[llvm] Allow MIN/MAX signedness flip when operands are known-negative (PR #174469)

[Simon Pilgrim via llvm-commits]

https://github.com/RKSimon requested changes to this pull request.

Please can you add similar <8 x i16>test coverage for umin/umax -> smin/smax to combine-umin.ll and combine-umax.ll? SSE2 supports uminw/umaxw but not sminw/smaxw.

https://github.com/llvm/llvm-project/pull/174469