Currently, the WeakPointerTo comparison operators compare the raw pointers, but this is not useful as it may cause a false equality if one weak pointer in the comparison is expired and points to memory that has since been reused.
Instead, we can define a comparison based on the control block pointer, which exists since the new weak pointer implementation in 0bb81a43c9. This is implemented in the owner_before method, matching C++11 std::weak_ptr semantics.
I would now recommend deprecating most comparison operators of WeakPointerTo or redefining them to make more sense, ie. comparing equal if they (once) referred to the same object and not if they simply point to the same memory address. This has not yet been done, though code that uses the comparison operators has been fixed in this commit.
Overloads of std::owner_less have been provided for creating a map or set with Weak(Const)PointerTo keys.
Only the basics are supported; the __members__ or iter interface is not supported at this time.
See also #351 for discussion on pulling in enum34 module.
If a class inherits from ReferenceCount, it can be destructed by downcasting to ReferenceCount, so it should not be an obstacle to properly clean it up when such a class is returned from C++.
This issue comes up when a CycleData is returned via MemoryUsagePointers, which is not exposed so is wrapped as NodeReferenceCount instead, which has a protected destructor.
If any code is relying on this, please let me know and I will add it back. It appears to be redundant, though, since one can access DtoolClassDict directly.
Symbol kept around temporarily in order to keep ABI compatibility for a short while as people may not update their interrogate and Panda in sync, but it can soon be removed.
I don't know if this is the right solution, but it does fix an issue accessing std::ios::openmode caused by the ios typedef being defined before ios_base is fully specified.
Rationale:
1. Per standard, fstream::open takes 2 arguments.
If platforms add a third, they're out of spec.
2. The only platform I could find that takes a file
mask specifically as the third argument is IRIX,
which Panda hasn't targeted in forever.
3. The mask being requested isn't even particularly
interesting - falling back to a platform default
is best.
4. When USE_PANDAFILESTREAM is defined, pfstream
is implemented as PandaFileStream, which doesn't
have a three-argument open() and breaks immediately.
5. makepanda doesn't ever define HAVE_OPEN_MASK
6. It's been broken for so long that, if it were
important to anybody, it would have been fixed by now.
There is a bug in clang versions before 3.2 (including the one shipped with Xcode) that makes it give a "conflicting types" compile error when there is a static constexpr function defined outside the class. The way to work around this is either to remove one of the "static" or "constexpr" keywords, or to simply put the definition inline.
See: https://stackoverflow.com/a/17494592/2135754
I would try and upgrade Xcode to version 5 to see if the problem is fixed, but the buildbot still runs OS X Lion (10.7) and the last version of Xcode that works on Lion is 4.6.3, so it seems easier to just apply these workarounds for now.
This renames acquire/release to lock/unlock in order to be compatible with std::lock_guard and std::unique_lock (which will eventually replace the *MutexHolder classes). It will also allow us to typedef MutexImpl to std::mutex later on.
To access a WeakPointerTo in a thread-safe way, use something like this:
if (auto ptr = weak_ptr.lock()) {
..use ptr as regular PointerTo
}
The new implementation no longer needs a reference to be stored to all weak pointers on the WeakReferenceList; a mere count of weak pointers is sufficient. Therefore, callbacks theoretically no longer require a WeakPointerTo to be constructed.
The WeakPointerTo class is not actually atomic; it could be made so, but I don't believe it's worth it at this time.
I have no idea why this error happens, but it does not seem worth the effort to investigate further, so I'm just reverting the previous change to this file.
I am not sure why the definition of NULL was removed; it might have been by mistake, but in any case it broke code like this:
img = PNMImage(w, h, color_space=CS_srgb)
since it would not understand the default value (NULL) for the argument preceding color_space.
This distinction allows us to better support Python 3, since it will raise exceptions when trying to put arbitrary binary data in a str object.
This also adds some convenience functions for efficiently initializing a Datagram or PTA_uchar from a vector_uchar.
This is designed to sanity-check the buildsystem, ensuring that the
expected BUILDING_ macros are defined at the expected time. It
also helps catch cases where the wrong BUILDING_/EXPCL_ macros
are used.
The GCC documentation states that, as of GCC 2.7.2, these aren't
necessary for proper program behavior. The documentation further
discourages their use because they don't suppress unnecessary code
duplication.
The Panda codebase these days uses "extern template class" instead,
which tells the compiler not to perform implicit template expansion
because an explicit template expansion is available for linking
elsewhere in the program. This is a more compiler-neutral way of
achieving the same thing as '#pragma interface', making
'#pragma interface' not only redundant, it could also mask problems
in the "extern template class" machinery.
As of f69354d9fa, this is just always
expected on all platforms. If it's not defined,
TemplateClass<X>::_type_handle won't link correctly with some compilers
(e.g. Clang) and the lack of 'extern template class' will result in
duplicated code in any case. In short, turning it off causes problems.
makepanda also always turns this on, so it doesn't represent any
optional feature or autodetected platform-specific behavior anyway.
This will help us catch cases wherein we inadvertently create a circular dependency between component libraries. Right now it only checks inheritance and typedefs, but this can be expanded in the future.
Also, sorts the module initialization by dependency topology. This isn't very important right now, but having this guarantee might help when refactoring module initialization in the future.
Typedefs are evidently always considered "fully defined", so we need to add another condition to the database merging code to prefer the type that is marked "global" when merging two types.
The binary path we get from /proc/self/exe isn't very useful; the path to the .apk is barely more useful but it still doesn't make a whole lot of sense. It might make more sense to set it to the path of the native .so that is being loaded by NativeActivity.
- allow setting API target with --target=android-21
- always link to libpython on Android, seems to be necessary
- support aarch64 (arm64-v8 ABI) architecture
- enable building on an Android machine (tested in termux)
[skip ci]
The new implementation uses a message-only window to register raw input events.
XInput is still used when possible, but raw input is being used to handle device detection. (XInput is only enabled when an XInput device is plugged in, which also prevents us from having to load and poll the XInput library unnecessarily.)
For raw devices, the Windows HID parser library is used to parse the raw data. Unfortunately, the Windows 7.1 SDK does not ship hid.lib (which is in the WDK 7.1.0), so I am dynamically loading hid.dll to prevent pulling in a dependency on the WDK. (But perhaps we can ship hid.lib in the thirdparty libraries?)
Input devices other than XInput gamepads and the 3D mouse have not been tested very well yet. Adding keyboard and mouse support is still a TODO.
This also splits out the Windows implementation of InputDeviceManager into a separate subclass; it is intended that the other implementations will follow suit.
Gets rid of properties defined as both MAKE_SEQ_PROPERTY/MAKE_MAP_PROPERTY, which are just a bad idea. Instead, adds a way for map properties to define a separate "keys" interface.
Fixes: #203
This remove support for coercing non-ReferenceCounted types that are neither default-constructible nor move-assignable, but it turns out none of the classes we really need it for matches that.
It further cuts down on the amount of code that is being generated to support coercion in cases where it makes absolutely no sense.
* handle static methods with explicit self
* fix len() and __setitem__ of mapping types
* fix inheritance of __getattr__ and __setattr__
* fix overload resolution error with nullptr_t arguments
* bool overloads now come after float/double overloads
* record whether class is final in interrogatedb
* add Dtool_EmptyTuple
* optimization for final classes: no need for downcast
rdb
Mitchell Stokes
Sam Edwards
Psychotropos
Younguk Kim
frainfreeze