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

[Yingwei Zheng via llvm-commits]

dtcxzyw wrote:

> This will likely cause performance issues unless they are mitigated.

Can you provide a test to demonstrate the performance issue? As I said before, vertical vector ops are always cheaper than horizontal ones. If not, we can move it into VectorCombine and use InstructionCost to see whether the transformation is a win.


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