Documentation

Edit

Rules

Status: Experimental. We may make breaking changes to the API, but we will help you update your code.

A rule defines a series of actions that Bazel should perform on inputs to get a set of outputs. For example, a C++ binary rule might take a set of .cpp files (the inputs), run g++ on them (the action), and return an executable file (the output).

Note that, from Bazel's perspective, g++ and the standard C++ libraries are also inputs to this rule. As a rule writer, you must consider not only the user-provided inputs to a rule, but also all of the tools and libraries required to execute the actions (called implicit inputs).

Before creating or modifying any rule, make sure you are familiar with the extensibility model (understand the three phases and the differences between macros and rules).

Rule creation

In a .bzl file, use the rule function to create a new rule and store it in a global variable:

my_rule = rule(...)

See the cookbook for examples. The rule can then be loaded by BUILD files:

load('//some/pkg:whatever.bzl', 'my_rule')

A custom rule can be used just like a native rule. It has a mandatory name attribute, you can refer to it with a label, and you can see it in bazel query.

The rule is analyzed when you explicitly build it, or if it is a dependency of the build. In this case, Bazel will execute its implementation function. This function decides what the outputs of the rule are and how to build them (using actions). During the analysis phase, no external command can be executed. Instead, actions are registered and will be run in the execution phase, if their output is needed for the build.

Attributes

An attribute is a rule argument, such as srcs or deps. You must list the attributes and their types when you define a rule.

sum = rule(
    implementation = _impl,
    attrs = {
        "number": attr.int(default = 1),
        "deps": attr.label_list(),
    },
)

The following attributes are implicitly added to every rule: deprecation, features, name, tags, testonly, visibility. Test rules also have the following attributes: args, flaky, local, shard_count, size, timeout.

Labels listed in attr will be inputs to the rule.

To access an attribute in a rule's implementation, use ctx.attr.<attribute_name>. The name and the package of a rule are available with ctx.label.name and ctx.label.package.

See an example of using attr in a rule.

Private Attributes

In Python, we use one leading underscore(_) for non-public methods and instance variables (see PEP-8).

Similarly, if an attribute name starts with _ it is private and users cannot set it. It is useful in particular for label attributes (your rule will have an implicit dependency on this label).

metal_compile = rule(
    implementation = _impl,
    attrs = {
        "srcs": attr.label_list(),
        "_compiler": attr.label(
            default = Label("//tools:metalc"),
            allow_single_file = True,
            executable = True,
        ),
    },
)

Implementation function

Every rule requires an implementation function. It contains the actual logic of the rule and is executed strictly in the analysis phase. The function has exactly one input parameter, ctx, and it may return the runfiles and providers of the rule. The input parameter ctx can be used to access attribute values, outputs and dependent targets, and files. It also has some helper functions. See the library for more context. Example:

def _impl(ctx):
  ...
  return [DefaultInfo(runfiles=...),  MyInfo(...)]

my_rule = rule(
    implementation = _impl,
    ...
)

Files

There are two kinds of files: files stored in the file system and generated files. For each generated file, there must be one and only one generating action, and each action must generate one or more output files. Bazel will throw an error otherwise.

Targets

Every build rule corresponds to exactly one target. A target can create actions, can have dependencies (which can be files or other build rules), output files (generated by its actions), and providers.

A target y depends on target x if y has a label or label list type attribute where x is declared:

my_rule(
    name = "x",
)

my_rule(
    name = "y",
    deps = [":x"],
)

In the above case, it's possible to access targets declared in my_rule.deps:

def _impl(ctx):
  for dep in ctx.attr.deps:
    # Do something with dep
  ...

my_rule = rule(
    implementation = _impl,
    attrs = {
        "deps": attr.label_list(),
    },
    ...
)

Output files

A target can declare output files, which must be generated by the target's actions. There are three ways to create output files:

  • If the rule is marked executable, it creates an output file of the same name as the rule's. See example

  • The rule can declare default outputs, which are always generated. See example

  • The rule can have output or output list type attributes. In that case the output files come from the actual attribute values. See example

Each output file must have exactly one generating action. See the library for more context.

Default outputs

Every rule has a set of default outputs. This is used:

  • When the user runs bazel build on your target. Bazel will build the default outputs of the rule.

  • When the target is used as a dependency of another rule. A rule can access the default outputs by using target.files. This is the case, for example, if you use a rule in the srcs attribute of a genrule.

Use the files provider to specify the default outputs of a rule. If left unspecified, it will contain all the declared outputs.

def _impl(ctx):
  # ...
  return DefaultInfo(files=depset([file1, file2]))

This can be useful for exposing files generated with ctx.actions.declare_file. You can also have "implicit outputs", i.e., files that are declared in the rule, but not in the default outputs (like _deploy.jar in java_binary).

Actions

An action describes how to generate a set of outputs from a set of inputs, for example "run gcc on hello.c and get hello.o". When an action is created, Bazel doesn't run the command immediately. It registers it in a graph of dependencies, because an action can depend on the output of another action (e.g. in C, the linker must be called after compilation). In the execution phase, Bazel decides which actions must be run and in which order.

All functions that create actions are defined in ctx.actions:

Actions take a set (which can be empty) of input files and generate a (non-empty) set of output files. The set of input and output files must be known during the analysis phase. It might depend on the value of attributes and information from dependencies, but it cannot depend on the result of the execution. For example, if your action runs the unzip command, you must specify which files you expect to be inflated (before running unzip).

Actions are comparable to pure functions: They should depend only on the provided inputs, and avoid accessing computer information, username, clock, network, or I/O devices (except for reading inputs and writing outputs). This is important because the output will be cached and reused.

If an action generates a file that is not listed in its outputs: This is fine, but the file will be ignored and cannot be used by other rules.

If an action does not generate a file that is listed in its outputs: This is an execution error and the build will fail. This happens for instance when a compilation fails.

If an action generates an unknown number of outputs and you want to keep them all, you must group them in a single file (e.g., a zip, tar, or other archive format). This way, you will be able to deterministically declare your outputs.

If an action does not list a file it uses as an input, the action execution will most likely result in an error. The file is not guaranteed to be available to the action, so if it is there, it's due to coincidence or error.

If an action lists a file as an input, but does not use it: This is fine. However, it can affect action execution order, resulting in sub-optimal performance.

Dependencies are resolved by Bazel, which will decide which actions are executed. It is an error if there is a cycle in the dependency graph. Creating an action does not guarantee that it will be executed: It depends on whether its outputs are needed for the build.

Configurations

Imagine that you want to build a C++ binary and target a different architecture. The build can be complex and involve multiple steps. Some of the intermediate binaries, like the compilers and code generators, have to run on your machine (the host); some of the binaries such the final output must be built for the target architecture.

For this reason, Bazel has a concept of "configurations" and transitions. The topmost targets (the ones requested on the command line) are built in the "target" configuration, while tools that should run locally on the host are built in the "host" configuration. Rules may generate different actions based on the configuration, for instance to change the cpu architecture that is passed to the compiler. In some cases, the same library may be needed for different configurations. If this happens, it will be analyzed and potentially built multiple times.

By default, Bazel builds the dependencies of a target in the same configuration as the target itself, i.e. without transitioning. When a target depends on a tool, the label attribute will specify a transition to the host configuration. This causes the tool and all of its dependencies to be built for the host machine, assuming those dependencies do not themselves have transitions.

For each label attribute, you can decide whether the dependency should be built in the same configuration, or transition to the host configuration (using cfg). If a label attribute has the flag executable=True, the configuration must be set explicitly. See example

In general, sources, dependent libraries, and executables that will be needed at runtime can use the same configuration.

Tools that are executed as part of the build (e.g., compilers, code generators) should be built for the host configuration. In this case, specify cfg="host" in the attribute.

Otherwise, executables that are used at runtime (e.g. as part of a test) should be built for the target configuration. In this case, specify cfg="target" in the attribute.

The configuration "data" is present for legacy reasons and should be used for the data attributes.

Configuration Fragments

Rules may access configuration fragments such as cpp, java and jvm. However, all required fragments must be declared in order to avoid access errors:

def _impl(ctx):
    # Using ctx.fragments.cpp would lead to an error since it was not declared.
    x = ctx.fragments.java
    ...

my_rule = rule(
    implementation = _impl,
    fragments = ["java"],      # Required fragments of the target configuration
    host_fragments = ["java"], # Required fragments of the host configuration
    ...
)

ctx.fragments only provides configuration fragments for the target configuration. If you want to access fragments for the host configuration, use ctx.host_fragments instead.

Providers

Providers are pieces of information that a rule exposes to other rules that depend on it. This data can include output files, libraries, parameters to pass on a tool's command line, or anything else the depending rule should know about. Providers are the only mechanism to exchange data between rules, and can be thought of as part of a rule's public interface (loosely analogous to a function's return value).

A rule can only see the providers of its direct dependencies. If there is a rule top that depends on middle, and middle depends on bottom, then we say that middle is a direct dependency of top, while bottom is a transitive dependency of top. In this case, top can see the providers of middle. The only way for top to see any information from bottom is if middle re-exports this information in its own providers; this is how transitive information can be accumulated from all dependencies. In such cases, consider using depsets to hold the data more efficiently without excessive copying.

Providers can be declared using the provider() function:

TransitiveDataInfo = provider()

Rule implementation function can then construct and return provider instances:

def rule_implementation(ctx):
  ...
  return [TransitiveDataInfo(value = ["a", "b", "c"])]

TransitiveDataInfo acts both as a constructor for provider instances and as a key to access them. A target serves as a map from each provider that the target supports, to the target's corresponding instance of that provider. A rule can access the providers of its dependencies using the square bracket notation ([]):

def dependent_rule_implementation(ctx):
  ...
  s = depset()
  for dep_target in ctx.attr.deps:
    s += dep_target[TransitiveDataInfo].value
  ...

All targets have a DefaultInfo provider that can be used to access some information relevant to all targets.

Providers are only available during the analysis phase. Examples of usage:

Note: Historically, Bazel also supported provider instances that are identified by strings and accessed as fields on the target object instead of as keys. This style is deprecated but still supported. Return legacy providers as follows:

def rule_implementation(ctx):
  ...
  modern_provider = TransitiveDataInfo(value = ["a", "b", "c"])
  # Legacy style.
  return struct(legacy_provider = struct(...),
                another_legacy_provider = struct(...),
                # The `providers` field contains provider instances that can be accessed
                # the "modern" way.
                providers = [modern_provider])

To access legacy providers, use the dot notation. Note that the same target can define both modern and legacy providers:

def dependent_rule_implementation(ctx):
  ...
  s = depset()
  for dep_target in ctx.attr.deps:
    x = dep_target.legacy_provider            # legacy style
    s += dep_target[TransitiveDataInfo].value # modern style
  ...

We recommend using modern providers for all future code.

Runfiles

Runfiles are a set of files used by the (often executable) output of a rule during runtime (as opposed to build time, i.e. when the binary itself is generated). During the execution phase, Bazel creates a directory tree containing symlinks pointing to the runfiles. This stages the environment for the binary so it can access the runfiles during runtime.

Runfiles can be added manually during rule creation and/or collected transitively from the rule's dependencies:

def _rule_implementation(ctx):
  ...
  transitive_runfiles = depset()
  for dep in ctx.attr.special_dependencies:
     transitive_runfiles += dep.transitive_runtime_files

  runfiles = ctx.runfiles(
      # Add some files manually.
      files = [ctx.file.some_data_file],
      # Add transitive files from dependencies manually.
      transitive_files = transitive_runfiles,
      # Collect runfiles from the common locations: transitively from srcs,
      # deps and data attributes.
      collect_default = True,
  )
  # Add a field named "runfiles" to the DefaultInfo provider in order to actually
  # create the symlink tree.
  return [DefaultInfo(runfiles=runfiles)]

Note that non-executable rule outputs can also have runfiles. For example, a library might need some external files during runtime, and every dependent binary should know about them.

Also note that if an action uses an executable, the executable's runfiles can be used when the action executes.

Normally, the relative path of a file in the runfiles tree is the same as the relative path of that file in the source tree or generated output tree. If these need to be different for some reason, you can specify the root_symlinks or symlinks arguments. The root_symlinks is a dictionary mapping paths to files, where the paths are relative to the root of the runfiles directory. The symlinks dictionary is the same, but paths are implicitly prefixed with the name of the workspace.

    ...
    runfiles = ctx.runfiles(
        root_symlinks = {"some/path/here.foo": ctx.file.some_data_file2}
        symlinks = {"some/path/here.bar": ctx.file.some_data_file3}
    )
    # Creates something like:
    # sometarget.runfiles/
    #     some/
    #         path/
    #             here.foo -> some_data_file2
    #     <workspace_name>/
    #         some/
    #             path/
    #                 here.bar -> some_data_file3

If symlinks or root_symlinks is used, be careful not to map two different files to the same path in the runfiles tree. This will cause the build to fail with an error describing the conflict. To fix, you will need to modify your ctx.runfiles arguments to remove the collision. This checking will be done for any targets using your rule, as well as targets of any kind that depend on those targets.

Output groups

By default Bazel builds a target's default outputs. However, a rule can also create other outputs that are not part of a typical build but might still be useful, such as debug information files. The facility for this is output groups.

A rule can declare that a certain file belongs to a certain output group by returning the OutputGroupInfo provider. Fields of that provider are output group names:

def _impl(ctx):
  name = ...
  binary = ctx.actions.declare_file(name)
  debug_file = ctx.actions.declare_file(name + ".pdb")
  # ... add actions to generate these files
  return [DefaultInfo(files = depset([binary])),
          OutputGroupInfo(debug_files = depset([debug_file]),
                          all_files = depset([binary, debug_file]))]

By default, only the binary file will be built. The user can specify an --output_groups=debug_files flag on the command line. In that case, only debug_file will be built. If the user specifies --output_groups=all_files, both binary and debug_file will be build.

Note: OutputGroupInfo is a regular provider, and dependencies of a target can examine it using the target[OutputGroupInfo] syntax.

Code coverage instrumentation

A rule can use the instrumented_files provider to provide information about which files should be measured when code coverage data collection is enabled:

def _rule_implementation(ctx):
  ...
  return struct(instrumented_files = struct(
      # Optional: File extensions used to filter files from source_attributes.
      # If not provided, then all files from source_attributes will be
      # added to instrumented files, if an empty list is provided, then
      # no files from source attributes will be added.
      extensions = ["ext1", "ext2"],
      # Optional: Attributes that contain source files for this rule.
      source_attributes = ["srcs"],
      # Optional: Attributes for dependencies that could include instrumented
      # files.
      dependency_attributes = ["data", "deps"]))

ctx.config.coverage_enabled notes whether coverage data collection is enabled for the current run in general (but says nothing about which files specifically should be instrumented). If a rule implementation needs to add coverage instrumentation at compile-time, it can determine if its sources should be instrumented with ctx.coverage_instrumented:

# Are this rule's sources instrumented?
if ctx.coverage_instrumented():
  # Do something to turn on coverage for this compile action

Note that function will always return false if ctx.config.coverage_enabled is false, so you don't need to check both.

If the rule directly includes sources from its dependencies before compilation (e.g. header files), it may also need to turn on compile-time instrumentation if the dependencies' sources should be instrumented. In this case, it may also be worth checking ctx.config.coverage_enabled so you can avoid looping over dependencies unnecessarily:

# Are this rule's sources or any of the sources for its direct dependencies
# in deps instrumented?
if ctx.config.coverage_enabled:
    if (ctx.coverage_instrumented() or
        any(ctx.coverage_instrumented(dep) for dep in ctx.attr.deps):
        # Do something to turn on coverage for this compile action

Executable rules

An executable rule is a rule that users can run using bazel run.

To make a rule executable, set executable=True in the rule function. The implementation function of the rule must generate the output file ctx.outputs.executable. See example

Test rules

Test rules are run using bazel test.

To create a test rule, set test=True in the rule function. The name of the rule must also end with _test. Test rules are implicitly executable, which means that the implementation function of the rule must generate the output file ctx.outputs.executable.

See example

Test rules inherit the following attributes: args, flaky, local, shard_count, size, timeout. The defaults of inherited attributes cannot be changed, but you can use a macro with default arguments:

def example_test(size="small", **kwargs):
  _example_test(size=size, **kwargs)

_example_test = rule(
 ...
)