1 /*-------------------------------------------------------------------------
2  *
3  * clauses.c
4  *	  routines to manipulate qualification clauses
5  *
6  * Portions Copyright (c) 1996-2018, PostgreSQL Global Development Group
7  * Portions Copyright (c) 1994, Regents of the University of California
8  *
9  *
10  * IDENTIFICATION
11  *	  src/backend/optimizer/util/clauses.c
12  *
13  * HISTORY
14  *	  AUTHOR			DATE			MAJOR EVENT
15  *	  Andrew Yu			Nov 3, 1994		clause.c and clauses.c combined
16  *
17  *-------------------------------------------------------------------------
18  */
19 
20 #include "postgres.h"
21 
22 #include "access/htup_details.h"
23 #include "catalog/pg_aggregate.h"
24 #include "catalog/pg_class.h"
25 #include "catalog/pg_language.h"
26 #include "catalog/pg_operator.h"
27 #include "catalog/pg_proc.h"
28 #include "catalog/pg_type.h"
29 #include "executor/executor.h"
30 #include "executor/functions.h"
31 #include "funcapi.h"
32 #include "miscadmin.h"
33 #include "nodes/makefuncs.h"
34 #include "nodes/nodeFuncs.h"
35 #include "optimizer/clauses.h"
36 #include "optimizer/cost.h"
37 #include "optimizer/planmain.h"
38 #include "optimizer/prep.h"
39 #include "optimizer/var.h"
40 #include "parser/analyze.h"
41 #include "parser/parse_agg.h"
42 #include "parser/parse_coerce.h"
43 #include "parser/parse_func.h"
44 #include "rewrite/rewriteManip.h"
45 #include "tcop/tcopprot.h"
46 #include "utils/acl.h"
47 #include "utils/builtins.h"
48 #include "utils/datum.h"
49 #include "utils/fmgroids.h"
50 #include "utils/lsyscache.h"
51 #include "utils/memutils.h"
52 #include "utils/syscache.h"
53 #include "utils/typcache.h"
54 
55 
56 typedef struct
57 {
58 	PlannerInfo *root;
59 	AggSplit	aggsplit;
60 	AggClauseCosts *costs;
61 } get_agg_clause_costs_context;
62 
63 typedef struct
64 {
65 	ParamListInfo boundParams;
66 	PlannerInfo *root;
67 	List	   *active_fns;
68 	Node	   *case_val;
69 	bool		estimate;
70 } eval_const_expressions_context;
71 
72 typedef struct
73 {
74 	int			nargs;
75 	List	   *args;
76 	int		   *usecounts;
77 } substitute_actual_parameters_context;
78 
79 typedef struct
80 {
81 	int			nargs;
82 	List	   *args;
83 	int			sublevels_up;
84 } substitute_actual_srf_parameters_context;
85 
86 typedef struct
87 {
88 	char	   *proname;
89 	char	   *prosrc;
90 } inline_error_callback_arg;
91 
92 typedef struct
93 {
94 	char		max_hazard;		/* worst proparallel hazard found so far */
95 	char		max_interesting;	/* worst proparallel hazard of interest */
96 	List	   *safe_param_ids; /* PARAM_EXEC Param IDs to treat as safe */
97 } max_parallel_hazard_context;
98 
99 static bool contain_agg_clause_walker(Node *node, void *context);
100 static bool get_agg_clause_costs_walker(Node *node,
101 							get_agg_clause_costs_context *context);
102 static bool find_window_functions_walker(Node *node, WindowFuncLists *lists);
103 static bool contain_subplans_walker(Node *node, void *context);
104 static bool contain_mutable_functions_walker(Node *node, void *context);
105 static bool contain_volatile_functions_walker(Node *node, void *context);
106 static bool contain_volatile_functions_not_nextval_walker(Node *node, void *context);
107 static bool max_parallel_hazard_walker(Node *node,
108 						   max_parallel_hazard_context *context);
109 static bool contain_nonstrict_functions_walker(Node *node, void *context);
110 static bool contain_exec_param_walker(Node *node, List *param_ids);
111 static bool contain_context_dependent_node(Node *clause);
112 static bool contain_context_dependent_node_walker(Node *node, int *flags);
113 static bool contain_leaked_vars_walker(Node *node, void *context);
114 static Relids find_nonnullable_rels_walker(Node *node, bool top_level);
115 static List *find_nonnullable_vars_walker(Node *node, bool top_level);
116 static bool is_strict_saop(ScalarArrayOpExpr *expr, bool falseOK);
117 static Node *eval_const_expressions_mutator(Node *node,
118 							   eval_const_expressions_context *context);
119 static bool contain_non_const_walker(Node *node, void *context);
120 static bool ece_function_is_safe(Oid funcid,
121 					 eval_const_expressions_context *context);
122 static List *simplify_or_arguments(List *args,
123 					  eval_const_expressions_context *context,
124 					  bool *haveNull, bool *forceTrue);
125 static List *simplify_and_arguments(List *args,
126 					   eval_const_expressions_context *context,
127 					   bool *haveNull, bool *forceFalse);
128 static Node *simplify_boolean_equality(Oid opno, List *args);
129 static Expr *simplify_function(Oid funcid,
130 				  Oid result_type, int32 result_typmod,
131 				  Oid result_collid, Oid input_collid, List **args_p,
132 				  bool funcvariadic, bool process_args, bool allow_non_const,
133 				  eval_const_expressions_context *context);
134 static List *reorder_function_arguments(List *args, HeapTuple func_tuple);
135 static List *add_function_defaults(List *args, HeapTuple func_tuple);
136 static List *fetch_function_defaults(HeapTuple func_tuple);
137 static void recheck_cast_function_args(List *args, Oid result_type,
138 						   HeapTuple func_tuple);
139 static Expr *evaluate_function(Oid funcid, Oid result_type, int32 result_typmod,
140 				  Oid result_collid, Oid input_collid, List *args,
141 				  bool funcvariadic,
142 				  HeapTuple func_tuple,
143 				  eval_const_expressions_context *context);
144 static Expr *inline_function(Oid funcid, Oid result_type, Oid result_collid,
145 				Oid input_collid, List *args,
146 				bool funcvariadic,
147 				HeapTuple func_tuple,
148 				eval_const_expressions_context *context);
149 static Node *substitute_actual_parameters(Node *expr, int nargs, List *args,
150 							 int *usecounts);
151 static Node *substitute_actual_parameters_mutator(Node *node,
152 									 substitute_actual_parameters_context *context);
153 static void sql_inline_error_callback(void *arg);
154 static Expr *evaluate_expr(Expr *expr, Oid result_type, int32 result_typmod,
155 			  Oid result_collation);
156 static Query *substitute_actual_srf_parameters(Query *expr,
157 								 int nargs, List *args);
158 static Node *substitute_actual_srf_parameters_mutator(Node *node,
159 										 substitute_actual_srf_parameters_context *context);
160 static bool tlist_matches_coltypelist(List *tlist, List *coltypelist);
161 
162 
163 /*****************************************************************************
164  *		OPERATOR clause functions
165  *****************************************************************************/
166 
167 /*
168  * make_opclause
169  *	  Creates an operator clause given its operator info, left operand
170  *	  and right operand (pass NULL to create single-operand clause),
171  *	  and collation info.
172  */
173 Expr *
make_opclause(Oid opno,Oid opresulttype,bool opretset,Expr * leftop,Expr * rightop,Oid opcollid,Oid inputcollid)174 make_opclause(Oid opno, Oid opresulttype, bool opretset,
175 			  Expr *leftop, Expr *rightop,
176 			  Oid opcollid, Oid inputcollid)
177 {
178 	OpExpr	   *expr = makeNode(OpExpr);
179 
180 	expr->opno = opno;
181 	expr->opfuncid = InvalidOid;
182 	expr->opresulttype = opresulttype;
183 	expr->opretset = opretset;
184 	expr->opcollid = opcollid;
185 	expr->inputcollid = inputcollid;
186 	if (rightop)
187 		expr->args = list_make2(leftop, rightop);
188 	else
189 		expr->args = list_make1(leftop);
190 	expr->location = -1;
191 	return (Expr *) expr;
192 }
193 
194 /*
195  * get_leftop
196  *
197  * Returns the left operand of a clause of the form (op expr expr)
198  *		or (op expr)
199  */
200 Node *
get_leftop(const Expr * clause)201 get_leftop(const Expr *clause)
202 {
203 	const OpExpr *expr = (const OpExpr *) clause;
204 
205 	if (expr->args != NIL)
206 		return linitial(expr->args);
207 	else
208 		return NULL;
209 }
210 
211 /*
212  * get_rightop
213  *
214  * Returns the right operand in a clause of the form (op expr expr).
215  * NB: result will be NULL if applied to a unary op clause.
216  */
217 Node *
get_rightop(const Expr * clause)218 get_rightop(const Expr *clause)
219 {
220 	const OpExpr *expr = (const OpExpr *) clause;
221 
222 	if (list_length(expr->args) >= 2)
223 		return lsecond(expr->args);
224 	else
225 		return NULL;
226 }
227 
228 /*****************************************************************************
229  *		NOT clause functions
230  *****************************************************************************/
231 
232 /*
233  * not_clause
234  *
235  * Returns t iff this is a 'not' clause: (NOT expr).
236  */
237 bool
not_clause(Node * clause)238 not_clause(Node *clause)
239 {
240 	return (clause != NULL &&
241 			IsA(clause, BoolExpr) &&
242 			((BoolExpr *) clause)->boolop == NOT_EXPR);
243 }
244 
245 /*
246  * make_notclause
247  *
248  * Create a 'not' clause given the expression to be negated.
249  */
250 Expr *
make_notclause(Expr * notclause)251 make_notclause(Expr *notclause)
252 {
253 	BoolExpr   *expr = makeNode(BoolExpr);
254 
255 	expr->boolop = NOT_EXPR;
256 	expr->args = list_make1(notclause);
257 	expr->location = -1;
258 	return (Expr *) expr;
259 }
260 
261 /*
262  * get_notclausearg
263  *
264  * Retrieve the clause within a 'not' clause
265  */
266 Expr *
get_notclausearg(Expr * notclause)267 get_notclausearg(Expr *notclause)
268 {
269 	return linitial(((BoolExpr *) notclause)->args);
270 }
271 
272 /*****************************************************************************
273  *		OR clause functions
274  *****************************************************************************/
275 
276 /*
277  * or_clause
278  *
279  * Returns t iff the clause is an 'or' clause: (OR { expr }).
280  */
281 bool
or_clause(Node * clause)282 or_clause(Node *clause)
283 {
284 	return (clause != NULL &&
285 			IsA(clause, BoolExpr) &&
286 			((BoolExpr *) clause)->boolop == OR_EXPR);
287 }
288 
289 /*
290  * make_orclause
291  *
292  * Creates an 'or' clause given a list of its subclauses.
293  */
294 Expr *
make_orclause(List * orclauses)295 make_orclause(List *orclauses)
296 {
297 	BoolExpr   *expr = makeNode(BoolExpr);
298 
299 	expr->boolop = OR_EXPR;
300 	expr->args = orclauses;
301 	expr->location = -1;
302 	return (Expr *) expr;
303 }
304 
305 /*****************************************************************************
306  *		AND clause functions
307  *****************************************************************************/
308 
309 
310 /*
311  * and_clause
312  *
313  * Returns t iff its argument is an 'and' clause: (AND { expr }).
314  */
315 bool
and_clause(Node * clause)316 and_clause(Node *clause)
317 {
318 	return (clause != NULL &&
319 			IsA(clause, BoolExpr) &&
320 			((BoolExpr *) clause)->boolop == AND_EXPR);
321 }
322 
323 /*
324  * make_andclause
325  *
326  * Creates an 'and' clause given a list of its subclauses.
327  */
328 Expr *
make_andclause(List * andclauses)329 make_andclause(List *andclauses)
330 {
331 	BoolExpr   *expr = makeNode(BoolExpr);
332 
333 	expr->boolop = AND_EXPR;
334 	expr->args = andclauses;
335 	expr->location = -1;
336 	return (Expr *) expr;
337 }
338 
339 /*
340  * make_and_qual
341  *
342  * Variant of make_andclause for ANDing two qual conditions together.
343  * Qual conditions have the property that a NULL nodetree is interpreted
344  * as 'true'.
345  *
346  * NB: this makes no attempt to preserve AND/OR flatness; so it should not
347  * be used on a qual that has already been run through prepqual.c.
348  */
349 Node *
make_and_qual(Node * qual1,Node * qual2)350 make_and_qual(Node *qual1, Node *qual2)
351 {
352 	if (qual1 == NULL)
353 		return qual2;
354 	if (qual2 == NULL)
355 		return qual1;
356 	return (Node *) make_andclause(list_make2(qual1, qual2));
357 }
358 
359 /*
360  * The planner frequently prefers to represent qualification expressions
361  * as lists of boolean expressions with implicit AND semantics.
362  *
363  * These functions convert between an AND-semantics expression list and the
364  * ordinary representation of a boolean expression.
365  *
366  * Note that an empty list is considered equivalent to TRUE.
367  */
368 Expr *
make_ands_explicit(List * andclauses)369 make_ands_explicit(List *andclauses)
370 {
371 	if (andclauses == NIL)
372 		return (Expr *) makeBoolConst(true, false);
373 	else if (list_length(andclauses) == 1)
374 		return (Expr *) linitial(andclauses);
375 	else
376 		return make_andclause(andclauses);
377 }
378 
379 List *
make_ands_implicit(Expr * clause)380 make_ands_implicit(Expr *clause)
381 {
382 	/*
383 	 * NB: because the parser sets the qual field to NULL in a query that has
384 	 * no WHERE clause, we must consider a NULL input clause as TRUE, even
385 	 * though one might more reasonably think it FALSE.  Grumble. If this
386 	 * causes trouble, consider changing the parser's behavior.
387 	 */
388 	if (clause == NULL)
389 		return NIL;				/* NULL -> NIL list == TRUE */
390 	else if (and_clause((Node *) clause))
391 		return ((BoolExpr *) clause)->args;
392 	else if (IsA(clause, Const) &&
393 			 !((Const *) clause)->constisnull &&
394 			 DatumGetBool(((Const *) clause)->constvalue))
395 		return NIL;				/* constant TRUE input -> NIL list */
396 	else
397 		return list_make1(clause);
398 }
399 
400 
401 /*****************************************************************************
402  *		Aggregate-function clause manipulation
403  *****************************************************************************/
404 
405 /*
406  * contain_agg_clause
407  *	  Recursively search for Aggref/GroupingFunc nodes within a clause.
408  *
409  *	  Returns true if any aggregate found.
410  *
411  * This does not descend into subqueries, and so should be used only after
412  * reduction of sublinks to subplans, or in contexts where it's known there
413  * are no subqueries.  There mustn't be outer-aggregate references either.
414  *
415  * (If you want something like this but able to deal with subqueries,
416  * see rewriteManip.c's contain_aggs_of_level().)
417  */
418 bool
contain_agg_clause(Node * clause)419 contain_agg_clause(Node *clause)
420 {
421 	return contain_agg_clause_walker(clause, NULL);
422 }
423 
424 static bool
contain_agg_clause_walker(Node * node,void * context)425 contain_agg_clause_walker(Node *node, void *context)
426 {
427 	if (node == NULL)
428 		return false;
429 	if (IsA(node, Aggref))
430 	{
431 		Assert(((Aggref *) node)->agglevelsup == 0);
432 		return true;			/* abort the tree traversal and return true */
433 	}
434 	if (IsA(node, GroupingFunc))
435 	{
436 		Assert(((GroupingFunc *) node)->agglevelsup == 0);
437 		return true;			/* abort the tree traversal and return true */
438 	}
439 	Assert(!IsA(node, SubLink));
440 	return expression_tree_walker(node, contain_agg_clause_walker, context);
441 }
442 
443 /*
444  * get_agg_clause_costs
445  *	  Recursively find the Aggref nodes in an expression tree, and
446  *	  accumulate cost information about them.
447  *
448  * 'aggsplit' tells us the expected partial-aggregation mode, which affects
449  * the cost estimates.
450  *
451  * NOTE that the counts/costs are ADDED to those already in *costs ... so
452  * the caller is responsible for zeroing the struct initially.
453  *
454  * We count the nodes, estimate their execution costs, and estimate the total
455  * space needed for their transition state values if all are evaluated in
456  * parallel (as would be done in a HashAgg plan).  Also, we check whether
457  * partial aggregation is feasible.  See AggClauseCosts for the exact set
458  * of statistics collected.
459  *
460  * In addition, we mark Aggref nodes with the correct aggtranstype, so
461  * that that doesn't need to be done repeatedly.  (That makes this function's
462  * name a bit of a misnomer.)
463  *
464  * This does not descend into subqueries, and so should be used only after
465  * reduction of sublinks to subplans, or in contexts where it's known there
466  * are no subqueries.  There mustn't be outer-aggregate references either.
467  */
468 void
get_agg_clause_costs(PlannerInfo * root,Node * clause,AggSplit aggsplit,AggClauseCosts * costs)469 get_agg_clause_costs(PlannerInfo *root, Node *clause, AggSplit aggsplit,
470 					 AggClauseCosts *costs)
471 {
472 	get_agg_clause_costs_context context;
473 
474 	context.root = root;
475 	context.aggsplit = aggsplit;
476 	context.costs = costs;
477 	(void) get_agg_clause_costs_walker(clause, &context);
478 }
479 
480 static bool
get_agg_clause_costs_walker(Node * node,get_agg_clause_costs_context * context)481 get_agg_clause_costs_walker(Node *node, get_agg_clause_costs_context *context)
482 {
483 	if (node == NULL)
484 		return false;
485 	if (IsA(node, Aggref))
486 	{
487 		Aggref	   *aggref = (Aggref *) node;
488 		AggClauseCosts *costs = context->costs;
489 		HeapTuple	aggTuple;
490 		Form_pg_aggregate aggform;
491 		Oid			aggtransfn;
492 		Oid			aggfinalfn;
493 		Oid			aggcombinefn;
494 		Oid			aggserialfn;
495 		Oid			aggdeserialfn;
496 		Oid			aggtranstype;
497 		int32		aggtransspace;
498 		QualCost	argcosts;
499 
500 		Assert(aggref->agglevelsup == 0);
501 
502 		/*
503 		 * Fetch info about aggregate from pg_aggregate.  Note it's correct to
504 		 * ignore the moving-aggregate variant, since what we're concerned
505 		 * with here is aggregates not window functions.
506 		 */
507 		aggTuple = SearchSysCache1(AGGFNOID,
508 								   ObjectIdGetDatum(aggref->aggfnoid));
509 		if (!HeapTupleIsValid(aggTuple))
510 			elog(ERROR, "cache lookup failed for aggregate %u",
511 				 aggref->aggfnoid);
512 		aggform = (Form_pg_aggregate) GETSTRUCT(aggTuple);
513 		aggtransfn = aggform->aggtransfn;
514 		aggfinalfn = aggform->aggfinalfn;
515 		aggcombinefn = aggform->aggcombinefn;
516 		aggserialfn = aggform->aggserialfn;
517 		aggdeserialfn = aggform->aggdeserialfn;
518 		aggtranstype = aggform->aggtranstype;
519 		aggtransspace = aggform->aggtransspace;
520 		ReleaseSysCache(aggTuple);
521 
522 		/*
523 		 * Resolve the possibly-polymorphic aggregate transition type, unless
524 		 * already done in a previous pass over the expression.
525 		 */
526 		if (OidIsValid(aggref->aggtranstype))
527 			aggtranstype = aggref->aggtranstype;
528 		else
529 		{
530 			Oid			inputTypes[FUNC_MAX_ARGS];
531 			int			numArguments;
532 
533 			/* extract argument types (ignoring any ORDER BY expressions) */
534 			numArguments = get_aggregate_argtypes(aggref, inputTypes);
535 
536 			/* resolve actual type of transition state, if polymorphic */
537 			aggtranstype = resolve_aggregate_transtype(aggref->aggfnoid,
538 													   aggtranstype,
539 													   inputTypes,
540 													   numArguments);
541 			aggref->aggtranstype = aggtranstype;
542 		}
543 
544 		/*
545 		 * Count it, and check for cases requiring ordered input.  Note that
546 		 * ordered-set aggs always have nonempty aggorder.  Any ordered-input
547 		 * case also defeats partial aggregation.
548 		 */
549 		costs->numAggs++;
550 		if (aggref->aggorder != NIL || aggref->aggdistinct != NIL)
551 		{
552 			costs->numOrderedAggs++;
553 			costs->hasNonPartial = true;
554 		}
555 
556 		/*
557 		 * Check whether partial aggregation is feasible, unless we already
558 		 * found out that we can't do it.
559 		 */
560 		if (!costs->hasNonPartial)
561 		{
562 			/*
563 			 * If there is no combine function, then partial aggregation is
564 			 * not possible.
565 			 */
566 			if (!OidIsValid(aggcombinefn))
567 				costs->hasNonPartial = true;
568 
569 			/*
570 			 * If we have any aggs with transtype INTERNAL then we must check
571 			 * whether they have serialization/deserialization functions; if
572 			 * not, we can't serialize partial-aggregation results.
573 			 */
574 			else if (aggtranstype == INTERNALOID &&
575 					 (!OidIsValid(aggserialfn) || !OidIsValid(aggdeserialfn)))
576 				costs->hasNonSerial = true;
577 		}
578 
579 		/*
580 		 * Add the appropriate component function execution costs to
581 		 * appropriate totals.
582 		 */
583 		if (DO_AGGSPLIT_COMBINE(context->aggsplit))
584 		{
585 			/* charge for combining previously aggregated states */
586 			costs->transCost.per_tuple += get_func_cost(aggcombinefn) * cpu_operator_cost;
587 		}
588 		else
589 			costs->transCost.per_tuple += get_func_cost(aggtransfn) * cpu_operator_cost;
590 		if (DO_AGGSPLIT_DESERIALIZE(context->aggsplit) &&
591 			OidIsValid(aggdeserialfn))
592 			costs->transCost.per_tuple += get_func_cost(aggdeserialfn) * cpu_operator_cost;
593 		if (DO_AGGSPLIT_SERIALIZE(context->aggsplit) &&
594 			OidIsValid(aggserialfn))
595 			costs->finalCost += get_func_cost(aggserialfn) * cpu_operator_cost;
596 		if (!DO_AGGSPLIT_SKIPFINAL(context->aggsplit) &&
597 			OidIsValid(aggfinalfn))
598 			costs->finalCost += get_func_cost(aggfinalfn) * cpu_operator_cost;
599 
600 		/*
601 		 * These costs are incurred only by the initial aggregate node, so we
602 		 * mustn't include them again at upper levels.
603 		 */
604 		if (!DO_AGGSPLIT_COMBINE(context->aggsplit))
605 		{
606 			/* add the input expressions' cost to per-input-row costs */
607 			cost_qual_eval_node(&argcosts, (Node *) aggref->args, context->root);
608 			costs->transCost.startup += argcosts.startup;
609 			costs->transCost.per_tuple += argcosts.per_tuple;
610 
611 			/*
612 			 * Add any filter's cost to per-input-row costs.
613 			 *
614 			 * XXX Ideally we should reduce input expression costs according
615 			 * to filter selectivity, but it's not clear it's worth the
616 			 * trouble.
617 			 */
618 			if (aggref->aggfilter)
619 			{
620 				cost_qual_eval_node(&argcosts, (Node *) aggref->aggfilter,
621 									context->root);
622 				costs->transCost.startup += argcosts.startup;
623 				costs->transCost.per_tuple += argcosts.per_tuple;
624 			}
625 		}
626 
627 		/*
628 		 * If there are direct arguments, treat their evaluation cost like the
629 		 * cost of the finalfn.
630 		 */
631 		if (aggref->aggdirectargs)
632 		{
633 			cost_qual_eval_node(&argcosts, (Node *) aggref->aggdirectargs,
634 								context->root);
635 			costs->transCost.startup += argcosts.startup;
636 			costs->finalCost += argcosts.per_tuple;
637 		}
638 
639 		/*
640 		 * If the transition type is pass-by-value then it doesn't add
641 		 * anything to the required size of the hashtable.  If it is
642 		 * pass-by-reference then we have to add the estimated size of the
643 		 * value itself, plus palloc overhead.
644 		 */
645 		if (!get_typbyval(aggtranstype))
646 		{
647 			int32		avgwidth;
648 
649 			/* Use average width if aggregate definition gave one */
650 			if (aggtransspace > 0)
651 				avgwidth = aggtransspace;
652 			else if (aggtransfn == F_ARRAY_APPEND)
653 			{
654 				/*
655 				 * If the transition function is array_append(), it'll use an
656 				 * expanded array as transvalue, which will occupy at least
657 				 * ALLOCSET_SMALL_INITSIZE and possibly more.  Use that as the
658 				 * estimate for lack of a better idea.
659 				 */
660 				avgwidth = ALLOCSET_SMALL_INITSIZE;
661 			}
662 			else
663 			{
664 				/*
665 				 * If transition state is of same type as first aggregated
666 				 * input, assume it's the same typmod (same width) as well.
667 				 * This works for cases like MAX/MIN and is probably somewhat
668 				 * reasonable otherwise.
669 				 */
670 				int32		aggtranstypmod = -1;
671 
672 				if (aggref->args)
673 				{
674 					TargetEntry *tle = (TargetEntry *) linitial(aggref->args);
675 
676 					if (aggtranstype == exprType((Node *) tle->expr))
677 						aggtranstypmod = exprTypmod((Node *) tle->expr);
678 				}
679 
680 				avgwidth = get_typavgwidth(aggtranstype, aggtranstypmod);
681 			}
682 
683 			avgwidth = MAXALIGN(avgwidth);
684 			costs->transitionSpace += avgwidth + 2 * sizeof(void *);
685 		}
686 		else if (aggtranstype == INTERNALOID)
687 		{
688 			/*
689 			 * INTERNAL transition type is a special case: although INTERNAL
690 			 * is pass-by-value, it's almost certainly being used as a pointer
691 			 * to some large data structure.  The aggregate definition can
692 			 * provide an estimate of the size.  If it doesn't, then we assume
693 			 * ALLOCSET_DEFAULT_INITSIZE, which is a good guess if the data is
694 			 * being kept in a private memory context, as is done by
695 			 * array_agg() for instance.
696 			 */
697 			if (aggtransspace > 0)
698 				costs->transitionSpace += aggtransspace;
699 			else
700 				costs->transitionSpace += ALLOCSET_DEFAULT_INITSIZE;
701 		}
702 
703 		/*
704 		 * We assume that the parser checked that there are no aggregates (of
705 		 * this level anyway) in the aggregated arguments, direct arguments,
706 		 * or filter clause.  Hence, we need not recurse into any of them.
707 		 */
708 		return false;
709 	}
710 	Assert(!IsA(node, SubLink));
711 	return expression_tree_walker(node, get_agg_clause_costs_walker,
712 								  (void *) context);
713 }
714 
715 
716 /*****************************************************************************
717  *		Window-function clause manipulation
718  *****************************************************************************/
719 
720 /*
721  * contain_window_function
722  *	  Recursively search for WindowFunc nodes within a clause.
723  *
724  * Since window functions don't have level fields, but are hard-wired to
725  * be associated with the current query level, this is just the same as
726  * rewriteManip.c's function.
727  */
728 bool
contain_window_function(Node * clause)729 contain_window_function(Node *clause)
730 {
731 	return contain_windowfuncs(clause);
732 }
733 
734 /*
735  * find_window_functions
736  *	  Locate all the WindowFunc nodes in an expression tree, and organize
737  *	  them by winref ID number.
738  *
739  * Caller must provide an upper bound on the winref IDs expected in the tree.
740  */
741 WindowFuncLists *
find_window_functions(Node * clause,Index maxWinRef)742 find_window_functions(Node *clause, Index maxWinRef)
743 {
744 	WindowFuncLists *lists = palloc(sizeof(WindowFuncLists));
745 
746 	lists->numWindowFuncs = 0;
747 	lists->maxWinRef = maxWinRef;
748 	lists->windowFuncs = (List **) palloc0((maxWinRef + 1) * sizeof(List *));
749 	(void) find_window_functions_walker(clause, lists);
750 	return lists;
751 }
752 
753 static bool
find_window_functions_walker(Node * node,WindowFuncLists * lists)754 find_window_functions_walker(Node *node, WindowFuncLists *lists)
755 {
756 	if (node == NULL)
757 		return false;
758 	if (IsA(node, WindowFunc))
759 	{
760 		WindowFunc *wfunc = (WindowFunc *) node;
761 
762 		/* winref is unsigned, so one-sided test is OK */
763 		if (wfunc->winref > lists->maxWinRef)
764 			elog(ERROR, "WindowFunc contains out-of-range winref %u",
765 				 wfunc->winref);
766 		/* eliminate duplicates, so that we avoid repeated computation */
767 		if (!list_member(lists->windowFuncs[wfunc->winref], wfunc))
768 		{
769 			lists->windowFuncs[wfunc->winref] =
770 				lappend(lists->windowFuncs[wfunc->winref], wfunc);
771 			lists->numWindowFuncs++;
772 		}
773 
774 		/*
775 		 * We assume that the parser checked that there are no window
776 		 * functions in the arguments or filter clause.  Hence, we need not
777 		 * recurse into them.  (If either the parser or the planner screws up
778 		 * on this point, the executor will still catch it; see ExecInitExpr.)
779 		 */
780 		return false;
781 	}
782 	Assert(!IsA(node, SubLink));
783 	return expression_tree_walker(node, find_window_functions_walker,
784 								  (void *) lists);
785 }
786 
787 
788 /*****************************************************************************
789  *		Support for expressions returning sets
790  *****************************************************************************/
791 
792 /*
793  * expression_returns_set_rows
794  *	  Estimate the number of rows returned by a set-returning expression.
795  *	  The result is 1 if it's not a set-returning expression.
796  *
797  * We should only examine the top-level function or operator; it used to be
798  * appropriate to recurse, but not anymore.  (Even if there are more SRFs in
799  * the function's inputs, their multipliers are accounted for separately.)
800  *
801  * Note: keep this in sync with expression_returns_set() in nodes/nodeFuncs.c.
802  */
803 double
expression_returns_set_rows(Node * clause)804 expression_returns_set_rows(Node *clause)
805 {
806 	if (clause == NULL)
807 		return 1.0;
808 	if (IsA(clause, FuncExpr))
809 	{
810 		FuncExpr   *expr = (FuncExpr *) clause;
811 
812 		if (expr->funcretset)
813 			return clamp_row_est(get_func_rows(expr->funcid));
814 	}
815 	if (IsA(clause, OpExpr))
816 	{
817 		OpExpr	   *expr = (OpExpr *) clause;
818 
819 		if (expr->opretset)
820 		{
821 			set_opfuncid(expr);
822 			return clamp_row_est(get_func_rows(expr->opfuncid));
823 		}
824 	}
825 	return 1.0;
826 }
827 
828 
829 /*****************************************************************************
830  *		Subplan clause manipulation
831  *****************************************************************************/
832 
833 /*
834  * contain_subplans
835  *	  Recursively search for subplan nodes within a clause.
836  *
837  * If we see a SubLink node, we will return true.  This is only possible if
838  * the expression tree hasn't yet been transformed by subselect.c.  We do not
839  * know whether the node will produce a true subplan or just an initplan,
840  * but we make the conservative assumption that it will be a subplan.
841  *
842  * Returns true if any subplan found.
843  */
844 bool
contain_subplans(Node * clause)845 contain_subplans(Node *clause)
846 {
847 	return contain_subplans_walker(clause, NULL);
848 }
849 
850 static bool
contain_subplans_walker(Node * node,void * context)851 contain_subplans_walker(Node *node, void *context)
852 {
853 	if (node == NULL)
854 		return false;
855 	if (IsA(node, SubPlan) ||
856 		IsA(node, AlternativeSubPlan) ||
857 		IsA(node, SubLink))
858 		return true;			/* abort the tree traversal and return true */
859 	return expression_tree_walker(node, contain_subplans_walker, context);
860 }
861 
862 
863 /*****************************************************************************
864  *		Check clauses for mutable functions
865  *****************************************************************************/
866 
867 /*
868  * contain_mutable_functions
869  *	  Recursively search for mutable functions within a clause.
870  *
871  * Returns true if any mutable function (or operator implemented by a
872  * mutable function) is found.  This test is needed so that we don't
873  * mistakenly think that something like "WHERE random() < 0.5" can be treated
874  * as a constant qualification.
875  *
876  * We will recursively look into Query nodes (i.e., SubLink sub-selects)
877  * but not into SubPlans.  See comments for contain_volatile_functions().
878  */
879 bool
contain_mutable_functions(Node * clause)880 contain_mutable_functions(Node *clause)
881 {
882 	return contain_mutable_functions_walker(clause, NULL);
883 }
884 
885 static bool
contain_mutable_functions_checker(Oid func_id,void * context)886 contain_mutable_functions_checker(Oid func_id, void *context)
887 {
888 	return (func_volatile(func_id) != PROVOLATILE_IMMUTABLE);
889 }
890 
891 static bool
contain_mutable_functions_walker(Node * node,void * context)892 contain_mutable_functions_walker(Node *node, void *context)
893 {
894 	if (node == NULL)
895 		return false;
896 	/* Check for mutable functions in node itself */
897 	if (check_functions_in_node(node, contain_mutable_functions_checker,
898 								context))
899 		return true;
900 
901 	if (IsA(node, SQLValueFunction))
902 	{
903 		/* all variants of SQLValueFunction are stable */
904 		return true;
905 	}
906 
907 	if (IsA(node, NextValueExpr))
908 	{
909 		/* NextValueExpr is volatile */
910 		return true;
911 	}
912 
913 	/*
914 	 * It should be safe to treat MinMaxExpr as immutable, because it will
915 	 * depend on a non-cross-type btree comparison function, and those should
916 	 * always be immutable.  Treating XmlExpr as immutable is more dubious,
917 	 * and treating CoerceToDomain as immutable is outright dangerous.  But we
918 	 * have done so historically, and changing this would probably cause more
919 	 * problems than it would fix.  In practice, if you have a non-immutable
920 	 * domain constraint you are in for pain anyhow.
921 	 */
922 
923 	/* Recurse to check arguments */
924 	if (IsA(node, Query))
925 	{
926 		/* Recurse into subselects */
927 		return query_tree_walker((Query *) node,
928 								 contain_mutable_functions_walker,
929 								 context, 0);
930 	}
931 	return expression_tree_walker(node, contain_mutable_functions_walker,
932 								  context);
933 }
934 
935 
936 /*****************************************************************************
937  *		Check clauses for volatile functions
938  *****************************************************************************/
939 
940 /*
941  * contain_volatile_functions
942  *	  Recursively search for volatile functions within a clause.
943  *
944  * Returns true if any volatile function (or operator implemented by a
945  * volatile function) is found. This test prevents, for example,
946  * invalid conversions of volatile expressions into indexscan quals.
947  *
948  * We will recursively look into Query nodes (i.e., SubLink sub-selects)
949  * but not into SubPlans.  This is a bit odd, but intentional.  If we are
950  * looking at a SubLink, we are probably deciding whether a query tree
951  * transformation is safe, and a contained sub-select should affect that;
952  * for example, duplicating a sub-select containing a volatile function
953  * would be bad.  However, once we've got to the stage of having SubPlans,
954  * subsequent planning need not consider volatility within those, since
955  * the executor won't change its evaluation rules for a SubPlan based on
956  * volatility.
957  */
958 bool
contain_volatile_functions(Node * clause)959 contain_volatile_functions(Node *clause)
960 {
961 	return contain_volatile_functions_walker(clause, NULL);
962 }
963 
964 static bool
contain_volatile_functions_checker(Oid func_id,void * context)965 contain_volatile_functions_checker(Oid func_id, void *context)
966 {
967 	return (func_volatile(func_id) == PROVOLATILE_VOLATILE);
968 }
969 
970 static bool
contain_volatile_functions_walker(Node * node,void * context)971 contain_volatile_functions_walker(Node *node, void *context)
972 {
973 	if (node == NULL)
974 		return false;
975 	/* Check for volatile functions in node itself */
976 	if (check_functions_in_node(node, contain_volatile_functions_checker,
977 								context))
978 		return true;
979 
980 	if (IsA(node, NextValueExpr))
981 	{
982 		/* NextValueExpr is volatile */
983 		return true;
984 	}
985 
986 	/*
987 	 * See notes in contain_mutable_functions_walker about why we treat
988 	 * MinMaxExpr, XmlExpr, and CoerceToDomain as immutable, while
989 	 * SQLValueFunction is stable.  Hence, none of them are of interest here.
990 	 */
991 
992 	/* Recurse to check arguments */
993 	if (IsA(node, Query))
994 	{
995 		/* Recurse into subselects */
996 		return query_tree_walker((Query *) node,
997 								 contain_volatile_functions_walker,
998 								 context, 0);
999 	}
1000 	return expression_tree_walker(node, contain_volatile_functions_walker,
1001 								  context);
1002 }
1003 
1004 /*
1005  * Special purpose version of contain_volatile_functions() for use in COPY:
1006  * ignore nextval(), but treat all other functions normally.
1007  */
1008 bool
contain_volatile_functions_not_nextval(Node * clause)1009 contain_volatile_functions_not_nextval(Node *clause)
1010 {
1011 	return contain_volatile_functions_not_nextval_walker(clause, NULL);
1012 }
1013 
1014 static bool
contain_volatile_functions_not_nextval_checker(Oid func_id,void * context)1015 contain_volatile_functions_not_nextval_checker(Oid func_id, void *context)
1016 {
1017 	return (func_id != F_NEXTVAL_OID &&
1018 			func_volatile(func_id) == PROVOLATILE_VOLATILE);
1019 }
1020 
1021 static bool
contain_volatile_functions_not_nextval_walker(Node * node,void * context)1022 contain_volatile_functions_not_nextval_walker(Node *node, void *context)
1023 {
1024 	if (node == NULL)
1025 		return false;
1026 	/* Check for volatile functions in node itself */
1027 	if (check_functions_in_node(node,
1028 								contain_volatile_functions_not_nextval_checker,
1029 								context))
1030 		return true;
1031 
1032 	/*
1033 	 * See notes in contain_mutable_functions_walker about why we treat
1034 	 * MinMaxExpr, XmlExpr, and CoerceToDomain as immutable, while
1035 	 * SQLValueFunction is stable.  Hence, none of them are of interest here.
1036 	 * Also, since we're intentionally ignoring nextval(), presumably we
1037 	 * should ignore NextValueExpr.
1038 	 */
1039 
1040 	/* Recurse to check arguments */
1041 	if (IsA(node, Query))
1042 	{
1043 		/* Recurse into subselects */
1044 		return query_tree_walker((Query *) node,
1045 								 contain_volatile_functions_not_nextval_walker,
1046 								 context, 0);
1047 	}
1048 	return expression_tree_walker(node,
1049 								  contain_volatile_functions_not_nextval_walker,
1050 								  context);
1051 }
1052 
1053 
1054 /*****************************************************************************
1055  *		Check queries for parallel unsafe and/or restricted constructs
1056  *****************************************************************************/
1057 
1058 /*
1059  * max_parallel_hazard
1060  *		Find the worst parallel-hazard level in the given query
1061  *
1062  * Returns the worst function hazard property (the earliest in this list:
1063  * PROPARALLEL_UNSAFE, PROPARALLEL_RESTRICTED, PROPARALLEL_SAFE) that can
1064  * be found in the given parsetree.  We use this to find out whether the query
1065  * can be parallelized at all.  The caller will also save the result in
1066  * PlannerGlobal so as to short-circuit checks of portions of the querytree
1067  * later, in the common case where everything is SAFE.
1068  */
1069 char
max_parallel_hazard(Query * parse)1070 max_parallel_hazard(Query *parse)
1071 {
1072 	max_parallel_hazard_context context;
1073 
1074 	context.max_hazard = PROPARALLEL_SAFE;
1075 	context.max_interesting = PROPARALLEL_UNSAFE;
1076 	context.safe_param_ids = NIL;
1077 	(void) max_parallel_hazard_walker((Node *) parse, &context);
1078 	return context.max_hazard;
1079 }
1080 
1081 /*
1082  * is_parallel_safe
1083  *		Detect whether the given expr contains only parallel-safe functions
1084  *
1085  * root->glob->maxParallelHazard must previously have been set to the
1086  * result of max_parallel_hazard() on the whole query.
1087  */
1088 bool
is_parallel_safe(PlannerInfo * root,Node * node)1089 is_parallel_safe(PlannerInfo *root, Node *node)
1090 {
1091 	max_parallel_hazard_context context;
1092 	PlannerInfo *proot;
1093 	ListCell   *l;
1094 
1095 	/*
1096 	 * Even if the original querytree contained nothing unsafe, we need to
1097 	 * search the expression if we have generated any PARAM_EXEC Params while
1098 	 * planning, because those are parallel-restricted and there might be one
1099 	 * in this expression.  But otherwise we don't need to look.
1100 	 */
1101 	if (root->glob->maxParallelHazard == PROPARALLEL_SAFE &&
1102 		root->glob->paramExecTypes == NIL)
1103 		return true;
1104 	/* Else use max_parallel_hazard's search logic, but stop on RESTRICTED */
1105 	context.max_hazard = PROPARALLEL_SAFE;
1106 	context.max_interesting = PROPARALLEL_RESTRICTED;
1107 	context.safe_param_ids = NIL;
1108 
1109 	/*
1110 	 * The params that refer to the same or parent query level are considered
1111 	 * parallel-safe.  The idea is that we compute such params at Gather or
1112 	 * Gather Merge node and pass their value to workers.
1113 	 */
1114 	for (proot = root; proot != NULL; proot = proot->parent_root)
1115 	{
1116 		foreach(l, proot->init_plans)
1117 		{
1118 			SubPlan    *initsubplan = (SubPlan *) lfirst(l);
1119 			ListCell   *l2;
1120 
1121 			foreach(l2, initsubplan->setParam)
1122 				context.safe_param_ids = lcons_int(lfirst_int(l2),
1123 												   context.safe_param_ids);
1124 		}
1125 	}
1126 
1127 	return !max_parallel_hazard_walker(node, &context);
1128 }
1129 
1130 /* core logic for all parallel-hazard checks */
1131 static bool
max_parallel_hazard_test(char proparallel,max_parallel_hazard_context * context)1132 max_parallel_hazard_test(char proparallel, max_parallel_hazard_context *context)
1133 {
1134 	switch (proparallel)
1135 	{
1136 		case PROPARALLEL_SAFE:
1137 			/* nothing to see here, move along */
1138 			break;
1139 		case PROPARALLEL_RESTRICTED:
1140 			/* increase max_hazard to RESTRICTED */
1141 			Assert(context->max_hazard != PROPARALLEL_UNSAFE);
1142 			context->max_hazard = proparallel;
1143 			/* done if we are not expecting any unsafe functions */
1144 			if (context->max_interesting == proparallel)
1145 				return true;
1146 			break;
1147 		case PROPARALLEL_UNSAFE:
1148 			context->max_hazard = proparallel;
1149 			/* we're always done at the first unsafe construct */
1150 			return true;
1151 		default:
1152 			elog(ERROR, "unrecognized proparallel value \"%c\"", proparallel);
1153 			break;
1154 	}
1155 	return false;
1156 }
1157 
1158 /* check_functions_in_node callback */
1159 static bool
max_parallel_hazard_checker(Oid func_id,void * context)1160 max_parallel_hazard_checker(Oid func_id, void *context)
1161 {
1162 	return max_parallel_hazard_test(func_parallel(func_id),
1163 									(max_parallel_hazard_context *) context);
1164 }
1165 
1166 static bool
max_parallel_hazard_walker(Node * node,max_parallel_hazard_context * context)1167 max_parallel_hazard_walker(Node *node, max_parallel_hazard_context *context)
1168 {
1169 	if (node == NULL)
1170 		return false;
1171 
1172 	/* Check for hazardous functions in node itself */
1173 	if (check_functions_in_node(node, max_parallel_hazard_checker,
1174 								context))
1175 		return true;
1176 
1177 	/*
1178 	 * It should be OK to treat MinMaxExpr as parallel-safe, since btree
1179 	 * opclass support functions are generally parallel-safe.  XmlExpr is a
1180 	 * bit more dubious but we can probably get away with it.  We err on the
1181 	 * side of caution by treating CoerceToDomain as parallel-restricted.
1182 	 * (Note: in principle that's wrong because a domain constraint could
1183 	 * contain a parallel-unsafe function; but useful constraints probably
1184 	 * never would have such, and assuming they do would cripple use of
1185 	 * parallel query in the presence of domain types.)  SQLValueFunction
1186 	 * should be safe in all cases.  NextValueExpr is parallel-unsafe.
1187 	 */
1188 	if (IsA(node, CoerceToDomain))
1189 	{
1190 		if (max_parallel_hazard_test(PROPARALLEL_RESTRICTED, context))
1191 			return true;
1192 	}
1193 
1194 	else if (IsA(node, NextValueExpr))
1195 	{
1196 		if (max_parallel_hazard_test(PROPARALLEL_UNSAFE, context))
1197 			return true;
1198 	}
1199 
1200 	/*
1201 	 * Treat window functions as parallel-restricted because we aren't sure
1202 	 * whether the input row ordering is fully deterministic, and the output
1203 	 * of window functions might vary across workers if not.  (In some cases,
1204 	 * like where the window frame orders by a primary key, we could relax
1205 	 * this restriction.  But it doesn't currently seem worth expending extra
1206 	 * effort to do so.)
1207 	 */
1208 	else if (IsA(node, WindowFunc))
1209 	{
1210 		if (max_parallel_hazard_test(PROPARALLEL_RESTRICTED, context))
1211 			return true;
1212 	}
1213 
1214 	/*
1215 	 * As a notational convenience for callers, look through RestrictInfo.
1216 	 */
1217 	else if (IsA(node, RestrictInfo))
1218 	{
1219 		RestrictInfo *rinfo = (RestrictInfo *) node;
1220 
1221 		return max_parallel_hazard_walker((Node *) rinfo->clause, context);
1222 	}
1223 
1224 	/*
1225 	 * Really we should not see SubLink during a max_interesting == restricted
1226 	 * scan, but if we do, return true.
1227 	 */
1228 	else if (IsA(node, SubLink))
1229 	{
1230 		if (max_parallel_hazard_test(PROPARALLEL_RESTRICTED, context))
1231 			return true;
1232 	}
1233 
1234 	/*
1235 	 * Only parallel-safe SubPlans can be sent to workers.  Within the
1236 	 * testexpr of the SubPlan, Params representing the output columns of the
1237 	 * subplan can be treated as parallel-safe, so temporarily add their IDs
1238 	 * to the safe_param_ids list while examining the testexpr.
1239 	 */
1240 	else if (IsA(node, SubPlan))
1241 	{
1242 		SubPlan    *subplan = (SubPlan *) node;
1243 		List	   *save_safe_param_ids;
1244 
1245 		if (!subplan->parallel_safe &&
1246 			max_parallel_hazard_test(PROPARALLEL_RESTRICTED, context))
1247 			return true;
1248 		save_safe_param_ids = context->safe_param_ids;
1249 		context->safe_param_ids = list_concat(list_copy(subplan->paramIds),
1250 											  context->safe_param_ids);
1251 		if (max_parallel_hazard_walker(subplan->testexpr, context))
1252 			return true;		/* no need to restore safe_param_ids */
1253 		context->safe_param_ids = save_safe_param_ids;
1254 		/* we must also check args, but no special Param treatment there */
1255 		if (max_parallel_hazard_walker((Node *) subplan->args, context))
1256 			return true;
1257 		/* don't want to recurse normally, so we're done */
1258 		return false;
1259 	}
1260 
1261 	/*
1262 	 * We can't pass Params to workers at the moment either, so they are also
1263 	 * parallel-restricted, unless they are PARAM_EXTERN Params or are
1264 	 * PARAM_EXEC Params listed in safe_param_ids, meaning they could be
1265 	 * either generated within the worker or can be computed in master and
1266 	 * then their value can be passed to the worker.
1267 	 */
1268 	else if (IsA(node, Param))
1269 	{
1270 		Param	   *param = (Param *) node;
1271 
1272 		if (param->paramkind == PARAM_EXTERN)
1273 			return false;
1274 
1275 		if (param->paramkind != PARAM_EXEC ||
1276 			!list_member_int(context->safe_param_ids, param->paramid))
1277 		{
1278 			if (max_parallel_hazard_test(PROPARALLEL_RESTRICTED, context))
1279 				return true;
1280 		}
1281 		return false;			/* nothing to recurse to */
1282 	}
1283 
1284 	/*
1285 	 * When we're first invoked on a completely unplanned tree, we must
1286 	 * recurse into subqueries so to as to locate parallel-unsafe constructs
1287 	 * anywhere in the tree.
1288 	 */
1289 	else if (IsA(node, Query))
1290 	{
1291 		Query	   *query = (Query *) node;
1292 
1293 		/* SELECT FOR UPDATE/SHARE must be treated as unsafe */
1294 		if (query->rowMarks != NULL)
1295 		{
1296 			context->max_hazard = PROPARALLEL_UNSAFE;
1297 			return true;
1298 		}
1299 
1300 		/* Recurse into subselects */
1301 		return query_tree_walker(query,
1302 								 max_parallel_hazard_walker,
1303 								 context, 0);
1304 	}
1305 
1306 	/* Recurse to check arguments */
1307 	return expression_tree_walker(node,
1308 								  max_parallel_hazard_walker,
1309 								  context);
1310 }
1311 
1312 
1313 /*****************************************************************************
1314  *		Check clauses for nonstrict functions
1315  *****************************************************************************/
1316 
1317 /*
1318  * contain_nonstrict_functions
1319  *	  Recursively search for nonstrict functions within a clause.
1320  *
1321  * Returns true if any nonstrict construct is found --- ie, anything that
1322  * could produce non-NULL output with a NULL input.
1323  *
1324  * The idea here is that the caller has verified that the expression contains
1325  * one or more Var or Param nodes (as appropriate for the caller's need), and
1326  * now wishes to prove that the expression result will be NULL if any of these
1327  * inputs is NULL.  If we return false, then the proof succeeded.
1328  */
1329 bool
contain_nonstrict_functions(Node * clause)1330 contain_nonstrict_functions(Node *clause)
1331 {
1332 	return contain_nonstrict_functions_walker(clause, NULL);
1333 }
1334 
1335 static bool
contain_nonstrict_functions_checker(Oid func_id,void * context)1336 contain_nonstrict_functions_checker(Oid func_id, void *context)
1337 {
1338 	return !func_strict(func_id);
1339 }
1340 
1341 static bool
contain_nonstrict_functions_walker(Node * node,void * context)1342 contain_nonstrict_functions_walker(Node *node, void *context)
1343 {
1344 	if (node == NULL)
1345 		return false;
1346 	if (IsA(node, Aggref))
1347 	{
1348 		/* an aggregate could return non-null with null input */
1349 		return true;
1350 	}
1351 	if (IsA(node, GroupingFunc))
1352 	{
1353 		/*
1354 		 * A GroupingFunc doesn't evaluate its arguments, and therefore must
1355 		 * be treated as nonstrict.
1356 		 */
1357 		return true;
1358 	}
1359 	if (IsA(node, WindowFunc))
1360 	{
1361 		/* a window function could return non-null with null input */
1362 		return true;
1363 	}
1364 	if (IsA(node, ArrayRef))
1365 	{
1366 		/* array assignment is nonstrict, but subscripting is strict */
1367 		if (((ArrayRef *) node)->refassgnexpr != NULL)
1368 			return true;
1369 		/* else fall through to check args */
1370 	}
1371 	if (IsA(node, DistinctExpr))
1372 	{
1373 		/* IS DISTINCT FROM is inherently non-strict */
1374 		return true;
1375 	}
1376 	if (IsA(node, NullIfExpr))
1377 	{
1378 		/* NULLIF is inherently non-strict */
1379 		return true;
1380 	}
1381 	if (IsA(node, BoolExpr))
1382 	{
1383 		BoolExpr   *expr = (BoolExpr *) node;
1384 
1385 		switch (expr->boolop)
1386 		{
1387 			case AND_EXPR:
1388 			case OR_EXPR:
1389 				/* AND, OR are inherently non-strict */
1390 				return true;
1391 			default:
1392 				break;
1393 		}
1394 	}
1395 	if (IsA(node, SubLink))
1396 	{
1397 		/* In some cases a sublink might be strict, but in general not */
1398 		return true;
1399 	}
1400 	if (IsA(node, SubPlan))
1401 		return true;
1402 	if (IsA(node, AlternativeSubPlan))
1403 		return true;
1404 	if (IsA(node, FieldStore))
1405 		return true;
1406 	if (IsA(node, ArrayCoerceExpr))
1407 	{
1408 		/*
1409 		 * ArrayCoerceExpr is strict at the array level, regardless of what
1410 		 * the per-element expression is; so we should ignore elemexpr and
1411 		 * recurse only into the arg.
1412 		 */
1413 		return contain_nonstrict_functions_walker((Node *) ((ArrayCoerceExpr *) node)->arg,
1414 												  context);
1415 	}
1416 	if (IsA(node, CaseExpr))
1417 		return true;
1418 	if (IsA(node, ArrayExpr))
1419 		return true;
1420 	if (IsA(node, RowExpr))
1421 		return true;
1422 	if (IsA(node, RowCompareExpr))
1423 		return true;
1424 	if (IsA(node, CoalesceExpr))
1425 		return true;
1426 	if (IsA(node, MinMaxExpr))
1427 		return true;
1428 	if (IsA(node, XmlExpr))
1429 		return true;
1430 	if (IsA(node, NullTest))
1431 		return true;
1432 	if (IsA(node, BooleanTest))
1433 		return true;
1434 
1435 	/* Check other function-containing nodes */
1436 	if (check_functions_in_node(node, contain_nonstrict_functions_checker,
1437 								context))
1438 		return true;
1439 
1440 	return expression_tree_walker(node, contain_nonstrict_functions_walker,
1441 								  context);
1442 }
1443 
1444 /*****************************************************************************
1445  *		Check clauses for Params
1446  *****************************************************************************/
1447 
1448 /*
1449  * contain_exec_param
1450  *	  Recursively search for PARAM_EXEC Params within a clause.
1451  *
1452  * Returns true if the clause contains any PARAM_EXEC Param with a paramid
1453  * appearing in the given list of Param IDs.  Does not descend into
1454  * subqueries!
1455  */
1456 bool
contain_exec_param(Node * clause,List * param_ids)1457 contain_exec_param(Node *clause, List *param_ids)
1458 {
1459 	return contain_exec_param_walker(clause, param_ids);
1460 }
1461 
1462 static bool
contain_exec_param_walker(Node * node,List * param_ids)1463 contain_exec_param_walker(Node *node, List *param_ids)
1464 {
1465 	if (node == NULL)
1466 		return false;
1467 	if (IsA(node, Param))
1468 	{
1469 		Param	   *p = (Param *) node;
1470 
1471 		if (p->paramkind == PARAM_EXEC &&
1472 			list_member_int(param_ids, p->paramid))
1473 			return true;
1474 	}
1475 	return expression_tree_walker(node, contain_exec_param_walker, param_ids);
1476 }
1477 
1478 /*****************************************************************************
1479  *		Check clauses for context-dependent nodes
1480  *****************************************************************************/
1481 
1482 /*
1483  * contain_context_dependent_node
1484  *	  Recursively search for context-dependent nodes within a clause.
1485  *
1486  * CaseTestExpr nodes must appear directly within the corresponding CaseExpr,
1487  * not nested within another one, or they'll see the wrong test value.  If one
1488  * appears "bare" in the arguments of a SQL function, then we can't inline the
1489  * SQL function for fear of creating such a situation.  The same applies for
1490  * CaseTestExpr used within the elemexpr of an ArrayCoerceExpr.
1491  *
1492  * CoerceToDomainValue would have the same issue if domain CHECK expressions
1493  * could get inlined into larger expressions, but presently that's impossible.
1494  * Still, it might be allowed in future, or other node types with similar
1495  * issues might get invented.  So give this function a generic name, and set
1496  * up the recursion state to allow multiple flag bits.
1497  */
1498 static bool
contain_context_dependent_node(Node * clause)1499 contain_context_dependent_node(Node *clause)
1500 {
1501 	int			flags = 0;
1502 
1503 	return contain_context_dependent_node_walker(clause, &flags);
1504 }
1505 
1506 #define CCDN_CASETESTEXPR_OK	0x0001	/* CaseTestExpr okay here? */
1507 
1508 static bool
contain_context_dependent_node_walker(Node * node,int * flags)1509 contain_context_dependent_node_walker(Node *node, int *flags)
1510 {
1511 	if (node == NULL)
1512 		return false;
1513 	if (IsA(node, CaseTestExpr))
1514 		return !(*flags & CCDN_CASETESTEXPR_OK);
1515 	else if (IsA(node, CaseExpr))
1516 	{
1517 		CaseExpr   *caseexpr = (CaseExpr *) node;
1518 
1519 		/*
1520 		 * If this CASE doesn't have a test expression, then it doesn't create
1521 		 * a context in which CaseTestExprs should appear, so just fall
1522 		 * through and treat it as a generic expression node.
1523 		 */
1524 		if (caseexpr->arg)
1525 		{
1526 			int			save_flags = *flags;
1527 			bool		res;
1528 
1529 			/*
1530 			 * Note: in principle, we could distinguish the various sub-parts
1531 			 * of a CASE construct and set the flag bit only for some of them,
1532 			 * since we are only expecting CaseTestExprs to appear in the
1533 			 * "expr" subtree of the CaseWhen nodes.  But it doesn't really
1534 			 * seem worth any extra code.  If there are any bare CaseTestExprs
1535 			 * elsewhere in the CASE, something's wrong already.
1536 			 */
1537 			*flags |= CCDN_CASETESTEXPR_OK;
1538 			res = expression_tree_walker(node,
1539 										 contain_context_dependent_node_walker,
1540 										 (void *) flags);
1541 			*flags = save_flags;
1542 			return res;
1543 		}
1544 	}
1545 	else if (IsA(node, ArrayCoerceExpr))
1546 	{
1547 		ArrayCoerceExpr *ac = (ArrayCoerceExpr *) node;
1548 		int			save_flags;
1549 		bool		res;
1550 
1551 		/* Check the array expression */
1552 		if (contain_context_dependent_node_walker((Node *) ac->arg, flags))
1553 			return true;
1554 
1555 		/* Check the elemexpr, which is allowed to contain CaseTestExpr */
1556 		save_flags = *flags;
1557 		*flags |= CCDN_CASETESTEXPR_OK;
1558 		res = contain_context_dependent_node_walker((Node *) ac->elemexpr,
1559 													flags);
1560 		*flags = save_flags;
1561 		return res;
1562 	}
1563 	return expression_tree_walker(node, contain_context_dependent_node_walker,
1564 								  (void *) flags);
1565 }
1566 
1567 /*****************************************************************************
1568  *		  Check clauses for Vars passed to non-leakproof functions
1569  *****************************************************************************/
1570 
1571 /*
1572  * contain_leaked_vars
1573  *		Recursively scan a clause to discover whether it contains any Var
1574  *		nodes (of the current query level) that are passed as arguments to
1575  *		leaky functions.
1576  *
1577  * Returns true if the clause contains any non-leakproof functions that are
1578  * passed Var nodes of the current query level, and which might therefore leak
1579  * data.  Such clauses must be applied after any lower-level security barrier
1580  * clauses.
1581  */
1582 bool
contain_leaked_vars(Node * clause)1583 contain_leaked_vars(Node *clause)
1584 {
1585 	return contain_leaked_vars_walker(clause, NULL);
1586 }
1587 
1588 static bool
contain_leaked_vars_checker(Oid func_id,void * context)1589 contain_leaked_vars_checker(Oid func_id, void *context)
1590 {
1591 	return !get_func_leakproof(func_id);
1592 }
1593 
1594 static bool
contain_leaked_vars_walker(Node * node,void * context)1595 contain_leaked_vars_walker(Node *node, void *context)
1596 {
1597 	if (node == NULL)
1598 		return false;
1599 
1600 	switch (nodeTag(node))
1601 	{
1602 		case T_Var:
1603 		case T_Const:
1604 		case T_Param:
1605 		case T_ArrayExpr:
1606 		case T_FieldSelect:
1607 		case T_FieldStore:
1608 		case T_NamedArgExpr:
1609 		case T_BoolExpr:
1610 		case T_RelabelType:
1611 		case T_CollateExpr:
1612 		case T_CaseExpr:
1613 		case T_CaseTestExpr:
1614 		case T_RowExpr:
1615 		case T_SQLValueFunction:
1616 		case T_NullTest:
1617 		case T_BooleanTest:
1618 		case T_NextValueExpr:
1619 		case T_List:
1620 
1621 			/*
1622 			 * We know these node types don't contain function calls; but
1623 			 * something further down in the node tree might.
1624 			 */
1625 			break;
1626 
1627 		case T_FuncExpr:
1628 		case T_OpExpr:
1629 		case T_DistinctExpr:
1630 		case T_NullIfExpr:
1631 		case T_ScalarArrayOpExpr:
1632 		case T_CoerceViaIO:
1633 		case T_ArrayCoerceExpr:
1634 
1635 			/*
1636 			 * If node contains a leaky function call, and there's any Var
1637 			 * underneath it, reject.
1638 			 */
1639 			if (check_functions_in_node(node, contain_leaked_vars_checker,
1640 										context) &&
1641 				contain_var_clause(node))
1642 				return true;
1643 			break;
1644 
1645 		case T_ArrayRef:
1646 			{
1647 				ArrayRef *aref = (ArrayRef *) node;
1648 
1649 				/*
1650 				 * array assignment is leaky, but subscripted fetches
1651 				 * are not
1652 				 */
1653 				if (aref->refassgnexpr != NULL)
1654 				{
1655 					/* Node is leaky, so reject if it contains Vars */
1656 					if (contain_var_clause(node))
1657 						return true;
1658 				}
1659 			}
1660 			break;
1661 
1662 		case T_RowCompareExpr:
1663 			{
1664 				/*
1665 				 * It's worth special-casing this because a leaky comparison
1666 				 * function only compromises one pair of row elements, which
1667 				 * might not contain Vars while others do.
1668 				 */
1669 				RowCompareExpr *rcexpr = (RowCompareExpr *) node;
1670 				ListCell   *opid;
1671 				ListCell   *larg;
1672 				ListCell   *rarg;
1673 
1674 				forthree(opid, rcexpr->opnos,
1675 						 larg, rcexpr->largs,
1676 						 rarg, rcexpr->rargs)
1677 				{
1678 					Oid			funcid = get_opcode(lfirst_oid(opid));
1679 
1680 					if (!get_func_leakproof(funcid) &&
1681 						(contain_var_clause((Node *) lfirst(larg)) ||
1682 						 contain_var_clause((Node *) lfirst(rarg))))
1683 						return true;
1684 				}
1685 			}
1686 			break;
1687 
1688 		case T_MinMaxExpr:
1689 			{
1690 				/*
1691 				 * MinMaxExpr is leakproof if the comparison function it calls
1692 				 * is leakproof.
1693 				 */
1694 				MinMaxExpr *minmaxexpr = (MinMaxExpr *) node;
1695 				TypeCacheEntry *typentry;
1696 				bool		leakproof;
1697 
1698 				/* Look up the btree comparison function for the datatype */
1699 				typentry = lookup_type_cache(minmaxexpr->minmaxtype,
1700 											 TYPECACHE_CMP_PROC);
1701 				if (OidIsValid(typentry->cmp_proc))
1702 					leakproof = get_func_leakproof(typentry->cmp_proc);
1703 				else
1704 				{
1705 					/*
1706 					 * The executor will throw an error, but here we just
1707 					 * treat the missing function as leaky.
1708 					 */
1709 					leakproof = false;
1710 				}
1711 
1712 				if (!leakproof &&
1713 					contain_var_clause((Node *) minmaxexpr->args))
1714 					return true;
1715 			}
1716 			break;
1717 
1718 		case T_CurrentOfExpr:
1719 
1720 			/*
1721 			 * WHERE CURRENT OF doesn't contain leaky function calls.
1722 			 * Moreover, it is essential that this is considered non-leaky,
1723 			 * since the planner must always generate a TID scan when CURRENT
1724 			 * OF is present -- cf. cost_tidscan.
1725 			 */
1726 			return false;
1727 
1728 		default:
1729 
1730 			/*
1731 			 * If we don't recognize the node tag, assume it might be leaky.
1732 			 * This prevents an unexpected security hole if someone adds a new
1733 			 * node type that can call a function.
1734 			 */
1735 			return true;
1736 	}
1737 	return expression_tree_walker(node, contain_leaked_vars_walker,
1738 								  context);
1739 }
1740 
1741 /*
1742  * find_nonnullable_rels
1743  *		Determine which base rels are forced nonnullable by given clause.
1744  *
1745  * Returns the set of all Relids that are referenced in the clause in such
1746  * a way that the clause cannot possibly return TRUE if any of these Relids
1747  * is an all-NULL row.  (It is OK to err on the side of conservatism; hence
1748  * the analysis here is simplistic.)
1749  *
1750  * The semantics here are subtly different from contain_nonstrict_functions:
1751  * that function is concerned with NULL results from arbitrary expressions,
1752  * but here we assume that the input is a Boolean expression, and wish to
1753  * see if NULL inputs will provably cause a FALSE-or-NULL result.  We expect
1754  * the expression to have been AND/OR flattened and converted to implicit-AND
1755  * format.
1756  *
1757  * Note: this function is largely duplicative of find_nonnullable_vars().
1758  * The reason not to simplify this function into a thin wrapper around
1759  * find_nonnullable_vars() is that the tested conditions really are different:
1760  * a clause like "t1.v1 IS NOT NULL OR t1.v2 IS NOT NULL" does not prove
1761  * that either v1 or v2 can't be NULL, but it does prove that the t1 row
1762  * as a whole can't be all-NULL.
1763  *
1764  * top_level is true while scanning top-level AND/OR structure; here, showing
1765  * the result is either FALSE or NULL is good enough.  top_level is false when
1766  * we have descended below a NOT or a strict function: now we must be able to
1767  * prove that the subexpression goes to NULL.
1768  *
1769  * We don't use expression_tree_walker here because we don't want to descend
1770  * through very many kinds of nodes; only the ones we can be sure are strict.
1771  */
1772 Relids
find_nonnullable_rels(Node * clause)1773 find_nonnullable_rels(Node *clause)
1774 {
1775 	return find_nonnullable_rels_walker(clause, true);
1776 }
1777 
1778 static Relids
find_nonnullable_rels_walker(Node * node,bool top_level)1779 find_nonnullable_rels_walker(Node *node, bool top_level)
1780 {
1781 	Relids		result = NULL;
1782 	ListCell   *l;
1783 
1784 	if (node == NULL)
1785 		return NULL;
1786 	if (IsA(node, Var))
1787 	{
1788 		Var		   *var = (Var *) node;
1789 
1790 		if (var->varlevelsup == 0)
1791 			result = bms_make_singleton(var->varno);
1792 	}
1793 	else if (IsA(node, List))
1794 	{
1795 		/*
1796 		 * At top level, we are examining an implicit-AND list: if any of the
1797 		 * arms produces FALSE-or-NULL then the result is FALSE-or-NULL. If
1798 		 * not at top level, we are examining the arguments of a strict
1799 		 * function: if any of them produce NULL then the result of the
1800 		 * function must be NULL.  So in both cases, the set of nonnullable
1801 		 * rels is the union of those found in the arms, and we pass down the
1802 		 * top_level flag unmodified.
1803 		 */
1804 		foreach(l, (List *) node)
1805 		{
1806 			result = bms_join(result,
1807 							  find_nonnullable_rels_walker(lfirst(l),
1808 														   top_level));
1809 		}
1810 	}
1811 	else if (IsA(node, FuncExpr))
1812 	{
1813 		FuncExpr   *expr = (FuncExpr *) node;
1814 
1815 		if (func_strict(expr->funcid))
1816 			result = find_nonnullable_rels_walker((Node *) expr->args, false);
1817 	}
1818 	else if (IsA(node, OpExpr))
1819 	{
1820 		OpExpr	   *expr = (OpExpr *) node;
1821 
1822 		set_opfuncid(expr);
1823 		if (func_strict(expr->opfuncid))
1824 			result = find_nonnullable_rels_walker((Node *) expr->args, false);
1825 	}
1826 	else if (IsA(node, ScalarArrayOpExpr))
1827 	{
1828 		ScalarArrayOpExpr *expr = (ScalarArrayOpExpr *) node;
1829 
1830 		if (is_strict_saop(expr, true))
1831 			result = find_nonnullable_rels_walker((Node *) expr->args, false);
1832 	}
1833 	else if (IsA(node, BoolExpr))
1834 	{
1835 		BoolExpr   *expr = (BoolExpr *) node;
1836 
1837 		switch (expr->boolop)
1838 		{
1839 			case AND_EXPR:
1840 				/* At top level we can just recurse (to the List case) */
1841 				if (top_level)
1842 				{
1843 					result = find_nonnullable_rels_walker((Node *) expr->args,
1844 														  top_level);
1845 					break;
1846 				}
1847 
1848 				/*
1849 				 * Below top level, even if one arm produces NULL, the result
1850 				 * could be FALSE (hence not NULL).  However, if *all* the
1851 				 * arms produce NULL then the result is NULL, so we can take
1852 				 * the intersection of the sets of nonnullable rels, just as
1853 				 * for OR.  Fall through to share code.
1854 				 */
1855 				/* FALL THRU */
1856 			case OR_EXPR:
1857 
1858 				/*
1859 				 * OR is strict if all of its arms are, so we can take the
1860 				 * intersection of the sets of nonnullable rels for each arm.
1861 				 * This works for both values of top_level.
1862 				 */
1863 				foreach(l, expr->args)
1864 				{
1865 					Relids		subresult;
1866 
1867 					subresult = find_nonnullable_rels_walker(lfirst(l),
1868 															 top_level);
1869 					if (result == NULL) /* first subresult? */
1870 						result = subresult;
1871 					else
1872 						result = bms_int_members(result, subresult);
1873 
1874 					/*
1875 					 * If the intersection is empty, we can stop looking. This
1876 					 * also justifies the test for first-subresult above.
1877 					 */
1878 					if (bms_is_empty(result))
1879 						break;
1880 				}
1881 				break;
1882 			case NOT_EXPR:
1883 				/* NOT will return null if its arg is null */
1884 				result = find_nonnullable_rels_walker((Node *) expr->args,
1885 													  false);
1886 				break;
1887 			default:
1888 				elog(ERROR, "unrecognized boolop: %d", (int) expr->boolop);
1889 				break;
1890 		}
1891 	}
1892 	else if (IsA(node, RelabelType))
1893 	{
1894 		RelabelType *expr = (RelabelType *) node;
1895 
1896 		result = find_nonnullable_rels_walker((Node *) expr->arg, top_level);
1897 	}
1898 	else if (IsA(node, CoerceViaIO))
1899 	{
1900 		/* not clear this is useful, but it can't hurt */
1901 		CoerceViaIO *expr = (CoerceViaIO *) node;
1902 
1903 		result = find_nonnullable_rels_walker((Node *) expr->arg, top_level);
1904 	}
1905 	else if (IsA(node, ArrayCoerceExpr))
1906 	{
1907 		/* ArrayCoerceExpr is strict at the array level; ignore elemexpr */
1908 		ArrayCoerceExpr *expr = (ArrayCoerceExpr *) node;
1909 
1910 		result = find_nonnullable_rels_walker((Node *) expr->arg, top_level);
1911 	}
1912 	else if (IsA(node, ConvertRowtypeExpr))
1913 	{
1914 		/* not clear this is useful, but it can't hurt */
1915 		ConvertRowtypeExpr *expr = (ConvertRowtypeExpr *) node;
1916 
1917 		result = find_nonnullable_rels_walker((Node *) expr->arg, top_level);
1918 	}
1919 	else if (IsA(node, CollateExpr))
1920 	{
1921 		CollateExpr *expr = (CollateExpr *) node;
1922 
1923 		result = find_nonnullable_rels_walker((Node *) expr->arg, top_level);
1924 	}
1925 	else if (IsA(node, NullTest))
1926 	{
1927 		/* IS NOT NULL can be considered strict, but only at top level */
1928 		NullTest   *expr = (NullTest *) node;
1929 
1930 		if (top_level && expr->nulltesttype == IS_NOT_NULL && !expr->argisrow)
1931 			result = find_nonnullable_rels_walker((Node *) expr->arg, false);
1932 	}
1933 	else if (IsA(node, BooleanTest))
1934 	{
1935 		/* Boolean tests that reject NULL are strict at top level */
1936 		BooleanTest *expr = (BooleanTest *) node;
1937 
1938 		if (top_level &&
1939 			(expr->booltesttype == IS_TRUE ||
1940 			 expr->booltesttype == IS_FALSE ||
1941 			 expr->booltesttype == IS_NOT_UNKNOWN))
1942 			result = find_nonnullable_rels_walker((Node *) expr->arg, false);
1943 	}
1944 	else if (IsA(node, PlaceHolderVar))
1945 	{
1946 		PlaceHolderVar *phv = (PlaceHolderVar *) node;
1947 
1948 		result = find_nonnullable_rels_walker((Node *) phv->phexpr, top_level);
1949 	}
1950 	return result;
1951 }
1952 
1953 /*
1954  * find_nonnullable_vars
1955  *		Determine which Vars are forced nonnullable by given clause.
1956  *
1957  * Returns a list of all level-zero Vars that are referenced in the clause in
1958  * such a way that the clause cannot possibly return TRUE if any of these Vars
1959  * is NULL.  (It is OK to err on the side of conservatism; hence the analysis
1960  * here is simplistic.)
1961  *
1962  * The semantics here are subtly different from contain_nonstrict_functions:
1963  * that function is concerned with NULL results from arbitrary expressions,
1964  * but here we assume that the input is a Boolean expression, and wish to
1965  * see if NULL inputs will provably cause a FALSE-or-NULL result.  We expect
1966  * the expression to have been AND/OR flattened and converted to implicit-AND
1967  * format.
1968  *
1969  * The result is a palloc'd List, but we have not copied the member Var nodes.
1970  * Also, we don't bother trying to eliminate duplicate entries.
1971  *
1972  * top_level is true while scanning top-level AND/OR structure; here, showing
1973  * the result is either FALSE or NULL is good enough.  top_level is false when
1974  * we have descended below a NOT or a strict function: now we must be able to
1975  * prove that the subexpression goes to NULL.
1976  *
1977  * We don't use expression_tree_walker here because we don't want to descend
1978  * through very many kinds of nodes; only the ones we can be sure are strict.
1979  */
1980 List *
find_nonnullable_vars(Node * clause)1981 find_nonnullable_vars(Node *clause)
1982 {
1983 	return find_nonnullable_vars_walker(clause, true);
1984 }
1985 
1986 static List *
find_nonnullable_vars_walker(Node * node,bool top_level)1987 find_nonnullable_vars_walker(Node *node, bool top_level)
1988 {
1989 	List	   *result = NIL;
1990 	ListCell   *l;
1991 
1992 	if (node == NULL)
1993 		return NIL;
1994 	if (IsA(node, Var))
1995 	{
1996 		Var		   *var = (Var *) node;
1997 
1998 		if (var->varlevelsup == 0)
1999 			result = list_make1(var);
2000 	}
2001 	else if (IsA(node, List))
2002 	{
2003 		/*
2004 		 * At top level, we are examining an implicit-AND list: if any of the
2005 		 * arms produces FALSE-or-NULL then the result is FALSE-or-NULL. If
2006 		 * not at top level, we are examining the arguments of a strict
2007 		 * function: if any of them produce NULL then the result of the
2008 		 * function must be NULL.  So in both cases, the set of nonnullable
2009 		 * vars is the union of those found in the arms, and we pass down the
2010 		 * top_level flag unmodified.
2011 		 */
2012 		foreach(l, (List *) node)
2013 		{
2014 			result = list_concat(result,
2015 								 find_nonnullable_vars_walker(lfirst(l),
2016 															  top_level));
2017 		}
2018 	}
2019 	else if (IsA(node, FuncExpr))
2020 	{
2021 		FuncExpr   *expr = (FuncExpr *) node;
2022 
2023 		if (func_strict(expr->funcid))
2024 			result = find_nonnullable_vars_walker((Node *) expr->args, false);
2025 	}
2026 	else if (IsA(node, OpExpr))
2027 	{
2028 		OpExpr	   *expr = (OpExpr *) node;
2029 
2030 		set_opfuncid(expr);
2031 		if (func_strict(expr->opfuncid))
2032 			result = find_nonnullable_vars_walker((Node *) expr->args, false);
2033 	}
2034 	else if (IsA(node, ScalarArrayOpExpr))
2035 	{
2036 		ScalarArrayOpExpr *expr = (ScalarArrayOpExpr *) node;
2037 
2038 		if (is_strict_saop(expr, true))
2039 			result = find_nonnullable_vars_walker((Node *) expr->args, false);
2040 	}
2041 	else if (IsA(node, BoolExpr))
2042 	{
2043 		BoolExpr   *expr = (BoolExpr *) node;
2044 
2045 		switch (expr->boolop)
2046 		{
2047 			case AND_EXPR:
2048 				/* At top level we can just recurse (to the List case) */
2049 				if (top_level)
2050 				{
2051 					result = find_nonnullable_vars_walker((Node *) expr->args,
2052 														  top_level);
2053 					break;
2054 				}
2055 
2056 				/*
2057 				 * Below top level, even if one arm produces NULL, the result
2058 				 * could be FALSE (hence not NULL).  However, if *all* the
2059 				 * arms produce NULL then the result is NULL, so we can take
2060 				 * the intersection of the sets of nonnullable vars, just as
2061 				 * for OR.  Fall through to share code.
2062 				 */
2063 				/* FALL THRU */
2064 			case OR_EXPR:
2065 
2066 				/*
2067 				 * OR is strict if all of its arms are, so we can take the
2068 				 * intersection of the sets of nonnullable vars for each arm.
2069 				 * This works for both values of top_level.
2070 				 */
2071 				foreach(l, expr->args)
2072 				{
2073 					List	   *subresult;
2074 
2075 					subresult = find_nonnullable_vars_walker(lfirst(l),
2076 															 top_level);
2077 					if (result == NIL)	/* first subresult? */
2078 						result = subresult;
2079 					else
2080 						result = list_intersection(result, subresult);
2081 
2082 					/*
2083 					 * If the intersection is empty, we can stop looking. This
2084 					 * also justifies the test for first-subresult above.
2085 					 */
2086 					if (result == NIL)
2087 						break;
2088 				}
2089 				break;
2090 			case NOT_EXPR:
2091 				/* NOT will return null if its arg is null */
2092 				result = find_nonnullable_vars_walker((Node *) expr->args,
2093 													  false);
2094 				break;
2095 			default:
2096 				elog(ERROR, "unrecognized boolop: %d", (int) expr->boolop);
2097 				break;
2098 		}
2099 	}
2100 	else if (IsA(node, RelabelType))
2101 	{
2102 		RelabelType *expr = (RelabelType *) node;
2103 
2104 		result = find_nonnullable_vars_walker((Node *) expr->arg, top_level);
2105 	}
2106 	else if (IsA(node, CoerceViaIO))
2107 	{
2108 		/* not clear this is useful, but it can't hurt */
2109 		CoerceViaIO *expr = (CoerceViaIO *) node;
2110 
2111 		result = find_nonnullable_vars_walker((Node *) expr->arg, false);
2112 	}
2113 	else if (IsA(node, ArrayCoerceExpr))
2114 	{
2115 		/* ArrayCoerceExpr is strict at the array level; ignore elemexpr */
2116 		ArrayCoerceExpr *expr = (ArrayCoerceExpr *) node;
2117 
2118 		result = find_nonnullable_vars_walker((Node *) expr->arg, top_level);
2119 	}
2120 	else if (IsA(node, ConvertRowtypeExpr))
2121 	{
2122 		/* not clear this is useful, but it can't hurt */
2123 		ConvertRowtypeExpr *expr = (ConvertRowtypeExpr *) node;
2124 
2125 		result = find_nonnullable_vars_walker((Node *) expr->arg, top_level);
2126 	}
2127 	else if (IsA(node, CollateExpr))
2128 	{
2129 		CollateExpr *expr = (CollateExpr *) node;
2130 
2131 		result = find_nonnullable_vars_walker((Node *) expr->arg, top_level);
2132 	}
2133 	else if (IsA(node, NullTest))
2134 	{
2135 		/* IS NOT NULL can be considered strict, but only at top level */
2136 		NullTest   *expr = (NullTest *) node;
2137 
2138 		if (top_level && expr->nulltesttype == IS_NOT_NULL && !expr->argisrow)
2139 			result = find_nonnullable_vars_walker((Node *) expr->arg, false);
2140 	}
2141 	else if (IsA(node, BooleanTest))
2142 	{
2143 		/* Boolean tests that reject NULL are strict at top level */
2144 		BooleanTest *expr = (BooleanTest *) node;
2145 
2146 		if (top_level &&
2147 			(expr->booltesttype == IS_TRUE ||
2148 			 expr->booltesttype == IS_FALSE ||
2149 			 expr->booltesttype == IS_NOT_UNKNOWN))
2150 			result = find_nonnullable_vars_walker((Node *) expr->arg, false);
2151 	}
2152 	else if (IsA(node, PlaceHolderVar))
2153 	{
2154 		PlaceHolderVar *phv = (PlaceHolderVar *) node;
2155 
2156 		result = find_nonnullable_vars_walker((Node *) phv->phexpr, top_level);
2157 	}
2158 	return result;
2159 }
2160 
2161 /*
2162  * find_forced_null_vars
2163  *		Determine which Vars must be NULL for the given clause to return TRUE.
2164  *
2165  * This is the complement of find_nonnullable_vars: find the level-zero Vars
2166  * that must be NULL for the clause to return TRUE.  (It is OK to err on the
2167  * side of conservatism; hence the analysis here is simplistic.  In fact,
2168  * we only detect simple "var IS NULL" tests at the top level.)
2169  *
2170  * The result is a palloc'd List, but we have not copied the member Var nodes.
2171  * Also, we don't bother trying to eliminate duplicate entries.
2172  */
2173 List *
find_forced_null_vars(Node * node)2174 find_forced_null_vars(Node *node)
2175 {
2176 	List	   *result = NIL;
2177 	Var		   *var;
2178 	ListCell   *l;
2179 
2180 	if (node == NULL)
2181 		return NIL;
2182 	/* Check single-clause cases using subroutine */
2183 	var = find_forced_null_var(node);
2184 	if (var)
2185 	{
2186 		result = list_make1(var);
2187 	}
2188 	/* Otherwise, handle AND-conditions */
2189 	else if (IsA(node, List))
2190 	{
2191 		/*
2192 		 * At top level, we are examining an implicit-AND list: if any of the
2193 		 * arms produces FALSE-or-NULL then the result is FALSE-or-NULL.
2194 		 */
2195 		foreach(l, (List *) node)
2196 		{
2197 			result = list_concat(result,
2198 								 find_forced_null_vars(lfirst(l)));
2199 		}
2200 	}
2201 	else if (IsA(node, BoolExpr))
2202 	{
2203 		BoolExpr   *expr = (BoolExpr *) node;
2204 
2205 		/*
2206 		 * We don't bother considering the OR case, because it's fairly
2207 		 * unlikely anyone would write "v1 IS NULL OR v1 IS NULL". Likewise,
2208 		 * the NOT case isn't worth expending code on.
2209 		 */
2210 		if (expr->boolop == AND_EXPR)
2211 		{
2212 			/* At top level we can just recurse (to the List case) */
2213 			result = find_forced_null_vars((Node *) expr->args);
2214 		}
2215 	}
2216 	return result;
2217 }
2218 
2219 /*
2220  * find_forced_null_var
2221  *		Return the Var forced null by the given clause, or NULL if it's
2222  *		not an IS NULL-type clause.  For success, the clause must enforce
2223  *		*only* nullness of the particular Var, not any other conditions.
2224  *
2225  * This is just the single-clause case of find_forced_null_vars(), without
2226  * any allowance for AND conditions.  It's used by initsplan.c on individual
2227  * qual clauses.  The reason for not just applying find_forced_null_vars()
2228  * is that if an AND of an IS NULL clause with something else were to somehow
2229  * survive AND/OR flattening, initsplan.c might get fooled into discarding
2230  * the whole clause when only the IS NULL part of it had been proved redundant.
2231  */
2232 Var *
find_forced_null_var(Node * node)2233 find_forced_null_var(Node *node)
2234 {
2235 	if (node == NULL)
2236 		return NULL;
2237 	if (IsA(node, NullTest))
2238 	{
2239 		/* check for var IS NULL */
2240 		NullTest   *expr = (NullTest *) node;
2241 
2242 		if (expr->nulltesttype == IS_NULL && !expr->argisrow)
2243 		{
2244 			Var		   *var = (Var *) expr->arg;
2245 
2246 			if (var && IsA(var, Var) &&
2247 				var->varlevelsup == 0)
2248 				return var;
2249 		}
2250 	}
2251 	else if (IsA(node, BooleanTest))
2252 	{
2253 		/* var IS UNKNOWN is equivalent to var IS NULL */
2254 		BooleanTest *expr = (BooleanTest *) node;
2255 
2256 		if (expr->booltesttype == IS_UNKNOWN)
2257 		{
2258 			Var		   *var = (Var *) expr->arg;
2259 
2260 			if (var && IsA(var, Var) &&
2261 				var->varlevelsup == 0)
2262 				return var;
2263 		}
2264 	}
2265 	return NULL;
2266 }
2267 
2268 /*
2269  * Can we treat a ScalarArrayOpExpr as strict?
2270  *
2271  * If "falseOK" is true, then a "false" result can be considered strict,
2272  * else we need to guarantee an actual NULL result for NULL input.
2273  *
2274  * "foo op ALL array" is strict if the op is strict *and* we can prove
2275  * that the array input isn't an empty array.  We can check that
2276  * for the cases of an array constant and an ARRAY[] construct.
2277  *
2278  * "foo op ANY array" is strict in the falseOK sense if the op is strict.
2279  * If not falseOK, the test is the same as for "foo op ALL array".
2280  */
2281 static bool
is_strict_saop(ScalarArrayOpExpr * expr,bool falseOK)2282 is_strict_saop(ScalarArrayOpExpr *expr, bool falseOK)
2283 {
2284 	Node	   *rightop;
2285 
2286 	/* The contained operator must be strict. */
2287 	set_sa_opfuncid(expr);
2288 	if (!func_strict(expr->opfuncid))
2289 		return false;
2290 	/* If ANY and falseOK, that's all we need to check. */
2291 	if (expr->useOr && falseOK)
2292 		return true;
2293 	/* Else, we have to see if the array is provably non-empty. */
2294 	Assert(list_length(expr->args) == 2);
2295 	rightop = (Node *) lsecond(expr->args);
2296 	if (rightop && IsA(rightop, Const))
2297 	{
2298 		Datum		arraydatum = ((Const *) rightop)->constvalue;
2299 		bool		arrayisnull = ((Const *) rightop)->constisnull;
2300 		ArrayType  *arrayval;
2301 		int			nitems;
2302 
2303 		if (arrayisnull)
2304 			return false;
2305 		arrayval = DatumGetArrayTypeP(arraydatum);
2306 		nitems = ArrayGetNItems(ARR_NDIM(arrayval), ARR_DIMS(arrayval));
2307 		if (nitems > 0)
2308 			return true;
2309 	}
2310 	else if (rightop && IsA(rightop, ArrayExpr))
2311 	{
2312 		ArrayExpr  *arrayexpr = (ArrayExpr *) rightop;
2313 
2314 		if (arrayexpr->elements != NIL && !arrayexpr->multidims)
2315 			return true;
2316 	}
2317 	return false;
2318 }
2319 
2320 
2321 /*****************************************************************************
2322  *		Check for "pseudo-constant" clauses
2323  *****************************************************************************/
2324 
2325 /*
2326  * is_pseudo_constant_clause
2327  *	  Detect whether an expression is "pseudo constant", ie, it contains no
2328  *	  variables of the current query level and no uses of volatile functions.
2329  *	  Such an expr is not necessarily a true constant: it can still contain
2330  *	  Params and outer-level Vars, not to mention functions whose results
2331  *	  may vary from one statement to the next.  However, the expr's value
2332  *	  will be constant over any one scan of the current query, so it can be
2333  *	  used as, eg, an indexscan key.
2334  *
2335  * CAUTION: this function omits to test for one very important class of
2336  * not-constant expressions, namely aggregates (Aggrefs).  In current usage
2337  * this is only applied to WHERE clauses and so a check for Aggrefs would be
2338  * a waste of cycles; but be sure to also check contain_agg_clause() if you
2339  * want to know about pseudo-constness in other contexts.  The same goes
2340  * for window functions (WindowFuncs).
2341  */
2342 bool
is_pseudo_constant_clause(Node * clause)2343 is_pseudo_constant_clause(Node *clause)
2344 {
2345 	/*
2346 	 * We could implement this check in one recursive scan.  But since the
2347 	 * check for volatile functions is both moderately expensive and unlikely
2348 	 * to fail, it seems better to look for Vars first and only check for
2349 	 * volatile functions if we find no Vars.
2350 	 */
2351 	if (!contain_var_clause(clause) &&
2352 		!contain_volatile_functions(clause))
2353 		return true;
2354 	return false;
2355 }
2356 
2357 /*
2358  * is_pseudo_constant_clause_relids
2359  *	  Same as above, except caller already has available the var membership
2360  *	  of the expression; this lets us avoid the contain_var_clause() scan.
2361  */
2362 bool
is_pseudo_constant_clause_relids(Node * clause,Relids relids)2363 is_pseudo_constant_clause_relids(Node *clause, Relids relids)
2364 {
2365 	if (bms_is_empty(relids) &&
2366 		!contain_volatile_functions(clause))
2367 		return true;
2368 	return false;
2369 }
2370 
2371 
2372 /*****************************************************************************
2373  *																			 *
2374  *		General clause-manipulating routines								 *
2375  *																			 *
2376  *****************************************************************************/
2377 
2378 /*
2379  * NumRelids
2380  *		(formerly clause_relids)
2381  *
2382  * Returns the number of different relations referenced in 'clause'.
2383  */
2384 int
NumRelids(Node * clause)2385 NumRelids(Node *clause)
2386 {
2387 	Relids		varnos = pull_varnos(clause);
2388 	int			result = bms_num_members(varnos);
2389 
2390 	bms_free(varnos);
2391 	return result;
2392 }
2393 
2394 /*
2395  * CommuteOpExpr: commute a binary operator clause
2396  *
2397  * XXX the clause is destructively modified!
2398  */
2399 void
CommuteOpExpr(OpExpr * clause)2400 CommuteOpExpr(OpExpr *clause)
2401 {
2402 	Oid			opoid;
2403 	Node	   *temp;
2404 
2405 	/* Sanity checks: caller is at fault if these fail */
2406 	if (!is_opclause(clause) ||
2407 		list_length(clause->args) != 2)
2408 		elog(ERROR, "cannot commute non-binary-operator clause");
2409 
2410 	opoid = get_commutator(clause->opno);
2411 
2412 	if (!OidIsValid(opoid))
2413 		elog(ERROR, "could not find commutator for operator %u",
2414 			 clause->opno);
2415 
2416 	/*
2417 	 * modify the clause in-place!
2418 	 */
2419 	clause->opno = opoid;
2420 	clause->opfuncid = InvalidOid;
2421 	/* opresulttype, opretset, opcollid, inputcollid need not change */
2422 
2423 	temp = linitial(clause->args);
2424 	linitial(clause->args) = lsecond(clause->args);
2425 	lsecond(clause->args) = temp;
2426 }
2427 
2428 /*
2429  * CommuteRowCompareExpr: commute a RowCompareExpr clause
2430  *
2431  * XXX the clause is destructively modified!
2432  */
2433 void
CommuteRowCompareExpr(RowCompareExpr * clause)2434 CommuteRowCompareExpr(RowCompareExpr *clause)
2435 {
2436 	List	   *newops;
2437 	List	   *temp;
2438 	ListCell   *l;
2439 
2440 	/* Sanity checks: caller is at fault if these fail */
2441 	if (!IsA(clause, RowCompareExpr))
2442 		elog(ERROR, "expected a RowCompareExpr");
2443 
2444 	/* Build list of commuted operators */
2445 	newops = NIL;
2446 	foreach(l, clause->opnos)
2447 	{
2448 		Oid			opoid = lfirst_oid(l);
2449 
2450 		opoid = get_commutator(opoid);
2451 		if (!OidIsValid(opoid))
2452 			elog(ERROR, "could not find commutator for operator %u",
2453 				 lfirst_oid(l));
2454 		newops = lappend_oid(newops, opoid);
2455 	}
2456 
2457 	/*
2458 	 * modify the clause in-place!
2459 	 */
2460 	switch (clause->rctype)
2461 	{
2462 		case ROWCOMPARE_LT:
2463 			clause->rctype = ROWCOMPARE_GT;
2464 			break;
2465 		case ROWCOMPARE_LE:
2466 			clause->rctype = ROWCOMPARE_GE;
2467 			break;
2468 		case ROWCOMPARE_GE:
2469 			clause->rctype = ROWCOMPARE_LE;
2470 			break;
2471 		case ROWCOMPARE_GT:
2472 			clause->rctype = ROWCOMPARE_LT;
2473 			break;
2474 		default:
2475 			elog(ERROR, "unexpected RowCompare type: %d",
2476 				 (int) clause->rctype);
2477 			break;
2478 	}
2479 
2480 	clause->opnos = newops;
2481 
2482 	/*
2483 	 * Note: we need not change the opfamilies list; we assume any btree
2484 	 * opfamily containing an operator will also contain its commutator.
2485 	 * Collations don't change either.
2486 	 */
2487 
2488 	temp = clause->largs;
2489 	clause->largs = clause->rargs;
2490 	clause->rargs = temp;
2491 }
2492 
2493 /*
2494  * Helper for eval_const_expressions: check that datatype of an attribute
2495  * is still what it was when the expression was parsed.  This is needed to
2496  * guard against improper simplification after ALTER COLUMN TYPE.  (XXX we
2497  * may well need to make similar checks elsewhere?)
2498  *
2499  * rowtypeid may come from a whole-row Var, and therefore it can be a domain
2500  * over composite, but for this purpose we only care about checking the type
2501  * of a contained field.
2502  */
2503 static bool
rowtype_field_matches(Oid rowtypeid,int fieldnum,Oid expectedtype,int32 expectedtypmod,Oid expectedcollation)2504 rowtype_field_matches(Oid rowtypeid, int fieldnum,
2505 					  Oid expectedtype, int32 expectedtypmod,
2506 					  Oid expectedcollation)
2507 {
2508 	TupleDesc	tupdesc;
2509 	Form_pg_attribute attr;
2510 
2511 	/* No issue for RECORD, since there is no way to ALTER such a type */
2512 	if (rowtypeid == RECORDOID)
2513 		return true;
2514 	tupdesc = lookup_rowtype_tupdesc_domain(rowtypeid, -1, false);
2515 	if (fieldnum <= 0 || fieldnum > tupdesc->natts)
2516 	{
2517 		ReleaseTupleDesc(tupdesc);
2518 		return false;
2519 	}
2520 	attr = TupleDescAttr(tupdesc, fieldnum - 1);
2521 	if (attr->attisdropped ||
2522 		attr->atttypid != expectedtype ||
2523 		attr->atttypmod != expectedtypmod ||
2524 		attr->attcollation != expectedcollation)
2525 	{
2526 		ReleaseTupleDesc(tupdesc);
2527 		return false;
2528 	}
2529 	ReleaseTupleDesc(tupdesc);
2530 	return true;
2531 }
2532 
2533 
2534 /*--------------------
2535  * eval_const_expressions
2536  *
2537  * Reduce any recognizably constant subexpressions of the given
2538  * expression tree, for example "2 + 2" => "4".  More interestingly,
2539  * we can reduce certain boolean expressions even when they contain
2540  * non-constant subexpressions: "x OR true" => "true" no matter what
2541  * the subexpression x is.  (XXX We assume that no such subexpression
2542  * will have important side-effects, which is not necessarily a good
2543  * assumption in the presence of user-defined functions; do we need a
2544  * pg_proc flag that prevents discarding the execution of a function?)
2545  *
2546  * We do understand that certain functions may deliver non-constant
2547  * results even with constant inputs, "nextval()" being the classic
2548  * example.  Functions that are not marked "immutable" in pg_proc
2549  * will not be pre-evaluated here, although we will reduce their
2550  * arguments as far as possible.
2551  *
2552  * Whenever a function is eliminated from the expression by means of
2553  * constant-expression evaluation or inlining, we add the function to
2554  * root->glob->invalItems.  This ensures the plan is known to depend on
2555  * such functions, even though they aren't referenced anymore.
2556  *
2557  * We assume that the tree has already been type-checked and contains
2558  * only operators and functions that are reasonable to try to execute.
2559  *
2560  * NOTE: "root" can be passed as NULL if the caller never wants to do any
2561  * Param substitutions nor receive info about inlined functions.
2562  *
2563  * NOTE: the planner assumes that this will always flatten nested AND and
2564  * OR clauses into N-argument form.  See comments in prepqual.c.
2565  *
2566  * NOTE: another critical effect is that any function calls that require
2567  * default arguments will be expanded, and named-argument calls will be
2568  * converted to positional notation.  The executor won't handle either.
2569  *--------------------
2570  */
2571 Node *
eval_const_expressions(PlannerInfo * root,Node * node)2572 eval_const_expressions(PlannerInfo *root, Node *node)
2573 {
2574 	eval_const_expressions_context context;
2575 
2576 	if (root)
2577 		context.boundParams = root->glob->boundParams;	/* bound Params */
2578 	else
2579 		context.boundParams = NULL;
2580 	context.root = root;		/* for inlined-function dependencies */
2581 	context.active_fns = NIL;	/* nothing being recursively simplified */
2582 	context.case_val = NULL;	/* no CASE being examined */
2583 	context.estimate = false;	/* safe transformations only */
2584 	return eval_const_expressions_mutator(node, &context);
2585 }
2586 
2587 /*--------------------
2588  * estimate_expression_value
2589  *
2590  * This function attempts to estimate the value of an expression for
2591  * planning purposes.  It is in essence a more aggressive version of
2592  * eval_const_expressions(): we will perform constant reductions that are
2593  * not necessarily 100% safe, but are reasonable for estimation purposes.
2594  *
2595  * Currently the extra steps that are taken in this mode are:
2596  * 1. Substitute values for Params, where a bound Param value has been made
2597  *	  available by the caller of planner(), even if the Param isn't marked
2598  *	  constant.  This effectively means that we plan using the first supplied
2599  *	  value of the Param.
2600  * 2. Fold stable, as well as immutable, functions to constants.
2601  * 3. Reduce PlaceHolderVar nodes to their contained expressions.
2602  *--------------------
2603  */
2604 Node *
estimate_expression_value(PlannerInfo * root,Node * node)2605 estimate_expression_value(PlannerInfo *root, Node *node)
2606 {
2607 	eval_const_expressions_context context;
2608 
2609 	context.boundParams = root->glob->boundParams;	/* bound Params */
2610 	/* we do not need to mark the plan as depending on inlined functions */
2611 	context.root = NULL;
2612 	context.active_fns = NIL;	/* nothing being recursively simplified */
2613 	context.case_val = NULL;	/* no CASE being examined */
2614 	context.estimate = true;	/* unsafe transformations OK */
2615 	return eval_const_expressions_mutator(node, &context);
2616 }
2617 
2618 /*
2619  * The generic case in eval_const_expressions_mutator is to recurse using
2620  * expression_tree_mutator, which will copy the given node unchanged but
2621  * const-simplify its arguments (if any) as far as possible.  If the node
2622  * itself does immutable processing, and each of its arguments were reduced
2623  * to a Const, we can then reduce it to a Const using evaluate_expr.  (Some
2624  * node types need more complicated logic; for example, a CASE expression
2625  * might be reducible to a constant even if not all its subtrees are.)
2626  */
2627 #define ece_generic_processing(node) \
2628 	expression_tree_mutator((Node *) (node), eval_const_expressions_mutator, \
2629 							(void *) context)
2630 
2631 /*
2632  * Check whether all arguments of the given node were reduced to Consts.
2633  * By going directly to expression_tree_walker, contain_non_const_walker
2634  * is not applied to the node itself, only to its children.
2635  */
2636 #define ece_all_arguments_const(node) \
2637 	(!expression_tree_walker((Node *) (node), contain_non_const_walker, NULL))
2638 
2639 /* Generic macro for applying evaluate_expr */
2640 #define ece_evaluate_expr(node) \
2641 	((Node *) evaluate_expr((Expr *) (node), \
2642 							exprType((Node *) (node)), \
2643 							exprTypmod((Node *) (node)), \
2644 							exprCollation((Node *) (node))))
2645 
2646 /*
2647  * Recursive guts of eval_const_expressions/estimate_expression_value
2648  */
2649 static Node *
eval_const_expressions_mutator(Node * node,eval_const_expressions_context * context)2650 eval_const_expressions_mutator(Node *node,
2651 							   eval_const_expressions_context *context)
2652 {
2653 	if (node == NULL)
2654 		return NULL;
2655 	switch (nodeTag(node))
2656 	{
2657 		case T_Param:
2658 			{
2659 				Param	   *param = (Param *) node;
2660 				ParamListInfo paramLI = context->boundParams;
2661 
2662 				/* Look to see if we've been given a value for this Param */
2663 				if (param->paramkind == PARAM_EXTERN &&
2664 					paramLI != NULL &&
2665 					param->paramid > 0 &&
2666 					param->paramid <= paramLI->numParams)
2667 				{
2668 					ParamExternData *prm;
2669 					ParamExternData prmdata;
2670 
2671 					/*
2672 					 * Give hook a chance in case parameter is dynamic.  Tell
2673 					 * it that this fetch is speculative, so it should avoid
2674 					 * erroring out if parameter is unavailable.
2675 					 */
2676 					if (paramLI->paramFetch != NULL)
2677 						prm = paramLI->paramFetch(paramLI, param->paramid,
2678 												  true, &prmdata);
2679 					else
2680 						prm = &paramLI->params[param->paramid - 1];
2681 
2682 					/*
2683 					 * We don't just check OidIsValid, but insist that the
2684 					 * fetched type match the Param, just in case the hook did
2685 					 * something unexpected.  No need to throw an error here
2686 					 * though; leave that for runtime.
2687 					 */
2688 					if (OidIsValid(prm->ptype) &&
2689 						prm->ptype == param->paramtype)
2690 					{
2691 						/* OK to substitute parameter value? */
2692 						if (context->estimate ||
2693 							(prm->pflags & PARAM_FLAG_CONST))
2694 						{
2695 							/*
2696 							 * Return a Const representing the param value.
2697 							 * Must copy pass-by-ref datatypes, since the
2698 							 * Param might be in a memory context
2699 							 * shorter-lived than our output plan should be.
2700 							 */
2701 							int16		typLen;
2702 							bool		typByVal;
2703 							Datum		pval;
2704 
2705 							get_typlenbyval(param->paramtype,
2706 											&typLen, &typByVal);
2707 							if (prm->isnull || typByVal)
2708 								pval = prm->value;
2709 							else
2710 								pval = datumCopy(prm->value, typByVal, typLen);
2711 							return (Node *) makeConst(param->paramtype,
2712 													  param->paramtypmod,
2713 													  param->paramcollid,
2714 													  (int) typLen,
2715 													  pval,
2716 													  prm->isnull,
2717 													  typByVal);
2718 						}
2719 					}
2720 				}
2721 
2722 				/*
2723 				 * Not replaceable, so just copy the Param (no need to
2724 				 * recurse)
2725 				 */
2726 				return (Node *) copyObject(param);
2727 			}
2728 		case T_WindowFunc:
2729 			{
2730 				WindowFunc *expr = (WindowFunc *) node;
2731 				Oid			funcid = expr->winfnoid;
2732 				List	   *args;
2733 				Expr	   *aggfilter;
2734 				HeapTuple	func_tuple;
2735 				WindowFunc *newexpr;
2736 
2737 				/*
2738 				 * We can't really simplify a WindowFunc node, but we mustn't
2739 				 * just fall through to the default processing, because we
2740 				 * have to apply expand_function_arguments to its argument
2741 				 * list.  That takes care of inserting default arguments and
2742 				 * expanding named-argument notation.
2743 				 */
2744 				func_tuple = SearchSysCache1(PROCOID, ObjectIdGetDatum(funcid));
2745 				if (!HeapTupleIsValid(func_tuple))
2746 					elog(ERROR, "cache lookup failed for function %u", funcid);
2747 
2748 				args = expand_function_arguments(expr->args, expr->wintype,
2749 												 func_tuple);
2750 
2751 				ReleaseSysCache(func_tuple);
2752 
2753 				/* Now, recursively simplify the args (which are a List) */
2754 				args = (List *)
2755 					expression_tree_mutator((Node *) args,
2756 											eval_const_expressions_mutator,
2757 											(void *) context);
2758 				/* ... and the filter expression, which isn't */
2759 				aggfilter = (Expr *)
2760 					eval_const_expressions_mutator((Node *) expr->aggfilter,
2761 												   context);
2762 
2763 				/* And build the replacement WindowFunc node */
2764 				newexpr = makeNode(WindowFunc);
2765 				newexpr->winfnoid = expr->winfnoid;
2766 				newexpr->wintype = expr->wintype;
2767 				newexpr->wincollid = expr->wincollid;
2768 				newexpr->inputcollid = expr->inputcollid;
2769 				newexpr->args = args;
2770 				newexpr->aggfilter = aggfilter;
2771 				newexpr->winref = expr->winref;
2772 				newexpr->winstar = expr->winstar;
2773 				newexpr->winagg = expr->winagg;
2774 				newexpr->location = expr->location;
2775 
2776 				return (Node *) newexpr;
2777 			}
2778 		case T_FuncExpr:
2779 			{
2780 				FuncExpr   *expr = (FuncExpr *) node;
2781 				List	   *args = expr->args;
2782 				Expr	   *simple;
2783 				FuncExpr   *newexpr;
2784 
2785 				/*
2786 				 * Code for op/func reduction is pretty bulky, so split it out
2787 				 * as a separate function.  Note: exprTypmod normally returns
2788 				 * -1 for a FuncExpr, but not when the node is recognizably a
2789 				 * length coercion; we want to preserve the typmod in the
2790 				 * eventual Const if so.
2791 				 */
2792 				simple = simplify_function(expr->funcid,
2793 										   expr->funcresulttype,
2794 										   exprTypmod(node),
2795 										   expr->funccollid,
2796 										   expr->inputcollid,
2797 										   &args,
2798 										   expr->funcvariadic,
2799 										   true,
2800 										   true,
2801 										   context);
2802 				if (simple)		/* successfully simplified it */
2803 					return (Node *) simple;
2804 
2805 				/*
2806 				 * The expression cannot be simplified any further, so build
2807 				 * and return a replacement FuncExpr node using the
2808 				 * possibly-simplified arguments.  Note that we have also
2809 				 * converted the argument list to positional notation.
2810 				 */
2811 				newexpr = makeNode(FuncExpr);
2812 				newexpr->funcid = expr->funcid;
2813 				newexpr->funcresulttype = expr->funcresulttype;
2814 				newexpr->funcretset = expr->funcretset;
2815 				newexpr->funcvariadic = expr->funcvariadic;
2816 				newexpr->funcformat = expr->funcformat;
2817 				newexpr->funccollid = expr->funccollid;
2818 				newexpr->inputcollid = expr->inputcollid;
2819 				newexpr->args = args;
2820 				newexpr->location = expr->location;
2821 				return (Node *) newexpr;
2822 			}
2823 		case T_OpExpr:
2824 			{
2825 				OpExpr	   *expr = (OpExpr *) node;
2826 				List	   *args = expr->args;
2827 				Expr	   *simple;
2828 				OpExpr	   *newexpr;
2829 
2830 				/*
2831 				 * Need to get OID of underlying function.  Okay to scribble
2832 				 * on input to this extent.
2833 				 */
2834 				set_opfuncid(expr);
2835 
2836 				/*
2837 				 * Code for op/func reduction is pretty bulky, so split it out
2838 				 * as a separate function.
2839 				 */
2840 				simple = simplify_function(expr->opfuncid,
2841 										   expr->opresulttype, -1,
2842 										   expr->opcollid,
2843 										   expr->inputcollid,
2844 										   &args,
2845 										   false,
2846 										   true,
2847 										   true,
2848 										   context);
2849 				if (simple)		/* successfully simplified it */
2850 					return (Node *) simple;
2851 
2852 				/*
2853 				 * If the operator is boolean equality or inequality, we know
2854 				 * how to simplify cases involving one constant and one
2855 				 * non-constant argument.
2856 				 */
2857 				if (expr->opno == BooleanEqualOperator ||
2858 					expr->opno == BooleanNotEqualOperator)
2859 				{
2860 					simple = (Expr *) simplify_boolean_equality(expr->opno,
2861 																args);
2862 					if (simple) /* successfully simplified it */
2863 						return (Node *) simple;
2864 				}
2865 
2866 				/*
2867 				 * The expression cannot be simplified any further, so build
2868 				 * and return a replacement OpExpr node using the
2869 				 * possibly-simplified arguments.
2870 				 */
2871 				newexpr = makeNode(OpExpr);
2872 				newexpr->opno = expr->opno;
2873 				newexpr->opfuncid = expr->opfuncid;
2874 				newexpr->opresulttype = expr->opresulttype;
2875 				newexpr->opretset = expr->opretset;
2876 				newexpr->opcollid = expr->opcollid;
2877 				newexpr->inputcollid = expr->inputcollid;
2878 				newexpr->args = args;
2879 				newexpr->location = expr->location;
2880 				return (Node *) newexpr;
2881 			}
2882 		case T_DistinctExpr:
2883 			{
2884 				DistinctExpr *expr = (DistinctExpr *) node;
2885 				List	   *args;
2886 				ListCell   *arg;
2887 				bool		has_null_input = false;
2888 				bool		all_null_input = true;
2889 				bool		has_nonconst_input = false;
2890 				Expr	   *simple;
2891 				DistinctExpr *newexpr;
2892 
2893 				/*
2894 				 * Reduce constants in the DistinctExpr's arguments.  We know
2895 				 * args is either NIL or a List node, so we can call
2896 				 * expression_tree_mutator directly rather than recursing to
2897 				 * self.
2898 				 */
2899 				args = (List *) expression_tree_mutator((Node *) expr->args,
2900 														eval_const_expressions_mutator,
2901 														(void *) context);
2902 
2903 				/*
2904 				 * We must do our own check for NULLs because DistinctExpr has
2905 				 * different results for NULL input than the underlying
2906 				 * operator does.
2907 				 */
2908 				foreach(arg, args)
2909 				{
2910 					if (IsA(lfirst(arg), Const))
2911 					{
2912 						has_null_input |= ((Const *) lfirst(arg))->constisnull;
2913 						all_null_input &= ((Const *) lfirst(arg))->constisnull;
2914 					}
2915 					else
2916 						has_nonconst_input = true;
2917 				}
2918 
2919 				/* all constants? then can optimize this out */
2920 				if (!has_nonconst_input)
2921 				{
2922 					/* all nulls? then not distinct */
2923 					if (all_null_input)
2924 						return makeBoolConst(false, false);
2925 
2926 					/* one null? then distinct */
2927 					if (has_null_input)
2928 						return makeBoolConst(true, false);
2929 
2930 					/* otherwise try to evaluate the '=' operator */
2931 					/* (NOT okay to try to inline it, though!) */
2932 
2933 					/*
2934 					 * Need to get OID of underlying function.  Okay to
2935 					 * scribble on input to this extent.
2936 					 */
2937 					set_opfuncid((OpExpr *) expr);	/* rely on struct
2938 													 * equivalence */
2939 
2940 					/*
2941 					 * Code for op/func reduction is pretty bulky, so split it
2942 					 * out as a separate function.
2943 					 */
2944 					simple = simplify_function(expr->opfuncid,
2945 											   expr->opresulttype, -1,
2946 											   expr->opcollid,
2947 											   expr->inputcollid,
2948 											   &args,
2949 											   false,
2950 											   false,
2951 											   false,
2952 											   context);
2953 					if (simple) /* successfully simplified it */
2954 					{
2955 						/*
2956 						 * Since the underlying operator is "=", must negate
2957 						 * its result
2958 						 */
2959 						Const	   *csimple = castNode(Const, simple);
2960 
2961 						csimple->constvalue =
2962 							BoolGetDatum(!DatumGetBool(csimple->constvalue));
2963 						return (Node *) csimple;
2964 					}
2965 				}
2966 
2967 				/*
2968 				 * The expression cannot be simplified any further, so build
2969 				 * and return a replacement DistinctExpr node using the
2970 				 * possibly-simplified arguments.
2971 				 */
2972 				newexpr = makeNode(DistinctExpr);
2973 				newexpr->opno = expr->opno;
2974 				newexpr->opfuncid = expr->opfuncid;
2975 				newexpr->opresulttype = expr->opresulttype;
2976 				newexpr->opretset = expr->opretset;
2977 				newexpr->opcollid = expr->opcollid;
2978 				newexpr->inputcollid = expr->inputcollid;
2979 				newexpr->args = args;
2980 				newexpr->location = expr->location;
2981 				return (Node *) newexpr;
2982 			}
2983 		case T_ScalarArrayOpExpr:
2984 			{
2985 				ScalarArrayOpExpr *saop;
2986 
2987 				/* Copy the node and const-simplify its arguments */
2988 				saop = (ScalarArrayOpExpr *) ece_generic_processing(node);
2989 
2990 				/* Make sure we know underlying function */
2991 				set_sa_opfuncid(saop);
2992 
2993 				/*
2994 				 * If all arguments are Consts, and it's a safe function, we
2995 				 * can fold to a constant
2996 				 */
2997 				if (ece_all_arguments_const(saop) &&
2998 					ece_function_is_safe(saop->opfuncid, context))
2999 					return ece_evaluate_expr(saop);
3000 				return (Node *) saop;
3001 			}
3002 		case T_BoolExpr:
3003 			{
3004 				BoolExpr   *expr = (BoolExpr *) node;
3005 
3006 				switch (expr->boolop)
3007 				{
3008 					case OR_EXPR:
3009 						{
3010 							List	   *newargs;
3011 							bool		haveNull = false;
3012 							bool		forceTrue = false;
3013 
3014 							newargs = simplify_or_arguments(expr->args,
3015 															context,
3016 															&haveNull,
3017 															&forceTrue);
3018 							if (forceTrue)
3019 								return makeBoolConst(true, false);
3020 							if (haveNull)
3021 								newargs = lappend(newargs,
3022 												  makeBoolConst(false, true));
3023 							/* If all the inputs are FALSE, result is FALSE */
3024 							if (newargs == NIL)
3025 								return makeBoolConst(false, false);
3026 
3027 							/*
3028 							 * If only one nonconst-or-NULL input, it's the
3029 							 * result
3030 							 */
3031 							if (list_length(newargs) == 1)
3032 								return (Node *) linitial(newargs);
3033 							/* Else we still need an OR node */
3034 							return (Node *) make_orclause(newargs);
3035 						}
3036 					case AND_EXPR:
3037 						{
3038 							List	   *newargs;
3039 							bool		haveNull = false;
3040 							bool		forceFalse = false;
3041 
3042 							newargs = simplify_and_arguments(expr->args,
3043 															 context,
3044 															 &haveNull,
3045 															 &forceFalse);
3046 							if (forceFalse)
3047 								return makeBoolConst(false, false);
3048 							if (haveNull)
3049 								newargs = lappend(newargs,
3050 												  makeBoolConst(false, true));
3051 							/* If all the inputs are TRUE, result is TRUE */
3052 							if (newargs == NIL)
3053 								return makeBoolConst(true, false);
3054 
3055 							/*
3056 							 * If only one nonconst-or-NULL input, it's the
3057 							 * result
3058 							 */
3059 							if (list_length(newargs) == 1)
3060 								return (Node *) linitial(newargs);
3061 							/* Else we still need an AND node */
3062 							return (Node *) make_andclause(newargs);
3063 						}
3064 					case NOT_EXPR:
3065 						{
3066 							Node	   *arg;
3067 
3068 							Assert(list_length(expr->args) == 1);
3069 							arg = eval_const_expressions_mutator(linitial(expr->args),
3070 																 context);
3071 
3072 							/*
3073 							 * Use negate_clause() to see if we can simplify
3074 							 * away the NOT.
3075 							 */
3076 							return negate_clause(arg);
3077 						}
3078 					default:
3079 						elog(ERROR, "unrecognized boolop: %d",
3080 							 (int) expr->boolop);
3081 						break;
3082 				}
3083 				break;
3084 			}
3085 		case T_SubPlan:
3086 		case T_AlternativeSubPlan:
3087 
3088 			/*
3089 			 * Return a SubPlan unchanged --- too late to do anything with it.
3090 			 *
3091 			 * XXX should we ereport() here instead?  Probably this routine
3092 			 * should never be invoked after SubPlan creation.
3093 			 */
3094 			return node;
3095 		case T_RelabelType:
3096 			{
3097 				/*
3098 				 * If we can simplify the input to a constant, then we don't
3099 				 * need the RelabelType node anymore: just change the type
3100 				 * field of the Const node.  Otherwise, must copy the
3101 				 * RelabelType node.
3102 				 */
3103 				RelabelType *relabel = (RelabelType *) node;
3104 				Node	   *arg;
3105 
3106 				arg = eval_const_expressions_mutator((Node *) relabel->arg,
3107 													 context);
3108 
3109 				/*
3110 				 * If we find stacked RelabelTypes (eg, from foo :: int ::
3111 				 * oid) we can discard all but the top one.
3112 				 */
3113 				while (arg && IsA(arg, RelabelType))
3114 					arg = (Node *) ((RelabelType *) arg)->arg;
3115 
3116 				if (arg && IsA(arg, Const))
3117 				{
3118 					Const	   *con = (Const *) arg;
3119 
3120 					con->consttype = relabel->resulttype;
3121 					con->consttypmod = relabel->resulttypmod;
3122 					con->constcollid = relabel->resultcollid;
3123 					return (Node *) con;
3124 				}
3125 				else
3126 				{
3127 					RelabelType *newrelabel = makeNode(RelabelType);
3128 
3129 					newrelabel->arg = (Expr *) arg;
3130 					newrelabel->resulttype = relabel->resulttype;
3131 					newrelabel->resulttypmod = relabel->resulttypmod;
3132 					newrelabel->resultcollid = relabel->resultcollid;
3133 					newrelabel->relabelformat = relabel->relabelformat;
3134 					newrelabel->location = relabel->location;
3135 					return (Node *) newrelabel;
3136 				}
3137 			}
3138 		case T_CoerceViaIO:
3139 			{
3140 				CoerceViaIO *expr = (CoerceViaIO *) node;
3141 				List	   *args;
3142 				Oid			outfunc;
3143 				bool		outtypisvarlena;
3144 				Oid			infunc;
3145 				Oid			intypioparam;
3146 				Expr	   *simple;
3147 				CoerceViaIO *newexpr;
3148 
3149 				/* Make a List so we can use simplify_function */
3150 				args = list_make1(expr->arg);
3151 
3152 				/*
3153 				 * CoerceViaIO represents calling the source type's output
3154 				 * function then the result type's input function.  So, try to
3155 				 * simplify it as though it were a stack of two such function
3156 				 * calls.  First we need to know what the functions are.
3157 				 *
3158 				 * Note that the coercion functions are assumed not to care
3159 				 * about input collation, so we just pass InvalidOid for that.
3160 				 */
3161 				getTypeOutputInfo(exprType((Node *) expr->arg),
3162 								  &outfunc, &outtypisvarlena);
3163 				getTypeInputInfo(expr->resulttype,
3164 								 &infunc, &intypioparam);
3165 
3166 				simple = simplify_function(outfunc,
3167 										   CSTRINGOID, -1,
3168 										   InvalidOid,
3169 										   InvalidOid,
3170 										   &args,
3171 										   false,
3172 										   true,
3173 										   true,
3174 										   context);
3175 				if (simple)		/* successfully simplified output fn */
3176 				{
3177 					/*
3178 					 * Input functions may want 1 to 3 arguments.  We always
3179 					 * supply all three, trusting that nothing downstream will
3180 					 * complain.
3181 					 */
3182 					args = list_make3(simple,
3183 									  makeConst(OIDOID,
3184 												-1,
3185 												InvalidOid,
3186 												sizeof(Oid),
3187 												ObjectIdGetDatum(intypioparam),
3188 												false,
3189 												true),
3190 									  makeConst(INT4OID,
3191 												-1,
3192 												InvalidOid,
3193 												sizeof(int32),
3194 												Int32GetDatum(-1),
3195 												false,
3196 												true));
3197 
3198 					simple = simplify_function(infunc,
3199 											   expr->resulttype, -1,
3200 											   expr->resultcollid,
3201 											   InvalidOid,
3202 											   &args,
3203 											   false,
3204 											   false,
3205 											   true,
3206 											   context);
3207 					if (simple) /* successfully simplified input fn */
3208 						return (Node *) simple;
3209 				}
3210 
3211 				/*
3212 				 * The expression cannot be simplified any further, so build
3213 				 * and return a replacement CoerceViaIO node using the
3214 				 * possibly-simplified argument.
3215 				 */
3216 				newexpr = makeNode(CoerceViaIO);
3217 				newexpr->arg = (Expr *) linitial(args);
3218 				newexpr->resulttype = expr->resulttype;
3219 				newexpr->resultcollid = expr->resultcollid;
3220 				newexpr->coerceformat = expr->coerceformat;
3221 				newexpr->location = expr->location;
3222 				return (Node *) newexpr;
3223 			}
3224 		case T_ArrayCoerceExpr:
3225 			{
3226 				ArrayCoerceExpr *ac = makeNode(ArrayCoerceExpr);
3227 				Node	   *save_case_val;
3228 
3229 				/*
3230 				 * Copy the node and const-simplify its arguments.  We can't
3231 				 * use ece_generic_processing() here because we need to mess
3232 				 * with case_val only while processing the elemexpr.
3233 				 */
3234 				memcpy(ac, node, sizeof(ArrayCoerceExpr));
3235 				ac->arg = (Expr *)
3236 					eval_const_expressions_mutator((Node *) ac->arg,
3237 												   context);
3238 
3239 				/*
3240 				 * Set up for the CaseTestExpr node contained in the elemexpr.
3241 				 * We must prevent it from absorbing any outer CASE value.
3242 				 */
3243 				save_case_val = context->case_val;
3244 				context->case_val = NULL;
3245 
3246 				ac->elemexpr = (Expr *)
3247 					eval_const_expressions_mutator((Node *) ac->elemexpr,
3248 												   context);
3249 
3250 				context->case_val = save_case_val;
3251 
3252 				/*
3253 				 * If constant argument and the per-element expression is
3254 				 * immutable, we can simplify the whole thing to a constant.
3255 				 * Exception: although contain_mutable_functions considers
3256 				 * CoerceToDomain immutable for historical reasons, let's not
3257 				 * do so here; this ensures coercion to an array-over-domain
3258 				 * does not apply the domain's constraints until runtime.
3259 				 */
3260 				if (ac->arg && IsA(ac->arg, Const) &&
3261 					ac->elemexpr && !IsA(ac->elemexpr, CoerceToDomain) &&
3262 					!contain_mutable_functions((Node *) ac->elemexpr))
3263 					return ece_evaluate_expr(ac);
3264 
3265 				return (Node *) ac;
3266 			}
3267 		case T_CollateExpr:
3268 			{
3269 				/*
3270 				 * If we can simplify the input to a constant, then we don't
3271 				 * need the CollateExpr node at all: just change the
3272 				 * constcollid field of the Const node.  Otherwise, replace
3273 				 * the CollateExpr with a RelabelType. (We do that so as to
3274 				 * improve uniformity of expression representation and thus
3275 				 * simplify comparison of expressions.)
3276 				 */
3277 				CollateExpr *collate = (CollateExpr *) node;
3278 				Node	   *arg;
3279 
3280 				arg = eval_const_expressions_mutator((Node *) collate->arg,
3281 													 context);
3282 
3283 				if (arg && IsA(arg, Const))
3284 				{
3285 					Const	   *con = (Const *) arg;
3286 
3287 					con->constcollid = collate->collOid;
3288 					return (Node *) con;
3289 				}
3290 				else if (collate->collOid == exprCollation(arg))
3291 				{
3292 					/* Don't need a RelabelType either... */
3293 					return arg;
3294 				}
3295 				else
3296 				{
3297 					RelabelType *relabel = makeNode(RelabelType);
3298 
3299 					relabel->resulttype = exprType(arg);
3300 					relabel->resulttypmod = exprTypmod(arg);
3301 					relabel->resultcollid = collate->collOid;
3302 					relabel->relabelformat = COERCE_IMPLICIT_CAST;
3303 					relabel->location = collate->location;
3304 
3305 					/* Don't create stacked RelabelTypes */
3306 					while (arg && IsA(arg, RelabelType))
3307 						arg = (Node *) ((RelabelType *) arg)->arg;
3308 					relabel->arg = (Expr *) arg;
3309 
3310 					return (Node *) relabel;
3311 				}
3312 			}
3313 		case T_CaseExpr:
3314 			{
3315 				/*----------
3316 				 * CASE expressions can be simplified if there are constant
3317 				 * condition clauses:
3318 				 *		FALSE (or NULL): drop the alternative
3319 				 *		TRUE: drop all remaining alternatives
3320 				 * If the first non-FALSE alternative is a constant TRUE,
3321 				 * we can simplify the entire CASE to that alternative's
3322 				 * expression.  If there are no non-FALSE alternatives,
3323 				 * we simplify the entire CASE to the default result (ELSE).
3324 				 *
3325 				 * If we have a simple-form CASE with constant test
3326 				 * expression, we substitute the constant value for contained
3327 				 * CaseTestExpr placeholder nodes, so that we have the
3328 				 * opportunity to reduce constant test conditions.  For
3329 				 * example this allows
3330 				 *		CASE 0 WHEN 0 THEN 1 ELSE 1/0 END
3331 				 * to reduce to 1 rather than drawing a divide-by-0 error.
3332 				 * Note that when the test expression is constant, we don't
3333 				 * have to include it in the resulting CASE; for example
3334 				 *		CASE 0 WHEN x THEN y ELSE z END
3335 				 * is transformed by the parser to
3336 				 *		CASE 0 WHEN CaseTestExpr = x THEN y ELSE z END
3337 				 * which we can simplify to
3338 				 *		CASE WHEN 0 = x THEN y ELSE z END
3339 				 * It is not necessary for the executor to evaluate the "arg"
3340 				 * expression when executing the CASE, since any contained
3341 				 * CaseTestExprs that might have referred to it will have been
3342 				 * replaced by the constant.
3343 				 *----------
3344 				 */
3345 				CaseExpr   *caseexpr = (CaseExpr *) node;
3346 				CaseExpr   *newcase;
3347 				Node	   *save_case_val;
3348 				Node	   *newarg;
3349 				List	   *newargs;
3350 				bool		const_true_cond;
3351 				Node	   *defresult = NULL;
3352 				ListCell   *arg;
3353 
3354 				/* Simplify the test expression, if any */
3355 				newarg = eval_const_expressions_mutator((Node *) caseexpr->arg,
3356 														context);
3357 
3358 				/* Set up for contained CaseTestExpr nodes */
3359 				save_case_val = context->case_val;
3360 				if (newarg && IsA(newarg, Const))
3361 				{
3362 					context->case_val = newarg;
3363 					newarg = NULL;	/* not needed anymore, see above */
3364 				}
3365 				else
3366 					context->case_val = NULL;
3367 
3368 				/* Simplify the WHEN clauses */
3369 				newargs = NIL;
3370 				const_true_cond = false;
3371 				foreach(arg, caseexpr->args)
3372 				{
3373 					CaseWhen   *oldcasewhen = lfirst_node(CaseWhen, arg);
3374 					Node	   *casecond;
3375 					Node	   *caseresult;
3376 
3377 					/* Simplify this alternative's test condition */
3378 					casecond = eval_const_expressions_mutator((Node *) oldcasewhen->expr,
3379 															  context);
3380 
3381 					/*
3382 					 * If the test condition is constant FALSE (or NULL), then
3383 					 * drop this WHEN clause completely, without processing
3384 					 * the result.
3385 					 */
3386 					if (casecond && IsA(casecond, Const))
3387 					{
3388 						Const	   *const_input = (Const *) casecond;
3389 
3390 						if (const_input->constisnull ||
3391 							!DatumGetBool(const_input->constvalue))
3392 							continue;	/* drop alternative with FALSE cond */
3393 						/* Else it's constant TRUE */
3394 						const_true_cond = true;
3395 					}
3396 
3397 					/* Simplify this alternative's result value */
3398 					caseresult = eval_const_expressions_mutator((Node *) oldcasewhen->result,
3399 																context);
3400 
3401 					/* If non-constant test condition, emit a new WHEN node */
3402 					if (!const_true_cond)
3403 					{
3404 						CaseWhen   *newcasewhen = makeNode(CaseWhen);
3405 
3406 						newcasewhen->expr = (Expr *) casecond;
3407 						newcasewhen->result = (Expr *) caseresult;
3408 						newcasewhen->location = oldcasewhen->location;
3409 						newargs = lappend(newargs, newcasewhen);
3410 						continue;
3411 					}
3412 
3413 					/*
3414 					 * Found a TRUE condition, so none of the remaining
3415 					 * alternatives can be reached.  We treat the result as
3416 					 * the default result.
3417 					 */
3418 					defresult = caseresult;
3419 					break;
3420 				}
3421 
3422 				/* Simplify the default result, unless we replaced it above */
3423 				if (!const_true_cond)
3424 					defresult = eval_const_expressions_mutator((Node *) caseexpr->defresult,
3425 															   context);
3426 
3427 				context->case_val = save_case_val;
3428 
3429 				/*
3430 				 * If no non-FALSE alternatives, CASE reduces to the default
3431 				 * result
3432 				 */
3433 				if (newargs == NIL)
3434 					return defresult;
3435 				/* Otherwise we need a new CASE node */
3436 				newcase = makeNode(CaseExpr);
3437 				newcase->casetype = caseexpr->casetype;
3438 				newcase->casecollid = caseexpr->casecollid;
3439 				newcase->arg = (Expr *) newarg;
3440 				newcase->args = newargs;
3441 				newcase->defresult = (Expr *) defresult;
3442 				newcase->location = caseexpr->location;
3443 				return (Node *) newcase;
3444 			}
3445 		case T_CaseTestExpr:
3446 			{
3447 				/*
3448 				 * If we know a constant test value for the current CASE
3449 				 * construct, substitute it for the placeholder.  Else just
3450 				 * return the placeholder as-is.
3451 				 */
3452 				if (context->case_val)
3453 					return copyObject(context->case_val);
3454 				else
3455 					return copyObject(node);
3456 			}
3457 		case T_ArrayRef:
3458 		case T_ArrayExpr:
3459 		case T_RowExpr:
3460 			{
3461 				/*
3462 				 * Generic handling for node types whose own processing is
3463 				 * known to be immutable, and for which we need no smarts
3464 				 * beyond "simplify if all inputs are constants".
3465 				 */
3466 
3467 				/* Copy the node and const-simplify its arguments */
3468 				node = ece_generic_processing(node);
3469 				/* If all arguments are Consts, we can fold to a constant */
3470 				if (ece_all_arguments_const(node))
3471 					return ece_evaluate_expr(node);
3472 				return node;
3473 			}
3474 		case T_CoalesceExpr:
3475 			{
3476 				CoalesceExpr *coalesceexpr = (CoalesceExpr *) node;
3477 				CoalesceExpr *newcoalesce;
3478 				List	   *newargs;
3479 				ListCell   *arg;
3480 
3481 				newargs = NIL;
3482 				foreach(arg, coalesceexpr->args)
3483 				{
3484 					Node	   *e;
3485 
3486 					e = eval_const_expressions_mutator((Node *) lfirst(arg),
3487 													   context);
3488 
3489 					/*
3490 					 * We can remove null constants from the list. For a
3491 					 * non-null constant, if it has not been preceded by any
3492 					 * other non-null-constant expressions then it is the
3493 					 * result. Otherwise, it's the next argument, but we can
3494 					 * drop following arguments since they will never be
3495 					 * reached.
3496 					 */
3497 					if (IsA(e, Const))
3498 					{
3499 						if (((Const *) e)->constisnull)
3500 							continue;	/* drop null constant */
3501 						if (newargs == NIL)
3502 							return e;	/* first expr */
3503 						newargs = lappend(newargs, e);
3504 						break;
3505 					}
3506 					newargs = lappend(newargs, e);
3507 				}
3508 
3509 				/*
3510 				 * If all the arguments were constant null, the result is just
3511 				 * null
3512 				 */
3513 				if (newargs == NIL)
3514 					return (Node *) makeNullConst(coalesceexpr->coalescetype,
3515 												  -1,
3516 												  coalesceexpr->coalescecollid);
3517 
3518 				newcoalesce = makeNode(CoalesceExpr);
3519 				newcoalesce->coalescetype = coalesceexpr->coalescetype;
3520 				newcoalesce->coalescecollid = coalesceexpr->coalescecollid;
3521 				newcoalesce->args = newargs;
3522 				newcoalesce->location = coalesceexpr->location;
3523 				return (Node *) newcoalesce;
3524 			}
3525 		case T_SQLValueFunction:
3526 			{
3527 				/*
3528 				 * All variants of SQLValueFunction are stable, so if we are
3529 				 * estimating the expression's value, we should evaluate the
3530 				 * current function value.  Otherwise just copy.
3531 				 */
3532 				SQLValueFunction *svf = (SQLValueFunction *) node;
3533 
3534 				if (context->estimate)
3535 					return (Node *) evaluate_expr((Expr *) svf,
3536 												  svf->type,
3537 												  svf->typmod,
3538 												  InvalidOid);
3539 				else
3540 					return copyObject((Node *) svf);
3541 			}
3542 		case T_FieldSelect:
3543 			{
3544 				/*
3545 				 * We can optimize field selection from a whole-row Var into a
3546 				 * simple Var.  (This case won't be generated directly by the
3547 				 * parser, because ParseComplexProjection short-circuits it.
3548 				 * But it can arise while simplifying functions.)  Also, we
3549 				 * can optimize field selection from a RowExpr construct, or
3550 				 * of course from a constant.
3551 				 *
3552 				 * However, replacing a whole-row Var in this way has a
3553 				 * pitfall: if we've already built the rel targetlist for the
3554 				 * source relation, then the whole-row Var is scheduled to be
3555 				 * produced by the relation scan, but the simple Var probably
3556 				 * isn't, which will lead to a failure in setrefs.c.  This is
3557 				 * not a problem when handling simple single-level queries, in
3558 				 * which expression simplification always happens first.  It
3559 				 * is a risk for lateral references from subqueries, though.
3560 				 * To avoid such failures, don't optimize uplevel references.
3561 				 *
3562 				 * We must also check that the declared type of the field is
3563 				 * still the same as when the FieldSelect was created --- this
3564 				 * can change if someone did ALTER COLUMN TYPE on the rowtype.
3565 				 * If it isn't, we skip the optimization; the case will
3566 				 * probably fail at runtime, but that's not our problem here.
3567 				 */
3568 				FieldSelect *fselect = (FieldSelect *) node;
3569 				FieldSelect *newfselect;
3570 				Node	   *arg;
3571 
3572 				arg = eval_const_expressions_mutator((Node *) fselect->arg,
3573 													 context);
3574 				if (arg && IsA(arg, Var) &&
3575 					((Var *) arg)->varattno == InvalidAttrNumber &&
3576 					((Var *) arg)->varlevelsup == 0)
3577 				{
3578 					if (rowtype_field_matches(((Var *) arg)->vartype,
3579 											  fselect->fieldnum,
3580 											  fselect->resulttype,
3581 											  fselect->resulttypmod,
3582 											  fselect->resultcollid))
3583 						return (Node *) makeVar(((Var *) arg)->varno,
3584 												fselect->fieldnum,
3585 												fselect->resulttype,
3586 												fselect->resulttypmod,
3587 												fselect->resultcollid,
3588 												((Var *) arg)->varlevelsup);
3589 				}
3590 				if (arg && IsA(arg, RowExpr))
3591 				{
3592 					RowExpr    *rowexpr = (RowExpr *) arg;
3593 
3594 					if (fselect->fieldnum > 0 &&
3595 						fselect->fieldnum <= list_length(rowexpr->args))
3596 					{
3597 						Node	   *fld = (Node *) list_nth(rowexpr->args,
3598 															fselect->fieldnum - 1);
3599 
3600 						if (rowtype_field_matches(rowexpr->row_typeid,
3601 												  fselect->fieldnum,
3602 												  fselect->resulttype,
3603 												  fselect->resulttypmod,
3604 												  fselect->resultcollid) &&
3605 							fselect->resulttype == exprType(fld) &&
3606 							fselect->resulttypmod == exprTypmod(fld) &&
3607 							fselect->resultcollid == exprCollation(fld))
3608 							return fld;
3609 					}
3610 				}
3611 				newfselect = makeNode(FieldSelect);
3612 				newfselect->arg = (Expr *) arg;
3613 				newfselect->fieldnum = fselect->fieldnum;
3614 				newfselect->resulttype = fselect->resulttype;
3615 				newfselect->resulttypmod = fselect->resulttypmod;
3616 				newfselect->resultcollid = fselect->resultcollid;
3617 				if (arg && IsA(arg, Const))
3618 				{
3619 					Const	   *con = (Const *) arg;
3620 
3621 					if (rowtype_field_matches(con->consttype,
3622 											  newfselect->fieldnum,
3623 											  newfselect->resulttype,
3624 											  newfselect->resulttypmod,
3625 											  newfselect->resultcollid))
3626 						return ece_evaluate_expr(newfselect);
3627 				}
3628 				return (Node *) newfselect;
3629 			}
3630 		case T_NullTest:
3631 			{
3632 				NullTest   *ntest = (NullTest *) node;
3633 				NullTest   *newntest;
3634 				Node	   *arg;
3635 
3636 				arg = eval_const_expressions_mutator((Node *) ntest->arg,
3637 													 context);
3638 				if (ntest->argisrow && arg && IsA(arg, RowExpr))
3639 				{
3640 					/*
3641 					 * We break ROW(...) IS [NOT] NULL into separate tests on
3642 					 * its component fields.  This form is usually more
3643 					 * efficient to evaluate, as well as being more amenable
3644 					 * to optimization.
3645 					 */
3646 					RowExpr    *rarg = (RowExpr *) arg;
3647 					List	   *newargs = NIL;
3648 					ListCell   *l;
3649 
3650 					foreach(l, rarg->args)
3651 					{
3652 						Node	   *relem = (Node *) lfirst(l);
3653 
3654 						/*
3655 						 * A constant field refutes the whole NullTest if it's
3656 						 * of the wrong nullness; else we can discard it.
3657 						 */
3658 						if (relem && IsA(relem, Const))
3659 						{
3660 							Const	   *carg = (Const *) relem;
3661 
3662 							if (carg->constisnull ?
3663 								(ntest->nulltesttype == IS_NOT_NULL) :
3664 								(ntest->nulltesttype == IS_NULL))
3665 								return makeBoolConst(false, false);
3666 							continue;
3667 						}
3668 
3669 						/*
3670 						 * Else, make a scalar (argisrow == false) NullTest
3671 						 * for this field.  Scalar semantics are required
3672 						 * because IS [NOT] NULL doesn't recurse; see comments
3673 						 * in ExecEvalRowNullInt().
3674 						 */
3675 						newntest = makeNode(NullTest);
3676 						newntest->arg = (Expr *) relem;
3677 						newntest->nulltesttype = ntest->nulltesttype;
3678 						newntest->argisrow = false;
3679 						newntest->location = ntest->location;
3680 						newargs = lappend(newargs, newntest);
3681 					}
3682 					/* If all the inputs were constants, result is TRUE */
3683 					if (newargs == NIL)
3684 						return makeBoolConst(true, false);
3685 					/* If only one nonconst input, it's the result */
3686 					if (list_length(newargs) == 1)
3687 						return (Node *) linitial(newargs);
3688 					/* Else we need an AND node */
3689 					return (Node *) make_andclause(newargs);
3690 				}
3691 				if (!ntest->argisrow && arg && IsA(arg, Const))
3692 				{
3693 					Const	   *carg = (Const *) arg;
3694 					bool		result;
3695 
3696 					switch (ntest->nulltesttype)
3697 					{
3698 						case IS_NULL:
3699 							result = carg->constisnull;
3700 							break;
3701 						case IS_NOT_NULL:
3702 							result = !carg->constisnull;
3703 							break;
3704 						default:
3705 							elog(ERROR, "unrecognized nulltesttype: %d",
3706 								 (int) ntest->nulltesttype);
3707 							result = false; /* keep compiler quiet */
3708 							break;
3709 					}
3710 
3711 					return makeBoolConst(result, false);
3712 				}
3713 
3714 				newntest = makeNode(NullTest);
3715 				newntest->arg = (Expr *) arg;
3716 				newntest->nulltesttype = ntest->nulltesttype;
3717 				newntest->argisrow = ntest->argisrow;
3718 				newntest->location = ntest->location;
3719 				return (Node *) newntest;
3720 			}
3721 		case T_BooleanTest:
3722 			{
3723 				/*
3724 				 * This case could be folded into the generic handling used
3725 				 * for ArrayRef etc.  But because the simplification logic is
3726 				 * so trivial, applying evaluate_expr() to perform it would be
3727 				 * a heavy overhead.  BooleanTest is probably common enough to
3728 				 * justify keeping this bespoke implementation.
3729 				 */
3730 				BooleanTest *btest = (BooleanTest *) node;
3731 				BooleanTest *newbtest;
3732 				Node	   *arg;
3733 
3734 				arg = eval_const_expressions_mutator((Node *) btest->arg,
3735 													 context);
3736 				if (arg && IsA(arg, Const))
3737 				{
3738 					Const	   *carg = (Const *) arg;
3739 					bool		result;
3740 
3741 					switch (btest->booltesttype)
3742 					{
3743 						case IS_TRUE:
3744 							result = (!carg->constisnull &&
3745 									  DatumGetBool(carg->constvalue));
3746 							break;
3747 						case IS_NOT_TRUE:
3748 							result = (carg->constisnull ||
3749 									  !DatumGetBool(carg->constvalue));
3750 							break;
3751 						case IS_FALSE:
3752 							result = (!carg->constisnull &&
3753 									  !DatumGetBool(carg->constvalue));
3754 							break;
3755 						case IS_NOT_FALSE:
3756 							result = (carg->constisnull ||
3757 									  DatumGetBool(carg->constvalue));
3758 							break;
3759 						case IS_UNKNOWN:
3760 							result = carg->constisnull;
3761 							break;
3762 						case IS_NOT_UNKNOWN:
3763 							result = !carg->constisnull;
3764 							break;
3765 						default:
3766 							elog(ERROR, "unrecognized booltesttype: %d",
3767 								 (int) btest->booltesttype);
3768 							result = false; /* keep compiler quiet */
3769 							break;
3770 					}
3771 
3772 					return makeBoolConst(result, false);
3773 				}
3774 
3775 				newbtest = makeNode(BooleanTest);
3776 				newbtest->arg = (Expr *) arg;
3777 				newbtest->booltesttype = btest->booltesttype;
3778 				newbtest->location = btest->location;
3779 				return (Node *) newbtest;
3780 			}
3781 		case T_PlaceHolderVar:
3782 
3783 			/*
3784 			 * In estimation mode, just strip the PlaceHolderVar node
3785 			 * altogether; this amounts to estimating that the contained value
3786 			 * won't be forced to null by an outer join.  In regular mode we
3787 			 * just use the default behavior (ie, simplify the expression but
3788 			 * leave the PlaceHolderVar node intact).
3789 			 */
3790 			if (context->estimate)
3791 			{
3792 				PlaceHolderVar *phv = (PlaceHolderVar *) node;
3793 
3794 				return eval_const_expressions_mutator((Node *) phv->phexpr,
3795 													  context);
3796 			}
3797 			break;
3798 		default:
3799 			break;
3800 	}
3801 
3802 	/*
3803 	 * For any node type not handled above, copy the node unchanged but
3804 	 * const-simplify its subexpressions.  This is the correct thing for node
3805 	 * types whose behavior might change between planning and execution, such
3806 	 * as CoerceToDomain.  It's also a safe default for new node types not
3807 	 * known to this routine.
3808 	 */
3809 	return ece_generic_processing(node);
3810 }
3811 
3812 /*
3813  * Subroutine for eval_const_expressions: check for non-Const nodes.
3814  *
3815  * We can abort recursion immediately on finding a non-Const node.  This is
3816  * critical for performance, else eval_const_expressions_mutator would take
3817  * O(N^2) time on non-simplifiable trees.  However, we do need to descend
3818  * into List nodes since expression_tree_walker sometimes invokes the walker
3819  * function directly on List subtrees.
3820  */
3821 static bool
contain_non_const_walker(Node * node,void * context)3822 contain_non_const_walker(Node *node, void *context)
3823 {
3824 	if (node == NULL)
3825 		return false;
3826 	if (IsA(node, Const))
3827 		return false;
3828 	if (IsA(node, List))
3829 		return expression_tree_walker(node, contain_non_const_walker, context);
3830 	/* Otherwise, abort the tree traversal and return true */
3831 	return true;
3832 }
3833 
3834 /*
3835  * Subroutine for eval_const_expressions: check if a function is OK to evaluate
3836  */
3837 static bool
ece_function_is_safe(Oid funcid,eval_const_expressions_context * context)3838 ece_function_is_safe(Oid funcid, eval_const_expressions_context *context)
3839 {
3840 	char		provolatile = func_volatile(funcid);
3841 
3842 	/*
3843 	 * Ordinarily we are only allowed to simplify immutable functions. But for
3844 	 * purposes of estimation, we consider it okay to simplify functions that
3845 	 * are merely stable; the risk that the result might change from planning
3846 	 * time to execution time is worth taking in preference to not being able
3847 	 * to estimate the value at all.
3848 	 */
3849 	if (provolatile == PROVOLATILE_IMMUTABLE)
3850 		return true;
3851 	if (context->estimate && provolatile == PROVOLATILE_STABLE)
3852 		return true;
3853 	return false;
3854 }
3855 
3856 /*
3857  * Subroutine for eval_const_expressions: process arguments of an OR clause
3858  *
3859  * This includes flattening of nested ORs as well as recursion to
3860  * eval_const_expressions to simplify the OR arguments.
3861  *
3862  * After simplification, OR arguments are handled as follows:
3863  *		non constant: keep
3864  *		FALSE: drop (does not affect result)
3865  *		TRUE: force result to TRUE
3866  *		NULL: keep only one
3867  * We must keep one NULL input because OR expressions evaluate to NULL when no
3868  * input is TRUE and at least one is NULL.  We don't actually include the NULL
3869  * here, that's supposed to be done by the caller.
3870  *
3871  * The output arguments *haveNull and *forceTrue must be initialized false
3872  * by the caller.  They will be set true if a NULL constant or TRUE constant,
3873  * respectively, is detected anywhere in the argument list.
3874  */
3875 static List *
simplify_or_arguments(List * args,eval_const_expressions_context * context,bool * haveNull,bool * forceTrue)3876 simplify_or_arguments(List *args,
3877 					  eval_const_expressions_context *context,
3878 					  bool *haveNull, bool *forceTrue)
3879 {
3880 	List	   *newargs = NIL;
3881 	List	   *unprocessed_args;
3882 
3883 	/*
3884 	 * We want to ensure that any OR immediately beneath another OR gets
3885 	 * flattened into a single OR-list, so as to simplify later reasoning.
3886 	 *
3887 	 * To avoid stack overflow from recursion of eval_const_expressions, we
3888 	 * resort to some tenseness here: we keep a list of not-yet-processed
3889 	 * inputs, and handle flattening of nested ORs by prepending to the to-do
3890 	 * list instead of recursing.  Now that the parser generates N-argument
3891 	 * ORs from simple lists, this complexity is probably less necessary than
3892 	 * it once was, but we might as well keep the logic.
3893 	 */
3894 	unprocessed_args = list_copy(args);
3895 	while (unprocessed_args)
3896 	{
3897 		Node	   *arg = (Node *) linitial(unprocessed_args);
3898 
3899 		unprocessed_args = list_delete_first(unprocessed_args);
3900 
3901 		/* flatten nested ORs as per above comment */
3902 		if (or_clause(arg))
3903 		{
3904 			List	   *subargs = list_copy(((BoolExpr *) arg)->args);
3905 
3906 			/* overly tense code to avoid leaking unused list header */
3907 			if (!unprocessed_args)
3908 				unprocessed_args = subargs;
3909 			else
3910 			{
3911 				List	   *oldhdr = unprocessed_args;
3912 
3913 				unprocessed_args = list_concat(subargs, unprocessed_args);
3914 				pfree(oldhdr);
3915 			}
3916 			continue;
3917 		}
3918 
3919 		/* If it's not an OR, simplify it */
3920 		arg = eval_const_expressions_mutator(arg, context);
3921 
3922 		/*
3923 		 * It is unlikely but not impossible for simplification of a non-OR
3924 		 * clause to produce an OR.  Recheck, but don't be too tense about it
3925 		 * since it's not a mainstream case. In particular we don't worry
3926 		 * about const-simplifying the input twice.
3927 		 */
3928 		if (or_clause(arg))
3929 		{
3930 			List	   *subargs = list_copy(((BoolExpr *) arg)->args);
3931 
3932 			unprocessed_args = list_concat(subargs, unprocessed_args);
3933 			continue;
3934 		}
3935 
3936 		/*
3937 		 * OK, we have a const-simplified non-OR argument.  Process it per
3938 		 * comments above.
3939 		 */
3940 		if (IsA(arg, Const))
3941 		{
3942 			Const	   *const_input = (Const *) arg;
3943 
3944 			if (const_input->constisnull)
3945 				*haveNull = true;
3946 			else if (DatumGetBool(const_input->constvalue))
3947 			{
3948 				*forceTrue = true;
3949 
3950 				/*
3951 				 * Once we detect a TRUE result we can just exit the loop
3952 				 * immediately.  However, if we ever add a notion of
3953 				 * non-removable functions, we'd need to keep scanning.
3954 				 */
3955 				return NIL;
3956 			}
3957 			/* otherwise, we can drop the constant-false input */
3958 			continue;
3959 		}
3960 
3961 		/* else emit the simplified arg into the result list */
3962 		newargs = lappend(newargs, arg);
3963 	}
3964 
3965 	return newargs;
3966 }
3967 
3968 /*
3969  * Subroutine for eval_const_expressions: process arguments of an AND clause
3970  *
3971  * This includes flattening of nested ANDs as well as recursion to
3972  * eval_const_expressions to simplify the AND arguments.
3973  *
3974  * After simplification, AND arguments are handled as follows:
3975  *		non constant: keep
3976  *		TRUE: drop (does not affect result)
3977  *		FALSE: force result to FALSE
3978  *		NULL: keep only one
3979  * We must keep one NULL input because AND expressions evaluate to NULL when
3980  * no input is FALSE and at least one is NULL.  We don't actually include the
3981  * NULL here, that's supposed to be done by the caller.
3982  *
3983  * The output arguments *haveNull and *forceFalse must be initialized false
3984  * by the caller.  They will be set true if a null constant or false constant,
3985  * respectively, is detected anywhere in the argument list.
3986  */
3987 static List *
simplify_and_arguments(List * args,eval_const_expressions_context * context,bool * haveNull,bool * forceFalse)3988 simplify_and_arguments(List *args,
3989 					   eval_const_expressions_context *context,
3990 					   bool *haveNull, bool *forceFalse)
3991 {
3992 	List	   *newargs = NIL;
3993 	List	   *unprocessed_args;
3994 
3995 	/* See comments in simplify_or_arguments */
3996 	unprocessed_args = list_copy(args);
3997 	while (unprocessed_args)
3998 	{
3999 		Node	   *arg = (Node *) linitial(unprocessed_args);
4000 
4001 		unprocessed_args = list_delete_first(unprocessed_args);
4002 
4003 		/* flatten nested ANDs as per above comment */
4004 		if (and_clause(arg))
4005 		{
4006 			List	   *subargs = list_copy(((BoolExpr *) arg)->args);
4007 
4008 			/* overly tense code to avoid leaking unused list header */
4009 			if (!unprocessed_args)
4010 				unprocessed_args = subargs;
4011 			else
4012 			{
4013 				List	   *oldhdr = unprocessed_args;
4014 
4015 				unprocessed_args = list_concat(subargs, unprocessed_args);
4016 				pfree(oldhdr);
4017 			}
4018 			continue;
4019 		}
4020 
4021 		/* If it's not an AND, simplify it */
4022 		arg = eval_const_expressions_mutator(arg, context);
4023 
4024 		/*
4025 		 * It is unlikely but not impossible for simplification of a non-AND
4026 		 * clause to produce an AND.  Recheck, but don't be too tense about it
4027 		 * since it's not a mainstream case. In particular we don't worry
4028 		 * about const-simplifying the input twice.
4029 		 */
4030 		if (and_clause(arg))
4031 		{
4032 			List	   *subargs = list_copy(((BoolExpr *) arg)->args);
4033 
4034 			unprocessed_args = list_concat(subargs, unprocessed_args);
4035 			continue;
4036 		}
4037 
4038 		/*
4039 		 * OK, we have a const-simplified non-AND argument.  Process it per
4040 		 * comments above.
4041 		 */
4042 		if (IsA(arg, Const))
4043 		{
4044 			Const	   *const_input = (Const *) arg;
4045 
4046 			if (const_input->constisnull)
4047 				*haveNull = true;
4048 			else if (!DatumGetBool(const_input->constvalue))
4049 			{
4050 				*forceFalse = true;
4051 
4052 				/*
4053 				 * Once we detect a FALSE result we can just exit the loop
4054 				 * immediately.  However, if we ever add a notion of
4055 				 * non-removable functions, we'd need to keep scanning.
4056 				 */
4057 				return NIL;
4058 			}
4059 			/* otherwise, we can drop the constant-true input */
4060 			continue;
4061 		}
4062 
4063 		/* else emit the simplified arg into the result list */
4064 		newargs = lappend(newargs, arg);
4065 	}
4066 
4067 	return newargs;
4068 }
4069 
4070 /*
4071  * Subroutine for eval_const_expressions: try to simplify boolean equality
4072  * or inequality condition
4073  *
4074  * Inputs are the operator OID and the simplified arguments to the operator.
4075  * Returns a simplified expression if successful, or NULL if cannot
4076  * simplify the expression.
4077  *
4078  * The idea here is to reduce "x = true" to "x" and "x = false" to "NOT x",
4079  * or similarly "x <> true" to "NOT x" and "x <> false" to "x".
4080  * This is only marginally useful in itself, but doing it in constant folding
4081  * ensures that we will recognize these forms as being equivalent in, for
4082  * example, partial index matching.
4083  *
4084  * We come here only if simplify_function has failed; therefore we cannot
4085  * see two constant inputs, nor a constant-NULL input.
4086  */
4087 static Node *
simplify_boolean_equality(Oid opno,List * args)4088 simplify_boolean_equality(Oid opno, List *args)
4089 {
4090 	Node	   *leftop;
4091 	Node	   *rightop;
4092 
4093 	Assert(list_length(args) == 2);
4094 	leftop = linitial(args);
4095 	rightop = lsecond(args);
4096 	if (leftop && IsA(leftop, Const))
4097 	{
4098 		Assert(!((Const *) leftop)->constisnull);
4099 		if (opno == BooleanEqualOperator)
4100 		{
4101 			if (DatumGetBool(((Const *) leftop)->constvalue))
4102 				return rightop; /* true = foo */
4103 			else
4104 				return negate_clause(rightop);	/* false = foo */
4105 		}
4106 		else
4107 		{
4108 			if (DatumGetBool(((Const *) leftop)->constvalue))
4109 				return negate_clause(rightop);	/* true <> foo */
4110 			else
4111 				return rightop; /* false <> foo */
4112 		}
4113 	}
4114 	if (rightop && IsA(rightop, Const))
4115 	{
4116 		Assert(!((Const *) rightop)->constisnull);
4117 		if (opno == BooleanEqualOperator)
4118 		{
4119 			if (DatumGetBool(((Const *) rightop)->constvalue))
4120 				return leftop;	/* foo = true */
4121 			else
4122 				return negate_clause(leftop);	/* foo = false */
4123 		}
4124 		else
4125 		{
4126 			if (DatumGetBool(((Const *) rightop)->constvalue))
4127 				return negate_clause(leftop);	/* foo <> true */
4128 			else
4129 				return leftop;	/* foo <> false */
4130 		}
4131 	}
4132 	return NULL;
4133 }
4134 
4135 /*
4136  * Subroutine for eval_const_expressions: try to simplify a function call
4137  * (which might originally have been an operator; we don't care)
4138  *
4139  * Inputs are the function OID, actual result type OID (which is needed for
4140  * polymorphic functions), result typmod, result collation, the input
4141  * collation to use for the function, the original argument list (not
4142  * const-simplified yet, unless process_args is false), and some flags;
4143  * also the context data for eval_const_expressions.
4144  *
4145  * Returns a simplified expression if successful, or NULL if cannot
4146  * simplify the function call.
4147  *
4148  * This function is also responsible for converting named-notation argument
4149  * lists into positional notation and/or adding any needed default argument
4150  * expressions; which is a bit grotty, but it avoids extra fetches of the
4151  * function's pg_proc tuple.  For this reason, the args list is
4152  * pass-by-reference.  Conversion and const-simplification of the args list
4153  * will be done even if simplification of the function call itself is not
4154  * possible.
4155  */
4156 static Expr *
simplify_function(Oid funcid,Oid result_type,int32 result_typmod,Oid result_collid,Oid input_collid,List ** args_p,bool funcvariadic,bool process_args,bool allow_non_const,eval_const_expressions_context * context)4157 simplify_function(Oid funcid, Oid result_type, int32 result_typmod,
4158 				  Oid result_collid, Oid input_collid, List **args_p,
4159 				  bool funcvariadic, bool process_args, bool allow_non_const,
4160 				  eval_const_expressions_context *context)
4161 {
4162 	List	   *args = *args_p;
4163 	HeapTuple	func_tuple;
4164 	Form_pg_proc func_form;
4165 	Expr	   *newexpr;
4166 
4167 	/*
4168 	 * We have three strategies for simplification: execute the function to
4169 	 * deliver a constant result, use a transform function to generate a
4170 	 * substitute node tree, or expand in-line the body of the function
4171 	 * definition (which only works for simple SQL-language functions, but
4172 	 * that is a common case).  Each case needs access to the function's
4173 	 * pg_proc tuple, so fetch it just once.
4174 	 *
4175 	 * Note: the allow_non_const flag suppresses both the second and third
4176 	 * strategies; so if !allow_non_const, simplify_function can only return a
4177 	 * Const or NULL.  Argument-list rewriting happens anyway, though.
4178 	 */
4179 	func_tuple = SearchSysCache1(PROCOID, ObjectIdGetDatum(funcid));
4180 	if (!HeapTupleIsValid(func_tuple))
4181 		elog(ERROR, "cache lookup failed for function %u", funcid);
4182 	func_form = (Form_pg_proc) GETSTRUCT(func_tuple);
4183 
4184 	/*
4185 	 * Process the function arguments, unless the caller did it already.
4186 	 *
4187 	 * Here we must deal with named or defaulted arguments, and then
4188 	 * recursively apply eval_const_expressions to the whole argument list.
4189 	 */
4190 	if (process_args)
4191 	{
4192 		args = expand_function_arguments(args, result_type, func_tuple);
4193 		args = (List *) expression_tree_mutator((Node *) args,
4194 												eval_const_expressions_mutator,
4195 												(void *) context);
4196 		/* Argument processing done, give it back to the caller */
4197 		*args_p = args;
4198 	}
4199 
4200 	/* Now attempt simplification of the function call proper. */
4201 
4202 	newexpr = evaluate_function(funcid, result_type, result_typmod,
4203 								result_collid, input_collid,
4204 								args, funcvariadic,
4205 								func_tuple, context);
4206 
4207 	if (!newexpr && allow_non_const && OidIsValid(func_form->protransform))
4208 	{
4209 		/*
4210 		 * Build a dummy FuncExpr node containing the simplified arg list.  We
4211 		 * use this approach to present a uniform interface to the transform
4212 		 * function regardless of how the function is actually being invoked.
4213 		 */
4214 		FuncExpr	fexpr;
4215 
4216 		fexpr.xpr.type = T_FuncExpr;
4217 		fexpr.funcid = funcid;
4218 		fexpr.funcresulttype = result_type;
4219 		fexpr.funcretset = func_form->proretset;
4220 		fexpr.funcvariadic = funcvariadic;
4221 		fexpr.funcformat = COERCE_EXPLICIT_CALL;
4222 		fexpr.funccollid = result_collid;
4223 		fexpr.inputcollid = input_collid;
4224 		fexpr.args = args;
4225 		fexpr.location = -1;
4226 
4227 		newexpr = (Expr *)
4228 			DatumGetPointer(OidFunctionCall1(func_form->protransform,
4229 											 PointerGetDatum(&fexpr)));
4230 	}
4231 
4232 	if (!newexpr && allow_non_const)
4233 		newexpr = inline_function(funcid, result_type, result_collid,
4234 								  input_collid, args, funcvariadic,
4235 								  func_tuple, context);
4236 
4237 	ReleaseSysCache(func_tuple);
4238 
4239 	return newexpr;
4240 }
4241 
4242 /*
4243  * expand_function_arguments: convert named-notation args to positional args
4244  * and/or insert default args, as needed
4245  *
4246  * If we need to change anything, the input argument list is copied, not
4247  * modified.
4248  *
4249  * Note: this gets applied to operator argument lists too, even though the
4250  * cases it handles should never occur there.  This should be OK since it
4251  * will fall through very quickly if there's nothing to do.
4252  */
4253 List *
expand_function_arguments(List * args,Oid result_type,HeapTuple func_tuple)4254 expand_function_arguments(List *args, Oid result_type, HeapTuple func_tuple)
4255 {
4256 	Form_pg_proc funcform = (Form_pg_proc) GETSTRUCT(func_tuple);
4257 	bool		has_named_args = false;
4258 	ListCell   *lc;
4259 
4260 	/* Do we have any named arguments? */
4261 	foreach(lc, args)
4262 	{
4263 		Node	   *arg = (Node *) lfirst(lc);
4264 
4265 		if (IsA(arg, NamedArgExpr))
4266 		{
4267 			has_named_args = true;
4268 			break;
4269 		}
4270 	}
4271 
4272 	/* If so, we must apply reorder_function_arguments */
4273 	if (has_named_args)
4274 	{
4275 		args = reorder_function_arguments(args, func_tuple);
4276 		/* Recheck argument types and add casts if needed */
4277 		recheck_cast_function_args(args, result_type, func_tuple);
4278 	}
4279 	else if (list_length(args) < funcform->pronargs)
4280 	{
4281 		/* No named args, but we seem to be short some defaults */
4282 		args = add_function_defaults(args, func_tuple);
4283 		/* Recheck argument types and add casts if needed */
4284 		recheck_cast_function_args(args, result_type, func_tuple);
4285 	}
4286 
4287 	return args;
4288 }
4289 
4290 /*
4291  * reorder_function_arguments: convert named-notation args to positional args
4292  *
4293  * This function also inserts default argument values as needed, since it's
4294  * impossible to form a truly valid positional call without that.
4295  */
4296 static List *
reorder_function_arguments(List * args,HeapTuple func_tuple)4297 reorder_function_arguments(List *args, HeapTuple func_tuple)
4298 {
4299 	Form_pg_proc funcform = (Form_pg_proc) GETSTRUCT(func_tuple);
4300 	int			pronargs = funcform->pronargs;
4301 	int			nargsprovided = list_length(args);
4302 	Node	   *argarray[FUNC_MAX_ARGS];
4303 	ListCell   *lc;
4304 	int			i;
4305 
4306 	Assert(nargsprovided <= pronargs);
4307 	if (pronargs < 0 || pronargs > FUNC_MAX_ARGS)
4308 		elog(ERROR, "too many function arguments");
4309 	memset(argarray, 0, pronargs * sizeof(Node *));
4310 
4311 	/* Deconstruct the argument list into an array indexed by argnumber */
4312 	i = 0;
4313 	foreach(lc, args)
4314 	{
4315 		Node	   *arg = (Node *) lfirst(lc);
4316 
4317 		if (!IsA(arg, NamedArgExpr))
4318 		{
4319 			/* positional argument, assumed to precede all named args */
4320 			Assert(argarray[i] == NULL);
4321 			argarray[i++] = arg;
4322 		}
4323 		else
4324 		{
4325 			NamedArgExpr *na = (NamedArgExpr *) arg;
4326 
4327 			Assert(argarray[na->argnumber] == NULL);
4328 			argarray[na->argnumber] = (Node *) na->arg;
4329 		}
4330 	}
4331 
4332 	/*
4333 	 * Fetch default expressions, if needed, and insert into array at proper
4334 	 * locations (they aren't necessarily consecutive or all used)
4335 	 */
4336 	if (nargsprovided < pronargs)
4337 	{
4338 		List	   *defaults = fetch_function_defaults(func_tuple);
4339 
4340 		i = pronargs - funcform->pronargdefaults;
4341 		foreach(lc, defaults)
4342 		{
4343 			if (argarray[i] == NULL)
4344 				argarray[i] = (Node *) lfirst(lc);
4345 			i++;
4346 		}
4347 	}
4348 
4349 	/* Now reconstruct the args list in proper order */
4350 	args = NIL;
4351 	for (i = 0; i < pronargs; i++)
4352 	{
4353 		Assert(argarray[i] != NULL);
4354 		args = lappend(args, argarray[i]);
4355 	}
4356 
4357 	return args;
4358 }
4359 
4360 /*
4361  * add_function_defaults: add missing function arguments from its defaults
4362  *
4363  * This is used only when the argument list was positional to begin with,
4364  * and so we know we just need to add defaults at the end.
4365  */
4366 static List *
add_function_defaults(List * args,HeapTuple func_tuple)4367 add_function_defaults(List *args, HeapTuple func_tuple)
4368 {
4369 	Form_pg_proc funcform = (Form_pg_proc) GETSTRUCT(func_tuple);
4370 	int			nargsprovided = list_length(args);
4371 	List	   *defaults;
4372 	int			ndelete;
4373 
4374 	/* Get all the default expressions from the pg_proc tuple */
4375 	defaults = fetch_function_defaults(func_tuple);
4376 
4377 	/* Delete any unused defaults from the list */
4378 	ndelete = nargsprovided + list_length(defaults) - funcform->pronargs;
4379 	if (ndelete < 0)
4380 		elog(ERROR, "not enough default arguments");
4381 	while (ndelete-- > 0)
4382 		defaults = list_delete_first(defaults);
4383 
4384 	/* And form the combined argument list, not modifying the input list */
4385 	return list_concat(list_copy(args), defaults);
4386 }
4387 
4388 /*
4389  * fetch_function_defaults: get function's default arguments as expression list
4390  */
4391 static List *
fetch_function_defaults(HeapTuple func_tuple)4392 fetch_function_defaults(HeapTuple func_tuple)
4393 {
4394 	List	   *defaults;
4395 	Datum		proargdefaults;
4396 	bool		isnull;
4397 	char	   *str;
4398 
4399 	/* The error cases here shouldn't happen, but check anyway */
4400 	proargdefaults = SysCacheGetAttr(PROCOID, func_tuple,
4401 									 Anum_pg_proc_proargdefaults,
4402 									 &isnull);
4403 	if (isnull)
4404 		elog(ERROR, "not enough default arguments");
4405 	str = TextDatumGetCString(proargdefaults);
4406 	defaults = castNode(List, stringToNode(str));
4407 	pfree(str);
4408 	return defaults;
4409 }
4410 
4411 /*
4412  * recheck_cast_function_args: recheck function args and typecast as needed
4413  * after adding defaults.
4414  *
4415  * It is possible for some of the defaulted arguments to be polymorphic;
4416  * therefore we can't assume that the default expressions have the correct
4417  * data types already.  We have to re-resolve polymorphics and do coercion
4418  * just like the parser did.
4419  *
4420  * This should be a no-op if there are no polymorphic arguments,
4421  * but we do it anyway to be sure.
4422  *
4423  * Note: if any casts are needed, the args list is modified in-place;
4424  * caller should have already copied the list structure.
4425  */
4426 static void
recheck_cast_function_args(List * args,Oid result_type,HeapTuple func_tuple)4427 recheck_cast_function_args(List *args, Oid result_type, HeapTuple func_tuple)
4428 {
4429 	Form_pg_proc funcform = (Form_pg_proc) GETSTRUCT(func_tuple);
4430 	int			nargs;
4431 	Oid			actual_arg_types[FUNC_MAX_ARGS];
4432 	Oid			declared_arg_types[FUNC_MAX_ARGS];
4433 	Oid			rettype;
4434 	ListCell   *lc;
4435 
4436 	if (list_length(args) > FUNC_MAX_ARGS)
4437 		elog(ERROR, "too many function arguments");
4438 	nargs = 0;
4439 	foreach(lc, args)
4440 	{
4441 		actual_arg_types[nargs++] = exprType((Node *) lfirst(lc));
4442 	}
4443 	Assert(nargs == funcform->pronargs);
4444 	memcpy(declared_arg_types, funcform->proargtypes.values,
4445 		   funcform->pronargs * sizeof(Oid));
4446 	rettype = enforce_generic_type_consistency(actual_arg_types,
4447 											   declared_arg_types,
4448 											   nargs,
4449 											   funcform->prorettype,
4450 											   false);
4451 	/* let's just check we got the same answer as the parser did ... */
4452 	if (rettype != result_type)
4453 		elog(ERROR, "function's resolved result type changed during planning");
4454 
4455 	/* perform any necessary typecasting of arguments */
4456 	make_fn_arguments(NULL, args, actual_arg_types, declared_arg_types);
4457 }
4458 
4459 /*
4460  * evaluate_function: try to pre-evaluate a function call
4461  *
4462  * We can do this if the function is strict and has any constant-null inputs
4463  * (just return a null constant), or if the function is immutable and has all
4464  * constant inputs (call it and return the result as a Const node).  In
4465  * estimation mode we are willing to pre-evaluate stable functions too.
4466  *
4467  * Returns a simplified expression if successful, or NULL if cannot
4468  * simplify the function.
4469  */
4470 static Expr *
evaluate_function(Oid funcid,Oid result_type,int32 result_typmod,Oid result_collid,Oid input_collid,List * args,bool funcvariadic,HeapTuple func_tuple,eval_const_expressions_context * context)4471 evaluate_function(Oid funcid, Oid result_type, int32 result_typmod,
4472 				  Oid result_collid, Oid input_collid, List *args,
4473 				  bool funcvariadic,
4474 				  HeapTuple func_tuple,
4475 				  eval_const_expressions_context *context)
4476 {
4477 	Form_pg_proc funcform = (Form_pg_proc) GETSTRUCT(func_tuple);
4478 	bool		has_nonconst_input = false;
4479 	bool		has_null_input = false;
4480 	ListCell   *arg;
4481 	FuncExpr   *newexpr;
4482 
4483 	/*
4484 	 * Can't simplify if it returns a set.
4485 	 */
4486 	if (funcform->proretset)
4487 		return NULL;
4488 
4489 	/*
4490 	 * Can't simplify if it returns RECORD.  The immediate problem is that it
4491 	 * will be needing an expected tupdesc which we can't supply here.
4492 	 *
4493 	 * In the case where it has OUT parameters, it could get by without an
4494 	 * expected tupdesc, but we still have issues: get_expr_result_type()
4495 	 * doesn't know how to extract type info from a RECORD constant, and in
4496 	 * the case of a NULL function result there doesn't seem to be any clean
4497 	 * way to fix that.  In view of the likelihood of there being still other
4498 	 * gotchas, seems best to leave the function call unreduced.
4499 	 */
4500 	if (funcform->prorettype == RECORDOID)
4501 		return NULL;
4502 
4503 	/*
4504 	 * Check for constant inputs and especially constant-NULL inputs.
4505 	 */
4506 	foreach(arg, args)
4507 	{
4508 		if (IsA(lfirst(arg), Const))
4509 			has_null_input |= ((Const *) lfirst(arg))->constisnull;
4510 		else
4511 			has_nonconst_input = true;
4512 	}
4513 
4514 	/*
4515 	 * If the function is strict and has a constant-NULL input, it will never
4516 	 * be called at all, so we can replace the call by a NULL constant, even
4517 	 * if there are other inputs that aren't constant, and even if the
4518 	 * function is not otherwise immutable.
4519 	 */
4520 	if (funcform->proisstrict && has_null_input)
4521 		return (Expr *) makeNullConst(result_type, result_typmod,
4522 									  result_collid);
4523 
4524 	/*
4525 	 * Otherwise, can simplify only if all inputs are constants. (For a
4526 	 * non-strict function, constant NULL inputs are treated the same as
4527 	 * constant non-NULL inputs.)
4528 	 */
4529 	if (has_nonconst_input)
4530 		return NULL;
4531 
4532 	/*
4533 	 * Ordinarily we are only allowed to simplify immutable functions. But for
4534 	 * purposes of estimation, we consider it okay to simplify functions that
4535 	 * are merely stable; the risk that the result might change from planning
4536 	 * time to execution time is worth taking in preference to not being able
4537 	 * to estimate the value at all.
4538 	 */
4539 	if (funcform->provolatile == PROVOLATILE_IMMUTABLE)
4540 		 /* okay */ ;
4541 	else if (context->estimate && funcform->provolatile == PROVOLATILE_STABLE)
4542 		 /* okay */ ;
4543 	else
4544 		return NULL;
4545 
4546 	/*
4547 	 * OK, looks like we can simplify this operator/function.
4548 	 *
4549 	 * Build a new FuncExpr node containing the already-simplified arguments.
4550 	 */
4551 	newexpr = makeNode(FuncExpr);
4552 	newexpr->funcid = funcid;
4553 	newexpr->funcresulttype = result_type;
4554 	newexpr->funcretset = false;
4555 	newexpr->funcvariadic = funcvariadic;
4556 	newexpr->funcformat = COERCE_EXPLICIT_CALL; /* doesn't matter */
4557 	newexpr->funccollid = result_collid;	/* doesn't matter */
4558 	newexpr->inputcollid = input_collid;
4559 	newexpr->args = args;
4560 	newexpr->location = -1;
4561 
4562 	return evaluate_expr((Expr *) newexpr, result_type, result_typmod,
4563 						 result_collid);
4564 }
4565 
4566 /*
4567  * inline_function: try to expand a function call inline
4568  *
4569  * If the function is a sufficiently simple SQL-language function
4570  * (just "SELECT expression"), then we can inline it and avoid the rather
4571  * high per-call overhead of SQL functions.  Furthermore, this can expose
4572  * opportunities for constant-folding within the function expression.
4573  *
4574  * We have to beware of some special cases however.  A directly or
4575  * indirectly recursive function would cause us to recurse forever,
4576  * so we keep track of which functions we are already expanding and
4577  * do not re-expand them.  Also, if a parameter is used more than once
4578  * in the SQL-function body, we require it not to contain any volatile
4579  * functions (volatiles might deliver inconsistent answers) nor to be
4580  * unreasonably expensive to evaluate.  The expensiveness check not only
4581  * prevents us from doing multiple evaluations of an expensive parameter
4582  * at runtime, but is a safety value to limit growth of an expression due
4583  * to repeated inlining.
4584  *
4585  * We must also beware of changing the volatility or strictness status of
4586  * functions by inlining them.
4587  *
4588  * Also, at the moment we can't inline functions returning RECORD.  This
4589  * doesn't work in the general case because it discards information such
4590  * as OUT-parameter declarations.
4591  *
4592  * Also, context-dependent expression nodes in the argument list are trouble.
4593  *
4594  * Returns a simplified expression if successful, or NULL if cannot
4595  * simplify the function.
4596  */
4597 static Expr *
inline_function(Oid funcid,Oid result_type,Oid result_collid,Oid input_collid,List * args,bool funcvariadic,HeapTuple func_tuple,eval_const_expressions_context * context)4598 inline_function(Oid funcid, Oid result_type, Oid result_collid,
4599 				Oid input_collid, List *args,
4600 				bool funcvariadic,
4601 				HeapTuple func_tuple,
4602 				eval_const_expressions_context *context)
4603 {
4604 	Form_pg_proc funcform = (Form_pg_proc) GETSTRUCT(func_tuple);
4605 	char	   *src;
4606 	Datum		tmp;
4607 	bool		isNull;
4608 	bool		modifyTargetList;
4609 	MemoryContext oldcxt;
4610 	MemoryContext mycxt;
4611 	inline_error_callback_arg callback_arg;
4612 	ErrorContextCallback sqlerrcontext;
4613 	FuncExpr   *fexpr;
4614 	SQLFunctionParseInfoPtr pinfo;
4615 	ParseState *pstate;
4616 	List	   *raw_parsetree_list;
4617 	Query	   *querytree;
4618 	Node	   *newexpr;
4619 	int		   *usecounts;
4620 	ListCell   *arg;
4621 	int			i;
4622 
4623 	/*
4624 	 * Forget it if the function is not SQL-language or has other showstopper
4625 	 * properties.  (The prokind and nargs checks are just paranoia.)
4626 	 */
4627 	if (funcform->prolang != SQLlanguageId ||
4628 		funcform->prokind != PROKIND_FUNCTION ||
4629 		funcform->prosecdef ||
4630 		funcform->proretset ||
4631 		funcform->prorettype == RECORDOID ||
4632 		!heap_attisnull(func_tuple, Anum_pg_proc_proconfig, NULL) ||
4633 		funcform->pronargs != list_length(args))
4634 		return NULL;
4635 
4636 	/* Check for recursive function, and give up trying to expand if so */
4637 	if (list_member_oid(context->active_fns, funcid))
4638 		return NULL;
4639 
4640 	/* Check permission to call function (fail later, if not) */
4641 	if (pg_proc_aclcheck(funcid, GetUserId(), ACL_EXECUTE) != ACLCHECK_OK)
4642 		return NULL;
4643 
4644 	/* Check whether a plugin wants to hook function entry/exit */
4645 	if (FmgrHookIsNeeded(funcid))
4646 		return NULL;
4647 
4648 	/*
4649 	 * Make a temporary memory context, so that we don't leak all the stuff
4650 	 * that parsing might create.
4651 	 */
4652 	mycxt = AllocSetContextCreate(CurrentMemoryContext,
4653 								  "inline_function",
4654 								  ALLOCSET_DEFAULT_SIZES);
4655 	oldcxt = MemoryContextSwitchTo(mycxt);
4656 
4657 	/* Fetch the function body */
4658 	tmp = SysCacheGetAttr(PROCOID,
4659 						  func_tuple,
4660 						  Anum_pg_proc_prosrc,
4661 						  &isNull);
4662 	if (isNull)
4663 		elog(ERROR, "null prosrc for function %u", funcid);
4664 	src = TextDatumGetCString(tmp);
4665 
4666 	/*
4667 	 * Setup error traceback support for ereport().  This is so that we can
4668 	 * finger the function that bad information came from.
4669 	 */
4670 	callback_arg.proname = NameStr(funcform->proname);
4671 	callback_arg.prosrc = src;
4672 
4673 	sqlerrcontext.callback = sql_inline_error_callback;
4674 	sqlerrcontext.arg = (void *) &callback_arg;
4675 	sqlerrcontext.previous = error_context_stack;
4676 	error_context_stack = &sqlerrcontext;
4677 
4678 	/*
4679 	 * Set up to handle parameters while parsing the function body.  We need a
4680 	 * dummy FuncExpr node containing the already-simplified arguments to pass
4681 	 * to prepare_sql_fn_parse_info.  (It is really only needed if there are
4682 	 * some polymorphic arguments, but for simplicity we always build it.)
4683 	 */
4684 	fexpr = makeNode(FuncExpr);
4685 	fexpr->funcid = funcid;
4686 	fexpr->funcresulttype = result_type;
4687 	fexpr->funcretset = false;
4688 	fexpr->funcvariadic = funcvariadic;
4689 	fexpr->funcformat = COERCE_EXPLICIT_CALL;	/* doesn't matter */
4690 	fexpr->funccollid = result_collid;	/* doesn't matter */
4691 	fexpr->inputcollid = input_collid;
4692 	fexpr->args = args;
4693 	fexpr->location = -1;
4694 
4695 	pinfo = prepare_sql_fn_parse_info(func_tuple,
4696 									  (Node *) fexpr,
4697 									  input_collid);
4698 
4699 	/*
4700 	 * We just do parsing and parse analysis, not rewriting, because rewriting
4701 	 * will not affect table-free-SELECT-only queries, which is all that we
4702 	 * care about.  Also, we can punt as soon as we detect more than one
4703 	 * command in the function body.
4704 	 */
4705 	raw_parsetree_list = pg_parse_query(src);
4706 	if (list_length(raw_parsetree_list) != 1)
4707 		goto fail;
4708 
4709 	pstate = make_parsestate(NULL);
4710 	pstate->p_sourcetext = src;
4711 	sql_fn_parser_setup(pstate, pinfo);
4712 
4713 	querytree = transformTopLevelStmt(pstate, linitial(raw_parsetree_list));
4714 
4715 	free_parsestate(pstate);
4716 
4717 	/*
4718 	 * The single command must be a simple "SELECT expression".
4719 	 *
4720 	 * Note: if you change the tests involved in this, see also plpgsql's
4721 	 * exec_simple_check_plan().  That generally needs to have the same idea
4722 	 * of what's a "simple expression", so that inlining a function that
4723 	 * previously wasn't inlined won't change plpgsql's conclusion.
4724 	 */
4725 	if (!IsA(querytree, Query) ||
4726 		querytree->commandType != CMD_SELECT ||
4727 		querytree->hasAggs ||
4728 		querytree->hasWindowFuncs ||
4729 		querytree->hasTargetSRFs ||
4730 		querytree->hasSubLinks ||
4731 		querytree->cteList ||
4732 		querytree->rtable ||
4733 		querytree->jointree->fromlist ||
4734 		querytree->jointree->quals ||
4735 		querytree->groupClause ||
4736 		querytree->groupingSets ||
4737 		querytree->havingQual ||
4738 		querytree->windowClause ||
4739 		querytree->distinctClause ||
4740 		querytree->sortClause ||
4741 		querytree->limitOffset ||
4742 		querytree->limitCount ||
4743 		querytree->setOperations ||
4744 		list_length(querytree->targetList) != 1)
4745 		goto fail;
4746 
4747 	/*
4748 	 * Make sure the function (still) returns what it's declared to.  This
4749 	 * will raise an error if wrong, but that's okay since the function would
4750 	 * fail at runtime anyway.  Note that check_sql_fn_retval will also insert
4751 	 * a RelabelType if needed to make the tlist expression match the declared
4752 	 * type of the function.
4753 	 *
4754 	 * Note: we do not try this until we have verified that no rewriting was
4755 	 * needed; that's probably not important, but let's be careful.
4756 	 */
4757 	if (check_sql_fn_retval(funcid, result_type, list_make1(querytree),
4758 							&modifyTargetList, NULL))
4759 		goto fail;				/* reject whole-tuple-result cases */
4760 
4761 	/* Now we can grab the tlist expression */
4762 	newexpr = (Node *) ((TargetEntry *) linitial(querytree->targetList))->expr;
4763 
4764 	/*
4765 	 * If the SQL function returns VOID, we can only inline it if it is a
4766 	 * SELECT of an expression returning VOID (ie, it's just a redirection to
4767 	 * another VOID-returning function).  In all non-VOID-returning cases,
4768 	 * check_sql_fn_retval should ensure that newexpr returns the function's
4769 	 * declared result type, so this test shouldn't fail otherwise; but we may
4770 	 * as well cope gracefully if it does.
4771 	 */
4772 	if (exprType(newexpr) != result_type)
4773 		goto fail;
4774 
4775 	/* check_sql_fn_retval couldn't have made any dangerous tlist changes */
4776 	Assert(!modifyTargetList);
4777 
4778 	/*
4779 	 * Additional validity checks on the expression.  It mustn't be more
4780 	 * volatile than the surrounding function (this is to avoid breaking hacks
4781 	 * that involve pretending a function is immutable when it really ain't).
4782 	 * If the surrounding function is declared strict, then the expression
4783 	 * must contain only strict constructs and must use all of the function
4784 	 * parameters (this is overkill, but an exact analysis is hard).
4785 	 */
4786 	if (funcform->provolatile == PROVOLATILE_IMMUTABLE &&
4787 		contain_mutable_functions(newexpr))
4788 		goto fail;
4789 	else if (funcform->provolatile == PROVOLATILE_STABLE &&
4790 			 contain_volatile_functions(newexpr))
4791 		goto fail;
4792 
4793 	if (funcform->proisstrict &&
4794 		contain_nonstrict_functions(newexpr))
4795 		goto fail;
4796 
4797 	/*
4798 	 * If any parameter expression contains a context-dependent node, we can't
4799 	 * inline, for fear of putting such a node into the wrong context.
4800 	 */
4801 	if (contain_context_dependent_node((Node *) args))
4802 		goto fail;
4803 
4804 	/*
4805 	 * We may be able to do it; there are still checks on parameter usage to
4806 	 * make, but those are most easily done in combination with the actual
4807 	 * substitution of the inputs.  So start building expression with inputs
4808 	 * substituted.
4809 	 */
4810 	usecounts = (int *) palloc0(funcform->pronargs * sizeof(int));
4811 	newexpr = substitute_actual_parameters(newexpr, funcform->pronargs,
4812 										   args, usecounts);
4813 
4814 	/* Now check for parameter usage */
4815 	i = 0;
4816 	foreach(arg, args)
4817 	{
4818 		Node	   *param = lfirst(arg);
4819 
4820 		if (usecounts[i] == 0)
4821 		{
4822 			/* Param not used at all: uncool if func is strict */
4823 			if (funcform->proisstrict)
4824 				goto fail;
4825 		}
4826 		else if (usecounts[i] != 1)
4827 		{
4828 			/* Param used multiple times: uncool if expensive or volatile */
4829 			QualCost	eval_cost;
4830 
4831 			/*
4832 			 * We define "expensive" as "contains any subplan or more than 10
4833 			 * operators".  Note that the subplan search has to be done
4834 			 * explicitly, since cost_qual_eval() will barf on unplanned
4835 			 * subselects.
4836 			 */
4837 			if (contain_subplans(param))
4838 				goto fail;
4839 			cost_qual_eval(&eval_cost, list_make1(param), NULL);
4840 			if (eval_cost.startup + eval_cost.per_tuple >
4841 				10 * cpu_operator_cost)
4842 				goto fail;
4843 
4844 			/*
4845 			 * Check volatility last since this is more expensive than the
4846 			 * above tests
4847 			 */
4848 			if (contain_volatile_functions(param))
4849 				goto fail;
4850 		}
4851 		i++;
4852 	}
4853 
4854 	/*
4855 	 * Whew --- we can make the substitution.  Copy the modified expression
4856 	 * out of the temporary memory context, and clean up.
4857 	 */
4858 	MemoryContextSwitchTo(oldcxt);
4859 
4860 	newexpr = copyObject(newexpr);
4861 
4862 	MemoryContextDelete(mycxt);
4863 
4864 	/*
4865 	 * If the result is of a collatable type, force the result to expose the
4866 	 * correct collation.  In most cases this does not matter, but it's
4867 	 * possible that the function result is used directly as a sort key or in
4868 	 * other places where we expect exprCollation() to tell the truth.
4869 	 */
4870 	if (OidIsValid(result_collid))
4871 	{
4872 		Oid			exprcoll = exprCollation(newexpr);
4873 
4874 		if (OidIsValid(exprcoll) && exprcoll != result_collid)
4875 		{
4876 			CollateExpr *newnode = makeNode(CollateExpr);
4877 
4878 			newnode->arg = (Expr *) newexpr;
4879 			newnode->collOid = result_collid;
4880 			newnode->location = -1;
4881 
4882 			newexpr = (Node *) newnode;
4883 		}
4884 	}
4885 
4886 	/*
4887 	 * Since there is now no trace of the function in the plan tree, we must
4888 	 * explicitly record the plan's dependency on the function.
4889 	 */
4890 	if (context->root)
4891 		record_plan_function_dependency(context->root, funcid);
4892 
4893 	/*
4894 	 * Recursively try to simplify the modified expression.  Here we must add
4895 	 * the current function to the context list of active functions.
4896 	 */
4897 	context->active_fns = lcons_oid(funcid, context->active_fns);
4898 	newexpr = eval_const_expressions_mutator(newexpr, context);
4899 	context->active_fns = list_delete_first(context->active_fns);
4900 
4901 	error_context_stack = sqlerrcontext.previous;
4902 
4903 	return (Expr *) newexpr;
4904 
4905 	/* Here if func is not inlinable: release temp memory and return NULL */
4906 fail:
4907 	MemoryContextSwitchTo(oldcxt);
4908 	MemoryContextDelete(mycxt);
4909 	error_context_stack = sqlerrcontext.previous;
4910 
4911 	return NULL;
4912 }
4913 
4914 /*
4915  * Replace Param nodes by appropriate actual parameters
4916  */
4917 static Node *
substitute_actual_parameters(Node * expr,int nargs,List * args,int * usecounts)4918 substitute_actual_parameters(Node *expr, int nargs, List *args,
4919 							 int *usecounts)
4920 {
4921 	substitute_actual_parameters_context context;
4922 
4923 	context.nargs = nargs;
4924 	context.args = args;
4925 	context.usecounts = usecounts;
4926 
4927 	return substitute_actual_parameters_mutator(expr, &context);
4928 }
4929 
4930 static Node *
substitute_actual_parameters_mutator(Node * node,substitute_actual_parameters_context * context)4931 substitute_actual_parameters_mutator(Node *node,
4932 									 substitute_actual_parameters_context *context)
4933 {
4934 	if (node == NULL)
4935 		return NULL;
4936 	if (IsA(node, Param))
4937 	{
4938 		Param	   *param = (Param *) node;
4939 
4940 		if (param->paramkind != PARAM_EXTERN)
4941 			elog(ERROR, "unexpected paramkind: %d", (int) param->paramkind);
4942 		if (param->paramid <= 0 || param->paramid > context->nargs)
4943 			elog(ERROR, "invalid paramid: %d", param->paramid);
4944 
4945 		/* Count usage of parameter */
4946 		context->usecounts[param->paramid - 1]++;
4947 
4948 		/* Select the appropriate actual arg and replace the Param with it */
4949 		/* We don't need to copy at this time (it'll get done later) */
4950 		return list_nth(context->args, param->paramid - 1);
4951 	}
4952 	return expression_tree_mutator(node, substitute_actual_parameters_mutator,
4953 								   (void *) context);
4954 }
4955 
4956 /*
4957  * error context callback to let us supply a call-stack traceback
4958  */
4959 static void
sql_inline_error_callback(void * arg)4960 sql_inline_error_callback(void *arg)
4961 {
4962 	inline_error_callback_arg *callback_arg = (inline_error_callback_arg *) arg;
4963 	int			syntaxerrposition;
4964 
4965 	/* If it's a syntax error, convert to internal syntax error report */
4966 	syntaxerrposition = geterrposition();
4967 	if (syntaxerrposition > 0)
4968 	{
4969 		errposition(0);
4970 		internalerrposition(syntaxerrposition);
4971 		internalerrquery(callback_arg->prosrc);
4972 	}
4973 
4974 	errcontext("SQL function \"%s\" during inlining", callback_arg->proname);
4975 }
4976 
4977 /*
4978  * evaluate_expr: pre-evaluate a constant expression
4979  *
4980  * We use the executor's routine ExecEvalExpr() to avoid duplication of
4981  * code and ensure we get the same result as the executor would get.
4982  */
4983 static Expr *
evaluate_expr(Expr * expr,Oid result_type,int32 result_typmod,Oid result_collation)4984 evaluate_expr(Expr *expr, Oid result_type, int32 result_typmod,
4985 			  Oid result_collation)
4986 {
4987 	EState	   *estate;
4988 	ExprState  *exprstate;
4989 	MemoryContext oldcontext;
4990 	Datum		const_val;
4991 	bool		const_is_null;
4992 	int16		resultTypLen;
4993 	bool		resultTypByVal;
4994 
4995 	/*
4996 	 * To use the executor, we need an EState.
4997 	 */
4998 	estate = CreateExecutorState();
4999 
5000 	/* We can use the estate's working context to avoid memory leaks. */
5001 	oldcontext = MemoryContextSwitchTo(estate->es_query_cxt);
5002 
5003 	/* Make sure any opfuncids are filled in. */
5004 	fix_opfuncids((Node *) expr);
5005 
5006 	/*
5007 	 * Prepare expr for execution.  (Note: we can't use ExecPrepareExpr
5008 	 * because it'd result in recursively invoking eval_const_expressions.)
5009 	 */
5010 	exprstate = ExecInitExpr(expr, NULL);
5011 
5012 	/*
5013 	 * And evaluate it.
5014 	 *
5015 	 * It is OK to use a default econtext because none of the ExecEvalExpr()
5016 	 * code used in this situation will use econtext.  That might seem
5017 	 * fortuitous, but it's not so unreasonable --- a constant expression does
5018 	 * not depend on context, by definition, n'est ce pas?
5019 	 */
5020 	const_val = ExecEvalExprSwitchContext(exprstate,
5021 										  GetPerTupleExprContext(estate),
5022 										  &const_is_null);
5023 
5024 	/* Get info needed about result datatype */
5025 	get_typlenbyval(result_type, &resultTypLen, &resultTypByVal);
5026 
5027 	/* Get back to outer memory context */
5028 	MemoryContextSwitchTo(oldcontext);
5029 
5030 	/*
5031 	 * Must copy result out of sub-context used by expression eval.
5032 	 *
5033 	 * Also, if it's varlena, forcibly detoast it.  This protects us against
5034 	 * storing TOAST pointers into plans that might outlive the referenced
5035 	 * data.  (makeConst would handle detoasting anyway, but it's worth a few
5036 	 * extra lines here so that we can do the copy and detoast in one step.)
5037 	 */
5038 	if (!const_is_null)
5039 	{
5040 		if (resultTypLen == -1)
5041 			const_val = PointerGetDatum(PG_DETOAST_DATUM_COPY(const_val));
5042 		else
5043 			const_val = datumCopy(const_val, resultTypByVal, resultTypLen);
5044 	}
5045 
5046 	/* Release all the junk we just created */
5047 	FreeExecutorState(estate);
5048 
5049 	/*
5050 	 * Make the constant result node.
5051 	 */
5052 	return (Expr *) makeConst(result_type, result_typmod, result_collation,
5053 							  resultTypLen,
5054 							  const_val, const_is_null,
5055 							  resultTypByVal);
5056 }
5057 
5058 
5059 /*
5060  * inline_set_returning_function
5061  *		Attempt to "inline" a set-returning function in the FROM clause.
5062  *
5063  * "rte" is an RTE_FUNCTION rangetable entry.  If it represents a call of a
5064  * set-returning SQL function that can safely be inlined, expand the function
5065  * and return the substitute Query structure.  Otherwise, return NULL.
5066  *
5067  * This has a good deal of similarity to inline_function(), but that's
5068  * for the non-set-returning case, and there are enough differences to
5069  * justify separate functions.
5070  */
5071 Query *
inline_set_returning_function(PlannerInfo * root,RangeTblEntry * rte)5072 inline_set_returning_function(PlannerInfo *root, RangeTblEntry *rte)
5073 {
5074 	RangeTblFunction *rtfunc;
5075 	FuncExpr   *fexpr;
5076 	Oid			func_oid;
5077 	HeapTuple	func_tuple;
5078 	Form_pg_proc funcform;
5079 	char	   *src;
5080 	Datum		tmp;
5081 	bool		isNull;
5082 	bool		modifyTargetList;
5083 	MemoryContext oldcxt;
5084 	MemoryContext mycxt;
5085 	List	   *saveInvalItems;
5086 	inline_error_callback_arg callback_arg;
5087 	ErrorContextCallback sqlerrcontext;
5088 	SQLFunctionParseInfoPtr pinfo;
5089 	List	   *raw_parsetree_list;
5090 	List	   *querytree_list;
5091 	Query	   *querytree;
5092 
5093 	Assert(rte->rtekind == RTE_FUNCTION);
5094 
5095 	/*
5096 	 * It doesn't make a lot of sense for a SQL SRF to refer to itself in its
5097 	 * own FROM clause, since that must cause infinite recursion at runtime.
5098 	 * It will cause this code to recurse too, so check for stack overflow.
5099 	 * (There's no need to do more.)
5100 	 */
5101 	check_stack_depth();
5102 
5103 	/* Fail if the RTE has ORDINALITY - we don't implement that here. */
5104 	if (rte->funcordinality)
5105 		return NULL;
5106 
5107 	/* Fail if RTE isn't a single, simple FuncExpr */
5108 	if (list_length(rte->functions) != 1)
5109 		return NULL;
5110 	rtfunc = (RangeTblFunction *) linitial(rte->functions);
5111 
5112 	if (!IsA(rtfunc->funcexpr, FuncExpr))
5113 		return NULL;
5114 	fexpr = (FuncExpr *) rtfunc->funcexpr;
5115 
5116 	func_oid = fexpr->funcid;
5117 
5118 	/*
5119 	 * The function must be declared to return a set, else inlining would
5120 	 * change the results if the contained SELECT didn't return exactly one
5121 	 * row.
5122 	 */
5123 	if (!fexpr->funcretset)
5124 		return NULL;
5125 
5126 	/*
5127 	 * Refuse to inline if the arguments contain any volatile functions or
5128 	 * sub-selects.  Volatile functions are rejected because inlining may
5129 	 * result in the arguments being evaluated multiple times, risking a
5130 	 * change in behavior.  Sub-selects are rejected partly for implementation
5131 	 * reasons (pushing them down another level might change their behavior)
5132 	 * and partly because they're likely to be expensive and so multiple
5133 	 * evaluation would be bad.
5134 	 */
5135 	if (contain_volatile_functions((Node *) fexpr->args) ||
5136 		contain_subplans((Node *) fexpr->args))
5137 		return NULL;
5138 
5139 	/* Check permission to call function (fail later, if not) */
5140 	if (pg_proc_aclcheck(func_oid, GetUserId(), ACL_EXECUTE) != ACLCHECK_OK)
5141 		return NULL;
5142 
5143 	/* Check whether a plugin wants to hook function entry/exit */
5144 	if (FmgrHookIsNeeded(func_oid))
5145 		return NULL;
5146 
5147 	/*
5148 	 * OK, let's take a look at the function's pg_proc entry.
5149 	 */
5150 	func_tuple = SearchSysCache1(PROCOID, ObjectIdGetDatum(func_oid));
5151 	if (!HeapTupleIsValid(func_tuple))
5152 		elog(ERROR, "cache lookup failed for function %u", func_oid);
5153 	funcform = (Form_pg_proc) GETSTRUCT(func_tuple);
5154 
5155 	/*
5156 	 * Forget it if the function is not SQL-language or has other showstopper
5157 	 * properties.  In particular it mustn't be declared STRICT, since we
5158 	 * couldn't enforce that.  It also mustn't be VOLATILE, because that is
5159 	 * supposed to cause it to be executed with its own snapshot, rather than
5160 	 * sharing the snapshot of the calling query.  We also disallow returning
5161 	 * SETOF VOID, because inlining would result in exposing the actual result
5162 	 * of the function's last SELECT, which should not happen in that case.
5163 	 * (Rechecking prokind and proretset is just paranoia.)
5164 	 */
5165 	if (funcform->prolang != SQLlanguageId ||
5166 		funcform->prokind != PROKIND_FUNCTION ||
5167 		funcform->proisstrict ||
5168 		funcform->provolatile == PROVOLATILE_VOLATILE ||
5169 		funcform->prorettype == VOIDOID ||
5170 		funcform->prosecdef ||
5171 		!funcform->proretset ||
5172 		!heap_attisnull(func_tuple, Anum_pg_proc_proconfig, NULL))
5173 	{
5174 		ReleaseSysCache(func_tuple);
5175 		return NULL;
5176 	}
5177 
5178 	/*
5179 	 * Make a temporary memory context, so that we don't leak all the stuff
5180 	 * that parsing might create.
5181 	 */
5182 	mycxt = AllocSetContextCreate(CurrentMemoryContext,
5183 								  "inline_set_returning_function",
5184 								  ALLOCSET_DEFAULT_SIZES);
5185 	oldcxt = MemoryContextSwitchTo(mycxt);
5186 
5187 	/*
5188 	 * When we call eval_const_expressions below, it might try to add items to
5189 	 * root->glob->invalItems.  Since it is running in the temp context, those
5190 	 * items will be in that context, and will need to be copied out if we're
5191 	 * successful.  Temporarily reset the list so that we can keep those items
5192 	 * separate from the pre-existing list contents.
5193 	 */
5194 	saveInvalItems = root->glob->invalItems;
5195 	root->glob->invalItems = NIL;
5196 
5197 	/* Fetch the function body */
5198 	tmp = SysCacheGetAttr(PROCOID,
5199 						  func_tuple,
5200 						  Anum_pg_proc_prosrc,
5201 						  &isNull);
5202 	if (isNull)
5203 		elog(ERROR, "null prosrc for function %u", func_oid);
5204 	src = TextDatumGetCString(tmp);
5205 
5206 	/*
5207 	 * Setup error traceback support for ereport().  This is so that we can
5208 	 * finger the function that bad information came from.
5209 	 */
5210 	callback_arg.proname = NameStr(funcform->proname);
5211 	callback_arg.prosrc = src;
5212 
5213 	sqlerrcontext.callback = sql_inline_error_callback;
5214 	sqlerrcontext.arg = (void *) &callback_arg;
5215 	sqlerrcontext.previous = error_context_stack;
5216 	error_context_stack = &sqlerrcontext;
5217 
5218 	/*
5219 	 * Run eval_const_expressions on the function call.  This is necessary to
5220 	 * ensure that named-argument notation is converted to positional notation
5221 	 * and any default arguments are inserted.  It's a bit of overkill for the
5222 	 * arguments, since they'll get processed again later, but no harm will be
5223 	 * done.
5224 	 */
5225 	fexpr = (FuncExpr *) eval_const_expressions(root, (Node *) fexpr);
5226 
5227 	/* It should still be a call of the same function, but let's check */
5228 	if (!IsA(fexpr, FuncExpr) ||
5229 		fexpr->funcid != func_oid)
5230 		goto fail;
5231 
5232 	/* Arg list length should now match the function */
5233 	if (list_length(fexpr->args) != funcform->pronargs)
5234 		goto fail;
5235 
5236 	/*
5237 	 * Set up to handle parameters while parsing the function body.  We can
5238 	 * use the FuncExpr just created as the input for
5239 	 * prepare_sql_fn_parse_info.
5240 	 */
5241 	pinfo = prepare_sql_fn_parse_info(func_tuple,
5242 									  (Node *) fexpr,
5243 									  fexpr->inputcollid);
5244 
5245 	/*
5246 	 * Parse, analyze, and rewrite (unlike inline_function(), we can't skip
5247 	 * rewriting here).  We can fail as soon as we find more than one query,
5248 	 * though.
5249 	 */
5250 	raw_parsetree_list = pg_parse_query(src);
5251 	if (list_length(raw_parsetree_list) != 1)
5252 		goto fail;
5253 
5254 	querytree_list = pg_analyze_and_rewrite_params(linitial(raw_parsetree_list),
5255 												   src,
5256 												   (ParserSetupHook) sql_fn_parser_setup,
5257 												   pinfo, NULL);
5258 	if (list_length(querytree_list) != 1)
5259 		goto fail;
5260 	querytree = linitial(querytree_list);
5261 
5262 	/*
5263 	 * The single command must be a plain SELECT.
5264 	 */
5265 	if (!IsA(querytree, Query) ||
5266 		querytree->commandType != CMD_SELECT)
5267 		goto fail;
5268 
5269 	/*
5270 	 * Make sure the function (still) returns what it's declared to.  This
5271 	 * will raise an error if wrong, but that's okay since the function would
5272 	 * fail at runtime anyway.  Note that check_sql_fn_retval will also insert
5273 	 * RelabelType(s) and/or NULL columns if needed to make the tlist
5274 	 * expression(s) match the declared type of the function.
5275 	 *
5276 	 * If the function returns a composite type, don't inline unless the check
5277 	 * shows it's returning a whole tuple result; otherwise what it's
5278 	 * returning is a single composite column which is not what we need. (Like
5279 	 * check_sql_fn_retval, we deliberately exclude domains over composite
5280 	 * here.)
5281 	 */
5282 	if (!check_sql_fn_retval(func_oid, fexpr->funcresulttype,
5283 							 querytree_list,
5284 							 &modifyTargetList, NULL) &&
5285 		(get_typtype(fexpr->funcresulttype) == TYPTYPE_COMPOSITE ||
5286 		 fexpr->funcresulttype == RECORDOID))
5287 		goto fail;				/* reject not-whole-tuple-result cases */
5288 
5289 	/*
5290 	 * If we had to modify the tlist to make it match, and the statement is
5291 	 * one in which changing the tlist contents could change semantics, we
5292 	 * have to punt and not inline.
5293 	 */
5294 	if (modifyTargetList)
5295 		goto fail;
5296 
5297 	/*
5298 	 * If it returns RECORD, we have to check against the column type list
5299 	 * provided in the RTE; check_sql_fn_retval can't do that.  (If no match,
5300 	 * we just fail to inline, rather than complaining; see notes for
5301 	 * tlist_matches_coltypelist.)	We don't have to do this for functions
5302 	 * with declared OUT parameters, even though their funcresulttype is
5303 	 * RECORDOID, so check get_func_result_type too.
5304 	 */
5305 	if (fexpr->funcresulttype == RECORDOID &&
5306 		get_func_result_type(func_oid, NULL, NULL) == TYPEFUNC_RECORD &&
5307 		!tlist_matches_coltypelist(querytree->targetList,
5308 								   rtfunc->funccoltypes))
5309 		goto fail;
5310 
5311 	/*
5312 	 * Looks good --- substitute parameters into the query.
5313 	 */
5314 	querytree = substitute_actual_srf_parameters(querytree,
5315 												 funcform->pronargs,
5316 												 fexpr->args);
5317 
5318 	/*
5319 	 * Copy the modified query out of the temporary memory context, and clean
5320 	 * up.
5321 	 */
5322 	MemoryContextSwitchTo(oldcxt);
5323 
5324 	querytree = copyObject(querytree);
5325 
5326 	/* copy up any new invalItems, too */
5327 	root->glob->invalItems = list_concat(saveInvalItems,
5328 										 copyObject(root->glob->invalItems));
5329 
5330 	MemoryContextDelete(mycxt);
5331 	error_context_stack = sqlerrcontext.previous;
5332 	ReleaseSysCache(func_tuple);
5333 
5334 	/*
5335 	 * We don't have to fix collations here because the upper query is already
5336 	 * parsed, ie, the collations in the RTE are what count.
5337 	 */
5338 
5339 	/*
5340 	 * Since there is now no trace of the function in the plan tree, we must
5341 	 * explicitly record the plan's dependency on the function.
5342 	 */
5343 	record_plan_function_dependency(root, func_oid);
5344 
5345 	return querytree;
5346 
5347 	/* Here if func is not inlinable: release temp memory and return NULL */
5348 fail:
5349 	MemoryContextSwitchTo(oldcxt);
5350 	root->glob->invalItems = saveInvalItems;
5351 	MemoryContextDelete(mycxt);
5352 	error_context_stack = sqlerrcontext.previous;
5353 	ReleaseSysCache(func_tuple);
5354 
5355 	return NULL;
5356 }
5357 
5358 /*
5359  * Replace Param nodes by appropriate actual parameters
5360  *
5361  * This is just enough different from substitute_actual_parameters()
5362  * that it needs its own code.
5363  */
5364 static Query *
substitute_actual_srf_parameters(Query * expr,int nargs,List * args)5365 substitute_actual_srf_parameters(Query *expr, int nargs, List *args)
5366 {
5367 	substitute_actual_srf_parameters_context context;
5368 
5369 	context.nargs = nargs;
5370 	context.args = args;
5371 	context.sublevels_up = 1;
5372 
5373 	return query_tree_mutator(expr,
5374 							  substitute_actual_srf_parameters_mutator,
5375 							  &context,
5376 							  0);
5377 }
5378 
5379 static Node *
substitute_actual_srf_parameters_mutator(Node * node,substitute_actual_srf_parameters_context * context)5380 substitute_actual_srf_parameters_mutator(Node *node,
5381 										 substitute_actual_srf_parameters_context *context)
5382 {
5383 	Node	   *result;
5384 
5385 	if (node == NULL)
5386 		return NULL;
5387 	if (IsA(node, Query))
5388 	{
5389 		context->sublevels_up++;
5390 		result = (Node *) query_tree_mutator((Query *) node,
5391 											 substitute_actual_srf_parameters_mutator,
5392 											 (void *) context,
5393 											 0);
5394 		context->sublevels_up--;
5395 		return result;
5396 	}
5397 	if (IsA(node, Param))
5398 	{
5399 		Param	   *param = (Param *) node;
5400 
5401 		if (param->paramkind == PARAM_EXTERN)
5402 		{
5403 			if (param->paramid <= 0 || param->paramid > context->nargs)
5404 				elog(ERROR, "invalid paramid: %d", param->paramid);
5405 
5406 			/*
5407 			 * Since the parameter is being inserted into a subquery, we must
5408 			 * adjust levels.
5409 			 */
5410 			result = copyObject(list_nth(context->args, param->paramid - 1));
5411 			IncrementVarSublevelsUp(result, context->sublevels_up, 0);
5412 			return result;
5413 		}
5414 	}
5415 	return expression_tree_mutator(node,
5416 								   substitute_actual_srf_parameters_mutator,
5417 								   (void *) context);
5418 }
5419 
5420 /*
5421  * Check whether a SELECT targetlist emits the specified column types,
5422  * to see if it's safe to inline a function returning record.
5423  *
5424  * We insist on exact match here.  The executor allows binary-coercible
5425  * cases too, but we don't have a way to preserve the correct column types
5426  * in the correct places if we inline the function in such a case.
5427  *
5428  * Note that we only check type OIDs not typmods; this agrees with what the
5429  * executor would do at runtime, and attributing a specific typmod to a
5430  * function result is largely wishful thinking anyway.
5431  */
5432 static bool
tlist_matches_coltypelist(List * tlist,List * coltypelist)5433 tlist_matches_coltypelist(List *tlist, List *coltypelist)
5434 {
5435 	ListCell   *tlistitem;
5436 	ListCell   *clistitem;
5437 
5438 	clistitem = list_head(coltypelist);
5439 	foreach(tlistitem, tlist)
5440 	{
5441 		TargetEntry *tle = (TargetEntry *) lfirst(tlistitem);
5442 		Oid			coltype;
5443 
5444 		if (tle->resjunk)
5445 			continue;			/* ignore junk columns */
5446 
5447 		if (clistitem == NULL)
5448 			return false;		/* too many tlist items */
5449 
5450 		coltype = lfirst_oid(clistitem);
5451 		clistitem = lnext(clistitem);
5452 
5453 		if (exprType((Node *) tle->expr) != coltype)
5454 			return false;		/* column type mismatch */
5455 	}
5456 
5457 	if (clistitem != NULL)
5458 		return false;			/* too few tlist items */
5459 
5460 	return true;
5461 }
5462