Accelerating Program Analyzers by Unleashing of Power of Compiler Optimizations

rainoftime@gmail.com

Problem statement

Given an input LLVM bitcode (the program) I and a static analyzer P, we aim to generate a compiler optimization strategy F that can transform I to a new bitcode I', such that P runs faster on I'.

Approach

Black-box optimization

Currently, we only use runtime as feedback and have implemented two modes:

• (parallel) random search
• genetic algorithm-based search

NOTE: The example code uses LLVM 3.6 (and only a subset of the options)

White-box (or Grey-box) Prediction

• Features + ML?
• Deep learning

Applications

• SVF: https://github.com/SVF-tools/SVF
• IKOS: https://github.com/NASA-SW-VnV/ikos