Introduce post processing framework.

* Introduce CodeOptimization interface that abstracts IRCode-level
optimization. Any optimizations that need to revisit certain methods
can specify what code optimizations to conduct. One trick is, a
collection of other CodeOptimizations can act like a CodeOptimization.

IRConverter's processMethod would be the default CodeOptimization.

* Introduce PostOptimization interface that abstracts optimizations that
require post processing. Through the interface, optimizations can
specify what methods they want to revisit and what CodeOptimizations
they want to conduct. For now, call site optimizations and double
inliner are implementing it and merged, asking IRConverter's process
method to conduct.

* Introduce MethodProcessor interface that encapsulates notions of
parallel processing, call site information, and phase. With that,
nasty arguments that always appear around BlahBlah#processBlahBlah,
but most of time their values are defaultish, are gone, including:
  - isProcessedConcurrently
  - callSiteInformation
  - outliner processing
This helps us have cleaner CodeOptimization surface.

* Introduce PostMethodProcessor and its Builder. Builder is passed to
PrimaryMethodProcessor to collect methods with a single call site.
Then, it collects methods from call site optimizations and (double)
inliner.  PostMethodProcessor builds IRCode _once_ and applies code
optimizations associated with the target method. For now, it's, again,
IRconverter's full processing.

* Introduce OneTimeMethodProcessor that receives a set of methods to
process and literally processes them one-time, with the given task.
This cleans up many defaultish use of IRConverter#processMethod with
defaultish arguments.

Bug: 127694949, 140766440
Change-Id: I2ebdf409bcf2cb3cab3df9e4c5a8790bf90d063b
17 files changed
tree: 81a0abfefbafb88650db86d6a3e44e80bdb9425f
  1. buildSrc/
  2. infra/
  3. library-licensing/
  4. scripts/
  5. src/
  6. tests/
  7. third_party/
  8. tools/
  9. .gitignore
  10. AUTHORS
  11. build.gradle
  12. codereview.settings
  13. compatibility-faq.md
  14. CONTRIBUTING.md
  15. copyAdditionalJctfCommonFiles.gradle
  16. LIBRARY-LICENSE
  17. LICENSE
  18. PRESUBMIT.py
  19. README.md
  20. settings.gradle
README.md

D8 dexer and R8 shrinker

The R8 repo contains two tools:

  • D8 is a dexer that converts java byte code to dex code.
  • R8 is a java program shrinking and minification tool that converts java byte code to optimized dex code.

D8 is a replacement for the DX dexer and R8 is a replacement for the Proguard shrinking and minification tool.

Downloading and building

The R8 project uses depot_tools from the chromium project to manage dependencies. Install depot_tools and add it to your path before proceeding.

The R8 project uses Java 8 language features and requires a Java 8 compiler and runtime system.

Typical steps to download and build:

$ git clone https://r8.googlesource.com/r8
$ cd r8
$ tools/gradle.py d8 r8

The tools/gradle.py script will bootstrap using depot_tools to download a version of gradle to use for building on the first run. This will produce two jar files: build/libs/d8.jar and build/libs/r8.jar.

Running D8

The D8 dexer has a simple command-line interface with only a few options.

The most important option is whether to build in debug or release mode. Debug is the default mode and includes debugging information in the resulting dex files. Debugging information contains information about local variables used when debugging dex code. This information is not useful when shipping final Android apps to users and therefore, final builds should use the --release flag to remove this debugging information to produce smaller dex files.

Typical invocations of D8 to produce dex file(s) in the out directoy:

Debug mode build:

$ java -jar build/libs/d8.jar --output out input.jar

Release mode build:

$ java -jar build/libs/d8.jar --release --output out input.jar

The full set of D8 options can be obtained by running the command line tool with the --help option.

Running R8

R8 is a Proguard replacement for whole-program optimization, shrinking and minification. R8 uses the Proguard keep rule format for specifying the entry points for an application.

Typical invocations of R8 to produce optimized dex file(s) in the out directory:

$ java -jar build/libs/r8.jar --release --output out --pg-conf proguard.cfg input.jar

The full set of R8 options can be obtained by running the command line tool with the --help option.

Testing

Typical steps to run tests:

$ tools/test.py --no_internal

The tools/test.py script will use depot_tools to download a lot of tests and test dependencies on the first run. This includes prebuilt version of the art runtime on which to validate the produced dex code.

Contributing

In order to contribute to D8/R8 you have to sign the Contributor License Agreement. If your contribution is owned by your employer you need the Corporate Contributor License Agreement.

Once the license agreement is in place, you can upload your patches using ‘git cl’ which is available in depot_tools. Once you have a change that you are happy with you should make sure that it passes all tests and then upload the change to our code review tool using:

$ git cl upload

On your first upload you will be asked to acquire credentials. Follow the instructions given by git cl upload.

On successful uploads a link to the code review is printed in the output of the upload command. In the code review tool you can assign reviewers and mark the change ready for review. At that point the code review tool will send emails to reviewers.

Getting help

For questions, reach out to us at r8-dev@googlegroups.com.

For D8, find known issues in the D8 issue tracker or file a new D8 bug report.

For R8, find known issues in the R8 issue tracker or file a new R8 bug report.