Hi Christian!
Christian Tismer wrote:
I spent the last two days hunting refcounting problems
with pypy generated extension modules.
After infinitely long staring at the final flow graphs,
I found out that our transition to the gctransform
is not as complete as I assumed. Impossible too find
by flow graphs, because we are still chiming extra increfs
in from the backend, see funcgen.py .
I tried to solve this by disabling all the extra increfs,
but this doesn't work, yet. Something else must be adjusted.
The problem is an inconsistency about how getattr is handled.
The code in funcgen.py always adds an incref when we access
a field or array item of a PyObject.
I am not exactly surprised. When we wrote the gctransformer we were
mostly annoyed by PyObjects and did the minimal thing to make all tests
that use them pass.
In gctransformation.py, there are deallocators created,
which produce code like
v_xxx = obj.xxx
pop_alive(v_xxx)
but this doesn't work for PyOnject, since funcget.py adds
an extra ref, and the object stays as alive as it was before.
I hoped to solve this, assuming that our gc code is generic
enough to get this right for PyObject as well. Maybe I'm missing
something, have to give up. I'm probably anyway interfering
with other people's work.
It would be quite helpful if you checked in a small test that shows the
faulty behaviour. It's a bit hard to look for the problem if you have no
way to reproduce it.
Anyway, as I said, the gctransformer is a complete mess right now and
needs a rewrite -- which we will tackle as soon as we found out why our
exception transformation crashes.
Cheers,
Carl Friedrich
_______________________________________________
[email protected]
http://codespeak.net/mailman/listinfo/pypy-dev
