[llvm] [LoongArch] Optimize extracting f32/f64 from 256-bit vector by using XVPICKVE. (PR #151914)

[via llvm-commits]

tangaac wrote:

> Better choice. Lower instruction latency than before.
> 
> By the way, maybe `extractelement+insertelement` sequence of `v8i32/v4i64` types can also use `xvpickve` and `xvinsve0` instructions?

Yes, but we don't need to an extra pattern for `extractelement+insertelement` now.
As you can see in the tests.

https://github.com/llvm/llvm-project/pull/151914