[llvm] [SROA] Extend SROA to support dynamic indexing (PR #217188)

[Yonah Goldberg via llvm-commits]

https://github.com/YonahGoldberg commented:

Looks interesting and relatively promising, although the llvm-opt-benchmark results look like a mixed bag. There's a lot of reviewer burden here because it's a large change, you're a first-time contributor, and you presumably used AI to write a lot of this, so I don't have high confidence in the quality of the code.

To start, it'd be good to fix the correctness issue Yingwei found and the failing AMDGPU test. Then I'd appreciate if you could go further in depth in your description on the context behind this, i.e. what was your motivation here? How did you find this enhancement? Is there some code you're interested in speeding up? You should also explain in more detail the changes you are making and the design decisions you made here.

Next, you should analyze the llvm-opt-benchmark regressions. For most IR optimizations like this we often can land them by looking at IR improvements statically rather than empirical results. I see some cases where we generate worse code as a result of this. Why is that? Is there anything in your PR you can fix so that we don't have any regressions? If you can't avoid regressions, are the improvements good enough to warrant landing this? It would be good to have an in-depth analysis of the results.

After all of this I'd be happy to read through more closely and I think other people probably would too :)

https://github.com/llvm/llvm-project/pull/217188