Profile-driven JIT compilers like those in most JVMs provide good performance on average, but they are a black box with highly nondeterministic behavior. Thus, achieving top performance, predictably and consistently, is very hard. This talk will present Lancet, a research JIT compiler written in Scala that allows the running application to take full control of the JIT process. This includes the ability to compile multiple specialized versions of code paths at runtime, which is key for removing abstraction overhead of high-level, generic code. Another key feature of Lancet are JIT macros, which enable execution of user code at JIT-compile time and thus open the door for "smart libraries" that come with domain-specific optimizations and checks. Unlike Scala macros, JIT macros are not a front-end system, but tightly integrated with regular JIT compiler optimizations and VM functionality like speculative optimization and re-optimization.

No outline is available