Configurable Query (cquery)

cquery is a variant of query that correctly handles select() and build options' effects on the build graph.

It achieves this by running over the results of Bazel's analysis phase, which integrates these effects. query, by constrast, runs over the results of Bazel's loading phase, before options are evaluated.

For example:

$ cat > tree/BUILD <<EOF
sh_library(
    name = "ash",
    deps = select({
        ":excelsior": [":manna-ash"],
        ":americana": [":white-ash"],
        "//conditions:default": [":common-ash"],
    }),
)
sh_library(name = "manna-ash")
sh_library(name = "white-ash")
sh_library(name = "common-ash")
config_setting(
    name = "excelsior",
    values = {"define": "species=excelsior"},
)
config_setting(
    name = "americana",
    values = {"define": "species=americana"},
)
EOF
# Traditional query: query doesn't know which select() branch will be chosen
# so it conservatively lists all of them.
$ bazel query "deps(//tree:ash)" --noimplicit_deps
//tree:ash
//tree:white-ash
//tree:manna-ash
//tree:common-ash

# cquery: cquery lets you set build options at the command line and chooses
# the exact dependencies that implies.
$ bazel cquery "deps(//tree:ash)" --define species=excelsior --noimplicit_deps
//tree:ash (9f87702)
//tree:manna-ash (9f87702)

Each result includes a unique identifier (9f87702) of the configuration the target is built with.

Since cquery runs over the configured target graph. it doesn't have insight into artifacts like build actions nor access to test_suite rules as they are not configured targets. For the former, see aquery.

Basic syntax

A simple cquery call looks like:

bazel cquery "function(//target)"

The query expression "function(//target)" consists of the following:

  • function(...) is the function to run on the target. cquery supports most of query's functions, plus a few new ones.
  • //target is the expression fed to the function. In this example, the expression is a simple target. But the query language also allows nesting of functions. See the Query How-To for examples.

cquery requires a target to run through the loading and analysis phases. Unless otherwise specified, cquery parses the target(s) listed in the query expression. See --universe_scope for querying dependencies of top-level build targets.

Configurations

The line:

//tree:ash (9f87702)

means //tree:ash was built in a configuration with ID 9f87702. For most targets, this is an opaque hash of the the build option values defining the configuration.

To see the configuration's complete contents, run:

$ bazel config 9f87702

The host configuration uses the special ID (HOST). Non-generated source files, like those commonly found in srcs, use the special ID (null) (because they don't need to be configured).

9f87702 is a prefix of the complete ID. This is because complete IDs are SHA-256 hashes, which are long and hard to follow. cquery understands any valid prefix of a complete ID, similar to Git short hashes. To see complete IDs, run $ bazel config.

Target pattern evaluation

//foo has a different meaning for cquery than for query. This is because cquery evaluates configured targets and the build graph may have multiple configured versions of //foo.

For cquery, a target pattern in the query expression evaluates to every configured target with a label that matches that pattern. Output is deterministic, but cquery makes no ordering guarantee beyond the core query ordering contract.

This produces subtler results for query expressions than with query. For example, the following can produce multiple results:

# Analyzes //foo in the target configuration, but also analyzes
# //genrule_with_foo_as_tool which depends on a host-configured
# //foo. So there are two configured target instances of //foo in
# the build graph.
$ bazel cquery //foo --universe_scope=//foo,//genrule_with_foo_as_tool
//foo (9f87702)
//foo (HOST)
  

If you want to precisely declare which instance to query over, use the config function.

See query's target pattern documentation for more information on target patterns.

Functions

Of the set of functions supported by query, cquery supports all but visible, siblings, buildfiles, and tests.

cquery also introduces the following new functions:

config

expr ::= config(expr, word)

The config operator attempts to find the configured target for the label denoted by the first argument and configuration specified by the second argument.

Valid values for the second argument are target, host, null, or a custom configuration hash. Hashes can be retrieved from $ bazel config or a prevous cquery's output.

Examples:

$ bazel cquery "config(//bar, host)" --universe_scope=//foo
$ bazel cquery "deps(//foo)"
//bar (HOST)
//baz (3732cc8)

$ bazel cquery "config(//baz, 3732cc8)"

If not all results of the first argument can be found in the specified configuration, only those that can be found are returned. If no results can be found in the specified configuration, the query fails.

Options

Build options

cquery runs over a regular Bazel build and thus inherits the set of options available during a build.

Using cquery options

--universe_scope (comma-separated list)

Often, the dependencies of configured targets go through transitions, which causes their configuration to differ from their dependent. This flag allows you to query a target as if it were built as a dependency or a transitive dependency of another target. For example:

# x/BUILD
genrule(
     name = "my_gen",
     srcs = ["x.in"],
     outs = ["x.cc"],
     cmd = "$(locations :tool) $< >$@",
     tools = [":tool"],
)
cc_library(
    name = "tool",
)

Genrules configure their tools in the host configuration so the following queries would produce the following outputs:

Query Target Built Output
bazel cquery "//x:tool" //x:tool //x:tool(targetconfig)
bazel cquery "//x:tool" --universe_scope="//x:my_gen" //x:my_gen //x:tool(hostconfig)

If this flag is set, its contents are built. If it's not set, all targets mentioned in the query expression are built instead. The transitive closure of the built targets are used as the universe of the query. Either way, the targets to be built must be buildable at the top level (that is, compatible with top-level options). cquery returns results in the transitive closure of these top-level targets.

Even if it's possible to build all targets in a query expression at the top level, it may be beneficial to not do so. For example, explicitly setting --universe_scope could prevent building targets multiple times in configurations you don't care about. It could also help specify which configuration version of a target you're looking for (since it's not currently possible to fully specify this any other way). You should set this flag if your query expression is more complex than deps(//foo).

--implicit_deps (boolean, default=True)

Setting this flag to false filters out all results that aren't explicitly set in the BUILD file and instead set elsewhere by Bazel. This includes filtering resolved toolchains.

--tool_deps (boolean, default=True)

Setting this flag to false filters out all configured targets for which the path from the queried target to them crosses a transition between the target configuration and the non-target configurations. If the queried target is in the target configuration, setting --notool_deps will only return targets that also are in the target configuration. If the queried target is in a non-target configuration, setting --notool_deps will only return targets also in non-target configurations. This setting generally does not affect filtering of resolved toolchains.

--include_aspects (boolean, default=True)

Aspects can add additional dependencies to a build. By default, cquery doesn't follow aspects because they make the queryable graph bigger, which uses more memory. But following them produces more accurate results.

If you're not worried about the memory impact of large queries, enable this flag by default in your bazelrc.

If you query with aspects disabled, you can experience a problem where target X fails while building target Y but cquery somepath(Y, X) and cquery deps(Y) | grep 'X' return no results because the dependency occurs through an aspect.

Output formats

By default, cquery outputs results in a dependency-ordered list of label and configuration pairs. There are other options for exposing the results as well.

Transitions

--transitions=lite
--transitions=full

Configuration transitions are used to build targets underneath the top level targets in different configurations than the top level targets.

For example, a target might impose a transition to the host configuration on all dependencies in its tools attribute. These are known as attribute transitions. Rules can also impose transitions on their own configurations, known as rule class transitions. This output format outputs information about these transitions such as what type they are and the effect they have on build options.

This output format is triggered by the --transitions flag which by default is set to NONE. It can be set to FULL or LITE mode. FULL mode outputs information about rule class transitions and attribute transitions including a detailed diff of the options before and after the transition. LITE mode outputs the same information without the options diff.

Protocol message output

--output=proto

This option causes the resulting targets to be printed in a binary protocol buffer form. The definition of the protocol buffer can be found at src/main/protobuf/analysis.proto.

--[no]proto:include_configurations

By default, cquery results return configuration information as part of each configured target. If you'd like to omit this information and get proto output that is formatted exactly like query's proto output, set this flag to false .

See query's proto output documentation for more proto output-related options.

Note: while selects are resolved both at the top level of returned targets and within attributes, all possible inputs for selects are still included as rule_input fields.

Graph output

--output=graph

This option generates output as a Graphviz-compatible .dot file. See query's graph output documentation for details. cquery also supports --graph:node_limit and --graph:factored.

Files output

--output=files

This option prints a list of the output files produced by each target matched by the query similar to the list printed at the end of a bazel build invocation. The output contains only the files advertised in the requested output groups as determined by the --output_groups flag and never contains source files. Note: The output of bazel cquery --output=files //pkg:foo contains the output files of //pkg:foo in all configurations that occur in the build (also see the section on target pattern evaluation. If that is not desired, wrap you query in config(..., target).

Defining the output format using Starlark

--output=starlark

This output format calls a Starlark function for each configured target in the query result, and prints the value returned by the call. The --starlark:file flag specifies the location of a Starlark file that defines a function named format with a single parameter, target. This function is called for each Target in the query result. Alternatively, for convenience, you may specify just the body of a function declared as def format(target): return expr by using the --starlark:expr flag.

'cquery' Starlark dialect

The cquery Starlark environment differs from a BUILD or .bzl file. It includes all core Starlark built-in constants and functions, plus a few cquery-specific ones described below, but not (for example) glob, native, or rule, and it does not support load statements.

build_options(target)

build_options(target) returns a map whose keys are build option identifiers (see Configurations) and whose values are their Starlark values. Build options whose values are not legal Starlark values are omitted from this map.

If the target is an input file, build_options(target) returns None, as input file targets have a null configuration.

providers(target)

providers(target) returns a map whose keys are names of providers (for example, "DefaultInfo") and whose values are their Starlark values. Providers whose values are not legal Starlark values are omitted from this map.

Examples

Print a space-separated list of the base names of all files produced by //foo:

  bazel cquery //foo --output=starlark \
    --starlark:expr="' '.join([f.basename for f in target.files.to_list()])"

Print a space-separated list of the paths of all files produced by rule targets in //bar and its subpackages:

  bazel cquery 'kind(rule, //bar/...)' --output=starlark \
    --starlark:expr="' '.join([f.path for f in target.files.to_list()])"

Print a list of the mnemonics of all actions registered by //foo.

  bazel cquery //foo --output=starlark \
    --starlark:expr="[a.mnemonic for a in target.actions]"

Print a list of compilation outputs registered by a cc_library //baz.

  bazel cquery //baz --output=starlark \
    --starlark:expr="[f.path for f in target.output_groups.compilation_outputs.to_list()]"

Print the value of the command line option --javacopt when building //foo.

  bazel cquery //foo --output=starlark \
    --starlark:expr="build_options(target)['//command_line_option:javacopt']"

Print the label of each target with exactly one output. This example uses Starlark functions defined in a file.

  $ cat example.cquery

  def has_one_output(target):
    return len(target.files.to_list()) == 1

  def format(target):
    if has_one_output(target):
      return target.label
    else:
      return ""


  $ bazel cquery //baz --output=starlark --starlark:file=example.cquery

Print the label of each target which is strictly Python 3. This example uses Starlark functions defined in a file.

  $ cat example.cquery

  def format(target):
    p = providers(target)
    py_info = p.get("PyInfo")
    if py_info and py_info.has_py3_only_sources:
      return target.label
    else:
      return ""


  $ bazel cquery //baz --output=starlark --starlark:file=example.cquery

Extract a value from a user defined Provider.

  $ cat some_package/my_rule.bzl

  MyRuleInfo = provider(fields={"color": "the name of a color"})

  def _my_rule_impl(ctx):
      ...
      return [MyRuleInfo(color="red")]

  my_rule = rule(
      implementation = _my_rule_impl,
      attrs = {...},
  )

  $ cat example.cquery

  def format(target):
    p = providers(target)
    my_rule_info = p.get("//some_package:my_rule.bzl%MyRuleInfo'")
    if my_rule_info:
      return my_rule_info.color
    return ""


  $ bazel cquery //baz --output=starlark --starlark:file=example.cquery

cquery vs. query

cquery and query complement each other and excel in different niches. Consider the following to decide which is right for you:

  • cquery follows specific select() branches to model the exact graph you build. query doesn't know which branch the build chooses, so overapproximates by including all branches.
  • cquery's precision requires building more of the graph than query does. Specifically, cquery evaluates configured targets while query only evaluates targets. This takes more time and uses more memory.
  • cquery's intepretation of the query language introduces ambiguity that query avoids. For example, if "//foo" exists in two configurations, which one should cquery "deps(//foo)" use? The config function can help with this.
  • As a newer tool, cquery lacks support for certain use cases. See Known issues for details.

Known issues

  • All targets that cquery "builds" must have the same configuration.

    Before evaluating queries, cquery triggers a build up to just before the point where build actions would execute. The targets it "builds" are by default selected from all labels that appear in the query expression (this can be overridden with --universe_scope). These must have the same configuration.

    While these generally share the top-level "target" configuration, rules can change their own configuration with incoming edge transitions. This is where cquery falls short.

    Workaround: If possible, set --universe_scope to a stricter scope. For example:

    # This command attempts to build the transitive closures of both //foo and
    # //bar. //bar uses an incoming edge transition to change its --cpu flag.
    $ bazel cquery 'somepath(//foo, //bar)'
    ERROR: Error doing post analysis query: Top-level targets //foo and //bar
    have different configurations (top-level targets with different
    configurations is not supported)
    
    # This command only builds the transitive closure of //foo, under which
    # //bar should exist in the correct configuration.
    $ bazel cquery 'somepath(//foo, //bar)' --universe_scope=//foo
       
  • No support for --output=xml.
  • Non-deterministic output.

    cquery does not automatically wipe the build graph from previous commands and is therefore prone to picking up results from past queries. For example, genquery exerts a host transition on its tools attribute - that is, it configures its tools in the host configuration.

    You can see the lingering effects of that transition below.

    $ cat > foo/BUILD <<<EOF
    genrule(
        name = "my_gen",
        srcs = ["x.in"],
        outs = ["x.cc"],
        cmd = "$(locations :tool) $< >$@",
        tools = [":tool"],
    )
    cc_library(
        name = "tool",
    )
    EOF
    
    $ bazel cquery "//foo:tool"
    tool(target_config)
    
    $ bazel cquery "deps(//foo:my_gen)"
    my_gen (target_config)
    tool (host_config)
    ...
    
    $ bazel cquery "//foo:tool"
    tool(host_config)
    

    Workaround: change any startup option to force re-analysis of configured targets. For example, add --test_arg=<whatever> to your build command.

Troubleshooting

Recursive target patterns (/...)

If you encounter:

$ bazel cquery --universe_scope=//foo:app "somepath(//foo:app, //foo/...)"
ERROR: Error doing post analysis query: Evaluation failed: Unable to load package '[foo]'
because package is not in scope. Check that all target patterns in query expression are within the
--universe_scope of this query.

this incorrectly suggests package //foo isn't in scope even though --universe_scope=//foo:app includes it. This is due to design limitations in cquery. As a workaround, explicitly include //foo/... in the universe scope:

$ bazel cquery --universe_scope=//foo:app,//foo/... "somepath(//foo:app, //foo/...)"

If that doesn't work (for example, because some target in //foo/... can't build with the chosen build flags), manually unwrap the pattern into its constituent packages with a pre-processing query:

# Replace "//foo/..." with a subshell query call (not cquery!) outputting each package, piped into
# a sed call converting "<pkg>" to "//<pkg>:*", piped into a "+"-delimited line merge.
# Output looks like "//foo:*+//foo/bar:*+//foo/baz".
#
$  bazel cquery --universe_scope=//foo:app "somepath(//foo:app, $(bazel query //foo/...
--output=package | sed -e 's/^/\/\//' -e 's/$/:*/' | paste -sd "+" -))"