[llvm] [NVPTX] Use v2.u64 to load/store 128-bit values (PR #136638)

[Alex MacLean via llvm-commits]

AlexMaclean wrote:

> PTX apparently supports 128-bit ld/st/mov on sm_70+ and it looks like we've already landed bits and pieces of support for it last fall, while I wasn't looking. We'll probably need to plumb it through to support 128-bit atomics on sm_90, and it would simplify things here, too (though we'll still need these changes for older GPUs).
> 
> @AlexMaclean is that something already on your radar? I was about to start poking at the atomics (#122760), but will be happy if it's about to be done by NVIDIA. :-)

Yea, I've been thinking about improving our 128-bit register support for a while now. I'm hoping to get some time in the next week or two to take a deeper look and clean the situation up a bit. 

Sorry to have ignored https://github.com/llvm/llvm-project/issues/122760! I'll look into that too since I'm already trying to do some cleanup around atomics as well.

https://github.com/llvm/llvm-project/pull/136638