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

[Artem Belevich via llvm-commits]

Artem-B wrote:

Here's my PR for making v4i8 legal: https://github.com/llvm/llvm-project/pull/67866

It does improve some cases where we were previously stuck with ld/st.v4.i8 (see unfold-masked-merge-vector-variablemask.ll in the PR above), but, AFAICT, does not help with your test case `combine_v16i8`. 


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