Line data Source code
1 : /*-------------------------------------------------------------------------
2 : *
3 : * nodeWindowAgg.c
4 : * routines to handle WindowAgg nodes.
5 : *
6 : * A WindowAgg node evaluates "window functions" across suitable partitions
7 : * of the input tuple set. Any one WindowAgg works for just a single window
8 : * specification, though it can evaluate multiple window functions sharing
9 : * identical window specifications. The input tuples are required to be
10 : * delivered in sorted order, with the PARTITION BY columns (if any) as
11 : * major sort keys and the ORDER BY columns (if any) as minor sort keys.
12 : * (The planner generates a stack of WindowAggs with intervening Sort nodes
13 : * as needed, if a query involves more than one window specification.)
14 : *
15 : * Since window functions can require access to any or all of the rows in
16 : * the current partition, we accumulate rows of the partition into a
17 : * tuplestore. The window functions are called using the WindowObject API
18 : * so that they can access those rows as needed.
19 : *
20 : * We also support using plain aggregate functions as window functions.
21 : * For these, the regular Agg-node environment is emulated for each partition.
22 : * As required by the SQL spec, the output represents the value of the
23 : * aggregate function over all rows in the current row's window frame.
24 : *
25 : *
26 : * Portions Copyright (c) 1996-2023, PostgreSQL Global Development Group
27 : * Portions Copyright (c) 1994, Regents of the University of California
28 : *
29 : * IDENTIFICATION
30 : * src/backend/executor/nodeWindowAgg.c
31 : *
32 : *-------------------------------------------------------------------------
33 : */
34 : #include "postgres.h"
35 :
36 : #include "access/htup_details.h"
37 : #include "catalog/objectaccess.h"
38 : #include "catalog/pg_aggregate.h"
39 : #include "catalog/pg_proc.h"
40 : #include "executor/executor.h"
41 : #include "executor/nodeWindowAgg.h"
42 : #include "miscadmin.h"
43 : #include "nodes/nodeFuncs.h"
44 : #include "optimizer/clauses.h"
45 : #include "optimizer/optimizer.h"
46 : #include "parser/parse_agg.h"
47 : #include "parser/parse_coerce.h"
48 : #include "utils/acl.h"
49 : #include "utils/builtins.h"
50 : #include "utils/datum.h"
51 : #include "utils/expandeddatum.h"
52 : #include "utils/lsyscache.h"
53 : #include "utils/memutils.h"
54 : #include "utils/regproc.h"
55 : #include "utils/syscache.h"
56 : #include "windowapi.h"
57 :
58 : /*
59 : * All the window function APIs are called with this object, which is passed
60 : * to window functions as fcinfo->context.
61 : */
62 : typedef struct WindowObjectData
63 : {
64 : NodeTag type;
65 : WindowAggState *winstate; /* parent WindowAggState */
66 : List *argstates; /* ExprState trees for fn's arguments */
67 : void *localmem; /* WinGetPartitionLocalMemory's chunk */
68 : int markptr; /* tuplestore mark pointer for this fn */
69 : int readptr; /* tuplestore read pointer for this fn */
70 : int64 markpos; /* row that markptr is positioned on */
71 : int64 seekpos; /* row that readptr is positioned on */
72 : } WindowObjectData;
73 :
74 : /*
75 : * We have one WindowStatePerFunc struct for each window function and
76 : * window aggregate handled by this node.
77 : */
78 : typedef struct WindowStatePerFuncData
79 : {
80 : /* Links to WindowFunc expr and state nodes this working state is for */
81 : WindowFuncExprState *wfuncstate;
82 : WindowFunc *wfunc;
83 :
84 : int numArguments; /* number of arguments */
85 :
86 : FmgrInfo flinfo; /* fmgr lookup data for window function */
87 :
88 : Oid winCollation; /* collation derived for window function */
89 :
90 : /*
91 : * We need the len and byval info for the result of each function in order
92 : * to know how to copy/delete values.
93 : */
94 : int16 resulttypeLen;
95 : bool resulttypeByVal;
96 :
97 : bool plain_agg; /* is it just a plain aggregate function? */
98 : int aggno; /* if so, index of its WindowStatePerAggData */
99 :
100 : WindowObject winobj; /* object used in window function API */
101 : } WindowStatePerFuncData;
102 :
103 : /*
104 : * For plain aggregate window functions, we also have one of these.
105 : */
106 : typedef struct WindowStatePerAggData
107 : {
108 : /* Oids of transition functions */
109 : Oid transfn_oid;
110 : Oid invtransfn_oid; /* may be InvalidOid */
111 : Oid finalfn_oid; /* may be InvalidOid */
112 :
113 : /*
114 : * fmgr lookup data for transition functions --- only valid when
115 : * corresponding oid is not InvalidOid. Note in particular that fn_strict
116 : * flags are kept here.
117 : */
118 : FmgrInfo transfn;
119 : FmgrInfo invtransfn;
120 : FmgrInfo finalfn;
121 :
122 : int numFinalArgs; /* number of arguments to pass to finalfn */
123 :
124 : /*
125 : * initial value from pg_aggregate entry
126 : */
127 : Datum initValue;
128 : bool initValueIsNull;
129 :
130 : /*
131 : * cached value for current frame boundaries
132 : */
133 : Datum resultValue;
134 : bool resultValueIsNull;
135 :
136 : /*
137 : * We need the len and byval info for the agg's input, result, and
138 : * transition data types in order to know how to copy/delete values.
139 : */
140 : int16 inputtypeLen,
141 : resulttypeLen,
142 : transtypeLen;
143 : bool inputtypeByVal,
144 : resulttypeByVal,
145 : transtypeByVal;
146 :
147 : int wfuncno; /* index of associated WindowStatePerFuncData */
148 :
149 : /* Context holding transition value and possibly other subsidiary data */
150 : MemoryContext aggcontext; /* may be private, or winstate->aggcontext */
151 :
152 : /* Current transition value */
153 : Datum transValue; /* current transition value */
154 : bool transValueIsNull;
155 :
156 : int64 transValueCount; /* number of currently-aggregated rows */
157 :
158 : /* Data local to eval_windowaggregates() */
159 : bool restart; /* need to restart this agg in this cycle? */
160 : } WindowStatePerAggData;
161 :
162 : static void initialize_windowaggregate(WindowAggState *winstate,
163 : WindowStatePerFunc perfuncstate,
164 : WindowStatePerAgg peraggstate);
165 : static void advance_windowaggregate(WindowAggState *winstate,
166 : WindowStatePerFunc perfuncstate,
167 : WindowStatePerAgg peraggstate);
168 : static bool advance_windowaggregate_base(WindowAggState *winstate,
169 : WindowStatePerFunc perfuncstate,
170 : WindowStatePerAgg peraggstate);
171 : static void finalize_windowaggregate(WindowAggState *winstate,
172 : WindowStatePerFunc perfuncstate,
173 : WindowStatePerAgg peraggstate,
174 : Datum *result, bool *isnull);
175 :
176 : static void eval_windowaggregates(WindowAggState *winstate);
177 : static void eval_windowfunction(WindowAggState *winstate,
178 : WindowStatePerFunc perfuncstate,
179 : Datum *result, bool *isnull);
180 :
181 : static void begin_partition(WindowAggState *winstate);
182 : static void spool_tuples(WindowAggState *winstate, int64 pos);
183 : static void release_partition(WindowAggState *winstate);
184 :
185 : static int row_is_in_frame(WindowAggState *winstate, int64 pos,
186 : TupleTableSlot *slot);
187 : static void update_frameheadpos(WindowAggState *winstate);
188 : static void update_frametailpos(WindowAggState *winstate);
189 : static void update_grouptailpos(WindowAggState *winstate);
190 :
191 : static WindowStatePerAggData *initialize_peragg(WindowAggState *winstate,
192 : WindowFunc *wfunc,
193 : WindowStatePerAgg peraggstate);
194 : static Datum GetAggInitVal(Datum textInitVal, Oid transtype);
195 :
196 : static bool are_peers(WindowAggState *winstate, TupleTableSlot *slot1,
197 : TupleTableSlot *slot2);
198 : static bool window_gettupleslot(WindowObject winobj, int64 pos,
199 : TupleTableSlot *slot);
200 :
201 :
202 : /*
203 : * initialize_windowaggregate
204 : * parallel to initialize_aggregates in nodeAgg.c
205 : */
206 : static void
207 4002 : initialize_windowaggregate(WindowAggState *winstate,
208 : WindowStatePerFunc perfuncstate,
209 : WindowStatePerAgg peraggstate)
210 : {
211 : MemoryContext oldContext;
212 :
213 : /*
214 : * If we're using a private aggcontext, we may reset it here. But if the
215 : * context is shared, we don't know which other aggregates may still need
216 : * it, so we must leave it to the caller to reset at an appropriate time.
217 : */
218 4002 : if (peraggstate->aggcontext != winstate->aggcontext)
219 3726 : MemoryContextReset(peraggstate->aggcontext);
220 :
221 4002 : if (peraggstate->initValueIsNull)
222 1034 : peraggstate->transValue = peraggstate->initValue;
223 : else
224 : {
225 2968 : oldContext = MemoryContextSwitchTo(peraggstate->aggcontext);
226 5936 : peraggstate->transValue = datumCopy(peraggstate->initValue,
227 2968 : peraggstate->transtypeByVal,
228 2968 : peraggstate->transtypeLen);
229 2968 : MemoryContextSwitchTo(oldContext);
230 : }
231 4002 : peraggstate->transValueIsNull = peraggstate->initValueIsNull;
232 4002 : peraggstate->transValueCount = 0;
233 4002 : peraggstate->resultValue = (Datum) 0;
234 4002 : peraggstate->resultValueIsNull = true;
235 4002 : }
236 :
237 : /*
238 : * advance_windowaggregate
239 : * parallel to advance_aggregates in nodeAgg.c
240 : */
241 : static void
242 148752 : advance_windowaggregate(WindowAggState *winstate,
243 : WindowStatePerFunc perfuncstate,
244 : WindowStatePerAgg peraggstate)
245 : {
246 148752 : LOCAL_FCINFO(fcinfo, FUNC_MAX_ARGS);
247 148752 : WindowFuncExprState *wfuncstate = perfuncstate->wfuncstate;
248 148752 : int numArguments = perfuncstate->numArguments;
249 : Datum newVal;
250 : ListCell *arg;
251 : int i;
252 : MemoryContext oldContext;
253 148752 : ExprContext *econtext = winstate->tmpcontext;
254 148752 : ExprState *filter = wfuncstate->aggfilter;
255 :
256 148752 : oldContext = MemoryContextSwitchTo(econtext->ecxt_per_tuple_memory);
257 :
258 : /* Skip anything FILTERed out */
259 148752 : if (filter)
260 : {
261 : bool isnull;
262 342 : Datum res = ExecEvalExpr(filter, econtext, &isnull);
263 :
264 342 : if (isnull || !DatumGetBool(res))
265 : {
266 162 : MemoryContextSwitchTo(oldContext);
267 162 : return;
268 : }
269 : }
270 :
271 : /* We start from 1, since the 0th arg will be the transition value */
272 148590 : i = 1;
273 236730 : foreach(arg, wfuncstate->args)
274 : {
275 88140 : ExprState *argstate = (ExprState *) lfirst(arg);
276 :
277 88140 : fcinfo->args[i].value = ExecEvalExpr(argstate, econtext,
278 : &fcinfo->args[i].isnull);
279 88140 : i++;
280 : }
281 :
282 148590 : if (peraggstate->transfn.fn_strict)
283 : {
284 : /*
285 : * For a strict transfn, nothing happens when there's a NULL input; we
286 : * just keep the prior transValue. Note transValueCount doesn't
287 : * change either.
288 : */
289 230540 : for (i = 1; i <= numArguments; i++)
290 : {
291 85108 : if (fcinfo->args[i].isnull)
292 : {
293 198 : MemoryContextSwitchTo(oldContext);
294 198 : return;
295 : }
296 : }
297 :
298 : /*
299 : * For strict transition functions with initial value NULL we use the
300 : * first non-NULL input as the initial state. (We already checked
301 : * that the agg's input type is binary-compatible with its transtype,
302 : * so straight copy here is OK.)
303 : *
304 : * We must copy the datum into aggcontext if it is pass-by-ref. We do
305 : * not need to pfree the old transValue, since it's NULL.
306 : */
307 145432 : if (peraggstate->transValueCount == 0 && peraggstate->transValueIsNull)
308 : {
309 470 : MemoryContextSwitchTo(peraggstate->aggcontext);
310 940 : peraggstate->transValue = datumCopy(fcinfo->args[1].value,
311 470 : peraggstate->transtypeByVal,
312 470 : peraggstate->transtypeLen);
313 470 : peraggstate->transValueIsNull = false;
314 470 : peraggstate->transValueCount = 1;
315 470 : MemoryContextSwitchTo(oldContext);
316 470 : return;
317 : }
318 :
319 144962 : if (peraggstate->transValueIsNull)
320 : {
321 : /*
322 : * Don't call a strict function with NULL inputs. Note it is
323 : * possible to get here despite the above tests, if the transfn is
324 : * strict *and* returned a NULL on a prior cycle. If that happens
325 : * we will propagate the NULL all the way to the end. That can
326 : * only happen if there's no inverse transition function, though,
327 : * since we disallow transitions back to NULL when there is one.
328 : */
329 0 : MemoryContextSwitchTo(oldContext);
330 : Assert(!OidIsValid(peraggstate->invtransfn_oid));
331 0 : return;
332 : }
333 : }
334 :
335 : /*
336 : * OK to call the transition function. Set winstate->curaggcontext while
337 : * calling it, for possible use by AggCheckCallContext.
338 : */
339 147922 : InitFunctionCallInfoData(*fcinfo, &(peraggstate->transfn),
340 : numArguments + 1,
341 : perfuncstate->winCollation,
342 : (void *) winstate, NULL);
343 147922 : fcinfo->args[0].value = peraggstate->transValue;
344 147922 : fcinfo->args[0].isnull = peraggstate->transValueIsNull;
345 147922 : winstate->curaggcontext = peraggstate->aggcontext;
346 147922 : newVal = FunctionCallInvoke(fcinfo);
347 147910 : winstate->curaggcontext = NULL;
348 :
349 : /*
350 : * Moving-aggregate transition functions must not return null, see
351 : * advance_windowaggregate_base().
352 : */
353 147910 : if (fcinfo->isnull && OidIsValid(peraggstate->invtransfn_oid))
354 0 : ereport(ERROR,
355 : (errcode(ERRCODE_NULL_VALUE_NOT_ALLOWED),
356 : errmsg("moving-aggregate transition function must not return null")));
357 :
358 : /*
359 : * We must track the number of rows included in transValue, since to
360 : * remove the last input, advance_windowaggregate_base() mustn't call the
361 : * inverse transition function, but simply reset transValue back to its
362 : * initial value.
363 : */
364 147910 : peraggstate->transValueCount++;
365 :
366 : /*
367 : * If pass-by-ref datatype, must copy the new value into aggcontext and
368 : * free the prior transValue. But if transfn returned a pointer to its
369 : * first input, we don't need to do anything. Also, if transfn returned a
370 : * pointer to a R/W expanded object that is already a child of the
371 : * aggcontext, assume we can adopt that value without copying it. (See
372 : * comments for ExecAggCopyTransValue, which this code duplicates.)
373 : */
374 219010 : if (!peraggstate->transtypeByVal &&
375 71100 : DatumGetPointer(newVal) != DatumGetPointer(peraggstate->transValue))
376 : {
377 960 : if (!fcinfo->isnull)
378 : {
379 960 : MemoryContextSwitchTo(peraggstate->aggcontext);
380 966 : if (DatumIsReadWriteExpandedObject(newVal,
381 : false,
382 960 : peraggstate->transtypeLen) &&
383 6 : MemoryContextGetParent(DatumGetEOHP(newVal)->eoh_context) == CurrentMemoryContext)
384 : /* do nothing */ ;
385 : else
386 954 : newVal = datumCopy(newVal,
387 954 : peraggstate->transtypeByVal,
388 954 : peraggstate->transtypeLen);
389 : }
390 960 : if (!peraggstate->transValueIsNull)
391 : {
392 900 : if (DatumIsReadWriteExpandedObject(peraggstate->transValue,
393 : false,
394 : peraggstate->transtypeLen))
395 0 : DeleteExpandedObject(peraggstate->transValue);
396 : else
397 900 : pfree(DatumGetPointer(peraggstate->transValue));
398 : }
399 : }
400 :
401 147910 : MemoryContextSwitchTo(oldContext);
402 147910 : peraggstate->transValue = newVal;
403 147910 : peraggstate->transValueIsNull = fcinfo->isnull;
404 : }
405 :
406 : /*
407 : * advance_windowaggregate_base
408 : * Remove the oldest tuple from an aggregation.
409 : *
410 : * This is very much like advance_windowaggregate, except that we will call
411 : * the inverse transition function (which caller must have checked is
412 : * available).
413 : *
414 : * Returns true if we successfully removed the current row from this
415 : * aggregate, false if not (in the latter case, caller is responsible
416 : * for cleaning up by restarting the aggregation).
417 : */
418 : static bool
419 4626 : advance_windowaggregate_base(WindowAggState *winstate,
420 : WindowStatePerFunc perfuncstate,
421 : WindowStatePerAgg peraggstate)
422 : {
423 4626 : LOCAL_FCINFO(fcinfo, FUNC_MAX_ARGS);
424 4626 : WindowFuncExprState *wfuncstate = perfuncstate->wfuncstate;
425 4626 : int numArguments = perfuncstate->numArguments;
426 : Datum newVal;
427 : ListCell *arg;
428 : int i;
429 : MemoryContext oldContext;
430 4626 : ExprContext *econtext = winstate->tmpcontext;
431 4626 : ExprState *filter = wfuncstate->aggfilter;
432 :
433 4626 : oldContext = MemoryContextSwitchTo(econtext->ecxt_per_tuple_memory);
434 :
435 : /* Skip anything FILTERed out */
436 4626 : if (filter)
437 : {
438 : bool isnull;
439 102 : Datum res = ExecEvalExpr(filter, econtext, &isnull);
440 :
441 102 : if (isnull || !DatumGetBool(res))
442 : {
443 48 : MemoryContextSwitchTo(oldContext);
444 48 : return true;
445 : }
446 : }
447 :
448 : /* We start from 1, since the 0th arg will be the transition value */
449 4578 : i = 1;
450 9138 : foreach(arg, wfuncstate->args)
451 : {
452 4560 : ExprState *argstate = (ExprState *) lfirst(arg);
453 :
454 4560 : fcinfo->args[i].value = ExecEvalExpr(argstate, econtext,
455 : &fcinfo->args[i].isnull);
456 4560 : i++;
457 : }
458 :
459 4578 : if (peraggstate->invtransfn.fn_strict)
460 : {
461 : /*
462 : * For a strict (inv)transfn, nothing happens when there's a NULL
463 : * input; we just keep the prior transValue. Note transValueCount
464 : * doesn't change either.
465 : */
466 5604 : for (i = 1; i <= numArguments; i++)
467 : {
468 2832 : if (fcinfo->args[i].isnull)
469 : {
470 78 : MemoryContextSwitchTo(oldContext);
471 78 : return true;
472 : }
473 : }
474 : }
475 :
476 : /* There should still be an added but not yet removed value */
477 : Assert(peraggstate->transValueCount > 0);
478 :
479 : /*
480 : * In moving-aggregate mode, the state must never be NULL, except possibly
481 : * before any rows have been aggregated (which is surely not the case at
482 : * this point). This restriction allows us to interpret a NULL result
483 : * from the inverse function as meaning "sorry, can't do an inverse
484 : * transition in this case". We already checked this in
485 : * advance_windowaggregate, but just for safety, check again.
486 : */
487 4500 : if (peraggstate->transValueIsNull)
488 0 : elog(ERROR, "aggregate transition value is NULL before inverse transition");
489 :
490 : /*
491 : * We mustn't use the inverse transition function to remove the last
492 : * input. Doing so would yield a non-NULL state, whereas we should be in
493 : * the initial state afterwards which may very well be NULL. So instead,
494 : * we simply re-initialize the aggregate in this case.
495 : */
496 4500 : if (peraggstate->transValueCount == 1)
497 : {
498 90 : MemoryContextSwitchTo(oldContext);
499 90 : initialize_windowaggregate(winstate,
500 90 : &winstate->perfunc[peraggstate->wfuncno],
501 : peraggstate);
502 90 : return true;
503 : }
504 :
505 : /*
506 : * OK to call the inverse transition function. Set
507 : * winstate->curaggcontext while calling it, for possible use by
508 : * AggCheckCallContext.
509 : */
510 4410 : InitFunctionCallInfoData(*fcinfo, &(peraggstate->invtransfn),
511 : numArguments + 1,
512 : perfuncstate->winCollation,
513 : (void *) winstate, NULL);
514 4410 : fcinfo->args[0].value = peraggstate->transValue;
515 4410 : fcinfo->args[0].isnull = peraggstate->transValueIsNull;
516 4410 : winstate->curaggcontext = peraggstate->aggcontext;
517 4410 : newVal = FunctionCallInvoke(fcinfo);
518 4410 : winstate->curaggcontext = NULL;
519 :
520 : /*
521 : * If the function returns NULL, report failure, forcing a restart.
522 : */
523 4410 : if (fcinfo->isnull)
524 : {
525 266 : MemoryContextSwitchTo(oldContext);
526 266 : return false;
527 : }
528 :
529 : /* Update number of rows included in transValue */
530 4144 : peraggstate->transValueCount--;
531 :
532 : /*
533 : * If pass-by-ref datatype, must copy the new value into aggcontext and
534 : * free the prior transValue. But if invtransfn returned a pointer to its
535 : * first input, we don't need to do anything. Also, if invtransfn
536 : * returned a pointer to a R/W expanded object that is already a child of
537 : * the aggcontext, assume we can adopt that value without copying it. (See
538 : * comments for ExecAggCopyTransValue, which this code duplicates.)
539 : *
540 : * Note: the checks for null values here will never fire, but it seems
541 : * best to have this stanza look just like advance_windowaggregate.
542 : */
543 6274 : if (!peraggstate->transtypeByVal &&
544 2130 : DatumGetPointer(newVal) != DatumGetPointer(peraggstate->transValue))
545 : {
546 666 : if (!fcinfo->isnull)
547 : {
548 666 : MemoryContextSwitchTo(peraggstate->aggcontext);
549 666 : if (DatumIsReadWriteExpandedObject(newVal,
550 : false,
551 666 : peraggstate->transtypeLen) &&
552 0 : MemoryContextGetParent(DatumGetEOHP(newVal)->eoh_context) == CurrentMemoryContext)
553 : /* do nothing */ ;
554 : else
555 666 : newVal = datumCopy(newVal,
556 666 : peraggstate->transtypeByVal,
557 666 : peraggstate->transtypeLen);
558 : }
559 666 : if (!peraggstate->transValueIsNull)
560 : {
561 666 : if (DatumIsReadWriteExpandedObject(peraggstate->transValue,
562 : false,
563 : peraggstate->transtypeLen))
564 0 : DeleteExpandedObject(peraggstate->transValue);
565 : else
566 666 : pfree(DatumGetPointer(peraggstate->transValue));
567 : }
568 : }
569 :
570 4144 : MemoryContextSwitchTo(oldContext);
571 4144 : peraggstate->transValue = newVal;
572 4144 : peraggstate->transValueIsNull = fcinfo->isnull;
573 :
574 4144 : return true;
575 : }
576 :
577 : /*
578 : * finalize_windowaggregate
579 : * parallel to finalize_aggregate in nodeAgg.c
580 : */
581 : static void
582 10562 : finalize_windowaggregate(WindowAggState *winstate,
583 : WindowStatePerFunc perfuncstate,
584 : WindowStatePerAgg peraggstate,
585 : Datum *result, bool *isnull)
586 : {
587 : MemoryContext oldContext;
588 :
589 10562 : oldContext = MemoryContextSwitchTo(winstate->ss.ps.ps_ExprContext->ecxt_per_tuple_memory);
590 :
591 : /*
592 : * Apply the agg's finalfn if one is provided, else return transValue.
593 : */
594 10562 : if (OidIsValid(peraggstate->finalfn_oid))
595 : {
596 7582 : LOCAL_FCINFO(fcinfo, FUNC_MAX_ARGS);
597 7582 : int numFinalArgs = peraggstate->numFinalArgs;
598 : bool anynull;
599 : int i;
600 :
601 7582 : InitFunctionCallInfoData(fcinfodata.fcinfo, &(peraggstate->finalfn),
602 : numFinalArgs,
603 : perfuncstate->winCollation,
604 : (void *) winstate, NULL);
605 7582 : fcinfo->args[0].value =
606 7582 : MakeExpandedObjectReadOnly(peraggstate->transValue,
607 : peraggstate->transValueIsNull,
608 : peraggstate->transtypeLen);
609 7582 : fcinfo->args[0].isnull = peraggstate->transValueIsNull;
610 7582 : anynull = peraggstate->transValueIsNull;
611 :
612 : /* Fill any remaining argument positions with nulls */
613 7682 : for (i = 1; i < numFinalArgs; i++)
614 : {
615 100 : fcinfo->args[i].value = (Datum) 0;
616 100 : fcinfo->args[i].isnull = true;
617 100 : anynull = true;
618 : }
619 :
620 7582 : if (fcinfo->flinfo->fn_strict && anynull)
621 : {
622 : /* don't call a strict function with NULL inputs */
623 0 : *result = (Datum) 0;
624 0 : *isnull = true;
625 : }
626 : else
627 : {
628 : Datum res;
629 :
630 7582 : winstate->curaggcontext = peraggstate->aggcontext;
631 7582 : res = FunctionCallInvoke(fcinfo);
632 7570 : winstate->curaggcontext = NULL;
633 7570 : *isnull = fcinfo->isnull;
634 7570 : *result = MakeExpandedObjectReadOnly(res,
635 : fcinfo->isnull,
636 : peraggstate->resulttypeLen);
637 : }
638 : }
639 : else
640 : {
641 2980 : *result =
642 2980 : MakeExpandedObjectReadOnly(peraggstate->transValue,
643 : peraggstate->transValueIsNull,
644 : peraggstate->transtypeLen);
645 2980 : *isnull = peraggstate->transValueIsNull;
646 : }
647 :
648 10550 : MemoryContextSwitchTo(oldContext);
649 10550 : }
650 :
651 : /*
652 : * eval_windowaggregates
653 : * evaluate plain aggregates being used as window functions
654 : *
655 : * This differs from nodeAgg.c in two ways. First, if the window's frame
656 : * start position moves, we use the inverse transition function (if it exists)
657 : * to remove rows from the transition value. And second, we expect to be
658 : * able to call aggregate final functions repeatedly after aggregating more
659 : * data onto the same transition value. This is not a behavior required by
660 : * nodeAgg.c.
661 : */
662 : static void
663 130104 : eval_windowaggregates(WindowAggState *winstate)
664 : {
665 : WindowStatePerAgg peraggstate;
666 : int wfuncno,
667 : numaggs,
668 : numaggs_restart,
669 : i;
670 : int64 aggregatedupto_nonrestarted;
671 : MemoryContext oldContext;
672 : ExprContext *econtext;
673 : WindowObject agg_winobj;
674 : TupleTableSlot *agg_row_slot;
675 : TupleTableSlot *temp_slot;
676 :
677 130104 : numaggs = winstate->numaggs;
678 130104 : if (numaggs == 0)
679 0 : return; /* nothing to do */
680 :
681 : /* final output execution is in ps_ExprContext */
682 130104 : econtext = winstate->ss.ps.ps_ExprContext;
683 130104 : agg_winobj = winstate->agg_winobj;
684 130104 : agg_row_slot = winstate->agg_row_slot;
685 130104 : temp_slot = winstate->temp_slot_1;
686 :
687 : /*
688 : * If the window's frame start clause is UNBOUNDED_PRECEDING and no
689 : * exclusion clause is specified, then the window frame consists of a
690 : * contiguous group of rows extending forward from the start of the
691 : * partition, and rows only enter the frame, never exit it, as the current
692 : * row advances forward. This makes it possible to use an incremental
693 : * strategy for evaluating aggregates: we run the transition function for
694 : * each row added to the frame, and run the final function whenever we
695 : * need the current aggregate value. This is considerably more efficient
696 : * than the naive approach of re-running the entire aggregate calculation
697 : * for each current row. It does assume that the final function doesn't
698 : * damage the running transition value, but we have the same assumption in
699 : * nodeAgg.c too (when it rescans an existing hash table).
700 : *
701 : * If the frame start does sometimes move, we can still optimize as above
702 : * whenever successive rows share the same frame head, but if the frame
703 : * head moves beyond the previous head we try to remove those rows using
704 : * the aggregate's inverse transition function. This function restores
705 : * the aggregate's current state to what it would be if the removed row
706 : * had never been aggregated in the first place. Inverse transition
707 : * functions may optionally return NULL, indicating that the function was
708 : * unable to remove the tuple from aggregation. If this happens, or if
709 : * the aggregate doesn't have an inverse transition function at all, we
710 : * must perform the aggregation all over again for all tuples within the
711 : * new frame boundaries.
712 : *
713 : * If there's any exclusion clause, then we may have to aggregate over a
714 : * non-contiguous set of rows, so we punt and recalculate for every row.
715 : * (For some frame end choices, it might be that the frame is always
716 : * contiguous anyway, but that's an optimization to investigate later.)
717 : *
718 : * In many common cases, multiple rows share the same frame and hence the
719 : * same aggregate value. (In particular, if there's no ORDER BY in a RANGE
720 : * window, then all rows are peers and so they all have window frame equal
721 : * to the whole partition.) We optimize such cases by calculating the
722 : * aggregate value once when we reach the first row of a peer group, and
723 : * then returning the saved value for all subsequent rows.
724 : *
725 : * 'aggregatedupto' keeps track of the first row that has not yet been
726 : * accumulated into the aggregate transition values. Whenever we start a
727 : * new peer group, we accumulate forward to the end of the peer group.
728 : */
729 :
730 : /*
731 : * First, update the frame head position.
732 : *
733 : * The frame head should never move backwards, and the code below wouldn't
734 : * cope if it did, so for safety we complain if it does.
735 : */
736 130104 : update_frameheadpos(winstate);
737 130098 : if (winstate->frameheadpos < winstate->aggregatedbase)
738 0 : elog(ERROR, "window frame head moved backward");
739 :
740 : /*
741 : * If the frame didn't change compared to the previous row, we can re-use
742 : * the result values that were previously saved at the bottom of this
743 : * function. Since we don't know the current frame's end yet, this is not
744 : * possible to check for fully. But if the frame end mode is UNBOUNDED
745 : * FOLLOWING or CURRENT ROW, no exclusion clause is specified, and the
746 : * current row lies within the previous row's frame, then the two frames'
747 : * ends must coincide. Note that on the first row aggregatedbase ==
748 : * aggregatedupto, meaning this test must fail, so we don't need to check
749 : * the "there was no previous row" case explicitly here.
750 : */
751 130098 : if (winstate->aggregatedbase == winstate->frameheadpos &&
752 126320 : (winstate->frameOptions & (FRAMEOPTION_END_UNBOUNDED_FOLLOWING |
753 124400 : FRAMEOPTION_END_CURRENT_ROW)) &&
754 124400 : !(winstate->frameOptions & FRAMEOPTION_EXCLUSION) &&
755 124220 : winstate->aggregatedbase <= winstate->currentpos &&
756 124184 : winstate->aggregatedupto > winstate->currentpos)
757 : {
758 242552 : for (i = 0; i < numaggs; i++)
759 : {
760 121282 : peraggstate = &winstate->peragg[i];
761 121282 : wfuncno = peraggstate->wfuncno;
762 121282 : econtext->ecxt_aggvalues[wfuncno] = peraggstate->resultValue;
763 121282 : econtext->ecxt_aggnulls[wfuncno] = peraggstate->resultValueIsNull;
764 : }
765 121270 : return;
766 : }
767 :
768 : /*----------
769 : * Initialize restart flags.
770 : *
771 : * We restart the aggregation:
772 : * - if we're processing the first row in the partition, or
773 : * - if the frame's head moved and we cannot use an inverse
774 : * transition function, or
775 : * - we have an EXCLUSION clause, or
776 : * - if the new frame doesn't overlap the old one
777 : *
778 : * Note that we don't strictly need to restart in the last case, but if
779 : * we're going to remove all rows from the aggregation anyway, a restart
780 : * surely is faster.
781 : *----------
782 : */
783 8828 : numaggs_restart = 0;
784 19414 : for (i = 0; i < numaggs; i++)
785 : {
786 10586 : peraggstate = &winstate->peragg[i];
787 10586 : if (winstate->currentpos == 0 ||
788 8600 : (winstate->aggregatedbase != winstate->frameheadpos &&
789 5206 : !OidIsValid(peraggstate->invtransfn_oid)) ||
790 8524 : (winstate->frameOptions & FRAMEOPTION_EXCLUSION) ||
791 7408 : winstate->aggregatedupto <= winstate->frameheadpos)
792 : {
793 3646 : peraggstate->restart = true;
794 3646 : numaggs_restart++;
795 : }
796 : else
797 6940 : peraggstate->restart = false;
798 : }
799 :
800 : /*
801 : * If we have any possibly-moving aggregates, attempt to advance
802 : * aggregatedbase to match the frame's head by removing input rows that
803 : * fell off the top of the frame from the aggregations. This can fail,
804 : * i.e. advance_windowaggregate_base() can return false, in which case
805 : * we'll restart that aggregate below.
806 : */
807 11996 : while (numaggs_restart < numaggs &&
808 8530 : winstate->aggregatedbase < winstate->frameheadpos)
809 : {
810 : /*
811 : * Fetch the next tuple of those being removed. This should never fail
812 : * as we should have been here before.
813 : */
814 3168 : if (!window_gettupleslot(agg_winobj, winstate->aggregatedbase,
815 : temp_slot))
816 0 : elog(ERROR, "could not re-fetch previously fetched frame row");
817 :
818 : /* Set tuple context for evaluation of aggregate arguments */
819 3168 : winstate->tmpcontext->ecxt_outertuple = temp_slot;
820 :
821 : /*
822 : * Perform the inverse transition for each aggregate function in the
823 : * window, unless it has already been marked as needing a restart.
824 : */
825 7806 : for (i = 0; i < numaggs; i++)
826 : {
827 : bool ok;
828 :
829 4638 : peraggstate = &winstate->peragg[i];
830 4638 : if (peraggstate->restart)
831 12 : continue;
832 :
833 4626 : wfuncno = peraggstate->wfuncno;
834 4626 : ok = advance_windowaggregate_base(winstate,
835 4626 : &winstate->perfunc[wfuncno],
836 : peraggstate);
837 4626 : if (!ok)
838 : {
839 : /* Inverse transition function has failed, must restart */
840 266 : peraggstate->restart = true;
841 266 : numaggs_restart++;
842 : }
843 : }
844 :
845 : /* Reset per-input-tuple context after each tuple */
846 3168 : ResetExprContext(winstate->tmpcontext);
847 :
848 : /* And advance the aggregated-row state */
849 3168 : winstate->aggregatedbase++;
850 3168 : ExecClearTuple(temp_slot);
851 : }
852 :
853 : /*
854 : * If we successfully advanced the base rows of all the aggregates,
855 : * aggregatedbase now equals frameheadpos; but if we failed for any, we
856 : * must forcibly update aggregatedbase.
857 : */
858 8828 : winstate->aggregatedbase = winstate->frameheadpos;
859 :
860 : /*
861 : * If we created a mark pointer for aggregates, keep it pushed up to frame
862 : * head, so that tuplestore can discard unnecessary rows.
863 : */
864 8828 : if (agg_winobj->markptr >= 0)
865 6224 : WinSetMarkPosition(agg_winobj, winstate->frameheadpos);
866 :
867 : /*
868 : * Now restart the aggregates that require it.
869 : *
870 : * We assume that aggregates using the shared context always restart if
871 : * *any* aggregate restarts, and we may thus clean up the shared
872 : * aggcontext if that is the case. Private aggcontexts are reset by
873 : * initialize_windowaggregate() if their owning aggregate restarts. If we
874 : * aren't restarting an aggregate, we need to free any previously saved
875 : * result for it, else we'll leak memory.
876 : */
877 8828 : if (numaggs_restart > 0)
878 3712 : MemoryContextReset(winstate->aggcontext);
879 19414 : for (i = 0; i < numaggs; i++)
880 : {
881 10586 : peraggstate = &winstate->peragg[i];
882 :
883 : /* Aggregates using the shared ctx must restart if *any* agg does */
884 : Assert(peraggstate->aggcontext != winstate->aggcontext ||
885 : numaggs_restart == 0 ||
886 : peraggstate->restart);
887 :
888 10586 : if (peraggstate->restart)
889 : {
890 3912 : wfuncno = peraggstate->wfuncno;
891 3912 : initialize_windowaggregate(winstate,
892 3912 : &winstate->perfunc[wfuncno],
893 : peraggstate);
894 : }
895 6674 : else if (!peraggstate->resultValueIsNull)
896 : {
897 6440 : if (!peraggstate->resulttypeByVal)
898 2152 : pfree(DatumGetPointer(peraggstate->resultValue));
899 6440 : peraggstate->resultValue = (Datum) 0;
900 6440 : peraggstate->resultValueIsNull = true;
901 : }
902 : }
903 :
904 : /*
905 : * Non-restarted aggregates now contain the rows between aggregatedbase
906 : * (i.e., frameheadpos) and aggregatedupto, while restarted aggregates
907 : * contain no rows. If there are any restarted aggregates, we must thus
908 : * begin aggregating anew at frameheadpos, otherwise we may simply
909 : * continue at aggregatedupto. We must remember the old value of
910 : * aggregatedupto to know how long to skip advancing non-restarted
911 : * aggregates. If we modify aggregatedupto, we must also clear
912 : * agg_row_slot, per the loop invariant below.
913 : */
914 8828 : aggregatedupto_nonrestarted = winstate->aggregatedupto;
915 8828 : if (numaggs_restart > 0 &&
916 3712 : winstate->aggregatedupto != winstate->frameheadpos)
917 : {
918 1432 : winstate->aggregatedupto = winstate->frameheadpos;
919 1432 : ExecClearTuple(agg_row_slot);
920 : }
921 :
922 : /*
923 : * Advance until we reach a row not in frame (or end of partition).
924 : *
925 : * Note the loop invariant: agg_row_slot is either empty or holds the row
926 : * at position aggregatedupto. We advance aggregatedupto after processing
927 : * a row.
928 : */
929 : for (;;)
930 147668 : {
931 : int ret;
932 :
933 : /* Fetch next row if we didn't already */
934 156496 : if (TupIsNull(agg_row_slot))
935 : {
936 152634 : if (!window_gettupleslot(agg_winobj, winstate->aggregatedupto,
937 : agg_row_slot))
938 4102 : break; /* must be end of partition */
939 : }
940 :
941 : /*
942 : * Exit loop if no more rows can be in frame. Skip aggregation if
943 : * current row is not in frame but there might be more in the frame.
944 : */
945 152394 : ret = row_is_in_frame(winstate, winstate->aggregatedupto, agg_row_slot);
946 152382 : if (ret < 0)
947 4702 : break;
948 147680 : if (ret == 0)
949 1896 : goto next_tuple;
950 :
951 : /* Set tuple context for evaluation of aggregate arguments */
952 145784 : winstate->tmpcontext->ecxt_outertuple = agg_row_slot;
953 :
954 : /* Accumulate row into the aggregates */
955 316746 : for (i = 0; i < numaggs; i++)
956 : {
957 170974 : peraggstate = &winstate->peragg[i];
958 :
959 : /* Non-restarted aggs skip until aggregatedupto_nonrestarted */
960 170974 : if (!peraggstate->restart &&
961 121874 : winstate->aggregatedupto < aggregatedupto_nonrestarted)
962 22222 : continue;
963 :
964 148752 : wfuncno = peraggstate->wfuncno;
965 148752 : advance_windowaggregate(winstate,
966 148752 : &winstate->perfunc[wfuncno],
967 : peraggstate);
968 : }
969 :
970 145772 : next_tuple:
971 : /* Reset per-input-tuple context after each tuple */
972 147668 : ResetExprContext(winstate->tmpcontext);
973 :
974 : /* And advance the aggregated-row state */
975 147668 : winstate->aggregatedupto++;
976 147668 : ExecClearTuple(agg_row_slot);
977 : }
978 :
979 : /* The frame's end is not supposed to move backwards, ever */
980 : Assert(aggregatedupto_nonrestarted <= winstate->aggregatedupto);
981 :
982 : /*
983 : * finalize aggregates and fill result/isnull fields.
984 : */
985 19354 : for (i = 0; i < numaggs; i++)
986 : {
987 : Datum *result;
988 : bool *isnull;
989 :
990 10562 : peraggstate = &winstate->peragg[i];
991 10562 : wfuncno = peraggstate->wfuncno;
992 10562 : result = &econtext->ecxt_aggvalues[wfuncno];
993 10562 : isnull = &econtext->ecxt_aggnulls[wfuncno];
994 10562 : finalize_windowaggregate(winstate,
995 10562 : &winstate->perfunc[wfuncno],
996 : peraggstate,
997 : result, isnull);
998 :
999 : /*
1000 : * save the result in case next row shares the same frame.
1001 : *
1002 : * XXX in some framing modes, eg ROWS/END_CURRENT_ROW, we can know in
1003 : * advance that the next row can't possibly share the same frame. Is
1004 : * it worth detecting that and skipping this code?
1005 : */
1006 10550 : if (!peraggstate->resulttypeByVal && !*isnull)
1007 : {
1008 2752 : oldContext = MemoryContextSwitchTo(peraggstate->aggcontext);
1009 2752 : peraggstate->resultValue =
1010 2752 : datumCopy(*result,
1011 2752 : peraggstate->resulttypeByVal,
1012 2752 : peraggstate->resulttypeLen);
1013 2752 : MemoryContextSwitchTo(oldContext);
1014 : }
1015 : else
1016 : {
1017 7798 : peraggstate->resultValue = *result;
1018 : }
1019 10550 : peraggstate->resultValueIsNull = *isnull;
1020 : }
1021 : }
1022 :
1023 : /*
1024 : * eval_windowfunction
1025 : *
1026 : * Arguments of window functions are not evaluated here, because a window
1027 : * function can need random access to arbitrary rows in the partition.
1028 : * The window function uses the special WinGetFuncArgInPartition and
1029 : * WinGetFuncArgInFrame functions to evaluate the arguments for the rows
1030 : * it wants.
1031 : */
1032 : static void
1033 796764 : eval_windowfunction(WindowAggState *winstate, WindowStatePerFunc perfuncstate,
1034 : Datum *result, bool *isnull)
1035 : {
1036 796764 : LOCAL_FCINFO(fcinfo, FUNC_MAX_ARGS);
1037 : MemoryContext oldContext;
1038 :
1039 796764 : oldContext = MemoryContextSwitchTo(winstate->ss.ps.ps_ExprContext->ecxt_per_tuple_memory);
1040 :
1041 : /*
1042 : * We don't pass any normal arguments to a window function, but we do pass
1043 : * it the number of arguments, in order to permit window function
1044 : * implementations to support varying numbers of arguments. The real info
1045 : * goes through the WindowObject, which is passed via fcinfo->context.
1046 : */
1047 796764 : InitFunctionCallInfoData(*fcinfo, &(perfuncstate->flinfo),
1048 : perfuncstate->numArguments,
1049 : perfuncstate->winCollation,
1050 : (void *) perfuncstate->winobj, NULL);
1051 : /* Just in case, make all the regular argument slots be null */
1052 989556 : for (int argno = 0; argno < perfuncstate->numArguments; argno++)
1053 192792 : fcinfo->args[argno].isnull = true;
1054 : /* Window functions don't have a current aggregate context, either */
1055 796764 : winstate->curaggcontext = NULL;
1056 :
1057 796764 : *result = FunctionCallInvoke(fcinfo);
1058 796674 : *isnull = fcinfo->isnull;
1059 :
1060 : /*
1061 : * The window function might have returned a pass-by-ref result that's
1062 : * just a pointer into one of the WindowObject's temporary slots. That's
1063 : * not a problem if it's the only window function using the WindowObject;
1064 : * but if there's more than one function, we'd better copy the result to
1065 : * ensure it's not clobbered by later window functions.
1066 : */
1067 796674 : if (!perfuncstate->resulttypeByVal && !fcinfo->isnull &&
1068 1008 : winstate->numfuncs > 1)
1069 96 : *result = datumCopy(*result,
1070 96 : perfuncstate->resulttypeByVal,
1071 96 : perfuncstate->resulttypeLen);
1072 :
1073 796674 : MemoryContextSwitchTo(oldContext);
1074 796674 : }
1075 :
1076 : /*
1077 : * begin_partition
1078 : * Start buffering rows of the next partition.
1079 : */
1080 : static void
1081 3306 : begin_partition(WindowAggState *winstate)
1082 : {
1083 3306 : WindowAgg *node = (WindowAgg *) winstate->ss.ps.plan;
1084 3306 : PlanState *outerPlan = outerPlanState(winstate);
1085 3306 : int frameOptions = winstate->frameOptions;
1086 3306 : int numfuncs = winstate->numfuncs;
1087 : int i;
1088 :
1089 3306 : winstate->partition_spooled = false;
1090 3306 : winstate->framehead_valid = false;
1091 3306 : winstate->frametail_valid = false;
1092 3306 : winstate->grouptail_valid = false;
1093 3306 : winstate->spooled_rows = 0;
1094 3306 : winstate->currentpos = 0;
1095 3306 : winstate->frameheadpos = 0;
1096 3306 : winstate->frametailpos = 0;
1097 3306 : winstate->currentgroup = 0;
1098 3306 : winstate->frameheadgroup = 0;
1099 3306 : winstate->frametailgroup = 0;
1100 3306 : winstate->groupheadpos = 0;
1101 3306 : winstate->grouptailpos = -1; /* see update_grouptailpos */
1102 3306 : ExecClearTuple(winstate->agg_row_slot);
1103 3306 : if (winstate->framehead_slot)
1104 1024 : ExecClearTuple(winstate->framehead_slot);
1105 3306 : if (winstate->frametail_slot)
1106 1696 : ExecClearTuple(winstate->frametail_slot);
1107 :
1108 : /*
1109 : * If this is the very first partition, we need to fetch the first input
1110 : * row to store in first_part_slot.
1111 : */
1112 3306 : if (TupIsNull(winstate->first_part_slot))
1113 : {
1114 2100 : TupleTableSlot *outerslot = ExecProcNode(outerPlan);
1115 :
1116 2100 : if (!TupIsNull(outerslot))
1117 2082 : ExecCopySlot(winstate->first_part_slot, outerslot);
1118 : else
1119 : {
1120 : /* outer plan is empty, so we have nothing to do */
1121 18 : winstate->partition_spooled = true;
1122 18 : winstate->more_partitions = false;
1123 18 : return;
1124 : }
1125 : }
1126 :
1127 : /* Create new tuplestore for this partition */
1128 3288 : winstate->buffer = tuplestore_begin_heap(false, false, work_mem);
1129 :
1130 : /*
1131 : * Set up read pointers for the tuplestore. The current pointer doesn't
1132 : * need BACKWARD capability, but the per-window-function read pointers do,
1133 : * and the aggregate pointer does if we might need to restart aggregation.
1134 : */
1135 3288 : winstate->current_ptr = 0; /* read pointer 0 is pre-allocated */
1136 :
1137 : /* reset default REWIND capability bit for current ptr */
1138 3288 : tuplestore_set_eflags(winstate->buffer, 0);
1139 :
1140 : /* create read pointers for aggregates, if needed */
1141 3288 : if (winstate->numaggs > 0)
1142 : {
1143 1818 : WindowObject agg_winobj = winstate->agg_winobj;
1144 1818 : int readptr_flags = 0;
1145 :
1146 : /*
1147 : * If the frame head is potentially movable, or we have an EXCLUSION
1148 : * clause, we might need to restart aggregation ...
1149 : */
1150 1818 : if (!(frameOptions & FRAMEOPTION_START_UNBOUNDED_PRECEDING) ||
1151 758 : (frameOptions & FRAMEOPTION_EXCLUSION))
1152 : {
1153 : /* ... so create a mark pointer to track the frame head */
1154 1078 : agg_winobj->markptr = tuplestore_alloc_read_pointer(winstate->buffer, 0);
1155 : /* and the read pointer will need BACKWARD capability */
1156 1078 : readptr_flags |= EXEC_FLAG_BACKWARD;
1157 : }
1158 :
1159 1818 : agg_winobj->readptr = tuplestore_alloc_read_pointer(winstate->buffer,
1160 : readptr_flags);
1161 1818 : agg_winobj->markpos = -1;
1162 1818 : agg_winobj->seekpos = -1;
1163 :
1164 : /* Also reset the row counters for aggregates */
1165 1818 : winstate->aggregatedbase = 0;
1166 1818 : winstate->aggregatedupto = 0;
1167 : }
1168 :
1169 : /* create mark and read pointers for each real window function */
1170 7404 : for (i = 0; i < numfuncs; i++)
1171 : {
1172 4116 : WindowStatePerFunc perfuncstate = &(winstate->perfunc[i]);
1173 :
1174 4116 : if (!perfuncstate->plain_agg)
1175 : {
1176 2124 : WindowObject winobj = perfuncstate->winobj;
1177 :
1178 2124 : winobj->markptr = tuplestore_alloc_read_pointer(winstate->buffer,
1179 : 0);
1180 2124 : winobj->readptr = tuplestore_alloc_read_pointer(winstate->buffer,
1181 : EXEC_FLAG_BACKWARD);
1182 2124 : winobj->markpos = -1;
1183 2124 : winobj->seekpos = -1;
1184 : }
1185 : }
1186 :
1187 : /*
1188 : * If we are in RANGE or GROUPS mode, then determining frame boundaries
1189 : * requires physical access to the frame endpoint rows, except in certain
1190 : * degenerate cases. We create read pointers to point to those rows, to
1191 : * simplify access and ensure that the tuplestore doesn't discard the
1192 : * endpoint rows prematurely. (Must create pointers in exactly the same
1193 : * cases that update_frameheadpos and update_frametailpos need them.)
1194 : */
1195 3288 : winstate->framehead_ptr = winstate->frametail_ptr = -1; /* if not used */
1196 :
1197 3288 : if (frameOptions & (FRAMEOPTION_RANGE | FRAMEOPTION_GROUPS))
1198 : {
1199 2142 : if (((frameOptions & FRAMEOPTION_START_CURRENT_ROW) &&
1200 82 : node->ordNumCols != 0) ||
1201 2060 : (frameOptions & FRAMEOPTION_START_OFFSET))
1202 1024 : winstate->framehead_ptr =
1203 1024 : tuplestore_alloc_read_pointer(winstate->buffer, 0);
1204 2142 : if (((frameOptions & FRAMEOPTION_END_CURRENT_ROW) &&
1205 1064 : node->ordNumCols != 0) ||
1206 1448 : (frameOptions & FRAMEOPTION_END_OFFSET))
1207 1696 : winstate->frametail_ptr =
1208 1696 : tuplestore_alloc_read_pointer(winstate->buffer, 0);
1209 : }
1210 :
1211 : /*
1212 : * If we have an exclusion clause that requires knowing the boundaries of
1213 : * the current row's peer group, we create a read pointer to track the
1214 : * tail position of the peer group (i.e., first row of the next peer
1215 : * group). The head position does not require its own pointer because we
1216 : * maintain that as a side effect of advancing the current row.
1217 : */
1218 3288 : winstate->grouptail_ptr = -1;
1219 :
1220 3288 : if ((frameOptions & (FRAMEOPTION_EXCLUDE_GROUP |
1221 288 : FRAMEOPTION_EXCLUDE_TIES)) &&
1222 288 : node->ordNumCols != 0)
1223 : {
1224 276 : winstate->grouptail_ptr =
1225 276 : tuplestore_alloc_read_pointer(winstate->buffer, 0);
1226 : }
1227 :
1228 : /*
1229 : * Store the first tuple into the tuplestore (it's always available now;
1230 : * we either read it above, or saved it at the end of previous partition)
1231 : */
1232 3288 : tuplestore_puttupleslot(winstate->buffer, winstate->first_part_slot);
1233 3288 : winstate->spooled_rows++;
1234 : }
1235 :
1236 : /*
1237 : * Read tuples from the outer node, up to and including position 'pos', and
1238 : * store them into the tuplestore. If pos is -1, reads the whole partition.
1239 : */
1240 : static void
1241 1617774 : spool_tuples(WindowAggState *winstate, int64 pos)
1242 : {
1243 1617774 : WindowAgg *node = (WindowAgg *) winstate->ss.ps.plan;
1244 : PlanState *outerPlan;
1245 : TupleTableSlot *outerslot;
1246 : MemoryContext oldcontext;
1247 :
1248 1617774 : if (!winstate->buffer)
1249 18 : return; /* just a safety check */
1250 1617756 : if (winstate->partition_spooled)
1251 82516 : return; /* whole partition done already */
1252 :
1253 : /*
1254 : * When in pass-through mode we can just exhaust all tuples in the current
1255 : * partition. We don't need these tuples for any further window function
1256 : * evaluation, however, we do need to keep them around if we're not the
1257 : * top-level window as another WindowAgg node above must see these.
1258 : */
1259 1535240 : if (winstate->status != WINDOWAGG_RUN)
1260 : {
1261 : Assert(winstate->status == WINDOWAGG_PASSTHROUGH ||
1262 : winstate->status == WINDOWAGG_PASSTHROUGH_STRICT);
1263 :
1264 18 : pos = -1;
1265 : }
1266 :
1267 : /*
1268 : * If the tuplestore has spilled to disk, alternate reading and writing
1269 : * becomes quite expensive due to frequent buffer flushes. It's cheaper
1270 : * to force the entire partition to get spooled in one go.
1271 : *
1272 : * XXX this is a horrid kluge --- it'd be better to fix the performance
1273 : * problem inside tuplestore. FIXME
1274 : */
1275 1535222 : else if (!tuplestore_in_memory(winstate->buffer))
1276 0 : pos = -1;
1277 :
1278 1535240 : outerPlan = outerPlanState(winstate);
1279 :
1280 : /* Must be in query context to call outerplan */
1281 1535240 : oldcontext = MemoryContextSwitchTo(winstate->ss.ps.ps_ExprContext->ecxt_per_query_memory);
1282 :
1283 2454314 : while (winstate->spooled_rows <= pos || pos == -1)
1284 : {
1285 922212 : outerslot = ExecProcNode(outerPlan);
1286 922212 : if (TupIsNull(outerslot))
1287 : {
1288 : /* reached the end of the last partition */
1289 1932 : winstate->partition_spooled = true;
1290 1932 : winstate->more_partitions = false;
1291 1932 : break;
1292 : }
1293 :
1294 920280 : if (node->partNumCols > 0)
1295 : {
1296 123624 : ExprContext *econtext = winstate->tmpcontext;
1297 :
1298 123624 : econtext->ecxt_innertuple = winstate->first_part_slot;
1299 123624 : econtext->ecxt_outertuple = outerslot;
1300 :
1301 : /* Check if this tuple still belongs to the current partition */
1302 123624 : if (!ExecQualAndReset(winstate->partEqfunction, econtext))
1303 : {
1304 : /*
1305 : * end of partition; copy the tuple for the next cycle.
1306 : */
1307 1206 : ExecCopySlot(winstate->first_part_slot, outerslot);
1308 1206 : winstate->partition_spooled = true;
1309 1206 : winstate->more_partitions = true;
1310 1206 : break;
1311 : }
1312 : }
1313 :
1314 : /*
1315 : * Remember the tuple unless we're the top-level window and we're in
1316 : * pass-through mode.
1317 : */
1318 919074 : if (winstate->status != WINDOWAGG_PASSTHROUGH_STRICT)
1319 : {
1320 : /* Still in partition, so save it into the tuplestore */
1321 919062 : tuplestore_puttupleslot(winstate->buffer, outerslot);
1322 919062 : winstate->spooled_rows++;
1323 : }
1324 : }
1325 :
1326 1535240 : MemoryContextSwitchTo(oldcontext);
1327 : }
1328 :
1329 : /*
1330 : * release_partition
1331 : * clear information kept within a partition, including
1332 : * tuplestore and aggregate results.
1333 : */
1334 : static void
1335 5478 : release_partition(WindowAggState *winstate)
1336 : {
1337 : int i;
1338 :
1339 12282 : for (i = 0; i < winstate->numfuncs; i++)
1340 : {
1341 6804 : WindowStatePerFunc perfuncstate = &(winstate->perfunc[i]);
1342 :
1343 : /* Release any partition-local state of this window function */
1344 6804 : if (perfuncstate->winobj)
1345 3390 : perfuncstate->winobj->localmem = NULL;
1346 : }
1347 :
1348 : /*
1349 : * Release all partition-local memory (in particular, any partition-local
1350 : * state that we might have trashed our pointers to in the above loop, and
1351 : * any aggregate temp data). We don't rely on retail pfree because some
1352 : * aggregates might have allocated data we don't have direct pointers to.
1353 : */
1354 5478 : MemoryContextReset(winstate->partcontext);
1355 5478 : MemoryContextReset(winstate->aggcontext);
1356 8892 : for (i = 0; i < winstate->numaggs; i++)
1357 : {
1358 3414 : if (winstate->peragg[i].aggcontext != winstate->aggcontext)
1359 2984 : MemoryContextReset(winstate->peragg[i].aggcontext);
1360 : }
1361 :
1362 5478 : if (winstate->buffer)
1363 3156 : tuplestore_end(winstate->buffer);
1364 5478 : winstate->buffer = NULL;
1365 5478 : winstate->partition_spooled = false;
1366 5478 : }
1367 :
1368 : /*
1369 : * row_is_in_frame
1370 : * Determine whether a row is in the current row's window frame according
1371 : * to our window framing rule
1372 : *
1373 : * The caller must have already determined that the row is in the partition
1374 : * and fetched it into a slot. This function just encapsulates the framing
1375 : * rules.
1376 : *
1377 : * Returns:
1378 : * -1, if the row is out of frame and no succeeding rows can be in frame
1379 : * 0, if the row is out of frame but succeeding rows might be in frame
1380 : * 1, if the row is in frame
1381 : *
1382 : * May clobber winstate->temp_slot_2.
1383 : */
1384 : static int
1385 160518 : row_is_in_frame(WindowAggState *winstate, int64 pos, TupleTableSlot *slot)
1386 : {
1387 160518 : int frameOptions = winstate->frameOptions;
1388 :
1389 : Assert(pos >= 0); /* else caller error */
1390 :
1391 : /*
1392 : * First, check frame starting conditions. We might as well delegate this
1393 : * to update_frameheadpos always; it doesn't add any notable cost.
1394 : */
1395 160518 : update_frameheadpos(winstate);
1396 160518 : if (pos < winstate->frameheadpos)
1397 144 : return 0;
1398 :
1399 : /*
1400 : * Okay so far, now check frame ending conditions. Here, we avoid calling
1401 : * update_frametailpos in simple cases, so as not to spool tuples further
1402 : * ahead than necessary.
1403 : */
1404 160374 : if (frameOptions & FRAMEOPTION_END_CURRENT_ROW)
1405 : {
1406 127982 : if (frameOptions & FRAMEOPTION_ROWS)
1407 : {
1408 : /* rows after current row are out of frame */
1409 2208 : if (pos > winstate->currentpos)
1410 972 : return -1;
1411 : }
1412 125774 : else if (frameOptions & (FRAMEOPTION_RANGE | FRAMEOPTION_GROUPS))
1413 : {
1414 : /* following row that is not peer is out of frame */
1415 125774 : if (pos > winstate->currentpos &&
1416 122450 : !are_peers(winstate, slot, winstate->ss.ss_ScanTupleSlot))
1417 1264 : return -1;
1418 : }
1419 : else
1420 : Assert(false);
1421 : }
1422 32392 : else if (frameOptions & FRAMEOPTION_END_OFFSET)
1423 : {
1424 17934 : if (frameOptions & FRAMEOPTION_ROWS)
1425 : {
1426 3948 : int64 offset = DatumGetInt64(winstate->endOffsetValue);
1427 :
1428 : /* rows after current row + offset are out of frame */
1429 3948 : if (frameOptions & FRAMEOPTION_END_OFFSET_PRECEDING)
1430 114 : offset = -offset;
1431 :
1432 3948 : if (pos > winstate->currentpos + offset)
1433 1152 : return -1;
1434 : }
1435 13986 : else if (frameOptions & (FRAMEOPTION_RANGE | FRAMEOPTION_GROUPS))
1436 : {
1437 : /* hard cases, so delegate to update_frametailpos */
1438 13986 : update_frametailpos(winstate);
1439 13944 : if (pos >= winstate->frametailpos)
1440 1470 : return -1;
1441 : }
1442 : else
1443 : Assert(false);
1444 : }
1445 :
1446 : /* Check exclusion clause */
1447 155474 : if (frameOptions & FRAMEOPTION_EXCLUDE_CURRENT_ROW)
1448 : {
1449 2466 : if (pos == winstate->currentpos)
1450 420 : return 0;
1451 : }
1452 153008 : else if ((frameOptions & FRAMEOPTION_EXCLUDE_GROUP) ||
1453 150146 : ((frameOptions & FRAMEOPTION_EXCLUDE_TIES) &&
1454 2970 : pos != winstate->currentpos))
1455 : {
1456 5292 : WindowAgg *node = (WindowAgg *) winstate->ss.ps.plan;
1457 :
1458 : /* If no ORDER BY, all rows are peers with each other */
1459 5292 : if (node->ordNumCols == 0)
1460 468 : return 0;
1461 : /* Otherwise, check the group boundaries */
1462 4824 : if (pos >= winstate->groupheadpos)
1463 : {
1464 2592 : update_grouptailpos(winstate);
1465 2592 : if (pos < winstate->grouptailpos)
1466 1008 : return 0;
1467 : }
1468 : }
1469 :
1470 : /* If we get here, it's in frame */
1471 153578 : return 1;
1472 : }
1473 :
1474 : /*
1475 : * update_frameheadpos
1476 : * make frameheadpos valid for the current row
1477 : *
1478 : * Note that frameheadpos is computed without regard for any window exclusion
1479 : * clause; the current row and/or its peers are considered part of the frame
1480 : * for this purpose even if they must be excluded later.
1481 : *
1482 : * May clobber winstate->temp_slot_2.
1483 : */
1484 : static void
1485 301544 : update_frameheadpos(WindowAggState *winstate)
1486 : {
1487 301544 : WindowAgg *node = (WindowAgg *) winstate->ss.ps.plan;
1488 301544 : int frameOptions = winstate->frameOptions;
1489 : MemoryContext oldcontext;
1490 :
1491 301544 : if (winstate->framehead_valid)
1492 166706 : return; /* already known for current row */
1493 :
1494 : /* We may be called in a short-lived context */
1495 134838 : oldcontext = MemoryContextSwitchTo(winstate->ss.ps.ps_ExprContext->ecxt_per_query_memory);
1496 :
1497 134838 : if (frameOptions & FRAMEOPTION_START_UNBOUNDED_PRECEDING)
1498 : {
1499 : /* In UNBOUNDED PRECEDING mode, frame head is always row 0 */
1500 124558 : winstate->frameheadpos = 0;
1501 124558 : winstate->framehead_valid = true;
1502 : }
1503 10280 : else if (frameOptions & FRAMEOPTION_START_CURRENT_ROW)
1504 : {
1505 2804 : if (frameOptions & FRAMEOPTION_ROWS)
1506 : {
1507 : /* In ROWS mode, frame head is the same as current */
1508 2376 : winstate->frameheadpos = winstate->currentpos;
1509 2376 : winstate->framehead_valid = true;
1510 : }
1511 428 : else if (frameOptions & (FRAMEOPTION_RANGE | FRAMEOPTION_GROUPS))
1512 : {
1513 : /* If no ORDER BY, all rows are peers with each other */
1514 428 : if (node->ordNumCols == 0)
1515 : {
1516 0 : winstate->frameheadpos = 0;
1517 0 : winstate->framehead_valid = true;
1518 0 : MemoryContextSwitchTo(oldcontext);
1519 0 : return;
1520 : }
1521 :
1522 : /*
1523 : * In RANGE or GROUPS START_CURRENT_ROW mode, frame head is the
1524 : * first row that is a peer of current row. We keep a copy of the
1525 : * last-known frame head row in framehead_slot, and advance as
1526 : * necessary. Note that if we reach end of partition, we will
1527 : * leave frameheadpos = end+1 and framehead_slot empty.
1528 : */
1529 428 : tuplestore_select_read_pointer(winstate->buffer,
1530 : winstate->framehead_ptr);
1531 428 : if (winstate->frameheadpos == 0 &&
1532 212 : TupIsNull(winstate->framehead_slot))
1533 : {
1534 : /* fetch first row into framehead_slot, if we didn't already */
1535 82 : if (!tuplestore_gettupleslot(winstate->buffer, true, true,
1536 : winstate->framehead_slot))
1537 0 : elog(ERROR, "unexpected end of tuplestore");
1538 : }
1539 :
1540 744 : while (!TupIsNull(winstate->framehead_slot))
1541 : {
1542 744 : if (are_peers(winstate, winstate->framehead_slot,
1543 : winstate->ss.ss_ScanTupleSlot))
1544 428 : break; /* this row is the correct frame head */
1545 : /* Note we advance frameheadpos even if the fetch fails */
1546 316 : winstate->frameheadpos++;
1547 316 : spool_tuples(winstate, winstate->frameheadpos);
1548 316 : if (!tuplestore_gettupleslot(winstate->buffer, true, true,
1549 : winstate->framehead_slot))
1550 0 : break; /* end of partition */
1551 : }
1552 428 : winstate->framehead_valid = true;
1553 : }
1554 : else
1555 : Assert(false);
1556 : }
1557 7476 : else if (frameOptions & FRAMEOPTION_START_OFFSET)
1558 : {
1559 7476 : if (frameOptions & FRAMEOPTION_ROWS)
1560 : {
1561 : /* In ROWS mode, bound is physically n before/after current */
1562 1572 : int64 offset = DatumGetInt64(winstate->startOffsetValue);
1563 :
1564 1572 : if (frameOptions & FRAMEOPTION_START_OFFSET_PRECEDING)
1565 1512 : offset = -offset;
1566 :
1567 1572 : winstate->frameheadpos = winstate->currentpos + offset;
1568 : /* frame head can't go before first row */
1569 1572 : if (winstate->frameheadpos < 0)
1570 228 : winstate->frameheadpos = 0;
1571 1344 : else if (winstate->frameheadpos > winstate->currentpos + 1)
1572 : {
1573 : /* make sure frameheadpos is not past end of partition */
1574 0 : spool_tuples(winstate, winstate->frameheadpos - 1);
1575 0 : if (winstate->frameheadpos > winstate->spooled_rows)
1576 0 : winstate->frameheadpos = winstate->spooled_rows;
1577 : }
1578 1572 : winstate->framehead_valid = true;
1579 : }
1580 5904 : else if (frameOptions & FRAMEOPTION_RANGE)
1581 : {
1582 : /*
1583 : * In RANGE START_OFFSET mode, frame head is the first row that
1584 : * satisfies the in_range constraint relative to the current row.
1585 : * We keep a copy of the last-known frame head row in
1586 : * framehead_slot, and advance as necessary. Note that if we
1587 : * reach end of partition, we will leave frameheadpos = end+1 and
1588 : * framehead_slot empty.
1589 : */
1590 4524 : int sortCol = node->ordColIdx[0];
1591 : bool sub,
1592 : less;
1593 :
1594 : /* We must have an ordering column */
1595 : Assert(node->ordNumCols == 1);
1596 :
1597 : /* Precompute flags for in_range checks */
1598 4524 : if (frameOptions & FRAMEOPTION_START_OFFSET_PRECEDING)
1599 3702 : sub = true; /* subtract startOffset from current row */
1600 : else
1601 822 : sub = false; /* add it */
1602 4524 : less = false; /* normally, we want frame head >= sum */
1603 : /* If sort order is descending, flip both flags */
1604 4524 : if (!winstate->inRangeAsc)
1605 : {
1606 654 : sub = !sub;
1607 654 : less = true;
1608 : }
1609 :
1610 4524 : tuplestore_select_read_pointer(winstate->buffer,
1611 : winstate->framehead_ptr);
1612 4524 : if (winstate->frameheadpos == 0 &&
1613 2502 : TupIsNull(winstate->framehead_slot))
1614 : {
1615 : /* fetch first row into framehead_slot, if we didn't already */
1616 570 : if (!tuplestore_gettupleslot(winstate->buffer, true, true,
1617 : winstate->framehead_slot))
1618 0 : elog(ERROR, "unexpected end of tuplestore");
1619 : }
1620 :
1621 7266 : while (!TupIsNull(winstate->framehead_slot))
1622 : {
1623 : Datum headval,
1624 : currval;
1625 : bool headisnull,
1626 : currisnull;
1627 :
1628 7062 : headval = slot_getattr(winstate->framehead_slot, sortCol,
1629 : &headisnull);
1630 7062 : currval = slot_getattr(winstate->ss.ss_ScanTupleSlot, sortCol,
1631 : &currisnull);
1632 7062 : if (headisnull || currisnull)
1633 : {
1634 : /* order of the rows depends only on nulls_first */
1635 108 : if (winstate->inRangeNullsFirst)
1636 : {
1637 : /* advance head if head is null and curr is not */
1638 48 : if (!headisnull || currisnull)
1639 : break;
1640 : }
1641 : else
1642 : {
1643 : /* advance head if head is not null and curr is null */
1644 60 : if (headisnull || !currisnull)
1645 : break;
1646 : }
1647 : }
1648 : else
1649 : {
1650 6954 : if (DatumGetBool(FunctionCall5Coll(&winstate->startInRangeFunc,
1651 : winstate->inRangeColl,
1652 : headval,
1653 : currval,
1654 : winstate->startOffsetValue,
1655 : BoolGetDatum(sub),
1656 : BoolGetDatum(less))))
1657 4170 : break; /* this row is the correct frame head */
1658 : }
1659 : /* Note we advance frameheadpos even if the fetch fails */
1660 2796 : winstate->frameheadpos++;
1661 2796 : spool_tuples(winstate, winstate->frameheadpos);
1662 2796 : if (!tuplestore_gettupleslot(winstate->buffer, true, true,
1663 : winstate->framehead_slot))
1664 54 : break; /* end of partition */
1665 : }
1666 4476 : winstate->framehead_valid = true;
1667 : }
1668 1380 : else if (frameOptions & FRAMEOPTION_GROUPS)
1669 : {
1670 : /*
1671 : * In GROUPS START_OFFSET mode, frame head is the first row of the
1672 : * first peer group whose number satisfies the offset constraint.
1673 : * We keep a copy of the last-known frame head row in
1674 : * framehead_slot, and advance as necessary. Note that if we
1675 : * reach end of partition, we will leave frameheadpos = end+1 and
1676 : * framehead_slot empty.
1677 : */
1678 1380 : int64 offset = DatumGetInt64(winstate->startOffsetValue);
1679 : int64 minheadgroup;
1680 :
1681 1380 : if (frameOptions & FRAMEOPTION_START_OFFSET_PRECEDING)
1682 1128 : minheadgroup = winstate->currentgroup - offset;
1683 : else
1684 252 : minheadgroup = winstate->currentgroup + offset;
1685 :
1686 1380 : tuplestore_select_read_pointer(winstate->buffer,
1687 : winstate->framehead_ptr);
1688 1380 : if (winstate->frameheadpos == 0 &&
1689 750 : TupIsNull(winstate->framehead_slot))
1690 : {
1691 : /* fetch first row into framehead_slot, if we didn't already */
1692 372 : if (!tuplestore_gettupleslot(winstate->buffer, true, true,
1693 : winstate->framehead_slot))
1694 0 : elog(ERROR, "unexpected end of tuplestore");
1695 : }
1696 :
1697 2142 : while (!TupIsNull(winstate->framehead_slot))
1698 : {
1699 2118 : if (winstate->frameheadgroup >= minheadgroup)
1700 1320 : break; /* this row is the correct frame head */
1701 798 : ExecCopySlot(winstate->temp_slot_2, winstate->framehead_slot);
1702 : /* Note we advance frameheadpos even if the fetch fails */
1703 798 : winstate->frameheadpos++;
1704 798 : spool_tuples(winstate, winstate->frameheadpos);
1705 798 : if (!tuplestore_gettupleslot(winstate->buffer, true, true,
1706 : winstate->framehead_slot))
1707 36 : break; /* end of partition */
1708 762 : if (!are_peers(winstate, winstate->temp_slot_2,
1709 : winstate->framehead_slot))
1710 522 : winstate->frameheadgroup++;
1711 : }
1712 1380 : ExecClearTuple(winstate->temp_slot_2);
1713 1380 : winstate->framehead_valid = true;
1714 : }
1715 : else
1716 : Assert(false);
1717 : }
1718 : else
1719 : Assert(false);
1720 :
1721 134790 : MemoryContextSwitchTo(oldcontext);
1722 : }
1723 :
1724 : /*
1725 : * update_frametailpos
1726 : * make frametailpos valid for the current row
1727 : *
1728 : * Note that frametailpos is computed without regard for any window exclusion
1729 : * clause; the current row and/or its peers are considered part of the frame
1730 : * for this purpose even if they must be excluded later.
1731 : *
1732 : * May clobber winstate->temp_slot_2.
1733 : */
1734 : static void
1735 148688 : update_frametailpos(WindowAggState *winstate)
1736 : {
1737 148688 : WindowAgg *node = (WindowAgg *) winstate->ss.ps.plan;
1738 148688 : int frameOptions = winstate->frameOptions;
1739 : MemoryContext oldcontext;
1740 :
1741 148688 : if (winstate->frametail_valid)
1742 17820 : return; /* already known for current row */
1743 :
1744 : /* We may be called in a short-lived context */
1745 130868 : oldcontext = MemoryContextSwitchTo(winstate->ss.ps.ps_ExprContext->ecxt_per_query_memory);
1746 :
1747 130868 : if (frameOptions & FRAMEOPTION_END_UNBOUNDED_FOLLOWING)
1748 : {
1749 : /* In UNBOUNDED FOLLOWING mode, all partition rows are in frame */
1750 180 : spool_tuples(winstate, -1);
1751 180 : winstate->frametailpos = winstate->spooled_rows;
1752 180 : winstate->frametail_valid = true;
1753 : }
1754 130688 : else if (frameOptions & FRAMEOPTION_END_CURRENT_ROW)
1755 : {
1756 124304 : if (frameOptions & FRAMEOPTION_ROWS)
1757 : {
1758 : /* In ROWS mode, exactly the rows up to current are in frame */
1759 120 : winstate->frametailpos = winstate->currentpos + 1;
1760 120 : winstate->frametail_valid = true;
1761 : }
1762 124184 : else if (frameOptions & (FRAMEOPTION_RANGE | FRAMEOPTION_GROUPS))
1763 : {
1764 : /* If no ORDER BY, all rows are peers with each other */
1765 124184 : if (node->ordNumCols == 0)
1766 : {
1767 60 : spool_tuples(winstate, -1);
1768 60 : winstate->frametailpos = winstate->spooled_rows;
1769 60 : winstate->frametail_valid = true;
1770 60 : MemoryContextSwitchTo(oldcontext);
1771 60 : return;
1772 : }
1773 :
1774 : /*
1775 : * In RANGE or GROUPS END_CURRENT_ROW mode, frame end is the last
1776 : * row that is a peer of current row, frame tail is the row after
1777 : * that (if any). We keep a copy of the last-known frame tail row
1778 : * in frametail_slot, and advance as necessary. Note that if we
1779 : * reach end of partition, we will leave frametailpos = end+1 and
1780 : * frametail_slot empty.
1781 : */
1782 124124 : tuplestore_select_read_pointer(winstate->buffer,
1783 : winstate->frametail_ptr);
1784 124124 : if (winstate->frametailpos == 0 &&
1785 676 : TupIsNull(winstate->frametail_slot))
1786 : {
1787 : /* fetch first row into frametail_slot, if we didn't already */
1788 676 : if (!tuplestore_gettupleslot(winstate->buffer, true, true,
1789 : winstate->frametail_slot))
1790 0 : elog(ERROR, "unexpected end of tuplestore");
1791 : }
1792 :
1793 247584 : while (!TupIsNull(winstate->frametail_slot))
1794 : {
1795 223464 : if (winstate->frametailpos > winstate->currentpos &&
1796 219700 : !are_peers(winstate, winstate->frametail_slot,
1797 : winstate->ss.ss_ScanTupleSlot))
1798 99340 : break; /* this row is the frame tail */
1799 : /* Note we advance frametailpos even if the fetch fails */
1800 124124 : winstate->frametailpos++;
1801 124124 : spool_tuples(winstate, winstate->frametailpos);
1802 124124 : if (!tuplestore_gettupleslot(winstate->buffer, true, true,
1803 : winstate->frametail_slot))
1804 664 : break; /* end of partition */
1805 : }
1806 124124 : winstate->frametail_valid = true;
1807 : }
1808 : else
1809 : Assert(false);
1810 : }
1811 6384 : else if (frameOptions & FRAMEOPTION_END_OFFSET)
1812 : {
1813 6384 : if (frameOptions & FRAMEOPTION_ROWS)
1814 : {
1815 : /* In ROWS mode, bound is physically n before/after current */
1816 180 : int64 offset = DatumGetInt64(winstate->endOffsetValue);
1817 :
1818 180 : if (frameOptions & FRAMEOPTION_END_OFFSET_PRECEDING)
1819 0 : offset = -offset;
1820 :
1821 180 : winstate->frametailpos = winstate->currentpos + offset + 1;
1822 : /* smallest allowable value of frametailpos is 0 */
1823 180 : if (winstate->frametailpos < 0)
1824 0 : winstate->frametailpos = 0;
1825 180 : else if (winstate->frametailpos > winstate->currentpos + 1)
1826 : {
1827 : /* make sure frametailpos is not past end of partition */
1828 180 : spool_tuples(winstate, winstate->frametailpos - 1);
1829 180 : if (winstate->frametailpos > winstate->spooled_rows)
1830 36 : winstate->frametailpos = winstate->spooled_rows;
1831 : }
1832 180 : winstate->frametail_valid = true;
1833 : }
1834 6204 : else if (frameOptions & FRAMEOPTION_RANGE)
1835 : {
1836 : /*
1837 : * In RANGE END_OFFSET mode, frame end is the last row that
1838 : * satisfies the in_range constraint relative to the current row,
1839 : * frame tail is the row after that (if any). We keep a copy of
1840 : * the last-known frame tail row in frametail_slot, and advance as
1841 : * necessary. Note that if we reach end of partition, we will
1842 : * leave frametailpos = end+1 and frametail_slot empty.
1843 : */
1844 4884 : int sortCol = node->ordColIdx[0];
1845 : bool sub,
1846 : less;
1847 :
1848 : /* We must have an ordering column */
1849 : Assert(node->ordNumCols == 1);
1850 :
1851 : /* Precompute flags for in_range checks */
1852 4884 : if (frameOptions & FRAMEOPTION_END_OFFSET_PRECEDING)
1853 912 : sub = true; /* subtract endOffset from current row */
1854 : else
1855 3972 : sub = false; /* add it */
1856 4884 : less = true; /* normally, we want frame tail <= sum */
1857 : /* If sort order is descending, flip both flags */
1858 4884 : if (!winstate->inRangeAsc)
1859 : {
1860 690 : sub = !sub;
1861 690 : less = false;
1862 : }
1863 :
1864 4884 : tuplestore_select_read_pointer(winstate->buffer,
1865 : winstate->frametail_ptr);
1866 4884 : if (winstate->frametailpos == 0 &&
1867 822 : TupIsNull(winstate->frametail_slot))
1868 : {
1869 : /* fetch first row into frametail_slot, if we didn't already */
1870 588 : if (!tuplestore_gettupleslot(winstate->buffer, true, true,
1871 : winstate->frametail_slot))
1872 0 : elog(ERROR, "unexpected end of tuplestore");
1873 : }
1874 :
1875 9006 : while (!TupIsNull(winstate->frametail_slot))
1876 : {
1877 : Datum tailval,
1878 : currval;
1879 : bool tailisnull,
1880 : currisnull;
1881 :
1882 7440 : tailval = slot_getattr(winstate->frametail_slot, sortCol,
1883 : &tailisnull);
1884 7440 : currval = slot_getattr(winstate->ss.ss_ScanTupleSlot, sortCol,
1885 : &currisnull);
1886 7440 : if (tailisnull || currisnull)
1887 : {
1888 : /* order of the rows depends only on nulls_first */
1889 108 : if (winstate->inRangeNullsFirst)
1890 : {
1891 : /* advance tail if tail is null or curr is not */
1892 48 : if (!tailisnull)
1893 3270 : break;
1894 : }
1895 : else
1896 : {
1897 : /* advance tail if tail is not null or curr is null */
1898 60 : if (!currisnull)
1899 36 : break;
1900 : }
1901 : }
1902 : else
1903 : {
1904 7332 : if (!DatumGetBool(FunctionCall5Coll(&winstate->endInRangeFunc,
1905 : winstate->inRangeColl,
1906 : tailval,
1907 : currval,
1908 : winstate->endOffsetValue,
1909 : BoolGetDatum(sub),
1910 : BoolGetDatum(less))))
1911 2730 : break; /* this row is the correct frame tail */
1912 : }
1913 : /* Note we advance frametailpos even if the fetch fails */
1914 4602 : winstate->frametailpos++;
1915 4602 : spool_tuples(winstate, winstate->frametailpos);
1916 4602 : if (!tuplestore_gettupleslot(winstate->buffer, true, true,
1917 : winstate->frametail_slot))
1918 480 : break; /* end of partition */
1919 : }
1920 4836 : winstate->frametail_valid = true;
1921 : }
1922 1320 : else if (frameOptions & FRAMEOPTION_GROUPS)
1923 : {
1924 : /*
1925 : * In GROUPS END_OFFSET mode, frame end is the last row of the
1926 : * last peer group whose number satisfies the offset constraint,
1927 : * and frame tail is the row after that (if any). We keep a copy
1928 : * of the last-known frame tail row in frametail_slot, and advance
1929 : * as necessary. Note that if we reach end of partition, we will
1930 : * leave frametailpos = end+1 and frametail_slot empty.
1931 : */
1932 1320 : int64 offset = DatumGetInt64(winstate->endOffsetValue);
1933 : int64 maxtailgroup;
1934 :
1935 1320 : if (frameOptions & FRAMEOPTION_END_OFFSET_PRECEDING)
1936 72 : maxtailgroup = winstate->currentgroup - offset;
1937 : else
1938 1248 : maxtailgroup = winstate->currentgroup + offset;
1939 :
1940 1320 : tuplestore_select_read_pointer(winstate->buffer,
1941 : winstate->frametail_ptr);
1942 1320 : if (winstate->frametailpos == 0 &&
1943 384 : TupIsNull(winstate->frametail_slot))
1944 : {
1945 : /* fetch first row into frametail_slot, if we didn't already */
1946 366 : if (!tuplestore_gettupleslot(winstate->buffer, true, true,
1947 : winstate->frametail_slot))
1948 0 : elog(ERROR, "unexpected end of tuplestore");
1949 : }
1950 :
1951 2268 : while (!TupIsNull(winstate->frametail_slot))
1952 : {
1953 2040 : if (winstate->frametailgroup > maxtailgroup)
1954 744 : break; /* this row is the correct frame tail */
1955 1296 : ExecCopySlot(winstate->temp_slot_2, winstate->frametail_slot);
1956 : /* Note we advance frametailpos even if the fetch fails */
1957 1296 : winstate->frametailpos++;
1958 1296 : spool_tuples(winstate, winstate->frametailpos);
1959 1296 : if (!tuplestore_gettupleslot(winstate->buffer, true, true,
1960 : winstate->frametail_slot))
1961 348 : break; /* end of partition */
1962 948 : if (!are_peers(winstate, winstate->temp_slot_2,
1963 : winstate->frametail_slot))
1964 600 : winstate->frametailgroup++;
1965 : }
1966 1320 : ExecClearTuple(winstate->temp_slot_2);
1967 1320 : winstate->frametail_valid = true;
1968 : }
1969 : else
1970 : Assert(false);
1971 : }
1972 : else
1973 : Assert(false);
1974 :
1975 130760 : MemoryContextSwitchTo(oldcontext);
1976 : }
1977 :
1978 : /*
1979 : * update_grouptailpos
1980 : * make grouptailpos valid for the current row
1981 : *
1982 : * May clobber winstate->temp_slot_2.
1983 : */
1984 : static void
1985 4872 : update_grouptailpos(WindowAggState *winstate)
1986 : {
1987 4872 : WindowAgg *node = (WindowAgg *) winstate->ss.ps.plan;
1988 : MemoryContext oldcontext;
1989 :
1990 4872 : if (winstate->grouptail_valid)
1991 3954 : return; /* already known for current row */
1992 :
1993 : /* We may be called in a short-lived context */
1994 918 : oldcontext = MemoryContextSwitchTo(winstate->ss.ps.ps_ExprContext->ecxt_per_query_memory);
1995 :
1996 : /* If no ORDER BY, all rows are peers with each other */
1997 918 : if (node->ordNumCols == 0)
1998 : {
1999 0 : spool_tuples(winstate, -1);
2000 0 : winstate->grouptailpos = winstate->spooled_rows;
2001 0 : winstate->grouptail_valid = true;
2002 0 : MemoryContextSwitchTo(oldcontext);
2003 0 : return;
2004 : }
2005 :
2006 : /*
2007 : * Because grouptail_valid is reset only when current row advances into a
2008 : * new peer group, we always reach here knowing that grouptailpos needs to
2009 : * be advanced by at least one row. Hence, unlike the otherwise similar
2010 : * case for frame tail tracking, we do not need persistent storage of the
2011 : * group tail row.
2012 : */
2013 : Assert(winstate->grouptailpos <= winstate->currentpos);
2014 918 : tuplestore_select_read_pointer(winstate->buffer,
2015 : winstate->grouptail_ptr);
2016 : for (;;)
2017 : {
2018 : /* Note we advance grouptailpos even if the fetch fails */
2019 1758 : winstate->grouptailpos++;
2020 1758 : spool_tuples(winstate, winstate->grouptailpos);
2021 1758 : if (!tuplestore_gettupleslot(winstate->buffer, true, true,
2022 : winstate->temp_slot_2))
2023 258 : break; /* end of partition */
2024 1500 : if (winstate->grouptailpos > winstate->currentpos &&
2025 1242 : !are_peers(winstate, winstate->temp_slot_2,
2026 : winstate->ss.ss_ScanTupleSlot))
2027 660 : break; /* this row is the group tail */
2028 : }
2029 918 : ExecClearTuple(winstate->temp_slot_2);
2030 918 : winstate->grouptail_valid = true;
2031 :
2032 918 : MemoryContextSwitchTo(oldcontext);
2033 : }
2034 :
2035 :
2036 : /* -----------------
2037 : * ExecWindowAgg
2038 : *
2039 : * ExecWindowAgg receives tuples from its outer subplan and
2040 : * stores them into a tuplestore, then processes window functions.
2041 : * This node doesn't reduce nor qualify any row so the number of
2042 : * returned rows is exactly the same as its outer subplan's result.
2043 : * -----------------
2044 : */
2045 : static TupleTableSlot *
2046 804102 : ExecWindowAgg(PlanState *pstate)
2047 : {
2048 804102 : WindowAggState *winstate = castNode(WindowAggState, pstate);
2049 : TupleTableSlot *slot;
2050 : ExprContext *econtext;
2051 : int i;
2052 : int numfuncs;
2053 :
2054 804102 : CHECK_FOR_INTERRUPTS();
2055 :
2056 804102 : if (winstate->status == WINDOWAGG_DONE)
2057 0 : return NULL;
2058 :
2059 : /*
2060 : * Compute frame offset values, if any, during first call (or after a
2061 : * rescan). These are assumed to hold constant throughout the scan; if
2062 : * user gives us a volatile expression, we'll only use its initial value.
2063 : */
2064 804102 : if (winstate->all_first)
2065 : {
2066 2100 : int frameOptions = winstate->frameOptions;
2067 : Datum value;
2068 : bool isnull;
2069 : int16 len;
2070 : bool byval;
2071 :
2072 2100 : econtext = winstate->ss.ps.ps_ExprContext;
2073 :
2074 2100 : if (frameOptions & FRAMEOPTION_START_OFFSET)
2075 : {
2076 : Assert(winstate->startOffset != NULL);
2077 816 : value = ExecEvalExprSwitchContext(winstate->startOffset,
2078 : econtext,
2079 : &isnull);
2080 816 : if (isnull)
2081 0 : ereport(ERROR,
2082 : (errcode(ERRCODE_NULL_VALUE_NOT_ALLOWED),
2083 : errmsg("frame starting offset must not be null")));
2084 : /* copy value into query-lifespan context */
2085 816 : get_typlenbyval(exprType((Node *) winstate->startOffset->expr),
2086 : &len, &byval);
2087 816 : winstate->startOffsetValue = datumCopy(value, byval, len);
2088 816 : if (frameOptions & (FRAMEOPTION_ROWS | FRAMEOPTION_GROUPS))
2089 : {
2090 : /* value is known to be int8 */
2091 300 : int64 offset = DatumGetInt64(value);
2092 :
2093 300 : if (offset < 0)
2094 0 : ereport(ERROR,
2095 : (errcode(ERRCODE_INVALID_PRECEDING_OR_FOLLOWING_SIZE),
2096 : errmsg("frame starting offset must not be negative")));
2097 : }
2098 : }
2099 2100 : if (frameOptions & FRAMEOPTION_END_OFFSET)
2100 : {
2101 : Assert(winstate->endOffset != NULL);
2102 912 : value = ExecEvalExprSwitchContext(winstate->endOffset,
2103 : econtext,
2104 : &isnull);
2105 912 : if (isnull)
2106 0 : ereport(ERROR,
2107 : (errcode(ERRCODE_NULL_VALUE_NOT_ALLOWED),
2108 : errmsg("frame ending offset must not be null")));
2109 : /* copy value into query-lifespan context */
2110 912 : get_typlenbyval(exprType((Node *) winstate->endOffset->expr),
2111 : &len, &byval);
2112 912 : winstate->endOffsetValue = datumCopy(value, byval, len);
2113 912 : if (frameOptions & (FRAMEOPTION_ROWS | FRAMEOPTION_GROUPS))
2114 : {
2115 : /* value is known to be int8 */
2116 330 : int64 offset = DatumGetInt64(value);
2117 :
2118 330 : if (offset < 0)
2119 0 : ereport(ERROR,
2120 : (errcode(ERRCODE_INVALID_PRECEDING_OR_FOLLOWING_SIZE),
2121 : errmsg("frame ending offset must not be negative")));
2122 : }
2123 : }
2124 2100 : winstate->all_first = false;
2125 : }
2126 :
2127 : /* We need to loop as the runCondition or qual may filter out tuples */
2128 : for (;;)
2129 : {
2130 804174 : if (winstate->buffer == NULL)
2131 : {
2132 : /* Initialize for first partition and set current row = 0 */
2133 2100 : begin_partition(winstate);
2134 : /* If there are no input rows, we'll detect that and exit below */
2135 : }
2136 : else
2137 : {
2138 : /* Advance current row within partition */
2139 802074 : winstate->currentpos++;
2140 : /* This might mean that the frame moves, too */
2141 802074 : winstate->framehead_valid = false;
2142 802074 : winstate->frametail_valid = false;
2143 : /* we don't need to invalidate grouptail here; see below */
2144 : }
2145 :
2146 : /*
2147 : * Spool all tuples up to and including the current row, if we haven't
2148 : * already
2149 : */
2150 804174 : spool_tuples(winstate, winstate->currentpos);
2151 :
2152 : /* Move to the next partition if we reached the end of this partition */
2153 804174 : if (winstate->partition_spooled &&
2154 32420 : winstate->currentpos >= winstate->spooled_rows)
2155 : {
2156 3126 : release_partition(winstate);
2157 :
2158 3126 : if (winstate->more_partitions)
2159 : {
2160 1206 : begin_partition(winstate);
2161 : Assert(winstate->spooled_rows > 0);
2162 :
2163 : /* Come out of pass-through mode when changing partition */
2164 1206 : winstate->status = WINDOWAGG_RUN;
2165 : }
2166 : else
2167 : {
2168 : /* No further partitions? We're done */
2169 1920 : winstate->status = WINDOWAGG_DONE;
2170 1920 : return NULL;
2171 : }
2172 : }
2173 :
2174 : /* final output execution is in ps_ExprContext */
2175 802254 : econtext = winstate->ss.ps.ps_ExprContext;
2176 :
2177 : /* Clear the per-output-tuple context for current row */
2178 802254 : ResetExprContext(econtext);
2179 :
2180 : /*
2181 : * Read the current row from the tuplestore, and save in
2182 : * ScanTupleSlot. (We can't rely on the outerplan's output slot
2183 : * because we may have to read beyond the current row. Also, we have
2184 : * to actually copy the row out of the tuplestore, since window
2185 : * function evaluation might cause the tuplestore to dump its state to
2186 : * disk.)
2187 : *
2188 : * In GROUPS mode, or when tracking a group-oriented exclusion clause,
2189 : * we must also detect entering a new peer group and update associated
2190 : * state when that happens. We use temp_slot_2 to temporarily hold
2191 : * the previous row for this purpose.
2192 : *
2193 : * Current row must be in the tuplestore, since we spooled it above.
2194 : */
2195 802254 : tuplestore_select_read_pointer(winstate->buffer, winstate->current_ptr);
2196 802254 : if ((winstate->frameOptions & (FRAMEOPTION_GROUPS |
2197 : FRAMEOPTION_EXCLUDE_GROUP |
2198 2898 : FRAMEOPTION_EXCLUDE_TIES)) &&
2199 2898 : winstate->currentpos > 0)
2200 : {
2201 2358 : ExecCopySlot(winstate->temp_slot_2, winstate->ss.ss_ScanTupleSlot);
2202 2358 : if (!tuplestore_gettupleslot(winstate->buffer, true, true,
2203 : winstate->ss.ss_ScanTupleSlot))
2204 0 : elog(ERROR, "unexpected end of tuplestore");
2205 2358 : if (!are_peers(winstate, winstate->temp_slot_2,
2206 : winstate->ss.ss_ScanTupleSlot))
2207 : {
2208 1242 : winstate->currentgroup++;
2209 1242 : winstate->groupheadpos = winstate->currentpos;
2210 1242 : winstate->grouptail_valid = false;
2211 : }
2212 2358 : ExecClearTuple(winstate->temp_slot_2);
2213 : }
2214 : else
2215 : {
2216 799896 : if (!tuplestore_gettupleslot(winstate->buffer, true, true,
2217 : winstate->ss.ss_ScanTupleSlot))
2218 0 : elog(ERROR, "unexpected end of tuplestore");
2219 : }
2220 :
2221 : /* don't evaluate the window functions when we're in pass-through mode */
2222 802254 : if (winstate->status == WINDOWAGG_RUN)
2223 : {
2224 : /*
2225 : * Evaluate true window functions
2226 : */
2227 802200 : numfuncs = winstate->numfuncs;
2228 1730748 : for (i = 0; i < numfuncs; i++)
2229 : {
2230 928638 : WindowStatePerFunc perfuncstate = &(winstate->perfunc[i]);
2231 :
2232 928638 : if (perfuncstate->plain_agg)
2233 131874 : continue;
2234 796764 : eval_windowfunction(winstate, perfuncstate,
2235 796764 : &(econtext->ecxt_aggvalues[perfuncstate->wfuncstate->wfuncno]),
2236 796764 : &(econtext->ecxt_aggnulls[perfuncstate->wfuncstate->wfuncno]));
2237 : }
2238 :
2239 : /*
2240 : * Evaluate aggregates
2241 : */
2242 802110 : if (winstate->numaggs > 0)
2243 130104 : eval_windowaggregates(winstate);
2244 : }
2245 :
2246 : /*
2247 : * If we have created auxiliary read pointers for the frame or group
2248 : * boundaries, force them to be kept up-to-date, because we don't know
2249 : * whether the window function(s) will do anything that requires that.
2250 : * Failing to advance the pointers would result in being unable to
2251 : * trim data from the tuplestore, which is bad. (If we could know in
2252 : * advance whether the window functions will use frame boundary info,
2253 : * we could skip creating these pointers in the first place ... but
2254 : * unfortunately the window function API doesn't require that.)
2255 : */
2256 802122 : if (winstate->framehead_ptr >= 0)
2257 6236 : update_frameheadpos(winstate);
2258 802122 : if (winstate->frametail_ptr >= 0)
2259 130280 : update_frametailpos(winstate);
2260 802122 : if (winstate->grouptail_ptr >= 0)
2261 1500 : update_grouptailpos(winstate);
2262 :
2263 : /*
2264 : * Truncate any no-longer-needed rows from the tuplestore.
2265 : */
2266 802122 : tuplestore_trim(winstate->buffer);
2267 :
2268 : /*
2269 : * Form and return a projection tuple using the windowfunc results and
2270 : * the current row. Setting ecxt_outertuple arranges that any Vars
2271 : * will be evaluated with respect to that row.
2272 : */
2273 802122 : econtext->ecxt_outertuple = winstate->ss.ss_ScanTupleSlot;
2274 :
2275 802122 : slot = ExecProject(winstate->ss.ps.ps_ProjInfo);
2276 :
2277 802122 : if (winstate->status == WINDOWAGG_RUN)
2278 : {
2279 802068 : econtext->ecxt_scantuple = slot;
2280 :
2281 : /*
2282 : * Now evaluate the run condition to see if we need to go into
2283 : * pass-through mode, or maybe stop completely.
2284 : */
2285 802068 : if (!ExecQual(winstate->runcondition, econtext))
2286 : {
2287 : /*
2288 : * Determine which mode to move into. If there is no
2289 : * PARTITION BY clause and we're the top-level WindowAgg then
2290 : * we're done. This tuple and any future tuples cannot
2291 : * possibly match the runcondition. However, when there is a
2292 : * PARTITION BY clause or we're not the top-level window we
2293 : * can't just stop as we need to either process other
2294 : * partitions or ensure WindowAgg nodes above us receive all
2295 : * of the tuples they need to process their WindowFuncs.
2296 : */
2297 84 : if (winstate->use_pass_through)
2298 : {
2299 : /*
2300 : * STRICT pass-through mode is required for the top window
2301 : * when there is a PARTITION BY clause. Otherwise we must
2302 : * ensure we store tuples that don't match the
2303 : * runcondition so they're available to WindowAggs above.
2304 : */
2305 42 : if (winstate->top_window)
2306 : {
2307 24 : winstate->status = WINDOWAGG_PASSTHROUGH_STRICT;
2308 24 : continue;
2309 : }
2310 : else
2311 : {
2312 18 : winstate->status = WINDOWAGG_PASSTHROUGH;
2313 :
2314 : /*
2315 : * If we're not the top-window, we'd better NULLify
2316 : * the aggregate results. In pass-through mode we no
2317 : * longer update these and this avoids the old stale
2318 : * results lingering. Some of these might be byref
2319 : * types so we can't have them pointing to free'd
2320 : * memory. The planner insisted that quals used in
2321 : * the runcondition are strict, so the top-level
2322 : * WindowAgg will filter these NULLs out in the filter
2323 : * clause.
2324 : */
2325 18 : numfuncs = winstate->numfuncs;
2326 72 : for (i = 0; i < numfuncs; i++)
2327 : {
2328 54 : econtext->ecxt_aggvalues[i] = (Datum) 0;
2329 54 : econtext->ecxt_aggnulls[i] = true;
2330 : }
2331 : }
2332 : }
2333 : else
2334 : {
2335 : /*
2336 : * Pass-through not required. We can just return NULL.
2337 : * Nothing else will match the runcondition.
2338 : */
2339 42 : winstate->status = WINDOWAGG_DONE;
2340 42 : return NULL;
2341 : }
2342 : }
2343 :
2344 : /*
2345 : * Filter out any tuples we don't need in the top-level WindowAgg.
2346 : */
2347 802002 : if (!ExecQual(winstate->ss.ps.qual, econtext))
2348 : {
2349 18 : InstrCountFiltered1(winstate, 1);
2350 18 : continue;
2351 : }
2352 :
2353 801984 : break;
2354 : }
2355 :
2356 : /*
2357 : * When not in WINDOWAGG_RUN mode, we must still return this tuple if
2358 : * we're anything apart from the top window.
2359 : */
2360 54 : else if (!winstate->top_window)
2361 24 : break;
2362 : }
2363 :
2364 802008 : return slot;
2365 : }
2366 :
2367 : /* -----------------
2368 : * ExecInitWindowAgg
2369 : *
2370 : * Creates the run-time information for the WindowAgg node produced by the
2371 : * planner and initializes its outer subtree
2372 : * -----------------
2373 : */
2374 : WindowAggState *
2375 2406 : ExecInitWindowAgg(WindowAgg *node, EState *estate, int eflags)
2376 : {
2377 : WindowAggState *winstate;
2378 : Plan *outerPlan;
2379 : ExprContext *econtext;
2380 : ExprContext *tmpcontext;
2381 : WindowStatePerFunc perfunc;
2382 : WindowStatePerAgg peragg;
2383 2406 : int frameOptions = node->frameOptions;
2384 : int numfuncs,
2385 : wfuncno,
2386 : numaggs,
2387 : aggno;
2388 : TupleDesc scanDesc;
2389 : ListCell *l;
2390 :
2391 : /* check for unsupported flags */
2392 : Assert(!(eflags & (EXEC_FLAG_BACKWARD | EXEC_FLAG_MARK)));
2393 :
2394 : /*
2395 : * create state structure
2396 : */
2397 2406 : winstate = makeNode(WindowAggState);
2398 2406 : winstate->ss.ps.plan = (Plan *) node;
2399 2406 : winstate->ss.ps.state = estate;
2400 2406 : winstate->ss.ps.ExecProcNode = ExecWindowAgg;
2401 :
2402 : /*
2403 : * Create expression contexts. We need two, one for per-input-tuple
2404 : * processing and one for per-output-tuple processing. We cheat a little
2405 : * by using ExecAssignExprContext() to build both.
2406 : */
2407 2406 : ExecAssignExprContext(estate, &winstate->ss.ps);
2408 2406 : tmpcontext = winstate->ss.ps.ps_ExprContext;
2409 2406 : winstate->tmpcontext = tmpcontext;
2410 2406 : ExecAssignExprContext(estate, &winstate->ss.ps);
2411 :
2412 : /* Create long-lived context for storage of partition-local memory etc */
2413 2406 : winstate->partcontext =
2414 2406 : AllocSetContextCreate(CurrentMemoryContext,
2415 : "WindowAgg Partition",
2416 : ALLOCSET_DEFAULT_SIZES);
2417 :
2418 : /*
2419 : * Create mid-lived context for aggregate trans values etc.
2420 : *
2421 : * Note that moving aggregates each use their own private context, not
2422 : * this one.
2423 : */
2424 2406 : winstate->aggcontext =
2425 2406 : AllocSetContextCreate(CurrentMemoryContext,
2426 : "WindowAgg Aggregates",
2427 : ALLOCSET_DEFAULT_SIZES);
2428 :
2429 : /* Only the top-level WindowAgg may have a qual */
2430 : Assert(node->plan.qual == NIL || node->topWindow);
2431 :
2432 : /* Initialize the qual */
2433 2406 : winstate->ss.ps.qual = ExecInitQual(node->plan.qual,
2434 : (PlanState *) winstate);
2435 :
2436 : /*
2437 : * Setup the run condition, if we received one from the query planner.
2438 : * When set, this may allow us to move into pass-through mode so that we
2439 : * don't have to perform any further evaluation of WindowFuncs in the
2440 : * current partition or possibly stop returning tuples altogether when all
2441 : * tuples are in the same partition.
2442 : */
2443 2406 : winstate->runcondition = ExecInitQual(node->runCondition,
2444 : (PlanState *) winstate);
2445 :
2446 : /*
2447 : * When we're not the top-level WindowAgg node or we are but have a
2448 : * PARTITION BY clause we must move into one of the WINDOWAGG_PASSTHROUGH*
2449 : * modes when the runCondition becomes false.
2450 : */
2451 2406 : winstate->use_pass_through = !node->topWindow || node->partNumCols > 0;
2452 :
2453 : /* remember if we're the top-window or we are below the top-window */
2454 2406 : winstate->top_window = node->topWindow;
2455 :
2456 : /*
2457 : * initialize child nodes
2458 : */
2459 2406 : outerPlan = outerPlan(node);
2460 2406 : outerPlanState(winstate) = ExecInitNode(outerPlan, estate, eflags);
2461 :
2462 : /*
2463 : * initialize source tuple type (which is also the tuple type that we'll
2464 : * store in the tuplestore and use in all our working slots).
2465 : */
2466 2406 : ExecCreateScanSlotFromOuterPlan(estate, &winstate->ss, &TTSOpsMinimalTuple);
2467 2406 : scanDesc = winstate->ss.ss_ScanTupleSlot->tts_tupleDescriptor;
2468 :
2469 : /* the outer tuple isn't the child's tuple, but always a minimal tuple */
2470 2406 : winstate->ss.ps.outeropsset = true;
2471 2406 : winstate->ss.ps.outerops = &TTSOpsMinimalTuple;
2472 2406 : winstate->ss.ps.outeropsfixed = true;
2473 :
2474 : /*
2475 : * tuple table initialization
2476 : */
2477 2406 : winstate->first_part_slot = ExecInitExtraTupleSlot(estate, scanDesc,
2478 : &TTSOpsMinimalTuple);
2479 2406 : winstate->agg_row_slot = ExecInitExtraTupleSlot(estate, scanDesc,
2480 : &TTSOpsMinimalTuple);
2481 2406 : winstate->temp_slot_1 = ExecInitExtraTupleSlot(estate, scanDesc,
2482 : &TTSOpsMinimalTuple);
2483 2406 : winstate->temp_slot_2 = ExecInitExtraTupleSlot(estate, scanDesc,
2484 : &TTSOpsMinimalTuple);
2485 :
2486 : /*
2487 : * create frame head and tail slots only if needed (must create slots in
2488 : * exactly the same cases that update_frameheadpos and update_frametailpos
2489 : * need them)
2490 : */
2491 2406 : winstate->framehead_slot = winstate->frametail_slot = NULL;
2492 :
2493 2406 : if (frameOptions & (FRAMEOPTION_RANGE | FRAMEOPTION_GROUPS))
2494 : {
2495 1446 : if (((frameOptions & FRAMEOPTION_START_CURRENT_ROW) &&
2496 76 : node->ordNumCols != 0) ||
2497 1370 : (frameOptions & FRAMEOPTION_START_OFFSET))
2498 742 : winstate->framehead_slot = ExecInitExtraTupleSlot(estate, scanDesc,
2499 : &TTSOpsMinimalTuple);
2500 1446 : if (((frameOptions & FRAMEOPTION_END_CURRENT_ROW) &&
2501 662 : node->ordNumCols != 0) ||
2502 1004 : (frameOptions & FRAMEOPTION_END_OFFSET))
2503 1168 : winstate->frametail_slot = ExecInitExtraTupleSlot(estate, scanDesc,
2504 : &TTSOpsMinimalTuple);
2505 : }
2506 :
2507 : /*
2508 : * Initialize result slot, type and projection.
2509 : */
2510 2406 : ExecInitResultTupleSlotTL(&winstate->ss.ps, &TTSOpsVirtual);
2511 2406 : ExecAssignProjectionInfo(&winstate->ss.ps, NULL);
2512 :
2513 : /* Set up data for comparing tuples */
2514 2406 : if (node->partNumCols > 0)
2515 600 : winstate->partEqfunction =
2516 600 : execTuplesMatchPrepare(scanDesc,
2517 : node->partNumCols,
2518 600 : node->partColIdx,
2519 600 : node->partOperators,
2520 600 : node->partCollations,
2521 : &winstate->ss.ps);
2522 :
2523 2406 : if (node->ordNumCols > 0)
2524 2054 : winstate->ordEqfunction =
2525 2054 : execTuplesMatchPrepare(scanDesc,
2526 : node->ordNumCols,
2527 2054 : node->ordColIdx,
2528 2054 : node->ordOperators,
2529 2054 : node->ordCollations,
2530 : &winstate->ss.ps);
2531 :
2532 : /*
2533 : * WindowAgg nodes use aggvalues and aggnulls as well as Agg nodes.
2534 : */
2535 2406 : numfuncs = winstate->numfuncs;
2536 2406 : numaggs = winstate->numaggs;
2537 2406 : econtext = winstate->ss.ps.ps_ExprContext;
2538 2406 : econtext->ecxt_aggvalues = (Datum *) palloc0(sizeof(Datum) * numfuncs);
2539 2406 : econtext->ecxt_aggnulls = (bool *) palloc0(sizeof(bool) * numfuncs);
2540 :
2541 : /*
2542 : * allocate per-wfunc/per-agg state information.
2543 : */
2544 2406 : perfunc = (WindowStatePerFunc) palloc0(sizeof(WindowStatePerFuncData) * numfuncs);
2545 2406 : peragg = (WindowStatePerAgg) palloc0(sizeof(WindowStatePerAggData) * numaggs);
2546 2406 : winstate->perfunc = perfunc;
2547 2406 : winstate->peragg = peragg;
2548 :
2549 2406 : wfuncno = -1;
2550 2406 : aggno = -1;
2551 5472 : foreach(l, winstate->funcs)
2552 : {
2553 3066 : WindowFuncExprState *wfuncstate = (WindowFuncExprState *) lfirst(l);
2554 3066 : WindowFunc *wfunc = wfuncstate->wfunc;
2555 : WindowStatePerFunc perfuncstate;
2556 : AclResult aclresult;
2557 : int i;
2558 :
2559 3066 : if (wfunc->winref != node->winref) /* planner screwed up? */
2560 0 : elog(ERROR, "WindowFunc with winref %u assigned to WindowAgg with winref %u",
2561 : wfunc->winref, node->winref);
2562 :
2563 : /* Look for a previous duplicate window function */
2564 3858 : for (i = 0; i <= wfuncno; i++)
2565 : {
2566 798 : if (equal(wfunc, perfunc[i].wfunc) &&
2567 6 : !contain_volatile_functions((Node *) wfunc))
2568 6 : break;
2569 : }
2570 3066 : if (i <= wfuncno)
2571 : {
2572 : /* Found a match to an existing entry, so just mark it */
2573 6 : wfuncstate->wfuncno = i;
2574 6 : continue;
2575 : }
2576 :
2577 : /* Nope, so assign a new PerAgg record */
2578 3060 : perfuncstate = &perfunc[++wfuncno];
2579 :
2580 : /* Mark WindowFunc state node with assigned index in the result array */
2581 3060 : wfuncstate->wfuncno = wfuncno;
2582 :
2583 : /* Check permission to call window function */
2584 3060 : aclresult = object_aclcheck(ProcedureRelationId, wfunc->winfnoid, GetUserId(),
2585 : ACL_EXECUTE);
2586 3060 : if (aclresult != ACLCHECK_OK)
2587 0 : aclcheck_error(aclresult, OBJECT_FUNCTION,
2588 0 : get_func_name(wfunc->winfnoid));
2589 3060 : InvokeFunctionExecuteHook(wfunc->winfnoid);
2590 :
2591 : /* Fill in the perfuncstate data */
2592 3060 : perfuncstate->wfuncstate = wfuncstate;
2593 3060 : perfuncstate->wfunc = wfunc;
2594 3060 : perfuncstate->numArguments = list_length(wfuncstate->args);
2595 3060 : perfuncstate->winCollation = wfunc->inputcollid;
2596 :
2597 3060 : get_typlenbyval(wfunc->wintype,
2598 : &perfuncstate->resulttypeLen,
2599 : &perfuncstate->resulttypeByVal);
2600 :
2601 : /*
2602 : * If it's really just a plain aggregate function, we'll emulate the
2603 : * Agg environment for it.
2604 : */
2605 3060 : perfuncstate->plain_agg = wfunc->winagg;
2606 3060 : if (wfunc->winagg)
2607 : {
2608 : WindowStatePerAgg peraggstate;
2609 :
2610 1428 : perfuncstate->aggno = ++aggno;
2611 1428 : peraggstate = &winstate->peragg[aggno];
2612 1428 : initialize_peragg(winstate, wfunc, peraggstate);
2613 1428 : peraggstate->wfuncno = wfuncno;
2614 : }
2615 : else
2616 : {
2617 1632 : WindowObject winobj = makeNode(WindowObjectData);
2618 :
2619 1632 : winobj->winstate = winstate;
2620 1632 : winobj->argstates = wfuncstate->args;
2621 1632 : winobj->localmem = NULL;
2622 1632 : perfuncstate->winobj = winobj;
2623 :
2624 : /* It's a real window function, so set up to call it. */
2625 1632 : fmgr_info_cxt(wfunc->winfnoid, &perfuncstate->flinfo,
2626 : econtext->ecxt_per_query_memory);
2627 1632 : fmgr_info_set_expr((Node *) wfunc, &perfuncstate->flinfo);
2628 : }
2629 : }
2630 :
2631 : /* Update numfuncs, numaggs to match number of unique functions found */
2632 2406 : winstate->numfuncs = wfuncno + 1;
2633 2406 : winstate->numaggs = aggno + 1;
2634 :
2635 : /* Set up WindowObject for aggregates, if needed */
2636 2406 : if (winstate->numaggs > 0)
2637 : {
2638 1344 : WindowObject agg_winobj = makeNode(WindowObjectData);
2639 :
2640 1344 : agg_winobj->winstate = winstate;
2641 1344 : agg_winobj->argstates = NIL;
2642 1344 : agg_winobj->localmem = NULL;
2643 : /* make sure markptr = -1 to invalidate. It may not get used */
2644 1344 : agg_winobj->markptr = -1;
2645 1344 : agg_winobj->readptr = -1;
2646 1344 : winstate->agg_winobj = agg_winobj;
2647 : }
2648 :
2649 : /* Set the status to running */
2650 2406 : winstate->status = WINDOWAGG_RUN;
2651 :
2652 : /* copy frame options to state node for easy access */
2653 2406 : winstate->frameOptions = frameOptions;
2654 :
2655 : /* initialize frame bound offset expressions */
2656 2406 : winstate->startOffset = ExecInitExpr((Expr *) node->startOffset,
2657 : (PlanState *) winstate);
2658 2406 : winstate->endOffset = ExecInitExpr((Expr *) node->endOffset,
2659 : (PlanState *) winstate);
2660 :
2661 : /* Lookup in_range support functions if needed */
2662 2406 : if (OidIsValid(node->startInRangeFunc))
2663 522 : fmgr_info(node->startInRangeFunc, &winstate->startInRangeFunc);
2664 2406 : if (OidIsValid(node->endInRangeFunc))
2665 588 : fmgr_info(node->endInRangeFunc, &winstate->endInRangeFunc);
2666 2406 : winstate->inRangeColl = node->inRangeColl;
2667 2406 : winstate->inRangeAsc = node->inRangeAsc;
2668 2406 : winstate->inRangeNullsFirst = node->inRangeNullsFirst;
2669 :
2670 2406 : winstate->all_first = true;
2671 2406 : winstate->partition_spooled = false;
2672 2406 : winstate->more_partitions = false;
2673 :
2674 2406 : return winstate;
2675 : }
2676 :
2677 : /* -----------------
2678 : * ExecEndWindowAgg
2679 : * -----------------
2680 : */
2681 : void
2682 2274 : ExecEndWindowAgg(WindowAggState *node)
2683 : {
2684 : PlanState *outerPlan;
2685 : int i;
2686 :
2687 2274 : release_partition(node);
2688 :
2689 3660 : for (i = 0; i < node->numaggs; i++)
2690 : {
2691 1386 : if (node->peragg[i].aggcontext != node->aggcontext)
2692 1216 : MemoryContextDelete(node->peragg[i].aggcontext);
2693 : }
2694 2274 : MemoryContextDelete(node->partcontext);
2695 2274 : MemoryContextDelete(node->aggcontext);
2696 :
2697 2274 : pfree(node->perfunc);
2698 2274 : pfree(node->peragg);
2699 :
2700 2274 : outerPlan = outerPlanState(node);
2701 2274 : ExecEndNode(outerPlan);
2702 2274 : }
2703 :
2704 : /* -----------------
2705 : * ExecReScanWindowAgg
2706 : * -----------------
2707 : */
2708 : void
2709 78 : ExecReScanWindowAgg(WindowAggState *node)
2710 : {
2711 78 : PlanState *outerPlan = outerPlanState(node);
2712 78 : ExprContext *econtext = node->ss.ps.ps_ExprContext;
2713 :
2714 78 : node->status = WINDOWAGG_RUN;
2715 78 : node->all_first = true;
2716 :
2717 : /* release tuplestore et al */
2718 78 : release_partition(node);
2719 :
2720 : /* release all temp tuples, but especially first_part_slot */
2721 78 : ExecClearTuple(node->ss.ss_ScanTupleSlot);
2722 78 : ExecClearTuple(node->first_part_slot);
2723 78 : ExecClearTuple(node->agg_row_slot);
2724 78 : ExecClearTuple(node->temp_slot_1);
2725 78 : ExecClearTuple(node->temp_slot_2);
2726 78 : if (node->framehead_slot)
2727 0 : ExecClearTuple(node->framehead_slot);
2728 78 : if (node->frametail_slot)
2729 6 : ExecClearTuple(node->frametail_slot);
2730 :
2731 : /* Forget current wfunc values */
2732 156 : MemSet(econtext->ecxt_aggvalues, 0, sizeof(Datum) * node->numfuncs);
2733 78 : MemSet(econtext->ecxt_aggnulls, 0, sizeof(bool) * node->numfuncs);
2734 :
2735 : /*
2736 : * if chgParam of subnode is not null then plan will be re-scanned by
2737 : * first ExecProcNode.
2738 : */
2739 78 : if (outerPlan->chgParam == NULL)
2740 6 : ExecReScan(outerPlan);
2741 78 : }
2742 :
2743 : /*
2744 : * initialize_peragg
2745 : *
2746 : * Almost same as in nodeAgg.c, except we don't support DISTINCT currently.
2747 : */
2748 : static WindowStatePerAggData *
2749 1428 : initialize_peragg(WindowAggState *winstate, WindowFunc *wfunc,
2750 : WindowStatePerAgg peraggstate)
2751 : {
2752 : Oid inputTypes[FUNC_MAX_ARGS];
2753 : int numArguments;
2754 : HeapTuple aggTuple;
2755 : Form_pg_aggregate aggform;
2756 : Oid aggtranstype;
2757 : AttrNumber initvalAttNo;
2758 : AclResult aclresult;
2759 : bool use_ma_code;
2760 : Oid transfn_oid,
2761 : invtransfn_oid,
2762 : finalfn_oid;
2763 : bool finalextra;
2764 : char finalmodify;
2765 : Expr *transfnexpr,
2766 : *invtransfnexpr,
2767 : *finalfnexpr;
2768 : Datum textInitVal;
2769 : int i;
2770 : ListCell *lc;
2771 :
2772 1428 : numArguments = list_length(wfunc->args);
2773 :
2774 1428 : i = 0;
2775 2730 : foreach(lc, wfunc->args)
2776 : {
2777 1302 : inputTypes[i++] = exprType((Node *) lfirst(lc));
2778 : }
2779 :
2780 1428 : aggTuple = SearchSysCache1(AGGFNOID, ObjectIdGetDatum(wfunc->winfnoid));
2781 1428 : if (!HeapTupleIsValid(aggTuple))
2782 0 : elog(ERROR, "cache lookup failed for aggregate %u",
2783 : wfunc->winfnoid);
2784 1428 : aggform = (Form_pg_aggregate) GETSTRUCT(aggTuple);
2785 :
2786 : /*
2787 : * Figure out whether we want to use the moving-aggregate implementation,
2788 : * and collect the right set of fields from the pg_attribute entry.
2789 : *
2790 : * It's possible that an aggregate would supply a safe moving-aggregate
2791 : * implementation and an unsafe normal one, in which case our hand is
2792 : * forced. Otherwise, if the frame head can't move, we don't need
2793 : * moving-aggregate code. Even if we'd like to use it, don't do so if the
2794 : * aggregate's arguments (and FILTER clause if any) contain any calls to
2795 : * volatile functions. Otherwise, the difference between restarting and
2796 : * not restarting the aggregation would be user-visible.
2797 : *
2798 : * We also don't risk using moving aggregates when there are subplans in
2799 : * the arguments or FILTER clause. This is partly because
2800 : * contain_volatile_functions() doesn't look inside subplans; but there
2801 : * are other reasons why a subplan's output might be volatile. For
2802 : * example, syncscan mode can render the results nonrepeatable.
2803 : */
2804 1428 : if (!OidIsValid(aggform->aggminvtransfn))
2805 182 : use_ma_code = false; /* sine qua non */
2806 1246 : else if (aggform->aggmfinalmodify == AGGMODIFY_READ_ONLY &&
2807 1246 : aggform->aggfinalmodify != AGGMODIFY_READ_ONLY)
2808 0 : use_ma_code = true; /* decision forced by safety */
2809 1246 : else if (winstate->frameOptions & FRAMEOPTION_START_UNBOUNDED_PRECEDING)
2810 0 : use_ma_code = false; /* non-moving frame head */
2811 1246 : else if (contain_volatile_functions((Node *) wfunc))
2812 18 : use_ma_code = false; /* avoid possible behavioral change */
2813 1228 : else if (contain_subplans((Node *) wfunc))
2814 0 : use_ma_code = false; /* subplans might contain volatile functions */
2815 : else
2816 1228 : use_ma_code = true; /* yes, let's use it */
2817 1428 : if (use_ma_code)
2818 : {
2819 1228 : peraggstate->transfn_oid = transfn_oid = aggform->aggmtransfn;
2820 1228 : peraggstate->invtransfn_oid = invtransfn_oid = aggform->aggminvtransfn;
2821 1228 : peraggstate->finalfn_oid = finalfn_oid = aggform->aggmfinalfn;
2822 1228 : finalextra = aggform->aggmfinalextra;
2823 1228 : finalmodify = aggform->aggmfinalmodify;
2824 1228 : aggtranstype = aggform->aggmtranstype;
2825 1228 : initvalAttNo = Anum_pg_aggregate_aggminitval;
2826 : }
2827 : else
2828 : {
2829 200 : peraggstate->transfn_oid = transfn_oid = aggform->aggtransfn;
2830 200 : peraggstate->invtransfn_oid = invtransfn_oid = InvalidOid;
2831 200 : peraggstate->finalfn_oid = finalfn_oid = aggform->aggfinalfn;
2832 200 : finalextra = aggform->aggfinalextra;
2833 200 : finalmodify = aggform->aggfinalmodify;
2834 200 : aggtranstype = aggform->aggtranstype;
2835 200 : initvalAttNo = Anum_pg_aggregate_agginitval;
2836 : }
2837 :
2838 : /*
2839 : * ExecInitWindowAgg already checked permission to call aggregate function
2840 : * ... but we still need to check the component functions
2841 : */
2842 :
2843 : /* Check that aggregate owner has permission to call component fns */
2844 : {
2845 : HeapTuple procTuple;
2846 : Oid aggOwner;
2847 :
2848 1428 : procTuple = SearchSysCache1(PROCOID,
2849 : ObjectIdGetDatum(wfunc->winfnoid));
2850 1428 : if (!HeapTupleIsValid(procTuple))
2851 0 : elog(ERROR, "cache lookup failed for function %u",
2852 : wfunc->winfnoid);
2853 1428 : aggOwner = ((Form_pg_proc) GETSTRUCT(procTuple))->proowner;
2854 1428 : ReleaseSysCache(procTuple);
2855 :
2856 1428 : aclresult = object_aclcheck(ProcedureRelationId, transfn_oid, aggOwner,
2857 : ACL_EXECUTE);
2858 1428 : if (aclresult != ACLCHECK_OK)
2859 0 : aclcheck_error(aclresult, OBJECT_FUNCTION,
2860 0 : get_func_name(transfn_oid));
2861 1428 : InvokeFunctionExecuteHook(transfn_oid);
2862 :
2863 1428 : if (OidIsValid(invtransfn_oid))
2864 : {
2865 1228 : aclresult = object_aclcheck(ProcedureRelationId, invtransfn_oid, aggOwner,
2866 : ACL_EXECUTE);
2867 1228 : if (aclresult != ACLCHECK_OK)
2868 0 : aclcheck_error(aclresult, OBJECT_FUNCTION,
2869 0 : get_func_name(invtransfn_oid));
2870 1228 : InvokeFunctionExecuteHook(invtransfn_oid);
2871 : }
2872 :
2873 1428 : if (OidIsValid(finalfn_oid))
2874 : {
2875 1010 : aclresult = object_aclcheck(ProcedureRelationId, finalfn_oid, aggOwner,
2876 : ACL_EXECUTE);
2877 1010 : if (aclresult != ACLCHECK_OK)
2878 0 : aclcheck_error(aclresult, OBJECT_FUNCTION,
2879 0 : get_func_name(finalfn_oid));
2880 1010 : InvokeFunctionExecuteHook(finalfn_oid);
2881 : }
2882 : }
2883 :
2884 : /*
2885 : * If the selected finalfn isn't read-only, we can't run this aggregate as
2886 : * a window function. This is a user-facing error, so we take a bit more
2887 : * care with the error message than elsewhere in this function.
2888 : */
2889 1428 : if (finalmodify != AGGMODIFY_READ_ONLY)
2890 0 : ereport(ERROR,
2891 : (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
2892 : errmsg("aggregate function %s does not support use as a window function",
2893 : format_procedure(wfunc->winfnoid))));
2894 :
2895 : /* Detect how many arguments to pass to the finalfn */
2896 1428 : if (finalextra)
2897 26 : peraggstate->numFinalArgs = numArguments + 1;
2898 : else
2899 1402 : peraggstate->numFinalArgs = 1;
2900 :
2901 : /* resolve actual type of transition state, if polymorphic */
2902 1428 : aggtranstype = resolve_aggregate_transtype(wfunc->winfnoid,
2903 : aggtranstype,
2904 : inputTypes,
2905 : numArguments);
2906 :
2907 : /* build expression trees using actual argument & result types */
2908 1428 : build_aggregate_transfn_expr(inputTypes,
2909 : numArguments,
2910 : 0, /* no ordered-set window functions yet */
2911 : false, /* no variadic window functions yet */
2912 : aggtranstype,
2913 : wfunc->inputcollid,
2914 : transfn_oid,
2915 : invtransfn_oid,
2916 : &transfnexpr,
2917 : &invtransfnexpr);
2918 :
2919 : /* set up infrastructure for calling the transfn(s) and finalfn */
2920 1428 : fmgr_info(transfn_oid, &peraggstate->transfn);
2921 1428 : fmgr_info_set_expr((Node *) transfnexpr, &peraggstate->transfn);
2922 :
2923 1428 : if (OidIsValid(invtransfn_oid))
2924 : {
2925 1228 : fmgr_info(invtransfn_oid, &peraggstate->invtransfn);
2926 1228 : fmgr_info_set_expr((Node *) invtransfnexpr, &peraggstate->invtransfn);
2927 : }
2928 :
2929 1428 : if (OidIsValid(finalfn_oid))
2930 : {
2931 1010 : build_aggregate_finalfn_expr(inputTypes,
2932 : peraggstate->numFinalArgs,
2933 : aggtranstype,
2934 : wfunc->wintype,
2935 : wfunc->inputcollid,
2936 : finalfn_oid,
2937 : &finalfnexpr);
2938 1010 : fmgr_info(finalfn_oid, &peraggstate->finalfn);
2939 1010 : fmgr_info_set_expr((Node *) finalfnexpr, &peraggstate->finalfn);
2940 : }
2941 :
2942 : /* get info about relevant datatypes */
2943 1428 : get_typlenbyval(wfunc->wintype,
2944 : &peraggstate->resulttypeLen,
2945 : &peraggstate->resulttypeByVal);
2946 1428 : get_typlenbyval(aggtranstype,
2947 : &peraggstate->transtypeLen,
2948 : &peraggstate->transtypeByVal);
2949 :
2950 : /*
2951 : * initval is potentially null, so don't try to access it as a struct
2952 : * field. Must do it the hard way with SysCacheGetAttr.
2953 : */
2954 1428 : textInitVal = SysCacheGetAttr(AGGFNOID, aggTuple, initvalAttNo,
2955 : &peraggstate->initValueIsNull);
2956 :
2957 1428 : if (peraggstate->initValueIsNull)
2958 560 : peraggstate->initValue = (Datum) 0;
2959 : else
2960 868 : peraggstate->initValue = GetAggInitVal(textInitVal,
2961 : aggtranstype);
2962 :
2963 : /*
2964 : * If the transfn is strict and the initval is NULL, make sure input type
2965 : * and transtype are the same (or at least binary-compatible), so that
2966 : * it's OK to use the first input value as the initial transValue. This
2967 : * should have been checked at agg definition time, but we must check
2968 : * again in case the transfn's strictness property has been changed.
2969 : */
2970 1428 : if (peraggstate->transfn.fn_strict && peraggstate->initValueIsNull)
2971 : {
2972 150 : if (numArguments < 1 ||
2973 150 : !IsBinaryCoercible(inputTypes[0], aggtranstype))
2974 0 : ereport(ERROR,
2975 : (errcode(ERRCODE_INVALID_FUNCTION_DEFINITION),
2976 : errmsg("aggregate %u needs to have compatible input type and transition type",
2977 : wfunc->winfnoid)));
2978 : }
2979 :
2980 : /*
2981 : * Insist that forward and inverse transition functions have the same
2982 : * strictness setting. Allowing them to differ would require handling
2983 : * more special cases in advance_windowaggregate and
2984 : * advance_windowaggregate_base, for no discernible benefit. This should
2985 : * have been checked at agg definition time, but we must check again in
2986 : * case either function's strictness property has been changed.
2987 : */
2988 1428 : if (OidIsValid(invtransfn_oid) &&
2989 1228 : peraggstate->transfn.fn_strict != peraggstate->invtransfn.fn_strict)
2990 0 : ereport(ERROR,
2991 : (errcode(ERRCODE_INVALID_FUNCTION_DEFINITION),
2992 : errmsg("strictness of aggregate's forward and inverse transition functions must match")));
2993 :
2994 : /*
2995 : * Moving aggregates use their own aggcontext.
2996 : *
2997 : * This is necessary because they might restart at different times, so we
2998 : * might never be able to reset the shared context otherwise. We can't
2999 : * make it the aggregates' responsibility to clean up after themselves,
3000 : * because strict aggregates must be restarted whenever we remove their
3001 : * last non-NULL input, which the aggregate won't be aware is happening.
3002 : * Also, just pfree()ing the transValue upon restarting wouldn't help,
3003 : * since we'd miss any indirectly referenced data. We could, in theory,
3004 : * make the memory allocation rules for moving aggregates different than
3005 : * they have historically been for plain aggregates, but that seems grotty
3006 : * and likely to lead to memory leaks.
3007 : */
3008 1428 : if (OidIsValid(invtransfn_oid))
3009 1228 : peraggstate->aggcontext =
3010 1228 : AllocSetContextCreate(CurrentMemoryContext,
3011 : "WindowAgg Per Aggregate",
3012 : ALLOCSET_DEFAULT_SIZES);
3013 : else
3014 200 : peraggstate->aggcontext = winstate->aggcontext;
3015 :
3016 1428 : ReleaseSysCache(aggTuple);
3017 :
3018 1428 : return peraggstate;
3019 : }
3020 :
3021 : static Datum
3022 868 : GetAggInitVal(Datum textInitVal, Oid transtype)
3023 : {
3024 : Oid typinput,
3025 : typioparam;
3026 : char *strInitVal;
3027 : Datum initVal;
3028 :
3029 868 : getTypeInputInfo(transtype, &typinput, &typioparam);
3030 868 : strInitVal = TextDatumGetCString(textInitVal);
3031 868 : initVal = OidInputFunctionCall(typinput, strInitVal,
3032 : typioparam, -1);
3033 868 : pfree(strInitVal);
3034 868 : return initVal;
3035 : }
3036 :
3037 : /*
3038 : * are_peers
3039 : * compare two rows to see if they are equal according to the ORDER BY clause
3040 : *
3041 : * NB: this does not consider the window frame mode.
3042 : */
3043 : static bool
3044 513474 : are_peers(WindowAggState *winstate, TupleTableSlot *slot1,
3045 : TupleTableSlot *slot2)
3046 : {
3047 513474 : WindowAgg *node = (WindowAgg *) winstate->ss.ps.plan;
3048 513474 : ExprContext *econtext = winstate->tmpcontext;
3049 :
3050 : /* If no ORDER BY, all rows are peers with each other */
3051 513474 : if (node->ordNumCols == 0)
3052 958 : return true;
3053 :
3054 512516 : econtext->ecxt_outertuple = slot1;
3055 512516 : econtext->ecxt_innertuple = slot2;
3056 512516 : return ExecQualAndReset(winstate->ordEqfunction, econtext);
3057 : }
3058 :
3059 : /*
3060 : * window_gettupleslot
3061 : * Fetch the pos'th tuple of the current partition into the slot,
3062 : * using the winobj's read pointer
3063 : *
3064 : * Returns true if successful, false if no such row
3065 : */
3066 : static bool
3067 677640 : window_gettupleslot(WindowObject winobj, int64 pos, TupleTableSlot *slot)
3068 : {
3069 677640 : WindowAggState *winstate = winobj->winstate;
3070 : MemoryContext oldcontext;
3071 :
3072 : /* often called repeatedly in a row */
3073 677640 : CHECK_FOR_INTERRUPTS();
3074 :
3075 : /* Don't allow passing -1 to spool_tuples here */
3076 677640 : if (pos < 0)
3077 312 : return false;
3078 :
3079 : /* If necessary, fetch the tuple into the spool */
3080 677328 : spool_tuples(winstate, pos);
3081 :
3082 677328 : if (pos >= winstate->spooled_rows)
3083 4504 : return false;
3084 :
3085 672824 : if (pos < winobj->markpos)
3086 0 : elog(ERROR, "cannot fetch row before WindowObject's mark position");
3087 :
3088 672824 : oldcontext = MemoryContextSwitchTo(winstate->ss.ps.ps_ExprContext->ecxt_per_query_memory);
3089 :
3090 672824 : tuplestore_select_read_pointer(winstate->buffer, winobj->readptr);
3091 :
3092 : /*
3093 : * Advance or rewind until we are within one tuple of the one we want.
3094 : */
3095 672824 : if (winobj->seekpos < pos - 1)
3096 : {
3097 2244 : if (!tuplestore_skiptuples(winstate->buffer,
3098 2244 : pos - 1 - winobj->seekpos,
3099 : true))
3100 0 : elog(ERROR, "unexpected end of tuplestore");
3101 2244 : winobj->seekpos = pos - 1;
3102 : }
3103 670580 : else if (winobj->seekpos > pos + 1)
3104 : {
3105 2752 : if (!tuplestore_skiptuples(winstate->buffer,
3106 2752 : winobj->seekpos - (pos + 1),
3107 : false))
3108 0 : elog(ERROR, "unexpected end of tuplestore");
3109 2752 : winobj->seekpos = pos + 1;
3110 : }
3111 667828 : else if (winobj->seekpos == pos)
3112 : {
3113 : /*
3114 : * There's no API to refetch the tuple at the current position. We
3115 : * have to move one tuple forward, and then one backward. (We don't
3116 : * do it the other way because we might try to fetch the row before
3117 : * our mark, which isn't allowed.) XXX this case could stand to be
3118 : * optimized.
3119 : */
3120 172218 : tuplestore_advance(winstate->buffer, true);
3121 172218 : winobj->seekpos++;
3122 : }
3123 :
3124 : /*
3125 : * Now we should be on the tuple immediately before or after the one we
3126 : * want, so just fetch forwards or backwards as appropriate.
3127 : *
3128 : * Notice that we tell tuplestore_gettupleslot to make a physical copy of
3129 : * the fetched tuple. This ensures that the slot's contents remain valid
3130 : * through manipulations of the tuplestore, which some callers depend on.
3131 : */
3132 672824 : if (winobj->seekpos > pos)
3133 : {
3134 175100 : if (!tuplestore_gettupleslot(winstate->buffer, false, true, slot))
3135 0 : elog(ERROR, "unexpected end of tuplestore");
3136 175100 : winobj->seekpos--;
3137 : }
3138 : else
3139 : {
3140 497724 : if (!tuplestore_gettupleslot(winstate->buffer, true, true, slot))
3141 0 : elog(ERROR, "unexpected end of tuplestore");
3142 497724 : winobj->seekpos++;
3143 : }
3144 :
3145 : Assert(winobj->seekpos == pos);
3146 :
3147 672824 : MemoryContextSwitchTo(oldcontext);
3148 :
3149 672824 : return true;
3150 : }
3151 :
3152 :
3153 : /***********************************************************************
3154 : * API exposed to window functions
3155 : ***********************************************************************/
3156 :
3157 :
3158 : /*
3159 : * WinGetPartitionLocalMemory
3160 : * Get working memory that lives till end of partition processing
3161 : *
3162 : * On first call within a given partition, this allocates and zeroes the
3163 : * requested amount of space. Subsequent calls just return the same chunk.
3164 : *
3165 : * Memory obtained this way is normally used to hold state that should be
3166 : * automatically reset for each new partition. If a window function wants
3167 : * to hold state across the whole query, fcinfo->fn_extra can be used in the
3168 : * usual way for that.
3169 : */
3170 : void *
3171 331374 : WinGetPartitionLocalMemory(WindowObject winobj, Size sz)
3172 : {
3173 : Assert(WindowObjectIsValid(winobj));
3174 331374 : if (winobj->localmem == NULL)
3175 432 : winobj->localmem =
3176 432 : MemoryContextAllocZero(winobj->winstate->partcontext, sz);
3177 331374 : return winobj->localmem;
3178 : }
3179 :
3180 : /*
3181 : * WinGetCurrentPosition
3182 : * Return the current row's position (counting from 0) within the current
3183 : * partition.
3184 : */
3185 : int64
3186 739454 : WinGetCurrentPosition(WindowObject winobj)
3187 : {
3188 : Assert(WindowObjectIsValid(winobj));
3189 739454 : return winobj->winstate->currentpos;
3190 : }
3191 :
3192 : /*
3193 : * WinGetPartitionRowCount
3194 : * Return total number of rows contained in the current partition.
3195 : *
3196 : * Note: this is a relatively expensive operation because it forces the
3197 : * whole partition to be "spooled" into the tuplestore at once. Once
3198 : * executed, however, additional calls within the same partition are cheap.
3199 : */
3200 : int64
3201 162 : WinGetPartitionRowCount(WindowObject winobj)
3202 : {
3203 : Assert(WindowObjectIsValid(winobj));
3204 162 : spool_tuples(winobj->winstate, -1);
3205 162 : return winobj->winstate->spooled_rows;
3206 : }
3207 :
3208 : /*
3209 : * WinSetMarkPosition
3210 : * Set the "mark" position for the window object, which is the oldest row
3211 : * number (counting from 0) it is allowed to fetch during all subsequent
3212 : * operations within the current partition.
3213 : *
3214 : * Window functions do not have to call this, but are encouraged to move the
3215 : * mark forward when possible to keep the tuplestore size down and prevent
3216 : * having to spill rows to disk.
3217 : */
3218 : void
3219 801518 : WinSetMarkPosition(WindowObject winobj, int64 markpos)
3220 : {
3221 : WindowAggState *winstate;
3222 :
3223 : Assert(WindowObjectIsValid(winobj));
3224 801518 : winstate = winobj->winstate;
3225 :
3226 801518 : if (markpos < winobj->markpos)
3227 0 : elog(ERROR, "cannot move WindowObject's mark position backward");
3228 801518 : tuplestore_select_read_pointer(winstate->buffer, winobj->markptr);
3229 801518 : if (markpos > winobj->markpos)
3230 : {
3231 795980 : tuplestore_skiptuples(winstate->buffer,
3232 795980 : markpos - winobj->markpos,
3233 : true);
3234 795980 : winobj->markpos = markpos;
3235 : }
3236 801518 : tuplestore_select_read_pointer(winstate->buffer, winobj->readptr);
3237 801518 : if (markpos > winobj->seekpos)
3238 : {
3239 444058 : tuplestore_skiptuples(winstate->buffer,
3240 444058 : markpos - winobj->seekpos,
3241 : true);
3242 444058 : winobj->seekpos = markpos;
3243 : }
3244 801518 : }
3245 :
3246 : /*
3247 : * WinRowsArePeers
3248 : * Compare two rows (specified by absolute position in partition) to see
3249 : * if they are equal according to the ORDER BY clause.
3250 : *
3251 : * NB: this does not consider the window frame mode.
3252 : */
3253 : bool
3254 165270 : WinRowsArePeers(WindowObject winobj, int64 pos1, int64 pos2)
3255 : {
3256 : WindowAggState *winstate;
3257 : WindowAgg *node;
3258 : TupleTableSlot *slot1;
3259 : TupleTableSlot *slot2;
3260 : bool res;
3261 :
3262 : Assert(WindowObjectIsValid(winobj));
3263 165270 : winstate = winobj->winstate;
3264 165270 : node = (WindowAgg *) winstate->ss.ps.plan;
3265 :
3266 : /* If no ORDER BY, all rows are peers; don't bother to fetch them */
3267 165270 : if (node->ordNumCols == 0)
3268 0 : return true;
3269 :
3270 : /*
3271 : * Note: OK to use temp_slot_2 here because we aren't calling any
3272 : * frame-related functions (those tend to clobber temp_slot_2).
3273 : */
3274 165270 : slot1 = winstate->temp_slot_1;
3275 165270 : slot2 = winstate->temp_slot_2;
3276 :
3277 165270 : if (!window_gettupleslot(winobj, pos1, slot1))
3278 0 : elog(ERROR, "specified position is out of window: " INT64_FORMAT,
3279 : pos1);
3280 165270 : if (!window_gettupleslot(winobj, pos2, slot2))
3281 0 : elog(ERROR, "specified position is out of window: " INT64_FORMAT,
3282 : pos2);
3283 :
3284 165270 : res = are_peers(winstate, slot1, slot2);
3285 :
3286 165270 : ExecClearTuple(slot1);
3287 165270 : ExecClearTuple(slot2);
3288 :
3289 165270 : return res;
3290 : }
3291 :
3292 : /*
3293 : * WinGetFuncArgInPartition
3294 : * Evaluate a window function's argument expression on a specified
3295 : * row of the partition. The row is identified in lseek(2) style,
3296 : * i.e. relative to the current, first, or last row.
3297 : *
3298 : * argno: argument number to evaluate (counted from 0)
3299 : * relpos: signed rowcount offset from the seek position
3300 : * seektype: WINDOW_SEEK_CURRENT, WINDOW_SEEK_HEAD, or WINDOW_SEEK_TAIL
3301 : * set_mark: If the row is found and set_mark is true, the mark is moved to
3302 : * the row as a side-effect.
3303 : * isnull: output argument, receives isnull status of result
3304 : * isout: output argument, set to indicate whether target row position
3305 : * is out of partition (can pass NULL if caller doesn't care about this)
3306 : *
3307 : * Specifying a nonexistent row is not an error, it just causes a null result
3308 : * (plus setting *isout true, if isout isn't NULL).
3309 : */
3310 : Datum
3311 182778 : WinGetFuncArgInPartition(WindowObject winobj, int argno,
3312 : int relpos, int seektype, bool set_mark,
3313 : bool *isnull, bool *isout)
3314 : {
3315 : WindowAggState *winstate;
3316 : ExprContext *econtext;
3317 : TupleTableSlot *slot;
3318 : bool gottuple;
3319 : int64 abs_pos;
3320 :
3321 : Assert(WindowObjectIsValid(winobj));
3322 182778 : winstate = winobj->winstate;
3323 182778 : econtext = winstate->ss.ps.ps_ExprContext;
3324 182778 : slot = winstate->temp_slot_1;
3325 :
3326 182778 : switch (seektype)
3327 : {
3328 182778 : case WINDOW_SEEK_CURRENT:
3329 182778 : abs_pos = winstate->currentpos + relpos;
3330 182778 : break;
3331 0 : case WINDOW_SEEK_HEAD:
3332 0 : abs_pos = relpos;
3333 0 : break;
3334 0 : case WINDOW_SEEK_TAIL:
3335 0 : spool_tuples(winstate, -1);
3336 0 : abs_pos = winstate->spooled_rows - 1 + relpos;
3337 0 : break;
3338 0 : default:
3339 0 : elog(ERROR, "unrecognized window seek type: %d", seektype);
3340 : abs_pos = 0; /* keep compiler quiet */
3341 : break;
3342 : }
3343 :
3344 182778 : gottuple = window_gettupleslot(winobj, abs_pos, slot);
3345 :
3346 182778 : if (!gottuple)
3347 : {
3348 318 : if (isout)
3349 318 : *isout = true;
3350 318 : *isnull = true;
3351 318 : return (Datum) 0;
3352 : }
3353 : else
3354 : {
3355 182460 : if (isout)
3356 182460 : *isout = false;
3357 182460 : if (set_mark)
3358 182304 : WinSetMarkPosition(winobj, abs_pos);
3359 182460 : econtext->ecxt_outertuple = slot;
3360 182460 : return ExecEvalExpr((ExprState *) list_nth(winobj->argstates, argno),
3361 : econtext, isnull);
3362 : }
3363 : }
3364 :
3365 : /*
3366 : * WinGetFuncArgInFrame
3367 : * Evaluate a window function's argument expression on a specified
3368 : * row of the window frame. The row is identified in lseek(2) style,
3369 : * i.e. relative to the first or last row of the frame. (We do not
3370 : * support WINDOW_SEEK_CURRENT here, because it's not very clear what
3371 : * that should mean if the current row isn't part of the frame.)
3372 : *
3373 : * argno: argument number to evaluate (counted from 0)
3374 : * relpos: signed rowcount offset from the seek position
3375 : * seektype: WINDOW_SEEK_HEAD or WINDOW_SEEK_TAIL
3376 : * set_mark: If the row is found/in frame and set_mark is true, the mark is
3377 : * moved to the row as a side-effect.
3378 : * isnull: output argument, receives isnull status of result
3379 : * isout: output argument, set to indicate whether target row position
3380 : * is out of frame (can pass NULL if caller doesn't care about this)
3381 : *
3382 : * Specifying a nonexistent or not-in-frame row is not an error, it just
3383 : * causes a null result (plus setting *isout true, if isout isn't NULL).
3384 : *
3385 : * Note that some exclusion-clause options lead to situations where the
3386 : * rows that are in-frame are not consecutive in the partition. But we
3387 : * count only in-frame rows when measuring relpos.
3388 : *
3389 : * The set_mark flag is interpreted as meaning that the caller will specify
3390 : * a constant (or, perhaps, monotonically increasing) relpos in successive
3391 : * calls, so that *if there is no exclusion clause* there will be no need
3392 : * to fetch a row before the previously fetched row. But we do not expect
3393 : * the caller to know how to account for exclusion clauses. Therefore,
3394 : * if there is an exclusion clause we take responsibility for adjusting the
3395 : * mark request to something that will be safe given the above assumption
3396 : * about relpos.
3397 : */
3398 : Datum
3399 8628 : WinGetFuncArgInFrame(WindowObject winobj, int argno,
3400 : int relpos, int seektype, bool set_mark,
3401 : bool *isnull, bool *isout)
3402 : {
3403 : WindowAggState *winstate;
3404 : ExprContext *econtext;
3405 : TupleTableSlot *slot;
3406 : int64 abs_pos;
3407 : int64 mark_pos;
3408 :
3409 : Assert(WindowObjectIsValid(winobj));
3410 8628 : winstate = winobj->winstate;
3411 8628 : econtext = winstate->ss.ps.ps_ExprContext;
3412 8628 : slot = winstate->temp_slot_1;
3413 :
3414 8628 : switch (seektype)
3415 : {
3416 0 : case WINDOW_SEEK_CURRENT:
3417 0 : elog(ERROR, "WINDOW_SEEK_CURRENT is not supported for WinGetFuncArgInFrame");
3418 : abs_pos = mark_pos = 0; /* keep compiler quiet */
3419 : break;
3420 4206 : case WINDOW_SEEK_HEAD:
3421 : /* rejecting relpos < 0 is easy and simplifies code below */
3422 4206 : if (relpos < 0)
3423 0 : goto out_of_frame;
3424 4206 : update_frameheadpos(winstate);
3425 4164 : abs_pos = winstate->frameheadpos + relpos;
3426 4164 : mark_pos = abs_pos;
3427 :
3428 : /*
3429 : * Account for exclusion option if one is active, but advance only
3430 : * abs_pos not mark_pos. This prevents changes of the current
3431 : * row's peer group from resulting in trying to fetch a row before
3432 : * some previous mark position.
3433 : *
3434 : * Note that in some corner cases such as current row being
3435 : * outside frame, these calculations are theoretically too simple,
3436 : * but it doesn't matter because we'll end up deciding the row is
3437 : * out of frame. We do not attempt to avoid fetching rows past
3438 : * end of frame; that would happen in some cases anyway.
3439 : */
3440 4164 : switch (winstate->frameOptions & FRAMEOPTION_EXCLUSION)
3441 : {
3442 3504 : case 0:
3443 : /* no adjustment needed */
3444 3504 : break;
3445 240 : case FRAMEOPTION_EXCLUDE_CURRENT_ROW:
3446 240 : if (abs_pos >= winstate->currentpos &&
3447 186 : winstate->currentpos >= winstate->frameheadpos)
3448 66 : abs_pos++;
3449 240 : break;
3450 120 : case FRAMEOPTION_EXCLUDE_GROUP:
3451 120 : update_grouptailpos(winstate);
3452 120 : if (abs_pos >= winstate->groupheadpos &&
3453 72 : winstate->grouptailpos > winstate->frameheadpos)
3454 : {
3455 72 : int64 overlapstart = Max(winstate->groupheadpos,
3456 : winstate->frameheadpos);
3457 :
3458 72 : abs_pos += winstate->grouptailpos - overlapstart;
3459 : }
3460 120 : break;
3461 300 : case FRAMEOPTION_EXCLUDE_TIES:
3462 300 : update_grouptailpos(winstate);
3463 300 : if (abs_pos >= winstate->groupheadpos &&
3464 204 : winstate->grouptailpos > winstate->frameheadpos)
3465 : {
3466 84 : int64 overlapstart = Max(winstate->groupheadpos,
3467 : winstate->frameheadpos);
3468 :
3469 84 : if (abs_pos == overlapstart)
3470 84 : abs_pos = winstate->currentpos;
3471 : else
3472 0 : abs_pos += winstate->grouptailpos - overlapstart - 1;
3473 : }
3474 300 : break;
3475 0 : default:
3476 0 : elog(ERROR, "unrecognized frame option state: 0x%x",
3477 : winstate->frameOptions);
3478 : break;
3479 : }
3480 4164 : break;
3481 4422 : case WINDOW_SEEK_TAIL:
3482 : /* rejecting relpos > 0 is easy and simplifies code below */
3483 4422 : if (relpos > 0)
3484 0 : goto out_of_frame;
3485 4422 : update_frametailpos(winstate);
3486 4416 : abs_pos = winstate->frametailpos - 1 + relpos;
3487 :
3488 : /*
3489 : * Account for exclusion option if one is active. If there is no
3490 : * exclusion, we can safely set the mark at the accessed row. But
3491 : * if there is, we can only mark the frame start, because we can't
3492 : * be sure how far back in the frame the exclusion might cause us
3493 : * to fetch in future. Furthermore, we have to actually check
3494 : * against frameheadpos here, since it's unsafe to try to fetch a
3495 : * row before frame start if the mark might be there already.
3496 : */
3497 4416 : switch (winstate->frameOptions & FRAMEOPTION_EXCLUSION)
3498 : {
3499 3936 : case 0:
3500 : /* no adjustment needed */
3501 3936 : mark_pos = abs_pos;
3502 3936 : break;
3503 120 : case FRAMEOPTION_EXCLUDE_CURRENT_ROW:
3504 120 : if (abs_pos <= winstate->currentpos &&
3505 12 : winstate->currentpos < winstate->frametailpos)
3506 12 : abs_pos--;
3507 120 : update_frameheadpos(winstate);
3508 120 : if (abs_pos < winstate->frameheadpos)
3509 6 : goto out_of_frame;
3510 114 : mark_pos = winstate->frameheadpos;
3511 114 : break;
3512 240 : case FRAMEOPTION_EXCLUDE_GROUP:
3513 240 : update_grouptailpos(winstate);
3514 240 : if (abs_pos < winstate->grouptailpos &&
3515 54 : winstate->groupheadpos < winstate->frametailpos)
3516 : {
3517 54 : int64 overlapend = Min(winstate->grouptailpos,
3518 : winstate->frametailpos);
3519 :
3520 54 : abs_pos -= overlapend - winstate->groupheadpos;
3521 : }
3522 240 : update_frameheadpos(winstate);
3523 240 : if (abs_pos < winstate->frameheadpos)
3524 54 : goto out_of_frame;
3525 186 : mark_pos = winstate->frameheadpos;
3526 186 : break;
3527 120 : case FRAMEOPTION_EXCLUDE_TIES:
3528 120 : update_grouptailpos(winstate);
3529 120 : if (abs_pos < winstate->grouptailpos &&
3530 36 : winstate->groupheadpos < winstate->frametailpos)
3531 : {
3532 36 : int64 overlapend = Min(winstate->grouptailpos,
3533 : winstate->frametailpos);
3534 :
3535 36 : if (abs_pos == overlapend - 1)
3536 36 : abs_pos = winstate->currentpos;
3537 : else
3538 0 : abs_pos -= overlapend - 1 - winstate->groupheadpos;
3539 : }
3540 120 : update_frameheadpos(winstate);
3541 120 : if (abs_pos < winstate->frameheadpos)
3542 0 : goto out_of_frame;
3543 120 : mark_pos = winstate->frameheadpos;
3544 120 : break;
3545 0 : default:
3546 0 : elog(ERROR, "unrecognized frame option state: 0x%x",
3547 : winstate->frameOptions);
3548 : mark_pos = 0; /* keep compiler quiet */
3549 : break;
3550 : }
3551 4356 : break;
3552 0 : default:
3553 0 : elog(ERROR, "unrecognized window seek type: %d", seektype);
3554 : abs_pos = mark_pos = 0; /* keep compiler quiet */
3555 : break;
3556 : }
3557 :
3558 8520 : if (!window_gettupleslot(winobj, abs_pos, slot))
3559 396 : goto out_of_frame;
3560 :
3561 : /* The code above does not detect all out-of-frame cases, so check */
3562 8124 : if (row_is_in_frame(winstate, abs_pos, slot) <= 0)
3563 300 : goto out_of_frame;
3564 :
3565 7794 : if (isout)
3566 0 : *isout = false;
3567 7794 : if (set_mark)
3568 7752 : WinSetMarkPosition(winobj, mark_pos);
3569 7794 : econtext->ecxt_outertuple = slot;
3570 7794 : return ExecEvalExpr((ExprState *) list_nth(winobj->argstates, argno),
3571 : econtext, isnull);
3572 :
3573 756 : out_of_frame:
3574 756 : if (isout)
3575 0 : *isout = true;
3576 756 : *isnull = true;
3577 756 : return (Datum) 0;
3578 : }
3579 :
3580 : /*
3581 : * WinGetFuncArgCurrent
3582 : * Evaluate a window function's argument expression on the current row.
3583 : *
3584 : * argno: argument number to evaluate (counted from 0)
3585 : * isnull: output argument, receives isnull status of result
3586 : *
3587 : * Note: this isn't quite equivalent to WinGetFuncArgInPartition or
3588 : * WinGetFuncArgInFrame targeting the current row, because it will succeed
3589 : * even if the WindowObject's mark has been set beyond the current row.
3590 : * This should generally be used for "ordinary" arguments of a window
3591 : * function, such as the offset argument of lead() or lag().
3592 : */
3593 : Datum
3594 1164 : WinGetFuncArgCurrent(WindowObject winobj, int argno, bool *isnull)
3595 : {
3596 : WindowAggState *winstate;
3597 : ExprContext *econtext;
3598 :
3599 : Assert(WindowObjectIsValid(winobj));
3600 1164 : winstate = winobj->winstate;
3601 :
3602 1164 : econtext = winstate->ss.ps.ps_ExprContext;
3603 :
3604 1164 : econtext->ecxt_outertuple = winstate->ss.ss_ScanTupleSlot;
3605 1164 : return ExecEvalExpr((ExprState *) list_nth(winobj->argstates, argno),
3606 : econtext, isnull);
3607 : }
|
