internal/encoding/tag: use proto3 defaults if proto3

proto3 legacy struct field tags are improperly parsed as proto2,
resulting in fd.L1.EditionFeatures.IsFieldPresence set to true when it
should have been false. This causes coderFieldInfo.isPointer to be true
for said fields. When the encode process (particularly
MessageInfo.sizePointerSlow) sees isPointer=true for the field, it
attempts to interpret the field as a pointer, which results in a "fatal
error: checkptr: misaligned pointer conversion" runtime error if the
field address happens to not be pointer-aligned. This only happens if
building with -race/-asan/-msan as they enable checkptr.

To avoid this issue, parse from proto3 edition defaults if the proto3
key is found in the struct field tag.

Change-Id: Id81ab38fe7efd437d0748a485653d27b1acc3afb
Reviewed-on: https://go-review.googlesource.com/c/protobuf/+/716360
TryBot-Bypass: Michael Stapelberg <stapelberg@google.com>
Auto-Submit: Nicolas Hillegeer <aktau@google.com>
Reviewed-by: Nicolas Hillegeer <aktau@google.com>
LUCI-TryBot-Result: Go LUCI <golang-scoped@luci-project-accounts.iam.gserviceaccount.com>
Run-TryBot: Damien Neil <dneil@google.com>
Reviewed-by: Damien Neil <dneil@google.com>
Reviewed-by: Chressie Himpel <chressie@google.com>
Reviewed-by: Michael Stapelberg <stapelberg@google.com>

Author: gh2o

internal/encoding/tag/tag.go

@@ -32,7 +32,7 @@ var byteType = reflect.TypeOf(byte(0))
 func Unmarshal(tag string, goType reflect.Type, evs protoreflect.EnumValueDescriptors) protoreflect.FieldDescriptor {
 	f := new(filedesc.Field)
 	f.L0.ParentFile = filedesc.SurrogateProto2
-	f.L1.EditionFeatures = f.L0.ParentFile.L1.EditionFeatures
+	packed := false
 	for len(tag) > 0 {
 		i := strings.IndexByte(tag, ',')
 		if i < 0 {
@@ -108,7 +108,7 @@ func Unmarshal(tag string, goType reflect.Type, evs protoreflect.EnumValueDescri
 				f.L1.StringName.InitJSON(jsonName)
 			}
 		case s == "packed":
-			f.L1.EditionFeatures.IsPacked = true
+			packed = true
 		case strings.HasPrefix(s, "def="):
 			// The default tag is special in that everything afterwards is the
 			// default regardless of the presence of commas.
@@ -121,6 +121,13 @@ func Unmarshal(tag string, goType reflect.Type, evs protoreflect.EnumValueDescri
 		tag = strings.TrimPrefix(tag[i:], ",")
 	}
 
+	// Update EditionFeatures after the loop and after we know whether this is
+	// a proto2 or proto3 field.
+	f.L1.EditionFeatures = f.L0.ParentFile.L1.EditionFeatures
+	if packed {
+		f.L1.EditionFeatures.IsPacked = true
+	}
+
 	// The generator uses the group message name instead of the field name.
 	// We obtain the real field name by lowercasing the group name.
 	if f.L1.Kind == protoreflect.GroupKind {

internal/encoding/tag/tag_test.go

@@ -23,6 +23,7 @@ func Test(t *testing.T) {
 	fd.L1.Cardinality = protoreflect.Repeated
 	fd.L1.Kind = protoreflect.BytesKind
 	fd.L1.Default = filedesc.DefaultValue(protoreflect.ValueOf([]byte("hello, \xde\xad\xbe\xef\n")), nil)
+	fd.L1.EditionFeatures = fd.L0.ParentFile.L1.EditionFeatures
 
 	// Marshal test.
 	gotTag := tag.Marshal(fd, "")
@@ -37,4 +38,7 @@ func Test(t *testing.T) {
 	if !proto.Equal(protodesc.ToFieldDescriptorProto(gotFD), protodesc.ToFieldDescriptorProto(wantFD)) {
 		t.Errorf("Umarshal() mismatch:\ngot  %v\nwant %v", gotFD, wantFD)
 	}
+	if gotFD.(*filedesc.Field).L1.EditionFeatures != wantFD.L1.EditionFeatures {
+		t.Errorf("Unmarshal() did not set proto3 edition features")
+	}
 }