[llvm] [ARM][SLP] Fix cost function for SLP Vectorization of ZExt/SExt (PR #122713)

[David Green via llvm-commits]

https://github.com/davemgreen commented:

The MVE instruction that getExtendedReductionCost is trying to model is vaddlv, which is a `i64 vecreduce(v4i64 sext(v4i32))`. So it would seem odd to add the cost of the extend again to it.  BaseT::getExtendedReductionCost includes the cast cost too, so it would look like this is double-counting.

This is an example of MVE using the vaddlv instruction, which looks decent in this small example: https://llvm.godbolt.org/z/oTrMTcM4E

https://github.com/llvm/llvm-project/pull/122713