About: A sidequest is an optional, not graded, not evaluated in anyway exploration into a topic. Sometimes it is just cool to know what is out there in the world. Sidequests come with a little descritions, presentations, related resources, and a sample code or exercise you can try. Side quests are likely a good area for further independent studies, research project, thesis, etc.

Machine Learning Compiler Optimizations

Problem Statement

There are many compiler optimizations provided by compilers like GCC, LLVM, MSVC, and Intel. Choosing which set of optimizations maximizes your code performance is challenging, though most compilers present -O1, -O2, -O3 to assist. Otherwise, an option is to run all of the compiler optimzations possible.

Proposed Exercise

Use a compiler (preferably LLVM) and the 'opt' tool to apply a variety of compiler optimizations.

Proposed Evaluation

• Find a benchmark (some C++ program that's big, and that you understand what it does. e.g. a program or database that runs a bunch of queries)
• Then fire up g++ and clang++ and run with different optimizations https://gcc.gnu.org/onlinedocs/gcc/Optimize-Options.html
  • Could just be for now -O0, -O1, -O2, -O3, etc.
• Then measure the performance (i.e. execution time with perf)
  • Note time, maybe other metrics like cache misses, cache hits, etc.

Key Resources

• [1] Improving Compiler Optimizations Using Machine Learning (Thesis)

Additional Resources

• [2] Machine Learning in Compiler Optimization (Survey Paper)
• [3] GCC List of compiler optimizations
• [4] Statistically Sound Performance Evaluation (Good paper to read to make sure our results are not noise)
• [5] Machine Guided Energy Efficient Compilation (Future diretions of project)
• [6] Deep Learning in Compilers
• [7] A Survey on Compiler Autotuning using Machine Learning