Improve performance of index-only scans with many index columns.

tglsfdc · tglsfdc · commit 80b9e9c4664a · 2019-03-03T16:57:14.000-05:00
StoreIndexTuple was a loop over index_getattr, which is O(N^2) if the index columns are variable-width, and the performance impact is already quite visible at ten columns. The obvious move is to replace that with a call to index_deform_tuple ... but that's *also* a loop over index_getattr. Improve it to be essentially a clone of heap_deform_tuple. (There are a few other places that loop over all index columns with index_getattr, and perhaps should be changed likewise, but most of them don't seem performance-critical. Anyway, the rest would mostly only be interested in the index key columns, which there aren't likely to be so many of. Wide index tuples are a new thing with INCLUDE.) Konstantin Knizhnik Discussion: https://postgr.es/m/e06b2d27-04fc-5c0e-bb8c-ecd72aa24959@postgrespro.ru
diff --git a/src/backend/access/common/indextuple.c b/src/backend/access/common/indextuple.c
@@ -418,19 +418,80 @@ nocache_index_getattr(IndexTuple tup,
  *
  * The caller must allocate sufficient storage for the output arrays.
  * (INDEX_MAX_KEYS entries should be enough.)
+ *
+ * This is nearly the same as heap_deform_tuple(), but for IndexTuples.
+ * One difference is that the tuple should never have any missing columns.
  */
 void
 index_deform_tuple(IndexTuple tup, TupleDesc tupleDescriptor,
 				   Datum *values, bool *isnull)
 {
-	int			i;
+	int			hasnulls = IndexTupleHasNulls(tup);
+	int			natts = tupleDescriptor->natts; /* number of atts to extract */
+	int			attnum;
+	char	   *tp;				/* ptr to tuple data */
+	int			off;			/* offset in tuple data */
+	bits8	   *bp;				/* ptr to null bitmap in tuple */
+	bool		slow = false;	/* can we use/set attcacheoff? */
 
 	/* Assert to protect callers who allocate fixed-size arrays */
-	Assert(tupleDescriptor->natts <= INDEX_MAX_KEYS);
+	Assert(natts <= INDEX_MAX_KEYS);
+
+	/* XXX "knows" t_bits are just after fixed tuple header! */
+	bp = (bits8 *) ((char *) tup + sizeof(IndexTupleData));
 
-	for (i = 0; i < tupleDescriptor->natts; i++)
+	tp = (char *) tup + IndexInfoFindDataOffset(tup->t_info);
+	off = 0;
+
+	for (attnum = 0; attnum < natts; attnum++)
 	{
-		values[i] = index_getattr(tup, i + 1, tupleDescriptor, &isnull[i]);
+		Form_pg_attribute thisatt = TupleDescAttr(tupleDescriptor, attnum);
+
+		if (hasnulls && att_isnull(attnum, bp))
+		{
+			values[attnum] = (Datum) 0;
+			isnull[attnum] = true;
+			slow = true;		/* can't use attcacheoff anymore */
+			continue;
+		}
+
+		isnull[attnum] = false;
+
+		if (!slow && thisatt->attcacheoff >= 0)
+			off = thisatt->attcacheoff;
+		else if (thisatt->attlen == -1)
+		{
+			/*
+			 * We can only cache the offset for a varlena attribute if the
+			 * offset is already suitably aligned, so that there would be no
+			 * pad bytes in any case: then the offset will be valid for either
+			 * an aligned or unaligned value.
+			 */
+			if (!slow &&
+				off == att_align_nominal(off, thisatt->attalign))
+				thisatt->attcacheoff = off;
+			else
+			{
+				off = att_align_pointer(off, thisatt->attalign, -1,
+										tp + off);
+				slow = true;
+			}
+		}
+		else
+		{
+			/* not varlena, so safe to use att_align_nominal */
+			off = att_align_nominal(off, thisatt->attalign);
+
+			if (!slow)
+				thisatt->attcacheoff = off;
+		}
+
+		values[attnum] = fetchatt(thisatt, tp + off);
+
+		off = att_addlength_pointer(off, thisatt->attlen, tp + off);
+
+		if (thisatt->attlen <= 0)
+			slow = true;		/* can't use attcacheoff anymore */
 	}
 }
 
diff --git a/src/backend/executor/nodeIndexonlyscan.c b/src/backend/executor/nodeIndexonlyscan.c
@@ -269,23 +269,17 @@ IndexOnlyNext(IndexOnlyScanState *node)
 static void
 StoreIndexTuple(TupleTableSlot *slot, IndexTuple itup, TupleDesc itupdesc)
 {
-	int			nindexatts = itupdesc->natts;
-	Datum	   *values = slot->tts_values;
-	bool	   *isnull = slot->tts_isnull;
-	int			i;
-
 	/*
-	 * Note: we must use the tupdesc supplied by the AM in index_getattr, not
-	 * the slot's tupdesc, in case the latter has different datatypes (this
-	 * happens for btree name_ops in particular).  They'd better have the same
-	 * number of columns though, as well as being datatype-compatible which is
-	 * something we can't so easily check.
+	 * Note: we must use the tupdesc supplied by the AM in index_deform_tuple,
+	 * not the slot's tupdesc, in case the latter has different datatypes
+	 * (this happens for btree name_ops in particular).  They'd better have
+	 * the same number of columns though, as well as being datatype-compatible
+	 * which is something we can't so easily check.
 	 */
-	Assert(slot->tts_tupleDescriptor->natts == nindexatts);
+	Assert(slot->tts_tupleDescriptor->natts == itupdesc->natts);
 
 	ExecClearTuple(slot);
-	for (i = 0; i < nindexatts; i++)
-		values[i] = index_getattr(itup, i + 1, itupdesc, &isnull[i]);
+	index_deform_tuple(itup, itupdesc, slot->tts_values, slot->tts_isnull);
 	ExecStoreVirtualTuple(slot);
 }
 
