Source-level debugging of compiled code only works when compilers generate
the necessary metadata. Currently, that means it rarely works well, at
least in optimising ahead-of-time compilers like LLVM and GCC. I'll give an
overview of how compiler-generated metadata enables source-level debugging,
the challenges of making it work for optimised code, and our recent work on
doing better. Whereas compilers have so far taken a "best-effort" approach
with no particular correctness criterion, I'll outline a correctness
condition for local variable information that seems to balance the relevant
trade-offs. I'll then describe a tool we've built that can use this to
mechanically find valid LLVM bugs capturing avoidable losses or corruptions
of debug info. A theme will be how the textbook framing of compiler
optimisations as "eliminating" code or variables could be more
constructively thought of as "residualising" them into debug info; I'll
finish with some thoughts on what that could mean for how compilers are
built. All this is joint work with J. Ryan Stinnett.
