[PyQt] [SIP] Critical issue since version 4.18

[Antoine Lambert]

Hi,

I am the maintainer of the Python bindings for the Tulip framework : an
open source graph visualization tool. I have recently upgraded to SIP
4.18.1 and one critical issue appears in our generated bindings leading to
a segfault.

The bug is the following. We use SIP in order to implement plugins in
Python for the Tulip framework. Implementing a plugin consists in writing a
Python class that derives from a Tulip wrapped plugin interface (see
http://pythonhosted.org/tulip-python/pythonplugins.html). For instance to
write a general graph algorithm plugin, one has to write a class that
derives from the 'tlp.Algorithm' Tulip Python class.
The 'tlp.Algorithm' class wraps the 'tlp::Algorithm" C++ class (
https://sourceforge.net/p/auber/code/HEAD/tree/tulip/
library/tulip-python/bindings/tulip-core/Algorithm.sip). The
"tlp::Algorithm" class derives from the "tlp::Plugin" class (
https://sourceforge.net/p/auber/code/HEAD/tree/tulip/
library/tulip-python/bindings/tulip-core/Plugin.sip), itself deriving from
"tlp::WithParameter" and "tlp::WithDependency" (https://sourceforge.net/p/
auber/code/HEAD/tree/tulip/library/tulip-python/bindings/
tulip-core/WithParameter.sip, https://sourceforge.net/p/
auber/code/HEAD/tree/tulip/library/tulip-python/bindings/
tulip-core/WithDependency.sip). The issue is that now every call to a
method from the "tlp.WithParameter" class inside a Tulip plugin source code
leads to a segfault during the plugin execution.
I managed to get back to the SIP revision that introduces the regression
through a session of "hg bisect" : this is revision 1425 (https://www.
riverbankcomputing.com/hg/sip/rev/14bfbaf7431a). It seems that some needed
class cast functions are no more generated since that revision which then
leads to segfault due to incorrect pointer cast (just my guess).
As a temporary workaround, i naively patch the SIP source code bundled in
the Tulip source tree forcing the generation of these cast functions (
https://sourceforge.net/p/auber/code/11726/).

I also found another issue that is not critical but still problematic. When
wrapping non const global C++ variables, setting their value in Python does
not modify the content of the wrapped variables. For instance, in the Tulip
Python bindings, we wrapped some global variables holding some file system
paths (e.g. "tlp.TulipBitmapDir", https://sourceforge.net/p/
auber/code/HEAD/tree/tulip/library/tulip-python/bindings/
tulip-core/TlpTools.sip).
Even if it seems that when modifying the variable content it is taken into
account (in particular when printing it), the content of the wrapped
variable is not changed.
I assumed there is some kind of cache for wrapped const global variables
but it should not be used for non const ones.
That issue was already present in SIP 4.17, I should have report it earlier
but its was not critical and I managed to workaround it using a Python
meta-class (https://sourceforge.net/p/auber/code/HEAD/tree/tulip/
library/tulip-python/bindings/tulip-core/__init__.py).

Hope that this report will help to fix those annoying issues for the next
SIP version.

Best regards,

Antoine Lambert
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://www.riverbankcomputing.com/pipermail/pyqt/attachments/20161114/d2a6589a/attachment.html>