llvm-project

[AMDGPU] Use subtarget feature for flat offset bit width instead of arch checks

#183742

Merged