[llvm] [DAGCombiner] Extend FP-to-Int cast without requiring nsz (PR #161093)

[Yi-Chi Lee via llvm-commits]

yichi170 wrote:

In the newest commit, I only handle the unsigned cases as @arsenm suggested, and I also added the test for the AMDGPU backend. It seems that it doesn't generate the expected codegen when the fp type is `double`. I'm not sure why.

I also adapted @paperchalice's suggestion to check the `nneg` flag.

https://github.com/llvm/llvm-project/pull/161093