Fix CommandLine class initialization deadlock.

[benjaminp]

#21566 (comment) shows a deadlock when initializing static variable CommandLine.EMPTY and another thread is initializing CommandLine's subclass AbstractCommandLine because CommandLine.EMPTY is an instance of AbstractCommandLine.

Fixes #21566.

[meteorcloudy]

@benjaminp Thanks so much for the fix!

/cc @justinhorvitz Looks like the bug was introduced in b82dcdc

[justinhorvitz]

@benjaminp thank you so much for the diagnosis and fix. Unfortunately we've already released the culprit change inside google yesterday and are seeing impact, but this PR allowed us to figure out the user complaints of stuck builds and start working on getting the fix released.

[justinhorvitz]

As a followup, I submitted a bug report to Intellij because the code inspections did not catch this: https://youtrack.jetbrains.com/issue/IDEA-348744/Static-initializer-references-subclass-inspection-doesnt-report-private-class. The irony is that initially I did get such a warning, but then rearranged the code to make the warning go away.

[benjaminp]

I've run into this type of issue a few times. It's surely one of darkest corners of the JVM especially given the superficial innocuousness of the code that unlocks deadlock gremlins in production. At least maybe this episode will yield better IDE and errorprone checks! IMO, hotspot should also have runtime deadlock detector for this issue. Static analysis cannot detect every case of the deadlock, and a "nice" JVM crash usually beats a hung process.