[llvm] [NVPTX] Fold vector FMAs in the IR peephole pass (PR #224018)

[Sjoerd Meijer via llvm-commits]

sjoerdmeijer wrote:

Sorry,  high level question first, but looking at the pattern, I was expecting DAGCombine to handle this case, and I think it does:

https://godbolt.org/z/4Ej1a5zM6

So, without this patch, we are already generating a `FMA_F32x2`.

I understand that with this patch, we generate a FMA earlier in the flow and that this may have some benefit, but that's why I wanted to ask about the motivation and if you do see benefits doing this?

https://github.com/llvm/llvm-project/pull/224018