[llvm] [VectorCombine] Fold binary op of reductions. (PR #121567)

[David Green via llvm-commits]

davemgreen wrote:

I added some quick cost-models for AArch64 extending reductions. I agree this isn't the most reliable thing, and having a more representative intrinsic might help. IMO this is better left for the backend, but not all architectures lower reduction nodes using sdag at the moment.

https://github.com/llvm/llvm-project/pull/121567