15:25 <graingert> this was a bit of a head scratcher

15:26 <graingert> this terminates on pypy2.7 and pypy3.10 but loops forever on cpython3.10 https://www.irccloud.com/pastebin/11FrLMyJ/demo.py

15:27 <graingert> same if you use keyword arguments: other(a=A())

15:30 <cfbolz> graingert: in cpy main keeps the *args tuple alive?

15:35 <graingert> python3.11 introduced the "move" behaviour for some function cals

15:35 <graingert> calls*

15:35 <graingert> but pypy already has move for kwargs or *(expr, )

15:38 <graingert> call_function does the pop before the call https://github.com/pypy/pypy/blob/d92d0bfd38318ede1cbaadadafd77da69d431fad/pypy/interpreter/pyopcode.py#L1208

15:43 <graingert> also the behaviour is different when run with the built pypy and the rpython pypy

15:44 <cfbolz> graingert: yeah, that's a little bit to be expected. some slightly weird things are going on in the non-translated emulated mode

15:45 <graingert> I think this https://github.com/pypy/pypy/blob/d92d0bfd38318ede1cbaadadafd77da69d431fad/pypy/interpreter/pyopcode.py#L1217 args = object is keeping the arguments alive

15:58 <cfbolz> graingert: you mean before translation?

15:59 <graingert> Yes

16:00 <cfbolz> yeah, that makes sense

16:02 <graingert> The clue was when I fixed the zpickle tests

16:02 <graingert> https://github.com/pypy/pypy/pull/5087/commits/91ae44d33199a47679ff49468587ed8993ad2e2f

16:02 <graingert> Clearly there's some branches where dropvalues doesn't need to be called it the function started being called

16:07 <cfbolz> graingert: I'm quite warming to this effort, fwiw. at first I thought that your changes wouldn't matter for the JIT too much. but I've now convinced myself that even if peak perf stays the same, warmup time should be slightly improved (might not be measurable, but still)

16:12 <graingert> I'm trying to work out what's up with this

16:12 <graingert> https://www.irccloud.com/pastebin/3nUI9FdP/demo_send.py

16:23 <graingert> I'm assuming it's something up with gateway but I don't know how rpython object lifetimes work

16:24 <cfbolz> graingert: the first send is to start the generator, right?

16:24 <graingert> Yes

16:25 <cfbolz> interesting. how does cpython do?

16:25 <graingert> I suspect whatever the fix is will fix the wired tuple in gc.get_referrers

16:25 <graingert> Weird*

16:26 <graingert> cpython3.11 12 and main hangs

16:26 <cfbolz> ok

16:26 <cfbolz> are you filing cpy bugs about some of these too, btw?

16:27 <graingert> Well maybe

16:28 <cfbolz> was mainly curious

16:30 <graingert> Needs everyone to agree what python is first

16:31 <graingert> I posted this https://discuss.python.org/t/visible-behaviour-change-in-3-11-python-to-python-calls/68534/6

16:32 <graingert> But everyone got bogged down in garbage collector differences

16:32 <graingert> It's not about the GC it's about reachability

16:33 <graingert> And I think reachability should be the same on all python implementations

16:33 <LarstiQ> cfbolz: was that "warming to this effort" a warmup pun? :)

16:33 <cfbolz> graingert: I'm sure there is a strong argument for why all of this is implementation defined. but I also think it's very useful to not be too horrible about it

16:34 <cfbolz> LarstiQ: only accidentally

16:35 <graingert> One interesting problem is it's hard to implement a function and maintain lsp

16:35 <graingert> class A:

16:35 <graingert> One moment

16:37 <graingert> https://www.irccloud.com/pastebin/i5PSMAG0/demo.py

16:38 <graingert> B().method keeps it's argument alive

16:41 <graingert> All a bit hairy really

16:49 <graingert> Is there a way to find out why an object is still alive when running a translated pypy-c?

18:16 <cfbolz> graingert: not easily. one problem is that the rpython object graph has a lot of intermediate objects that aren't really valid python objects. we have an api exposed to introspect those a bit from python anyway, but you don't get a lot of interesting information out of them

18:16 <cfbolz> that could be fixed, but it would be quite a bit of work I think

18:58 <graingert> The extra reference is here https://github.com/pypy/pypy/blob/d92d0bfd38318ede1cbaadadafd77da69d431fad/pypy/interpreter/gateway.py#L792

19:19 <cfbolz> graingert: ah

19:20 <cfbolz> how did you find it? unfortunately it's pretty hard to get rid of, I think

19:35 <graingert> Manually reading the code cfbolz

19:35 <cfbolz> right :-\

19:36 <graingert> I think it's a case of moving all the DescrMismatch processing into the behavior

19:36 <graingert> Bit of a pain, but it's the same behavior as cpython so maybe not so much of a problem

19:41 <cfbolz> that's a lot of changes at once. let's maybe finalize the call_valuestack change first

19:47 <graingert> Yeah I think so

19:50 <graingert> Why doesn't the finally: frame.dropvalues(args + 2) get run when using interp_pickle?

19:50 <graingert> Eg why was this hack needed? https://github.com/pypy/pypy/pull/5087/commits/91ae44d33199a47679ff49468587ed8993ad2e2f#diff-55060a0771af9366a19e7cb604ce593c13e4ff841e47135a7fe5e397e112a3b1R117

20:00 <graingert> Also can one run the zpickle tests without using cpython2.7? Eg what happens if you can't get cpython2.7 to run anymore?