[llvm] [NVPTX] fold movs into loads and stores (PR #144581)

[Princeton Ferro via llvm-commits]

Prince781 wrote:

> One thing I wonder about is how this changes the PTX semantics of a program.  If we change from loading something as a b32 to a v2.b16 will this impact the memory consistency guarantees in PTX?

I'm not aware of anything in the spec that leads to different semantics. I would think since the underlying data being accessed and the alignment requirements are the same in both cases, they are essentially the "same operation" using different amounts of registers.

https://github.com/llvm/llvm-project/pull/144581