IssueTracker

Regarding the first issue, that is because NoosaScriptNoLighting instances are only ever created by reflection in Script I think. R8 does not know what a reflective new instance is instantiating. Therefore, it will assume that there are no instances of NoosaScriptNoLighting and therefore, it will conclude that the shader() method call in NoosaScript can never hit the shader method in the subclass as no instance ever exist of that.

The fix is to add an explicit keep rule for anything that is only instantiated by reflection. Something like:

-keep class com.watabou.noosa.NoosaScript
-keep class com.watabou.noosa.NoosaScriptNoLighting

should do it.

Regarding the second issue, I have only validated that the reset method is actually in the output and that it does assign a new empty array to the field. Not yet sure why that doesn't take effect.

@2 Thank you, that's a good point regarding the script classes, but it's worth noting that it's the shader() method in NoosaScript that is giving incorrect behavior, not the one in NoosaScriptNoLighting. I just did a quick test and adding a pointless call to new NoosaScript() does fix the issue, adding in new NoosaScriptNoLighting() does not. Using the keep rule does fix the issue as well, of course.

There are also various other game elements that are only ever instantiated via reflection though. For example many of the game's items are only ever instantiated via reflection in Generator.java. These have always worked in prior R8 versions, and while I haven't tested them exhaustively they seem to be working with R8 2.1.52 as well.

Does R8 only optimize away members of classes created exclusively via reflection in some cases?

What is happening is that R8 determines that there is only one shader() method that can possibly be hit at runtime. Therefore, it inlines that method at the call-site. Therefore, in the resulting APK there are no shader() methods at all and the same shader string is always used. The keep rules tells R8 that both classes are instantiated, which let's R8 see that the shader() method call has two possible targets, and therefore the methods remain and different strings are used for the two classes. I hope that makes sense.

In general, if things are only instantiated via reflection you typically need some kind of keep rule to tell R8 that this class will be instantiated at runtime. Otherwise, when we add further optimizations to R8, we might end up deducing that there can be no instances of these classes and remove things that are actually needed (like leaving static methods but removing instance methods as the instance methods can never be hit if there is no instance).

The second issue definitely looks like a bug in R8. The reset call seems to disappear from Dungeon.init which is inline into InterlevelScene.descend (which again is inline into an accessibility bridge). I'll continue looking into that. Thank you for the report!

Alright, at least I know what the problem is now.

In the Dungeon.init() method R8 decides to class inline the reset call on the QuickSlot instance as it thinks the QuickSlot instance doesn't escape and therefore the side effects on the object is not visible elsewhere. That is of course incorrect as the QuickSlot instance is used directly from other classes and methods as well.

The effect is that the reset call is incorrectly removed. Stay tuned for a fix.

@6 excellent, thank you very much for looking into the reset bug, and for the explanation regarding the script classes! I'll make adjustments to my app's R8 config and will keep that in mind when testing new R8 versions.

I have similar issues on my app with R8 2.0.x

And I can confirm my issue is fixed since R8 2.1.59

Can google provide this fix on R8 2.0.x ?

The fix is now also available on the 2.0 branch, starting from R8 version 2.0.99.