py38-scikit-learn fails to build on macOS 11 Big Sur

[smeingast (Stefan Meingast)]

After the latest fix to clang-9.0, I tried to build scikit-learn and immediately ran into a build error. Log attached.

[smeingast (Stefan Meingast)]

log

[cjones051073 (Chris Jones)]

The port is using macports clang 9.0 compiler to build (not clear exactly why to me ?) but setting the deployment target to 11.0, which this older version of the compiler does not understand

:info:build clang: error: invalid version number in 'MACOSX_DEPLOYMENT_TARGET=11.0'

The solution is to either move to a new compiler that knows about the 11.0 target, or set the target for this build to 10.16, e.g. something like

 if { ${os.major} == 20 } {
    macosx_deployment_target 10.16
}

maybe with additional checks to only do that if macports clang 9.0 (or 10) is used. clang 11 (coming soon) might well understand 11.0.

[cjones051073 (Chris Jones)]

Another option is to patch the macports clang 9.0 and 10 compilers to understand 11.0 as a deployment target. I am not sure if this is just a matter of adjusting some logic to handle 11.x as well as 10.x, or more involved. Ken, thoughts ?

[cjones051073 (Chris Jones)]

This might form the basis of a patch

​https://github.com/llvm/llvm-project/commit/f724ce0d73eb3f85364e346a036588825bc47567#diff-edd349ed46d0559be15de04eb20dca7a0bf9a60e694128ece9de6b7fa62dae5a

[cjones051073 (Chris Jones)]

patch for clang-9.0 seems to have worked fine. Without patch

Oberon ~/cernbox/MacPorts/c++17 > MACOSX_DEPLOYMENT_TARGET=11.0 clang++-mp-9.0 -std=c++17 test.cpp
clang: error: invalid version number in 'MACOSX_DEPLOYMENT_TARGET=11.0'

with

Oberon ~/cernbox/MacPorts/c++17 > MACOSX_DEPLOYMENT_TARGET=11.0 clang++-mp-9.0 -std=c++17 test.cpp
Oberon ~/cernbox/MacPorts/c++17 > 

will test similar one for clang 10, and then look to preparing a PR.

[kencu (Ken)]

seems to me the safest approach, and a simple backported one-liner. Might have to also apply in the lldb subport, NB.

I have another minor patch for clang 9 & 10 to do as well, so I can put a few of these small ones together and avoid everyone rebuilding clang too many times. It's the reinplace issue I fixed in the recent clang-devel.

There's a more complicated reinplace issue involving a debug variant of clang that I might never fix...

[cjones051073 (Chris Jones)]

Yeah, definitely a good idea to avoid rev-bumping more than necessary...

I'm just test building clang-10. Once done I'll run few basic checks and then push what I have to a branch in GitHub, so you can take it from there...

[cjones051073 (Chris Jones)]

See this branch Ken

​https://github.com/cjones051073/macports-ports/commits/clang-9-10-support-darwin-11.0-target

this commit

​https://github.com/cjones051073/macports-ports/commit/e7464761d4c9eb1d498adbc9059c1a2963f79469

[kencu (Ken)]

thanks. will do it this afternoon.

[cjones051073 (Chris Jones)]

py-scikit-learn now builds fine, with the patched clang-9.0. So when committed can be tagged to close this ticket.

[smeingast (Stefan Meingast)]

have my thanks guys, you are just wonderful (and quick)! :)

[cjones051073 (Chris Jones)]

Ken, could these clang fixes be committed asap ? There is a growing number of other tickets for this, e.g. ..

​https://trac.macports.org/ticket/61549

[kencu (Ken)]

should be fixed by 99b6a78ddf9eabc4f76fb3deb00615721100092d/macports-ports

[cjones051073 (Chris Jones)]

Thanks, although there is a typo in that link as it takes me to a ccache change from 4 years ago;)

​https://github.com/macports/macports-ports/commit/99b6a78ddf9eabc4f76fb3deb00615721100092d

Looks like the right one