Line data Source code
1 : /*-------------------------------------------------------------------------
2 : *
3 : * parse_collate.c
4 : * Routines for assigning collation information.
5 : *
6 : * We choose to handle collation analysis in a post-pass over the output
7 : * of expression parse analysis. This is because we need more state to
8 : * perform this processing than is needed in the finished tree. If we
9 : * did it on-the-fly while building the tree, all that state would have
10 : * to be kept in expression node trees permanently. This way, the extra
11 : * storage is just local variables in this recursive routine.
12 : *
13 : * The info that is actually saved in the finished tree is:
14 : * 1. The output collation of each expression node, or InvalidOid if it
15 : * returns a noncollatable data type. This can also be InvalidOid if the
16 : * result type is collatable but the collation is indeterminate.
17 : * 2. The collation to be used in executing each function. InvalidOid means
18 : * that there are no collatable inputs or their collation is indeterminate.
19 : * This value is only stored in node types that might call collation-using
20 : * functions.
21 : *
22 : * You might think we could get away with storing only one collation per
23 : * node, but the two concepts really need to be kept distinct. Otherwise
24 : * it's too confusing when a function produces a collatable output type but
25 : * has no collatable inputs or produces noncollatable output from collatable
26 : * inputs.
27 : *
28 : * Cases with indeterminate collation might result in an error being thrown
29 : * at runtime. If we knew exactly which functions require collation
30 : * information, we could throw those errors at parse time instead.
31 : *
32 : * Portions Copyright (c) 1996-2026, PostgreSQL Global Development Group
33 : * Portions Copyright (c) 1994, Regents of the University of California
34 : *
35 : *
36 : * IDENTIFICATION
37 : * src/backend/parser/parse_collate.c
38 : *
39 : *-------------------------------------------------------------------------
40 : */
41 : #include "postgres.h"
42 :
43 : #include "catalog/pg_aggregate.h"
44 : #include "catalog/pg_collation.h"
45 : #include "nodes/makefuncs.h"
46 : #include "nodes/nodeFuncs.h"
47 : #include "parser/parse_collate.h"
48 : #include "utils/lsyscache.h"
49 :
50 :
51 : /*
52 : * Collation strength (the SQL standard calls this "derivation"). Order is
53 : * chosen to allow comparisons to work usefully. Note: the standard doesn't
54 : * seem to distinguish between NONE and CONFLICT.
55 : */
56 : typedef enum
57 : {
58 : COLLATE_NONE, /* expression is of a noncollatable datatype */
59 : COLLATE_IMPLICIT, /* collation was derived implicitly */
60 : COLLATE_CONFLICT, /* we had a conflict of implicit collations */
61 : COLLATE_EXPLICIT, /* collation was derived explicitly */
62 : } CollateStrength;
63 :
64 : typedef struct
65 : {
66 : ParseState *pstate; /* parse state (for error reporting) */
67 : Oid collation; /* OID of current collation, if any */
68 : CollateStrength strength; /* strength of current collation choice */
69 : int location; /* location of expr that set collation */
70 : /* Remaining fields are only valid when strength == COLLATE_CONFLICT */
71 : Oid collation2; /* OID of conflicting collation */
72 : int location2; /* location of expr that set collation2 */
73 : } assign_collations_context;
74 :
75 : static bool assign_query_collations_walker(Node *node, ParseState *pstate);
76 : static bool assign_collations_walker(Node *node,
77 : assign_collations_context *context);
78 : static void merge_collation_state(Oid collation,
79 : CollateStrength strength,
80 : int location,
81 : Oid collation2,
82 : int location2,
83 : assign_collations_context *context);
84 : static void assign_aggregate_collations(Aggref *aggref,
85 : assign_collations_context *loccontext);
86 : static void assign_ordered_set_collations(Aggref *aggref,
87 : assign_collations_context *loccontext);
88 : static void assign_hypothetical_collations(Aggref *aggref,
89 : assign_collations_context *loccontext);
90 :
91 :
92 : /*
93 : * assign_query_collations()
94 : * Mark all expressions in the given Query with collation information.
95 : *
96 : * This should be applied to each Query after completion of parse analysis
97 : * for expressions. Note that we do not recurse into sub-Queries, since
98 : * those should have been processed when built.
99 : */
100 : void
101 303344 : assign_query_collations(ParseState *pstate, Query *query)
102 : {
103 : /*
104 : * We just use query_tree_walker() to visit all the contained expressions.
105 : * We can skip the rangetable and CTE subqueries, though, since RTEs and
106 : * subqueries had better have been processed already (else Vars referring
107 : * to them would not get created with the right collation).
108 : */
109 303344 : (void) query_tree_walker(query,
110 : assign_query_collations_walker,
111 : pstate,
112 : QTW_IGNORE_RANGE_TABLE |
113 : QTW_IGNORE_CTE_SUBQUERIES);
114 303323 : }
115 :
116 : /*
117 : * Walker for assign_query_collations
118 : *
119 : * Each expression found by query_tree_walker is processed independently.
120 : * Note that query_tree_walker may pass us a whole List, such as the
121 : * targetlist, in which case each subexpression must be processed
122 : * independently --- we don't want to bleat if two different targetentries
123 : * have different collations.
124 : */
125 : static bool
126 3339664 : assign_query_collations_walker(Node *node, ParseState *pstate)
127 : {
128 : /* Need do nothing for empty subexpressions */
129 3339664 : if (node == NULL)
130 2726151 : return false;
131 :
132 : /*
133 : * We don't want to recurse into a set-operations tree; it's already been
134 : * fully processed in transformSetOperationStmt.
135 : */
136 613513 : if (IsA(node, SetOperationStmt))
137 6694 : return false;
138 :
139 606819 : if (IsA(node, List))
140 297115 : assign_list_collations(pstate, (List *) node);
141 : else
142 309704 : assign_expr_collations(pstate, node);
143 :
144 606798 : return false;
145 : }
146 :
147 : /*
148 : * assign_list_collations()
149 : * Mark all nodes in the list of expressions with collation information.
150 : *
151 : * The list member expressions are processed independently; they do not have
152 : * to share a common collation.
153 : */
154 : void
155 333048 : assign_list_collations(ParseState *pstate, List *exprs)
156 : {
157 : ListCell *lc;
158 :
159 1194035 : foreach(lc, exprs)
160 : {
161 861002 : Node *node = (Node *) lfirst(lc);
162 :
163 861002 : assign_expr_collations(pstate, node);
164 : }
165 333033 : }
166 :
167 : /*
168 : * assign_expr_collations()
169 : * Mark all nodes in the given expression tree with collation information.
170 : *
171 : * This is exported for the benefit of various utility commands that process
172 : * expressions without building a complete Query. It should be applied after
173 : * calling transformExpr() plus any expression-modifying operations such as
174 : * coerce_to_boolean().
175 : */
176 : void
177 1242886 : assign_expr_collations(ParseState *pstate, Node *expr)
178 : {
179 : assign_collations_context context;
180 :
181 : /* initialize context for tree walk */
182 1242886 : context.pstate = pstate;
183 1242886 : context.collation = InvalidOid;
184 1242886 : context.strength = COLLATE_NONE;
185 1242886 : context.location = -1;
186 :
187 : /* and away we go */
188 1242886 : (void) assign_collations_walker(expr, &context);
189 1242862 : }
190 :
191 : /*
192 : * select_common_collation()
193 : * Identify a common collation for a list of expressions.
194 : *
195 : * The expressions should all return the same datatype, else this is not
196 : * terribly meaningful.
197 : *
198 : * none_ok means that it is permitted to return InvalidOid, indicating that
199 : * no common collation could be identified, even for collatable datatypes.
200 : * Otherwise, an error is thrown for conflict of implicit collations.
201 : *
202 : * In theory, none_ok = true reflects the rules of SQL standard clause "Result
203 : * of data type combinations", none_ok = false reflects the rules of clause
204 : * "Collation determination" (in some cases invoked via "Grouping
205 : * operations").
206 : */
207 : Oid
208 38071 : select_common_collation(ParseState *pstate, List *exprs, bool none_ok)
209 : {
210 : assign_collations_context context;
211 :
212 : /* initialize context for tree walk */
213 38071 : context.pstate = pstate;
214 38071 : context.collation = InvalidOid;
215 38071 : context.strength = COLLATE_NONE;
216 38071 : context.location = -1;
217 :
218 : /* and away we go */
219 38071 : (void) assign_collations_walker((Node *) exprs, &context);
220 :
221 : /* deal with collation conflict */
222 38071 : if (context.strength == COLLATE_CONFLICT)
223 : {
224 36 : if (none_ok)
225 18 : return InvalidOid;
226 18 : ereport(ERROR,
227 : (errcode(ERRCODE_COLLATION_MISMATCH),
228 : errmsg("collation mismatch between implicit collations \"%s\" and \"%s\"",
229 : get_collation_name(context.collation),
230 : get_collation_name(context.collation2)),
231 : errhint("You can choose the collation by applying the COLLATE clause to one or both expressions."),
232 : parser_errposition(context.pstate, context.location2)));
233 : }
234 :
235 : /*
236 : * Note: if strength is still COLLATE_NONE, we'll return InvalidOid, but
237 : * that's okay because it must mean none of the expressions returned
238 : * collatable datatypes.
239 : */
240 38035 : return context.collation;
241 : }
242 :
243 : /*
244 : * assign_collations_walker()
245 : * Recursive guts of collation processing.
246 : *
247 : * Nodes with no children (eg, Vars, Consts, Params) must have been marked
248 : * when built. All upper-level nodes are marked here.
249 : *
250 : * Note: if this is invoked directly on a List, it will attempt to infer a
251 : * common collation for all the list members. In particular, it will throw
252 : * error if there are conflicting explicit collations for different members.
253 : */
254 : static bool
255 4661797 : assign_collations_walker(Node *node, assign_collations_context *context)
256 : {
257 : assign_collations_context loccontext;
258 : Oid collation;
259 : CollateStrength strength;
260 : int location;
261 :
262 : /* Need do nothing for empty subexpressions */
263 4661797 : if (node == NULL)
264 255950 : return false;
265 :
266 : /*
267 : * Prepare for recursion. For most node types, though not all, the first
268 : * thing we do is recurse to process all nodes below this one. Each level
269 : * of the tree has its own local context.
270 : */
271 4405847 : loccontext.pstate = context->pstate;
272 4405847 : loccontext.collation = InvalidOid;
273 4405847 : loccontext.strength = COLLATE_NONE;
274 4405847 : loccontext.location = -1;
275 : /* Set these fields just to suppress uninitialized-value warnings: */
276 4405847 : loccontext.collation2 = InvalidOid;
277 4405847 : loccontext.location2 = -1;
278 :
279 : /*
280 : * Recurse if appropriate, then determine the collation for this node.
281 : *
282 : * Note: the general cases are at the bottom of the switch, after various
283 : * special cases.
284 : */
285 4405847 : switch (nodeTag(node))
286 : {
287 4729 : case T_CollateExpr:
288 : {
289 : /*
290 : * COLLATE sets an explicitly derived collation, regardless of
291 : * what the child state is. But we must recurse to set up
292 : * collation info below here.
293 : */
294 4729 : CollateExpr *expr = (CollateExpr *) node;
295 :
296 4729 : (void) expression_tree_walker(node,
297 : assign_collations_walker,
298 : &loccontext);
299 :
300 4729 : collation = expr->collOid;
301 : Assert(OidIsValid(collation));
302 4729 : strength = COLLATE_EXPLICIT;
303 4729 : location = expr->location;
304 : }
305 4729 : break;
306 10916 : case T_FieldSelect:
307 : {
308 : /*
309 : * For FieldSelect, the result has the field's declared
310 : * collation, independently of what happened in the arguments.
311 : * (The immediate argument must be composite and thus not
312 : * collatable, anyhow.) The field's collation was already
313 : * looked up and saved in the node.
314 : */
315 10916 : FieldSelect *expr = (FieldSelect *) node;
316 :
317 : /* ... but first, recurse */
318 10916 : (void) expression_tree_walker(node,
319 : assign_collations_walker,
320 : &loccontext);
321 :
322 10916 : if (OidIsValid(expr->resultcollid))
323 : {
324 : /* Node's result type is collatable. */
325 : /* Pass up field's collation as an implicit choice. */
326 1528 : collation = expr->resultcollid;
327 1528 : strength = COLLATE_IMPLICIT;
328 1528 : location = exprLocation(node);
329 : }
330 : else
331 : {
332 : /* Node's result type isn't collatable. */
333 9388 : collation = InvalidOid;
334 9388 : strength = COLLATE_NONE;
335 9388 : location = -1; /* won't be used */
336 : }
337 : }
338 10916 : break;
339 2092 : case T_RowExpr:
340 : {
341 : /*
342 : * RowExpr is a special case because the subexpressions are
343 : * independent: we don't want to complain if some of them have
344 : * incompatible explicit collations.
345 : */
346 2092 : RowExpr *expr = (RowExpr *) node;
347 :
348 2092 : assign_list_collations(context->pstate, expr->args);
349 :
350 : /*
351 : * Since the result is always composite and therefore never
352 : * has a collation, we can just stop here: this node has no
353 : * impact on the collation of its parent.
354 : */
355 2092 : return false; /* done */
356 : }
357 135 : case T_RowCompareExpr:
358 : {
359 : /*
360 : * For RowCompare, we have to find the common collation of
361 : * each pair of input columns and build a list. If we can't
362 : * find a common collation, we just put InvalidOid into the
363 : * list, which may or may not cause an error at runtime.
364 : */
365 135 : RowCompareExpr *expr = (RowCompareExpr *) node;
366 135 : List *colls = NIL;
367 : ListCell *l;
368 : ListCell *r;
369 :
370 432 : forboth(l, expr->largs, r, expr->rargs)
371 : {
372 297 : Node *le = (Node *) lfirst(l);
373 297 : Node *re = (Node *) lfirst(r);
374 : Oid coll;
375 :
376 297 : coll = select_common_collation(context->pstate,
377 297 : list_make2(le, re),
378 : true);
379 297 : colls = lappend_oid(colls, coll);
380 : }
381 135 : expr->inputcollids = colls;
382 :
383 : /*
384 : * Since the result is always boolean and therefore never has
385 : * a collation, we can just stop here: this node has no impact
386 : * on the collation of its parent.
387 : */
388 135 : return false; /* done */
389 : }
390 45778 : case T_CoerceToDomain:
391 : {
392 : /*
393 : * If the domain declaration included a non-default COLLATE
394 : * spec, then use that collation as the output collation of
395 : * the coercion. Otherwise allow the input collation to
396 : * bubble up. (The input should be of the domain's base type,
397 : * therefore we don't need to worry about it not being
398 : * collatable when the domain is.)
399 : */
400 45778 : CoerceToDomain *expr = (CoerceToDomain *) node;
401 45778 : Oid typcollation = get_typcollation(expr->resulttype);
402 :
403 : /* ... but first, recurse */
404 45778 : (void) expression_tree_walker(node,
405 : assign_collations_walker,
406 : &loccontext);
407 :
408 45778 : if (OidIsValid(typcollation))
409 : {
410 : /* Node's result type is collatable. */
411 38280 : if (typcollation == DEFAULT_COLLATION_OID)
412 : {
413 : /* Collation state bubbles up from child. */
414 306 : collation = loccontext.collation;
415 306 : strength = loccontext.strength;
416 306 : location = loccontext.location;
417 : }
418 : else
419 : {
420 : /* Use domain's collation as an implicit choice. */
421 37974 : collation = typcollation;
422 37974 : strength = COLLATE_IMPLICIT;
423 37974 : location = exprLocation(node);
424 : }
425 : }
426 : else
427 : {
428 : /* Node's result type isn't collatable. */
429 7498 : collation = InvalidOid;
430 7498 : strength = COLLATE_NONE;
431 7498 : location = -1; /* won't be used */
432 : }
433 :
434 : /*
435 : * Save the state into the expression node. We know it
436 : * doesn't care about input collation.
437 : */
438 45778 : if (strength == COLLATE_CONFLICT)
439 0 : exprSetCollation(node, InvalidOid);
440 : else
441 45778 : exprSetCollation(node, collation);
442 : }
443 45778 : break;
444 835096 : case T_TargetEntry:
445 835096 : (void) expression_tree_walker(node,
446 : assign_collations_walker,
447 : &loccontext);
448 :
449 : /*
450 : * TargetEntry can have only one child, and should bubble that
451 : * state up to its parent. We can't use the general-case code
452 : * below because exprType and friends don't work on TargetEntry.
453 : */
454 835087 : collation = loccontext.collation;
455 835087 : strength = loccontext.strength;
456 835087 : location = loccontext.location;
457 :
458 : /*
459 : * Throw error if the collation is indeterminate for a TargetEntry
460 : * that is a sort/group target. We prefer to do this now, instead
461 : * of leaving the comparison functions to fail at runtime, because
462 : * we can give a syntax error pointer to help locate the problem.
463 : * There are some cases where there might not be a failure, for
464 : * example if the planner chooses to use hash aggregation instead
465 : * of sorting for grouping; but it seems better to predictably
466 : * throw an error. (Compare transformSetOperationTree, which will
467 : * throw error for indeterminate collation of set-op columns, even
468 : * though the planner might be able to implement the set-op
469 : * without sorting.)
470 : */
471 835087 : if (strength == COLLATE_CONFLICT &&
472 15 : ((TargetEntry *) node)->ressortgroupref != 0)
473 9 : ereport(ERROR,
474 : (errcode(ERRCODE_COLLATION_MISMATCH),
475 : errmsg("collation mismatch between implicit collations \"%s\" and \"%s\"",
476 : get_collation_name(loccontext.collation),
477 : get_collation_name(loccontext.collation2)),
478 : errhint("You can choose the collation by applying the COLLATE clause to one or both expressions."),
479 : parser_errposition(context->pstate,
480 : loccontext.location2)));
481 835078 : break;
482 619993 : case T_InferenceElem:
483 : case T_RangeTblRef:
484 : case T_JoinExpr:
485 : case T_FromExpr:
486 : case T_OnConflictExpr:
487 : case T_SortGroupClause:
488 : case T_MergeAction:
489 619993 : (void) expression_tree_walker(node,
490 : assign_collations_walker,
491 : &loccontext);
492 :
493 : /*
494 : * When we're invoked on a query's jointree, we don't need to do
495 : * anything with join nodes except recurse through them to process
496 : * WHERE/ON expressions. So just stop here. Likewise, we don't
497 : * need to do anything when invoked on sort/group lists.
498 : */
499 619987 : return false;
500 27477 : case T_Query:
501 : {
502 : /*
503 : * We get here when we're invoked on the Query belonging to a
504 : * SubLink. Act as though the Query returns its first output
505 : * column, which indeed is what it does for EXPR_SUBLINK and
506 : * ARRAY_SUBLINK cases. In the cases where the SubLink
507 : * returns boolean, this info will be ignored. Special case:
508 : * in EXISTS, the Query might return no columns, in which case
509 : * we need do nothing.
510 : *
511 : * We needn't recurse, since the Query is already processed.
512 : */
513 27477 : Query *qtree = (Query *) node;
514 : TargetEntry *tent;
515 :
516 27477 : if (qtree->targetList == NIL)
517 3 : return false;
518 27474 : tent = linitial_node(TargetEntry, qtree->targetList);
519 27474 : if (tent->resjunk)
520 0 : return false;
521 :
522 27474 : collation = exprCollation((Node *) tent->expr);
523 : /* collation doesn't change if it's converted to array */
524 27474 : strength = COLLATE_IMPLICIT;
525 27474 : location = exprLocation((Node *) tent->expr);
526 : }
527 27474 : break;
528 70783 : case T_List:
529 70783 : (void) expression_tree_walker(node,
530 : assign_collations_walker,
531 : &loccontext);
532 :
533 : /*
534 : * When processing a list, collation state just bubbles up from
535 : * the list elements.
536 : */
537 70783 : collation = loccontext.collation;
538 70783 : strength = loccontext.strength;
539 70783 : location = loccontext.location;
540 70783 : break;
541 :
542 1880669 : case T_Var:
543 : case T_Const:
544 : case T_Param:
545 : case T_CoerceToDomainValue:
546 : case T_CaseTestExpr:
547 : case T_SetToDefault:
548 : case T_CurrentOfExpr:
549 :
550 : /*
551 : * General case for childless expression nodes. These should
552 : * already have a collation assigned; it is not this function's
553 : * responsibility to look into the catalogs for base-case
554 : * information.
555 : */
556 1880669 : collation = exprCollation(node);
557 :
558 : /*
559 : * Note: in most cases, there will be an assigned collation
560 : * whenever type_is_collatable(exprType(node)); but an exception
561 : * occurs for a Var referencing a subquery output column for which
562 : * a unique collation was not determinable. That may lead to a
563 : * runtime failure if a collation-sensitive function is applied to
564 : * the Var.
565 : */
566 :
567 1880669 : if (OidIsValid(collation))
568 451302 : strength = COLLATE_IMPLICIT;
569 : else
570 1429367 : strength = COLLATE_NONE;
571 1880669 : location = exprLocation(node);
572 1880669 : break;
573 :
574 908179 : default:
575 : {
576 : /*
577 : * General case for most expression nodes with children. First
578 : * recurse, then figure out what to assign to this node.
579 : */
580 : Oid typcollation;
581 :
582 : /*
583 : * For most node types, we want to treat all the child
584 : * expressions alike; but there are a few exceptions, hence
585 : * this inner switch.
586 : */
587 908179 : switch (nodeTag(node))
588 : {
589 25041 : case T_Aggref:
590 : {
591 : /*
592 : * Aggref is messy enough that we give it its own
593 : * function, in fact three of them. The FILTER
594 : * clause is independent of the rest of the
595 : * aggregate, however, so it can be processed
596 : * separately.
597 : */
598 25041 : Aggref *aggref = (Aggref *) node;
599 :
600 25041 : switch (aggref->aggkind)
601 : {
602 24888 : case AGGKIND_NORMAL:
603 24888 : assign_aggregate_collations(aggref,
604 : &loccontext);
605 24882 : break;
606 90 : case AGGKIND_ORDERED_SET:
607 90 : assign_ordered_set_collations(aggref,
608 : &loccontext);
609 90 : break;
610 63 : case AGGKIND_HYPOTHETICAL:
611 63 : assign_hypothetical_collations(aggref,
612 : &loccontext);
613 60 : break;
614 0 : default:
615 0 : elog(ERROR, "unrecognized aggkind: %d",
616 : (int) aggref->aggkind);
617 : }
618 :
619 25032 : assign_expr_collations(context->pstate,
620 25032 : (Node *) aggref->aggfilter);
621 : }
622 25032 : break;
623 1962 : case T_WindowFunc:
624 : {
625 : /*
626 : * WindowFunc requires special processing only for
627 : * its aggfilter clause, as for aggregates.
628 : */
629 1962 : WindowFunc *wfunc = (WindowFunc *) node;
630 :
631 1962 : (void) assign_collations_walker((Node *) wfunc->args,
632 : &loccontext);
633 :
634 1962 : assign_expr_collations(context->pstate,
635 1962 : (Node *) wfunc->aggfilter);
636 : }
637 1962 : break;
638 24297 : case T_CaseExpr:
639 : {
640 : /*
641 : * CaseExpr is a special case because we do not
642 : * want to recurse into the test expression (if
643 : * any). It was already marked with collations
644 : * during transformCaseExpr, and furthermore its
645 : * collation is not relevant to the result of the
646 : * CASE --- only the output expressions are.
647 : */
648 24297 : CaseExpr *expr = (CaseExpr *) node;
649 : ListCell *lc;
650 :
651 65581 : foreach(lc, expr->args)
652 : {
653 41284 : CaseWhen *when = lfirst_node(CaseWhen, lc);
654 :
655 : /*
656 : * The condition expressions mustn't affect
657 : * the CASE's result collation either; but
658 : * since they are known to yield boolean, it's
659 : * safe to recurse directly on them --- they
660 : * won't change loccontext.
661 : */
662 41284 : (void) assign_collations_walker((Node *) when->expr,
663 : &loccontext);
664 41284 : (void) assign_collations_walker((Node *) when->result,
665 : &loccontext);
666 : }
667 24297 : (void) assign_collations_walker((Node *) expr->defresult,
668 : &loccontext);
669 : }
670 24297 : break;
671 6945 : case T_SubscriptingRef:
672 : {
673 : /*
674 : * The subscripts are treated as independent
675 : * expressions not contributing to the node's
676 : * collation. Only the container, and the source
677 : * expression if any, contribute. (This models
678 : * the old behavior, in which the subscripts could
679 : * be counted on to be integers and thus not
680 : * contribute anything.)
681 : */
682 6945 : SubscriptingRef *sbsref = (SubscriptingRef *) node;
683 :
684 6945 : assign_expr_collations(context->pstate,
685 6945 : (Node *) sbsref->refupperindexpr);
686 6945 : assign_expr_collations(context->pstate,
687 6945 : (Node *) sbsref->reflowerindexpr);
688 6945 : (void) assign_collations_walker((Node *) sbsref->refexpr,
689 : &loccontext);
690 6945 : (void) assign_collations_walker((Node *) sbsref->refassgnexpr,
691 : &loccontext);
692 : }
693 6945 : break;
694 849934 : default:
695 :
696 : /*
697 : * Normal case: all child expressions contribute
698 : * equally to loccontext.
699 : */
700 849934 : (void) expression_tree_walker(node,
701 : assign_collations_walker,
702 : &loccontext);
703 849928 : break;
704 : }
705 :
706 : /*
707 : * Now figure out what collation to assign to this node.
708 : */
709 908164 : typcollation = get_typcollation(exprType(node));
710 908164 : if (OidIsValid(typcollation))
711 : {
712 : /* Node's result is collatable; what about its input? */
713 160242 : if (loccontext.strength > COLLATE_NONE)
714 : {
715 : /* Collation state bubbles up from children. */
716 120019 : collation = loccontext.collation;
717 120019 : strength = loccontext.strength;
718 120019 : location = loccontext.location;
719 : }
720 : else
721 : {
722 : /*
723 : * Collatable output produced without any collatable
724 : * input. Use the type's collation (which is usually
725 : * DEFAULT_COLLATION_OID, but might be different for a
726 : * domain).
727 : */
728 40223 : collation = typcollation;
729 40223 : strength = COLLATE_IMPLICIT;
730 40223 : location = exprLocation(node);
731 : }
732 : }
733 : else
734 : {
735 : /* Node's result type isn't collatable. */
736 747922 : collation = InvalidOid;
737 747922 : strength = COLLATE_NONE;
738 747922 : location = -1; /* won't be used */
739 : }
740 :
741 : /*
742 : * Save the result collation into the expression node. If the
743 : * state is COLLATE_CONFLICT, we'll set the collation to
744 : * InvalidOid, which might result in an error at runtime.
745 : */
746 908164 : if (strength == COLLATE_CONFLICT)
747 15 : exprSetCollation(node, InvalidOid);
748 : else
749 908149 : exprSetCollation(node, collation);
750 :
751 : /*
752 : * Likewise save the input collation, which is the one that
753 : * any function called by this node should use.
754 : */
755 908164 : if (loccontext.strength == COLLATE_CONFLICT)
756 54 : exprSetInputCollation(node, InvalidOid);
757 : else
758 908110 : exprSetInputCollation(node, loccontext.collation);
759 : }
760 908164 : break;
761 : }
762 :
763 : /*
764 : * Now, merge my information into my parent's state.
765 : */
766 3783591 : merge_collation_state(collation,
767 : strength,
768 : location,
769 : loccontext.collation2,
770 : loccontext.location2,
771 : context);
772 :
773 3783579 : return false;
774 : }
775 :
776 : /*
777 : * Merge collation state of a subexpression into the context for its parent.
778 : */
779 : static void
780 3783591 : merge_collation_state(Oid collation,
781 : CollateStrength strength,
782 : int location,
783 : Oid collation2,
784 : int location2,
785 : assign_collations_context *context)
786 : {
787 : /*
788 : * If the collation strength for this node is different from what's
789 : * already in *context, then this node either dominates or is dominated by
790 : * earlier siblings.
791 : */
792 3783591 : if (strength > context->strength)
793 : {
794 : /* Override previous parent state */
795 864201 : context->collation = collation;
796 864201 : context->strength = strength;
797 864201 : context->location = location;
798 : /* Bubble up error info if applicable */
799 864201 : if (strength == COLLATE_CONFLICT)
800 : {
801 57 : context->collation2 = collation2;
802 57 : context->location2 = location2;
803 : }
804 : }
805 2919390 : else if (strength == context->strength)
806 : {
807 : /* Merge, or detect error if there's a collation conflict */
808 2859051 : switch (strength)
809 : {
810 2732419 : case COLLATE_NONE:
811 : /* Nothing + nothing is still nothing */
812 2732419 : break;
813 126587 : case COLLATE_IMPLICIT:
814 126587 : if (collation != context->collation)
815 : {
816 : /*
817 : * Non-default implicit collation always beats default.
818 : */
819 27695 : if (context->collation == DEFAULT_COLLATION_OID)
820 : {
821 : /* Override previous parent state */
822 7458 : context->collation = collation;
823 7458 : context->strength = strength;
824 7458 : context->location = location;
825 : }
826 20237 : else if (collation != DEFAULT_COLLATION_OID)
827 : {
828 : /*
829 : * Oops, we have a conflict. We cannot throw error
830 : * here, since the conflict could be resolved by a
831 : * later sibling CollateExpr, or the parent might not
832 : * care about collation anyway. Return enough info to
833 : * throw the error later, if needed.
834 : */
835 90 : context->strength = COLLATE_CONFLICT;
836 90 : context->collation2 = collation;
837 90 : context->location2 = location;
838 : }
839 : }
840 126587 : break;
841 0 : case COLLATE_CONFLICT:
842 : /* We're still conflicted ... */
843 0 : break;
844 45 : case COLLATE_EXPLICIT:
845 45 : if (collation != context->collation)
846 : {
847 : /*
848 : * Oops, we have a conflict of explicit COLLATE clauses.
849 : * Here we choose to throw error immediately; that is what
850 : * the SQL standard says to do, and there's no good reason
851 : * to be less strict.
852 : */
853 12 : ereport(ERROR,
854 : (errcode(ERRCODE_COLLATION_MISMATCH),
855 : errmsg("collation mismatch between explicit collations \"%s\" and \"%s\"",
856 : get_collation_name(context->collation),
857 : get_collation_name(collation)),
858 : parser_errposition(context->pstate, location)));
859 : }
860 33 : break;
861 : }
862 : }
863 3783579 : }
864 :
865 : /*
866 : * Aggref is a special case because expressions used only for ordering
867 : * shouldn't be taken to conflict with each other or with regular args,
868 : * indeed shouldn't affect the aggregate's result collation at all.
869 : * We handle this by applying assign_expr_collations() to them rather than
870 : * passing down our loccontext.
871 : *
872 : * Note that we recurse to each TargetEntry, not directly to its contained
873 : * expression, so that the case above for T_TargetEntry will complain if we
874 : * can't resolve a collation for an ORDER BY item (whether or not it is also
875 : * a normal aggregate arg).
876 : *
877 : * We need not recurse into the aggorder or aggdistinct lists, because those
878 : * contain only SortGroupClause nodes which we need not process.
879 : */
880 : static void
881 24888 : assign_aggregate_collations(Aggref *aggref,
882 : assign_collations_context *loccontext)
883 : {
884 : ListCell *lc;
885 :
886 : /* Plain aggregates have no direct args */
887 : Assert(aggref->aggdirectargs == NIL);
888 :
889 : /* Process aggregated args, holding resjunk ones at arm's length */
890 43747 : foreach(lc, aggref->args)
891 : {
892 18865 : TargetEntry *tle = lfirst_node(TargetEntry, lc);
893 :
894 18865 : if (tle->resjunk)
895 917 : assign_expr_collations(loccontext->pstate, (Node *) tle);
896 : else
897 17948 : (void) assign_collations_walker((Node *) tle, loccontext);
898 : }
899 24882 : }
900 :
901 : /*
902 : * For ordered-set aggregates, it's somewhat unclear how best to proceed.
903 : * The spec-defined inverse distribution functions have only one sort column
904 : * and don't return collatable types, but this is clearly too restrictive in
905 : * the general case. Our solution is to consider that the aggregate's direct
906 : * arguments contribute normally to determination of the aggregate's own
907 : * collation, while aggregated arguments contribute only when the aggregate
908 : * is designed to have exactly one aggregated argument (i.e., it has a single
909 : * aggregated argument and is non-variadic). If it can have more than one
910 : * aggregated argument, we process the aggregated arguments as independent
911 : * sort columns. This avoids throwing error for something like
912 : * agg(...) within group (order by x collate "foo", y collate "bar")
913 : * while also guaranteeing that variadic aggregates don't change in behavior
914 : * depending on how many sort columns a particular call happens to have.
915 : *
916 : * Otherwise this is much like the plain-aggregate case.
917 : */
918 : static void
919 90 : assign_ordered_set_collations(Aggref *aggref,
920 : assign_collations_context *loccontext)
921 : {
922 : bool merge_sort_collations;
923 : ListCell *lc;
924 :
925 : /* Merge sort collations to parent only if there can be only one */
926 180 : merge_sort_collations = (list_length(aggref->args) == 1 &&
927 90 : get_func_variadictype(aggref->aggfnoid) == InvalidOid);
928 :
929 : /* Direct args, if any, are normal children of the Aggref node */
930 90 : (void) assign_collations_walker((Node *) aggref->aggdirectargs,
931 : loccontext);
932 :
933 : /* Process aggregated args appropriately */
934 180 : foreach(lc, aggref->args)
935 : {
936 90 : TargetEntry *tle = lfirst_node(TargetEntry, lc);
937 :
938 90 : if (merge_sort_collations)
939 90 : (void) assign_collations_walker((Node *) tle, loccontext);
940 : else
941 0 : assign_expr_collations(loccontext->pstate, (Node *) tle);
942 : }
943 90 : }
944 :
945 : /*
946 : * Hypothetical-set aggregates are even more special: per spec, we need to
947 : * unify the collations of each pair of hypothetical and aggregated args.
948 : * And we need to force the choice of collation down into the sort column
949 : * to ensure that the sort happens with the chosen collation. Other than
950 : * that, the behavior is like regular ordered-set aggregates. Note that
951 : * hypothetical direct arguments contribute to the aggregate collation
952 : * only when their partner aggregated arguments do.
953 : */
954 : static void
955 63 : assign_hypothetical_collations(Aggref *aggref,
956 : assign_collations_context *loccontext)
957 : {
958 63 : ListCell *h_cell = list_head(aggref->aggdirectargs);
959 63 : ListCell *s_cell = list_head(aggref->args);
960 : bool merge_sort_collations;
961 : int extra_args;
962 :
963 : /* Merge sort collations to parent only if there can be only one */
964 111 : merge_sort_collations = (list_length(aggref->args) == 1 &&
965 48 : get_func_variadictype(aggref->aggfnoid) == InvalidOid);
966 :
967 : /* Process any non-hypothetical direct args */
968 63 : extra_args = list_length(aggref->aggdirectargs) - list_length(aggref->args);
969 : Assert(extra_args >= 0);
970 63 : while (extra_args-- > 0)
971 : {
972 0 : (void) assign_collations_walker((Node *) lfirst(h_cell), loccontext);
973 0 : h_cell = lnext(aggref->aggdirectargs, h_cell);
974 : }
975 :
976 : /* Scan hypothetical args and aggregated args in parallel */
977 150 : while (h_cell && s_cell)
978 : {
979 90 : Node *h_arg = (Node *) lfirst(h_cell);
980 90 : TargetEntry *s_tle = (TargetEntry *) lfirst(s_cell);
981 : assign_collations_context paircontext;
982 :
983 : /*
984 : * Assign collations internally in this pair of expressions, then
985 : * choose a common collation for them. This should match
986 : * select_common_collation(), but we can't use that function as-is
987 : * because we need access to the whole collation state so we can
988 : * bubble it up to the aggregate function's level.
989 : */
990 90 : paircontext.pstate = loccontext->pstate;
991 90 : paircontext.collation = InvalidOid;
992 90 : paircontext.strength = COLLATE_NONE;
993 90 : paircontext.location = -1;
994 : /* Set these fields just to suppress uninitialized-value warnings: */
995 90 : paircontext.collation2 = InvalidOid;
996 90 : paircontext.location2 = -1;
997 :
998 90 : (void) assign_collations_walker(h_arg, &paircontext);
999 90 : (void) assign_collations_walker((Node *) s_tle->expr, &paircontext);
1000 :
1001 : /* deal with collation conflict */
1002 87 : if (paircontext.strength == COLLATE_CONFLICT)
1003 0 : ereport(ERROR,
1004 : (errcode(ERRCODE_COLLATION_MISMATCH),
1005 : errmsg("collation mismatch between implicit collations \"%s\" and \"%s\"",
1006 : get_collation_name(paircontext.collation),
1007 : get_collation_name(paircontext.collation2)),
1008 : errhint("You can choose the collation by applying the COLLATE clause to one or both expressions."),
1009 : parser_errposition(paircontext.pstate,
1010 : paircontext.location2)));
1011 :
1012 : /*
1013 : * At this point paircontext.collation can be InvalidOid only if the
1014 : * type is not collatable; no need to do anything in that case. If we
1015 : * do have to change the sort column's collation, do it by inserting a
1016 : * RelabelType node into the sort column TLE.
1017 : *
1018 : * XXX This is pretty grotty for a couple of reasons:
1019 : * assign_collations_walker isn't supposed to be changing the
1020 : * expression structure like this, and a parse-time change of
1021 : * collation ought to be signaled by a CollateExpr not a RelabelType
1022 : * (the use of RelabelType for collation marking is supposed to be a
1023 : * planner/executor thing only). But we have no better alternative.
1024 : * In particular, injecting a CollateExpr could result in the
1025 : * expression being interpreted differently after dump/reload, since
1026 : * we might be effectively promoting an implicit collation to
1027 : * explicit. This kluge is relying on ruleutils.c not printing a
1028 : * COLLATE clause for a RelabelType, and probably on some other
1029 : * fragile behaviors.
1030 : */
1031 102 : if (OidIsValid(paircontext.collation) &&
1032 15 : paircontext.collation != exprCollation((Node *) s_tle->expr))
1033 : {
1034 0 : s_tle->expr = (Expr *)
1035 0 : makeRelabelType(s_tle->expr,
1036 0 : exprType((Node *) s_tle->expr),
1037 0 : exprTypmod((Node *) s_tle->expr),
1038 : paircontext.collation,
1039 : COERCE_IMPLICIT_CAST);
1040 : }
1041 :
1042 : /*
1043 : * If appropriate, merge this column's collation state up to the
1044 : * aggregate function.
1045 : */
1046 87 : if (merge_sort_collations)
1047 0 : merge_collation_state(paircontext.collation,
1048 : paircontext.strength,
1049 : paircontext.location,
1050 : paircontext.collation2,
1051 : paircontext.location2,
1052 : loccontext);
1053 :
1054 87 : h_cell = lnext(aggref->aggdirectargs, h_cell);
1055 87 : s_cell = lnext(aggref->args, s_cell);
1056 : }
1057 : Assert(h_cell == NULL && s_cell == NULL);
1058 60 : }
|
