Optimize alignment calculations in tuple form/deform

Here we convert CompactAttribute.attalign from a char, which is directly
derived from pg_attribute.attalign into a uint8, which stores the number
of bytes to align the column's value by in the tuple.

This allows tuple deformation and tuple size calculations to move away
from using the inefficient att_align_nominal() macro, which manually
checks each TYPALIGN_* char to translate that into the alignment bytes
for the given type.  Effectively, this commit changes those to TYPEALIGN
calls, which are branchless and only perform some simple arithmetic with
some bit-twiddling.

The removed branches were often mispredicted by CPUs, especially so in
real-world tables which often contain a mishmash of different types
with different alignment requirements.

Author: David Rowley
Reviewed-by: Andres Freund, Victor Yegorov
Discussion: https://postgr.es/m/CAApHDvrBztXP3yx=NKNmo3xwFAFhEdyPnvrDg3=M0RhDs+4vYw@mail.gmail.com

contrib/amcheck/verify_heapam.c

@@ -1596,7 +1596,7 @@ check_tuple_attribute(HeapCheckContext *ctx)
 	/* Skip non-varlena values, but update offset first */
 	if (thisatt->attlen != -1)
 	{
-		ctx->offset = att_align_nominal(ctx->offset, thisatt->attalign);
+		ctx->offset = att_nominal_alignby(ctx->offset, thisatt->attalignby);
 		ctx->offset = att_addlength_pointer(ctx->offset, thisatt->attlen,
 											tp + ctx->offset);
 		if (ctx->tuphdr->t_hoff + ctx->offset > ctx->lp_len)
@@ -1612,7 +1612,7 @@ check_tuple_attribute(HeapCheckContext *ctx)
 	}
 
 	/* Ok, we're looking at a varlena attribute. */
-	ctx->offset = att_align_pointer(ctx->offset, thisatt->attalign, -1,
+	ctx->offset = att_pointer_alignby(ctx->offset, thisatt->attalignby, -1,
 									  tp + ctx->offset);
 
 	/* Get the (possibly corrupt) varlena datum */

contrib/pageinspect/heapfuncs.c

@@ -357,7 +357,7 @@ tuple_data_split_internal(Oid relid, char *tupdata,
 
 			if (attr->attlen == -1)
 			{
-				off = att_align_pointer(off, attr->attalign, -1,
+				off = att_pointer_alignby(off, attr->attalignby, -1,
 										  tupdata + off);
 
 				/*
@@ -376,7 +376,7 @@ tuple_data_split_internal(Oid relid, char *tupdata,
 			}
 			else
 			{
-				off = att_align_nominal(off, attr->attalign);
+				off = att_nominal_alignby(off, attr->attalignby);
 				len = attr->attlen;
 			}
 

src/backend/access/brin/brin_tuple.c

@@ -703,13 +703,15 @@ brin_deconstruct_tuple(BrinDesc *brdesc,
 
 			if (thisatt->attlen == -1)
 			{
-				off = att_align_pointer(off, thisatt->attalign, -1,
+				off = att_pointer_alignby(off,
+										  thisatt->attalignby,
+										  -1,
 										  tp + off);
 			}
 			else
 			{
-				/* not varlena, so safe to use att_align_nominal */
-				off = att_align_nominal(off, thisatt->attalign);
+				/* not varlena, so safe to use att_nominal_alignby */
+				off = att_nominal_alignby(off, thisatt->attalignby);
 			}
 
 			values[stored++] = fetchatt(thisatt, tp + off);

src/backend/access/common/attmap.c

@@ -315,13 +315,13 @@ check_attrmap_match(TupleDesc indesc,
 
 		/*
 		 * If it's a dropped column and the corresponding input column is also
-		 * dropped, we don't need a conversion.  However, attlen and attalign
-		 * must agree.
+		 * dropped, we don't need a conversion.  However, attlen and
+		 * attalignby must agree.
 		 */
 		if (attrMap->attnums[i] == 0 &&
 			inatt->attisdropped &&
 			inatt->attlen == outatt->attlen &&
-			inatt->attalign == outatt->attalign)
+			inatt->attalignby == outatt->attalignby)
 			continue;
 
 		return false;

src/backend/access/common/heaptuple.c

@@ -251,12 +251,12 @@ heap_compute_data_size(TupleDesc tupleDesc,
 			 * we want to flatten the expanded value so that the constructed
 			 * tuple doesn't depend on it
 			 */
-			data_length = att_align_nominal(data_length, atti->attalign);
+			data_length = att_nominal_alignby(data_length, atti->attalignby);
 			data_length += EOH_get_flat_size(DatumGetEOHP(val));
 		}
 		else
 		{
-			data_length = att_align_datum(data_length, atti->attalign,
+			data_length = att_datum_alignby(data_length, atti->attalignby,
 											atti->attlen, val);
 			data_length = att_addlength_datum(data_length, atti->attlen,
 											  val);
@@ -308,13 +308,13 @@ fill_val(CompactAttribute *att,
 	}
 
 	/*
-	 * XXX we use the att_align macros on the pointer value itself, not on an
-	 * offset.  This is a bit of a hack.
+	 * XXX we use the att_nominal_alignby macro on the pointer value itself,
+	 * not on an offset.  This is a bit of a hack.
 	 */
 	if (att->attbyval)
 	{
 		/* pass-by-value */
-		data = (char *) att_align_nominal(data, att->attalign);
+		data = (char *) att_nominal_alignby(data, att->attalignby);
 		store_att_byval(data, datum, att->attlen);
 		data_length = att->attlen;
 	}
@@ -334,8 +334,7 @@ fill_val(CompactAttribute *att,
 				 */
 				ExpandedObjectHeader *eoh = DatumGetEOHP(datum);
 
-				data = (char *) att_align_nominal(data,
-												  att->attalign);
+				data = (char *) att_nominal_alignby(data, att->attalignby);
 				data_length = EOH_get_flat_size(eoh);
 				EOH_flatten_into(eoh, data, data_length);
 			}
@@ -363,8 +362,7 @@ fill_val(CompactAttribute *att,
 		else
 		{
 			/* full 4-byte header varlena */
-			data = (char *) att_align_nominal(data,
-											  att->attalign);
+			data = (char *) att_nominal_alignby(data, att->attalignby);
 			data_length = VARSIZE(val);
 			memcpy(data, val, data_length);
 		}
@@ -373,14 +371,14 @@ fill_val(CompactAttribute *att,
 	{
 		/* cstring ... never needs alignment */
 		*infomask |= HEAP_HASVARWIDTH;
-		Assert(att->attalign == TYPALIGN_CHAR);
+		Assert(att->attalignby == sizeof(char));
 		data_length = strlen(DatumGetCString(datum)) + 1;
 		memcpy(data, DatumGetPointer(datum), data_length);
 	}
 	else
 	{
 		/* fixed-length pass-by-reference */
-		data = (char *) att_align_nominal(data, att->attalign);
+		data = (char *) att_nominal_alignby(data, att->attalignby);
 		Assert(att->attlen > 0);
 		data_length = att->attlen;
 		memcpy(data, DatumGetPointer(datum), data_length);
@@ -634,7 +632,7 @@ nocachegetattr(HeapTuple tup,
 			if (att->attlen <= 0)
 				break;
 
-			off = att_align_nominal(off, att->attalign);
+			off = att_nominal_alignby(off, att->attalignby);
 
 			att->attcacheoff = off;
 
@@ -683,19 +681,19 @@ nocachegetattr(HeapTuple tup,
 				 * either an aligned or unaligned value.
 				 */
 				if (usecache &&
-					off == att_align_nominal(off, att->attalign))
+					off == att_nominal_alignby(off, att->attalignby))
 					att->attcacheoff = off;
 				else
 				{
-					off = att_align_pointer(off, att->attalign, -1,
+					off = att_pointer_alignby(off, att->attalignby, -1,
 											  tp + off);
 					usecache = false;
 				}
 			}
 			else
 			{
-				/* not varlena, so safe to use att_align_nominal */
-				off = att_align_nominal(off, att->attalign);
+				/* not varlena, so safe to use att_nominal_alignby */
+				off = att_nominal_alignby(off, att->attalignby);
 
 				if (usecache)
 					att->attcacheoff = off;
@@ -898,8 +896,8 @@ expand_tuple(HeapTuple *targetHeapTuple,
 			{
 				CompactAttribute *att = TupleDescCompactAttr(tupleDesc, attnum);
 
-				targetDataLen = att_align_datum(targetDataLen,
-												att->attalign,
+				targetDataLen = att_datum_alignby(targetDataLen,
+												  att->attalignby,
 												  att->attlen,
 												  attrmiss[attnum].am_value);
 
@@ -1396,19 +1394,19 @@ heap_deform_tuple(HeapTuple tuple, TupleDesc tupleDesc,
 			 * an aligned or unaligned value.
 			 */
 			if (!slow &&
-				off == att_align_nominal(off, thisatt->attalign))
+				off == att_nominal_alignby(off, thisatt->attalignby))
 				thisatt->attcacheoff = off;
 			else
 			{
-				off = att_align_pointer(off, thisatt->attalign, -1,
+				off = att_pointer_alignby(off, thisatt->attalignby, -1,
 										  tp + off);
 				slow = true;
 			}
 		}
 		else
 		{
-			/* not varlena, so safe to use att_align_nominal */
-			off = att_align_nominal(off, thisatt->attalign);
+			/* not varlena, so safe to use att_nominal_alignby */
+			off = att_nominal_alignby(off, thisatt->attalignby);
 
 			if (!slow)
 				thisatt->attcacheoff = off;

src/backend/access/common/indextuple.c

@@ -363,7 +363,7 @@ nocache_index_getattr(IndexTuple tup,
 			if (att->attlen <= 0)
 				break;
 
-			off = att_align_nominal(off, att->attalign);
+			off = att_nominal_alignby(off, att->attalignby);
 
 			att->attcacheoff = off;
 
@@ -412,19 +412,19 @@ nocache_index_getattr(IndexTuple tup,
 				 * either an aligned or unaligned value.
 				 */
 				if (usecache &&
-					off == att_align_nominal(off, att->attalign))
+					off == att_nominal_alignby(off, att->attalignby))
 					att->attcacheoff = off;
 				else
 				{
-					off = att_align_pointer(off, att->attalign, -1,
+					off = att_pointer_alignby(off, att->attalignby, -1,
 											  tp + off);
 					usecache = false;
 				}
 			}
 			else
 			{
-				/* not varlena, so safe to use att_align_nominal */
-				off = att_align_nominal(off, att->attalign);
+				/* not varlena, so safe to use att_nominal_alignby */
+				off = att_nominal_alignby(off, att->attalignby);
 
 				if (usecache)
 					att->attcacheoff = off;
@@ -513,19 +513,19 @@ index_deform_tuple_internal(TupleDesc tupleDescriptor,
 			 * an aligned or unaligned value.
 			 */
 			if (!slow &&
-				off == att_align_nominal(off, thisatt->attalign))
+				off == att_nominal_alignby(off, thisatt->attalignby))
 				thisatt->attcacheoff = off;
 			else
 			{
-				off = att_align_pointer(off, thisatt->attalign, -1,
+				off = att_pointer_alignby(off, thisatt->attalignby, -1,
 										  tp + off);
 				slow = true;
 			}
 		}
 		else
 		{
-			/* not varlena, so safe to use att_align_nominal */
-			off = att_align_nominal(off, thisatt->attalign);
+			/* not varlena, so safe to use att_nominal_alignby */
+			off = att_nominal_alignby(off, thisatt->attalignby);
 
 			if (!slow)
 				thisatt->attcacheoff = off;

src/backend/access/common/tupdesc.c

@@ -80,7 +80,25 @@ populate_compact_attribute(TupleDesc tupdesc, int attnum)
 	dst->attgenerated = (src->attgenerated != '\0');
 	dst->attnotnull = src->attnotnull;
 
-	dst->attalign = src->attalign;
+	switch (src->attalign)
+	{
+		case TYPALIGN_INT:
+			dst->attalignby = ALIGNOF_INT;
+			break;
+		case TYPALIGN_CHAR:
+			dst->attalignby = sizeof(char);
+			break;
+		case TYPALIGN_DOUBLE:
+			dst->attalignby = ALIGNOF_DOUBLE;
+			break;
+		case TYPALIGN_SHORT:
+			dst->attalignby = ALIGNOF_SHORT;
+			break;
+		default:
+			dst->attalignby = 0;
+			elog(ERROR, "invalid attalign value: %c", src->attalign);
+			break;
+	}
 }
 
 /*

src/backend/executor/execExprInterp.c

@@ -5304,7 +5304,7 @@ ExecEvalWholeRowVar(ExprState *state, ExprEvalStep *op, ExprContext *econtext)
 			if (slot->tts_isnull[i])
 				continue;		/* null is always okay */
 			if (vattr->attlen != sattr->attlen ||
-				vattr->attalign != sattr->attalign)
+				vattr->attalignby != sattr->attalignby)
 				ereport(ERROR,
 						(errcode(ERRCODE_DATATYPE_MISMATCH),
 						 errmsg("table row type and query-specified row type do not match"),

src/backend/executor/execTuples.c

@@ -202,12 +202,12 @@ tts_virtual_materialize(TupleTableSlot *slot)
 			 * We want to flatten the expanded value so that the materialized
 			 * slot doesn't depend on it.
 			 */
-			sz = att_align_nominal(sz, att->attalign);
+			sz = att_nominal_alignby(sz, att->attalignby);
 			sz += EOH_get_flat_size(DatumGetEOHP(val));
 		}
 		else
 		{
-			sz = att_align_nominal(sz, att->attalign);
+			sz = att_nominal_alignby(sz, att->attalignby);
 			sz = att_addlength_datum(sz, att->attlen, val);
 		}
 	}
@@ -242,8 +242,8 @@ tts_virtual_materialize(TupleTableSlot *slot)
 			 */
 			ExpandedObjectHeader *eoh = DatumGetEOHP(val);
 
-			data = (char *) att_align_nominal(data,
-											  att->attalign);
+			data = (char *) att_nominal_alignby(data,
+												att->attalignby);
 			data_length = EOH_get_flat_size(eoh);
 			EOH_flatten_into(eoh, data, data_length);
 
@@ -254,7 +254,7 @@ tts_virtual_materialize(TupleTableSlot *slot)
 		{
 			Size		data_length = 0;
 
-			data = (char *) att_align_nominal(data, att->attalign);
+			data = (char *) att_nominal_alignby(data, att->attalignby);
 			data_length = att_addlength_datum(data_length, att->attlen, val);
 
 			memcpy(data, DatumGetPointer(val), data_length);
@@ -1067,19 +1067,19 @@ slot_deform_heap_tuple(TupleTableSlot *slot, HeapTuple tuple, uint32 *offp,
 			 * an aligned or unaligned value.
 			 */
 			if (!slow &&
-				off == att_align_nominal(off, thisatt->attalign))
+				off == att_nominal_alignby(off, thisatt->attalignby))
 				thisatt->attcacheoff = off;
 			else
 			{
-				off = att_align_pointer(off, thisatt->attalign, -1,
+				off = att_pointer_alignby(off, thisatt->attalignby, -1,
 										  tp + off);
 				slow = true;
 			}
 		}
 		else
 		{
-			/* not varlena, so safe to use att_align_nominal */
-			off = att_align_nominal(off, thisatt->attalign);
+			/* not varlena, so safe to use att_nominal_alignby */
+			off = att_nominal_alignby(off, thisatt->attalignby);
 
 			if (!slow)
 				thisatt->attcacheoff = off;

src/backend/jit/llvm/llvmjit_deform.c

@@ -395,7 +395,7 @@ slot_compile_deform(LLVMJitContext *context, TupleDesc desc,
 	{
 		CompactAttribute *att = TupleDescCompactAttr(desc, attnum);
 		LLVMValueRef v_incby;
-		int			alignto;
+		int			alignto = att->attalignby;
 		LLVMValueRef l_attno = l_int16_const(lc, attnum);
 		LLVMValueRef v_attdatap;
 		LLVMValueRef v_resultp;
@@ -494,21 +494,6 @@ slot_compile_deform(LLVMJitContext *context, TupleDesc desc,
 		}
 		LLVMPositionBuilderAtEnd(b, attcheckalignblocks[attnum]);
 
-		/* determine required alignment */
-		if (att->attalign == TYPALIGN_INT)
-			alignto = ALIGNOF_INT;
-		else if (att->attalign == TYPALIGN_CHAR)
-			alignto = 1;
-		else if (att->attalign == TYPALIGN_DOUBLE)
-			alignto = ALIGNOF_DOUBLE;
-		else if (att->attalign == TYPALIGN_SHORT)
-			alignto = ALIGNOF_SHORT;
-		else
-		{
-			elog(ERROR, "unknown alignment");
-			alignto = 0;
-		}
-
 		/* ------
 		 * Even if alignment is required, we can skip doing it if provably
 		 * unnecessary:

src/include/access/tupdesc.h

@@ -75,7 +75,7 @@ typedef struct CompactAttribute
 	bool		attisdropped;	/* as FormData_pg_attribute.attisdropped */
 	bool		attgenerated;	/* FormData_pg_attribute.attgenerated != '\0' */
 	bool		attnotnull;		/* as FormData_pg_attribute.attnotnull */
-	char		attalign;		/* alignment requirement */
+	uint8		attalignby;		/* alignment requirement in bytes */
 } CompactAttribute;
 
 /*

src/include/access/tupmacs.h

@@ -91,6 +91,16 @@ fetch_att(const void *T, bool attbyval, int attlen)
 	att_align_nominal(cur_offset, attalign) \
 )
 
+/*
+ * Similar to att_align_datum, but accepts a number of bytes, typically from
+ * CompactAttribute.attalignby to align the Datum by.
+ */
+#define att_datum_alignby(cur_offset, attalignby, attlen, attdatum) \
+	( \
+	((attlen) == -1 && VARATT_IS_SHORT(DatumGetPointer(attdatum))) ? \
+	(uintptr_t) (cur_offset) : \
+	TYPEALIGN(attalignby, cur_offset))
+
 /*
  * att_align_pointer performs the same calculation as att_align_datum,
  * but is used when walking a tuple.  attptr is the current actual data
@@ -112,6 +122,16 @@ fetch_att(const void *T, bool attbyval, int attlen)
 	att_align_nominal(cur_offset, attalign) \
 )
 
+/*
+ * Similar to att_align_pointer, but accepts a number of bytes, typically from
+ * CompactAttribute.attalignby to align the pointer by.
+ */
+#define att_pointer_alignby(cur_offset, attalignby, attlen, attptr) \
+	( \
+	((attlen) == -1 && VARATT_NOT_PAD_BYTE(attptr)) ? \
+	(uintptr_t) (cur_offset) : \
+	TYPEALIGN(attalignby, cur_offset))
+
 /*
  * att_align_nominal aligns the given offset as needed for a datum of alignment
  * requirement attalign, ignoring any consideration of packed varlena datums.
@@ -138,6 +158,13 @@ fetch_att(const void *T, bool attbyval, int attlen)
 	   ))) \
 )
 
+/*
+ * Similar to att_align_nominal, but accepts a number of bytes, typically from
+ * CompactAttribute.attalignby to align the offset by.
+ */
+#define att_nominal_alignby(cur_offset, attalignby) \
+	TYPEALIGN(attalignby, cur_offset)
+
 /*
  * att_addlength_datum increments the given offset by the space needed for
  * the given Datum variable.  attdatum is only accessed if we are dealing