[llvm] [DAGCombiner] Avoid illegal scalar nodes in estimate refinement (PR #219394)

[QIU Chaofan via llvm-commits]

ecnelises wrote:

Is X86 the only target having this problem? If so, I'd like to understand if this could be handled in X86-specific lowering code.

BTW, is there any reason why `f32` is illegal on i686+SSE1? Would changing it be a viable alternative solution?

https://github.com/llvm/llvm-project/pull/219394