Design Document: Extensibility For Native Rules

Design documents are not descriptions of the current functionality of Bazel. Always go to the documentation for current information.

Status: Reviewed, not yet implemeted

Author: Dmitry Lomov

Design document published: 04 August 2016


There is a number of requests that require Skylark API to access functionality of native rules from Skylark rules. Typical scenarios can be illustrated by the following "sandwich":

bread_library(name = "top", )
java_library(name = "meat", deps = [":top", ], )
bread_library(name = "bottom", deps = [":meat", ])

Here bread_library is a rule written in Skylark. Here we need three things:

  • implementation of bread_library should be able to produce jar files and other artifacts in the same way as native java_library rule would; in other words the implementation should be able to delegate to the a native implementation
  • java_library should allow depending on bread_library; importantly, that dependence should be meaningful, that is if bread_library produces Java artifacts, such as jar files, java_library should be able to compile against those like it would against a java_library dependency
  • bread_library should be able to depend on java_library; it should be able to access all information it needs; when delegating to native implementation, there should be a simple way to pass information from a dependency to to the native implementation

In this document we present some ideas about what how this all might look like, and suggest some practical steps we can take to get there.

Extensible Native Rules

This proposal assumes Declared Providers are implemented. Here is how the implementation of bread_library might look like:

# Implementation of a rule that transpiles to Java and invokes a native
# compilation
def _bread_library_impl(ctx):
    bread_sources = [f for src in ctx.attrs.src for f in src.files]
    generated_java_files = _invoke_bread_transpiler(ctx, bread_sources)
    # is a declared provider for Java
    java_deps = [target[] for target in ctx.attrs.deps]
    # create a native compilation action
    java_p =,
                srcs = generated_java_files,
                # information about dependencies is just a
                deps = java_deps,
    # java_p is a representing the result of compilation
    # action we return that provider and immediately java_libary rule can depend
    # on us
    return [java_p, ...]

# Implementation of a rule that compiles to JVM bytecode directly
def _scala_library_impl(ctx):
    # collect dependency jars to pass to the compile action
    dep_jars = [dep[java.lang.provider].jar for dep in ctx.attrs.deps]
    jar_output = ctx.new_file(...)
    ... construct compilation actions ...
    # build a provider that passes all transitive information
    transitive_p =
                      [dep[java.lang.provider] for dep in ctx.attrs.deps])
    java_p =
                jar = jar_output,
                # update transitive information that we care about
                transitive_jars =
                    transitive_p.transitive_jars | set(jar_output),
                    ... whatever other information is needed ...)
    # return java.lang.provider
    return [java_p, ...]

The provider is the glue ("butter") that connects Skylark rules to native rules and also to the native rule implementations exposed to Skylark. Note how the native rule implementation ( both consumes the entire providers from dependencies and returns the provider that needs to be returned from the rule. is a function that passes all the transitive information correctly from dependencies. The existing '.java' provider becomes the same thing as

Note: for the sake for this document we are placing things in There are other alternatives to this, e.g. "magical" .bzl files from which java_provider and java_compile function are exported.

How to get there

Our current native rules are not as neat as described above. Making them extensible in one go is a difficult and long term project (or rather, projects: one for each language). Here is a suggested steps for extensibility of particular language implementation (we continue to use Java as a running example).

Phase 1: Expose native compilation actions

At this step, is a black box. Skylark rules cannot construct the directly: the only way to create it is to invoke function.

Native implementations of Java rules are rewritten so that they can link to deps that return and that they return The implementation of the provider can just be a bag of all providers that Java rules normally return - since that bag is not openable by Skylark, we can refactor it later without much difficulty .

Native compilation function (java.lang.compile) is pretty much JavaLibrary.create refactored so that it gets its dependent providers not from attributes but from a list of bags. JavaLibrary.create just collects the bags from deps and passes those to that function.

At the end of this phase, implementing code generators (and code generating aspects) such as java_proto_library becomes possible. This also covers many (most?) use cases where people use macros to delegate to native rule implementations

No huge refactoring of language rule implementation is needed, but the stage is set for gradual opening up in the future.

Phase 1a: Implementing JavaSkylarkApiProvider on top of black box

(Optional) As is just a bag of existing providers, it is easy to just implement everything in '' on top of it, if desired.

Phase 2: Evolving the API and opening up

The next step in the API evolution is making the black box provider less black. This means introducing a constructor for as well as accessors to fields.

The API can be designed gradually and thoughtfully, only exposing the things we need and adding carefully: as an example sequence first just java libraries, then resources, then JNI, then support for tests. Existence of is crucial at this stage as it allows merging of transitive information from dependencies that is not yet exposed to Skylark.

As API exposure gradually progresses, the exposed Skylark API reaches parity with internal API.

Through the execution of this phase, more and more use cases are covered, and at the end the rules are fully extensible.