[llvm] [NVPTX] add combiner rule for final packed op in reduction (PR #143943)

[Alex MacLean via llvm-commits]

AlexMaclean wrote:

I too wonder if this can be made more general. Is there a reason this optimization needs to be NVPTX specific? Maybe we could update the lowering of these operations based on if vector shuffle is legal. Or maybe we could add a DAG combiner rule that looks for BUILD_VECTOR <- [BIN_OP] <- extract_vector_elt and coverts to extract_vector_elt <- [BIN_OP].

https://github.com/llvm/llvm-project/pull/143943