The OpenJDK mailing list has been given an update on the progress of Project Valhalla, the OpenJDK project designed to "explore and incubate" advanced Java VM and language feature candidates such as value types and generic specialization.
Brian Goetz, Oracle's Java language architect, provided the update in a post to the OpenJDK mailing list giving details on progress so far. Valhalla started back in 2014, and in the intervending five years Goetz has overseen a number of prototypes to explore options.
In his post, Goetz said:
"We believe we are now at the point where we have a clear and coherent path to enhance the Java language and virtual machine with value types, have them interoperate cleanly with existing generics, and have a compatible path for migrating our existing value-based classes to inline classes and our existing generic classes to specialized generics."
The core feature that is of interest to the Valhalla group is inline types, but as Goetz says:
"such a fundamental perturbation of the JVM type system brings with it a host of other features and challenges, such as specialized generics, as well as tools for enabling compatible migration of existing APIs to inline types and specialized generics."
The most recent prototype of Valhalla, dubbed “L World” (because inline classes can share the L carrier with object references,managed to unify references and inlines under one set of type descriptors, bytecodes, and types, without significant compromises. This is what has led to Goetz's statement about having a clear and coherent path to enhance the Java language and virtual machine with value types.
Goetz says the goal is to:
"Give developers the control to match data layouts with the performance model of today's hardware, providing Java developers with an easier path to flat (cache-efficient) and dense (memory-efficient) data layouts without compromising abstraction or type safety."
The post provides a full update on Valhalla, including links to other documents and some nice diagrams.
IBM has released new Granite models that it says provide state-of-the-art performance relative to model size. The Granite 3.0 collection includes a new, instruction-tuned, dense decoder-only LLM.
