Build System Changes for Writers

COPY_HEADERS usage now produces warnings {#copy_headers}

We‘ve considered BUILD_COPY_HEADERS/LOCAL_COPY_HEADERS to be deprecated for a long time, and the places where it’s been able to be used have shrinked over the last several releases. Equivalent functionality is not available in Soong.

See the build/soong/docs/ for more information about how best to handle headers in Android.

m4 is not available on $PATH

There is a prebuilt of it available in prebuilts/build-tools, and a make variable M4 that contains the path.

Beyond the direct usage, whenever you use bison or flex directly, they call m4 behind the scene, so you must set the M4 environment variable (and depend upon it for incremental build correctness):

$(intermediates)/foo.c: .KATI_IMPLICIT_OUTPUTS := $(intermediates)/foo.h
$(intermediates)/foo.c: $(LOCAL_PATH)/foo.y $(M4) $(BISON) $(BISON_DATA)
	M4=$(M4) $(BISON) ...

Rules executed within limited environment

With ALLOW_NINJA_ENV=false (soon to be the default), ninja, and all the rules/actions executed within it will only have access to a limited number of environment variables. Ninja does not track when environment variables change in order to trigger rebuilds, so changing behavior based on arbitrary variables is not safe with incremental builds.

Kati and Soong can safely use environment variables, so the expectation is that you'd embed any environment variables that you need to use within the command line generated by those tools. See the export section below for examples.

For a temporary workaround, you can set ALLOW_NINJA_ENV=true in your environment to restore the previous behavior, or set BUILD_BROKEN_NINJA_USES_ENV_VAR := <var> <var2> ... in your to allow specific variables to be passed through until you've fixed the rules.

LOCAL_C_INCLUDES outside the source/output trees are an error {#BUILD_BROKEN_OUTSIDE_INCLUDE_DIRS}

Include directories are expected to be within the source tree (or in the output directory, generated during the build). This has been checked in some form since Oreo, but now has better checks.

There‘s now a BUILD_BROKEN_OUTSIDE_INCLUDE_DIRS variable, that when set, will turn these errors into warnings temporarily. I don’t expect this to last more than a release, since they're fairly easy to clean up.

Neither of these cases are supported by Soong, and will produce errors when converting your module.

Absolute paths

This has been checked since Oreo. The common reason to hit this is because a makefile is calculating a path, and ran abspath/realpath/etc. This is a problem because it makes your build non-reproducible. It's very unlikely that your source path is the same on every machine.

Using ../ to leave the source/output directories

This is the new check that has been added. In every case I've found, this has been a mistake in the -- assuming that LOCAL_C_INCLUDES (which is relative to the top of the source tree) acts like LOCAL_SRC_FILES (which is relative to LOCAL_PATH).

Since this usually isn't a valid path, you can almost always just remove the offending line.


Define proper HIDL / Stable AIDL HAL instead.


PRODUCT_STATIC_BOOT_CONTROL_HAL was the workaround to allow sideloading with statically linked boot control HAL, before shared library HALs were supported under recovery. Android Q has added such support (HALs will be loaded in passthrough mode), and the workarounds are being removed. Targets should build and install the recovery variant of boot control HAL modules into recovery image, similar to the ones installed for normal boot. See the change to crosshatch for example of this:

Deprecation of BUILD_* module types

See build/make/ for the current status.


Previously, adding a module to PRODUCT_PACKAGES that supported both the host and the target (host_supported in Android.bp; two modules with the same name in would cause both to be built and installed. In many cases you only want either the host or target versions to be built/installed by default, and would be over-building with both. So PRODUCT_PACKAGES will be changing to just affect target modules, while PRODUCT_HOST_PACKAGES is being added for host modules.

Functional differences between PRODUCT_PACKAGES and PRODUCT_HOST_PACKAGES:

  • PRODUCT_HOST_PACKAGES does not have _ENG/_DEBUG variants, as that's a property of the target, not the host.
  • PRODUCT_HOST_PACKAGES requires listed modules to exist, and be host modules. (Unless ALLOW_MISSING_DEPENDENCIES is set)

This is still an active migration, so currently it still uses PRODUCT_PACKAGES to make installation decisions, but verifies that if we used PRODUCT_HOST_PACKAGES, it would trigger installation for all of the same host packages. This check ignores shared libraries, as those are not normally necessary in PRODUCT_*PACKAGES, and tended to be over-built (especially the 32-bit variants).

Future changes will switch installation decisions to PRODUCT_HOST_PACKAGES for host modules, error when there's a host-only module in PRODUCT_PACKAGES, and do some further cleanup where LOCAL_REQUIRED_MODULES are still merged between host and target modules with the same name.

*.c.arm / *.cpp.arm deprecation {#file_arm}

In files, you used to be able to change LOCAL_ARM_MODE for each source file by appending .arm to the end of the filename in LOCAL_SRC_FILES.

Soong does not support this uncommonly used behavior, instead expecting those files to be split out into a separate static library that chooses arm over thumb for the entire library. This must now also be done in files.

Windows cross-compiles no longer supported in

Modules that build for Windows (our only HOST_CROSS OS currently) must now be defined in Android.bp files.

LOCAL_MODULE_TAGS := eng debug are obsolete {#LOCAL_MODULE_TAGS}

LOCAL_MODULE_TAGS value eng and debug are now obsolete. They allowed modules to specify that they should always be installed on -eng, or -eng and -userdebug builds. This conflicted with the ability for products to specify which modules should be installed, effectively making it impossible to build a stripped down product configuration that did not include those modules.

For the equivalent functionality, specify the modules in PRODUCT_PACKAGES_ENG or PRODUCT_PACKAGES_DEBUG in the appropriate product makefiles.

Core android packages like su got added to the list in build/make/target/product/, but for device-specific modules there are often better base product makefiles to use instead.

USER deprecation

USER will soon be nobody in many cases due to the addition of a sandbox around the Android build. Most of the time you shouldn‘t need to know the identity of the user running the build, but if you do, it’s available in the make variable BUILD_USERNAME for now.

Similarly, the hostname tool will also be returning a more consistent value of android-build. The real value is available as BUILD_HOSTNAME.


BUILD_NUMBER should not be used directly in files, as it would trigger them to be re-read every time the BUILD_NUMBER changes (which it does on every build server build). If possible, just remove the use so that your builds are more reproducible. If you do need it, use BUILD_NUMBER_FROM_FILE:

	mytool --build_number $(BUILD_NUMBER_FROM_FILE) -o $@

That will expand out to a subshell that will read the current BUILD_NUMBER whenever it's run. It will not re-run your command if the build number has changed, so incremental builds will have the build number from the last time the particular output was rebuilt.

DIST_DIR, dist_goal, and dist-for-goals

DIST_DIR and dist_goal are no longer available when reading files (or other build tasks). Always use dist-for-goals instead, which takes a PHONY goal, and a list of files to copy to $DIST_DIR. Whenever dist is specified, and the goal would be built (either explicitly on the command line, or as a dependency of something on the command line), that file will be copied into $DIST_DIR. For example,

$(call dist-for-goals,foo,bar/baz)

will copy bar/baz into $DIST_DIR/baz when m foo dist is run.

Renames during copy

Instead of specifying just a file, a destination name can be specified, including subdirectories:

$(call dist-for-goals,foo,bar/baz:logs/foo.log)

will copy bar/baz into $DIST_DIR/logs/foo.log when m foo dist is run.

.PHONY rule enforcement {#phony_targets}

There are several new warnings/errors meant to ensure the proper use of .PHONY targets in order to improve the speed and reliability of incremental builds.

.PHONY-marked targets are often used as shortcuts to provide “friendly” names for real files to be built, but any target marked with .PHONY is also always considered dirty, needing to be rebuilt every build. This isn't a problem for aliases or one-off user-requested operations, but if real builds steps depend on a .PHONY target, it can get quite expensive for what should be a tiny build. warning: PHONY target "out/.../foo" looks like a real file (contains a "/")

Between this warning and the next, we're requiring that .PHONY targets do not have “/” in them, and real file targets do have a “/”. This makes it more obvious when reading makefiles what is happening, and will help the build system differentiate these in the future too. warning: writing to readonly directory: "kernel-modules"

This warning will show up for one of two reasons:

  1. The target isn't intended to be a real file, and should be marked with .PHONY. This would be the case for this example.
  2. The target is a real file, but it's outside the output directories. All outputs from the build system should be within the output directory, otherwise m clean is unable to clean the build, and future builds may not work properly. warning: real file "out/.../foo" depends on PHONY target "buildbins"

If the first target isn‘t intended to be a real file, then it should be marked with .PHONY, which will satisfy this warning. This isn’t the case for this example, as we require .PHONY targets not to have ‘/’ in them.

If the second (PHONY) target is a real file, it may unnecessarily be marked with .PHONY.

.PHONY and calling other build systems

One common pattern (mostly outside AOSP) that we‘ve seen hit these warning is when building with external build systems (firmware, bootloader, kernel, etc). Those are often marked as .PHONY because the Android build system doesn’t have enough dependencies to know when to run the other build system again during an incremental build.

We recommend to build these outside of Android, and deliver prebuilts into the Android tree instead of decreasing the speed and reliability of the incremental Android build.

In cases where that's not desired, to preserve the speed of Android incrementals, over-specifying dependencies is likely a better option than marking it with .PHONY:

out/target/.../zImage: $(sort $(shell find -L $(KERNEL_SRCDIR)))

For reliability, many of these other build systems do not guarantee the same level of incremental build assurances as the Android Build is attempting to do -- without custom checks, Make doesn‘t rebuild objects when CFLAGS change, etc. In order to fix this, our recommendation is to do clean builds for each of these external build systems every time anything they rely on changes. For relatively smaller builds (like the kernel), this may be reasonable as long as you’re not trying to actively debug the kernel.

export and unexport deprecation {#export_keyword}

The export and unexport keywords are obsolete, and will throw errors when used.

Device specific configuration should not be able to affect common core build steps -- we're looking at triggering build steps to be invalidated if the set of environment variables they can access changes. If device specific configuration is allowed to change those, switching devices with the same output directory could become significantly more expensive than it already can be.

If used during files, and later tasks, it is increasingly likely that they are being used incorrectly. Attempting to change the environment for a single build step, and instead setting it for hundreds of thousands.

It is not recommended to just move the environment variable setting outside of the build (in, or some other configuration script or wrapper). We expect to limit the environment variables that the build respects in the future, others will be cleared. (There will be methods to get custom variables into the build, just not to every build step)

Instead, write the export commands into the rule command lines themselves:

	rm -rf $@
	export MY_ENV_A="$(MY_A)"; make ...

If you want to set many environment variables, and/or use them many times, write them out to a script and source the script:

envsh := $(intermediates)/
	rm -rf $@
	echo 'export MY_ENV_A="$(MY_A)"' >$@
	echo 'export MY_ENV_B="$(MY_B)"' >>$@

$(intermediates)/generated_output.img: PRIVATE_ENV := $(envsh)
$(intermediates)/generated_output.img: $(envsh) a/b/c/
	rm -rf $@
	source $(PRIVATE_ENV); make ...
	source $(PRIVATE_ENV); a/b/c/ ...

Implicit make rules are obsolete {#implicit_rules}

Implicit rules look something like the following:


%.o :

These can have wide ranging effects across unrelated modules, so they're now obsolete. Instead, use static pattern rules, which are similar, but explicitly match the specified outputs:

libs := $(foreach lib,libfoo libbar,$(TARGET_OUT_SHARED_LIBRARIES)/$(lib)

files := $(wildcard $(LOCAL_PATH)/*.foo)
gen := $(patsubst $(LOCAL_PATH)/,$(intermediates)/%.o,$(files))
$(gen): %.o :

Removing ‘/’ from Valid Module Names {#name_slash}

The build system uses module names in path names in many places. Having an extra ‘/’ or ‘../’ being inserted can cause problems -- and not just build breaks, but stranger invalid behavior.

In every case we‘ve seen, the fix is relatively simple: move the directory into LOCAL_MODULE_RELATIVE_PATH (or LOCAL_MODULE_PATH if you’re still using it). If this causes multiple modules to be named the same, use unique module names and LOCAL_MODULE_STEM to change the installed file name:

include $(CLEAR_VARS)
LOCAL_MODULE := ver1/code.bin

include $(CLEAR_VARS)
LOCAL_MODULE := ver2/code.bin

Can be rewritten as:

include $(CLEAR_VARS)
LOCAL_MODULE := ver1_code.bin

include $(CLEAR_VARS)
LOCAL_MODULE := ver2_code.bin

You just need to make sure that any other references (PRODUCT_PACKAGES, LOCAL_REQUIRED_MODULES, etc) are converted to the new names.

Valid Module Names

We've adopted lexical requirements very similar to Bazel's requirements for target names. Valid characters are a-z, A-Z, 0-9, and the special characters _.+-=,@~. This currently applies to LOCAL_PACKAGE_NAME, LOCAL_MODULE, and LOCAL_MODULE_SUFFIX, and LOCAL_MODULE_STEM*.

Many other characters already caused problems if you used them, so we don't expect this to have a large effect.

PATH Tools {#PATH_Tools}

The build has started restricting the external host tools usable inside the build. This will help ensure that build results are reproducible across different machines, and catch mistakes before they become larger issues.

To start with, this includes replacing the $PATH with our own directory of tools, mirroring that of the host PATH. The only difference so far is the removal of the host GCC tools. Anything that is not explicitly in the configuration as allowed will continue functioning, but will generate a log message. This is expected to become more restrictive over time.

The configuration is located in build/soong/ui/build/paths/config.go, and contains all the common tools in use in many builds. Anything not in that list will currently print a warning in the $OUT_DIR/soong.log file, including the command and arguments used, and the process tree in order to help locate the usage.

In order to fix any issues brought up by these checks, the best way to fix them is to use tools checked into the tree -- either as prebuilts, or building them as host tools during the build.

As a temporary measure, you can set TEMPORARY_DISABLE_PATH_RESTRICTIONS=true in your environment to temporarily turn off the error checks and allow any tool to be used (with logging). Beware that GCC didn't work well with the interposer used for logging, so this may not help in all cases.

Deprecating / obsoleting variables in Makefiles

It is not required to source before running a build. Many scripts, including a majority of our automated build systems, do not do so. Make will transparently make every environment variable available as a make variable. This means that relying on environment variables only set up in will produce different output for local users and scripted users.

Many of these variables also include absolute path names, which we‘d like to keep out of the generated files, so that you don’t need to do a full rebuild if you move the source tree.

To fix this, we‘re marking the variables that are set in as deprecated in the makefiles. This will trigger a warning every time one is read (or written) inside Kati. Once all the warnings have been removed for a particular variable, we’ll switch it to obsolete, and any references will become errors. variables with make equivalents

instead ofuse

All of the make variables may be relative paths from the current directory, or absolute paths if the output directory was specified as an absolute path. If you need an absolute variable, convert it to absolute during a rule, so that it's not expanded into the generated ninja file:

$(PRODUCT_OUT)/gen.img: my/src/path/
	export PRODUCT_OUT=$$(cd $(PRODUCT_OUT); pwd); cd my/src/path; ./ -o $${PRODUCT_OUT}/gen.img


In files, you can always assume that the current directory is the root of the source tree, so this can just be replaced with ‘.’ (which is what $TOP is hardcoded to), or removed entirely. If you need an absolute path, see the instructions above.

Stop using PATH directly

This isn‘t only set by, but it is modified by it. Due to that it’s rather easy for this to change between different shells, and it's not ideal to reread the makefiles every time this changes.

In most cases, you shouldn‘t need to touch PATH at all. When you need to have a rule reference a particular binary that’s part of the source tree or outputs, it's preferrable to just use the path to the file itself (since you should already be adding that as a dependency).

Depending on the rule, passing the file path itself may not be feasible due to layers of unchangable scripts/binaries. In that case, be sure to add the dependency, but modify the PATH within the rule itself:

$(TARGET): myscript my/path/binary
	PATH=my/path:$$PATH myscript -o $@

Stop using PYTHONPATH directly

Like PATH, this isn‘t only set by, but it is modified by it. Due to that it’s rather easy for this to change between different shells, and it's not ideal to reread the makefiles every time.

The best solution here is to start switching to Soong's python building support, which packages the python interpreter, libraries, and script all into one file that no longer needs PYTHONPATH. See fontchain_lint for examples of this:

If you still need to use PYTHONPATH, do so within the rule itself, just like path:

$(TARGET): $(sort $(shell find my/python/lib -name '*.py'))
	PYTHONPATH=my/python/lib:$$PYTHONPATH -o $@


Specify Framework Compatibility Matrix Version in device manifest by adding a target-level attribute to the root element <manifest>. If PRODUCT_COMPATIBILITY_MATRIX_LEVEL_OVERRIDE is 26 or 27, you can add "target-level"="1" to your device manifest instead.


Clang is the default and only supported Android compiler, so there is no reason for this option to exist.

Other variables {#other_envsetup_variables}


These are all exported from, but don‘t have clear equivalents within the makefile system. If you need one of them, you’ll have to set up your own version.