case-studies.tex

% !TEX root = thesis.tex
\chapter{Case Studies}
\label{ch:case-studies}

In this chapter we discuss three case studies to explore the end-to-end utility of the techniques proposed in \mychapter\ref{ch:profiling} and \mychapter\ref{ch:continuous}. In the first case study, we show an example of profile-driven optimization based on our multi-iteration path profiling techniques. In particular, we show that the code of a masked convolution filter used for image processing can be adapted to exploit the characteristics of the workload, achieving two-digit speedups in our experiments. We also discuss a possible application of multi-iteration path profiles to trace schedulers.

The second case study, which has been repeated and endorsed by the joint Artifact Evaluation process of CGO-PPoPP 2016, presents an example of adaptive type specialization enabled by \osrkit. The flexibility offered by its compensation code abstraction allowed us to implement an optimization mechanism for a higher-order construct of the MATLAB language that significantly advances the state of the art compared to previous approaches, nearly matching the performance of code optimized by hand.

%Finally, the third case study illustrates an application of our algorithms to automatically build OSR mappings to the context of source-level debugging of optimized code. On prominent C benchmarks, the \reconstruct\ algorithm is able to recovery the expected source-level values for the vast majority of user variables that become endangered due to the effects of classic compiler optimizations.

Finally, in the third case study we show that our OSR (On-Stack Replacement) compensation code algorithms also provide useful novel building blocks to integrate in optimized code debuggers. On prominent C benchmarks, the \reconstruct\ algorithm is able to recovery the expected source-level values for the vast majority of user variables that become endangered due to the effects of classic compiler optimizations.

%In this section we show that our OSR compensation code techniques provide novel building blocks that could be integrated in optimized code debuggers."

\input{case-study-kblpp}
\input{case-study-mcvm}
\input{case-study-debugging}

\section{Conclusions}
The series of case studies illustrated in this chapter give further evidence of the utility of our performance profiling and continuous program optimization techniques. Improved profile accuracy is useful to enhance the program understanding and optimization process: multi-iteration path profiles can reveal interesting optimization opportunities that an acyclic profiler would miss. On-stack replacement is essential technology for dynamic optimization and debugging: our OSR with compensation code abstraction enables transformations that change the program state deeply, while our algorithms for automatic OSR mapping construction can help debuggers in providing expected behavior for an optimized program to the user.