }
4342 :
4343 86318 : for (transno = 0; transno < node->numtrans; transno++)
4344 : {
4345 44308 : AggStatePerTrans pertrans = &node->pertrans[transno];
4346 :
4347 89618 : for (setno = 0; setno < numGroupingSets; setno++)
4348 : {
4349 45310 : if (pertrans->sortstates[setno])
4350 0 : tuplesort_end(pertrans->sortstates[setno]);
4351 : }
4352 : }
4353 :
4354 : /* And ensure any agg shutdown callbacks have been called */
4355 84848 : for (setno = 0; setno < numGroupingSets; setno++)
4356 42838 : ReScanExprContext(node->aggcontexts[setno]);
4357 42010 : if (node->hashcontext)
4358 7556 : ReScanExprContext(node->hashcontext);
4359 :
4360 42010 : outerPlan = outerPlanState(node);
4361 42010 : ExecEndNode(outerPlan);
4362 42010 : }
4363 :
4364 : void
4365 41930 : ExecReScanAgg(AggState *node)
4366 : {
4367 41930 : ExprContext *econtext = node->ss.ps.ps_ExprContext;
4368 41930 : PlanState *outerPlan = outerPlanState(node);
4369 41930 : Agg *aggnode = (Agg *) node->ss.ps.plan;
4370 : int transno;
4371 41930 : int numGroupingSets = Max(node->maxsets, 1);
4372 : int setno;
4373 :
4374 41930 : node->agg_done = false;
4375 :
4376 41930 : if (node->aggstrategy == AGG_HASHED)
4377 : {
4378 : /*
4379 : * In the hashed case, if we haven't yet built the hash table then we
4380 : * can just return; nothing done yet, so nothing to undo. If subnode's
4381 : * chgParam is not NULL then it will be re-scanned by ExecProcNode,
4382 : * else no reason to re-scan it at all.
4383 : */
4384 8682 : if (!node->table_filled)
4385 92 : return;
4386 :
4387 : /*
4388 : * If we do have the hash table, and it never spilled, and the subplan
4389 : * does not have any parameter changes, and none of our own parameter
4390 : * changes affect input expressions of the aggregated functions, then
4391 : * we can just rescan the existing hash table; no need to build it
4392 : * again.
4393 : */
4394 8590 : if (outerPlan->chgParam == NULL && !node->hash_ever_spilled &&
4395 946 : !bms_overlap(node->ss.ps.chgParam, aggnode->aggParams))
4396 : {
4397 922 : ResetTupleHashIterator(node->perhash[0].hashtable,
4398 : &node->perhash[0].hashiter);
4399 922 : select_current_set(node, 0, true);
4400 922 : return;
4401 : }
4402 : }
4403 :
4404 : /* Make sure we have closed any open tuplesorts */
4405 101232 : for (transno = 0; transno < node->numtrans; transno++)
4406 : {
4407 120668 : for (setno = 0; setno < numGroupingSets; setno++)
4408 : {
4409 60352 : AggStatePerTrans pertrans = &node->pertrans[transno];
4410 :
4411 60352 : if (pertrans->sortstates[setno])
4412 : {
4413 0 : tuplesort_end(pertrans->sortstates[setno]);
4414 0 : pertrans->sortstates[setno] = NULL;
4415 : }
4416 : }
4417 : }
4418 :
4419 : /*
4420 : * We don't need to ReScanExprContext the output tuple context here;
4421 : * ExecReScan already did it. But we do need to reset our per-grouping-set
4422 : * contexts, which may have transvalues stored in them. (We use rescan
4423 : * rather than just reset because transfns may have registered callbacks
4424 : * that need to be run now.) For the AGG_HASHED case, see below.
4425 : */
4426 :
4427 81868 : for (setno = 0; setno < numGroupingSets; setno++)
4428 : {
4429 40952 : ReScanExprContext(node->aggcontexts[setno]);
4430 : }
4431 :
4432 : /* Release first tuple of group, if we have made a copy */
4433 40916 : if (node->grp_firstTuple != NULL)
4434 : {
4435 0 : heap_freetuple(node->grp_firstTuple);
4436 0 : node->grp_firstTuple = NULL;
4437 : }
4438 40916 : ExecClearTuple(node->ss.ss_ScanTupleSlot);
4439 :
4440 : /* Forget current agg values */
4441 101232 : MemSet(econtext->ecxt_aggvalues, 0, sizeof(Datum) * node->numaggs);
4442 40916 : MemSet(econtext->ecxt_aggnulls, 0, sizeof(bool) * node->numaggs);
4443 :
4444 : /*
4445 : * With AGG_HASHED/MIXED, the hash table is allocated in a sub-context of
4446 : * the hashcontext. This used to be an issue, but now, resetting a context
4447 : * automatically deletes sub-contexts too.
4448 : */
4449 40916 : if (node->aggstrategy == AGG_HASHED || node->aggstrategy == AGG_MIXED)
4450 : {
4451 7698 : hashagg_reset_spill_state(node);
4452 :
4453 7698 : node->hash_ever_spilled = false;
4454 7698 : node->hash_spill_mode = false;
4455 7698 : node->hash_ngroups_current = 0;
4456 :
4457 7698 : ReScanExprContext(node->hashcontext);
4458 : /* Rebuild an empty hash table */
4459 7698 : build_hash_tables(node);
4460 7698 : node->table_filled = false;
4461 : /* iterator will be reset when the table is filled */
4462 :
4463 7698 : hashagg_recompile_expressions(node, false, false);
4464 : }
4465 :
4466 40916 : if (node->aggstrategy != AGG_HASHED)
4467 : {
4468 : /*
4469 : * Reset the per-group state (in particular, mark transvalues null)
4470 : */
4471 66532 : for (setno = 0; setno < numGroupingSets; setno++)
4472 : {
4473 153892 : MemSet(node->pergroups[setno], 0,
4474 : sizeof(AggStatePerGroupData) * node->numaggs);
4475 : }
4476 :
4477 : /* reset to phase 1 */
4478 33248 : initialize_phase(node, 1);
4479 :
4480 33248 : node->input_done = false;
4481 33248 : node->projected_set = -1;
4482 : }
4483 :
4484 40916 : if (outerPlan->chgParam == NULL)
4485 152 : ExecReScan(outerPlan);
4486 : }
4487 :
4488 :
4489 : /***********************************************************************
4490 : * API exposed to aggregate functions
4491 : ***********************************************************************/
4492 :
4493 :
4494 : /*
4495 : * AggCheckCallContext - test if a SQL function is being called as an aggregate
4496 : *
4497 : * The transition and/or final functions of an aggregate may want to verify
4498 : * that they are being called as aggregates, rather than as plain SQL
4499 : * functions. They should use this function to do so. The return value
4500 : * is nonzero if being called as an aggregate, or zero if not. (Specific
4501 : * nonzero values are AGG_CONTEXT_AGGREGATE or AGG_CONTEXT_WINDOW, but more
4502 : * values could conceivably appear in future.)
4503 : *
4504 : * If aggcontext isn't NULL, the function also stores at *aggcontext the
4505 : * identity of the memory context that aggregate transition values are being
4506 : * stored in. Note that the same aggregate call site (flinfo) may be called
4507 : * interleaved on different transition values in different contexts, so it's
4508 : * not kosher to cache aggcontext under fn_extra. It is, however, kosher to
4509 : * cache it in the transvalue itself (for internal-type transvalues).
4510 : */
4511 : int
4512 4487550 : AggCheckCallContext(FunctionCallInfo fcinfo, MemoryContext *aggcontext)
4513 : {
4514 4487550 : if (fcinfo->context && IsA(fcinfo->context, AggState))
4515 : {
4516 4413428 : if (aggcontext)
4517 : {
4518 1826020 : AggState *aggstate = ((AggState *) fcinfo->context);
4519 1826020 : ExprContext *cxt = aggstate->curaggcontext;
4520 :
4521 1826020 : *aggcontext = cxt->ecxt_per_tuple_memory;
4522 : }
4523 4413428 : return AGG_CONTEXT_AGGREGATE;
4524 : }
4525 74122 : if (fcinfo->context && IsA(fcinfo->context, WindowAggState))
4526 : {
4527 72254 : if (aggcontext)
4528 710 : *aggcontext = ((WindowAggState *) fcinfo->context)->curaggcontext;
4529 72254 : return AGG_CONTEXT_WINDOW;
4530 : }
4531 :
4532 : /* this is just to prevent "uninitialized variable" warnings */
4533 1868 : if (aggcontext)
4534 1820 : *aggcontext = NULL;
4535 1868 : return 0;
4536 : }
4537 :
4538 : /*
4539 : * AggGetAggref - allow an aggregate support function to get its Aggref
4540 : *
4541 : * If the function is being called as an aggregate support function,
4542 : * return the Aggref node for the aggregate call. Otherwise, return NULL.
4543 : *
4544 : * Aggregates sharing the same inputs and transition functions can get
4545 : * merged into a single transition calculation. If the transition function
4546 : * calls AggGetAggref, it will get some one of the Aggrefs for which it is
4547 : * executing. It must therefore not pay attention to the Aggref fields that
4548 : * relate to the final function, as those are indeterminate. But if a final
4549 : * function calls AggGetAggref, it will get a precise result.
4550 : *
4551 : * Note that if an aggregate is being used as a window function, this will
4552 : * return NULL. We could provide a similar function to return the relevant
4553 : * WindowFunc node in such cases, but it's not needed yet.
4554 : */
4555 : Aggref *
4556 246 : AggGetAggref(FunctionCallInfo fcinfo)
4557 : {
4558 246 : if (fcinfo->context && IsA(fcinfo->context, AggState))
4559 : {
4560 246 : AggState *aggstate = (AggState *) fcinfo->context;
4561 : AggStatePerAgg curperagg;
4562 : AggStatePerTrans curpertrans;
4563 :
4564 : /* check curperagg (valid when in a final function) */
4565 246 : curperagg = aggstate->curperagg;
4566 :
4567 246 : if (curperagg)
4568 0 : return curperagg->aggref;
4569 :
4570 : /* check curpertrans (valid when in a transition function) */
4571 246 : curpertrans = aggstate->curpertrans;
4572 :
4573 246 : if (curpertrans)
4574 246 : return curpertrans->aggref;
4575 : }
4576 0 : return NULL;
4577 : }
4578 :
4579 : /*
4580 : * AggGetTempMemoryContext - fetch short-term memory context for aggregates
4581 : *
4582 : * This is useful in agg final functions; the context returned is one that
4583 : * the final function can safely reset as desired. This isn't useful for
4584 : * transition functions, since the context returned MAY (we don't promise)
4585 : * be the same as the context those are called in.
4586 : *
4587 : * As above, this is currently not useful for aggs called as window functions.
4588 : */
4589 : MemoryContext
4590 0 : AggGetTempMemoryContext(FunctionCallInfo fcinfo)
4591 : {
4592 0 : if (fcinfo->context && IsA(fcinfo->context, AggState))
4593 : {
4594 0 : AggState *aggstate = (AggState *) fcinfo->context;
4595 :
4596 0 : return aggstate->tmpcontext->ecxt_per_tuple_memory;
4597 : }
4598 0 : return NULL;
4599 : }
4600 :
4601 : /*
4602 : * AggStateIsShared - find out whether transition state is shared
4603 : *
4604 : * If the function is being called as an aggregate support function,
4605 : * return true if the aggregate's transition state is shared across
4606 : * multiple aggregates, false if it is not.
4607 : *
4608 : * Returns true if not called as an aggregate support function.
4609 : * This is intended as a conservative answer, ie "no you'd better not
4610 : * scribble on your input". In particular, will return true if the
4611 : * aggregate is being used as a window function, which is a scenario
4612 : * in which changing the transition state is a bad idea. We might
4613 : * want to refine the behavior for the window case in future.
4614 : */
4615 : bool
4616 246 : AggStateIsShared(FunctionCallInfo fcinfo)
4617 : {
4618 246 : if (fcinfo->context && IsA(fcinfo->context, AggState))
4619 : {
4620 246 : AggState *aggstate = (AggState *) fcinfo->context;
4621 : AggStatePerAgg curperagg;
4622 : AggStatePerTrans curpertrans;
4623 :
4624 : /* check curperagg (valid when in a final function) */
4625 246 : curperagg = aggstate->curperagg;
4626 :
4627 246 : if (curperagg)
4628 0 : return aggstate->pertrans[curperagg->transno].aggshared;
4629 :
4630 : /* check curpertrans (valid when in a transition function) */
4631 246 : curpertrans = aggstate->curpertrans;
4632 :
4633 246 : if (curpertrans)
4634 246 : return curpertrans->aggshared;
4635 : }
4636 0 : return true;
4637 : }
4638 :
4639 : /*
4640 : * AggRegisterCallback - register a cleanup callback for an aggregate
4641 : *
4642 : * This is useful for aggs to register shutdown callbacks, which will ensure
4643 : * that non-memory resources are freed. The callback will occur just before
4644 : * the associated aggcontext (as returned by AggCheckCallContext) is reset,
4645 : * either between groups or as a result of rescanning the query. The callback
4646 : * will NOT be called on error paths. The typical use-case is for freeing of
4647 : * tuplestores or tuplesorts maintained in aggcontext, or pins held by slots
4648 : * created by the agg functions. (The callback will not be called until after
4649 : * the result of the finalfn is no longer needed, so it's safe for the finalfn
4650 : * to return data that will be freed by the callback.)
4651 : *
4652 : * As above, this is currently not useful for aggs called as window functions.
4653 : */
4654 : void
4655 660 : AggRegisterCallback(FunctionCallInfo fcinfo,
4656 : ExprContextCallbackFunction func,
4657 : Datum arg)
4658 : {
4659 660 : if (fcinfo->context && IsA(fcinfo->context, AggState))
4660 : {
4661 660 : AggState *aggstate = (AggState *) fcinfo->context;
4662 660 : ExprContext *cxt = aggstate->curaggcontext;
4663 :
4664 660 : RegisterExprContextCallback(cxt, func, arg);
4665 :
4666 660 : return;
4667 : }
4668 0 : elog(ERROR, "aggregate function cannot register a callback in this context");
4669 : }
4670 :
4671 :
4672 : /* ----------------------------------------------------------------
4673 : * Parallel Query Support
4674 : * ----------------------------------------------------------------
4675 : */
4676 :
4677 : /* ----------------------------------------------------------------
4678 : * ExecAggEstimate
4679 : *
4680 : * Estimate space required to propagate aggregate statistics.
4681 : * ----------------------------------------------------------------
4682 : */
4683 : void
4684 558 : ExecAggEstimate(AggState *node, ParallelContext *pcxt)
4685 : {
4686 : Size size;
4687 :
4688 : /* don't need this if not instrumenting or no workers */
4689 558 : if (!node->ss.ps.instrument || pcxt->nworkers == 0)
4690 450 : return;
4691 :
4692 108 : size = mul_size(pcxt->nworkers, sizeof(AggregateInstrumentation));
4693 108 : size = add_size(size, offsetof(SharedAggInfo, sinstrument));
4694 108 : shm_toc_estimate_chunk(&pcxt->estimator, size);
4695 108 : shm_toc_estimate_keys(&pcxt->estimator, 1);
4696 : }
4697 :
4698 : /* ----------------------------------------------------------------
4699 : * ExecAggInitializeDSM
4700 : *
4701 : * Initialize DSM space for aggregate statistics.
4702 : * ----------------------------------------------------------------
4703 : */
4704 : void
4705 558 : ExecAggInitializeDSM(AggState *node, ParallelContext *pcxt)
4706 : {
4707 : Size size;
4708 :
4709 : /* don't need this if not instrumenting or no workers */
4710 558 : if (!node->ss.ps.instrument || pcxt->nworkers == 0)
4711 450 : return;
4712 :
4713 108 : size = offsetof(SharedAggInfo, sinstrument)
4714 108 : + pcxt->nworkers * sizeof(AggregateInstrumentation);
4715 108 : node->shared_info = shm_toc_allocate(pcxt->toc, size);
4716 : /* ensure any unfilled slots will contain zeroes */
4717 108 : memset(node->shared_info, 0, size);
4718 108 : node->shared_info->num_workers = pcxt->nworkers;
4719 108 : shm_toc_insert(pcxt->toc, node->ss.ps.plan->plan_node_id,
4720 108 : node->shared_info);
4721 : }
4722 :
4723 : /* ----------------------------------------------------------------
4724 : * ExecAggInitializeWorker
4725 : *
4726 : * Attach worker to DSM space for aggregate statistics.
4727 : * ----------------------------------------------------------------
4728 : */
4729 : void
4730 1532 : ExecAggInitializeWorker(AggState *node, ParallelWorkerContext *pwcxt)
4731 : {
4732 1532 : node->shared_info =
4733 1532 : shm_toc_lookup(pwcxt->toc, node->ss.ps.plan->plan_node_id, true);
4734 1532 : }
4735 :
4736 : /* ----------------------------------------------------------------
4737 : * ExecAggRetrieveInstrumentation
4738 : *
4739 : * Transfer aggregate statistics from DSM to private memory.
4740 : * ----------------------------------------------------------------
4741 : */
4742 : void
4743 108 : ExecAggRetrieveInstrumentation(AggState *node)
4744 : {
4745 : Size size;
4746 : SharedAggInfo *si;
4747 :
4748 108 : if (node->shared_info == NULL)
4749 0 : return;
4750 :
4751 108 : size = offsetof(SharedAggInfo, sinstrument)
4752 108 : + node->shared_info->num_workers * sizeof(AggregateInstrumentation);
4753 108 : si = palloc(size);
4754 108 : memcpy(si, node->shared_info, size);
4755 108 : node->shared_info = si;
4756 : }
|