[llvm] [RISCV] Avoid VMNOT by swapping VMERGE operands (PR #126751)

[Piotr Fusik via llvm-commits]

pfusik wrote:

In llvm-codegen-benchmark I noticed hundreds of patterns such as https://github.com/dtcxzyw/llvm-codegen-benchmark/blob/12a783bf6af60c93d100354b71dc6c0c4dea4eb7/result/rvv/0002ccdf5e758f9a.S
This crude proof-of-concept turns it into:

        vsetivli        zero, 8, e32, m2, ta, ma
        vmv.v.i v10, 1
        vmerge.vim      v10, v10, 0, v0
        vand.vi v8, v8, 3
        vor.vv  v8, v8, v10
        vmseq.vi        v0, v8, 0
        ret

Questions:
1. Do we want this optimization?
2. Do we want it at this stage or maybe there's a better place?

Further work:
I think in many scenarios we could emit a mask-undisturbed operation (here: `vor.vi`) instead of `vmv`+`vmerge`+vector op.


https://github.com/llvm/llvm-project/pull/126751