[llvm] [CodeGen] Expansion of scalable vector reductions (PR #129214)

[David Green via llvm-commits]

davemgreen wrote:

We should be expanding as many reductions as possible in ISel.

So far for scalable vectors there are certain operations that are not supported and return an invalid cost. frem is one of them, as is vecreduce.fmul but there are other types too.  The reasoning is that expanding via a loop is so expensive that you might as well have generated scalar / something else anyway.  If we are going to change that then it would be worth having a single point where that happens for any operation. It could do things like generate a single loop for multiple different operations, as opposed to having to generate multiple loops in isolation.

https://github.com/llvm/llvm-project/pull/129214