11:19 <mattip> an HPy handle can be assigned to a global variable in C, and in CPython the handle is preserved over different C function calls

11:19 <mattip> like "HPy keep" in this test

11:19 <mattip> https://foss.heptapod.net/pypy/pypy/-/blob/branch/hpy-0.0.4/extra_tests/hpy_tests/_vendored/debug/test_handles_invalid.py#L114

11:21 <mattip> but I thought the point of handles is to make sure to close all the arguments to a function call

11:22 <mattip> so how can 'keep' stay alive between the function calls?

11:42 <cfbolz> mattip: isn't this a test about

11:42 <cfbolz> an illegal use of the API?

11:44 <mattip> right, but how does "mod.g() == len(s)" even pass on CPython?

11:44 <mattip> mod.g() should raise, since it should be using an invalid handle

11:44 <mattip> (it used to crash on PyPy, but now raises)

11:45 <mattip> maybe I should ask on the hpy IRC channel

11:46 <cfbolz> mattip: I suspect it works by accident, because a handle is just a PyObject* outside debug mode

11:51 <mattip> ok, let's go with the "works by accident", I fixed the test assuming that is correct

13:28 <mattip> progress: extra_tests/hpy_tests does not crash, there are less that 10 failures/errors

13:28 <mattip> on the hpy-0.0.4 branch

13:35 <phlebas> Is there a way for me to open a merge request for pypy/benchmarks on heptapod? It seems I cannot a repository without manual (human) action on their side, so I cannot create a merge request without access to the repo itself to push my branch there?

13:41 <phlebas> s/I cannot/I cannot fork/

13:42 <Hodgestar> mattip: antocuni will know better than me, but I think reading the comment at the top of hpy/debug/src/debug_internal.h might help. If I understand debug mode correctly, then, in that test, "keep" is a DHPy handle (i.e. an HPy containing a pointer to a DebugHandle). Thus it's valid to store it for the implementation of the debug context within the HPy repo. I don't know too much about debug mode in PyPy (maybe it is the same, maybe not).

13:44 <Hodgestar> Outside of debug mode, of course storing the handle in keep is totally broken and could do almost anything if used later.

14:02 <mattip> so maybe this is a bug with how PyPy manages handles. There is a rpython-level handle manager on top of the c-level one

14:03 <mattip> the rpython debug handle manager does not have a closed_handles list

14:36 <ronan> phlebas: heptapod doesn't do forks, you need to ask for access to the repo

14:37 <Hodgestar> mattip: Maybe the PyPy debug handle manager hasn't had the feature added?

14:39 <mattip> was this a 0.0.4 addition?

20:04 <Hodgestar> mattip: It was.

20:05 <Hodgestar> Or at least the test was.

20:10 <Hodgestar> The general structure of debug mode hasn't changed much since 0.0.3, but the new tests relies on the hpy repo implementation. We could probably revisit the semantics a bit if needed.