And add tests in this region.

Fix functions in region between previous
overflow bound and maximum float.

Accuracy is maintained by relying on high
accuracy log (in branch), however performance
drops slightly. Ideally the core approximation
should also work above threshold.

Issue #600 highlighted a limit of current
implementation where early overflow is noticed,
which diverges from the C99 standard.

This patch adds test to catch this type of cases.

It only checks a few points therefore we don't
know for sure if overflow occurs at the right time
but we know if something is wrong in the overflow
region.

Fix both scalar and simd implementation (double precision)

Fixes #523.

To test for the RISC-V vector extension (RVV), we currently run two
small programs -- one for each value of LMUL=1,2 -- that use some
vector intrinsics from the v0.11.x draft of the intrinsics spec. The
library itself however needs newer intrinsics than the ones that we
test for. This can lead to a problem because GCC 13 supports the
intrinsics that we test for, but not (all of) the ones we intend to
use. In short, the build can fail with GCC 13:

  https://github.com/shibatch/sleef/issues/579

This commit changes the detection mechanism. Instead of compiling a
program that makes use of the intrinsics, we now just check the two
values, __riscv_v and __riscv_v_intrinsic, to ensure that we have
version 1.0 of the vector extension and version 0.12 of the
intrinsics. These should be the stable versions of each.

Note: the "m1" and "m2" checks are now identical because they are both
covered by the official spec.

Fixes #579

Minor fixup to previous change

Just checking that it is working.

Ubuntu-24.04 GH-runner still has a buggy version of gcc 12.
Therefore we jump to gcc 13.

And complete other toolchain file.

This will ultimately unlock the use of latest versions of
gcc 12, then gcc 13, and finally gcc 14.
Gcc 14 will unlock more testing.
This PR does not update version of gcc yet, just to make sure
old version still pass tests.
Remove some expensive tests.

Error: va_list 'ap2' was opened but not closed by va_end().

Error: memleakOnRealloc
Fix: Free is realloc fails (and at end of main).

Error: integer overflow
Fix: Use appropriate intermediate type. UL should be big enough
     on all platforms.

Error: uninitialized variables
Fix: Only use mpfr_zinit when mpfr_t already initialized
Note: The use of mpfr_zinit is inconsistent anyway.
      Gencoef needs a big cleanup and possibly porting to CMake

Deleted old benchmark system and replaced with most up to date one
which uses googlebench.
Changes to CMakeLists were carried accordingly.
Deleted section in documentation on legacy benchmark system.

Used to be scalar, now is sve.

When benchmarks are built, SLEEF_ENABLE_CXX should be
set to ON, since they use cpp.

Calls to functions which return two values (for example sincos),
were being optimized away in new benchmark tool. Fixed by
decreasing optimization level (O3 -> O2).

Add paragraph on multiple output format feature.

This patch fixes the following two problems:

* On aarch32, cbrt(4) was returned instead of M_PI/2 when +-INFINITY is given to atanf.
* Due to a bug in gcc-13, sign of tanf(+-0) was wrong on PUREC_SCALAR and VXE.

This patch also adds testing for the first problem.
The second problem is detected by the existing tester.

---------

Co-authored-by: Naoki Shibata <shibatch.sf.net@gmail.com>

Added new benchmark tool to SLEEF project using googlebench
framework.
In theory this tool can benchmark any unary and binary function
in SLEEF.
Benchmark is enabled in all functions listed in benchsleef.cpp.
This list uses macros in benchmark_callers.cpp, so that we can
enable benchmarking in multiple precisions using single lines
of code. It is also possible to list the exact function we want,
as each of the macros in benchmark_callers.cpp can be called
independently and connected like building blocks.
The tool is integrated with SLEEF via CMake, meaning it can
be built automatically when SLEEF is built. To enable that,
pass CMake argument -DSLEEF_BUILD_BENCH=ON.

This tool depends on C++17 standard.
Tested on aarch64 for scalar, vector and SVE routines
Tested on x86 for different vector length extensions.
Tested for llvm-17, gcc-11 and gcc-14.

It adds explicit typecasting so that C++ compilers can compile the inline headers generated from these helpers.

Co-authored-by: Naoki Shibata <shibatch.sf.net@gmail.com>

Propagate change from #489 to quad-tester.

Manually rebased.

Co-authored-by: Naoki Shibata <shibatch.sf.net@gmail.com>

same

Bump version to 3.6.1, as this version has a compatibility fix.

Add a new workflow to detect unexpected backward-incompatible
changes at precommit, by comparing current version against a
reference version, both generated on Linux with gcc.
It uses abidiff to compare ELF-format shared libraries.
This is the tool used by Fedora to report a recent breach in
API #534. Test should fail until it is fixed by #545.

Documentation for these runners here https://gitlab.arm.com/tooling/gha-runner-docs
Note that only SLEEF upstream repo is allowed to use these runners.

Integrate runners via strategy.matrix.os, in order to facilitate customisation
of step (e.g. config) based on runner label.

Install OpenSSL since tester3 is enforced.
New runners do not seem to come with openssl or cmake pre-installed.

Use SVE enabled instances (c7g)
Use 2xlarge instances so tests run slightly faster without compromising on availability.

Remove aarch64 from cross build and tests (on Linux).
And update documentation/README.

Define a custom style classes for dates and posts.

Bump version in CMakeLists.

Only Linux supports the AArch64 vector ABI, so it doesn't make sense
to expose these symbols anywhere else for AArch64. Without any
compelling reason to provide them on x86_64, it is better to just
revert this patch and deal with this more carefully after the release.

This reverts commit 0e47fcc3.

Restructured some of the sentences and added commands to
install dependencies for sleef's internal microbenchmark tool.

The acosh function is only defined for values greater or
equal to 1. Documentation was updated according to this.
Also, rephrased a confusing sentence.

Store the result of the function call during its benchmark
to prevent it from being optimized away.
Also changed name of NITER1 and NITER2 macros to more obvious
acronymes.
Tested on aarch64 and x86 so far.