[llvm] [NVPTX] Preserve v16i8 vector loads when legalizing (PR #67322)

[Drew Kersnar via llvm-commits]

dakersnar wrote:

Sounds good, I opened an issue here: https://github.com/llvm/llvm-project/issues/118851, we can continue the discussion there 👍.

One note I'll comment on here:

> Conversion often kill any benefit of vectorizing the wide load/store, so it may not be worth it in general case, where the rest of the code would use loaded data in element-wise fashion.

My understanding is that chunks of `load/store v4i8` gets lowered to `u32` currently, so the conversion/extraction is already happening regardless, but currently v8i8 gets lowered to two u32s instead of a v2.u32, which seems like a clear loss. See the repro in my issue.


https://github.com/llvm/llvm-project/pull/67322