[llvm] [CodeGen] Emit a more efficient magic number multiplication for exact udivs (PR #87161)

[Matt Arsenault via llvm-commits]

arsenm wrote:

> > Probably need a scalable vector test for the SPLAT_VECTOR case?
> 
> I did that in the .mir example

Would be better to have it in the end-to-end IR test. I think we're overusing MIR tests, especially for optimizations like this. Too many things can go wrong in the pipeline as a whole such that the optimization isn't useful when testing MIR directly 

https://github.com/llvm/llvm-project/pull/87161