[llvm] [VPlan] Impl VPlan-based pattern match for ExtendedRed and MulAccRed (NFCI) (PR #113903)

[David Green via llvm-commits]

davemgreen wrote:

> If the legacy-cost estimates are worse, I don't think we need to try hard to match them, if we can make use of more accurate modeling in VPlan. If that's the case, I'd use `planContainsAdditionalSimplifications` to allow divergences between legacy and VPlan-based cost model for plans that contain extended reduction recipes.

Yeah I think that makes sense in this regard, if we can avoid the assert. The old cost model found it difficult to deal with things that the new method can more easily handle, and I've tried to verify the performance looks OK with the new costmodel.

https://github.com/llvm/llvm-project/pull/113903