Port Sleef to AIX

[KamathForAIX]

Closes #628.

[shibatch]

The cross-compilation test is not passing.
Could you take a look at it a little bit?

https://shibatch.dns-cloud.net/jenkins8086/job/sleef/job/PR-629/1/consoleFull

[219/248] Linking C executable bin/qiutdspx2
FAILED: bin/qiutdspx2 
: && /usr/bin/powerpc64le-linux-gnu-gcc-14 -Wall -Wno-unused-function -Wno-attributes -Wno-unused-result -Wno-psabi -ffp-contract=off -fno-math-errno -fno-trapping-math -fno-strict-aliasing -O3 -DNDEBUG  src/quad-tester/CMakeFiles/qiutdspx2.dir/qiutsimd.c.o src/quad-tester/CMakeFiles/qiutdspx2.dir/__/common/main_checkfeature.c.o src/quad-tester/CMakeFiles/qiutdspx2.dir/qtesterutil.c.o -o bin/qiutdspx2 -L/home/jenkins/workspace/sleef_PR-629/build/submodules/tlfloat/lib   -L/home/jenkins/workspace/sleef_PR-629/build/lib   -L/home/jenkins/workspace/sleef_PR-629/build/src/common -Wl,-rpath,/home/jenkins/workspace/sleef_PR-629/build/submodules/tlfloat/lib:/home/jenkins/workspace/sleef_PR-629/build/lib:/home/jenkins/workspace/sleef_PR-629/build/src/common  lib/libsleefquad.a  lib/libsleef.a  /usr/powerpc64le-linux-gnu/lib/libm.so  /usr/powerpc64le-linux-gnu/lib/librt.a && :
/usr/lib/gcc-cross/powerpc64le-linux-gnu/14/../../../../powerpc64le-linux-gnu/bin/ld: lib/libsleefquad.a(qdispx2.c.o):(.toc+0x0): undefined reference to `Sleef_fmaq2_u05vsx'
/usr/lib/gcc-cross/powerpc64le-linux-gnu/14/../../../../powerpc64le-linux-gnu/bin/ld: lib/libsleefquad.a(qdispx2.c.o):(.toc+0x8): undefined reference to `Sleef_ilogbq2_vsx'

[KamathForAIX]

@shibatch I am not sure why this happened because when I tried in PPC Linux with the same commands as in the Jenkins build, it worked. I have made a guess looking at the library name that it might have to do something with the change in src/common/quaddef.h.

Can we check again with the recent change I made?

[shibatch]

The culprit seems to be this?

-- Performing Test COMPILER_SUPPORTS_VSX - Failed

Is VSX support removed in gcc-14?

[shibatch]

You don't need this #endif here.

[KamathForAIX]

@shibatch I have corrected this. Thank you for pointing out.

[KamathForAIX]

@shibatch It is possible since I see
-- Performing Test COMPILER_SUPPORTS_FLOAT128 - Success
-- Performing Test COMPILER_SUPPORTS_QUADMATH
-- Performing Test COMPILER_SUPPORTS_QUADMATH - Success
-- Performing Test COMPILER_SUPPORTS_VSX
-- Performing Test COMPILER_SUPPORTS_VSX - Success
-- Performing Test COMPILER_SUPPORTS_VSX3

In my PPC Linux LPAR.

[shibatch]

I am sorry, but the AIX support will be removed before SLEEF 4.0 is released.

#673

You may think that the code for AIX support is minimal and untested, so the cost may seem low.
However, even if the code itself for AIX support is small, having a lot of such code makes the code dirty and difficult to maintain.

Especially for IBM related architectures, there are already two, powerpc and S390X, and S390X is a big endian architecture, which is rare these days, so there is an extra maintenance cost. These architectures do not have a large market share, and we cannot accept additional cost increases on top of the maintenance costs we are currently paying for free.

[KamathForAIX]

@shibatch I understand your concern.

But we do need sleef to be working on AIX for PyTorch. Otherwise our customers and users will not be able to do pip3 install torch since the sleef submodule will not compile.

AIX is putting an effort to modernize, and we are supported in CMake, ONNXRuntime, Scipy, Numpy, Pandas, Meson, and even in talks with the PyTorch community via PRs for AIX support.

You can feel free to change anything you want to in the future in Sleef without worrying about AIX. We have CIs internally running and will make sure Sleef is working in AIX, and if something breaks, I will raise a PR to fix it.

You can also connect with @ranjitshs, @ayappanec or me (@KamathForAIX) for any queries related to AIX.

Having said that, is there anything we can do that would work for AIX and you that you have in mind? Kindly let me know. We are open to suggestions.

[shibatch]

At this time, the SLEEF project has no sponsor and is basically maintained entirely by myself.
As I have repeatedly stated in the CoC below and elsewhere, it is my view that if a company is profiting from the deliverables of this project, they deserve to be a sponsor of this project.
As I mentioned above, as far as IBM is concerned, I already maintain two architectures, powerpc and S390x, for free, and I do not see how it makes sense for me to provide any more effort for free.
IBM is asked for sponsorship.

https://github.com/shibatch/nofreelunch?tab=coc-ov-file

#673