Hi Andrew, Yes, I can help with that!
-- Josh Tynjala Bowler Hat LLC <https://bowlerhat.dev> On Mon, Nov 9, 2020 at 3:22 PM Andrew Wetmore <cottag...@gmail.com> wrote: > Josh, this is very interesting. I would like to include an actionable > amount of this information in our user documentation. If I create a page in > the help docs for it, can you help me populate instructions based on your > researchs? > > Thanks! > > Andrew > > On Mon, Nov 9, 2020 at 6:16 PM Josh Tynjala <joshtynj...@bowlerhat.dev> > wrote: > > > Hi all, > > > > Some of you have probably been wondering about my changes to the compiler > > over the last year or more. I apologize again for occasionally breaking > > things for short periods. It's been quite a challenge getting this stuff > > working, but I'm excited to finally be able to report some real > > improvements that pretty much anyone should be able to take advantage of > > when building a Royale app. > > > > First some background. A while back, Harbs asked me to look into ways of > > reducing the file size of release builds. As you may know, we use > Google's > > Closure compiler to optimize our generated JavaScript. Closure can be > very > > aggressive in its optimizations, by renaming symbols (things like > variable > > and function names) and removing "dead code" that is detected as never > > being called. > > > > Closure's optimizations are good, but they also require developers to be > > very careful about how they write their JavaScript code. When you enable > > Closure's full optimizations, you are not allowed to use certain > JavaScript > > features because Closure cannot analyze them properly. For instance, > > consider the following code: > > > > var propName= "myProp"; > > var value = obj[propName]; > > > > When you dynamically access a property with a string, Closure cannot > > reliably know that the property exists and will be accessed at runtime. > It > > may decide to rename or remove that property, which would break things at > > runtime. > > > > ActionScript supports many of the same restricted dynamic features too, > so > > if you want to support the entire AS3 language, we can't let Closure do > its > > full optimization. Luckily, Closure also provides a bit of a backdoor: it > > allows you to "export" symbols, which means that they won't be renamed > and > > they won't be removed as dead code. Traditionally, we have made heavy use > > of this exporting feature in Royale. > > > > Harbs wanted to know if we absolutely needed to export everything that we > > currently export, and if we could potentially allow developers to turn > off > > exporting entirely, as long as they follow the stricter rules required by > > Closure. > > > > I won't go into all of the details, but over the last several months, > I've > > been changing the compiler to give developers more control over release > > builds. In particular, control over which symbols get exported, but also > > the ability to block Closure from renaming symbols that haven't been > > exported. > > > > Now, for some of the results. I'm going to share the output file size of > > the release build for several Royale projects with various different > > compiler options. > > > > For the example projects included with Royale, I built royale-asjs commit > > 94f12ed0e564b0b443834400dc2fc06d61b90a8a from October 26, 2020. If you > want > > to try building these examples yourself, the file sizes of release builds > > may be slightly different, if you use a different commit. > > > > SpectrumBrowser is a project developed by Harbs and his team. I used > commit > > d25a3def972b15ec029ae838f1a8a677d2d158bd from October 20 for the results > > below. Repo: https://github.com/unhurdle/spectrum-royale/ > > > > To establish a baseline, I built all of these projects with the older > > Royale 0.9.7 compiler first. > > > > ========== > > Baseline: royale-compiler 0.9.7 > > ========== > > > > HelloWorld: 68 KB > > ASDoc: 231 KB > > TourDeJewel: 1074 KB > > SpectrumBrowser: 900 KB > > > > Again, I am building the same AS3/MXML code every time, but these first > > numbers are from building with the older compiler. All apps build and run > > successfully. > > > > ----- > > > > The rest of the results are built with royale-compiler commit > > df8bd9f686f1bbf89539e545377b2797c646172c from November 3. > > > > All results below include the difference in KB and %. These values are > > always in comparison to the baseline numbers above. > > > > ========== > > Result 1: 0.9.8 default options > > ========== > > > > HelloWorld: 84 KB (+10 KB / +24%) > > ASDoc: 254 KB (+23 KB / +10%) > > TourDeJewel: 1105 KB (+31 KB / +3%) > > SpectrumBrowser: 936 KB (+36 KB / +4%) > > > > These examples are slightly larger when built with the newer compiler. > > That's expected. It's not ideal, but in the process of testing a > multitude > > of things to be sure that nothing had broken after my compiler changes, I > > discovered some cases where exporting a symbol didn't actually work > > correctly in 0.9.7! To properly fix the bug and export these symbols, > there > > was no choice but to make the file size a bit larger. > > > > ========== > > Result 2: Disable export > > ========== > > > > HelloWorld: 74 KB (+6 KB / +9%) > > ASDoc: 227 KB (-4 KB / -2%) > > TourDeJewel: 942 KB (-132 KB / -12%) > > SpectrumBrowser: 882 KB (-18 KB / -2%) > > > > In this round, I added the *-export-public-symbols=false* compiler > option. > > You may recall that I said earlier that I also modified the compiler to > > allow a symbol not to be exported, but still prevent it from being > renamed. > > With that in mind, -export-public-symbols=false basically tells the > > compiler that it still can't rename things, but it is allowed to remove > > what it perceives as dead code. > > > > HelloWorld is still slightly larger than 0.9.7, but the three other > > examples are now slightly smaller than 0.9.7. > > > > Most developers should be able to safely add -export-public-symbols=false > > to their compiler options when building a Royale app. The only time that > > you might still want this exporting is if you have external JavaScript in > > your page that isn't part of your Royale app, but it needs to call > > functions/classes in your Royale app. > > > > ========== > > Result 3: Allow non-public things to be renamed > > ========== > > > > HelloWorld: 72 KB (+4 KB / +6%) > > ASDoc: 221 KB (-10 KB / -4%) > > TourDeJewel: 918 KB (-156 KB / -15%) > > SpectrumBrowser: 861 KB (-39 KB / -4%) > > > > In this round, I used the following compiler options: > > > > -export-public-symbols=false > > > > > > > *-prevent-rename-protected-symbols=false-prevent-rename-internal-symbols=false* > > > > The two new options allow Closure compiler to rename protected and > internal > > symbols. Once again, HelloWorld is still slightly larger than 0.9.7, but > > the other three examples have gotten smaller again. > > > > While -prevent-rename-public-symbols=false exists too, we cannot use it. > > The examples would not work correctly at runtime. This option would > > probably work in a pure AS3 app, but our implementation of MXML in Royale > > uses dynamic language features that Closure restricts. Unless that is > > fixed, we need to avoid renaming certain public symbols. > > > > Again, most developers should be able to add > > -prevent-rename-protected-symbols=false > > and -prevent-rename-internal-symbols=false to their Royale app's compiler > > options. You might need to prevent renaming of protected/internal symbols > > if you access them dynamically. However, in my experience, people are > much > > more likely to access public symbols dynamically. > > > > ----- > > > > ========== > > Result 4: Allow public methods to be renamed > > ========== > > > > HelloWorld: 64 KB (-4 KB / -6%) > > ASDoc: 206 KB (-25 KB / -11%) > > TourDeJewel: 881 KB (-193 KB / -18%) > > SpectrumBrowser: 828 KB (-72 KB / -8%) > > > > In this round, I used the following compiler options: > > > > -export-public-symbols=false > > -prevent-rename-protected-symbols=false > > -prevent-rename-internal-symbols=false > > > > > > > *-prevent-rename-public-static-methods=false-prevent-rename-public-instance-methods=false > > * > > > > The two new options allow Closure to rename methods that are public. Now, > > all four examples are smaller than 0.9.7, and the file size difference is > > getting even more dramatic. > > > > Once again, -prevent-rename-public-static-methods=false and > > -prevent-rename-public-instance-methods=false should be safe for most > > developers to enable when compiling their Royale app. In my experience, > > calling methods dynamically is rare. > > > > ========== > > More new compiler options > > ========== > > > > There are some additional new compiler options available, but using them > is > > likely to break most Royale apps. > > > > -prevent-rename-public-static-variables=false > > -prevent-rename-public-instance-variables=false > > -prevent-rename-public-static-accessors=false > > -prevent-rename-public-instance-accessors=false > > > > These options control whether Closure allows variables or accessors > > (getters and setters) to be renamed. There are also similarly-named > options > > for protected and internal symbols, if you want more control over those > > too, instead of using -prevent-rename-protected-symbols=false and > > -prevent-rename-internal-symbols=false. > > > > Unfortunately, renaming public variables/accessors is usually not > possible > > without breaking the app at runtime. In some apps, you might be able to > > allow public static members to be renamed. However, in my experience, > > binding to static constants is pretty common, and renaming breaks those > > bindings. > > > > ========== > > Next Steps > > ========== > > > > Ideally, I'd like to make it possible for developers to be able to tell > > Closure that it's allowed to rename all symbols, including public ones. I > > believe that we could see even more file size savings in release builds > if > > Closure works with full optimizations for all symbols. Obviously, > > ActionScript developers would be required to strictly follow Closure's > > rules, if they opt into renaming of public symbols, but that's a choice > > that they should be allowed to make. > > > > As I mentioned above, our implementation of MXML and binding uses dynamic > > access, which is not compatible with Closure's full optimizations. To > > support those optimizations, I will need to explore changes to how we > > generate JS for MXML, and how it gets parsed at runtime. > > > > We previously discussed this subject a bit in this older thread from > > January 2020: > > > > > > > https://lists.apache.org/thread.html/r843e55252e37967b71b1430a2a904719791d698f3e5e2a79de74e493%40%3Cdev.royale.apache.org%3E > > > > At the time, I tried out some ideas that we came up with while > > brainstorming, but all had various downsides that didn't make for an > > obvious winner. In the end, I decided to set further investigation aside > > and first focus on exporting/renaming stuff. Now, I'm ready to take a > > second look with a fresh perspective, and maybe we'll have some new ideas > > to try. > > > > ----- > > > > That was really long, so thank you for reading, if you made it to the > end! > > > > TL;DR: By enabling certain, new compiler options, most Royale developers > > can make their app release builds smaller. Additionally, I plan to keep > > investigating, and I expect to find more ways to reduce file size in the > > future. > > > > -- > > Josh Tynjala > > Bowler Hat LLC <https://bowlerhat.dev> > > > > > -- > Andrew Wetmore > > http://cottage14.blogspot.com/ >
