Back

@kristerw super cool. using my TV tool for LLVM's AArch64 backend, I've had the best luck not using yarpgen, but using a custom LLVM IR mutation tool that we wrote. an analogous tool for GIMPLE would be super useful. here's a writeup about ours:

https://users.cs.utah.edu/~regehr/papers/cgo24.pdf

@kristerw also, I've mostly finished a paper about our tool. if you're interested, I'll share a draft once it's in the final editing phase.

@zwarich @kristerw that's the goal for sure! but real functions (after inlining and unrolling and such) are pretty huge and many of them are well beyond what we can do using Alive right now.

@zwarich @kristerw the answer (as far as we know) is cutpoints, which break up the TV problem into separate parts. so far Alive eschews these because they require significant knowledge about what optimizations are actually doing, and we want to treat LLVM as a black box. but we're now seeing the limits of the black box approach, so we'll need to do cutpoints. nobody is working on this but Nuno has some plans.