[llvm] [AMDGPU] Quit PromoteAllocaToVector if intrinsic is used (PR #68744)

[Pierre van Houtryve via llvm-commits]

https://github.com/Pierre-vh commented:

Can you rename the change to something like "Bail if assume-like intrinsic is used in PromoteAllocaToVector"? Current title sounds like you want to disallow intrinsics altogether :) 

About the change itself I have no strong opinion, I'm not familiar with those intrinsics so I added other reviewers. I'm wondering if we could maybe pre-calculate the value of those intrinsics and replace them? If the value is just `sizeof` we can replace it with a `ConstantInt` of the size of the VectorType for instance?

It all depends on how common these are. If they're very rare then we can just ignore them I think, and only implement better handling if we have data to prove we need to do better with them.

https://github.com/llvm/llvm-project/pull/68744