[llvm] [VPlan] Use VPlan-based costs for all integer IVs. (PR #217766)

[Mel Chen via llvm-commits]

https://github.com/Mel-Chen commented:

This looks like it could fix some of the regressions we ran into while retiring our downstream strided access work. :)
Because the strided access transformation happens after the legacy CM's uniform analysis, getLegacyCost() fails to account for the scalar IV cost, causing the induction variable cost to be severely overestimated.

https://github.com/llvm/llvm-project/pull/217766