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

[David Sherwood via llvm-commits]

david-arm wrote:

It looks like if we force vectorisation with VF=2 and disable vector combine, the saxpy_5 loop is still slightly faster than scalar. I think what really needs to happen is for the loop vectoriser to apply the same vplan transformation that takes place in VectorCombine.cpp. However, that might take a while to fix and won't solve the regression in the short term. So for now I've got a couple of options:

1. Sacrifice generic improvements to the vplan cost model and revert this PR, but potentially making other loops worse, or
2. Find a way to reduce the cost for VF=2. It does seem like a cost of 28 for the interleave groups are a bit high.

@fhahn any thoughts?

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