[llvm] [NVPTX] Don't use stack memory when bitcasting to/from v2i8 (PR #113928)

[Justin Fargnoli via llvm-commits]

https://github.com/justinfargnoli commented:

Overall, thanks for this PR. The current lowering is definitely not what we want! 

---

> we hit the default legalization rules which perform the bitcast in stack memory

Do you know where this default legalization rule is implemented? 

---

Note: [`mov` vector-to-scalar (pack) or scalar-to-vector (unpack)](https://docs.nvidia.com/cuda/parallel-thread-execution/index.html#data-movement-and-conversion-instructions-mov-2) doesn't support `.b8`. 



https://github.com/llvm/llvm-project/pull/113928