[llvm] [NVPTX] Add IR pass for FMA transformation in the llc pipeline (PR #154735)

[Alex MacLean via llvm-commits]

https://github.com/AlexMaclean commented:

Looks reasonable to me. 

While I'm not convinced this is needed for <2 x float> vectorization, I think it can be simpler to just fold to an intrinsic in the IR as opposed to trying to keep a multi-instruction idiom recognizable and intact through other transformations. I know we've moved this direction with integer min/max and I think it makes sense here as well. 

https://github.com/llvm/llvm-project/pull/154735