Profile-Guided, Multi-Version Binary Rewriting
This work addresses performance optimization for users of binary rewriting tools, offering a workflow to reduce instrumentation overhead, though it is incremental as it builds on existing binary rewriting techniques.
The paper tackles the high runtime overhead of static binary instrumentation by introducing a profile-guided, multi-version binary rewriting approach, which reduces the geometric overhead of shadow stack and block coverage from 7.6% and 161.3% to 1.4% and 4.0% on SPEC CPU 2017, and shadow stack overhead from about 20% to 3.5% on Apache HTTP Server.
The static instrumentation of machine code, also known as binary rewriting, is a power technique, but suffers from high runtime overhead compared to compiler-level instrumentation. Recent research has shown that tools can achieve near-to-zero overhead when rewriting binaries (excluding the overhead from the application specific instrumentation). However, the users of binary rewriting tools often have difficulties in understanding why their instrumentation is slow and how to optimize their instrumentation. We are inspired by a traditional program optimization workflow, where one can profile the program execution to identify performance hot spots, modify the source code or apply suitable compiler optimizations, and even apply profile-guided optimization. We present profile-guided, Multi-Version Binary Rewriting to enable this optimization workflow for static binary instrumentation. Our new techniques include three components. First, we augment existing binary rewriting to support call path profiling; one can interactively view instrumentation costs and understand the calling contexts where the costs incur. Second, we present Versioned Structure Binary Editing, which is a general binary transformation technique. Third, we use call path profiles to guide the application of binary transformation. We apply our new techniques to shadow stack and basic block code coverage. Our instrumentation optimization workflow helps us identify several opportunities with regard to code transformation and instrumentation data layout. Our evaluation on SPEC CPU 2017 shows that the geometric overhead of shadow stack and block coverage is reduced from 7.6% and 161.3% to 1.4% and 4.0%, respectively. We also achieve promising results on Apache HTTP Server, where the shadow stack overhead is reduced from about 20% to 3.5%.