# How Protobuf supports multiple C++ build systems

This document explains how the Protobuf project supports multiple C++ build
systems.

## Background

Protobuf primarily uses [Bazel](https://bazel.build) to build the Protobuf C++
runtime and Protobuf compiler[^historical_sot]. However, there are several
different build systems in common use for C++, each one of which requires
essentially a complete copy of the same build definitions.

[^historical_sot]: On a historical note, prior to its
    [release as Open Source Software](https://opensource.googleblog.com/2008/07/protocol-buffers-googles-data.html),
    the Protobuf project was developed using Google's internal
    build system, which was the predecessor to Bazel (the vast
    majority of Google's contributions continue to be developed
    this way). The Open Source Protobuf project, however,
    historically used Autoconf to build the C++ implementation.
    Over time, other build systems (including Bazel) have been
    added, thanks in large part to substantial contributions from
    the Open Source community. Since the Protobuf project deals
    with multiple languages (all of which ultimately rely upon
    C++, for the Protobuf compiler), Bazel is a natural choice
    for a project-wide build system -- in fact, Bazel (and its
    predecessor, Blaze) was designed in large part to support
    exactly this type of rich, multi-language build.

Currently, C++ Protobuf can be built with Bazel and CMake. Each of these build
systems has different semantics and structure, but share in common the list of
files needed to build the runtime and compiler.

## Design

### Extracting information from Bazel

Bazel's Starlark API provides [aspects](https://bazel.build/rules/aspects) to
traverse the build graph, inspect build rules, define additional actions, and
expose information through
[providers](https://bazel.build/rules/rules#providers). For example, the
`cc_proto_library` rule uses an aspect to traverse the dependency graph of
`proto_library` rules, and dynamically attaches actions to generate C++ code
using the Protobuf compiler and compile using the C++ compiler.

In order to support multiple build systems, the overall build structure is
defined once for each system, and expose frequently-changing metadata from Bazel
in a way that can be included from the build definition. Primarily, this means
exposing the list of source files in a way that can be included in other build
definitions.

Two aspects are used to extract this information from the Bazel build
definitions:

*   `cc_file_list_aspect` extracts `srcs`, `hdrs`, and `textual_hdrs` from build
    rules like `cc_library`. The sources are exposed through a provider named
    `CcFileList`.
*   `proto_file_list_aspect` extracts the `srcs` from a `proto_library`, and
    also generates the expected filenames that would be generated by the
    Protobuf compiler. This information is exposed through a provider named
    `ProtoFileList`.

On their own, these aspects have limited utility. However, they can be
instantiated by custom rules, so that an ordinary `BUILD.bazel` target can
produce outputs based on the information gleaned from these aspects.

### (Aside) Distribution libraries

Bazel's native `cc_library` rule is typically used on a "fine-grained" level, so
that, for example, lightweight unit tests can be written with narrow scope.
Although Bazel does build library artifacts (such as `.so` and `.a` files on
Linux), they correspond to `cc_library` rules.

Since the entire "Protobuf library" includes many constituent `cc_library`
rules, a special rule, `cc_dist_library`, combines several fine-grained
libraries into a single, monolithic library.

For the Protobuf project, these "distribution libraries" are intended to match
the granularity of the CMake-based builds. Since the Bazel-built distribution
library covers the rules with the source files needed by other builds, the
`cc_dist_library` rule invokes the `cc_file_list_aspect` on its input libraries.
The result is that a `cc_dist_library` rule not only produces composite library
artifacts, but also collect and provide the list of sources that were inputs.

For example:

```
$ cat cc_dist_library_example/BUILD.bazel
load("@rules_cc//cc:defs.bzl", "cc_library")
load("//pkg:cc_dist_library.bzl", "cc_dist_library")

cc_library(
    name = "a",
    srcs = ["a.cc"],
)

cc_library(
    name = "b",
    srcs = ["b.cc"],
    deps = [":c"],
)

# N.B.: not part of the cc_dist_library, even though it is in the deps of 'b':
cc_library(
    name = "c",
    srcs = ["c.cc"],
)

cc_dist_library(
    name = "lib",
    deps = [
        ":a",
        ":b",
    ],
    visibility = ["//visibility:public"],
)

# Note: the output below has been formatted for clarity:
$ bazel cquery //cc_dist_library_example:lib \
    --output=starlark \
    --starlark:expr='providers(target)["//pkg:cc_dist_library.bzl%CcFileList"]'
struct(
    hdrs = depset([]),
    internal_hdrs = depset([]),
    srcs = depset([
        <source file cc_dist_library_example/a.cc>,
        <source file cc_dist_library_example/b.cc>,
    ]),
    textual_hdrs = depset([]),
)
```

The upshot is that the "coarse-grained" library can be defined by the Bazel
build, and then export the list of source files that are needed to reproduce the
library in a different build system.

One major difference from most Bazel rule types is that the file list aspects do
not propagate. In other words, they only expose the immediate dependency's
sources, not transitive sources. This is for two reasons:

1.  Immediate dependencies are conceptually simple, while transitivity requires
    substantially more thought. For example, if transitive dependencies were
    considered, then some way would be needed to exclude dependencies that
    should not be part of the final library (for example, a distribution library
    for `//:protobuf` could be defined not to include all of
    `//:protobuf_lite`). While dependency elision is an interesting design
    problem, the protobuf library is small enough that directly listing
    dependencies should not be problematic.
2.  Dealing only with immediate dependencies gives finer-grained control over
    what goes into the composite library. For example, a Starlark `select()`
    could conditionally add fine-grained libraries to some builds, but not
    others.

Another subtlety for tests is due to Bazel internals. Internally, a slightly
different configuration is used when evaluating `cc_test` rules as compared to
`cc_dist_library`. If `cc_test` targets are included in a `cc_dist_library`
rule, and both are evaluated by Bazel, this can result in a build-time error:
the config used for the test contains additional options that tell Bazel how to
execute the test that the `cc_file_list_aspect` build config does not. Bazel
detects this as two conflicting actions generating the same outputs. (For
`cc_test` rules, the simplest workaround is to provide sources through a
`filegroup` or similar.)

### File list generation

Lists of input files are generated by Bazel in a format that can be imported to
other build systems. Currently only CMake-style files can be generated.

The lists of files are derived from Bazel build targets. The sources can be: *
`cc_dist_library` rules (as described above) * `proto_library` rules *
individual files * `filegroup` rules * `pkg_files` or `pkg_filegroup` rules from
https://github.com/bazelbuild/rules_pkg

For example:

```
$ cat gen_file_lists_example/BUILD.bazel
load("@protobuf//bazel:proto_library.bzl", "proto_library")
load("//pkg:build_systems.bzl", "gen_cmake_file_lists")

filegroup(
    name = "doc_files",
    srcs = [
        "README.md",
        "englilsh_paper.md",
    ],
)

proto_library(
    name = "message",
    srcs = ["message.proto"],
)

gen_cmake_file_lists(
    name = "source_lists",
    out = "source_lists.cmake",
    src_libs = {
        ":doc_files": "docs",
        ":message": "buff",
        "//cc_dist_library_example:c": "distlib",
    },
)

$ bazel build gen_file_lists_example:source_lists
$ cat bazel-bin/gen_file_lists_example/source_lists.cmake
# Auto-generated by //gen_file_lists_example:source_lists
#
# This file contains lists of sources based on Bazel rules. It should
# be included from a hand-written CMake file that defines targets.
#
# Changes to this file will be overwritten based on Bazel definitions.

if(${CMAKE_VERSION} VERSION_GREATER 3.10 OR ${CMAKE_VERSION} VERSION_EQUAL 3.10)
  include_guard()
endif()

# //gen_file_lists_example:doc_files
set(docs_files
  gen_file_lists_example/README.md
  gen_file_lists_example/englilsh_paper.md
)

# //gen_file_lists_example:message
set(buff_proto_srcs
  gen_file_lists_example/message.proto
)

# //gen_file_lists_example:message
set(buff_srcs
  gen_file_lists_example/message.proto.pb.cc
)

# //gen_file_lists_example:message
set(buff_hdrs
  gen_file_lists_example/message.proto.pb.h
)

# //gen_file_lists_example:message
set(buff_files
  gen_file_lists_example/message-descriptor-set.proto.bin
)

# //cc_dist_library_example:c
set(distlib_srcs
  cc_dist_library_example/a.cc
  cc_dist_library_example/b.cc
)

# //cc_dist_library_example:c
set(distlib_hdrs

)
```

A hand-written CMake build rule could then use the generated file to define
libraries, such as:

```
include(source_lists.cmake)
add_library(distlib ${distlib_srcs} ${buff_srcs})
```

### Protobuf usage

The main C++ runtimes (lite and full) and the Protobuf compiler use their
corresponding `cc_dist_library` rules to generate file lists. For
`proto_library` targets, the file list generation can extract the source files
directly. For other targets, notably `cc_test` targets, the file list generators
use `filegroup` rules.

In general, adding new targets to a non-Bazel build system in Protobuf (or
adding a new build system altogether) requires some one-time setup:

1.  The overall structure of the new build system has to be defined. It should
    import lists of files and refer to them by variable, instead of listing
    files directly.
2.  (Only if the build system is new) A new rule type has to be added to
    `//pkg:build_systems.bzl`. Most of the implementation is shared, but a
    "fragment generator" is need to declare a file list variable, and the rule
    type itself has to be defined and call the shared implementation.

When files are added or deleted, or when the Protobuf Bazel structure is
changed, these changes may need to be reflected in the file list logic. These
are some example scenarios:

*   Files are added to (or removed from) the `srcs` of an existing `cc_library`:
    no changes needed. If the `cc_library` is already part of a
    `cc_dist_library`, then regenerating the source lists will reflect the
    change.
*   A `cc_library` is added: the new target may need to be added to the Protobuf
    `cc_dist_library` targets, as appropriate.
*   A `cc_library` is deleted: if a `cc_dist_library` depends upon the deleted
    target, then a build-time error will result. The library needs to be removed
    from the `cc_dist_library`.
*   A `cc_test` is added or deleted: test sources are handled by `filegroup`
    rules defined in the same package as the `cc_test` rule. The `filegroup`s
    are usually given a name like `"test_srcs"`, and often use `glob()` to find
    sources. This means that adding or removing a test may not require any extra
    work, but this can be verified within the same package as the test rule.
*   Test-only proto files are added: the `proto_library` might need to be added
    to the file list map in `//pkg:BUILD.bazel`, and then the file added to
    various build systems. However, most test-only protos are already exposed
    through libraries like `//src/google/protobuf:test_protos`.

If there are changes, then the regenerated file lists need to be copied back
into the repo. That way, the corresponding build systems can be used with a git
checkout, without needing to run Bazel first.

### (Aside) Distribution archives

A very similar set of rules is defined in `//pkg` to build source distribution
archives for releases. In addition to the full sources, Protobuf releases also
include source archives sliced by language, so that, for example, a Ruby-based
project can get just the sources needed to build the Ruby runtime. (The
per-language slices also include sources needed to build the protobuf compiler,
so they all effectively include the C++ runtime.)

These archives are defined using rules from the
[rules_pkg](https://github.com/bazelbuild/rules_pkg) project. Although they are
similar to `cc_dist_library` and the file list generation rules, the goals are
different: the build system file lists described above only apply to C++, and
are organized according to what should or should not be included in different
parts of the build (e.g., no tests are included in the main library). On the
other hand, the distribution archives deal with languages other than C++, and
contain all the files that need to be distributed as part of a release (even for
C++, this is more than just the C++ sources).

While it might be possible to use information from the `CcFileList` and
`ProtoFileList` providers to define the distribution files, additional files
(such as the various `BUILD.bazel` files) are also needed in the distribution
archive. The lists of distribution files can usually be generated by `glob()`,
anyhow, so sharing logic with the file list aspects may not be beneficial.

Currently, all of the file lists are checked in. However, it would be possible
to build the file lists on-the-fly and include them in the distribution
archives, rather than checking them in.