rules_cc/cc/toolchains/README.md

Writing a custom rule_based C++ toolchain with rule-based definition.

Work in progress!

This document serves two purposes:

• Until complete, this serves as an agreement for the final user-facing API.
• Once complete, this will serve as onboarding documentation.

This section will be removed once complete.

Step 1: Define tools

A tool is simply a binary. Just like any other bazel binary, a tool can specify additional files required to run.

We can use any bazel binary as an input to anything that requires tools. In the example below, you could use both clang and ld as tools.

# @sysroot//:BUILD
cc_tool(
    name = "clang",
    exe = ":bin/clang",
    execution_requirements = ["requires-mem:24g"],
    data = [...],
)

sh_binary(
    name = "ld",
    srcs = ["ld_wrapper.sh"],
    data = [":bin/ld"],
)
    

Step 2: Generate action configs from those tools

An action config is a mapping from action to:

• A list of tools, (the first one matching the execution requirements is used).
• A list of flags and features that are always enabled for the action
• A set of additional files required for the action

Each action can only be specified once in the toolchain. Specifying multiple actions in a single cc_action_config is just a shorthand for specifying the same config for every one of those actions.

If you're already familiar with how to define toolchains, the additional files is a replacement for compile_files, link_files, etc.

Additionally, to replace all_files, we add cc_additional_files_for_actions. This allows you to specify that particular files are required for particular actions.

We provide additional_files on the cc_action_config as a shorthand for specifying cc_additional_files_for_actions

Warning: Implying a feature that is not listed directly in the toolchain will throw an error. This is to ensure you don't accidentally add a feature to the toolchain.

cc_action_config(
    name  = "c_compile",
    actions = ["@rules_cc//actions:all_c_compile"],
    tools = ["@sysroot//:clang"],
    flag_sets = [":my_flag_set"],
    implies = [":my_feature"],
    additional_files = ["@sysroot//:all_header_files"],
)

cc_additional_files_for_actions(
    name = "all_action_files",
    actions = ["@rules_cc//actions:all_actions"],
    additional_files = ["@sysroot//:always_needed_files"]
)

Step 3: Define some flag sets

Flag sets are just sets of flags to be associated with actions. Most flag sets are simple, so we provide the shorthand flags. However, sometimes you need to do more complex things, for which we support flag_groups instead.

Flag groups work exactly the same as the existing toolchain definition.

Flag sets are a combination of both flag_set and env_set from the existing toolchain definition.

cc_flag_set_list is simply a list of flag sets. This can be used to group flag sets together, and preserves ordering.

cc_flag_set(
    name = "simple",
    actions = ["@rules_cc//actions:all_cpp_compile_actions"],
    flags = ["--foo"],
    envs = {"FOO": "bar"},
)

cc_flag_group(
    name = "complex_flag_group",
    # API TBD
)
cc_flag_set(
    name = "complex",
    actions = ["@rules_cc//actions:c_compile"],
    flag_groups = [":complex_flag_group"],
)

cc_flag_set_list(
    name = "all_flags",
    flag_sets = [":simple", ":complex"],
)

Step 4: Define some features

A feature is a set of flags and configurations that can be enabled or disabled.

Although the existing toolchain recommends using features to avoid duplication of definitions, we recommend avoiding using features unless you want the user to be able to enable / disable the feature themselves. This is because we provide alternatives such as cc_flag_set_list to allow combining flag sets and specifying them on each action in the action config.

cc_feature(
    name = "my_feature",
    feature_name = "my_feature",
    flag_sets = [":all_flags"],
    implies = [":other_feature"],
)

Step 5: Generate the toolchain

The cc_toolchain macro:

• Performs validation on the inputs (eg. no two action configs for a single action)
• Converts the type-safe providers to the unsafe ones in cc_toolchain_config_lib.bzl
• Generates a set of providers for each of the filegroups respectively
• Generates the appropriate native.cc_toolchain invocation.

cc_toolchain(
    name = "toolchain",
    features = [":my_feature"]
    unconditional_flag_sets = [":all_warnings"],
    action_configs = [":c_compile"],
    additional_files = [":all_action_files"],
)

Ancillary components for type-safe toolchains.

Well-known features

Well-known features will be defined in @rules_cc//features/well_known:*. Any feature with feature_name in the well known features will have to specify overrides.

cc_toolchain is aware of the builtin / well-known features. In order to ensure that a user understands that this overrides the builtin opt feature (I originally thought that it added extra flags to opt, but you still got the default ones, so that can definitely happen), and to ensure that they don't accidentally do so, we will force them to explicitly specify that it overrides the builtin one. This is essentially just an acknowledgement of "I know what I'm doing".

Warning: Specifying two features with the same name is an error, unless one overrides the other.

cc_feature(
    name = "opt",
    ...,
    overrides = "@rules_cc//features/well_known:opt",
)

In addition to well-known features, we could also consider in future iterations to also use known features for partial migrations, where you still imply a feature that's still defined by the legacy API:

# Implementation
def cc_legacy_features(name, features):
  for feature in features:
    cc_known_feature(name = name + "_" + feature.name)
  cc_legacy_features(name = name, features = FEATURES)


# Build file
FOO = feature(name = "foo", flag_sets=[flag_group(...)])
FEATURES = [FOO]
cc_legacy_features(name = "legacy_features", features = FEATURES)

cc_feature(name = "bar", implies = [":legacy_features_foo"])

cc_toolchain(
  name = "toolchain",
  legacy_features = ":legacy_features",
  features = [":bar"],
)

Mutual exclusion

Features can be mutually exclusive.

We allow two approaches to mutual exclusion - via features or via categories.

The existing toolchain uses provides for both of these. We rename it so that it makes more sense semantically.

cc_feature(
   name = "incompatible_with_my_feature",
   feature_name = "bar",
   mutually_exclusive = [":my_feature"],
)


# This is an example of how we would define compilation mode.
# Since it already exists, this wouldn't work.
cc_mutual_exclusion_category(
    name = "compilation_mode",
)

cc_feature(
    name = "opt",
    ...
    mutually_exclusive = [":compilation_mode"],
)
cc_feature(
    name = "dbg",
    ...
    mutually_exclusive = [":compilation_mode"],
)

Feature requirements

Feature requirements can come in two formats.

For example:

• Features can require some subset of features to be enabled.
• Flag sets can require some subset of features to be enabled, but others to be disabled.

This is very confusing for toolchain authors, so we will simplify things with the use of providers:

• cc_feature will provide feature, feature_set, and with_feature
• cc_feature_set will provide feature_set and with_feature.
• cc_feature_constraint will provide with_features only.

We will rename all with_features and requires to requires_any_of, to make it very clear that only one of the requirements needs to be met.

cc_feature_set(
    name = "my_feature_set",
    all_of = [":my_feature"],
)

cc_feature_constraint(
    name = "my_feature_constraint",
    all_of = [":my_feature"],
    none_of = [":my_other_feature"],
)

cc_flag_set(
   name = "foo",
   # All of these provide with_feature.
   requires_any_of = [":my_feature", ":my_feature_set", ":my_feature_constraint"]
)

# my_feature_constraint would be an error here.
cc_feature(
   name = "foo",
   # Both of these provide feature_set.
   requires_any_of = [":my_feature", ":my_feature_set"]
   implies = [":my_other_feature", :my_other_feature_set"],
)