[llvm] [RISCV] Lower prefix slidedown shuffles using VL + tail-undisturbed (PR #210927)

[Philip Reames via llvm-commits]

preames wrote:

> > Do you have any evidence the new lowering is profitable? I agree you can use a VL toggle instead of a mask here, but should you? That's not clear to me.
> 
> When building the llvm-test-suite with rva23u64 --O3 -ffast-math, I get some optimizations, especially for MultiSource/Benchmarks/tramp3d-v4.

Is this a static statement - i.e. you see the code gen changing?  It sounds like it is.  I'm not asking about whether this actually triggers.  I'm asking if you have evidence that when it triggers the new code is actually faster.  I suspect it is worse on the whole.  

https://github.com/llvm/llvm-project/pull/210927