[llvm] [RFC][BPF] Report Unreachable Behavior from IR (PR #126858)

[via llvm-commits]

yonghong-song wrote:

> On a conventional target, in general, code with undefined behavior is fine, as long as it doesn't actually execute at runtime. People write source code based on this, and transforms duplicate code based on this. And this makes diagnostics in the backend impractical.
> 
> I understand that BPF operates under different constraints than conventional targets... in particular, BPF-C is a subset of C. It only allows programs that pass the BPF verifier. So maybe you don't run into exactly the same issues with late diagnostics. But, I'm not sure what the implications are here for transforms. For example, can JumpThreading introduce an "unreachable" that triggers the error here? If so, do we need to restrict JumpThreading on BPF?

JumpThreading works for BPF target as well and the middle end optimization does not need to restrict JumpThreading for BPF target. The bpf backend pass implemented in this patch will handle such code patterns (BPF backend will do some analysis if necessary). If in the future, middle end optimization has new cases which introduce unreachable, BPF backend will try to handle handle such new patterns if necessary and this should be fine as they typically will be in the same release.

So in summary, middle end optimization does not need to do any special things for BPF backend.



https://github.com/llvm/llvm-project/pull/126858