"Make" Variables

"Make" variables are a special class of expandable string variables available to attributes marked as "Subject to 'Make variable' substitution".

These can be used, for example, to inject specific toolchain paths into user-constructed build actions.

Bazel provides both predefined variables, which are available to all targets, and custom variables, which are defined in dependency targets and only available to targets that depend on them.

The reason for the term "Make" is historical: the syntax and semantics of these variables were originally intended to match GNU Make.

Use

Attributes marked as "Subject to 'Make variable' substitution" can reference the "Make" variable FOO as follows:

my_attr = "prefix $(FOO) suffix"

In other words, any substring matching $(FOO) gets expanded to FOO's value. If that value is "bar", the final string becomes:

my_attr = "prefix bar suffix"

If FOO doesn't correspond to a variable known to the consuming target, Bazel fails with an error.

"Make" variables whose names are non-letter symbols, such as @, can also be referenced using only a dollar sign, without the parentheses. For example:

my_attr = "prefix $@ suffix"

To write $ as a a string literal (i.e. to prevent variable expansion), write $$.

Predefined variables

Predefined "Make" variables can be referenced by any attribute marked as "Subject to 'Make variable' substitution" on any target.

To see the list of these variables and their values for a given set of build options, run

bazel info --show_make_env [build options]

and look at the top output lines with capital letters.

See an example of predefined variables.

Toolchain option variables

Path variables

  • BINDIR: The base of the generated binary tree for the target architecture.

    Note that a different tree may be used for programs that run during the build on the host architecture, to support cross-compiling.

    If you want to run a tool from within a genrule, the recommended way to get its path is $(execpath toolname), where toolname must be listed in the genrule's tools attribute.

  • GENDIR: The base of the generated code tree for the target architecture.

Machine architecture variables

  • TARGET_CPU: The target architecture's CPU, e.g. k8.

Predefined genrule variables

The following are specially available to genrule's cmd attribute and are generally important for making that attribute work.

See an example of predefined genrule variables.

  • OUTS: The genrule's outs list. If you have only one output file, you can also use $@.
  • SRCS: The genrule's srcs list (or more precisely: the path names of the files corresponding to labels in the srcs list). If you have only one source file, you can also use $<.
  • <: SRCS, if it is a single file. Else triggers a build error.
  • @: OUTS, if it is a single file. Else triggers a build error.

  • RULEDIR: The output directory of the target, that is, the directory corresponding to the name of the package containing the target under the genfiles or bin tree. For //my/pkg:my_genrule this always ends in my/pkg, even if //my/pkg:my_genrule's outputs are in subdirectories.

  • @D: The output directory. If outs has one entry, this expands to the directory containing that file. If it has multiple entries, this expands to the package's root directory in the genfiles tree, even if all output files are in the same subdirectory!

    Note: Use RULEDIR over @D because RULEDIR has simpler semantics and behaves the same way regardless of the number of output files.

    If the genrule needs to generate temporary intermediate files (perhaps as a result of using some other tool like a compiler), it should attempt to write them to @D (although /tmp will also be writable) and remove them before finishing.

    Especially avoid writing to directories containing inputs. They may be on read-only filesystems. Even if not, doing so would trash the source tree.

Predefined source/output path variables

The predefined variables execpath, execpaths, rootpath, rootpaths, location, and locations take label parameters (e.g. $(execpath //foo:bar)) and substitute the file paths denoted by that label.

For source files, this is the path relative to your workspace root. For files that are outputs of rules, this is the file's output path (see the explanation of output files below).

See an example of predefined path variables.

  • execpath: Denotes the path beneath the execroot where Bazel runs build actions.

    In the above example, Bazel runs all build actions in the directory linked by the bazel-myproject symlink in your workspace root. The source file empty.source is linked at the path bazel-myproject/testapp/empty.source. So its exec path (which is the subpath below the root) is testapp/empty.source. This is the path build actions can use to find the file.

    Output files are staged similarly, but are also prefixed with the subpath bazel-out/cpu-compilation_mode/bin (or for certain outputs: bazel-out/cpu-compilation_mode/genfiles, or for the outputs of host tools: bazel-out/host/bin). In the above example, //testapp:app is a host tool because it appears in show_app_output's tools attribute. So its output file app is written to bazel-myproject/bazel-out/host/bin/testapp/app. The exec path is thus bazel-out/host/bin/testapp/app. This extra prefix makes it possible to build the same target for, say, two different CPUs in the same build without the results clobbering each other.

    The label passed to this variable must represent exactly one file. For labels representing source files, this is automatically true. For labels representing rules, the rule must generate exactly one output. If this is false or the label is malformed, the build fails with an error.

  • rootpath: Denotes the runfiles path that a built binary can use to find its dependencies at runtime.

    This is the same as execpath but strips the output prefixes described above. In the above example this means both empty.source and app use pure workspace-relative paths: testapp/empty.source and testapp/app.

    This has the same "one output only" requirements as execpath.

  • location: A synonym for either execpath or rootpath, depending on the attribute being expanded. This is legacy pre-Starlark behavior and not recommended unless you really know what it does for a particular rule. See #2475 for details.

execpaths, rootpaths, and locations are the plural variations of execpath, rootpath, and location, respectively. They support labels producing multiple outputs, in which case each output is listed separated by a space. Zero-output rules and malformed labels produce build errors.

All referenced labels must appear in the consuming target's srcs, output files, or deps. Otherwise the build fails. C++ targets can also reference labels in data.

Labels don't have to be in canonical form: foo, :foo and //somepkg:foo are all fine.

Custom variables

Custom "Make" variables can be referenced by any attribute marked as "Subject to 'Make variable' substitution", but only on targets that depend on other targets that define these variables.

As best practice all variables should be custom unless there's a really good reason to bake them into core Bazel. This saves Bazel from having to load potentially expensive dependencies to supply variables consuming tarets may not care about.

C++ toolchain variables

The following are defined in C++ toolchain rules and available to any rule that sets toolchains = ["@bazel_tools//tools/cpp:current_cc_toolchain"] (or "@bazel_tools//tools/cpp:current_cc_host_toolchain" for the host toolchain equivalent). Some rules, like java_binary, implicitly include the C++ toolchain in their rule definition. They inherit these variables automatically.

The built-in C++ rules are much more sophisticated than "run the compiler on it". In order to support compilation modes as diverse as *SAN, ThinLTO, with/without modules, and carefully optimized binaries at the same time as fast running tests on multiple platforms, the built-in rules go to great lengths to ensure the correct inputs, outputs, and command-line flags are set on each of potentially multiple internally generated actions.

These variables are a fallback mechanism to be used by language experts in rare cases. If you are tempted to use them, please contact the Bazel devs first.

  • ABI: The C++ ABI version.
  • AR: The "ar" command from crosstool.
  • C_COMPILER: The C/C++ compiler identifier, e.g. llvm.

  • CC: The C and C++ compiler command.

    We strongly recommended always using CC_FLAGS in combination with CC. Fail to do so at your own risk.

  • CC_FLAGS: A minimal set of flags for the C/C++ compiler to be usable by genrules. In particular, this contains flags to select the correct architecture if CC supports multiple architectures.
  • NM: The "nm" command from crosstool.
  • OBJCOPY: The objcopy command from the same suite as the C/C++ compiler.
  • STRIP: The strip command from the same suite as the C/C++ compiler.

Java toolchain variables

The following are defined in Java toolchain rules and available to any rule that sets toolchains = ["@bazel_tools//tools/jdk:current_java_runtime"] (or "@bazel_tools//tools/jdk:current_host_java_runtime" for the host toolchain equivalent).

Most of the tools in the JDK should not be used directly. The built-in Java rules use much more sophisticated approaches to Java compilation and packaging than upstream tools can express, such as interface Jars, header interface Jars, and highly optimized Jar packaging and merging implementations.

These variables are a fallback mechanism to be used by language experts in rare cases. If you are tempted to use them, please contact the Bazel devs first.

  • JAVA: The "java" command (a Java virtual machine). Avoid this, and use a java_binary rule instead where possible. May be a relative path. If you must change directories before invoking java, you need to capture the working directory before changing it.
  • JAVABASE: The base directory containing the Java utilities. May be a relative path. It will have a "bin" subdirectory.

Starlark-defined variables

Rule and toolchain writers can define completely custom variables by returning a TemplateVariableInfo provider. Any rules depending on these through the toolchains attribute can then read their values:

See an example of Starlark-defined variables.