[llvm] [LV] Implement integer part of VPDerivedIV cost model (PR #198252)

[David Sherwood via llvm-commits]

david-arm wrote:

Hi @ElvinaYakubova, thanks! What the saxpy test has shown is that the cost of interleaved memory ops with factor 5 is too high in general. I've been testing out some micro-benchmarks based off saxpy and other interleave-factor-5 loops. In general it looks like we should be vectorising even when the vector combine pass doesn't optimise the IR. I am testing a downstream patch. If I can't land a fix for this soon, I'll probably end up reverting the PR as I can't see any other obvious solution.

https://github.com/llvm/llvm-project/pull/198252