Chromium's Java Toolchain

This doc aims to describe the Chrome build process that takes a set of .java files and turns them into a classes.dex file.

Core GN Target Types

The following have supports_android and requires_android set to false by default:

• java_library(): Compiles .java -> .jar
• java_prebuilt(): Imports a prebuilt .jar file.

The following have supports_android and requires_android set to true. They also have a default jar_excluded_patterns set (more on that later):

• android_library()
• android_java_prebuilt()

All target names must end with “_java” so that the build system can distinguish them from non-java targets (or other variations).

Most targets produce two separate .jar files:

• Device .jar: Used to produce .dex.jar, which is used on-device.
• Host .jar: For use on the host machine (junit_binary / java_binary).
  • Host .jar files live in lib.java/ so that they are archived in builder/tester bots (which do not archive obj/).

From Source to Final Dex

Step 1: Create interface .jar with turbine or ijar

For prebuilt .jar files, use //third_party/ijar to create interface .jar from prebuilt .jar.

For non-prebuilt targets, use //third_party/turbine to create interface .jar from .java source files. Turbine is much faster than javac, and so enables full compilation to happen more concurrently.

What are interface jars?:

• The contain .class files with all non-public symbols and function bodies removed.
• Dependant targets use interface .jar files to skip having to be rebuilt when only private implementation details change.

Step 2a: Compile with javac

This step is the only step that does not apply to prebuilt targets.

• All .java files in a target are compiled by javac into .class files.
  • This includes .java files that live within .srcjar files, referenced through srcjar_deps.
• The classpath used when compiling a target is comprised of .jar files of its deps.
  • When deps are library targets, the Step 1 .jar file is used.
  • When deps are prebuilt targets, the original .jar file is used.
  • All .jar processing done in subsequent steps does not impact compilation classpath.
• .class files are zipped into an output .jar file.
• There is no support for incremental compilation at this level.
  • If one source file changes within a library, then the entire library is recompiled.
  • Prefer smaller targets to avoid slow compiles.

Step 2b: Compile with ErrorProne

This step can be disabled via GN arg: use_errorprone_java_compiler = false

• Concurrently with step 1a: ErrorProne compiles java files and checks for bug patterns, including some custom to Chromium.
• ErrorProne used to replace step 1a, but was changed to a concurrent step after being identified as being slower.

Step 3: Desugaring (Device .jar Only)

This step happens only when targets have supports_android = true. It is not applied to .jar files used by junit_binary.

• //third_party/bazel/desugar converts certain Java 8 constructs, such as lambdas and default interface methods, into constructs that are compatible with Java 7.

Step 4: Instrumenting (Device .jar Only)

This step happens only when this GN arg is set: use_jacoco_coverage = true

• Jacoco adds instrumentation hooks to methods.

Step 5: Filtering

This step happens only when targets that have jar_excluded_patterns or jar_included_patterns set (e.g. all android_ targets).

• Remove .class files that match the filters from the .jar. These .class files are generally those that are re-created with different implementations further on in the build process.
  • E.g.: R.class files - a part of Android Resources.
  • E.g.: GEN_JNI.class - a part of our JNI glue.
  • E.g.: AppHooksImpl.class - how chrome_java wires up different implementations for non-public builds.

Step 6: Per-Library Dexing

This step happens only when targets have supports_android = true.

• d8 converts .jar files containing .class files into .dex.jar files containing classes.dex files.
• Dexing is incremental - it will reuse dex'ed classes from a previous build if the corresponding .class file is unchanged.
• These per-library .dex.jar files are used directly by incremental install, and are inputs to the Apk step when enable_proguard = false.
  • Even when is_java_debug = false, many apk targets do not enable ProGuard (e.g. unit tests).

Step 7: Apk / Bundle Module Compile

• Each android_apk and android_bundle_module template has a nested java_library target. The nested library includes final copies of files stripped out by prior filtering steps. These files include:
  • Final R.java files, created by compile_resources.py.
  • Final GEN_JNI.java for JNI glue.
  • BuildConfig.java and NativeLibraries.java (//base dependencies).

Step 8: Final Dexing

This step is skipped when building using Incremental Install.

When is_java_debug = true:

• d8 merges all library .dex.jar files into a final .mergeddex.jar.

When is_java_debug = false:

• R8 performs whole-program optimization on all library lib.java .jar files and outputs a final .r8dex.jar.
  • For App Bundles, R8 creates a .r8dex.jar for each module.

Test APKs with apk_under_test

Test APKs are normal APKs that contain an <instrumentation> tag within their AndroidManifest.xml. If this tag specifies an android:targetPackage different from itself, then Android will add that package‘s classes.dex to the test APK’s Java classpath when run. In GN, you can enable this behavior using the apk_under_test parameter on instrumentation_test_apk targets. Using it is discouraged if APKs have proguard_enabled=true.

Difference in Final Dex

When enable_proguard=false:

• Any library depended on by the test APK that is also depended on by the apk-under-test is excluded from the test APK's final dex step.

When enable_proguard=true:

• Test APKs cannot make use of the apk-under-test‘s dex because only symbols explicitly kept by -keep directives are guaranteed to exist after ProGuarding. As a work-around, test APKs include all of the apk-under-test’s libraries directly in its own final dex such that the under-test apk‘s Java code is never used (because it is entirely shadowed by the test apk’s dex).
  • We've found this configuration to be fragile, and are trying to move away from it.

Difference in GEN_JNI.java

• Calling native methods using JNI glue requires that a GEN_JNI.java class be generated that contains all native methods for an APK. There cannot be conflicting GEN_JNI classes in both the test apk and the apk-under-test, so only the apk-under-test has one generated for it. As a result this, instrumentation test APKs that use apk-under-test cannot use native methods that aren't already part of the apk-under-test.

How to Generate Java Source Code

There are two ways to go about generating source files: Annotation Processors and custom build steps.

Annotation Processors

• These are run by javac as part of the compile step.
• They cannot modify the source files that they apply to. They can only generate new sources.
• Use these when:
  • an existing Annotation Processor does what you want (E.g. Dagger, AutoService, etc.), or
  • you need to understand Java types to do generation.

Custom Build Steps

• These use discrete build actions to generate source files.
  • Some generate .java directly, but most generate a zip file of sources (called a .srcjar) to simplify the number of inputs / outputs.
• Examples of existing templates:
  • jinja_template: Generates source files using Jinja.
  • java_cpp_template: Generates source files using the C preprocessor.
  • java_cpp_enum: Generates @IntDefs based on enums within .h files.
  • java_cpp_strings: Generates String constants based on strings defined in .cc files.
• Custom build steps are preferred over Annotation Processors because they are generally easier to understand, and can run in parallel with other steps (rather than being tied to compiles).

Static Analysis & Code Checks

We use several tools for static analysis.

ErrorProne

• Runs as part of normal compilation. Controlled by GN arg: use_errorprone_java_compiler.
• Most useful check:
  • Enforcement of @GuardedBy annotations.
• List of enabled / disabled checks exists within javac.py
  • Many checks are currently disabled because there is work involved in fixing violations they introduce. Please help!
• Custom checks for Chrome:
  • //tools/android/errorprone_plugin/src/org/chromium/tools/errorprone/plugin/
• Use ErrorProne checks when you need something more sophisticated than pattern matching.
• Checks run on the entire codebase, not only on changed lines.
• Does not run when chromium_code = false (e.g. for //third_party).

Android Lint

• Runs as part of normal compilation. Controlled by GN arg: disable_android_lint
• Most useful check:
  • Enforcing @TargetApi annotations (ensure you don't call a function that does not exist on all versions of Android unless guarded by an version check).
• List of disabled checks:
  • //build/android/lint/suppressions.xml
• Custom lint checks are possible, but we don't have any.
• Checks run on the entire codebase, not only on changed lines.
• Does not run when chromium_code = false (e.g. for //third_party).

Bytecode Processor

• Performs a single check:
  • That target deps are not missing any entries.
  • In other words: Enforces that targets do not rely on indirect dependencies to populate their classpath.
• Checks run on the entire codebase, not only on changed lines.

PRESUBMIT.py:

• Checks for banned patterns via _BANNED_JAVA_FUNCTIONS.
  • (These should likely be moved to checkstyle).
• Checks for a random set of things in ChecksAndroidSpecificOnUpload().
  • Including running Checkstyle.
  • (Some of these other checks should likely also be moved to checkstyle).
• Checks run only on changed lines.

Checkstyle

• Checks Java style rules that are not covered by clang-format.
  • E.g.: Unused imports and naming conventions.
• Allows custom checks to be added via XML. Here is ours.
• Preferred over adding checks directly in PRESUBMIT.py because the tool understands @SuppressWarnings annotations.
• Checks run only on changed lines.

clang-format

• Formats .java files via git cl format.
• Can be toggle on/off with code comments.

  // clang-format off
  ... non-formatted code here ...
  // clang-format on
  

• Does not work great for multiple annotations or on some lambda expressions, but is generally agreed it is better than not having it at all.