java valuetype_Value type and Packaged Object for Java

前段时间有人问了我下关于Java 中Value Type的概念，并转了我一篇王垠去年的一篇博客Java 有值类型吗？

Java语言层面上并不能算是支持value type。Java有两种类型： primitive type and object type。前者如我们大家经常见到的 `int` `float`，从JVM的角度上看，变量的数据是直接在JVM operand stack上面的。而后者，所有的对象则是从heap上申请，每个对象的field则是通过reference 来联系起来。

当然这篇博文主要不是针对王垠的博文(自己个人理解垠神在PL 的Type System方面的学识远远超过了本人)。这篇文章还是还是比较value type and packaged object.

Value Type

最早关于Java Value Type这个词的时候是先前有一次旁听Doug lea和某博士的关于Packaged object的讨论会，后来我的合作方给我稍微普及, 并给了IBM clear过的链接 Valhalla(Open JDK experiment项目)以及他们最新的 Minimal Value Types.

Value Type的思想是关于Object Model。每个Java Object有一个Header(有的地方叫做Metadata)：包含lock-word, ptr, 等信息。通过lock-word 可以实现object synchronization (e.g., lock-word)，而　ptr可以获取object相应的相Class信息。Value Type的目标是Headerless that is no more header( or no identify ) for an object on the heap. The benefits would be on GC, memory usage, potential cache locality (Refer Valhalla home for more).

当然，任何的好处都不是没有代价的。Because of no header and identify, it would be problem to call methods of a value type object: hashcode, getClass, wait, notify. Another problem is about the object synchronization, which strongly depends on marker bit of an object's header that is no longer available.

对于普通的开发人员来讲，value type的出炉需要保持对原有Java应用的支持，数据模型的变更应该对使用者是透明的。

Packaged Object

Packaged Object is an experimental enhancement for J9 JVM to improve serialization, Java object I/O cross on-heap and off-heap (native memory), and object layout alignment. In the packaged object data model, data are organized in a multi-tier manner (like the nested C structs), in which the object’s data is nested in its enclosing Packed Object instead of being pointed to by an object reference as in the traditional Java object data model.

This project stopped at 2014(?). The main disclosure document is Marcel Mitran's ``IBM Java PackedObjects: An Overview`` (You get get it via google), and some publications about synchronization from IBM CAS. My post will only based on these document.

The packaged object's goals：Remove object metadata. This point is similar to the Value Type.

Off-head and on-head serialization. As what is shown at marcel's slides，the copy and serialization can be avoided avoided by using a packaged header at on-heap to reference the packaged object at off-heap directly.

...

Because the header(the lock-word) is removed, the existing Java synchronization does not work for packaged object directly. Developer has to achieve the thread-safe via:在packaged object上一层使用java 已有的lock机制。

Concurrent Multi-way tree 以及其衍生的方法。这个方法主要是IBM CAS中的Dr Bing Yang提出来实现的，大概意思是将根据multi-tier PackedObject data structure来构建一个multiple-way tree, 在读写 packaged data 的时候通过操作(Search, add, remove)这个multiple-way tree来实现同步。具体的方式可以参考下CASCON 2016 best paper: Bing Yang: A concurrent multiway tree using the lazy locking mechanism.

本来这篇博文还安排了一部分关于GraphJIT, a jit bytecode compiler that simplify graph structure. GraphJIT 也能达到类似Packaged object和value type的部分效果。