[llvm] [NVPTX] Use PRMT more widely, and improve folding around this instruction (PR #148261)

[Alex MacLean via llvm-commits]

================
@@ -1699,25 +1692,33 @@ def cond_not_signed : PatLeaf<(cond), [{
 // comparisons of i8 extracted with BFE as i32
 // It's faster to do comparison directly on i32 extracted by BFE,
 // instead of the long conversion and sign extending.
----------------
AlexMaclean wrote:

Fixed

https://github.com/llvm/llvm-project/pull/148261