OpenTK now directly calculates the elapsed time between UpdateFrame
(RenderFrame) events and compares that directly to TargetUpdatePeriod
(TargetRenderPeriod). This significantly simplifies the implementation
and improves timing stability.
PlatformFactoryBase provides a common base interface for platform
backends. Platform backends should inherit from PlatformFactoryBase in
order to reduce code duplication.
LegacyJoystickDriver implements the legacy IJoystickDriver interface
(GameWindow.Joysticks) in terms of the new IJoystickDriver2 interface
(OpenTK.Input.Joystick).
This removes a large chunk of code from each platform backend, as they
no longer need to implement IJoystickDriver themselves. Additionally,
it adds support for device hot plugging which was previously missing.
This patch modifies GameWindow.Run() to use a single stopwatch instead
of two separate stopwatches for timing UpdateFrame and RenderFrame
events.
It improves timing accuracy for issue #20 (FrameEventArgs.Time
Inconsistencies)
DeviceAdded already checks that devices conform to the desired usage
pages. Checking again in DeviceRemoved is unnecessary - if a device
exists, then it has already passed muster.
JoystickAxis/Button.Last is used internally to allocate the correct
amount of storage for joystick axes and buttons. JoystickAxis.Axis10 is
required to support the maximum number of axes available on Mac OS X.
SDL2 uses a weird system of device ids and instance ids to report
joystick events, where the ADDED event uses a device id and the rest use
instance ids.
The SDL2 joystick driver is now fixed to correctly distinguish between
the two, which fixes hotplugging support for joystick devices.
SDL GameControllerAxis and GamePadAxes are not interchangeable. The
driver will now correctly interpret incoming SDL messages and update
the GamePadState for the relevant axis.
GamePadState.SetAxis() receives a GamePadAxes enumeration, which is a
bitmask of the axes we wish to set. SetAxis now correctly decodes the
bitmask to apply the values we are interested in.
This test uses SDL2 to create a window and an OpenGL context. It then
uses OpenTK to render into the external SDL2 context. If everything is
working correctly, a black window should appear and gradually turn
white before disappearing.
When combining OpenTK with a third-party OpenGL toolkit, it is now
possible to implement a suitable GetAddress() and GetCurrentContext()
implementation in terms of the third-party toolkit. If no
implementation is specified, then OpenTK will try to guess the most
suitable implementation within its own platform backends.
If no custom implementation is defined, and if no suitable
implementation can be found, then OpenTK will throw a
PlatformNotSupportedException. If a suitable implementation is found or
defined, then OpenTK will attempt to load OpenGL entry points using
that implementation.
In this case third-party toolkit remains solely responsible for
managing its context via its MakeCurrent(), SwapBuffers(), etc
implementations.
DummyGLContext will now attempt to load OpenGL and OpenGL ES entry
points when a suitable OpenGL context is current on the calling thread.
This allows OpenTK to be used on contexts created through third-party
toolkits.
CreateGetAddress() constructs a GraphicsContext.GetAddressDelegate that
is suitable for the current platform. This can be used when combining
OpenTK with an OpenGL context created through a third-party toolkit.
The functions defined in libdl.dylib are useful in more places than
just AglContext. Moving them to their own class ensures we can access
these from wherever we might need them.
CGL appears to work with both AGL and NSOpenGL contexts, whereas AGL is
limited to AGL contexts. This allows us to be more flexible in terms of
implementation (i.e. we can use Cgl.GetCurrentContext to retrieve a
handle to a context created through SDL, which uses NSOpenGL
internally.)
When the user requests a GraphicsMode that is not directly supported
by the GPU/drivers, we should relax the requested parameters until
we find a supported mode. An exception should only be thrown when
there is no usable mode.
This makes the X11 backend match the behavior of Windows. The SDL/X11
backend works a little bit differently, in that it falls back to the
a default mode directly if the requested mode is not available. There
is nothing we can do about that.
The _NET_FRAME_EXTENTS atom is implemented differently by
different window managers, when window decorations are hidden
with Motif. Unity returns a 0 size, while Gnome 3 returns the
previous size.
This patch removes that ambiguity: when decorations are hidden,
border size becomes zero. This should work everywhere, unless
some window manager decides to troll us by decorating the window
when we explicitly request no decorations. Sigh...
Windows can now be resized programmatically, even when they have
WindowBorder.Fixed. All resizing logic is now consolidated in the
Bounds property, and ConfigureNotify messages are now handled
correctly depending on their source (StructureNotify or
SubStructureNotify.)
WinMM is optimized for general joystick use, not for the canonical
GamePad layout. Instead of exposing IGamePadDriver directly, it should
expose IJoystickDriver2 and use a mapping driver to get GamePad support.
GraphicsMode.Index is set by the platform-specific context constructor,
which is invoked after the X11GLControl constructor. It does not make
sense to check GraphicsMode.Index in the X11GLControl constructor, as it
is never set at that point.
Without this flag, OpenGL rendering does not work as expected.
Additionally, all WGL_ARB_pixel_format attributes are expected to be
specified in key-value pairs. Fixed double-buffering and stereoscoping
rendering attributes.
Fixed WGL_ARB_pixel_format attribute selection for doublebuffering,
stereoscopic rendering and hardware acceleration. Implemented
minimization strategy to select the optimal PixelFormatDescriptor in the
fallback path.
Instead of creating a list of all available formats and iterating
through that, we let the driver decide which is the best accelerated
format to use for the user parameters. If no such format exists, we fall
back to generic acceleration or software acceleration, in turn.
This affects issue #21
Starting with OpenGL 4.2, strings passed to GL.ShaderSource are allowed
to contain multi-byte characters in comments (issue #18). This patch
modifies the marshaling code to use UTF8.GetBytes in order to marshal
strings, instead of Marshal.StringToHGlobalAnsi().
Don't filter window messages passed to our window (see
http://blogs.msdn.com/b/oldnewthing/archive/2005/02/09/369804.aspx).
Additionally, return the correct values for all messages we are actually
handling and clean up unmanaged memory after we are done with the
window.
Instead of combining PeekMessage+GetMessage, we can simply call
PeekMessage(Remove) to achieve the same effect. This also allows us to
remove the IsIdle property, which is no longer used anywhere.
The temporary context is now retained until the actual context has been
constructed. If we don't do this, then WGL_ARB_create_context may fail
to work correctly on specific GPUs (e.g. Intel). This may affect issue
#19.
The correct way to query number of available pixel formats is to use
Wgl.Arb.GetPixelFormatAttrib(NumberPixelFormatsArb), not
Wgl.Arb.ChoosePixelFormats. This fixes an issue where Intel drivers
would fail to report any pixel formats in GetModesARB, even when
WGL_ARB_pixel_format is supported.
On many/most platforms, GraphicsContexts can only be released by the
thread where they are current. This means that the user must call
GraphicsContext.Dispose() or risk a resource leak.
Since we cannot release contexts on the finalizer thread, we should keep
strong references, instead of weak references, until the user explicitly
calls Dispose().
This patch fixes issues with SDL2 crashing when running the MonoGame
WindowsGL test suite.
Misbehaving clients that shall not be named here may call
GameWindow.Close() inside the GameWindow.Closing event. This causes
recursion in SDL2, crashing the application.
This patch adds a guard to protect against recursion when calling
GameWindow.Close().
Several projects are still using the last svn revision from the
sourceforge repository (r3127). These overloads provide an upgrade path
from r3127 to OpenTK 1.1.
The Mono 2.10 compiler fails when compiling extern methods that are not
marked as DllImport. We fix that by adding a method body that throws a
NotImplementedException instead.
Additionally, MonoDevelop 2.8 cannot open sln files with ToolsVersion 12.
The fix is to change ToolsVersion to 11.
Tools now go to the Binaries/Tools/[Debug|Release] directory. OpenTK
remains at BInaries/OpenTK/[Debug|Release].
Mono.Cecil and IKVM now reside under the Dependencies/managed/
directory.
Instead of modifying the name of an OpenGL symbol on the managed side,
before copying it to the unmanaged side, we perform the modification
directly on the unmanaged side. This reduces the total amount of
allocations in OpenTK by ~30% (673496 bytes in 10750 objects compared
to 930272 bytes in 15243 objects before this modification.)
WS_CLIPCHILDREN and WS_CLIPSIBLINGS appear to cause flickering on
specific video cards. OpenGL appears to work correctly without these, so
we'll disable them to return to OpenTK 1.0 behavior.
It is now possible to indicate that an application is not DPI-aware. In
that case, OpenTK will let the operating system handle DPI scaling. This
results in worse visuals (pixel doubling) but allows non DPI-aware
applications to continue working.
It appears that calli callsites cannot be decorated with the
“platformapi” calling convention like DllImport signatures can. This is
problematic since Windows uses stdcall by default and most other
platforms use cdecl.
There are three approaches to this issue, without going back to
delegate calls: (a) generate an unmanaged thunk that cleans up the
stack after a GL call; (b) use libFFI; (c) use cdecl *or* stdcall
everywhere and hope that the runtime can cope.
.Net 2.0 can detect and fix stdcall functions invoked through a cdecl
callsite. .Net 4.0 adds a configuration option to enable or disable
this fixup (faster p/invoke if disabled) and raise a MDA exception when
this condition is detected. (This affects x86 only.)
Mono appears to be able to cope with cdecl functions invoked through a
stdcall callsite.
More testing is required.
We should be able to use static pinvokes on platforms that do not
provide or require extensions and calli instructions on platforms with
extension APIs. This dinstiction will be implemented as a parameter in
the rewriter.
By using untyped integers instead of typed integers in the unmanaged
callsites, we allow monolinker to keep the exact set of enums that are
used by the user. Without this, we’d have to keep every single enum in
place to avoid missing type exceptions.
This does not affect the public signatures or the generated code in any
way.
This includes arrays of primitives and arrays of generics. Our code is
similar to the code generated by the Mono C# compiler for the "fixed"
construct. The .Net compiler produces slightly different code (two local
variables instead of one) - more research is required.
Default results in a managed calling convention which does not generate
unmanaged thunking code for parameter marshaling.
System.Runtime.InteropServices.CallingConvention.Winapi appears to
correspond to StdCall for calli callsites (this might be different for
pinvoke, which supports an unmanaged "platformapi" calling convention.)
Needs more testing to prove this is doing the right thing on non-Windows
platforms.
WGL was autogenerated a few years ago but never touched after that.
Since we use a tiny fraction of all available methods, it makes sense to
remove the unused ones. This reduces dll size and improves startup
times.
The rewriter will patch the body of methods marked with [AutoGenerated].
Methods that are implemented manually (e.g. various math helper
overloads) should avoid this attribute.
.Net will happily execute a calli with a generic return type, whereas
Mono will refuse to. Mono is probably doing the right thing here. Fixed
by resolving the generic return into a concrete type.
On Windows, entry points for OpenGL 1.0 and 1.1 are not exposed by
wglGetProcAddress. We fall back to LoadLibrary+GetProcAddress when
wglProcAddress fails.
OpenTK normally uses reflection to load bindings, instead of generating
huge constructors. Although reflection is faster on first load (thanks
to reduced JIT overhead), it fails to work correctly with monolinker.
This branch explores the performance of a direct binding.
When we enter the modal resize loop on Windows with ClipCursor set, we
cause a feedback loop where every resize causes the cursor to move and
every move causes a new resize. To fix this, we need to ungrab the
cursor when we are enter the modal loop.
Implementations may reuse OpenGL context handles that have been
destroyed. If a context is finalized but not Disposed, then OpenTK may
keep a reference to the old context handle, causing a crash when the
same handle is returned for a new context. To fix that, new context
handles will now replace old handles in case of a clash.
SDL_DestroyWindow must be called on the main thread. If the window is
finalized, the finalizer will push a CLOSE event to the event loop
(thread-safe) and the window will be destroyed on the main thread.
Sdl2InputDriver.Dispose() would call SDL_DelEventWatch with a different
"user_data" parameter than SDL_AdEventWatch. This caused the EventFilter
to remain registered and subsequently crash when closing and reopening a
window.
Trim regex will now correctly match GetInteger64 and other functions
ending in "64". It also uses a correct ending anchor to avoid matches
in the middle of a function name.
Scan through the list of wrappers once, instead of multiple times, in
order to find out which functions use which enums. This speeds up enum
generation tremendously.
GetBoolean, GetInteger6, GetFixedvOES and Delete* are now matched in
the convenience wrapper generator. Methods returning vectors of fixed
size (e.g. 4 ints) are no longer matched.
The lookup for function overrides and overloads now tries to work
around extension case mismatches (e.g. IBM vs Ibm). This fixes a few
specific cases of missing overrides.
It is now possible to specify multiple overloads for the same function.
This is helpful for maintaining backwards compatibility with previous
releases.
SDL does not currently support embedding into Windows Forms (this is an
upstream limitation.) To ensure that existing WinForms applications
continue to function even if SDL is installed, GLControl will now try to
initialize OpenTK with a native backend. The user can still override
this behavior using OpenTK.Toolkit.Init(ToolkitOptions), as normal.
OpenTK.Toolkit will now initialize OpenTK.Configuration and
OpenTK.Platform.Factory explicitly. It can also receive an optional
ToolkitOptions parameter to influence the OpenTK.Platform implementation
that will be chosen. Finally, it explicitly implements IDisposable to
clean up after itself.
This significantly cleans up the startup sequence on all platforms:
- X11 is not detected on non-Linux platforms unless the user explicitly
requests it
- Supports selection of platform abstractions (SDL) vs native
implementations.
- Returns correct flags on Android and iOS.
This contains a semantic change: OpenTK.Configuration will not return
correct values until OpenTK.Toolkit.Init() has been called, either
directly or indirectly (e.g. by creating a window.)
By mistake, this code would always create a desktop context. The correct
approach is to create an embedded (EGL) context and only fallback to
desktop if that doesn't work.
OpenTK 1.0 and Xamarin Android/iOS do not use strongly-typed enums for
OpenGL ES. Generate overloads with the "All" enum in order to maintain
compatibility.
Only Get*, Gen*, Delete* and New* functions get convenience overloads.
This avoids issues with functions such as Rect() that have similar
signatures but cannot use such overloads.
This restriction will be relaxed in the future.
These are convenience parameters for function receiving a size and an
array parameter, like DeleteTextures(int n, int[] ids). The generator
will now add overloads taking a single parameter, such as
DeleteTexture(int id).
CreateCLSCompliantWrappers must always change return types into
cls-compliant types. The reason is that we cannot overload on return
type alone, so we should always choose the compliant version.
Functions returning a value or array via an 'out' parameter will now get a convenience overload that returns the result via a return statement. In the case of arrays, only single-valued arrays will be supported. For example:
void GetIntegerv(enum pname, out int value)
will be get an overload of
int GetIntegerv(enum pname)
This will reduce the amount of helper overloads that must be maintained manually in GLHelpers.cs.
New WrapperTypes for convenience functions: ConvenienceReturnType to
replace an "out" parameter by a return value, and
ConvenienceArrayReturnType to replace an out array parameter by a
single return value (array count of 1 only).
It is now possible to pass a non-realized GraphicsMode to the X11GLContext and X11GLNative constructors. A non-realized GraphicsMode is a GraphicsMode with a null Index (i.e. which has not passed through SelectGraphicsMode()).
According to the Linux OpenGL ABI, glXGetProcAddressARB must be statically exported by libGL. This does *not* hold true for glXGetProcAddress. We must used the ARB version instead.
Furthermore, glx entry points, unlike wgl, do not depend on any specific OpenGL context. This means we can load them in the constructor of the Glx class.
EnableCap.ColorArray is part of the client state and must be enabled
with GL.EnableClientState, not GL.Enalbe. This is a potential fix for
http://www.opentk.com/node/3430 "Picking example's problem"
WinGraphicsMode no longer creates a temporary context in order to create
the list of available modes. Instead, it requires to be passed an
existing context in its constructor.
WinGLContext now creates one temporary context in its static constructor
and hands that to WinGraphicsMode.
WinFactory no longer supports the CreateGraphicsMode API. This API will
be removed in the future, because the link because contexts and modes
cannot be separated in the general case.
The static initializer of GetCurrentContext would always default to the
desktop PlatformFactory, which caused problems when trying to run OpenGL
ES code on the desktop. The initializer is now removed and
GetCurrentContext is set in the context constructor, before creating any
contexts.
glXGetProcAddress may return a non-null value even if a function is not
supported by the server. We need to check the extension string using
any GLX extensions. Fixes issue http://www.opentk.com/node/3111 "GLX
extension support is not checked correctly".
Furthermore, mode selection is now performed explicitly by the
X11GLContext constructor.
GLX entry points are not bound to a specific context. This means that, unlike WGL, GLX does not require a temporary context in order to load its entry points!
The IGraphicsMode interface is gradually being removed and the
MacOSFactory will now throw an exception if an instance is requested.
AglContext no longer duplicates MacOSGraphicsMode functionality.
Calling SDL_GL_GetAttribute when context construction has failed leads
to erroneous behavior. This call should only be made when a context has
been constructed correctly.
We should only generate delegates for actual OpenGL entry points, not
for overloaded functions that resolve to the same entry point. This
improves loading speeds and reduces the size of the compiled dll.
SDL_GL_GetCurrentContext will allow us to replace the weird
implementation in Sdl2Factory.CreateGetCurrentGraphicsContext()
and the latter to disable mouse emulation if we wish to.
Conflicts:
Source/OpenTK/Platform/SDL2/Sdl2.cs
WinGraphicsMode no longer creates a temporary context in order to create
the list of available modes. Instead, it requires to be passed an
existing context in its constructor.
WinGLContext now creates one temporary context in its static constructor
and hands that to WinGraphicsMode.
WinFactory no longer supports the CreateGraphicsMode API. This API will
be removed in the future, because the link because contexts and modes
cannot be separated in the general case.
The static initializer of GetCurrentContext would always default to the
desktop PlatformFactory, which caused problems when trying to run OpenGL
ES code on the desktop. The initializer is now removed and
GetCurrentContext is set in the context constructor, before creating any
contexts.
GraphicsMode.Default used to be set to
(DisplayDevice.Default.BitsPerPixel, 16, 0, 0, 0, 2, false) for improved
compatibility with older systems. However, this appears to be causing
issues with specific modern GPUs. Switch the default mode to (32, 24, 8)
until a more proper solution can be found.
Paths that don't define a "version" attribute will now match all
possible versions. This will make it easier to add support for newer
APIs as they are introduced.
After the previous commit, several ES 2.0 functions would refer to
enums found in ES 3.0. These enums have been copied to ES 2.0, either
as core enums or as extensions.
ES 3.0 includes ES 2.0 verbatim. We can significantly reduce
duplication by using the same <replace> node for both APIs. Note that
the enumerations must remain separate, as ES 2.0 and 3.0 support
different tokens.
ContextAttribute.DOUBLEBUFFER is a boolean in SDL (false->single
buffering, true->double buffering). We need to adjust the number of
buffers accordingly (single buffering->1 buffer, double buffering->2
buffers).
The issue is that some display devices report a BitsPerPel value of 0.
It is not clear whether this is a bug in WinDisplayDevice.cs or some
strange windows issue. The implemented workaround adds an entry to the
debug log and hardcodes BitsPerPel to 32 whenever this condition is
encountered. More investigation required.
The <overload> element simplifies the addition of overloads for
backwards compatibility. It is defined similar to the <replace>
element, but instead of replacing the parameters of a function
in-place, it adds a new overload and modifies the overload instead.
Sometimes an enum may reuse the tokens of another enum verbatim
(possibly adding a few extra tokens.) The reuse directive simplifies
the handling of this case:
<enum name="Foo">
<reuse enum="Bar" />
</enum>
Added strongly-typed enums for sections: Vertices, Shaders and
Programs. Added a number of missing enums for ES 2.0. Normalized
several APIs between OpenGL and OpenGL ES.
When a Delegate is defined multiple times in the spec, we should check
if any of these definitions contains a proper Delegate.Version and
store that. This improves the self-documentation aspect of the bindings.
Signatures were split into 4 APIs before (gl, glcore, gles1, gles2).
However, gles2 contains bindings for both version 2.0 and 3.0. The
version information is now maintained, which allows us to cut down on
the number of generated enumerations.
Now that we support function overloads, it is safe to ignore functions
that are defined multiple times. We just merge their Category
properties if they are not identical.
It is now possible to define multiple overloads of the same function,
each with different parameters. This is extremely useful for
maintaining backwards compatibility in the face of the changes between
GL 4.3 and 4.4.
Tokens are now sorted according to their value, exactly is in the
upstream gl.xml spec. Additionally, gles1 and gles2 now include the
pre-defined groups in the specification.
Sometimes elements in overrides.xml contain extra spaces due to typos,
which are quite difficult to track down. The XmlSpecReader can now
cope with that.
OpenTK does not currently support generating bindings for pointers of
order 3 or higher. No OpenGL or OpenGL ES API currently uses such
pointers, so we just issue a warning message if such an API is
encountered in the future.
The registry reuses the "gles2" apiname for both OpenGL ES 2.0 and 3.0.
The generator will now use the apiversion attribute to distinguish
between the two APIs.
As a sideeffect of the group element definition, the parser will
convert a ptype of "const GLubyte *" with a group of "String" to
"String *", which is not the correct result. GLXmlParser will now
detect and fix this condition that affects the GetString function of
families.
GetOverridesPath encapsulates the code that retrieves an override enum
or function from overrides.xml. Additionally, it now supports names and
extensions with multiple values (e.g. extension="Core|Ext").
GLXmlParser now explicitly marks its output with version="2", while
the older .spec parsers produce version="1" signatures. The binding
generator uses the newer API to support multiple apinames in the same
file.
Explicit Settings objects are now created for each generator. A new
-mode:all option has been added, which generates bindings for all APIs
in one go (useful for the post-GL4.4 specs, which collect all APIs in a
single file.)
ISpecWriter implementations must now store explicit references to a
Settings object. Additionally, all code generation is now handled inside
the ISpecWriter implementation (it used to be scattered over all Type,
Parameter, etc classes resulting in unmaintainable spaghetti code.)
IBind implementations must now store explicit references to Settings,
GLTypes and CSTypes instances. This allows us to use multiple
configurations in the same process.
FuncProcessor now implements all translation logic for types, parameters
and functions. This used to be scattered in the various classes (Type,
Parameter, etc) resulting in a rather disgusting spaghetti. Code
generation has been removed from FuncProcessor and is now handled by the
various language-specific ISpecWriter backends.
This is part of a long-due series of source cleanup patches. ToString()
is no longer used for code generation (code generation is handled by an
ISpecWriter instance.) Enum is no longer public. EnumCollection now
exposes its backing store through an interface, rather than a concrete
class (simpler to change backing store in the future.)
This is part of a long-due series of source cleanup patches. ToString()
is no longer used for code generation (this is handled by an ISpecWriter
implementation). This class is no longer public.
This is part of a long-due series of source cleanup patches. All
translation logic is now part of FuncProcessor. Code generation is now
handled by an ISpecWriter implementation. Minor improvements to
IEquatable and ToString() implementations. Reduced the surface of the
public API.
This is part of a long-due cleanup patch series. All translation logic
is now part of the FuncProcessor. Language-specific code generation is
now part of the ISpecWriter, not the delegate class. Implemented the
IEquatable interface.
This is part of a long-due source cleanup series. All Translate*()
methods are now part of the FuncProcessor. Additionally, ToString() has
been improved and the IEquatable interface is now implemented.
ParameterCollection now has better control of when its cache should be
rebuilt.
This is part of a long-due source cleanup operation. GLTypes and CSTypes
are no longer global singletons, but must now be accessed through an
IBind instance. All Translate*() methods are now part of the
FuncProcessor. ToString() has been improved and the IEquatable interface
is now explicitly implemented.
Settings, GLTypes and CSTypes are no longer global singletons. Instead,
explicit instances must be stored in the IBind implementation. This
allows us to use multiple configurations in the same process.
Utilities.Keywords now accepts an explicit parameter, instead of
accessing Settings directly. This allows us to use multiple
configurations in the same process.
Additionally, the Utilities class is no longer public (the Bind project
is not meant to be consumed as a dll.)
A Settings object is now passed directly as a parameter, in order to
support different configurations in the same process. The XmlSpecReader
can now distinguish between, and support both, pre-GL4.4 and post-GL4.4
specs.
Deprecated elements are marked as such in the compatibility profile and
are completely absent from the core profile. This is in-line with the
new glcore headers from Khronos.
A function redefinition usually stems from a bug in the OpenGL specs, or
a bug in the spec converter. The binding generator now logs a warning,
instead of crashing, when a redefinition is encountered.
Use a stable sort for generated elements, in order to make the generated
file work better with version control. Extension names are now extracted
directly from function names - solves issues with functions that are
defined in multiple extensions.
GL4.4 renamed PixelInternalFormat to InternalFormat. The missing tokens
are now re-added, since a lot of functions depend on them. Additionally,
a number of 'open' <use> elements now define a specific enum reference.
This is not necessary, but is good for documentation purposes.
Instead of recursing, we use a simple do..while loop to resolve the
transitive reference of a constant. If there is a loop at any point, we
stop and use a brute force search over all tokens. If this still fails
to resolve the reference, then we report this reference as unresolved.
