Documentation for Bazel master
Bazel is still in Beta and new releases may include backward incompatible changes. As we make changes and polish the extension mechanism, old features may be removed and new features that are not backward compatible may be added.
Backward incompatible changes are introduced gradually:
false.true. You
can still use the flag to disable the change.To check if your code will be compatible with future releases you can:
--all_incompatible_changes. This flag
enables all backward incompatible changes, and so you can ensure your code
is compatible with upcoming changes.The following are the backward incompatible changes that are implemented and guarded behind flags in the current release:
General Starlark
Starlark Rules
cfg = "data"Objc
External repositories
Java
Misc
C++
We are removing the + operator on dictionaries. This includes the += form
where the left-hand side is a dictionary. This is done to improve compatibility
with Python. A possible workaround is to use the .update method instead.
--incompatible_disallow_dict_plusfalsePreviously, the load statement could appear anywhere in a .bzl file so long
as it was at the top level. With this change, for .bzl files, load must
appear at the beginning of the file, i.e. before any other non-load statement.
--incompatible_bzl_disallow_load_after_statementfalseWhen the flag is set to true, depset objects are not treated as iterable. This
prohibits directly iterating over depsets in for loops, taking its size via
len(), and passing it to many functions such as list, tuple, min, max,
sorted, all, and any. It does not prohibit checking for emptiness by
converting the depset to a boolean.
The goal of this change is to avoid accidental iteration on depset, which can
be expensive. If you really need
to iterate over a depset, you can call the .to_list() method to obtain a
flattened list of its contents.
deps = depset()
[x.path for x in deps] # deprecated
[x.path for x in deps.to_list()] # recommended
sorted(deps) # deprecated
sorted(deps.to_list()) # recommended
--incompatible_depset_is_not_iterablefalseTo merge two sets, the following examples used to be supported, but are now deprecated:
depset1 + depset2 # deprecated
depset1 | depset2 # deprecated
depset1.union(depset2) # deprecated
The recommended solution is to use the depset constructor:
depset(transtive = [depset1, depset2])
See the depset documentation for more information.
--incompatible_depset_unionfalseWhen the flag is set to true, string objects are not treated as iterable. This
affects for loops and many functions, e.g. list, tuple, min, max,
sorted, all, and any. String iteration has been a source of errors and
confusion, such as this error:
def my_macro(name, srcs):
for src in srcs:
# do something with src
# equivalent to: my_macro("hello", ["f", "o", "o", ".", "c", "c"])
my_macro(
name = "hello",
srcs = "foo.cc",
)
String indexing and len are still allowed. If you need to iterate over a
string, you may explicitly use:
my_string = "hello world"
for i in range(len(my_string)):
char = my_string[i]
# do something with char
--incompatible_string_is_not_iterablefalse//Integer division operator is now // instead of /. This aligns with
Python 3 and it highlights the fact it is a floor division.
x = 7 / 2 # deprecated
x = 7 // 2 # x is 3
--incompatible_disallow_slash_operatortrueThe current package name should be retrieved by calling package_name() in
BUILD files or native.package_name() in .bzl files. The old way of referring
to the magic PACKAGE_NAME variable bends the language since it is neither a
parameter, local variable, nor global variable.
Likewise, the magic REPOSITORY_NAME variable is replaced by
repository_name() and native.repository_name(). Both deprecations use the
same flag.
--incompatible_package_name_is_a_functionfalseThe FileType function is going away. The main use-case was as an argument to the rule function. It’s no longer needed, you can simply pass a list of strings to restrict the file types the rule accepts.
--incompatible_disallow_filetypefalseThis change removes the old methods for registering actions within rules, and requires that you use the new methods instead. The deprecated methods and their replacements are as follows.
ctx.new_file(...) –> ctx.actions.declare_file(...)ctx.experimental_new_directory(...) –>
ctx.actions.declare_directory(...)ctx.action(...) –> either ctx.actions.run(...) or
ctx.actions.run_shell(...)ctx.file_action(...) –> ctx.actions.write(...)ctx.empty_action(...) –> ctx.actions.do_nothing(...)ctx.template_action(...) –> ctx.actions.expand_template(...)--incompatible_new_actions_apifalseThe Args object returned by ctx.actions.args() has dedicated
methods for appending the contents of a list or depset to the command line.
Previously these use cases were lumped into its add()
method, resulting in a more cluttered API.
With this flag, add() only works for scalar values, and its deprecated
parameters are disabled. To add many arguments at once you must use add_all()
or add_joined() instead.
--incompatible_disallow_old_style_args_addfalseWhen set, the result of range(...) function is a lazy range type instead of
a list. Because of this repetitions using * operator are no longer
supported and range slices are also lazy range instances.
--incompatible_range_typetrueThis flag disables certain deprecated resource fields on ObjcProvider.
--incompatible_objc_provider_resourcesfalseThis flag disables the output_group field on the Target Starlark type.
Use OutputGroupInfo instead.
For example, replace:
dep_bin = ctx.attr.dep.output_group.bin
with:
dep_bin = ctx.attr.dep[OutputGroupInfo].bin
--incompatible_no_target_output_groupfalseThis flag disables the default parameter on attr.output and
attr.output_list. Use Starlark macros to specify defaults for these attributes
instead.
For example, replace:
my_rule = rule(
...
attrs = {"out" : attr.output(default = "foo.txt")}
...
with:
# myrule.bzl
my_rule = rule(
...
attrs = {"out" : attr.output()}
...
# mymacro.bzl
load(":myrule.bzl", _my_rule = "my_rule")
def my_rule(name):
_my_rule(
name = name,
output = "%s_out.txt" % name
)
The previous default parameter of these attribute types was severely
bug-prone, as two targets of the same rule would be unable to exist in the same
package under default behavior. (Two targets both generating foo.txt in the
same package would conflict.)
--incompatible_no_output_attr_defaultfalseWhen set, the native git_repository and new_git_repository rules are
disabled. The Starlark versions
load("@bazel_tools//tools/build_defs/repo:git.bzl",
"git_repository", "new_git_repository")
should be used instead. These are drop-in replacements of the corresponding
native rules, however with the additional requirement that all label arguments
be provided as a fully qualified label (usually starting with @//),
for example: build_file = "@//third_party:repo.BUILD".
--incompatible_remove_native_git_repositorytrueWhen set, the native http_archive and all related rules are disabled.
The Starlark version
load("@bazel_tools//tools/build_defs/repo:http.bzl", "http_archive")
should be used instead. This is a drop-in replacement, however with the
additional requirement that all label arguments be provided as
fully qualified labels (usually starting with @//). The Starlark http_archive
is also a drop-in replacement for the native new_http_archive (with
the same proviso). http.bzl also
provides http_jar and http_file (the latter only supports the urls
parameter, not url).
--incompatible_remove_native_http_archivetrueWhen set, java_common.create_provider and certain arguments to JavaInfo are deprecated. The
deprecated arguments are: actions, sources, source_jars, use_ijar, java_toolchain,
and host_javabase.
Example migration from create_provider:
# Before
provider = java_common.create_provider(
ctx.actions,
compile_time_jars = [output_jar],
use_ijar = True,
java_toolchain = ctx.attr._java_toolchain,
transitive_compile_time_jars = transitive_compile_time,
transitive_runtime_jars = transitive_runtime_jars,
)
# After
compile_jar = java_common.run_ijar(
ctx.actions,
jar = output_jar,
target_label = ctx.label,
java_toolchain = ctx.attr._java_toolchain,
)
provider = JavaInfo(
output_jar = output_jar,
compile_jar = compile_jar,
deps = deps,
runtime_deps = runtime_deps,
)
Example migration from deprecated JavaInfo arguments:
# Before
provider = JavaInfo(
output_jar = my_jar,
use_ijar = True,
sources = my_sources,
deps = my_compile_deps,
runtime_deps = my_runtime_deps,
actions = ctx.actions,
java_toolchain = my_java_toolchain,
host_javabase = my_host_javabase,
)
# After
my_ijar = java_common.run_ijar(
ctx.actions,
jar = my_jar,
target_label = ctx.label,
java_toolchain, my_java_toolchain,
)
my_source_jar = java_common.pack_sources(
ctx.actions,
sources = my_sources,
java_toolchain = my_java_toolchain,
host_javabase = my_host_javabase,
)
provider = JavaInfo(
output_jar = my_jar,
compile_jar = my_ijar,
source_jar = my_source_jar,
deps = my_compile_deps,
runtime_deps = my_runtime_deps,
)
A tool is an input coming from an attribute of type label
where the attribute has been marked executable = True. In order for an action
to run a tool, it needs access to its runfiles.
Under the old API, tools are passed to ctx.actions.run() and
ctx.actions.run_shell() via their inputs parameter. Bazel scans this
argument (which may be a large depset) to find all the inputs that are tools,
and adds their runfiles automatically.
In the new API, tools are instead passed to a dedicated tools parameter. The
inputs are not scanned. If a tool is accidentally put in inputs instead of
tools, the action will fail during the execution phase with an error due to
missing runfiles. This may be somewhat cryptic.
To support a gradual transition, all actions with a tools argument are opted
into the new API, while all actions without a tools argument still follow the
old one. In the future (when this flag is removed), all actions will use the new
API unconditionally.
This flag turns on a safety check that is useful for migrating existing code.
The safety check applies to all actions that do not have a tools argument. It
scans the inputs looking for tools, and if it finds any, it raises an error
during the analysis phase that clearly identifies the offending tools.
In the rare case that your action requires a tool as input, but does not
actually run the tool and therefore does not need its runfiles, the safety check
will fail even though the action would have succeeded. In this case, you can
bypass the check by adding a (possibly empty) tools argument to your action.
Note that once an action has been modified to take a tools argument, you will
no longer get helpful analysis-time errors for any remaining tools that should
have been migrated from inputs.
--incompatible_no_support_tools_in_action_inputsfalsePreviously, directories created by
ctx.actions.declare_directory expanded
to the path of the directory when added to an Args object.
With this flag enabled, directories are instead replaced by the full file
contents of that directory when passed to args.add_all() or
args.add_joined(). (Directories may not be passed to args.add().)
If you want the old behavior on a case-by-case basis (perhaps your tool can
handle directories on the command line), you can pass expand_directories=False
to the args.add_all() or args.add_joined() call.
d = ctx.action.declare_directory("dir")
# ... Some action runs and produces ["dir/file1", "dir/file2"] ...
f = ctx.action.declare_file("file")
args = ctx.action.args()
args.add_all([d, f])
# -> Used to expand to ["dir", "file"]
# Now expands to ["dir/file1", "dir/file2", "file"]
--incompatible_expand_directoriesfalseWhen the flag is set, use a saner way to resolve variables. The previous behavior was buggy in a number of subtle ways. See the proposal for background and examples.
The proposal is not fully implemented yet.
--incompatible_static_name_resolutionfalseWhen the flag is set, load can only import symbols that were explicitly
defined in the target file, using either = or def.
When the flag is unset (legacy behavior), load may also import symbols that
come from other load statements.
In other words, the x below is exported only if the flag is unset:
load(":file.bzl", "x")
y = 1
--incompatible_no_transitive_loadsfalse0.19.0If false, Bazel constructs an in-memory //tools/defaults package based on the
command line options. If true, //tools/defaults:* is resolved from file system
as a regular package.
--incompatible_disable_tools_defaults_packagefalse//tools/default was initially created as virtual in-memory package. It
generates content dynamically based on current configuration. There is no need
of having //tools/defaults any more as LateBoundAlias can do dynamic
configuration-based label resolving. Also, having //tools/default makes
negative impact on performance, and introduces unnecessary code complexity.
All references to //tools/defaults:* targets should be removed or replaced
to corresponding target in @bazel_tools//tools/jdk: and
@bazel_tools//tools/cpp: packages.
Targets in //tools/default will not exist any more. If you have any references
inside your BUILD or *.bzl files to any of its, then bazel will fail to resolve.
Please replace all occurrences:
//tools/defaults:jdk
@bazel_tools//tools/jdk:current_java_runtime@bazel_tools//tools/jdk:current_host_java_runtime//tools/defaults:java_toolchain
@bazel_tools//tools/jdk:current_java_toolchain//tools/defaults:crosstool
@bazel_tools//tools/cpp:current_cc_toolchain@bazel_tools//tools/cpp:current_cc_host_toolchainlibc_top, then @bazel_tools//tools/cpp:current_libc_topThese targets will not be supported any more:
//tools/defaults:coverage_report_generator//tools/defaults:coverage_supportIf true, Bazel will stop retrieving the value of compiler from the cpp configuration when
--compiler is not specified. This will cause a config_setting that have
values = {"compiler": "x"} to not work properly when --compiler is not specified at command
line.
The former behavior can be achieved by changing the config_setting to use
flag_values = {"@bazel_tools//tools/cpp:compiler": "x"} instead:
# Before
config_setting(
name = "cpu_x_compiler_y",
values = {
"cpu": "x",
"compiler": "y",
},
)
# After
config_setting(
name = "cpu_x_compiler_y",
values = {
"cpu": "x",
},
flag_values = {
"@bazel_tools//tools/cpp:compiler": "y",
},
)
--incompatible_disable_late_bound_option_defaultsfalse0.18.0If true, Bazel will no longer accept depsets in user_compile_flags for
create_compile_variables,
and in user_link_flags for
create_link_variables.
Use plain lists instead.
--incompatible_disable_depset_in_cc_user_flagsfalse0.18.0Currently Bazel selects the cc_toolchain to use from the toolchains
dictionary attribute of cc_toolchain_suite. The key it uses is constructed
the following way:
--compiler option is specified, the key is --cpu|--compiler. Bazel
errors out if the entry doesn’t exist.--compiler option was not specified on command line, Bazel checks if
an entry with the key --cpu exists, and uses it if it does. If such an
entry doesn’t exist, it loops through the default_toolchain list in the
CROSSTOOL file, selects the first one that matches the --cpu option,
finds the CToolchain whose identifier matches the
default_toolchain.toolchain_identifier field, and then uses the key
CToolchain.targetCpu|Ctoolchain.compiler. It errors out if the entry
doesn’t exist.We’re making selection of the cc_toolchain label independent of the
CROSSTOOL file: when the flag is set to True, Bazel will no longer loop
through the default_toolchain list in order to construct a key for selecting
a cc_toolchain label from cc_toolchain_suite.toolchains, but throw an error
instead.
In order to not be affected by this change, one should add entries in the
cc_toolchain_suite.toolchains for the potential values of --cpu:
# Before
cc_toolchain_suite(
toolchains = {
"cpu1|compiler1": ":cc_toolchain_label1",
"cpu2|compiler2": ":cc_tolchain_label2",
}
)
# After
cc_toolchain_suite(
toolchains = {
"cpu1|compiler1": ":cc_toolchain_label1",
"cpu2|compiler2": ":cc_toolchain_label2",
"cpu1": ":cc_toolchain_label3",
"cpu2": ":cc_tolchain_label4",
}
)
Before, it could happen that the same cc_toolchain is used with multiple
CToolchains from the CROSSTOOL through default_toolchains. This is no longer
allowed, each cc_toolchain must point to at most one CToolchain by:
cc_toolchain.toolchain_identifier equal to
CToolchain.toolchain_identifiercompiler)
specifying cc_toolchain.cpu and cc_toolchain.compiler fields that match
CToolchain.target_cpu and CToolchain.compiler respectively.--cpu and --compiler options.Using cc_toolchain.toolchain_identifier will save you one migration in the
future.
--incompatible_disable_cc_toolchain_label_from_crosstool_protofalse0.18.0Currently Bazel allows rule authors to access certain Make variables that are implicitly provided to every rule by the CppConfiguration. This causes every target to implicitly depend on CppConfiguration, which creates an undesirable number of extra, unused, dependencies.
We are removing the implicit provision of these Make variables, and requiring rules and targets that use these Make variables to explicitly depend on a C++ toolchain in order to access them.
The list of Make variables is:
In order to not be affected by this change, add a C++ toolchain to the
toolchains attribute for targets, or to the_toolchains attribute for
Starlark rules. The best choice for this value is
the alias target @bazel_tools//tools/cpp:current_cc_toolchain, which will
always resolve to the currently selected C++ toolchain.
Genrules will still have access to these Make variables for the time being because that information is plumbed not through CppConfiguration, but through an implicit dependency on the C++ toolchain. That will also be removed at some point in the future, so it’s considered good practice to add an explicit dependency on the toolchain as demonstrated below.
For genrules and other targets using C++ Make Variables:
# Before
genrule(
cmd = "$(STRIP) file-to-be-stripped.o",
)
# After
genrule(
cmd = "$(STRIP) file-to-be-stripped.o",
toolchains = ["@bazel_tools//tools/cpp:current_cc_toolchain"],
)
For Starlark rules using C++ Make Variables:
# Before
def _impl(ctx):
strip = ctx.vars["STRIP"]
...
my_rule = rule(
implementation = _impl,
attrs = {
},
)
# After
def _impl(ctx):
strip = ctx.vars["STRIP"]
...
my_rule = rule(
implementation = _impl,
attrs = {
"_toolchains": attr.label_list(default = [Label("@bazel_tools//tools/cpp:current_cc_toolchain")]),
},
)
--incompatible_disable_cc_configuration_make_variablesfalse0.18.0This turns off legacy Starlark access to cc toolchain information via the
ctx.fragments.cpp fragment. Instead of declaring dependency on the ctx.fragments.cpp using the
fragments attribute declare a dependency on the @bazel_tools//tools/cpp:current_cc_toolchain
via implicit attribute named _cc_toolchain (see example below). Use find_cpp_toolchain from
@bazel_tools//tools/cpp:toolchain_utils.bzl to get the current C++ toolchain in the rule
implementation.
# Before
def _impl(ctx):
...
ctx.fragments.cpp.compiler_options()
foo = rule(
implementation = _impl,
fragments = ["cpp"],
...
)
# After
load("@bazel_tools//tools/cpp:toolchain_utils.bzl", "find_cpp_toolchain")
def _impl(ctx):
...
cc_toolchain = find_cpp_toolchain(ctx)
cc_toolchain.compiler_options()
foo = rule(
implementation = _impl,
attrs = {
"_cc_toolchain": attr.label(
default = Label("@bazel_tools//tools/cpp:current_cc_toolchain")
),
},
)
List of all legacy fields and their corresponding cc_toolchain alternative:
ctx.fragments.cpp |
cc_toolchain |
|---|---|
ar_executable |
ar_executable() |
built_in_include_directories |
built_in_include_directories |
c_options |
c_options() |
compiler |
compiler |
compiler_executable |
compiler_executable() |
compiler_options(unused_arg) |
compiler_options() |
cpu |
cpu |
cxx_options(unused_arg) |
cxx_options() |
dynamic_link_options(unused_arg, bool) |
dynamic_link_options(bool) |
fully_static_link_options(unused_arg, True) |
fully_static_link_options(True) |
ld_executable |
ld_executable() |
link_options |
link_options_do_not_use |
mostly_static_link_options(unused_arg, bool) |
mostly_static_link_options(bool) |
nm_executable |
nm_executable() |
objcopy_executable |
objcopy_executable() |
objdump_executable |
objdump_executable() |
preprocessor_executable |
preprocessor_executable() |
strip_executable |
strip_executable() |
sysroot |
sysroot |
target_gnu_system_name |
target_gnu_system_name |
unfiltered_compiler_options(unused_arg) |
unfiltered_compiler_options(unused_arg) |
--incompatible_disable_legacy_cpp_toolchain_skylark_apifalse0.18.0We have deprecated the cc_toolchain Starlark API returning legacy CROSSTOOL fields:
Use the new API from cc_common
# Before:
load("@bazel_tools//tools/cpp:toolchain_utils.bzl", "find_cpp_toolchain")
def _impl(ctx):
cc_toolchain = find_cc_toolchain(ctx)
compiler_options = (
cc_toolchain.compiler_options() +
cc_toolchain.unfiltered_compiler_options([]) +
["-w", "-Wno-error"]
)
link_options = (
["-shared", "-static-libgcc"] +
cc_toolchain.mostly_static_link_options(True) +
["-Wl,-whole-archive"] +
[l.path for l in libs] +
["-Wl,-no-whole-archive"] +
cc_toolchain.link_options_do_not_use
)
# After
load("@bazel_tools//tools/cpp:toolchain_utils.bzl", "find_cpp_toolchain")
load(
"@bazel_tools//tools/build_defs/cc:action_names.bzl",
"CPP_LINK_DYNAMIC_LIBRARY_ACTION_NAME",
"C_COMPILE_ACTION_NAME",
)
def _impl(ctx):
cc_toolchain = find_cc_toolchain(ctx)
feature_configuration = cc_common.configure_features(
cc_toolchain = cc_toolchain,
requested_features = ctx.features,
unsupported_features = ctx.disabled_features,
)
compile_variables = cc_common.create_compile_variables(
feature_configuration = feature_configuration,
cc_toolchain = cc_toolchain,
user_compile_flags = depset(["-w", "-Wno-error"]),
)
compiler_options = cc_common.get_memory_inefficient_command_line(
feature_configuration = feature_configuration,
action_name = C_COMPILE_ACTION_NAME,
variables = compile_variables,
)
link_variables = cc_common.create_link_variables(
feature_configuration = feature_configuration,
cc_toolchain = cc_toolchain,
is_linking_dynamic_library = True,
user_link_flags =
["-static-libgcc"] +
["-Wl,-whole-archive"] +
[lib.path for lib in libs] +
["-Wl,-no-whole-archive"],
)
link_flags = cc_common.get_memory_inefficient_command_line(
feature_configuration = feature_configuration,
action_name = CPP_LINK_DYNAMIC_LIBRARY_ACTION_NAME,
variables = link_variables,
)
--incompatible_disable_legacy_flags_cc_toolchain_apifalse0.19.0cfg = "data"cfg = "data" is a no-op that incorrectly gives the impression dependencies under
it are built in a distinct “data” mode:
my_rule = rule(
...
"some_attr": attr.label_list(
cfg = "data" # This line does nothing
)
)
The original semantics were unclear and were removed in 0.16.0.
Because this syntax is non-functional and confusing, it’s being removed outright (#6153). The functionality it implies will be provided by Starlark build configuration.
When --incompatible_disallow_data_transition=true, builds using this syntax
fail with an error.
--incompatible_disallow_data_transitionfalse0.16.0If set to true, disallow rule implementation functions from returning multiple instances of the same type of provider. If false, only the last in the list will be used.
incompatible_disallow_conflicting_providerstruePreviously, the label argument to the load statement (the first argument) was
checked to ensure that it referenced an existing package but it was not checked
to ensure that it didn’t cross a package boundary.
For example, in
load("//a:b/c.bzl", "doesntmatter")
if this flag is set to true, the above statement will be in error if //a/b
is a package; in such a case, the correct way to reference c.bzl via a label
would be //a/b:c.bzl.
--incompatible_disallow_load_labels_to_cross_package_boundariesfalse