[llvm] Adding change to IndVarSimplify pass to optimize IVs stuck in trivial vector operations (PR #122248)

[via llvm-commits]

anilavakundu007 wrote:

> Thanks. I think in your particular case, I suspect it probably makes sense to scalarize the vector op independently of whether it contains an induction variable? This case creates a splat vector, does a vector add and then immediately extracts both elements. It seems unlikely that this is going to be faster than just doing two adds on any architecture.

The reason this primarily focused on IVs as once the vector operation is broken LLVM can basically strip everything down to couple of instructions (following loop unroll). 

https://github.com/llvm/llvm-project/pull/122248