[llvm] Partition MUBUF load/store offsets into vector and scalar parts (PR #142328)

[Jay Foad via llvm-commits]

jayfoad wrote:

> Instruction selection seems to be taking different paths for the same buffer_load_dword instructions, example, source node could be load and @llvm.amdgcn.s.buffer.load. This would required adding patterns in multiple places, wouldn't making changes post Instruction Selection be simpler?

No, it's still better to do it during instruction selection. That's where all the addressing mode matching happens. If that causes lots of code duplication then we can look at ways of commoning up some of the selection patterns.

https://github.com/llvm/llvm-project/pull/142328