[llvm] [AMDGPU] Prefer vector i8s in PHI Nodes (PR #91016)

[Jeffrey Byrnes via llvm-commits]

jrbyrnes wrote:

Rebase + Removed getPHIScalarizationOverhead and vectorizing i8 shuffles in favor of https://github.com/llvm/llvm-project/pull/91016

This PR now simply tells SLP to analyze i8s by changing the vectorization factor. The cost model seems okay as is, with perhaps some overestimation of vector cost for some ops. The vectorization in the lit tests is due to reduced cost of vectorizing vs extracting/inserting for identity vectors.



https://github.com/llvm/llvm-project/pull/91016