[Python-Dev] new failing test -- test_compiler

[Tim Peters]

[Armin Rigo]
> ...
> This is now fixed by skipping the optimization if we got an interned string.
> While messing with ob_refcnt is always a bit fragile I am still convinced that
> the += optimization is safe again.

Ya, this bug wasn't scary (an unlikely but clear-in-hindsight problem,
specifically due to a subsystem lying about true refcounts -- not much
of that goes on, thank God).

Alas, test_compiler still fails (but in an ordinary "doesn't work"
way, no longer in an internal-error way), so testing with -uall is
still a bit broken.