[llvm-branch-commits] [RISCV] Combine vwaddu_wv+vabd(u) to vwabda(u) (PR #184603)

[Luke Lau via llvm-branch-commits]

lukel97 wrote:

> @wangpc-pp I assume this will eventually lead to generating vwaddu instruction in x264? Can you please give your high-level overview of your plan? Have you considered doing this at IR level with intrinsics?

RVV intrinsics only support scalable vectors, and if it's coming from x264 it will be generated by SLP which will need fixed length vectors I think. Did you have a different approach downstream?

https://github.com/llvm/llvm-project/pull/184603