Create a "sort support" interface API for faster sorting.

This patch creates an API whereby a btree index opclass can optionally provide non-SQL-callable support functions for sorting. In the initial patch, we only use this to provide a directly-callable comparator function, which can be invoked with a bit less overhead than the traditional SQL-callable comparator. While that should be of value in itself, the real reason for doing this is to provide a datatype-extensible framework for more aggressive optimizations, as in Peter Geoghegan's recent work. Robert Haas and Tom Lane
2025-10-16 17:07:43 +03:00 · 2011-12-07 00:18:38 -05:00
parent d2a662182e
commit c6e3ac11b6
30 changed files with 869 additions and 421 deletions
--- a/src/backend/commands/analyze.c
+++ b/src/backend/commands/analyze.c
@@ -47,8 +47,8 @@
 #include "utils/lsyscache.h"
 #include "utils/memutils.h"
 #include "utils/pg_rusage.h"
+#include "utils/sortsupport.h"
 #include "utils/syscache.h"
-#include "utils/tuplesort.h"
 #include "utils/timestamp.h"
 #include "utils/tqual.h"

@@ -1774,8 +1774,7 @@ typedef struct

 typedef struct
 {
-	FmgrInfo   *cmpFn;
-	int			cmpFlags;
+	SortSupport ssup;
 	int		   *tupnoLink;
 } CompareScalarsContext;

@@ -2222,9 +2221,7 @@ compute_scalar_stats(VacAttrStatsP stats,
 	bool		is_varwidth = (!stats->attrtype->typbyval &&
 							   stats->attrtype->typlen < 0);
 	double		corr_xysum;
-	Oid			cmpFn;
-	int			cmpFlags;
-	FmgrInfo	f_cmpfn;
+	SortSupportData ssup;
 	ScalarItem *values;
 	int			values_cnt = 0;
 	int		   *tupnoLink;
@@ -2238,8 +2235,13 @@ compute_scalar_stats(VacAttrStatsP stats,
 	tupnoLink = (int *) palloc(samplerows * sizeof(int));
 	track = (ScalarMCVItem *) palloc(num_mcv * sizeof(ScalarMCVItem));

-	SelectSortFunction(mystats->ltopr, false, &cmpFn, &cmpFlags);
-	fmgr_info(cmpFn, &f_cmpfn);
+	memset(&ssup, 0, sizeof(ssup));
+	ssup.ssup_cxt = CurrentMemoryContext;
+	/* We always use the default collation for statistics */
+	ssup.ssup_collation = DEFAULT_COLLATION_OID;
+	ssup.ssup_nulls_first = false;
+
+	PrepareSortSupportFromOrderingOp(mystats->ltopr, &ssup);

 	/* Initial scan to find sortable values */
 	for (i = 0; i < samplerows; i++)
@@ -2307,8 +2309,7 @@ compute_scalar_stats(VacAttrStatsP stats,
 		CompareScalarsContext cxt;

 		/* Sort the collected values */
-		cxt.cmpFn = &f_cmpfn;
-		cxt.cmpFlags = cmpFlags;
+		cxt.ssup = &ssup;
 		cxt.tupnoLink = tupnoLink;
 		qsort_arg((void *) values, values_cnt, sizeof(ScalarItem),
 				  compare_scalars, (void *) &cxt);
@@ -2712,12 +2713,9 @@ compare_scalars(const void *a, const void *b, void *arg)
 	Datum		db = ((const ScalarItem *) b)->value;
 	int			tb = ((const ScalarItem *) b)->tupno;
 	CompareScalarsContext *cxt = (CompareScalarsContext *) arg;
-	int32		compare;
+	int			compare;

-	/* We always use the default collation for statistics */
-	compare = ApplySortFunction(cxt->cmpFn, cxt->cmpFlags,
-								DEFAULT_COLLATION_OID,
-								da, false, db, false);
+	compare = ApplySortComparator(da, false, db, false, cxt->ssup);
 	if (compare != 0)
 		return compare;

--- a/src/backend/commands/opclasscmds.c
+++ b/src/backend/commands/opclasscmds.c
@@ -19,6 +19,7 @@

 #include "access/genam.h"
 #include "access/heapam.h"
+#include "access/nbtree.h"
 #include "access/sysattr.h"
 #include "catalog/dependency.h"
 #include "catalog/indexing.h"
@@ -1151,27 +1152,48 @@ assignProcTypes(OpFamilyMember *member, Oid amoid, Oid typeoid)
 	procform = (Form_pg_proc) GETSTRUCT(proctup);

 	/*
-	 * btree support procs must be 2-arg procs returning int4; hash support
-	 * procs must be 1-arg procs returning int4; otherwise we don't know.
+	 * btree comparison procs must be 2-arg procs returning int4, while btree
+	 * sortsupport procs must take internal and return void.  hash support
+	 * procs must be 1-arg procs returning int4.  Otherwise we don't know.
 	 */
 	if (amoid == BTREE_AM_OID)
 	{
-		if (procform->pronargs != 2)
-			ereport(ERROR,
-					(errcode(ERRCODE_INVALID_OBJECT_DEFINITION),
-					 errmsg("btree procedures must have two arguments")));
-		if (procform->prorettype != INT4OID)
-			ereport(ERROR,
-					(errcode(ERRCODE_INVALID_OBJECT_DEFINITION),
-					 errmsg("btree procedures must return integer")));
+		if (member->number == BTORDER_PROC)
+		{
+			if (procform->pronargs != 2)
+				ereport(ERROR,
+						(errcode(ERRCODE_INVALID_OBJECT_DEFINITION),
+						 errmsg("btree comparison procedures must have two arguments")));
+			if (procform->prorettype != INT4OID)
+				ereport(ERROR,
+						(errcode(ERRCODE_INVALID_OBJECT_DEFINITION),
+						 errmsg("btree comparison procedures must return integer")));

-		/*
-		 * If lefttype/righttype isn't specified, use the proc's input types
-		 */
-		if (!OidIsValid(member->lefttype))
-			member->lefttype = procform->proargtypes.values[0];
-		if (!OidIsValid(member->righttype))
-			member->righttype = procform->proargtypes.values[1];
+			/*
+			 * If lefttype/righttype isn't specified, use the proc's input
+			 * types
+			 */
+			if (!OidIsValid(member->lefttype))
+				member->lefttype = procform->proargtypes.values[0];
+			if (!OidIsValid(member->righttype))
+				member->righttype = procform->proargtypes.values[1];
+		}
+		else if (member->number == BTSORTSUPPORT_PROC)
+		{
+			if (procform->pronargs != 1 ||
+				procform->proargtypes.values[0] != INTERNALOID)
+				ereport(ERROR,
+						(errcode(ERRCODE_INVALID_OBJECT_DEFINITION),
+						 errmsg("btree sort support procedures must accept type \"internal\"")));
+			if (procform->prorettype != VOIDOID)
+				ereport(ERROR,
+						(errcode(ERRCODE_INVALID_OBJECT_DEFINITION),
+						 errmsg("btree sort support procedures must return void")));
+
+			/*
+			 * Can't infer lefttype/righttype from proc, so use default rule
+			 */
+		}
 	}
 	else if (amoid == HASH_AM_OID)
 	{
@@ -1192,23 +1214,21 @@ assignProcTypes(OpFamilyMember *member, Oid amoid, Oid typeoid)
 		if (!OidIsValid(member->righttype))
 			member->righttype = procform->proargtypes.values[0];
 	}
-	else
-	{
-		/*
-		 * The default for GiST and GIN in CREATE OPERATOR CLASS is to use the
-		 * class' opcintype as lefttype and righttype.  In CREATE or ALTER
-		 * OPERATOR FAMILY, opcintype isn't available, so make the user
-		 * specify the types.
-		 */
-		if (!OidIsValid(member->lefttype))
-			member->lefttype = typeoid;
-		if (!OidIsValid(member->righttype))
-			member->righttype = typeoid;
-		if (!OidIsValid(member->lefttype) || !OidIsValid(member->righttype))
-			ereport(ERROR,
-					(errcode(ERRCODE_INVALID_OBJECT_DEFINITION),
-					 errmsg("associated data types must be specified for index support procedure")));
-	}
+
+	/*
+	 * The default in CREATE OPERATOR CLASS is to use the class' opcintype as
+	 * lefttype and righttype.  In CREATE or ALTER OPERATOR FAMILY, opcintype
+	 * isn't available, so make the user specify the types.
+	 */
+	if (!OidIsValid(member->lefttype))
+		member->lefttype = typeoid;
+	if (!OidIsValid(member->righttype))
+		member->righttype = typeoid;
+
+	if (!OidIsValid(member->lefttype) || !OidIsValid(member->righttype))
+		ereport(ERROR,
+				(errcode(ERRCODE_INVALID_OBJECT_DEFINITION),
+				 errmsg("associated data types must be specified for index support procedure")));

 	ReleaseSysCache(proctup);
 }