Use output dir for empty packages to be hermetic

[joeljeske]

What type of PR is this?

Bug fix

What does this PR do? Why is it needed?

Some packages are not deterministic, rather they have different content hashes when rebuilt. This occurs when rules_go creates an _empty.go file to satisfy the compiler. In those cases, the empty file is created in the OS temp dir, which changes across compiler invocations. The path of this file changes the content hash of the resulting package, thus breaking caches and downstream builds.

This PR changes the location of this tmp file to the output path of the package, such that it is the same output path across compiler invocations.

Reproduction steps:

bazel build //tests/core/go_library:empty 
md5 bazel-bin/tests/core/go_library/empty.a

rm -f  bazel-bin/tests/core/go_library/empty.a

bazel build //tests/core/go_library:empty 
md5 bazel-bin/tests/core/go_library/empty.a

Other notes for review

I noted a broad use of tmpdir or workdir within rules_go, and this issue may be prevalent in other conditions. If determinism is a priority, perhaps an audit of tmp dir uses may be warranted.

It would be useful to have a test that validations reproducibility of this kind, but I'm not aware of a good way to structure that type of test. Reproducibility test has been added

[joeljeske]

This strategy is already being used here https://github.com/bazelbuild/rules_go/blob/master/go/tools/builders/compile.go#L98

[joeljeske]

There is a failure with windows resolving the empty filepath. Presumably it might be fixed by resolving the absolute path and using that instead of the relative path, but then I'm concerned it would suffer from the same non determinism.

I'm open to suggestions. I also don't have a windows box for testing this

[achew22]

@joeljeske, it seems like you're onto something here. WRT writing a test that covers this, I think we can use the preexisting reproducibility_test.go. It would be nice to have a test that failed before merging this in, would. you be willing to take a whack at it?

[joeljeske]

Fantastic, @achew22, thanks for the tip. I had not seen those tests. I added a simple empty lib test that fails at its commit.

[achew22]

Great contribution, thank you @joeljeske for helping out!