change into a greater Android developer with Compiler Explorer

Posted by Shai Barack – Android Platform Efficiency lead

Introducing Android help in Compiler Explorer

In a earlier weblog put up you discovered how Android engineers constantly enhance the Android Runtime (ART) in ways in which increase app efficiency on consumer units. These modifications to the compiler make system and app code sooner or smaller. Builders don’t want to vary their code and rebuild their apps to learn from new optimizations, and customers get a greater expertise. On this weblog put up I’ll take you contained in the compiler with a software referred to as Compiler Explorer and witness a few of these optimizations in motion.

Compiler Explorer is an interactive web site for finding out how compilers work. It’s an open supply mission that anybody can contribute to. This 12 months, our engineers added help to Compiler Explorer for the Java and Kotlin programming languages on Android.

You should utilize Compiler Explorer to know how your supply code is translated to meeting language, and the way high-level programming language constructs in a language like Kotlin change into low-level directions that run on the processor.

At Google our engineers use this software to check completely different coding patterns for effectivity, to see how present compiler optimizations work, to share new optimization alternatives, and to show and be taught.
Studying is greatest when it’s carried out by instruments, not guidelines. As a substitute of instructing builders to memorize completely different guidelines for how you can write environment friendly code or what the compiler may or may not optimize, give the engineers the instruments to search out out for themselves what occurs once they write their code in several methods, and allow them to experiment and be taught. Let’s be taught collectively!

Begin by going to godbolt.org. By default we see C++ pattern code, so click on the dropdown that claims C++ and choose Android Java. You must see this pattern code:

class Sq. {
   static int sq.(int num) {
       return num * num;
   }
}

On the left you’ll see a quite simple program. You may say that it is a one line program. However this isn’t a significant assertion when it comes to efficiency – what number of strains of code there are doesn’t inform us how lengthy this program will take to run, or how a lot reminiscence shall be occupied by the code when this system is loaded.

On the best you’ll see a disassembly of the compiler output. That is expressed when it comes to meeting language for the goal structure, the place each line is a CPU instruction. Wanting on the directions, we will say that the implementation of the sq.(int num) methodology consists of two directions within the goal structure. The quantity and sort of directions give us a greater thought for how briskly this system is than the variety of strains of supply code. For the reason that goal structure is AArch64 aka ARM64, each instruction is 4 bytes, which implies that our program’s code occupies 8 bytes in RAM when this system is compiled and loaded.

Let’s take a quick detour and introduce some Android toolchain ideas.

The Android construct toolchain (briefly)

If you write your Android app, you’re sometimes writing supply code within the Java or Kotlin programming languages. If you construct your app in Android Studio, it’s initially compiled by a language-specific compiler into language-agnostic JVM bytecode in a .jar. Then the Android construct instruments remodel the .jar into Dalvik bytecode in .dex recordsdata, which is what the Android Runtime executes on Android units. Sometimes builders use d8 of their Debug builds, and r8 for optimized Launch builds. The .dex recordsdata go within the .apk that you simply push to check units or add to an app retailer. As soon as the .apk is put in on the consumer’s machine, an on-device compiler which is aware of the particular goal machine structure can convert the bytecode to directions for the machine’s CPU.

We are able to use Compiler Explorer to learn the way all these instruments come collectively, and to experiment with completely different inputs and see how they have an effect on the outputs.

Going again to our default view for Android Java, on the left is Java supply code and on the best is the disassembly for the on-device compiler dex2oat, the final step in our toolchain diagram. The goal structure is ARM64 as that is the most typical CPU structure in use immediately by Android units.

The ARM64 Instruction Set Structure presents many directions and extensions, however as you learn disassemblies you’ll find that you simply solely must memorize a couple of key directions. You possibly can search for ARM64 Fast Reference playing cards on-line that will help you learn disassemblies.

At Google we examine the output of dex2oat in Compiler Explorer for various causes, akin to:

• Gaining instinct for what optimizations the compiler performs so as to consider how you can write extra environment friendly code.
• Estimating how a lot reminiscence shall be required when a program with this snippet of code is loaded into reminiscence.
• Figuring out optimization alternatives within the compiler – methods to generate directions for a similar code which are extra environment friendly, leading to sooner execution or in decrease reminiscence utilization with out requiring app builders to vary and rebuild their code.
• Troubleshooting compiler bugs! 🐞

Compiler optimizations demystified

Let’s take a look at an actual instance of compiler optimizations in follow. Within the earlier weblog put up you’ll be able to examine compiler optimizations that the ART group lately added, akin to coalescing returns. Now you’ll be able to see the optimization, with Compiler Explorer!

Let’s load this instance:

class CoalescingReturnsDemo {
   String intToString(int num) {
       change (num) {
           case 1:
               return "1";
           case 2:
               return "2";
           case 3:
               return "3";           
           default:
               return "different";
       }
   }
}

How would a compiler implement this code in CPU directions? Each case could be a department goal, with a case physique that has some distinctive directions (akin to referencing the particular string) and a few frequent directions (akin to assigning the string reference to a register and returning to the caller). Coalescing returns implies that some directions on the tail of every case physique may be shared throughout all instances. The advantages develop for bigger switches, proportional to the variety of the instances.

You possibly can see the optimization in motion! Merely create two compiler home windows, one for dex2oat from the October 2022 launch (the final launch earlier than the optimization was added), and one other for dex2oat from the November 2023 launch (the primary launch after the optimization was added). You must see that earlier than the optimization, the dimensions of the tactic physique for intToString was 124 bytes. After the optimization, it’s down to simply 76 bytes.

That is in fact a contrived instance for simplicity’s sake. However this sample is quite common in Android code. As an example take into account an implementation of Handler.handleMessage(Message), the place you may implement a change assertion over the worth of Message#what.

How does the compiler implement optimizations akin to this? Compiler Explorer lets us look contained in the compiler’s pipeline of optimization passes. In a compiler window, click on Add New > Decide Pipeline. A brand new window will open, displaying the Excessive-level Inner Illustration (HIR) that the compiler makes use of for this system, and the way it’s reworked at each step.

In the event you take a look at the code_sinking move you will notice that the November 2023 compiler replaces Return HIR directions with Goto directions.

Many of the passes are hidden when Filters > Disguise Inconsequential Passes is checked. You possibly can uncheck this selection and see all optimization passes, together with ones that didn’t change the HIR (i.e. don’t have any “diff” over the HIR).

Let’s examine one other easy optimization, and look contained in the optimization pipeline to see it in motion. Take into account this code:

class ConstantFoldingDemo {
   static int demo(int num) {
       int outcome = num;
       if (num == 2) {
           outcome = num + 2;
       }
       return outcome;
   }
}

The above is functionally equal to the beneath:

class ConstantFoldingDemo {
   static int demo(int num) {
       int outcome = num;
       if (num == 2) {
           outcome = 4;
       }
       return outcome;
   }
}

Can the compiler make this optimization for us? Let’s load it in Compiler Explorer and switch to the Decide Pipeline Viewer for solutions.

The disassembly exhibits us that the compiler by no means bothers with “two plus two”, it is aware of that if num is 2 then outcome must be 4. This optimization is named fixed folding. Contained in the conditional block the place we all know that num == 2 we propagate the fixed 2 into the symbolic identify num, then fold num + 2 into the fixed 4.

You possibly can see this optimization taking place over the compiler’s IR by deciding on the constant_folding move within the Decide Pipeline Viewer.

Kotlin and Java, aspect by aspect

Now that we’ve seen the directions for Java code, strive altering the language to Android Kotlin. You must see this pattern code, the Kotlin equal of the essential Java pattern we’ve seen earlier than:

enjoyable sq.(num: Int): Int = num * num

You’ll discover that the supply code is completely different however the pattern program is functionally an identical, and so is the output from dex2oat. Discovering the sq. of a quantity leads to the identical directions, whether or not you write your supply code in Java or in Kotlin.

You possibly can take this chance to check attention-grabbing language options and uncover how they work. As an example, let’s evaluate Java String concatenation with Kotlin String interpolation.

In Java, you may write your code as follows:

class StringConcatenationDemo {
   void stringConcatenationDemo(String myVal) {
       System.out.println("The worth of myVal is " + myVal);
   }
}

Let’s learn how Java String concatenation truly works by attempting this instance in Compiler Explorer.

First you’ll discover that we modified the output compiler from dex2oat to d8. Studying Dalvik bytecode, which is the output from d8, is often simpler than studying the ARM64 directions that dex2oat outputs. It’s because Dalvik bytecode makes use of increased stage ideas. Certainly you’ll be able to see the names of sorts and strategies from the supply code on the left aspect mirrored within the bytecode on the best aspect. Attempt altering the compiler to dex2oat and again to see the distinction.

As you learn the d8 output you could notice that Java String concatenation is definitely carried out by rewriting your supply code to make use of a StringBuilder. The supply code above is rewritten internally by the Java compiler as follows:

class StringConcatenationDemo {
   void stringConcatenationDemo(String myVal) {
       StringBuilder sb = new StringBuilder();
       sb.append("The worth of myVal is ");
       sb.append(myVal);
       System.out.println(sb.toString());
  }
}

In Kotlin, we will use String interpolation:

enjoyable stringInterpolationDemo(myVal: String) {
   System.out.println("The worth of myVal is $myVal");
}

The Kotlin syntax is less complicated to learn and write, however does this comfort come at a price? In the event you do this instance in Compiler Explorer, you could discover that the Dalvik bytecode output is roughly the identical! On this case we see that Kotlin presents an improved syntax, whereas the compiler emits comparable bytecode.

At Google we examine examples of language options in Compiler Explorer to study how high-level language options are carried out in lower-level phrases, and to raised inform ourselves on the completely different tradeoffs that we’d make in selecting whether or not and how you can undertake these language options. Recall our studying precept: instruments, not guidelines. Somewhat than memorizing guidelines for a way you must write your code, use the instruments that can show you how to perceive the upsides and drawbacks of various options, after which make an knowledgeable resolution.

What occurs if you minify your app?

Talking of constructing knowledgeable choices as an app developer, you have to be minifying your apps with R8 when constructing your Launch APK. Minifying typically does three issues to optimize your app to make it smaller and sooner:

1. Useless code elimination: discover all of the stay code (code that’s reachable from well-known program entry factors), which tells us that the remaining code is just not used, and subsequently may be eliminated.

2. Bytecode optimization: varied specialised optimizations that rewrite your app’s bytecode to make it functionally an identical however sooner and/or smaller.

3. Obfuscation: renaming every type, strategies, and fields in your program that aren’t accessed by reflection (and subsequently may be safely renamed) from their names in supply code (com.instance.MyVeryLongFooFactorySingleton) to shorter names that slot in much less reminiscence (a.b.c).

Let’s see an instance of all three advantages! Begin by loading this view in Compiler Explorer.

First you’ll discover that we’re referencing sorts from the Android SDK. You are able to do this in Compiler Explorer by clicking Libraries and including Android API stubs.

Second, you’ll discover that this view has a number of supply recordsdata open. The Kotlin supply code is in instance.kt, however there’s one other file referred to as proguard.cfg.

-keep class MinifyDemo {
   public void goToSite(...);
}

Wanting inside this file, you’ll see directives within the format of Proguard configuration flags, which is the legacy format for configuring what to maintain when minifying your app. You possibly can see that we’re asking to maintain a sure methodology of MinifyDemo. “Conserving” on this context means don’t shrink (we inform the minifier that this code is stay). Let’s say we’re growing a library and we’d like to supply our buyer a prebuilt .jar the place they’ll name this methodology, so we’re preserving this as a part of our API contract.

We arrange a view that can allow us to see the advantages of minifying. On one aspect you’ll see d8, displaying the dex code with out minification, and on the opposite aspect r8, displaying the dex code with minification. By evaluating the 2 outputs, we will see minification in motion:

1. Useless code elimination: R8 eliminated all of the logging code, because it by no means executes (as DEBUG is all the time false). We eliminated not simply the calls to android.util.Log, but in addition the related strings.

2. Bytecode optimization: because the specialised strategies goToGodbolt, goToAndroidDevelopers, and goToGoogleIo simply name goToUrl with a hardcoded parameter, R8 inlined the calls to goToUrl into the decision websites in goToSite. This inlining saves us the overhead of defining a way, invoking the tactic, and getting back from the tactic.

3. Obfuscation: we instructed R8 to maintain the general public methodology goToSite, and it did. R8 additionally determined to maintain the tactic goToUrl because it’s utilized by goToSite, however you’ll discover that R8 renamed that methodology to a. This methodology’s identify is an inside implementation element, so obfuscating its identify saved us a couple of valuable bytes.

You should utilize R8 in Compiler Explorer to know how minification impacts your app, and to experiment with alternative ways to configure R8.

At Google our engineers use R8 in Compiler Explorer to check how minification works on small samples. The authoritative software for finding out how an actual app compiles is the APK Analyzer in Android Studio, as optimization is a whole-program drawback and a snippet may not seize each nuance. However iterating on launch builds of an actual app is gradual, so finding out pattern code in Compiler Explorer helps our engineers rapidly be taught and iterate.

Google engineers construct very giant apps which are utilized by billions of individuals on completely different units, in order that they care deeply about these sorts of optimizations, and attempt to take advantage of use out of optimizing instruments. However a lot of our apps are additionally very giant, and so altering the configuration and rebuilding takes a really very long time. Our engineers can now use Compiler Explorer to experiment with minification below completely different configurations and see leads to seconds, not minutes.

Chances are you’ll marvel what would occur if we modified our code to rename goToSite? Sadly our construct would break, except we additionally renamed the reference to that methodology within the Proguard flags. Fortuitously, R8 now natively helps Hold Annotations as a substitute for Proguard flags. We are able to modify our program to make use of Hold Annotations:

@UsedByReflection(type = KeepItemKind.CLASS_AND_METHODS)
public static void goToSite(Context context, String web site) {
    ...
}

Right here is the full instance. You’ll discover that we eliminated the proguard.cfg file, and below Libraries we added “R8 keep-annotations”, which is how we’re importing @UsedByReflection.

At Google our engineers favor annotations over flags. Right here we’ve seen one good thing about annotations – preserving the details about the code in a single place moderately than two makes refactors simpler. One other is that the annotations have a self-documenting facet to them. As an example if this methodology was stored often because it’s referred to as from native code, we might annotate it as @UsedByNative as an alternative.

Baseline profiles and also you

Lastly, let’s contact on baseline profiles. To date you noticed some demos the place we checked out dex code, and others the place we checked out ARM64 directions. In the event you toggle between the completely different codecs you’ll discover that the high-level dex bytecode is rather more compact than low-level CPU directions. There’s an attention-grabbing tradeoff to discover right here – whether or not, and when, to compile bytecode to CPU directions?

For any program methodology, the Android Runtime has three compilation choices:

1. Compile the tactic Simply in Time (JIT).

2. Compile the tactic Forward of Time (AOT).

3. Don’t compile the tactic in any respect, as an alternative use a bytecode interpreter.

Working code in an interpreter is an order of magnitude slower, however doesn’t incur the price of loading the illustration of the tactic as CPU directions which as we’ve seen is extra verbose. That is greatest used for “chilly” code – code that runs solely as soon as, and isn’t important to consumer interactions.

When ART detects {that a} methodology is “scorching”, it is going to be JIT-compiled if it’s not already been AOT compiled. JIT compilation accelerates execution occasions, however pays the one-time price of compilation throughout app runtime. That is the place baseline profiles are available. Utilizing baseline profiles, you because the app developer can provide ART a touch as to which strategies are going to be scorching or in any other case price compiling. ART will use that trace earlier than runtime, compiling the code AOT (often at set up time, or when the machine is idle) moderately than at runtime. This is the reason apps that use Baseline Profiles see sooner startup occasions.

With Compiler Explorer we will see Baseline Profiles in motion.

Let’s open this instance.

The Java supply code has two methodology definitions, factorial and fibonacci. This instance is about up with a guide baseline profile, listed within the file profile.prof.txt. You’ll discover that the profile solely references the factorial methodology. Consequently, the dex2oat output will solely present compiled code for factorial, whereas fibonacci exhibits within the output with no directions and a measurement of 0 bytes.

Within the context of compilation modes, which means factorial is compiled AOT, and fibonacci shall be compiled JIT or interpreted. It’s because we utilized a special compiler filter within the profile pattern. That is mirrored within the dex2oat output, which reads: “Compiler filter: speed-profile” (AOT compile solely profile code), the place earlier examples learn “Compiler filter: pace” (AOT compile every part).

Conclusion

Compiler Explorer is a good software for understanding what occurs after you write your supply code however earlier than it may run on a goal machine. The software is simple to make use of, interactive, and shareable. Compiler Explorer is greatest used with pattern code, nevertheless it goes by the identical procedures as constructing an actual app, so you’ll be able to see the impression of all steps within the toolchain.

By studying how you can use instruments like this to find how the compiler works below the hood, moderately than memorizing a bunch of guidelines of optimization greatest practices, you can also make extra knowledgeable choices.

Now that you’ve got seen how you can use the Java and Kotlin programming languages and the Android toolchain in Compiler Explorer, you’ll be able to stage up your Android improvement abilities.

Lastly, do not forget that Compiler Explorer is an open supply mission on GitHub. If there’s a function you’d prefer to see then it is only a Pull Request away.

Java and OpenJDK are logos or registered logos of Oracle and/or its associates.