b(1) General Commands Manual b(1)

b - build system driver

b --help
b --version
b [options] [variables] [buildspec]

buildspec = meta-operation(operation(target...[,parameters])...)...

The build2 build system driver executes a set of meta-operations on operations on targets according to the build specification, or buildspec for short. This process can be controlled by specifying driver options and build system variables.

Note that options, variables, and buildspec fragments can be specified in any order. To avoid treating an argument that starts with '-' as an option, add the '--' separator. To avoid treating an argument that contains '=' as a variable, add the second '--' separator.

All components in the buildspec can be omitted. If meta-operation is omitted, then it defaults to perform. If operation is omitted, then it defaults to the default operation for this meta-operation. For perform it is update. Finally, if target is omitted, then it defaults to the current working directory. A meta-operation on operation is called an action. Some operations and meta-operations may take additional parameters. For example:

$ b                       # perform(update(./))
$ b foo/                  # perform(update(foo/))
$ b foo/ bar/             # perform(update(foo/ bar/))
$ b update                # perform(update(./))
$ b 'clean(../)'          # perform(clean(../))
$ b perform               # perform(update(./))
$ b configure             # configure(?(./))
$ b 'configure(../)'      # configure(?(../))
$ b clean update          # perform(clean(./) update(./))
$ b configure update      # configure(?(./)) perform(update(./))
$ b 'create(conf/, cxx)'  # create(?(conf/), cxx)

Notice the question mark used to show the (imaginary) default operation for the configure meta-operation. For configure the default operation is "all operations". That is, it will configure all the operations for the specified target.

You can also "generate" multiple operations for the same set of targets. Compare:

$ b 'clean(foo/ bar/)' 'update(foo/ bar/)'
$ b '{clean update}(foo/ bar/)'

Some more useful buildspec examples:

$ b '{clean update}(...)'        # rebuild
$ b '{clean update clean}(...)'  # make sure builds
$ b '{clean test clean}(...)'    # make sure passes tests
$ b '{clean disfigure}(...)'     # similar to distclean

In POSIX shells parenthesis are special characters and must be quoted when used in a buildspec. Besides being an inconvenience in itself, quoting also inhibits path auto-completion. To help with this situation a shortcut syntax is available for executing a single operation or meta-operation, for example:

$ b clean: foo/ bar/                # clean(foo/ bar/)
$ b configure: src/@out/            # configure(src/@out/)
$ b create: conf/, cxx              # create(conf/, cxx)
$ b configure: config.cxx=g++ src/  # configure(src/) config.cxx=g++

To activate the shortcut syntax the first buildspec argument must start with an operation or meta-operation name and end with a colon (:). To transform the shortcut syntax to the normal buildspec syntax the colon is replaced with the opening parenthesis ('('), the rest of the buildspec arguments are treated as is, and the final closing parenthesis (')') is added.

For each target the driver expects to find buildfile either in the target's directory or, if the directory is part of the out tree (out_base), in the corresponding src directory (src_base).

For example, assuming foo/ is the source directory of a project:

$ b foo/              # out_base=src_base=foo/
$ b foo-out/          # out_base=foo-out/ src_base=foo/
$ b foo-out/exe{foo}  # out_base=foo-out/ src_base=foo/

An exception to this requirement is a directory target in which case, provided the directory has subdirectories, an implied buildfile with the following content is assumed:

# Implied directory buildfile: build all subdirectories.
#
./: */

In the above example, we assumed that the build system driver was able to determine the association between out_base and src_base. In case src_base and out_base are not the same directory, this is achieved in one of two ways: the config module (which implements the configure, disfigure, and create meta-operations) saves this association as part of the configuration process. If, however, the association hasn't been saved, then we have to specify src_base explicitly using the following extended target syntax:

src-base/@target

Continuing with the previous example:

$ b foo/@foo-out/exe{foo}  # out_base=foo-out/ src_base=foo/

Normally, you would need to specify src_base explicitly only once, during configuration. For example, a typical usage would be:

$ b configure: foo/@foo-out/  # src_base is saved
$ b foo-out/                  # no need to specify src_base
$ b clean: foo-out/exe{foo}   # no need to specify src_base

Besides in and out of source builds, build2 also supports configuring a project's source directory as forwarded to an out of source build. With such a forwarded configuration in place, if we run the build system driver from the source directory, it will automatically build in the output directory and backlink (using symlinks or another suitable mechanism) certain "interesting" targets (executables, documentation, etc) to the source directory for easy access. Continuing with the previous example:

$ b configure: foo/@foo-out/,forward  # foo/ forwarded to foo-out/
$ cd foo/
$ b                                   # build in foo-out/
$ ./foo                               # symlink to foo-out/foo

The ability to specify build2 variables as part of the command line is normally used to pass configuration values, for example:

$ b config.cxx=clang++ config.cxx.coptions=-O3

Similar to buildspec, POSIX shells often inhibit path auto-completion on the right hand side of a variable assignment. To help with this situation the assignment can be broken down into three separate command line arguments, for example:

$ b config.import.libhello = ../libhello/

The build system has the following built-in and pre-defined meta-operations:


Perform an operation.

Configure all operations supported by a project and save the result in the project's build/config.build file. Implemented by the config module. For example: 
$ b configure                      \
    config.cxx=clang++             \
    config.cxx.coptions=-O3        \
    config.install.root=/usr/local \
    config.install.root.sudo=sudo

Use the forward parameter to instead configure a source directory as forwarded to an out of source build. For example:

$ b configure: src/@out/,forward

Disfigure all operations supported by a project and remove the project's build/config.build file. Implemented by the config module.

Use the forward parameter to instead disfigure forwarding of a source directory to an out of source build. For example:

$ b disfigure: src/,forward

Create and configure a configuration project. Implemented by the config module.

Normally a build2 project is created manually by writing the bootstrap.build and config.build files, adding source files, and so on. However, a special kind of project, which we call configuration, is often useful. Such a project doesn't have any source files of its own. Instead, it serves as an amalgamation for building other projects as part of it. Doing it this way has two major benefits: sub-projects automatically resolve their imports to other projects in the amalgamation and sub-projects inherits their configuration from the amalgamation (which means if we want to change something, we only need to do it in one place).

As an example, let's assume we have two C++ projects: the libhello library in libhello/ and the hello executable that imports it in hello/. And we want to build hello with clang++.

One way to do it would be to configure and build each project in its own directory, for example:

$ b configure: libhello/@libhello-clang/ config.cxx=clang++
$ b configure: hello/@hello-clang/ config.cxx=clang++ \
    config.import.libhello=libhello-clang/

The two drawbacks, as mentioned above, are the need to explicitly resolve the import and having to make changes in multiple places should, for example, we want to switch from clang++ to g++.

We can, however, achieve the same end result but without any of the drawbacks using the configuration project:

$ b create: clang/,cxx config.cxx=clang++  # Creates clang/.
$ b configure: libhello/@clang/libhello/
$ b configure: hello/@clang/hello/

The targets passed to the create meta-operation must be directories which should either not exist or be empty. For each such directory create first initializes a project as described below and then configures it by executing the configure meta-operation.

The first optional parameter to create is the list of modules to load in root.build. By default, create appends .config to the names of these modules so that only their configurations are loaded. You can override this behavior by specifying the period (.) after the module name. You can also instruct create to use the optional module load by prefixing the module name with the question mark (?).

The second optional parameter is the list of modules to load in bootstrap.build. If not specified, then the test, dist, and install modules are loaded by default. The config module is always loaded first.

Besides creating project's bootstrap.build and root.build, create also writes the root buildfile with the following contents:

./: {*/ -build/}

If used, this buildfile will build all the sub-projects currently present in the configuration.


Prepare a distribution containing all files necessary to perform all operations in a project. Implemented by the dist module.

Print basic information (name, version, source and output directories, etc) about one or more projects to stdout, separating multiple projects with a blank line. Each project is identified by its root directory target. For example (some output is omitted): 
$ b info: libfoo/ libbar/
project: libfoo
version: 1.0.0
src_root: /tmp/libfoo
out_root: /tmp/libfoo
subprojects: @tests
project: libbar
version: 2.0.0
src_root: /tmp/libbar
out_root: /tmp/libbar-out
subprojects: @tests

To omit discovering and printing subprojects information, use the no_subprojects parameter, for example:

$ b info: libfoo/,no_subprojects

To instead print this information in the JSON format, use the json parameter, for example:

$ b info: libfoo/,json

In this case the output is a JSON array of objects which are the serialized representation of the following C++ struct project_info:

struct subproject
{
  string           path;
  optional<string> name;
};
struct project_info
{
  optional<string>   project;
  optional<string>   version;
  optional<string>   summary;
  optional<string>   url;
  string             src_root;
  string             out_root;
  optional<string>   amalgamation;
  vector<subproject> subprojects;
  vector<string>     operations;
  vector<string>     meta_operations;
  vector<string>     modules;
};

For example:

[
  {
    "project": "libfoo",
    "version": "1.0.0",
    "summary": "libfoo C++ library",
    "src_root": "/tmp/libfoo",
    "out_root": "/tmp/gcc-debug/libfoo",
    "amalgamation": "..",
    "subprojects": [
      {
        "path": "tests"
      }
    ],
    "operations": [
      "update",
      "clean",
      "test",
      "update-for-test",
      "install",
      "uninstall",
      "update-for-install"
    ],
    "meta-operations": [
      "perform",
      "configure",
      "disfigure",
      "dist",
      "info"
    ],
    "modules": [
      "version",
      "config",
      "test",
      "install",
      "dist"
    ]
  }
]

See the JSON OUTPUT section below for details on the overall properties of this format and the semantics of the struct serialization.

The build system has the following built-in and pre-defined operations:


Update a target.

Clean a target.

Test a target. Performs update as a pre-operation. Implemented by the test module.

Update a target for testing. This operation is equivalent to the update pre-operation as executed by the test operation and can be used to only update what is necessary for testing. Implemented by the test module.

Install a target. Performs update as a pre-operation. Implemented by the install module.

Uninstall a target. Performs update as a pre-operation. Implemented by the install module.

Update a target for installation. This operation is equivalent to the update pre-operation as executed by the install operation and can be used to only update what is necessary for installation. Implemented by the install module.

Note that buildspec and command line variable values are treated as buildfile fragments and so can use quoting and escaping as well as contain variable expansions and evaluation contexts. However, to be more usable on various platforms, escaping in these two situations is limited to the effective sequences of \', \", \\, \$, and \( with all other sequences interpreted as is. Together with double-quoting this is sufficient to represent any value. For example:

$ b config.install.root=c:\projects\install
$ b "config.install.root='c:\Program Files\test\'"
$ b 'config.cxx.poptions=-DFOO_STR="foo"'

Print actual commands being executed. This options is equivalent to --verbose 2.
Print all underlying commands being executed. This options is equivalent to --verbose 3.
Run quietly, only printing error messages in most contexts. In certain contexts (for example, while updating build system modules) this verbosity level may be ignored. Use --silent to run quietly in all contexts. This option is equivalent to --verbose 0.
Run quietly, only printing error messages in all contexts.
Set the diagnostics verbosity to level between 0 and 6. Level 0 disables any non-error messages (but see the difference between --quiet and --silent) while level 6 produces lots of information, with level 1 being the default. The following additional types of diagnostics are produced at each level:
1.
High-level information messages.
2.
Essential underlying commands being executed.
3.
All underlying commands being executed.
4.
Information that could be helpful to the user.
5.
Information that could be helpful to the developer.
6.
Even more detailed information.
Display build statistics.
Display build progress. If printing to a terminal the progress is displayed by default for low verbosity levels. Use --no-progress to suppress.
Don't display build progress.
Use color in diagnostics. If printing to a terminal the color is used by default provided the terminal is not dumb. Use --no-diag-color to suppress.

This option affects the diagnostics printed by the build system itself. Some rules may also choose to propagate its value to tools (such as compilers) that they invoke.

Don't use color in diagnostics.
Number of active jobs to perform in parallel. This includes both the number of active threads inside the build system as well as the number of external commands (compilers, linkers, etc) started but not yet finished. If this option is not specified or specified with the 0 value, then the number of available hardware threads is used.
Maximum number of jobs (threads) to create. The default is 8x the number of active jobs (--jobs|j) on 32-bit architectures and 32x on 64-bit. See the build system scheduler implementation for details.
The queue depth as a multiplier over the number of active jobs. Normally we want a deeper queue if the jobs take long (for example, compilation) and shorter if they are quick (for example, simple tests). The default is 4. See the build system scheduler implementation for details.
File cache implementation to use for intermediate build results. Valid values are noop (no caching or compression) and sync-lz4 (no caching with synchronous LZ4 on-disk compression). If this option is not specified, then a suitable default implementation is used (currently sync-lz4).
The maximum stack size in KBytes to allow for newly created threads. For pthreads-based systems the driver queries the stack size of the main thread and uses the same size for creating additional threads. This allows adjusting the stack size using familiar mechanisms, such as ulimit. Sometimes, however, the stack size of the main thread is excessively large. As a result, the driver checks if it is greater than a predefined limit (64MB on 64-bit systems and 32MB on 32-bit ones) and caps it to a more sensible value (8MB) if that's the case. This option allows you to override this check with the special zero value indicating that the main thread stack size should be used as is.
Run serially and stop at the first error. This mode is useful to investigate build failures that are caused by build system errors rather than compilation errors. Note that if you don't want to keep going but still want parallel execution, add --jobs|-j (for example -j 0 for default concurrency). Note also that during serial execution there is no diagnostics buffering and child process' stderr is a terminal (unless redirected; see --no-diag-buffer for details).
Print commands without actually executing them. Note that commands that are required to create an accurate build state will still be executed and the extracted auxiliary dependency information saved. In other words, this is not the "don't touch the filesystem" mode but rather "do minimum amount of work to show what needs to be done". Note also that only the perform meta-operation supports this mode.
Do not buffer diagnostics from child processes. By default, unless running serially, such diagnostics is buffered and printed all at once after each child exits in order to prevent interleaving. However, this can have side-effects since the child process' stderr is no longer a terminal. Most notably, the use of color in diagnostics may be disabled by some programs. On the other hand, depending on the platform and programs invoked, the interleaving diagnostics may not break lines and thus could be tolerable.
Match the rules without executing the operation. This mode is primarily useful for profiling and dumping the build system state.
Match the rules only to alias{} targets ignoring other targets and without executing the operation. In particular, this has the effect of loading all the subdirectory buildfiles that are not explicitly included. Note that this option can only be used with the perform(update) action on an alias{} target, usually dir{}.
Don't load external modules during project bootstrap. Note that this option can only be used with meta-operations that do not load the project's buildfiles, such as info.
Write the result of execution in a structured form. In this mode, instead of printing to stderr diagnostics messages about the outcome of executing actions on targets, the driver writes to stdout a machine-readable result description in the specified format. Valid values for this option are lines and json. Note that currently only the perform meta-operation supports the structured result output.

If the output format is lines, then the result is written one line per the buildspec action/target pair. Each line has the following form:

state meta-operation operation target

Where state can be one of unchanged, changed, or failed. If the action is a pre or post operation, then the outer operation is specified in parenthesis. For example:

unchanged perform update(test) /tmp/hello/hello/exe{hello}
changed perform test /tmp/hello/hello/exe{hello}

If the output format is json, then the output is a JSON array of objects which are the serialized representation of the following C++ struct target_action_result:

struct target_action_result
{
  string           target;
  string           display_target;
  string           target_type;
  optional<string> target_path;
  string           meta_operation;
  string           operation;
  optional<string> outer_operation;
  string           state;
};

For example:

[
  {
    "target": "/tmp/hello/hello/exe{hello.}",
    "display_target": "/tmp/hello/hello/exe{hello}",
    "target_type": "exe",
    "target_path": "/tmp/hello/hello/hello",
    "meta_operation": "perform",
    "operation": "update",
    "outer_operation": "test",
    "state": "unchanged"
  },
  {
    "target": "/tmp/hello/hello/exe{hello.}",
    "display_target": "/tmp/hello/hello/exe{hello}",
    "target_type": "exe",
    "target_path": "/tmp/hello/hello/hello",
    "meta_operation": "perform",
    "operation": "test",
    "state": "changed"
  }
]

See the JSON OUTPUT section below for details on the overall properties of this format and the semantics of the struct serialization.

The target member is the target name that is qualified with the extension (if applicable) and, if required, is quoted so that it can be passed back to the build system driver on the command line. The display_target member is the unqualified and unquoted "display" target name, the same as in the lines format. The target_type member is the type of target. The target_path member is an absolute path to the target if the target type is path-based or dir.

Perform file modification time sanity checks. These checks can be helpful in diagnosing spurious rebuilds and are enabled by default on Windows (which is known not to guarantee monotonically increasing mtimes) and for the staged version of the build system on other platforms. Use --no-mtime-check to disable.
Don't perform file modification time sanity checks. See --mtime-check for details.
Dump the build system state after the specified phase. Valid phase values are load (after loading buildfiles) and match (after matching rules to targets). The match value also has the match-pre and match-post variants to dump the state for the pre/post-operations (match dumps the main operation only). Repeat this option to dump the state after multiple phases/variants. By default the entire build state is dumped but this behavior can be altered with the --dump-scope and --dump-target options. See also the --match-only and --load-only options.
Representation format and output stream to use when dumping the build system state. Valid values for this option are buildfile (a human-readable, Buildfile-like format written to stderr; this is the default), and json-v0.1 (machine-readable, JSON-based format written to stdout). For details on the buildfile format, see Diagnostics and Debugging (#intro-diag-debug). For details on the json-v0.1 format, see the JSON OUTPUT section below (overall properties) and JSON Dump Format (#json-dump) (format specifics). Note that the JSON format is currently unstable (thus the temporary -v0.1 suffix).

Note that because it's possible to end up with multiple dumps (for example, by specifying the --dump-scope and/or --dump-target options multiple times), the JSON output is in the "JSON Lines" form, that is, without pretty-printing and with the top-level JSON objects delimited by newlines. Note also that if the JSON dump output is combined with --structured-result=json, then the structured result is the last line.

Dump the build system state for the specified scope only. Repeat this option to dump the state of multiple scopes.
Dump the build system state for the specified target only. Repeat this option to dump the state of multiple targets.
Trace rule matching for the specified target. This is primarily useful during troubleshooting. Repeat this option to trace multiple targets.
Trace rule execution for the specified target. This is primarily useful during troubleshooting. Repeat this option to trace multiple targets.
Don't print column numbers in diagnostics.
Don't print line and column numbers in diagnostics.
The alternative file to read build information from. The default is buildfile or build2file, depending on the project's build file/directory naming scheme. If path is '-', then read from stdin. Note that this option only affects the files read as part of the buildspec processing. Specifically, it has no effect on the source and include directives. As a result, this option is primarily intended for testing rather than changing the build file names in real projects.
The path to the config.guess(1) script that should be used to guess the host machine triplet. If this option is not specified, then b will fall back on to using the target it was built for as host.
The path to the config.sub(1) script that should be used to canonicalize machine triplets. If this option is not specified, then b will use its built-in canonicalization support which should be sufficient for commonly-used platforms.
The pager program to be used to show long text. Commonly used pager programs are less and more. You can also specify additional options that should be passed to the pager program with --pager-option. If an empty string is specified as the pager program, then no pager will be used. If the pager program is not explicitly specified, then b will try to use less. If it is not available, then no pager will be used.
Additional option to be passed to the pager program. See --pager for more information on the pager program. Repeat this option to specify multiple pager options.
Read additional options from file. Each option should appear on a separate line optionally followed by space or equal sign (=) and an option value. Empty lines and lines starting with # are ignored. Option values can be enclosed in double (") or single (') quotes to preserve leading and trailing whitespaces as well as to specify empty values. If the value itself contains trailing or leading quotes, enclose it with an extra pair of quotes, for example '"x"'. Non-leading and non-trailing quotes are interpreted as being part of the option value.

The semantics of providing options in a file is equivalent to providing the same set of options in the same order on the command line at the point where the --options-file option is specified except that the shell escaping and quoting is not required. Repeat this option to specify more than one options file.

The directory to load additional default options files from.
Don't load default options files.
Print usage information and exit.
Print version and exit.

Instead of having a separate config file format for tool configuration, the build2 toolchain uses default options files which contain the same options as what can be specified on the command line. The default options files are like options files that one can specify with --options-file except that they are loaded by default.

The default options files for the build system driver are called b.options and are searched for in the .build2/ subdirectory of the home directory and in the system directory (for example, /etc/build2/) if configured. Note that besides options these files can also contain global variable overrides.

Once the search is complete, the files are loaded in the reverse order, that is, beginning from the system directory (if any), followed by the home directory, and finishing off with the options specified on the command line. In other words, the files are loaded from the more generic to the more specific with the command line options having the ability to override any values specified in the default options files.

If a default options file contains --no-default-options, then the search is stopped at the directory containing this file and no outer files are loaded. If this option is specified on the command line, then none of the default options files are searched for or loaded.

An additional directory containing default options files can be specified with --default-options. Its configuration files are loaded after the home directory.

The order in which default options files are loaded is traced at the verbosity level 3 (-V option) or higher.

Commands that support the JSON output specify their formats as a serialized representation of a C++ struct or an array thereof. For example:

struct package
{
  string name;
};
struct configuration
{
  uint64_t         id;
  string           path;
  optional<string> name;
  bool             default;
  vector<package>  packages;
};

An example of the serialized JSON representation of struct configuration:

{
  "id": 1,
  "path": "/tmp/hello-gcc",
  "name": "gcc",
  "default": true,
  "packages": [
    {
      "name": "hello"
    }
  ]
}

This sections provides details on the overall properties of such formats and the semantics of the struct serialization.

The order of members in a JSON object is fixed as specified in the corresponding struct. While new members may be added in the future (and should be ignored by older consumers), the semantics of the existing members (including whether the top-level entry is an object or array) may not change.

An object member is required unless its type is optional<>, bool, or vector<> (array). For bool members absent means false. For vector<> members absent means empty. An empty top-level array is always present.

For example, the following JSON text is a possible serialization of the above struct configuration:

{
  "id": 1,
  "path": "/tmp/hello-gcc"
}

Non-zero exit status is returned in case of an error.

The HOME environment variable is used to determine the user's home directory. If it is not set, then getpwuid(3) is used instead. This value is used to shorten paths printed in diagnostics by replacing the home directory with ~/. It is also made available to buildfile's as the build.home variable.

The BUILD2_VAR_OVR environment variable is used to propagate global variable overrides to nested build system driver invocations. Its value is a list of global variable assignments separated with newlines.

The BUILD2_DEF_OPT environment variable is used to suppress loading of default options files in nested build system driver invocations. Its values are false or 0 to suppress and true or 1 to load.

Send bug reports to the users@build2.org mailing list.

Copyright (c) 2014-2024 the build2 authors.

Permission is granted to copy, distribute and/or modify this document under the terms of the MIT License.

June 2024 build2 0.17.0