LIL: an architecture-neutral language for virtual-machine stubs

High-performance MREs (managed runtime environments) that run either Java or CLI applications require machine code sequences, called stubs, to implement such runtime support operations as object allocation, synchronization, and native method invocation. Due to the frequency of these operations, implementing stubs efficiently is critical for performance. Also, the number of different stubs that have to be created and maintained makes stub creation a sizable part of an MRE's implementation. Stubs typically require access to low-level resources such as registers and the call stack, and often must be specialized at runtime for particular classes or methods. Although stubs can be implemented by generating hand-crafted machine code at runtime, this approach is tedious and error-prone, and leads to stubs that are non-portable and difficult to maintain. To address these problems, we designed a domain-specific language, called LIL, for implementing stubs. LIL is low-level but architecture-neutral, allowing the creation of stubs that are both portable and efficient. LIL also abstracts away many implementation details, making stubs easier to read. It is lightweight enough to be used for dynamic stub generation. LIL's validity checker helps us to catch many errors early. Our preliminary experience using LIL indicates that it greatly eases development and maintenance of stubs without sacrificing performance.