OpenFL devlog: HashLink/C compilation

I thought I’d share another update about the new features that I’ve been working on recently for the next versions of OpenFL and Lime. In today’s post, I’m going to talk a bit about expanding OpenFL’s HashLink target to support not just HashLink/JIT, as it has for a while, but also add support for HashLink/C on Windows, macOS, and Linux.

If you’re not familiar, OpenFL is an implementation of the APIs available in Adobe Flash Player (and Adobe AIR) using the Haxe programming language. Projects built with Haxe can be cross-compiled to JavaScript, C++, and many other targets — making it available on the web, desktop apps, mobile apps, game consoles, and more. No browser plugins required, and fully native C++ performance.


HashLink is a virtual machine for the Haxe programming language. As I understand it, HashLink is considered similar to Adobe AIR in many ways. It has a bytecode format with JIT and garbage collection, which is compiled from a high-level language with a mix of object-oriented and functional programming paradigms. Its APIs are designed to be cross-platform — including window management, graphics, and audio. And it can be extended by loading native code libraries and exposing them to higher-level code.

One interesting thing about HashLink is that it can compile to two different formats. As mentioned above, HashLink supports its own bytecode format, which makes compilation fast and runtime performance relatively decent for day-to-day development. However, it can also cross-compile to low-level C code, which you can compile into fully native machine code with your favorite C compiler on a variety of operating systems (there are HashLink games running on desktop, mobile phones/tablets, and even game consoles). The HL/C workflow may take a bit longer for each individual build, but runtime performance is even better than with HL/JIT. Best of all, your Haxe code will behave exactly the same either way, whether compiled to bytecode or to C.

As of Lime 8.1 and OpenFL 9.3, the HashLink target supports only HL/JIT. As I mentioned, the HL/JIT target performs relatively well and offers a nice development workflow. It’s perfectly fine to release apps using HL/JIT, if it meets your needs, so that made HL/JIT good enough for previous OpenFL contributors to implement as a starting point. After a bit of research, I figured it wouldn’t be too difficult to support HL/C as well, so I started working on it for the upcoming Lime 8.2 release.

Step 1: Haxe compilation for HL/C

From the Haxe side, switching between HL/JIT and HL/C is a minor change. The example below compiles to bytecode for HL/JIT:

haxe -hl bin/MyApp.hl -main MyApp

All you need to do is change the output file extension from .hl to .c when calling the Haxe compiler:

haxe -hl bin/MyApp.c -main MyApp

I added a new lime build hlc command to the Lime tools (or lime build hl -hlc, if you prefer), and that makes Lime generate HL/C code instead of HL/JIT bytecode.

Step 2: C compilation for HL/C

However, an extra step is required after that. Once you have generated some C code, it needs to be passed to a C compiler. Since macOS is my current primary platform, I decided to start there and see how easily I could compile HL/C with the gcc compiler included with Apple’s Xcode.

The Haxe and HashLink documentation both provide pretty good starting points for how to compile the C code. I ended up with something similar to the following command (some file system paths have been simplified a bit, to keep the full command from being overly long):

gcc -O3 -o bin/MyApp -std=c11 -I obj obj/MyApp.c lib/libhl.dylib lib/fmt.hdll lib/mysql.hdll lib/sqlite.hdll lib/ssl.hdll lib/ui.hdll lib/uv.hdll lib/lime.hdll

Basically, you run gcc with a high level of optimization, using the 2011 version of the C language, and including all of the C code and headers generated by Haxe. Additionally, the executable should load libhl.dylib (the core HashLink dynamic library), and a number of .hdll files (other dynamic libraries for HashLink).

On macOS, executables are usually wrapped in .app bundles that macOS treats as a single file, even though it’s more like a directory containing many files. Lime already generated an .app bundle for HashLink/JIT, and that required a bit of finessing to work properly with HashLink/C. The executable generated by gcc was having some trouble finding libhl.dylib and the .hdll files when they were all bundled together inside the .app bundle. I ended up using install_name_tool to force the executable to find the dynamic libraries in the same directory as itself using the @executable_name directory value in the paths. The command looks something like this:

install_name_tool -change lime.hdll @executable_path/lime.hdll bin/

Figuring out exactly which install_name_tool sub-command to use (and determining even if that was the correct approach) took a bit of trial and error. To add to the complexity, I was seeing slightly different behavior between the version of HashLink that Lime bundles, and a nightly build of HashLink that I downloaded from the source. Even though Lime bundles HashLink, we also want to support new updates or custom builds, whenever possible. In the end, I got both the bundled and custom versions working, and I was ready to jump to other operating systems.

I thought Linux would be super easy after macOS, since it would also use gcc and both operating systems are Unix-ish. However, I quickly ran into similar issues where the executable couldn’t find the dynamic libraries. After a few hours of banging my head on the desk, I decided to take a break and try Windows instead, and I’d revisit Linux later. Sometimes, when I get stuck, allowing my brain do some background processing for a few hours (or days) can yield surprising successes.

Compiling HL/C on Windows had its own issues, of course. I initially settled on requiring gcc on Windows too, even if it’s not the default C compiler for most Windows developers. I tried to support Visual Studio’s C compiler, but it was failing due to a Haxe compiler issue that I needed to report (more on that in a second).

Anyway, it’s relatively easy to get gcc for Windows thanks to the MinGW-w64 project. I ended up installing it quickly in a terminal with the Chocolatey package manager.

choco install mingw

You could probably also get it with the Scoop package manager, or download the MinGW executables/installer directly from the source.

The gcc command for Windows was very similar. As far as differences go, I mainly needed to set the “windows” subsystem instead of the default “console” subsystem so that an extra terminal didn’t open when double-clicking the .exe file, and I needed to add dbghelp.dll as a dependency, for some reason (if you understand why, I’d be curious for an explanation or workaround). Thankfully, I didn’t need any extra commands similar to install_name_tool on Windows to get it to find the .hdll files.

gcc -O3 -o bin/MyApp -std=c11 -Wl,-subsystem,windows -I obj obj/MyApp.c C:/Windows/System32/dbghelp.dll lib/libhl.dll lib/fmt.hdll lib/mysql.hdll lib/sqlite.hdll lib/ssl.hdll lib/ui.hdll lib/uv.hdll lib/lime.hdll

I was actually able to get a cl.exe (Visual Studio’s C compiler) command working locally, if I did a find/replace for _restrict (which is generated from OpenFL’s TextField.restrict property) on the output C code. It seems that cl.exe treats some things differently than gcc. I submitted a bug report to the Haxe compiler (which has been partially addressed, which I’ll talk about in a second), and I declared MinGW support on Windows good enough to start with.

After I got Windows with MinGW working, I went back to Linux. One morning, I recalled reading about something called rpath, which is used to find dynamic libraries on both macOS and Linux. However, since I hadn’t used it on macOS, I didn’t think much of it when I saw it mentioned with Linux too. This time around, I decided to use rpath as my starting point for web searches, and eventually, I found some random example that mentioned an $ORIGIN value that could be passed to the linker options that worked similarly to @executable_name on macOS.

I also had to specify the libraries a bit differently using -L for the parent directory and listing the .so and .hdll files by name instead of full paths.

gcc -O3 -o bin/MyApp -std=c11 -Wl,-rpath,$ORIGIN -I obj obj/MyApp.c -L lib libhl.dylib fmt.hdll mysql.hdll sqlite.hdll ssl.hdll ui.hdll uv.hdll lime.hdll

I tried doing the same thing with rpath and -L on macOS, to make the commands more consistent, but apparently, gcc (or something else in the operating system) works differently with dynamic libraries, so the macOS command(s) had to remain different from Linux. Not a big deal, but I just wish they were more similar.

At this point, I had a few extra moments to spare, and I realized that I could probably allow HL/C code to be compiled with clang instead of gcc too. It turns out that I was right, and the same command line options worked without changes. Only the command name changed from gcc to clang. At least on macOS and Linux. I ran into some strange issues with certain Windows APIs not being recognized, for some reason, and I honestly don’t know enough about clang (or C development in general) to know how to fix that. Perhaps someone else can contribute that, and we’ll have a full range of compilers supported on all platforms.

There was one other detail that came up. On Apple Silicon CPUs for macOS, gcc and clang default to compiling for Apple Silicon, but currently, Lime’s and HashLink’s .hdll libraries are strictly for Intel CPUs. So I needed to precede the gcc command with arch command to ensure that it compiles for Intel to be compatible with the libraries.

arch -x86_64 gcc [options]

Lime and HashLink for Intel both work well on Apple Silicon with Rosetta, by the way. However, I don’t doubt that we’ll revisit full Apple Silicon support in the future.

After I submitted the bug report about certain HL/C code failing to compile with Visual Studio, the Haxe compiler team fixed the issue right away (at least, partially) and released Haxe 4.3.3 soon after. Most vanilla OpenFL apps should compile with this specific Haxe version, but there are some edge cases affecting Feathers UI that will require another small update to Haxe (or some renamed variables in Feathers UI).

With the other C compilers working on all platforms, I went back to trying to get cl.exe from Visual Studio to compile. I decided that the first step would be to get the cl.exe command working in Visual Studio’s special developer console, which automatically sets up various environment variables for you. I knew I’d eventually want to be able to build OpenFL apps from any random Windows terminal window, but I could look into that later. One step at a time.

I ended up with a command that looks something like this:

cl.exe /Ox /Fe:bin/MyApp.exe -I obj obj/MyApp.c -L lib libhl.lib fmt.lib mysql.lib sqlite.lib ssl.lib ui.lib uv.lib lime.lib /link /subsystem:windows

When an executable is built on Windows with Visual Studio, and it uses a .dll file (or .hdll, in the case of HashLink libraries), the .dll file is associated with a seperate .lib file. gcc and clang seem to be able to link directly to a HashLink .hdll file, but Visual Studio’s cl.exe wants the .lib file instead. HashLink is distributed with .hdll and .lib files for its core library. Lime wasn’t bundling those .lib files in 8.1 and older, but I made a minor tweak to the Lime build process to copy those next to the .hdll files that we were already bundling (the .lib files were already built, but they were simply not copied because HL/JIT doesn’t need them).

With that working, the next step was to figure out how to set up Visual Studio’s build environment in a random command prompt instead of requiring OpenFL users to open the Visual Studio developer console.

First, I had to figure out where Visual Studio was installed. I learned that Visual Studio comes with an executable called vswhere.exe, that is guaranteed to be available at the following location if Visual Studio is installed, regardless of which version of Visual Studio we’re dealing with:

C:\Program Files (x86)\Microsoft Visual Studio\Installer\vswhere.exe

Running vswhere.exe with the following options should print the location of Visual Studio:

vswhere.exe -latest -products * -requires Microsoft.VisualStudio.Component.VC.Tools.x86.x64 -property installationPath

Visual Studio includes a batch file, called vcvarsall.bat, which can be used to set the appropriate environment variables. Using the installation path returned by vswhere.exe, you just need to append VC\Auxiliary\Build\vcvarsall.bat to the end. It takes a few options, but the only required one is the archtecture. Lime uses the x86_64 architecture (sometimes abbreviated to x64) by default on Windows, so I can pass x64 for the architecture option.

vcvarsall.bat x64

Finally, I needed to combine everything together into a single command that Lime’s tools could run. This one was a little trickier than I expected, and it took me some trial and error, but I ended up launching a separate cmd.exe instance to make things work. The command looks something like this (cl.exe options omitted for brevity):

cmd.exe /s /c vcvarsall.bat x64 && cl.exe [options]

Whew! And that was the last C compiler that I wanted to get working! Now, users won’t need to manually run all of the commands above. They’ll just need to run this one, and Lime’s tooling will do all of the heavy lifting:

lime build hlc

Much better!

Where can I find the code?

If you want to try out the new hlc target for OpenFL, you’ll need to check out Lime’s 8.2.0-Dev branch on Github, or you can download both the lime-haxelib artifact from a successful Github Actions Lime 8.2.0-Dev nightly build. If you check out Lime’s source code from Github, you’ll need to manually rebuild the .hdll binary files and Lime’s command line tools, so you may find it easier to go with the pre-built artifacts. Of course, nightly builds are not necessarily ready for release to Haxelib, so use at your own risk in production. You may encounter some bugs, but we’d love any feedback you can give. Thanks, and happy coding!

OpenFL devlog: updateAfterEvent(), FileReference improvements, option to use Terser minifier

I thought I’d share another update about the new features that I’ve been working on recently for the next versions of OpenFL and Lime. In today’s post, I’m going to talk about implementing updateAfterEvent(), which makes it possible to influence OpenFL’s rendering frequency, improvements with downloading and uploading files with, and adding the Terser minifier to Lime as an option for the html5 target.

If you’re not familiar, OpenFL is an implementation of the APIs available in Adobe Flash Player (and Adobe AIR) using the Haxe programming language. Projects built with Haxe can be cross-compiled to JavaScript, C++, and many other targets — making it available on the web, desktop apps, mobile apps, and more. No browser plugins required, and fully native C++ performance.

Implementing updateAfterEvent()

In Flash, the MouseEvent, TouchEvent, KeyboardEvent, and TimerEvent classes all have a method named updateAfterEvent(). According to the API reference for Adobe AIR, this is what the method does:

Instructs Flash Player or Adobe AIR to render after processing of this event completes, if the display list has been modified.

That description may not be clear enough, so let me expand a bit. Calling updateAfterEvent() inside certain event listeners asks Flash to render more often than the requested frame rate would normally allow. Basically, this method allows developers to set the frame rate very low to improve idle performance (12 FPS was common back in the day, and it might have even been the default in the Flash authoring tool, but you could potentially go as low as 1 FPS). However, when the user is interacting with mouse/keyboard/touch, you can call updateAfterEvent() to effectively increase the frame rate temporarily so that user interaction feels very smooth.

Implementing updateAfterEvent() in OpenFL mostly involved some changes to the openfl.display.Stage class. The stage relies on lime.ui.Window to manage the requested frame rate. Every time that the Window is ready to render a new frame, it will dispatch its onRender event. The stage listens for this event, and that kicks off not just rendering all of the display objects, but also the dispatching of its own events, like Event.ENTER_FRAME, Event.FRAME_CONSTRUCTED, and Event.EXIT_FRAME that are used heavily by developers for animations and things.

OpenFL’s handling of mouse/touch/keyboard events also happens within the Stage object, so forcing a render outside of the normal frame rate is relatively convenient to implement since there is no need to communicate with other classes. However, within the Stage class, I did need to refactor a little to separate the rendering code from the code that dispatches of events like ENTER_FRAME because updateAfterEvent() should not cause MovieClip or other display objects to advance their frame playheads. It simply renders them in their existing frame again (if necessary).

I had to add one hack to make this work on native targets, like C++, Neko, and HashLink. On these targets, before Lime dispatches onRender on a window, Lime also calls window.__backend.render(). Then, after the dispatch, it calls window.__backend.contextFlip(). If these backend methods aren’t called, the state of OpenFL updates, but nothing visually changes on screen. These methods should be internal to Lime, but OpenFL currently has to break encapsulation to call them in order to force Lime to render outside of the normal frame rate. Ideally, Lime would have a public API that could force a render. This should be revisited in the future.

If you want to try out updateAfterEvent() with mouse events at a low frame rate, I should mention one thing that you’ll need to do first. OpenFL has an optimization (enabled by default) that throttles MouseEvent.MOUSE_MOVE events so that they don’t get dispatched more than once between two consecutive frames (technically, that’s not always true, but it’s close enough for our purposes). Since you’ll get MOUSE_MOVE only once between frames, updateAfterEvent() doesn’t add a lot of benefit because it can only trigger one extra render, at most. However, the person who implemented this MOUSE_MOVE optimization provided a define that will restore Flash’s original behavior that allowed multiple MOUSE_MOVE events between frames.

Add the following to your OpenFL project.xml file, and you’re good to go:

<haxedef name="openfl_always_dispatch_mouse_events"/>

OpenFL does not throttle TouchEvent or KeyboardEvent in the same way, so calling updateAfterEvent() on one of those classes will work automatically, without a special define.

FileReference improvements

The class is intended for downloading from and uploading files to a server. It conveniently provides native file system dialogs to choose where to save a downloaded file or to allow a user to select an existing file to upload.

The first improvement that I implemented was ensuring that FileReference could successfully download binary files with its download() method. Internally, FileReference uses a basic By default, the dataFormat property of URLLoader is set to URLLoaderDataFormat.TEXT, and the previous implementation did not change the dataFormat from that default value. If a binary file is loaded as a String instead of a ByteArray, data may be corrupted when attempting to save it to a file. On the other hand, a plain text file can be loaded as a ByteArray without any data loss. The fix to allow downloading binary files, while preserving the ability to download text files too, was as simple as using URLLoaderDataFormat.BINARY instead.

Until now, OpenFL’s implementation of FileReference did not provide an upload() method implementation at all. It simply threw an exception. Constructing the HTTP request to upload a file to a server is kind of complicated, so I understand why this didn’t exist yet, but my client needed it, so I needed to wade into the murky waters and figure out the multipart/form-data content type.

Most HTTP POST requests are submitted using the application/x-www-form-urlencoded content type. Parameters are sent separated by the & character with names and values separated by the = character, and certain characters are URL encoded, like this:


To submit a file in an HTTP POST request wouldn’t work well with that format, especially if it were binary instead of text, so that’s why the multipart/form-data content type is needed. It cleanly separates each parameter with more complex boundaries than the & character, and it doesn’t make any modifications to the values.

Normally, URLLoader generates the HTTP protocol request on its own. However, when using multipart/form-data, it’s necessary to do part of it manually. I ended up creating a ByteArray and writing most of the request with multiple calls to writeUTFBytes(), writing the file data with writeBytes(), and passed the resulting ByteArray to the data property of the URLLoader.

If the original URLRequest contains URLVariables, these also need to be taken into consideration. If the request’s method is set to URLRequestMethod.GET, the variables can be appended to the URL’s search query. However, if the method is URLRequestMethod.POST instead (which is most common for file uploads), the variables need to be inserted into the ByteArray that gets generated above.

Optionally using Terser to minify JS

By default Lime uses Closure Compiler from Google to minify JavaScript for the html5 target when using either the -final or -minify command line flags. Closure Compiler often generates smaller minified JavaScript than most other minifiers, so it makes good sense to use Closure Compiler as the default for Lime and OpenFL. However, there are other popular minifiers out there, and for some JavaScript code, sometimes those other minifiers are the better choice, so I thought it would be nice if Lime could provide more options.

In fact, Lime has long had the ability to use YUI Compressor instead of Closure Compiler, by adding the -yui command line flag. YUI Compressor was once one of the leading minifiers for JavaScript, but it eventually stopped being as useful because it stopped being updated with the latest JavaScript syntax, so it may fail when it sees code that it doesn’t understand. In my experience, using -yui with OpenFL is impossible because OpenFL includes some JavaScript libraries that YUI Compressor can’t deal with properly. I’m just mentioning it to show that Lime already supports multiple minifiers, so why not one more?

One of the most popular minifiers in the JavaScript ecosystem today is called Terser. Terser runs on Node.js and is included with Webpack by default, so a lot of JavaScript projects use Terser, even if they don’t realize it. Considering its popularity, in my mind, Terser was an obvious choice to include as an alternative to the default Closure Compiler.

Lime already bundled Node.js binaries to launch an HTTP server when you run the lime test html5 command. However, the included Node.js version had gotten pretty far out of date, so I updated to Node.js v18, which is the current recommended LTS release at this time. Then, I installed Terser and all of its dependencies locally, and I copied the necessary files into Lime’s repository. Inside Lime’s command line tools, I implemented a new -terser command line flag to tell Lime to use Terser instead of Closure Compiler when minifying for the html5 target. To build, run lime build html5 -minify -terser, and you’re good to go (and it also works with test instead of build)

I also added a new -npx command line flag. This tells Lime’s command line tools to use the npx tool from Node.js to run the latest versions of the google-closure-compiler or terser npm modules instead of the ones bundled with Lime. This should allow developers to take advantage of the latest minifier features (or possibly work around any bugs) without waiting for Lime to update its bundled dependencies and release an update.

Where can I find the code?

If you want to try out the new updateAfterEvent() feature, the improved FileReference class, or the Terser minifier, you’ll need to check out Lime’s 8.1.0-Dev branch and OpenFL’s 9.3.0-Dev branch on Github, or you can download both the lime-haxelib artifact from a successful Github Actions Lime 8.1.0-Dev nightly build and the openfl-haxelib artifact from a successful Github Actions OpenFL 9.3.0-Dev nightly build. If you check out Lime’s source code from Github, you’ll need to rebuild the .ndll binary files and Lime tools, so you may find it easier to go with the pre-built artifacts. Of course, nightly builds are not necessarily ready for release to Haxelib, so use at your own risk in production. You may encounter some bugs, but we’d love any feedback you can give. Thanks, and happy coding!

OpenFL devlog: NativeWindow, Lime’s Haxe and C++ bridge, and improved cleanup

Recently, as a member of OpenFL’s leadership team, I decided that I wanted to do more to keep the community up-to-date with what new features I’m working on for the next versions of OpenFL and Lime. In today’s post, I’m going to talk a bit about the upcoming implementation of the openfl.display.NativeWindow class, which is based on a core API introduced way back in Adobe AIR 1.0. The purpose of the NativeWindow class is to open a new window (with its own stage and display list) on desktop operating systems, like Windows, macOS, and Linux.

If you’re not familiar, OpenFL is an implementation of the APIs available in Adobe Flash Player (and Adobe AIR) using the Haxe programming language. Projects built with Haxe can be cross-compiled to JavaScript, C++, and many other targets — making it available on the web, desktop apps, mobile apps, and more. No browser plugins required, and fully native C++ apps.

Implementing NativeWindow

Lime, which provides OpenFL’s integration with the native operating system, has the ability to open multiple windows through its use of SDL. SDL is a C++ library that provides a cross-platform interface for things like graphics, sound, gamepads/joysticks, and more. It’s commonly used in games, but isn’t necessarily restricted to that category of program.

Lime’s Window class was actually already exposed in OpenFL, including creating a new stage for each window. However, one of my goals as a contributor is to ensure that developers adopting OpenFL can easily migrate from Flash or AIR. So I needed to wrap Lime’s API with a new class that offers the familiar interface of AIR’s NativeWindow. In the process, I discovered that I had to expose a few new APIs that Lime didn’t provide yet. And, I was able to find some ways to optimize and improve cleanup of a window and its stage (and the entire application… more on this below).

For the most part, implementing NativeWindow was pretty straightforward because SDL provides many similar APIs to what is available in AIR, and many of these were already exposed through Lime. However, not all of these SDL APIs were fully exposed yet, and there were important differences in behavior between Lime’s Window and AIR’s NativeWindow where Lime needed some tweaks to support both behaviors. For instance, by default, Lime’s Window immediately opened on creation, but AIR keeps a window hidden until the visible property is set to true, or the activate() method is called.

Lime had actually exposed a way to hide a window by default, but there wasn’t yet a property/method to reveal it later. So I needed to dive into Lime’s C++ code and expose a new window visible property to Haxe that could be changed at runtime. Similarly, AIR allows developers to optionally set minimum and maximum dimensions of a NativeWindow, and Lime hadn’t yet exposed SDL’s functions for that feature yet.

Bridging Haxe and C++ in Lime

Adding new C++ code to Lime that can be called from Haxe involves changes in several files.

  1. The .hx class that provides the public API. In the case of the new visible property, that’s Window.hx.
  2. Lime APIs are usually powered by a separate “backend” class for each target. In this case, Window.hx calls into NativeWindow.hx (to be clear, that’s lime._internal.backend.NativeWindow, which is different from openfl.display.NativeWindow).
  3. C++ backend classes call methods defined in the NativeCFFI.hx class. NativeCFFI tells Haxe how to find the C++ APIs exposed by the compiled .ndll binary (or .hdll binary for HashLink). It requires separate declarations for Neko, HashLink, and hxcpp, so it’s kind of a confusing file to work with because there’s a lot of duplication.
  4. The ExternalInterface.cpp file, which exposes the public C++ API for the .ndll and .hdll binaries. This file usually doesn’t contain the concrete implementation, but instead, is more of the glue that talks to other C++ classes that do the actual work.
  5. The .h and .cpp file where the actual C++ concrete implementation exists. In this case, that’s SDLWindow.h and SDLWindow.cpp.

Adding new Haxe APIs to Lime that are implemented in C++ isn’t actually very hard, but it is (admittedly) a complicated process because there are so many places that can be affected by a single API change.

Improved disposal/cleanup

As I was implementing OpenFL’s NativeWindow class, I took some time to consider memory management. Many OpenFL apps have one window only, including mobile and JS/web apps (yes, Lime and OpenFL still have a “window” concept when building for web browsers). With that in mind, there’s often only one Stage and display list too. However, with NativeWindow, each additional window creates a new Stage. When closing a window, it’s important to clean up that stage as well, to make sure that there aren’t any memory leaks, or worse, memory leaks plus extra code running without end — even if it isn’t needed anymore. (The first thing I checked was that Event.ENTER_FRAME wasn’t still firing out of control, but thankfully, that was already working properly.)

Ideally, when an OpenFL app exits, it should close all of its windows and remove any references to itself from static variables. Doing this manually isn’t necessary on desktop and mobile because the operating system will clean everything up automatically when the app exits. However, when targeting JS/web, if you were to manually “exit” an OpenFL application (such as by calling Lime’s System.exit() method), or if you were to manually call close() on the “window”, that doesn’t necessarily mean the browser has navigated somewhere else. The app could be embedded in a page with other content, and it may be important to be able to load a new instance of the same OpenFL app again, or even replace it with a separate OpenFL app (or something entirely different), inside the same parent HTML DOM element at a later time.

The easiest improvement was setting the static properties and openfl.desktop.NativeApplication.nativeApplication both to null on exit. It wouldn’t be possible for the application instance to be garbage collected without this change. Similarly, a number of event listeners defined by the application are added to static dispatchers that expose notifications from C++. These should be removed when the application exits too, so that they don’t prevent garbage collection.


On the JS/HTML side, “closing” the “window” should also clean up the HTML DOM. Lime creates a <canvas> element that either renders to WebGL or falls back to software rendering. It also adds some event listeners to the canvas for mouse and touch. When the window is closed, or the application exits, the canvas should be removed from the DOM, and any listeners added to DOM elements should be removed too. Again, we need things to be garbage collected, and everything returned as close to the original state as possible.

var element = parent.element;
if (element != null)
	if (canvas != null)
		if (element != cast canvas)
		canvas = null;

After all of these changes, Lime and OpenFL feel like much better HTML/JavaScript citizens, and it opens up some new possibilities for more complex websites where OpenFL is one small part of a larger system.

Where can I find the code?

If you want to try out the new NativeWindow feature, or the improved application/window/stage disposal, you’ll need to check out both Lime’s 8.1.0-Dev branch and OpenFL’s 9.3.0-Dev branch on Github, or download both the lime-haxelib artifact from a successful Github Actions Lime 8.1.0-Dev nightly build and the openfl-haxelib artifact from a successful Github Actions OpenFL 9.3.0-Dev nightly build. Of course, this code is not yet ready for release to Haxelib, so use at your own risk in production. There may still be some bugs.