llc

The goal of the decompy project is essentially to make decompiled code more accurate by using machine learning to “move” the program through possible representations to an easily understandable state.

The current working method of the team is to use swaps performed by do what is essentially a code equivalence recognition against a known database of snippets. Using these “equivalencies” the idea is to swap for improved versions.

Our team is looking into various ways to generate these equivalencies, but has often settled on using opt, since that is essentially the root of what we are trying to undo (kind of similar to taking a function, plugging in a huge amount of inputs, generating the outputs and infering the inverse function from the results). One issue with this is that many of the issues we would be able to fix with minimal work using this method are isolated changes. I.E. the infamous fcd problem

c = b - a;
if (c != 0 && (c & 9223372036854775808) == (((b ^ c) & (b ^ a) & 9223372036854775808)))

to

if (b > a)

Unfortunately, these changes mostly occur not in the llvm itself during opt transformations, but at the level of target specific optimizations which occur during llc’s passes and which do not produce llvm.