[llvm] AMDGPU: Fix handling of -0 in round lowering (PR #65761)

[Jay Foad via llvm-commits]

jayfoad wrote:

> The GlobalISel output [...] fails to form BFI.

That's not visible in any of the codegen tests is it? We should have something that shows the final codegen for globalisel.

Code change looks good.

https://github.com/llvm/llvm-project/pull/65761