Hybrid Specialization: A Trade-off Between Static and Dynamic Specialization

Code specialization is a well-known technique used to produce more efficient code from a generic one. This tech-nique is widely used, from high level languages to opti-mized libraries like FFTW. It works by substituting a for-mal input value by an effective value, and can be done ei-ther statically or dynamically. Static specialization makes use of data that is expected to be frequently used, whereas, dynamic specialization uses the actual values at run-time. We propose a novel hybrid method combining both the static and the dynamic specialization. With the code specialized for different values, the com-piler is able to generate highly optimized code. This lets the compiler fully exploit the ILP provided by modern pro-cessors. Code specialization is most effectively performed at runtime due to the unavailability of input values. Thi