[llvm] [AMDGPU][InstCombine] Optimize constant shuffle patterns (PR #192246)

[Jay Foad via llvm-commits]

jayfoad wrote:

> This avoids the expensive ds_bpermute fallback, which requires address computation and LDS memory access,

ds_bpermute doesn't do any memory access. It just uses the DS instruction encoding format.


https://github.com/llvm/llvm-project/pull/192246