Configurable Build Attributes

Contents

 

Configurable attributes, commonly known as select(), is a Bazel feature that lets users toggle the values of BUILD rule attributes at the command line.

This can be used, for example, for a multiplatform library that automatically chooses the appropriate implementation for the architecture, or for a feature-configurable binary that can be customized at build time.

Example

# myapp/BUILD

cc_binary(
    name = "mybinary",
    srcs = ["main.cc"],
    deps = select({
        ":arm_build": [":arm_lib"],
        ":x86_debug_build": [":x86_dev_lib"],
        "//conditions:default": [":generic_lib"],
    }),
)

config_setting(
    name = "arm_build",
    values = {"cpu": "arm"},
)

config_setting(
    name = "x86_debug_build",
    values = {
        "cpu": "x86",
        "compilation_mode": "dbg",
    },
)

This declares a cc_binary that “chooses” its deps based on the flags at the command line. Specficially, deps becomes:

Command deps =
bazel build //myapp:mybinary --cpu=arm [":arm_lib"]
bazel build //myapp:mybinary -c dbg --cpu=x86 [":x86_dev_lib"]
bazel build //myapp:mybinary --cpu=ppc [":generic_lib"]
bazel build //myapp:mybinary -c dbg --cpu=ppc [":generic_lib"]

select() serves as a placeholder for a value that will be chosen based on configuration conditions. These conditions are labels that refer to config_setting targets. By using select() in a configurable attribute, the attribute effectively takes on different values when different conditions hold.

Matches must be unambiguous: either exactly one condition must match or, if multiple conditions match, one’s values must be a strict superset of all others’. For example, values = {"cpu": "x86", "compilation_mode": "dbg"} is an unambiguous specialization of values = {"cpu": "x86"}. The built-in condition //conditions:default automatically matches when nothing else does.

This example uses deps. But select() works just as well on srcs, resources, cmd, or practically any other attribute. Only a small number of attributes are non-configurable, and those are clearly annotated; for instance, config_setting’s own values attribute is non-configurable.

Certain attributes, like the tools of a genrule, have the effect of changing the build parameters (such as the cpu) for all targets that transitively appear beneath them. This will affect how conditions are matched within those targets but not within the attribute that causes the change. That is, a select in the tools attribute of a genrule will work the same as a select in the srcs.

Configuration Conditions

Each key in a configurable attribute is a label reference to a config_setting target. This is just a collection of expected command line flag settings. By encapsulating these in a target, it’s easy to maintain “standard” conditions that can be referenced across targets and BUILD files.

The core config_setting syntax is:

config_setting(
    name = "meaningful_condition_name",
    values = {
        "flag1": "expected_value1",
        "flag2": "expected_value2",
        ...
    },
)

flagN is an arbitrary Bazel command line flag. value is the expected value for that flag. A config_setting matches when all of its flags match (order is irrelevant).

values entries use the same parsing logic as at the actual command line. This means:

  • values = { "compilation_mode": "opt" } matches bazel build -c opt ...
  • values = { "java_header_compilation": "true" } matches bazel build --java_header_compilation=1 ...
  • values = { "java_header_compilation": "0" } matches bazel build --nojava_header_compilation ...

config_setting only works with flags that affect build rule output. For example, --show_progress isn’t allowed because this only affects how Bazel reports progress to the user.

config_setting semantics are intentionally simple. For example, there’s no direct support for OR chaining (although a convenience function provides this). Consider writing macros for complicated flag logic.

Defaults

The built-in condition //conditions:default matches when no other condition matches.

Because of the “exactly one match” rule, a configurable attribute with no match and no default condition triggers a "no matching conditions" error. This can protect against silent failures from unexpected build flags:

# foo/BUILD

config_setting(
    name = "foobar",
    values = {"define": "foo=bar"},
)

cc_library(
    name = "my_lib",
    srcs = select({
        ":foobar": ["foobar_lib.cc"],
    }),
)
$ bazel build //foo:my_lib --define foo=baz
ERROR: Configurable attribute "srcs" doesn't match this configuration (would
a default condition help?).
Conditions checked:
  //foo:foobar

select() can include a no_match_error for custom failure messages.

Custom Keys

Since config_setting currently only supports built-in Bazel flags, the level of custom conditioning it can support is limited. For example, there’s no Bazel flag for IncludeSpecialProjectFeatureX.

Plans for truly custom flags are underway. In the meantime, --define is the best approach for these purposes. --define is a bit awkward to use and wasn’t originally designed for this purpose. We recommend using it sparingly until true custom flags are available. For example, don’t use --define to specify multiple variants of top-level binary. Just use multiple targets instead.

To trigger an arbitrary condition with --define, write

config_setting(
    name = "bar",
    values = {"define": "foo=bar"},
)

config_setting(
    name = "baz",
    values = {"define": "foo=baz"},
)

and run $ bazel build //my:target --define foo=baz.

The values attribute can’t contain multiple defines. This is because each instance has the same dictionary key. To solve this, use define_values:

config_setting(
    name = "bar_and_baz",
    define_values = {
        "foo": "bar",  # matches --define foo=bar
        "baz": "bat",  # matches --define baz=bat
    },
)

When defines appear in both values and define_values, all must match for the config_setting to match.

Platforms

While the ability to specify multiple flags on the command line provides flexibility, it can also be burdensome to individually set each one every time you want to build a target. Platforms allow you to consolidate these into simple bundles.

# myapp/BUILD

sh_binary(
    name = "my_rocks",
    srcs = select({
        ":basalt": ["pyroxene.sh"],
        ":marble": ["calcite.sh"],
        "//conditions:default": ["feldspar.sh"],
    }),
)

config_setting(
    name = "basalt",
    constraint_values = [
        ":black",
        ":igneous",
    ],
)

config_setting(
    name = "marble",
    constraint_values = [
        ":white",
        ":metamorphic",
    ],
)

# constraint_setting acts as an enum type, and constraint_value as an enum value.
constraint_setting(name = "color")
constraint_value(name = "black", constraint_setting = "color")
constraint_value(name = "white", constraint_setting = "color")
constraint_setting(name = "texture")
constraint_value(name = "smooth", constraint_setting = "texture")
constraint_setting(name = "type")
constraint_value(name = "igneous", constraint_setting = "type")
constraint_value(name = "metamorphic", constraint_setting = "type")

platform(
    name = "basalt_platform",
    constraint_values = [
        ":black",
        ":igneous",
    ],
)

platform(
    name = "marble_platform",
    constraint_values = [
        ":white",
        ":smooth",
        ":metamorphic",
    ],
)

The platform can be specified on the command line. It activates the config_settings that contain a subset of the platform’s constraint_values, allowing those config_settings to match in select() expressions.

For example, in order to set the srcs attribute of my_rocks to calcite.sh, we can simply run

bazel build //my_app:my_rocks --platforms=//myapp:marble_platform

Without platforms, this might look something like

bazel build //my_app:my_rocks --define color=white --define texture=smooth --define type=metamorphic

Platforms are still under development. See the documentation and roadmap for details.

Short Keys

Since configuration keys are target labels, their names can get long and unwieldy. This can be mitigated with local variable definitions:

Before:

sh_binary(
    name = "my_target",
    srcs = select({
        "//my/project/my/team/configs:config1": ["my_target_1.sh"],
        "//my/project/my/team/configs:config2": ["my_target_2.sh"],
    }),
)

After:

CONFIG1="//my/project/my/team/configs:config1"
CONFIG2="//my/project/my/team/configs:config2"

sh_binary(
    name = "my_target",
    srcs = select({
        CONFIG1: ["my_target_1.sh"],
        CONFIG2: ["my_target_2.sh"],
    })
)

For more complex expressions, you can use macros:

Before:

# foo/BUILD

genrule(
    name = "my_target",
    srcs = [],
    outs = ["my_target.out"],
    cmd = select({
        "//my/project/my/team/configs/config1": "echo custom val: this > $@",
        "//my/project/my/team/configs/config2": "echo custom val: that > $@",
        "//conditions:default": "echo default output > $@",
    }),
)

After:

# foo/genrule_select.bzl

def select_echo(input_dict):
    echo_cmd = "echo %s > $@"
    out_dict = {"//conditions:default": echo_cmd % "default output"}
    for (key, val) in input_dict.items():
        cmd = echo_cmd % ("custom val: " + val)
        out_dict["//my/project/my/team/configs/config" + key] = cmd
    return select(out_dict)
# foo/BUILD

load("//foo:genrule_select.bzl", "select_echo")

genrule(
    name = "my_target",
    srcs = [],
    outs = ["my_target.out"],
    cmd = select_echo({
        "1": "this",
        "2": "that",
    }),
)

Multiple Selects

select can appear multiple times in the same attribute:

sh_binary(
    name = "my_target",
    srcs = ["always_include.sh"] +
           select({
               ":armeabi_mode": ["armeabi_src.sh"],
               ":x86_mode": ["x86_src.sh"],
           }) +
           select({
               ":opt_mode": ["opt_extras.sh"],
               ":dbg_mode": ["dbg_extras.sh"],
           }),
)

select cannot appear inside another select (i.e. AND chaining). If you need to AND selects together, either use an intermediate target:

sh_binary(
    name = "my_target",
    srcs = ["always_include.sh"],
    deps = select({
        ":armeabi_mode": [":armeabi_lib"],
        ...
    }),
)

sh_library(
    name = "armeabi_lib",
    srcs = select({
        ":opt_mode": ["armeabi_with_opt.sh"],
        ...
    }),
)

or write a macro to do the same thing automatically.

This approach doesn’t work for non-deps attributes (like genrule:cmd). For these, extra config_settings may be necessary:

config_setting(
    name = "armeabi_and_opt",
    values = {
        "cpu": "armeabi",
        "compilation_mode": "opt",
    },
)

OR Chaining

Consider the following:

sh_binary(
    name = "my_target",
    srcs = ["always_include.sh"],
    deps = select({
        ":config1": [":standard_lib"],
        ":config2": [":standard_lib"],
        ":config3": [":standard_lib"],
        ":config4": [":special_lib"],
    }),
)

Most conditions evaluate to the same dep. But this syntax is verbose, hard to maintain, and refactoring-unfriendly. It would be nice to not have to repeat [":standard_lib"] over and over.

One option is to predefine the declaration as a BUILD variable:

STANDARD_DEP = [":standard_lib"]

sh_binary(
    name = "my_target",
    srcs = ["always_include.sh"],
    deps = select({
        ":config1": STANDARD_DEP,
        ":config2": STANDARD_DEP,
        ":config3": STANDARD_DEP,
        ":config4": [":special_lib"],
    }),
)

This makes it easier to manage the dependency. But it still adds unnecessary duplication.

select() doesn’t support native syntax for ORed conditions. For this, use one of the following:

selects.with_or

The Skylib utility selects defines a Starlark macro that emulates OR behavior:

load("@bazel_skylib//:lib.bzl", "selects")
sh_binary(
    name = "my_target",
    srcs = ["always_include.sh"],
    deps = selects.with_or({
        (":config1", ":config2", ":config3"): [":standard_lib"],
        ":config4": [":special_lib"],
    }),
)

This automatically expands the select to the original syntax above.

config_settingAliasing

If you’d like to OR conditions under a proper config_setting that any rule can reference, you can use a selectable alias that matches any of the desired conditions:

alias(
    name = "config1_or_2_or_3",
    actual = select({
        # When the build matches :config1, this alias *becomes* :config1.
        # So it too matches by definition. The same applies for :config2
        # and :config3.
        ":config1": ":config1",
        ":config2": ":config2",
        ":config3": ":config3",
        # The default condition represents this alias "not matching" (i.e.
        # none of the conditions that we care about above match). In this
        # case, bind the alias to any of those conditions. By definition
        # it won't match.
        "//conditions:default": ":config2", # Arbitrarily chosen from above.
    }),
)

sh_binary(
    name = "my_target",
    srcs = ["always_include.sh"],
    deps = select({
        ":config1_or_2_or_3": [":standard_lib"],
        ":config4": [":special_lib"],
    }),
)

Unlike selects.with_or, different rules can select on :config1_or_2_or_3 with different values.

Note that it’s an error for multiple conditions to match unless one is a “specialization” of the other. See select() documentation for details.

For AND chaining, see here.

Custom Error Messages

By default, when no condition matches, the owning target fails with the error:

ERROR: Configurable attribute "deps" doesn't match this configuration (would
a default condition help?).
Conditions checked:
  //tools/cc_target_os:darwin
  //tools/cc_target_os:android

This can be customized with no_match_error:

cc_library(
    name = "my_lib",
    deps = select(
        {
            "//tools/cc_target_os:android": [":android_deps"],
            "//tools/cc_target_os:windows": [":windows_deps"],
        },
        no_match_error = "Please build with an Android or Windows toolchain",
    ),
)
$ bazel build //foo:my_lib
ERROR: Configurable attribute "deps" doesn't match this configuration: Please
build with an Android or Windows toolchain

Rules Compatibility

Rule implementations receive the resolved values of configurable attributes. For example, given:

# myproject/BUILD

some_rule(
    name = "my_target",
    some_attr = select({
        ":foo_mode": [":foo"],
        ":bar_mode": [":bar"],
    }),
)
$ bazel build //myproject/my_target --define mode=foo

Rule implementation code sees ctx.attr.some_attr as [":foo"].

Macros can accept select() clauses and pass them through to native rules. But they cannot directly manipulate them. For example, there’s no way for a macro to convert

select({"foo": "val"}, ...)

to

select({"foo": "val_with_suffix"}, ...)

This is for two reasons.

First, macros that need to know which path a select will choose cannot work because macros are evaluated in Bazel’s loading phase, which occurs before flag values are known. This is a core Bazel design restriction that’s unlikely to change any time soon.

Second, macros that just need to iterate over all select paths, while technically feasible, lack a coherent UI. Further design is necessary to change this.

Bazel Query and Cquery

Bazel query operates over Bazel’s loading phase. This means it doesn’t know what command line flags will be applied to a target since those flags aren’t evaluated until later in the build (during the analysis phase). So the query command can’t accurately determine which path a configurable attribute will follow.

Bazel cquery has the advantage of being able to parse build flags and operating post-analysis phase so it correctly resolves configurable attributes. It doesn’t have full feature parity with query but supports most major functionality and is actively being worked on. Querying the following build file…

# myproject/BUILD

cc_library(
    name = "my_lib",
    deps = select({
        ":long": [":foo_dep"],
        ":short": [":bar_dep"],
    }),
)

config_setting(
    name = "long",
    values = {"define": "dog=dachshund"},
)

config_setting(
    name = "short",
    values = {"define": "dog=pug"},
)

…would return the following results.

$ bazel query 'deps(//myproject:my_lib)'
//myproject:my_lib
//myproject:foo_dep
//myproject:bar_dep

$ bazel cquery 'deps(//myproject:my_lib)' --define dog=pug
//myproject:my_lib
//myproject:bar_dep

FAQ

Why doesn’t select() work in macros?

select() does work in rules! See Rules compatibility for details.

The key issue this question usually means is that select() doesn’t work in macros. These are different than rules. See the documentation on rules and macros to understand the difference. Here’s an end-to-end example:

Define a rule and macro:

# myproject/defs.bzl

# Rule implementation: when an attribute is read, all select()s have already
# been resolved. So it looks like a plain old attribute just like any other.
def _impl(ctx):
    name = ctx.attr.name
    allcaps = ctx.attr.my_config_string.upper()  # This works fine on all values.
    print("My name is " + name + " with custom message: " + allcaps)

# Rule declaration:
my_custom_bazel_rule = rule(
    implementation = _impl,
    attrs = {"my_config_string": attr.string()},
)

# Macro declaration:
def my_custom_bazel_macro(name, my_config_string):
    allcaps = my_config_string.upper()  # This line won't work with select(s).
    print("My name is " + name + " with custom message: " + allcaps)

Instantiate the rule and macro:

# myproject/BUILD

load("//myproject:defs.bzl", "my_custom_bazel_rule")
load("//myproject:defs.bzl", "my_custom_bazel_macro")

my_custom_bazel_rule(
    name = "happy_rule",
    my_config_string = select({
        "//tools/target_cpu:x86": "first string",
        "//tools/target_cpu:ppc": "second string",
    }),
)

my_custom_bazel_macro(
    name = "happy_macro",
    my_config_string = "fixed string",
)

my_custom_bazel_macro(
    name = "sad_macro",
    my_config_string = select({
        "//tools/target_cpu:x86": "first string",
        "//tools/target_cpu:ppc": "other string",
    }),
)

Building fails because sad_macro can’t process the select():

$ bazel build //myproject:all
ERROR: /myworkspace/myproject/BUILD:17:1: Traceback
  (most recent call last):
File "/myworkspace/myproject/BUILD", line 17
my_custom_bazel_macro(name = "sad_macro", my_config_stri..."}))
File "/myworkspace/myproject/defs.bzl", line 4, in
  my_custom_bazel_macro
my_config_string.upper()
type 'select' has no method upper().
ERROR: error loading package 'myproject': Package 'myproject' contains errors.

Building succeeds when we comment out sad_macro:

# Comment out sad_macro so it doesn't mess up the build.
$ bazel build //myproject:all
DEBUG: /myworkspace/myproject/defs.bzl:5:3: My name is happy_macro with custom message: FIXED STRING.
DEBUG: /myworkspace/myproject/hi.bzl:15:3: My name is happy_rule with custom message: FIRST STRING.

This is impossible to change because by definition macros are evaluated before Bazel reads the build’s command line flags. That means there isn’t enough information to evaluate select()s.

Macros can, however, pass select()s as opaque blobs to rules:

# myproject/defs.bzl

def my_custom_bazel_macro(name, my_config_string):
    print("Invoking macro " + name)
    my_custom_bazel_rule(
        name = name + "_as_target",
        my_config_string = my_config_string,
    )
$ bazel build //myproject:sad_macro_less_sad
DEBUG: /myworkspace/myproject/defs.bzl:23:3: Invoking macro sad_macro_less_sad.
DEBUG: /myworkspace/myproject/defs.bzl:15:3: My name is sad_macro_less_sad with custom message: FIRST STRING.

Why does select() always return true?

Because macros (but not rules) by definition can’t evaluate select(s), any attempt to do so usually produces a an error:

ERROR: /myworkspace/myproject/BUILD:17:1: Traceback
  (most recent call last):
File "/myworkspace/myproject/BUILD", line 17
my_custom_bazel_macro(name = "sad_macro", my_config_stri..."}))
File "/myworkspace/myproject/defs.bzl", line 4, in
  my_custom_bazel_macro
my_config_string.upper()
type 'select' has no method upper().

Booleans are a special case that fail silently, so you should be particularly vigilant with them:

$ cat myproject/defs.bzl
def my_boolean_macro(boolval):
  print("TRUE" if boolval else "FALSE")

$ cat myproject/BUILD
load("//myproject:defs.bzl", "my_boolean_macro")
my_boolean_macro(
    boolval = select({
        "//tools/target_cpu:x86": True,
        "//tools/target_cpu:ppc": False,
    }),
)

$ bazel build //myproject:all --cpu=x86
DEBUG: /myworkspace/myproject/defs.bzl:4:3: TRUE.
$ bazel build //myproject:all --cpu=ppc
DEBUG: /myworkspace/myproject/defs.bzl:4:3: TRUE.

This happens because macros don’t understand the contents of select(). So what they’re really evaluting is the select() object itself. According to Pythonic design standards, all objects aside from a very small number of exceptions automatically return true.

Can I read select() like a dict?

Fine. Macros can’t evaluate select(s) because macros are evaluated before Bazel knows what the command line flags are.

Can macros at least read the select()’s dictionary, say, to add an extra suffix to each branch?

Conceptually this is possible. But this isn’t yet implemented and is not currently prioritized. What you can do today is prepare a straight dictionary, then feed it into a select():

$ cat myproject/defs.bzl
def selecty_genrule(name, select_cmd):
  for key in select_cmd.keys():
    select_cmd[key] += " WITH SUFFIX"
  native.genrule(
      name = name,
      outs = [name + ".out"],
      srcs = [],
      cmd = "echo " + select(select_cmd + {"//conditions:default": "default"})
        + " > $@"
  )

$ cat myproject/BUILD
selecty_genrule(
    name = "selecty",
    select_cmd = {
        "//tools/target_cpu:x86": "x86 mode",
    },
)

$ bazel build //testapp:selecty --cpu=x86 && cat bazel-genfiles/testapp/selecty.out
x86 mode WITH SUFFIX

If you’d like to support both select() and native types, you can do this:

$ cat myproject/defs.bzl
def selecty_genrule(name, select_cmd):
    cmd_suffix = ""
    if type(select_cmd) == "string":
        cmd_suffix = select_cmd + " WITH SUFFIX"
    elif type(select_cmd) == "dict":
        for key in select_cmd.keys():
            select_cmd[key] += " WITH SUFFIX"
        cmd_suffix = select(select_cmd + {"//conditions:default": "default"})

    native.genrule(
        name = name,
        outs = [name + ".out"],
        srcs = [],
        cmd = "echo " + cmd_suffix + "> $@",
    )

Why doesn’t select() work with bind()?

Because bind() is a WORKSPACE rule, not a BUILD rule.

Workspace rules do not have a specific configuration, and aren’t evaluated in the same way as BUILD rules. Therefore, a select() in a bind() can’t actually evaluate to any specific branch.

Instead, you should use alias(), with a select() in the actual attribute, to perform this type of run-time determination. This works correctly, since alias() is a BUILD rule, and is evaluated with a specific configuration.

You can even have a bind() target point to an alias(), if needed.

$ cat WORKSPACE
workspace(name = "myproject")
bind(name = "openssl", actual = "//:ssl")
http_archive(name = "alternative", ...)
http_archive(name = "boringssl", ...)

$ cat BUILD
config_setting(
    name = "alt_ssl",
    define_values = {
        "ssl_library": "alternative",
    },
)

alias(
    name = "ssl",
    actual = select({
        "//:alt_ssl": "@alternative//:ssl",
        "//conditions:default": "@boringssl//:ssl",
    }),
)

With this setup, you can pass --define ssl_library=alternative, and any target that depends on either //:ssl or //external:ssl will see the alternative located at @alternative//:ssl.