[llvm] [InstCombine] Teach foldSelectOpOp about samesign (PR #122723)

[Alexander Kornienko via llvm-commits]

alexfh wrote:

This diff is a result of the whole -O3 optimization pipeline. It could be that the change here may affect further passes. We couldn't use `-opt-bisec-limit` to get to a specific optimization pass, since there's a large range of -opt-bisect-limit values where Clang just crashes for the input file in question.

https://github.com/llvm/llvm-project/pull/122723