Fixed floating point issue with asinh function
This should fixes#63271 , in which `asinh(-0.0)` returns `0.0` instead of `-0.0`.
according to @nagisa
>
>
> IEEE-754 (2008), section 9.2.1:
>
> > For the functions expm1, exp2m1, exp10m1, logp1, log2p1, log10p1, sin, tan, sinPi, atanPi, asin, atan, sinh, tanh, asinh, and atanh, f(+0) is +0 and f(−0) is −0 with no exception.
>
> and
>
> > sinh(±∞) and asinh(±∞) are ±∞ with no exception.
After ensuring that the function `asinh` is the only function affected (functions like `sin`, `sinh` are all based on `cmath` library or `llvm` intrinsics), and that `atanh` always gives the correct result. The only function to modify is `asinh`.
This patch adds a `copysign` function to the float primitive types.
It is an exceptionally useful function for writing efficient numeric
code, as it often avoids branches, is auto-vectorizable, and there
are efficient intrinsics for most platforms.
I think this might work as-is, as the relevant `copysign` intrinsic
is already used internally for the implementation of `signum`. It's
possible that an implementation might be needed in japaric/libm for
portability across all platforms, in which case I'll do that also.
Part of the work towards #55107
… previously in the unstable core::num::Float trait.
Per https://github.com/rust-lang/rust/issues/32110#issuecomment-379503183,
the `abs`, `signum`, and `powi` methods are *not* included for now
since they rely on LLVM intrinsics and we haven’t determined yet whether
those instrinsics lower to calls to libm functions on any platform.
The current implementation/documentation was made to avoid sNaN because of
potential safety issues implied by old/bad LLVM documentation. These issues
aren't real, so we can just make the implementation transmute (as permitted
by the existing documentation of this method).
Also the documentation didn't actually match the behaviour: it said we may
change sNaNs, but in fact we canonicalized *all* NaNs.
Also an example in the documentation was wrong: it said we *always* change
sNaNs, when the documentation was explicitly written to indicate it was
implementation-defined.
This makes to_bits and from_bits perfectly roundtrip cross-platform, except
for one caveat: although the 2008 edition of IEEE-754 specifies how to
interpet the signaling bit, earlier editions didn't. This lead to some platforms
picking the opposite interpretation, so all signaling NaNs on x86/ARM are quiet
on MIPS, and vice-versa.
NaN-boxing is a fairly important optimization, while we don't even guarantee
that float operations properly preserve signalingness. As such, this seems like
the more natural strategy to take (as opposed to trying to mangle the signaling
bit on a per-platform basis).
This implementation is also, of course, faster.
