long double functions, that are identical to double functions, usually
define themselves as weak references to their double counterpart.
However, the piece of assembler written for that produces
.equ <name>l, <name>
but on SuperH C-exported symbols have leading underscores. This commit
adds an SH3 exception to use instead
.equ _<name>l, _<name>
which fixes a number of long double functions.
My previous Apple Silicon build went through, so I thought it already
worked, but it turns out it accidentally built an armv7 build instead.
This actually fixes the Apple Silicon build. One thing to note in
particular is that on Apple Silicong `long double` is the same as
`double` while on Linux `long double` is a 128 bit double-double
format.
Co-authored-by: Elliot Saba <staticfloat@gmail.com>
__strong_alias is an ELF feature that is not available on Darwin/MachO.
We have openlibm_strong_reference to paper over these differences,
but we weren't using it consistently. In particular, we were still
using it to define long double -> double aliases on platforms where
these are the same type (e.g. aarch64). This didn't used to matter,
because the only such platform we supported was always Linux and
thus ELF. This patch fixes these instances thus fixing the build
on Apple Silicon.
On x86_64, openlibm_fenv.h is silently coupled to openlibm_complex.h due to the inclusion of math_private.h inside of openlibm_fenv_amd64.h. The math_private.h header was included to gain access to the OLM_DLLEXPORT macro. However, this macro is redefined, since it is originally defined in openlibm_math.h. This inclusion also exposes math_private.h to public clients, which is undesirable.
Instead, a standalone header is created to contain the OLM_DLLEXPORT definition. This header is used included in the appropriate files. openlibm_fenv_amd64.h is no longer coupled to math_private.h.
Changes needed for mips suport have been added.
Code for openlibm_fenv_mips.h, mips_fpmath.h and fenv.c
was taken from https://github.com/freebsd/freebsd
As a component of atan2(y, x), the case of x == 1.0 is farmed out to
atan(y). The current implementation of this comparison is vulnerable
to signed integer underflow (that is, undefined behavior), and it's
performed in a somewhat more complicated way than it need be. Change
it to not be quite so cute, rather directly comparing the high/low
bits of x to the specific IEEE-754 bit pattern that encodes 1.0.
Note that while there are three different e_atan* files in the
relevant directory, only this one needs fixing. e_atan2f.c already
compares against the full bit pattern encoding 1.0f, while
e_atan2l.cuses bitwise-ands/ors/nots and so doesn't require a change.
Incorporated from FreeBSD source tree. See:
b21ccf63f2
In particular, OpenBSD already defines __GNUC_PREREQ__, __BEGIN_DECLS,
and __warn_references. Hiding the definitions behind #ifndef guards
avoids warnings when building.
* Previously, behavior differed if the same value of `ARCH` was defined
within `Make.inc` or defined on the command line. Don't do that.
* Provide saner defaults for `ARCH` and `MARCH`, and more importantly,
allow for the proper overriding of both.
* Split `AArch64` code further away from the other `arm` code.
- Align DLLEXPORT in definitions and declations. There is still a few
cases left, where the declation in the compiler's complex.h disagrees
with the implementation here. For now we can't do anything about that,
but maybe should be revisited in the future.
- Fix the syntax on an .ascii directive that gcc accepted mistakingly, but
clang does not.
_LITTLE_ENDIAN and _BIG_ENDIAN are built-in on some platforms/versions.
Better use __ORDER_LITTLE_ENDIAN__, __ORDER_BIG_ENDIAN__ and __BYTE_ORDER__,
which are standard for gcc and clang, and define them when they are missing.
Also remove the special-case for FreeBSD, which is apprently not needed.
Clang doesn't like it if you bitshift negative numbers, as the behaviour
of it is undefined. Solve this by first shifting, followed by negating.
This was fixed similarly in FreeBSD.
The current tradition of openlibm is to hide all of its internal symbols
into the reserved system namespace. CloudABI has a check in place to
ensure that its C library (which contains openlibm) to not place any
unwanted symbols into the public namespace. openlibm seems to leak
_scan_nan() in there, so we'd better add an additional underscore.