py-tensorflow @2.6.0 on M1: Architecture confusion for dependencies

[p-bro]

When attempting to install the py-tensorflow port on MacOS 11.6 on M1, port install py-tensorflow command bails out with the following error:

Error: Cannot install py39-tensorflow for the arch 'x86_64' because
Error: its dependency bazel-3.7 does not build for the required arch by default
Error: and does not have a universal variant.
Error: Follow https://guide.macports.org/#project.tickets if you believe there is a bug.
Error: Processing of port py-tensorflow failed

A full log is attached. For some reason ports is confused about the architecture it's running on: At no point I requested the x86_64 architecture. All my configuration files are the default.

Directly requesting the py39 version with the command port install py39-tensorflow does not have this issue and proceeds to install dependencies in the correct architectures.

I encountered the bug on py-tensorflow, but there may be something more fundamental going on here, and the bug may actually be with base.

[p-bro]

output of sudo port -d install py-tensorflow +native

[jmroot (Joshua Root)]

Replying to p-bro:

For some reason ports is confused about the architecture it's running on: At no point I requested the x86_64 architecture.

From the Portfile:

# Not yet ready for arm
supported_archs     x86_64

That should only be set in the subports, since py-tensorflow itself is noarch. But that's unlikely to be the whole problem.

It looks like bazel-3.7 can build for arm64 or x86_64, but not both (no universal variant). That makes it tricky to use as a dependency of x86_64-only ports on an arm64 system.

[jmroot (Joshua Root)]

If bazel's architecture doesn't matter to its dependents, they can use depends_skip_archcheck to indicate that (and that could be done in the portgroup). I don't know if that's the case.