[llvm] [RISCV] Ensure the valid vtype during copyPhysReg (PR #118252)

[Piyou Chen via llvm-commits]

BeMg wrote:

> I think we still need to take into account whole vector register moves that might be emitted after vsetvli insertion, i.e. #118283 might not be enough.
> 
> After #118283 do we know how frequent these are or where they might come from?

I have a branch https://github.com/BeMg/llvm-project/commit/eca5123051e7e2994af6bffd494d2e5069abdcc2
 for creating a standalone pass (borrowing part of insert vsetvl pass for VSETVLInfo analysis) to handle the COPY added after the insert vsetvl pass, but it hasn't found any examples that need to emit new vsetvl so far.

My idea is to wait until we actually encounter a RVVReg COPY pseudo being added after the insert vsetvl pass, and then discuss how to resolve that case.

cc @kito-cheng @topperc 




https://github.com/llvm/llvm-project/pull/118252