In the former form, the negation of NAN is hidden inside the
optimization blocking macro/wrapper. After
9be66e9d22, F(-NAN) works just as
well as -F(NAN).
This avoids test issues with recent optimizations in LLVM.
https://github.com/llvm/llvm-project/pull/193586 seems to make
the compiler consider known symmetry for hyperbolic math functions,
exploiting the fact that cosh(-x) is equal to cosh(x) (for values
other than NANs). By hiding the negation inside the F() wrapper,
the compiler can't reason about is existence.