[llvm] [KnownFPClass] Add KnownFPClass deductions for `pow` (PR #215592)

[via llvm-commits]

ZERICO2005 wrote:

I guess part of it may be on what part of LLVM will be handling the special cases.

Should `KnownFPClass` always assume that it will handle inputs that are never `only-special-cases`?

For a comparison of extremes, we can look at something like the `acosh(x)` function, which returns non-NaN iff `x >= +1.0`. 

A useful deduction would rule out positive infinity iff `x` is not positive infinity. It is useful since it helps in cases where the input could be normal.

A useless deduction would rule out NaN iff `x` is always positive infinity. It is useless because it requires the input to be known never-normal, which is very unlikely.



https://github.com/llvm/llvm-project/pull/215592