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

[David Green via llvm-commits]

davemgreen wrote:

> > Does this fail for you under aarch64? I didn't look into what might be going on - it was strange that it was trying to do any vectorization with fp128 long doubles. https://godbolt.org/z/hP4eYooTK
> 
> Hi @davemgreen I am not sure I understood correctly. Did you mean that the cost model should return InvalidCost or the loop vectorizer should not try to vectorize loop entirely when the type is fp128?

I just meant that I don't think there is any point in vectorizing fp128 - there is no native support so it will all become libcalls anyway and there isn't much point in "vectorizing" (or interleaving with a vf==1). The cost should maybe be higher than they are but not infinite. It is hopefully not too important though if the error is coming from the fmuladd on a double, and this code just happens to trigger the mismatch in costs.

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