[llvm] [NVPTX] Prefer prmt.b32 over bfi.b32 (PR #110766)

[Artem Belevich via llvm-commits]

Artem-B wrote:

@justinfargnoli -- any chance NVIDIA could publish more details on the instruction timings. It's kind of hard to tell how fast something is expected to be by looking at the mnemonics alone.


https://github.com/llvm/llvm-project/pull/110766