[llvm] [RISCV] Reduce the VL of both operands in VMERGE_VVM (PR #144759)

[Min-Yih Hsu via llvm-commits]

mshockwave wrote:

> Why doesn't the VL optimizer handle this?

Because VMERGE_VVM would be eliminated by another optimization in VectorPeephole, so by the time we reached VL optimizer, we couldn't correctly perform this VL reduction (because we're not sure if we actually need VLMAX number of splat element or not at that point)

https://github.com/llvm/llvm-project/pull/144759