Section: New Results

Fast and reliable unwinding via DWARF tables

Participants : Theophile Bastian, Francesco Zappa Nardelli.

DWARF is a widely-used debugging data format. DWARF is obviously relied upon by debuggers, but it plays an unexpected role in the runtime of high-level programming languages and in the implementation of program analysis tools. The debug information itself can be pervaded by subtle bugs, making the whole infrastructure unreliable. In this project we are investigating techniques and tools to perform validation and synthesis of the DWARF stack unwinding tables, to speedup DWARF-based unwinding, as well as exploring adventurous projects that can be built on top of reliable DWARF information.

At the time of writing, we have a tool that can validate DWARF unwind tables generated by mainstream compilers; the approach is effective, we found a problem in Clang table generation and several in GLIBC inline-assembly snippets. We also designed and implemented a tool that can synthesise DWARF unwind tables from binary that lacks them (e.g. because the compiler did not generate them - immediate applications: JITs assembly, inline assembly, ...). Additionally we have designed and implemented a ahead-of-time compiler of DWARF unwind tables to assembly, and an ad-hoc unwinder integrated with the defacto standard unwinder libuwind. It can speed up unwinding by a factor between 25x and 60x (depending on application), with a 2.5x size overhead for unwind information.

Discussion is in progress to get these tools included in mainstream tool (e.g. the GNU profiler Perf).