The motivation for this change is to give tree-style adapters precedence order in registration. This fixes the test I committed earlier today, where registration order was not honored.
This renamed ParameterizedTypeHandlerMap to the shorter 'TypeMap'. For type adapters, this is now only used for type hierarchy. We still need non-hierarchy support in TypeMap for instance creators; I'll be looking for workarounds to see if further simplification is possible here.
The most notable impact of this change is that it adds several new public APIs:
- The TypeAdapter abstract class
- The TypeAdapter.Factory interface
- Four new methods on GsonBuilder to register streaming TypeAdapters (via Factory, via Class, via TypeToken, and as a type hierarchy)
- Three new methods on Gson to lookup streaming TypeAdapters (by type, by class, and to get the next type adapter)
Still outstanding:
- Write beautiful prose to document the new APIs above
- Change GsonBuilder's precedence so that both old and new-style type adapters are registered in one lot
Replaced DEFAULT_SERIALIZERS, DEFAULT_DESERIALIZERS and DEFAULT_INSTANCE_CREATORS with a single EMPTY_MAP.
Removed obsoleted TODO from Gson.
made ParameterizedTypeHandlerMap.makeUnmodifiable a builder method that returns this instance.
Add setSerializeNulls() to JsonWriter, so nulls can be skipped from serialization. This does not yet impact JsonElementWriter.
One change in behavior: if the only value is skipped, we now emit "null" rather than "".
GSON 1.x applies different rules for versioning for classes vs fields. So, if you deserialize a
JSON into a field that is supposed to be skipped, the field is set to null (or default value).
However, if you deserialize it to a top-level class, a default instance is returned.
GSON 2.x returns null for the top-level class.
Previously we would set the field if the created instance is being assigned to a field of another object. We wouldn't set it when the created instance is a collection element or the top-level object.
Created a new class for character/char tests, PrimitiveCharacterTest. This is the start of breaking down PrimitiveTest into smaller chunks.
When MiniGson can not find a type adapter for a type, revised the message to say can't handle instead of can't serialize.
Got rid of default Hierarchy Serializer/Deserializers from DefaultTypeAdapters.
Got rid of methods for gettting default serializers/deserializers/instance creators. Instead we reuse the static final instances.
Fixed warnings in TypeAdapters where a parameterized type T was hiding the parameterized T in methods.
Removed support to unwrap single element array of enums into enum values. Also removed the test that verifies this behavior.
Consolidated all of the different code paths that we use to construct instances. We now have an ObjectConstructor class that knows what type it constructs; this means that we don't need to ever do reflection to lookup a constructor at construction time.
Cleaned up some buggy type adapters, particularly around handling of null.
Removed dead code for object graph navigation.
Moved some classes into 'internal' so they are visible to the 'bind' subpackage.
Turned some TypeAdapterFactory/TypeAdapter pairs inside out so that the TypeAdapter is now the inner class. This is necessary so that the factories can take parameters.
Added an API to request the 'next' type adapter for a type. This allows type adapters to compose other type adapters. We're using this in two places:
- where the user has excluded a type from serialization but not deserialization, we need to use the "default" deserialization but interpose null on serialization. We create a type adapter that delegates for one and returns null for the other.
- similarly when a DOM type serializer is registered but no deserializer, or vice versa.
This is the biggest change to the MiniGson core.
For backwards compatibility, return null for the empty string.
Simplify JsonSerializationContext/JsonDeserializationContext to simply call through to GSON. SerializeDefault is currently unsupported.
More useful error messages when calling getAsBoolean on a JsonNull.
Remove currently unused MemoryRefStack. We might need this back again, though wiring it back in will be much more difficult because we don't interject ourselves between the users' various type adapters.
This is probably a break from the past: previous versions of Gson allowed truncating a floating point into a long or int. However, it wasn't consistent in this behavior. It disallowed converting a BigDecimal value into BigInteger, int or long. Refusing to deserialize such values is aligned with fail-fast approach of uncovering bugs.
This avoids needing to parse number if the equivalent object field doesn't exist.
It also avoids the performance penalty of trying to parse it eagerly as a big decimal, float etc.
