NativeBuferring,一种零分配的数据类型[下篇]
上文说到Unmanaged、BufferedBinary和BufferedString是NativeBuffering支持的三个基本数据类型,其实我们也可以说NativeBuffering只支持Unmanaged和IReadOnlyBufferedObject<T>两种类型,BufferedString、NativeBuffering和通过Source Generator生成的BufferedMessage类型,以及下面介绍的几种集合和字典类型,都实现了IReadOnlyBufferedObject<T>接口。
一、IReadOnlyBufferedObject<T>
二、集合
三、字典
四、为什么不直接返回接口?
一、IReadOnlyBufferedObject<T>
顾名思义,IReadOnlyBufferedObject<T>表示一个针对缓冲字节序列创建的只读数据类型。如下面的代码片段所示,该接口只定义了一个名为Parse的静态方法,意味着对于任何一个实现了该接口的类型,对应的实例都可以利用一个代表缓冲字节序列的NativeBuffer的对象进行创建。
public interface IReadOnlyBufferedObject<T> where T: IReadOnlyBufferedObject<T> { static abstract T Parse(NativeBuffer buffer); } public unsafe readonly struct NativeBuffer { public byte[] Bytes { get; } public void* Start { get; } public NativeBuffer(byte[] bytes, void* start) { Bytes = bytes ?? throw new ArgumentNullException(nameof(bytes)); Start = start; } public NativeBuffer(byte[] bytes, int index = 0) { Bytes = bytes ?? throw new ArgumentNullException(nameof(bytes)); Start = Unsafe.AsPointer(ref bytes[index]); } }
由于IReadOnlyBufferedObject<T>是NativeBuffering支持的基础类型,而生成的BufferedMessage类型也实现了这个接口。通过这种“无限嵌套”的形式,我们可以定义一个具有任意结构的数据类型。比如我们具有如下这个表示联系人的Contact类型,我们需要利用它作为“源类型”生成对应BufferedMessage类型。
[BufferedMessageSource] public partial class Contact { public Contact(string id, string name, Address address) { Id = id; Name = name; ShipAddress = address; } public string Id { get; } public string Name { get; } public Address ShipAddress { get; } } [BufferedMessageSource] public partial class Address { public string Province { get; } public string City { get; } public string District { get; } public string Street { get; } public Address(string province, string city, string district, string street) { Province = province ?? throw new ArgumentNullException(nameof(province)); City = city ?? throw new ArgumentNullException(nameof(city)); District = district ?? throw new ArgumentNullException(nameof(district)); Street = street ?? throw new ArgumentNullException(nameof(street)); } }
Contact具有Id、Name和ShipAddress 三个数据成员,ShipAddress 对应的Address又是一个复合类型,具有四个表示省、市、区和介绍的字符串类型成员。现在我们为Contact和Address这两个类型生成对应的ContactBufferedMessage和AddressBufferedMessage。
public unsafe readonly struct ContactBufferedMessage : IReadOnlyBufferedObject<ContactBufferedMessage> { public NativeBuffer Buffer { get; } public ContactBufferedMessage(NativeBuffer buffer) => Buffer = buffer; public static ContactBufferedMessage Parse(NativeBuffer buffer) => new ContactBufferedMessage(buffer); public BufferedString Id => Buffer.ReadBufferedObjectField<BufferedString>(0); public BufferedString Name => Buffer.ReadBufferedObjectField<BufferedString>(1); public AddressBufferedMessage ShipAddress => Buffer.ReadBufferedObjectField<AddressBufferedMessage>(2); } public unsafe readonly struct AddressBufferedMessage : IReadOnlyBufferedObject<AddressBufferedMessage> { public NativeBuffer Buffer { get; } public AddressBufferedMessage(NativeBuffer buffer) => Buffer = buffer; public static AddressBufferedMessage Parse(NativeBuffer buffer) => new AddressBufferedMessage(buffer); public BufferedString Province => Buffer.ReadBufferedObjectField<BufferedString>(0); public BufferedString City => Buffer.ReadBufferedObjectField<BufferedString>(1); public BufferedString District => Buffer.ReadBufferedObjectField<BufferedString>(2); public BufferedString Street => Buffer.ReadBufferedObjectField<BufferedString>(3); }
如下的程序演示了如何将一个Contact对象转换成字节数组,然后利用这这段字节序列生成一个ContactBufferedMessage对象。给出的调试断言验证了Contact和ContactBufferedMessage对象承载了一样的数据,fixed关键字是为了将字节数组“固定住”。(源代码从这里下载)
using NativeBuffering; using System.Diagnostics; var address = new Address("Jiangsu", "Suzhou", "Industory Park", "#328, Xinghu St"); var contact = new Contact("123456789", "John Doe", address); var size = contact.CalculateSize(); var bytes = new byte[size]; var context = new BufferedObjectWriteContext(bytes); contact.Write(context); unsafe { fixed (byte* _ = bytes) { var contactMessage = ContactBufferedMessage.Parse(new NativeBuffer(bytes)); Debug.Assert(contactMessage.Id == "123456789"); Debug.Assert(contactMessage.Name == "John Doe"); Debug.Assert(contactMessage.ShipAddress.Province == "Jiangsu"); Debug.Assert(contactMessage.ShipAddress.City == "Suzhou"); Debug.Assert(contactMessage.ShipAddress.District == "Industory Park"); Debug.Assert(contactMessage.ShipAddress.Street == "#328, Xinghu St"); } }
二、集合
NativeBuffering同样支持集合。由于Unmanaged和IReadOnlyBufferedObject<T>是两种基本的数据类型,它们的根据区别在于:前者的长度有类型本身决定,是固定长度类型,后者则是可变长度类型。元素类型为Unmanaged和IReadOnlyBufferedObject<T>的集合分别通过ReadOnlyFixedLengthTypedList<T>和ReadOnlyVaraibleLengthTypedList<T>类型(结构体)表示,它们同样实现了IReadOnlyBufferedObject<T>接口。ReadOnlyFixedLengthTypedList<T>采用如下的字节布局:集合元素数量(4字节整数)+所有元素的字节内容(下图-上)。对于ReadOnlyVaraibleLengthTypedList<T>类型,我们会在前面为每个元素添加一个索引(4字节的整数),该索引指向目标元素在整个缓冲区的偏移量(下图-下)。
以如下所示的Entity为例,它具有两个数组类型的属性成员Collection1和Collection2,数组元素类型分别为Foobar和double,它们分别代表了上述的两种集合类型。
[BufferedMessageSource] public partial class Entity { public Foobar[] Collection1 { get; } public double[] Collection2 { get; } public Entity(Foobar[] collection1, double[] collection2) { Collection1 = collection1; Collection2 = collection2; } } [BufferedMessageSource] public partial class Foobar { public int Foo { get; } public string Bar { get; } public Foobar(int foo, string bar) { Foo = foo; Bar = bar; } }
NativeBuffering.Generator会将作为“源类型”的Entity和Foobar类型的生成对应的BufferedMessage类型(EntityBufferredMessage和FoobarBufferedMessage)。从EntityBufferredMessage类型的定义可以看出,两个集合属性的分别是ReadOnlyVariableLengthTypeList<FoobarBufferedMessage>和ReadOnlyFixedLengthTypedList<double>。
public unsafe readonly struct EntityBufferedMessage : IReadOnlyBufferedObject<EntityBufferedMessage> { public NativeBuffer Buffer { get; } public EntityBufferedMessage(NativeBuffer buffer) => Buffer = buffer; public static EntityBufferedMessage Parse(NativeBuffer buffer) => new EntityBufferedMessage(buffer); public ReadOnlyVariableLengthTypeList<FoobarBufferedMessage> Collection1 => Buffer.ReadBufferedObjectCollectionField<FoobarBufferedMessage>(0); public ReadOnlyFixedLengthTypedList<System.Double> Collection2 => Buffer.ReadUnmanagedCollectionField<System.Double>(1); } public unsafe readonly struct FoobarBufferedMessage : IReadOnlyBufferedObject<FoobarBufferedMessage> { public NativeBuffer Buffer { get; } public FoobarBufferedMessage(NativeBuffer buffer) => Buffer = buffer; public static FoobarBufferedMessage Parse(NativeBuffer buffer) => new FoobarBufferedMessage(buffer); public System.Int32 Foo => Buffer.ReadUnmanagedField<System.Int32>(0); public BufferedString Bar => Buffer.ReadBufferedObjectField<BufferedString>(1); }
两个集合类型都实现了IEnumerable<T>接口,还提供了索引。下面的代码演示了以索引的形式提取集合元素(源代码从这里下载)。
using NativeBuffering; using System.Diagnostics; var entity = new Entity( collection1: new Foobar[] { new Foobar(1, "foo"), new Foobar(2, "bar") }, collection2: new double[] { 1.1, 2.2 }); var bytes = new byte[entity.CalculateSize()]; var context = new BufferedObjectWriteContext(bytes); entity.Write(context); unsafe { fixed (byte* p = bytes) { var entityMessage = EntityBufferedMessage.Parse(new NativeBuffer(bytes)); var foobar = entityMessage.Collection1[0]; Debug.Assert(foobar.Foo == 1); Debug.Assert(foobar.Bar == "foo"); foobar = entityMessage.Collection1[1]; Debug.Assert(foobar.Foo == 2); Debug.Assert(foobar.Bar == "bar"); Debug.Assert(entityMessage.Collection2[0] == 1.1); Debug.Assert(entityMessage.Collection2[1] == 2.2); } }
三、字典
从数据的存储来看,字典就是键值对的集合,所以我们采用与集合一致的存储形式。NativeBuffering对集合的Key作了限制,要求其类型只能是Unmanaged和字符串(String/BufferredString)。按照Key和Value的类型组合,我们一共定义了四种类型的字典类型,它们分别是:
- ReadOnlyUnmanagedUnmanagedDictionary<TKey, TValue>:Key=Unmanaged; Value = Unmanaged
- ReadOnlyUnmanagedBufferedObjectDictionary<TKey, TValue>:Key=Unmanaged; Value = IReadOnlyBufferedObject<TValue>
- ReadOnlyStringUnmanagedDictionary<TValue>:Key=String/BufferredString; Value = Unmanaged
- ReadOnlyStringBufferedObjectDictionary<TValue>:Key=String/BufferredString; Value = IReadOnlyBufferedObject<TValue>
如果Key和Value的类型都是Unmanaged,键值对就是定长类型,所以我们会采用类似于ReadOnlyFixedLengthTypedList<T>的字节布局方式(下图-上),至于其他三种字典类型,则采用类似于ReadOnlyVaraibleLengthTypedList<T>的字节布局形式(下图-下)。
但是这仅仅解决了字段数据存储的问题,字典基于哈希检索定位的功能是没有办法实现的。这里我们不得不作出妥协,四种字典的索引均不能提供时间复杂度O(1)的哈希检索方式。为了在现有的数据结构上使针对Key的查找尽可能高效,在生成字节内容之前,我们会按照Key对键值对进行排序,这样我们至少可以采用二分法的形式进行检索,所以四种类型的字典的索引在根据指定的Key查找对应Value,对应的时间复杂度为Log(N)。如果字典包含的元素比较多,这样的查找方式不能满足我们的需求,我们可以I将它们转换成普通的Dictionary<TKey, TValue>类型,但是这就没法避免内存分配了。
我们照例编写一个简答的程序来演示针对字典的使用。我们定义了如下这个Entity作为“源类型”,它的四个属性对应的字典类型刚好对应上述四种键值对的组合。从生成的EntityBufferedMessage类型可以看出,四个成员的类型正好对应上述的四种字典类型。
[BufferedMessageSource] public partial class Entity { public Dictionary<int, long> Dictionary1 { get; set; } public Dictionary<int, string> Dictionary2 { get; set; } public Dictionary<string, long> Dictionary3 { get; set; } public Dictionary<string, string> Dictionary4 { get; set; } } public unsafe readonly struct EntityBufferedMessage : IReadOnlyBufferedObject<EntityBufferedMessage> { public NativeBuffer Buffer { get; } public EntityBufferedMessage(NativeBuffer buffer) => Buffer = buffer; public static EntityBufferedMessage Parse(NativeBuffer buffer) => new EntityBufferedMessage(buffer); public ReadOnlyUnmanagedUnmanagedDictionary<System.Int32, System.Int64> Dictionary1 => Buffer.ReadUnmanagedUnmanagedDictionaryField<System.Int32, System.Int64>(0); public ReadOnlyUnmanagedBufferedObjectDictionary<System.Int32, BufferedString> Dictionary2 => Buffer.ReadUnmanagedBufferedObjectDictionaryField<System.Int32, BufferedString>(1); public ReadOnlyStringUnmanagedDictionary<System.Int64> Dictionary3 => Buffer.ReadStringUnmanagedDictionaryField<System.Int64>(2); public ReadOnlyStringBufferedObjectDictionary<BufferedString> Dictionary4 => Buffer.ReadStringBufferedObjectDictionaryField<BufferedString>(3); }
如下的代码演示了基于四种字典类型基于“索引”的检索方式(源代码从这里下载)。
using NativeBuffering; using System.Diagnostics; var entity = new Entity { Dictionary1 = new Dictionary<int, long> { { 1, 1 }, { 2, 2 }, { 3, 3 } }, Dictionary2 = new Dictionary<int, string> { { 1, "foo" }, { 2, "bar" }, { 3, "baz" } }, Dictionary3 = new Dictionary<string, long> { { "foo", 1 }, { "bar", 2 }, { "baz", 3 } }, Dictionary4 = new Dictionary<string, string> { { "a", "foo" }, { "b", "bar" }, { "c", "baz" } } }; var bytes = new byte[entity.CalculateSize()]; var context = new BufferedObjectWriteContext(bytes); entity.Write(context); unsafe { fixed (void* _ = bytes) { var bufferedMessage = EntityBufferedMessage.Parse(new NativeBuffer(bytes)); ref var value1 = ref bufferedMessage.Dictionary1.AsRef(1); Debug.Assert(value1 == 1); ref var value2 = ref bufferedMessage.Dictionary3.AsRef("baz"); Debug.Assert(value2 == 3); var dictionary1 = bufferedMessage.Dictionary1; Debug.Assert(dictionary1[1] == 1); Debug.Assert(dictionary1[2] == 2); Debug.Assert(dictionary1[3] == 3); var dictionary2 = bufferedMessage.Dictionary2; Debug.Assert(dictionary2[1] == "foo"); Debug.Assert(dictionary2[2] == "bar"); Debug.Assert(dictionary2[3] == "baz"); var dictionary3 = bufferedMessage.Dictionary3; Debug.Assert(dictionary3["foo"] == 1); Debug.Assert(dictionary3["bar"] == 2); Debug.Assert(dictionary3["baz"] == 3); var dictionary4 = bufferedMessage.Dictionary4; Debug.Assert(dictionary4["a"] == "foo"); Debug.Assert(dictionary4["b"] == "bar"); Debug.Assert(dictionary4["c"] == "baz"); } }
四、为什么不直接返回接口
针对集合,NativeBuffering提供了两种类型;针对字典,更是定义了四种类型,为什么不直接返回IList<T>/IDictionary<TKey,TValue>(或者IReadOnlyList<T>/IReadOnlyDictionary<TKey,TValue>)接口呢?这主要有两个原因,第一:为了尽可能地减少内存占用,我们将四种字典类型都定义成了结构体,如果使用接口的话会导致装箱;第二,四种字典类型的提供的API是有差异的,比如ReadOnlyFixedLengthTypedList<T> 和ReadOnlyUnmanagedUnmanagedDictionary<TKey, TValue>都提供了一个额外的AsRef方法,它直接返回值的引用(只读)。如果这个值被定义成一个成员较多的结构体,传引用的方式可以避免较多的拷贝。
public readonly unsafe struct ReadOnlyFixedLengthTypedList<T> : IReadOnlyList<T>, IReadOnlyBufferedObject<ReadOnlyFixedLengthTypedList<T>> where T: unmanaged { public readonly ref T AsRef(int index); ... } public unsafe readonly struct ReadOnlyUnmanagedUnmanagedDictionary<TKey, TValue> : IReadOnlyDictionary<TKey, TValue>, IReadOnlyBufferedObject<ReadOnlyUnmanagedUnmanagedDictionary<TKey, TValue>> where TKey : unmanaged, IComparable<TKey> where TValue : unmanaged { public readonly ref TValue AsRef(TKey index) ; ... }