2022-07-28 –, Purple
Cthulhu.jl is a highly useful tool for performance engineering as well as general debugging of Julia programs. However, as the name implies, one can quickly descend into the abyss that is Julia's compilation pipeline and get lost in the vast amounts of code even modest looking Julia functions may end up generating. I present a combination of Cthulhu.jl with a step-by-step debugger, showing concrete results every step along the type lattice to make compilation more interpretable.
One of the main motivations for this work was the use case of debugging source-to-source automatic differentiation on scientific codes as exemplified by Zygote.jl, which emit code that is significantly more complex than that of the original program. That can make it difficult to correctly identify intermediate steps. This is often complicated by the fact that the more complicated code can lead to results not being inferred to a concrete type anymore.
I leverage the already existing infrastructure in JuliaInterpreter.jl to enable explorative analysis of what the code does by interpreting the program based on concrete input values. Because interpretation works on a statement-by-statement basis just as inference does, this allows the user to go back and look at what the interpreter computed for any intermediate steps or see which branch actually ended up being taken.
One of the main challenges was the fact that JuliaInterpreter.jl was designed to interpret untyped Julia IR, which has slightly different semantics to IR after inference which again differs from the semantics of IR after all other Julia-specific optimizations such as inlining. A prototype currently exists in https://github.com/JuliaDebug/Cthulhu.jl/pull/214. I plan to introduce a flexible plugin infrastructure for this to be able to develop most of this outside of Cthulhu first. Support for step-by-step execution and for interpreting optimized Julia IR is also being worked on.
In this talk I aim to first give an overview on how Cthulhu.jl differs from a debugger and the various advantages and disadvantages of both approaches. I will then explain how I combined the two and how users can take advantage of these new capabilities in their own workflows. While I will be primarily targeting intermediate to advanced Julia users, I believe this could even be of use to those who have not used Cthulhu.jl before, because it allows for a much more interactive exploration of the intricacies of Julia IR.
I am currently a master's student in Computational Science and Engineering at the MIT Julia Lab. I have been a Julia user since 2018 and am one of the maintainers of the Julia language. Among other parts of the language, I work on Julia's front end and contribute to packages such as Cthulhu.jl, JuliaInterpreter.jl, and ArtifactUtils.jl.