[llvm] [InstCombine] Handle isNanOrInf idioms (PR #80414)

[Andy Kaylor via llvm-commits]

andykaylor wrote:

> I will reverse the transform in SDAG to fix the problem. See also [#76338 (comment)](https://github.com/llvm/llvm-project/pull/76338#issuecomment-1929678129).

Thanks! This broke the build of some downstream libraries I work with because they didn't want to link with the runtime library. We disabled the transformations for fp128 locally to work around the problem, but I think this is a good transformation and I'd like to re-enable it when we can.

You may already be considering this, but x86_fp80 also has issues, in that after your changes it's using x87 instructions where it used to use a faster integer implementation: https://godbolt.org/z/bPfrnW368

https://github.com/llvm/llvm-project/pull/80414