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.
This commit moves the `f32::cmath` and `f64::cmath` modules into the
`sys` module. Note that these are not publicly exported modules, simply
implementation details. These modules are already platform-specific with
shims on MSVC and this is mostly just a reflection of that reality. This
should also help cut down on `#[cfg]` traffic if platforms are brought on
which don't directly support these functions.
IEEE 754-1985 specifies the encoding of NaN floating point numbers,
but while it mentions that NaNs can be subdivided into signaling
and quiet ones, it doesn't fix the encoding of signaling NaNs in binary
formats. This led to different implementations (CPUs) having different
encodings. IEEE 754-2008 finally specified the encoding of signaling NaNs
but some architectures are compatible with it, while others aren't.
Certain MIPS and PA-RISC CPUs have different encodings for signaling
NaNs.
In order to have the float <-> binary cast feature of the std library be
portable to them, we don't mask any quiet NaNs like we did before (only
being compliant to IEEE 754-2008 and nothing else), but instead we
simply pass a known good NaN instead.
Note that in the code removed there was a bug; the 64 bit mask for quiet
NaNs should have been `0x0008000000000000` instead of the specified
`0x0001000000000000`.
Fix NaN handling in is_sign_negative/positive
This would be my proposed fix for the #42425 provided we decide it is indeed a problem.
Note this would technically be a breaking change to a stable API. We might want to consider deprecating these methods and adding new ones.
