| This thesis demonstrates that careful selection of compiler transformations can improve the output and reduce the compile-time cost of adaptive compilation. Compiler effectiveness depends on the order of code transformations applied. Adaptive compilation, then, uses empirical search to tune the transformation sequence for each program. This method achieves higher performance than traditional compilers, but often requires large compilation times. Previous research reduces compilation time by tuning the search process. This thesis, instead, tunes the search space by adding a loop unroller, addressing a deficiency in our compiler. Despite increasing the searchspace size, this change results in more effective and efficient searching. Averaged across nine benchmarks, the adaptive compiler produces code, more quickly, that is 10% faster. Unfortunately, implementing a loop unroller is non-trivial for our low-level intermediate language. Therefore, this thesis also contributes an algorithm for identifying and unrolling loops in the absence of high-level loop structures. |
