Section: New Results

Dynamic Languages: Virtual Machines

Benzo: Reflective Glue for Low-level Programming. The goal of high-level low-level programming is to bring the abstraction capabilities of high-level languages to the system programming domain, such as virtual machines (VMs) and language runtimes. However, existing solutions are bound to compilation time and expose limited possibilities to be changed at runtime and from language-side. They do not fit well with fully reflective languages and environments. We propose Benzo1, a lightweight framework for high-level low-level programming that allows developers to generate and execute at runtime low-level code. It promotes the implementation, and dynamic modification, of system components with high-level language tools outperforming existing dynamic solutions. Since Benzo is a general framework we choose three applications that cover an important range of the spectrum of system programming for validating the infrastructure: a Foreign Function Interface (FFI), primitives instrumentation and a just-in-time bytecode compiler (JIT). With Benzo we show that these typical VM-level components are feasible as reflective language-side implementations. Due to its unique combination of high-level reflection and low-level programming, Benzo shows better performance for these three applications than the comparable high-level implementations. [30]

A bytecode set for adaptive optimizations. The Cog virtual machine features a bytecode interpreter and a baseline Just-in-time compiler. To reach the performance level of industrial quality virtual machines such as Java HotSpot, it needs to employ an adaptive inlining com-piler, a tool that on the fly aggressively optimizes frequently executed portions of code. We decided to implement such a tool as a bytecode to bytecode optimizer, implemented above the virtual machine, where it can be written and developed in Smalltalk. The optimizer we plan needs to extend the operations encoded in the bytecode set and its quality heavily depends on the bytecode set quality. The current bytecode set understood by the virtual machine is old and lacks any room to add new operations. We decided to implement a new bytecode set, which includes additional bytecodes that allow the Just-in-time compiler to generate less generic, and hence simpler and faster code sequences for frequently executed primitives. The new bytecode set includes traps for validating speculative inlining de-cisions and is extensible without compromising optimization opportunities. In addition, we took advantage of this work to solve limitations of the current bytecode set such as the maximum number of instance variable per class, or number of literals per method. We plan to have it in production in the Cog virtual machine and its Pharo, Squeak and Newspeak clients in the coming year. [43]