[llvm] [InstCombine] Do not fold gep operand into phi unless not poison (PR #115901)

[Antonio Frighetto via llvm-commits]

antoniofrighetto wrote:

> I don't think this is the right fix. As far as I can tell this is just a bug in the flag intersection, because we end up ignoring the flags on the first instruction. Would doing this fix the issue?

Updated, thank you! Admittedly, I thought about intersecting flags here, but felt like dropping flags (losing inbounds in the gep) at this point might have been unfavourable for later optimizations (so thought of preventing the folding altogether was "sort of" more robust).

https://github.com/llvm/llvm-project/pull/115901