Use ExprStates for hashing in GROUP BY and SubPlans

This speeds up obtaining hash values for GROUP BY and hashed SubPlans by
using the ExprState support for hashing, thus allowing JIT compilation for
obtaining hash values for these operations.

This, even without JIT compilation, has been shown to improve Hash
Aggregate performance in some cases by around 15% and hashed NOT IN
queries in one case by over 30%, however, real-world cases are likely to
see smaller gains as the test cases used were purposefully designed to
have high hashing overheads by keeping the hash table small to prevent
additional memory overheads that would be a factor when working with large
hash tables.

In passing, fix a hypothetical bug in ExecBuildHash32Expr() so that the
initial value is stored directly in the ExprState's result field if
there are no expressions to hash.  None of the current users of this
function use an initial value, so the bug is only hypothetical.

Reviewed-by: Andrei Lepikhov <lepihov@gmail.com>
Discussion: https://postgr.es/m/CAApHDvpYSO3kc9UryMevWqthTBrxgfd9djiAjKHMPUSQeX9vdQ@mail.gmail.com

Author: david-rowley

src/backend/executor/execExpr.c

@@ -3978,6 +3978,161 @@ ExecBuildAggTransCall(ExprState *state, AggState *aggstate,
 	}
 }
 
+/*
+ * Build an ExprState that calls the given hash function(s) on the attnums
+ * given by 'keyColIdx' .  When numCols > 1, the hash values returned by each
+ * hash function are combined to produce a single hash value.
+ *
+ * desc: tuple descriptor for the to-be-hashed columns
+ * ops: TupleTableSlotOps to use for the give TupleDesc
+ * hashfunctions: FmgrInfos for each hash function to call, one per numCols.
+ * These are used directly in the returned ExprState so must remain allocated.
+ * collations: collation to use when calling the hash function.
+ * numCols: array length of hashfunctions, collations and keyColIdx.
+ * parent: PlanState node that the resulting ExprState will be evaluated at
+ * init_value: Normally 0, but can be set to other values to seed the hash
+ * with.  Non-zero is marginally slower, so best to only use if it's provably
+ * worthwhile.
+ */
+ExprState *
+ExecBuildHash32FromAttrs(TupleDesc desc, const TupleTableSlotOps *ops,
+						 FmgrInfo *hashfunctions, Oid *collations,
+						 int numCols, AttrNumber *keyColIdx,
+						 PlanState *parent, uint32 init_value)
+{
+	ExprState  *state = makeNode(ExprState);
+	ExprEvalStep scratch = {0};
+	NullableDatum *iresult = NULL;
+	intptr_t	opcode;
+	AttrNumber	last_attnum = 0;
+
+	Assert(numCols >= 0);
+
+	state->parent = parent;
+
+	/*
+	 * Make a place to store intermediate hash values between subsequent
+	 * hashing of individual columns.  We only need this if there is more than
+	 * one column to hash or an initial value plus one column.
+	 */
+	if ((int64) numCols + (init_value != 0) > 1)
+		iresult = palloc(sizeof(NullableDatum));
+
+	/* find the highest attnum so we deform the tuple to that point */
+	for (int i = 0; i < numCols; i++)
+		last_attnum = Max(last_attnum, keyColIdx[i]);
+
+	scratch.opcode = EEOP_INNER_FETCHSOME;
+	scratch.d.fetch.last_var = last_attnum;
+	scratch.d.fetch.fixed = false;
+	scratch.d.fetch.kind = ops;
+	scratch.d.fetch.known_desc = desc;
+	if (ExecComputeSlotInfo(state, &scratch))
+		ExprEvalPushStep(state, &scratch);
+
+	if (init_value == 0)
+	{
+		/*
+		 * No initial value, so we can assign the result of the hash function
+		 * for the first attribute without having to concern ourselves with
+		 * combining the result with any initial value.
+		 */
+		opcode = EEOP_HASHDATUM_FIRST;
+	}
+	else
+	{
+		/*
+		 * Set up operation to set the initial value.  Normally we store this
+		 * in the intermediate hash value location, but if there are no
+		 * columns to hash, store it in the ExprState's result field.
+		 */
+		scratch.opcode = EEOP_HASHDATUM_SET_INITVAL;
+		scratch.d.hashdatum_initvalue.init_value = UInt32GetDatum(init_value);
+		scratch.resvalue = numCols > 0 ? &iresult->value : &state->resvalue;
+		scratch.resnull = numCols > 0 ? &iresult->isnull : &state->resnull;
+
+		ExprEvalPushStep(state, &scratch);
+
+		/*
+		 * When using an initial value use the NEXT32 ops as the FIRST ops
+		 * would overwrite the stored initial value.
+		 */
+		opcode = EEOP_HASHDATUM_NEXT32;
+	}
+
+	for (int i = 0; i < numCols; i++)
+	{
+		FmgrInfo   *finfo;
+		FunctionCallInfo fcinfo;
+		Oid			inputcollid = collations[i];
+		AttrNumber	attnum = keyColIdx[i] - 1;
+
+		finfo = &hashfunctions[i];
+		fcinfo = palloc0(SizeForFunctionCallInfo(1));
+
+		/* Initialize function call parameter structure too */
+		InitFunctionCallInfoData(*fcinfo, finfo, 1, inputcollid, NULL, NULL);
+
+		/*
+		 * Fetch inner Var for this attnum and store it in the 1st arg of the
+		 * hash func.
+		 */
+		scratch.opcode = EEOP_INNER_VAR;
+		scratch.resvalue = &fcinfo->args[0].value;
+		scratch.resnull = &fcinfo->args[0].isnull;
+		scratch.d.var.attnum = attnum;
+		scratch.d.var.vartype = TupleDescAttr(desc, attnum)->atttypid;
+
+		ExprEvalPushStep(state, &scratch);
+
+		/* Call the hash function */
+		scratch.opcode = opcode;
+
+		if (i == numCols - 1)
+		{
+			/*
+			 * The result for hashing the final column is stored in the
+			 * ExprState.
+			 */
+			scratch.resvalue = &state->resvalue;
+			scratch.resnull = &state->resnull;
+		}
+		else
+		{
+			Assert(iresult != NULL);
+
+			/* intermediate values are stored in an intermediate result */
+			scratch.resvalue = &iresult->value;
+			scratch.resnull = &iresult->isnull;
+		}
+
+		/*
+		 * NEXT32 opcodes need to look at the intermediate result.  We might
+		 * as well just set this for all ops.  FIRSTs won't look at it.
+		 */
+		scratch.d.hashdatum.iresult = iresult;
+
+		scratch.d.hashdatum.finfo = finfo;
+		scratch.d.hashdatum.fcinfo_data = fcinfo;
+		scratch.d.hashdatum.fn_addr = finfo->fn_addr;
+		scratch.d.hashdatum.jumpdone = -1;
+
+		ExprEvalPushStep(state, &scratch);
+
+		/* subsequent attnums must be combined with the previous */
+		opcode = EEOP_HASHDATUM_NEXT32;
+	}
+
+	scratch.resvalue = NULL;
+	scratch.resnull = NULL;
+	scratch.opcode = EEOP_DONE;
+	ExprEvalPushStep(state, &scratch);
+
+	ExecReadyExpr(state);
+
+	return state;
+}
+
 /*
  * Build an ExprState that calls the given hash function(s) on the given
  * 'hash_exprs'.  When multiple expressions are present, the hash values

src/backend/executor/execGrouping.c

@@ -169,6 +169,7 @@ BuildTupleHashTableExt(PlanState *parent,
 	Size		hash_mem_limit;
 	MemoryContext oldcontext;
 	bool		allow_jit;
+	uint32		hash_iv = 0;
 
 	Assert(nbuckets > 0);
 
@@ -183,14 +184,13 @@ BuildTupleHashTableExt(PlanState *parent,
 
 	hashtable->numCols = numCols;
 	hashtable->keyColIdx = keyColIdx;
-	hashtable->tab_hash_funcs = hashfunctions;
 	hashtable->tab_collations = collations;
 	hashtable->tablecxt = tablecxt;
 	hashtable->tempcxt = tempcxt;
 	hashtable->entrysize = entrysize;
 	hashtable->tableslot = NULL;	/* will be made on first lookup */
 	hashtable->inputslot = NULL;
-	hashtable->in_hash_funcs = NULL;
+	hashtable->in_hash_expr = NULL;
 	hashtable->cur_eq_func = NULL;
 
 	/*
@@ -202,9 +202,7 @@ BuildTupleHashTableExt(PlanState *parent,
 	 * underestimated.
 	 */
 	if (use_variable_hash_iv)
-		hashtable->hash_iv = murmurhash32(ParallelWorkerNumber);
-	else
-		hashtable->hash_iv = 0;
+		hash_iv = murmurhash32(ParallelWorkerNumber);
 
 	hashtable->hashtab = tuplehash_create(metacxt, nbuckets, hashtable);
 
@@ -225,6 +223,16 @@ BuildTupleHashTableExt(PlanState *parent,
 	 */
 	allow_jit = metacxt != tablecxt;
 
+	/* build hash ExprState for all columns */
+	hashtable->tab_hash_expr = ExecBuildHash32FromAttrs(inputDesc,
+														&TTSOpsMinimalTuple,
+														hashfunctions,
+														collations,
+														numCols,
+														keyColIdx,
+														allow_jit ? parent : NULL,
+														hash_iv);
+
 	/* build comparator for all columns */
 	/* XXX: should we support non-minimal tuples for the inputslot? */
 	hashtable->tab_eq_func = ExecBuildGroupingEqual(inputDesc, inputDesc,
@@ -316,7 +324,7 @@ LookupTupleHashEntry(TupleHashTable hashtable, TupleTableSlot *slot,
 
 	/* set up data needed by hash and match functions */
 	hashtable->inputslot = slot;
-	hashtable->in_hash_funcs = hashtable->tab_hash_funcs;
+	hashtable->in_hash_expr = hashtable->tab_hash_expr;
 	hashtable->cur_eq_func = hashtable->tab_eq_func;
 
 	local_hash = TupleHashTableHash_internal(hashtable->hashtab, NULL);
@@ -342,7 +350,7 @@ TupleHashTableHash(TupleHashTable hashtable, TupleTableSlot *slot)
 	uint32		hash;
 
 	hashtable->inputslot = slot;
-	hashtable->in_hash_funcs = hashtable->tab_hash_funcs;
+	hashtable->in_hash_expr = hashtable->tab_hash_expr;
 
 	/* Need to run the hash functions in short-lived context */
 	oldContext = MemoryContextSwitchTo(hashtable->tempcxt);
@@ -370,7 +378,7 @@ LookupTupleHashEntryHash(TupleHashTable hashtable, TupleTableSlot *slot,
 
 	/* set up data needed by hash and match functions */
 	hashtable->inputslot = slot;
-	hashtable->in_hash_funcs = hashtable->tab_hash_funcs;
+	hashtable->in_hash_expr = hashtable->tab_hash_expr;
 	hashtable->cur_eq_func = hashtable->tab_eq_func;
 
 	entry = LookupTupleHashEntry_internal(hashtable, slot, isnew, hash);
@@ -386,14 +394,14 @@ LookupTupleHashEntryHash(TupleHashTable hashtable, TupleTableSlot *slot,
  * created if there's not a match.  This is similar to the non-creating
  * case of LookupTupleHashEntry, except that it supports cross-type
  * comparisons, in which the given tuple is not of the same type as the
- * table entries.  The caller must provide the hash functions to use for
- * the input tuple, as well as the equality functions, since these may be
+ * table entries.  The caller must provide the hash ExprState to use for
+ * the input tuple, as well as the equality ExprState, since these may be
  * different from the table's internal functions.
  */
 TupleHashEntry
 FindTupleHashEntry(TupleHashTable hashtable, TupleTableSlot *slot,
 				   ExprState *eqcomp,
-				   FmgrInfo *hashfunctions)
+				   ExprState *hashexpr)
 {
 	TupleHashEntry entry;
 	MemoryContext oldContext;
@@ -404,7 +412,7 @@ FindTupleHashEntry(TupleHashTable hashtable, TupleTableSlot *slot,
 
 	/* Set up data needed by hash and match functions */
 	hashtable->inputslot = slot;
-	hashtable->in_hash_funcs = hashfunctions;
+	hashtable->in_hash_expr = hashexpr;
 	hashtable->cur_eq_func = eqcomp;
 
 	/* Search the hash table */
@@ -421,25 +429,24 @@ FindTupleHashEntry(TupleHashTable hashtable, TupleTableSlot *slot,
  * copied into the table.
  *
  * Also, the caller must select an appropriate memory context for running
- * the hash functions. (dynahash.c doesn't change CurrentMemoryContext.)
+ * the hash functions.
  */
 static uint32
 TupleHashTableHash_internal(struct tuplehash_hash *tb,
 							const MinimalTuple tuple)
 {
 	TupleHashTable hashtable = (TupleHashTable) tb->private_data;
-	int			numCols = hashtable->numCols;
-	AttrNumber *keyColIdx = hashtable->keyColIdx;
-	uint32		hashkey = hashtable->hash_iv;
+	uint32		hashkey;
 	TupleTableSlot *slot;
-	FmgrInfo   *hashfunctions;
-	int			i;
+	bool		isnull;
 
 	if (tuple == NULL)
 	{
 		/* Process the current input tuple for the table */
-		slot = hashtable->inputslot;
-		hashfunctions = hashtable->in_hash_funcs;
+		hashtable->exprcontext->ecxt_innertuple = hashtable->inputslot;
+		hashkey = DatumGetUInt32(ExecEvalExpr(hashtable->in_hash_expr,
+											  hashtable->exprcontext,
+											  &isnull));
 	}
 	else
 	{
@@ -449,38 +456,17 @@ TupleHashTableHash_internal(struct tuplehash_hash *tb,
 		 * (this case never actually occurs due to the way simplehash.h is
 		 * used, as the hash-value is stored in the entries)
 		 */
-		slot = hashtable->tableslot;
+		slot = hashtable->exprcontext->ecxt_innertuple = hashtable->tableslot;
 		ExecStoreMinimalTuple(tuple, slot, false);
-		hashfunctions = hashtable->tab_hash_funcs;
-	}
-
-	for (i = 0; i < numCols; i++)
-	{
-		AttrNumber	att = keyColIdx[i];
-		Datum		attr;
-		bool		isNull;
-
-		/* combine successive hashkeys by rotating */
-		hashkey = pg_rotate_left32(hashkey, 1);
-
-		attr = slot_getattr(slot, att, &isNull);
-
-		if (!isNull)			/* treat nulls as having hash key 0 */
-		{
-			uint32		hkey;
-
-			hkey = DatumGetUInt32(FunctionCall1Coll(&hashfunctions[i],
-													hashtable->tab_collations[i],
-													attr));
-			hashkey ^= hkey;
-		}
+		hashkey = DatumGetUInt32(ExecEvalExpr(hashtable->tab_hash_expr,
+											  hashtable->exprcontext,
+											  &isnull));
 	}
 
 	/*
-	 * The way hashes are combined above, among each other and with the IV,
-	 * doesn't lead to good bit perturbation. As the IV's goal is to lead to
-	 * achieve that, perform a round of hashing of the combined hash -
-	 * resulting in near perfect perturbation.
+	 * The hashing done above, even with an initial value, doesn't tend to
+	 * result in good hash perturbation.  Running the value produced above
+	 * through murmurhash32 leads to near perfect hash perturbation.
 	 */
 	return murmurhash32(hashkey);
 }