[llvm] [RISCV] Fold vmv.s.x into load from stack (PR #109774)

[Philip Reames via llvm-commits]

preames wrote:

A couple of observations on other cases we can fold.  Dropping these here not because they related to the review, but merely cause I need some place to find them in the future.
* vfmv.f.s - basically same logic as this, but for FP
* vmv.v.x/vfmv.v.f as strided load - specifically, only useful in *scalar* register allocation, and should be conditional on tuning flag.  Effectively combines with existing remat (for vector register allocation) of same.

https://github.com/llvm/llvm-project/pull/109774