[llvm] [RISCV] Optimize divide by constant for VP intrinsics (PR #125991)

[Craig Topper via llvm-commits]

topperc wrote:

> > RISC-V seems to be the only active development of VP intrinsics in tree. We the RISC-V community, should agree on what's important to us and prioritize where we spend effort.
> 
> Agreed. If the only use case for non-trapping VP intrinsics was EVL tail folding in the loop vectorizer then adding optimisations for them is less of a priority. We should probably talk about this in the RISC-V LLVM sync up on Thursday.
> 
> It's worth noting though that the loop vectorizer still emits load/store VP intrinsics, and they might be missing combines. Is this something you've looked at already in your downstream?
> 
> > Is this question about the max case or the entire patch?
> 
> The entire patch, I just chose a random case to check if instcombine did any strength reduction

InstCombine for sure does not do the BuildVPUDIV/BuildVPSDIV part of this code that does the main optimization for arbritrary constants.

https://github.com/llvm/llvm-project/pull/125991