[PyKDE] sipTypeDef structure layout and invalid casts

[Phil Thompson]

On Wednesday 07 April 2004 8:26 pm, Dave Cuthbert wrote:
> [Apologies if this appears twice -- I didn't realize that posting via
> gmane wasn't the right way of sending to this list...]
>
>
> I'm currently debugging some custom QCanvasItem subclasses wrapped by
> PyQt -- this is a bit of an unfriendly case, since QCanvasItem doesn't
> use the standard MOC metaobjects, but has its own rtti() method <sigh>...
>
> Anyway.  We're on commercial versions of SIP 4.0rc3 + latest patches,
> PyQt 3.11 + some patches to allow us to subclass from QCanvasItems.  Our
> module is named "tokyo", which I'll use in my code excerpts below.
>
>
> The issue I'm running into is the casting of sipTypeDef structures into
> sipWrapperTypes.  siptokyocmodule.cpp contains an area like:
>
> ====
> static sipWrapperType *typesTable[] = {
> 	(sipWrapperType *)&sipType_TextUI,
> 	(sipWrapperType *)&sipType_EditBuffer,
> 	(sipWrapperType *)&sipType_GSelectionTableItem,
> 	/* ... */
> 	0
> };
> ====
>
> Each sipType_XXXX structure is a sipTypeDef defined something like (for
> TextUI):
>
> ====
> /* The main type data structure. */
> sipTypeDef sipType_TextUI = {
> 	0,
> 	"tokyo.TextUI",
> 	-1,
> 	supers_TextUI,
> 	0,
> 	5, methods_TextUI,
> 	0, 0,
> 	0,
> 	init_TextUI,
> 	dealloc_TextUI,
> 	cast_TextUI,
> 	forceConvertTo_TextUI,
> 	0,
> 	sipNewProxy_tokyo,
> 	signals_TextUI,
> 	{0, 0, 0, 0, 0, 0}
> };
> ====
>
> The problem is that this layout doesn't match sipWrapperType, which is
> defined in sip.h as:
>
> ====
> typedef struct _sipWrapperType {
> 	PyHeapTypeObject super;		/* The super-metatype. */
> 	struct _sipTypeDef *type;	/* The additional type information. */
> } sipWrapperType;
> ====
>
> I'm not sure how this is supposed to work.  A sipWrapperType "inherits"
> from PyHeapTypeObject, which inherits from PyTypeObject; sipTypeDef
> looks nothing like this, so I think the casts are invalid.
>
> My crash is happening when we try to downcast and end up sending these
> structures into Python as a PyTypeObject (we either see serious stack
> corruption or coredumps deep in Python itself).
>
> Any insight?  I think this is a SIP bug, but I could just be
> misunderstanding something.

You are (understandably) misunderstanding. You need to look at the hack in 
siplib.c where createType() is called. The sipTypeDef pointer is passed to 
createType() and the returned sipWrapperType pointer replaces the original 
pointer in the table.

Phil