[llvm-branch-commits] [llvm] AMDGPU: Fix buffer load/store of pointers (PR #95379)

[Krzysztof Drewniak via llvm-branch-commits]

krzysz00 wrote:

So, general question on this patch series:

Wouldn't it be more reasonable to, instead of having separate handling for all the possible register types, always do loads as `i8`, `i16`, `i32` `<2 x i32>`, `<3 x i32>, or `<4 x i32>` and then `bitcast`/`merge_values`/... the results back to their type?

Or at least to have that fallback path - if we don't know what a type is, load/store it as its bits?

(Then we wouldn't need to, for example, go back and add a `<16 x i8>` case if someone realizes they want that)

https://github.com/llvm/llvm-project/pull/95379