[llvm] [NVPTX] Vectorize and lower 256-bit global loads/stores for sm_100+/ptx88+ (PR #139292)

[Drew Kersnar via llvm-commits]

dakersnar wrote:

> Unfortunately, we have lots of brittle code which uses hard-coded operand indices for various machine instructions. Here is one example I personally am guilty of introducing:

Have you thought about how hard it would be to fix this problem across the codebase? I'd agree that the hard-coded operand indices seem to be one of the more dangerous patterns I've noticed in our code, but I also wouldn't be surprised if it would be too much churn to fix.

I'm curious, does anyone know if other LLVM backends (especially ones that use the selectiondag framework fully as intended) manage to avoid this issue, or is it a flaw baked into the framework?

https://github.com/llvm/llvm-project/pull/139292