add aarch64, s390x, ppc64le manylinux2014 support

[mattip]

No description provided.

[mattip]

these changes are to better support searching for "manylinux1" and "manylinux2010"

[mattip]

I only see appveyor checks, shouldn't PRs trigger travis as well?

[mattip]

manylinux2014 supports and tests aarch64, and will soon get ppc64le support too. Would be nice to support that as well here.

[mattip]

The suppress test is failing since i is not a valid option in bash to set? How does this pass on other platform? The functioning of set_opts is a bit obscure, what is it meant to do?

[mattip]

According to the documentation, bash's set does not accept i

[hrw]

aarch64 build fails because component you are building has too old 'gnu-config' files.

[mattip]

it seems that is libtiff. Again (like the question about set -i on bash) - how does this pass on other platforms?

[hrw]

Simple: other platforms are older than aarch64 so libtiff's copy of 'gnu-config' knows them.

Update config.{guess,sub} before running "./configure" and it will work.

[mattip]

This PR is now stuck on openblas builds. NumPy is using builds for aarch64 and s390x from packages built and uploaded by @tylerjereddy to a dropbox location. I think it would make sense to copy those to the GF_LIB_URL used here.

[hrw]

What is wrong with openblas CentOS packages?

[mattip]

Yum is broken in the image. See the details below. I could download the openblas as used in numpy, but it is a different version. Ideally I would like to get openblas built and uploaded via https://github.com/MacPython/openblas-libs, but in order to do that this PR needs to go in first.

++(/io/./library_builders.sh:98): build_openblas(): yum install epel-release -y

There was a problem importing one of the Python modules

required to run yum. The error leading to this problem was:

   /usr/local/lib/liblzma.so.5: version `XZ_5.1.2alpha' not found (required by /lib64/librpmio.so.3)

Please install a package which provides this module, or

verify that the module is installed correctly.

It's possible that the above module doesn't match the

current version of Python, which is:

2.7.5 (default, Aug  7 2019, 00:57:09) 

[GCC 4.8.5 20150623 (Red Hat 4.8.5-39)]

If you cannot solve this problem yourself, please go to 

the yum faq at:

  http://yum.baseurl.org/wiki/Faq

[hrw]

Remove /usr/local/lib from LD_LIBRARY_PATH when you call yum?

I do not get why you insist on rebuilding most of libraries by hand instead of using distro ones - would make it easier.

[mattip]

@hrw for NumPy we would like to use OpenBLAS 3.7, which is not available in the distros.

[hrw]

https://github.com/hrw/multibuild/tree/hrw-devel has gnu-config change.

for me openblas build fails due to fetching not existing file

[hrw]

https://travis-ci.org/hrw/multibuild/jobs/620353346 is travis job for aarch64

[mattip]

I was so encouraged by 6989374 passing tests that I tried adding s390x and ppc64le as well, so this still should not be merged

[hrw]

Looks like ppc64le needs PLAT.

[hrw]

ppc64le wants config.{guess,sub} update for yaml ;(

[mattip]

libyaml was 0.1.5, latest is 0.2.2. The Major release announcement did not mention any API changes, so I will upgrade

[mattip]

Now I think this is ready: the s390x, ppc64le, aarch64 builds passed in isolation. I restored the other CI test runs (lots of macOS there) and updated the README

[mattip]

The only failing build is a macOS one with the message

warning: The i386 architecture is deprecated for macOS (remove from the Xcode build setting: ARCHS)

Let's leave that for a different "macOS test cleanup" PR (which I have no idea how to do since I know nothing about macOS).

[mattip]

@matthew-brett any thoughts?

[matthew-brett]

Seems reasonable to me. Any other takers for review? @radarhere ?

[matthew-brett]

I guess the test failure here is because lcms2-2.9 has dropped support for dual arch build or 32-bit build, on MacOS. I wonder if this will get fixed by the just-merged default-to-macos-10.9 PR - #279

[mattip]

All builds are green.

[matthew-brett]

Let's give it a shot ...

[robbuckley]

I guess the test failure here is because lcms2-2.9 has dropped support for dual arch build or 32-bit build, on MacOS. I wonder if this will get fixed by the just-merged default-to-macos-10.9 PR - #279

@matthew-brett yes, this fail seems to have been due to the policy for macOS builds prior to #279 when it would default to building for 10.6 / dual arch for everything except MB_PYTHON_VER=3.8

see https://travis-ci.org/matthew-brett/multibuild/jobs/620819251#L389