We rely on a weird, GNU specific behaviour of objcopy:
$ aarch64-linux-android-objcopy -w libgcc.a stripped.a --strip-unneeded-symbol=*
$ llvm-nm --defined-only stripped.a | grep multi
0000000000000000 T __multi3
$ aarch64-linux-android-objcopy -w libgcc.a stripped.a --strip-unneeded-symbol=* --keep-symbol=FOO
$ llvm-nm --defined-only stripped.a | grep multi
llvm-objcopy has a different, more consistent behaviour that does not
mark exported symbols as unneeded when "--keep-symbol" flag exists.
However this behaviour will leave most symbols not deleted in the
libgcc_stripped library.
Revert the change while I'm implementing an alternative solution.
Test: presubmit
Bug: 142585047
Change-Id: I175e811f8a1f6afc6558267fc54e159ad5e12acf
A new sysprop neverallow rules are mandatory only for devices launching
with R or later. For devices already launched, neverallow rules can be
relaxed with adding following line to BoardConfig.mk:
BUILD_BROKEN_TREBLE_SYSPROP_NEVERALLOW := true
Bug: 131162102
Test: Set PRODUCT_SHIPPING_API_LEVEL := 30 and try building with
changing some system_public_prop to system_internal_prop
Test: m cts sepolicy_tests
Change-Id: I5e1640f7b43fd47863bc7bd163c2f43b01fa3326
Merged-In: I5e1640f7b43fd47863bc7bd163c2f43b01fa3326
(cherry picked from commit e36f52754f)
With __ANDROID_APEX_<NAME>__ definition, native modules may have
different behavior when it is built for a specific apex module.
Previously, the name is passed as value of definition __ANDROID_APEX__
like -D__ANDROID_APEX__=com.android.foo. But it is difficult to do
conditional compilation with it.
Now, since the name is incorporated into definition itself, it gets
easier to set #ifdef condition.
Bug: 142582178
Test: m (soong test added)
Change-Id: I3c90c789fa692a19addf2e5a7c8d4cc571cde112
Older VNDK libraries are provided as vndk_prebuilt_shared modules. Those
are added to corresponding VNDK APEX as dependencies.
With VNDK APEX installed, VNDK libs are unnecessary. By the way, since
there can be vendor modules which depend on VNDK libs, Make targets are
still emitted with UNINSTALLABLE=true.
Android.mk has additional modules for vndk libraries which are named
with apex name as suffices. For example, if libfoo is a vndk library,
then libfoo.vendor is its vendor variant and it would be in
/system/lib/vndk. But with vndk apex, it has additional
libfoo.com.android.vndk.current variant.
Bug: 141451661
Bug: 139772411
Test: m (soong tests)
Test: boot with aosp_arm64 system image on Q vendor device
Change-Id: I269c28a4d4c4e2f1518bd51df558438fe5316774
AIDEGen collects the complied sources which are generated from build
system, the sources includes not only java/kt files but also srcjar
files. Since the srcjar files are record in srcjars parameter in json
file, we should keep only java or kt files in the srcs parameter.
The size diff of module_bp_java_deps.json:
Without this patch: 15,298,369 Bytes
With this patch: 15,044,804 Bytes
The build time diff:
Without this patch: 2m31.345
With this patch: 2m32.662
Bug: 141528361
Test: 1. m clean -j
2. aidegen tradefed
3. Open out/soong/module_bp_java_deps.json
4. Find the module CtsSyncManagerCommon and check the
sync_manager_cts.srcjar files doesn't exist in srcs but exists
in srcjars section.
Change-Id: I43fc5c05b657473054e632cae4795220907dc711
APK in a flattened APEX is installed as 'ETC' class module, instead of
the 'APP' class. This is to prevent Make from doing app-specific
amendments (e.g. such as adding module name after my_module_path)
to the paths and filenames which are all correctly set in the Soong
side.
Test: add `installable:true` to development/samples/Snake/Android.mk
and add 'Snake' to the `apps` property of an APEX. The build is
successful.
Test: build the APEX on a device with TARGET_FLATTEN_APEX=true
The APK is at system/apex/<apexname>/app/Snake/Snake.apk
Bug: 142537672
Change-Id: I314bf1ab7abd4d4a4e9fa210442c004f54d8ccca
This allows droidstubs to use the same system modules to create the
stubs that will be used to compile them. It improves consistency and
avoids droidstubs having to duplicate the libraries that make up the
system modules on its libs property.
Adds systemModules() to the sdkContext which allows consistent error
checking behavior between droidstubs and java_library.
Bug: 142534789
Test: m checkbuild
Change-Id: Ib2006906d9528a900f16851f50b62152ffb51a1b
This fixes the bug: flattened apex is installed under
/system/apex/<soong module name> not /system/apex/<apex name>.
Flattened apexes can be seen as pre-activated under /system/apex, which
is bind-mount to /apex. Therefore, install dir for flattend apex should
be the same with the activation point(or mountpoint) of non-flattened apex.
When apex_name: is set, it should be used as its name instead of name:.
This change also emits makevar APEX_FILE_CONTEXTS_INFOS which is list of
pairs(<apex_name>:<file_contexts>) and is used by sepolicy to modify
file_contexts files for flattened apexes properly.
Bug: 123314817
Bug: 142300241
Test: add a test apex module(foo) with apex_name(foobar)
Test: OVERRIDE_TARGET_FLATTEN_APEX=true m foo
Test: see if the apex is installed correctly(/system/apex/foobar)
Test: see if the file_contexts for foo is correctly modified
Change-Id: I97a07de5cc772cd1d91e95ae059d282038028916
Should unbreak the mac build while I work on toybox xargs some more.
Test: suck it and see (since there's no mac presubmit)
Change-Id: I48954f801fc980a1174f93e474d73a24ad12ce66
With this change, setting the environment variable
EXPERIMENTAL_USE_OPENJDK11_TOOLCHAIN=true switches from the OpenJDK 9
toolchain to the OpenJDK 11 one (prebuilts/jdk/jdk11).
Since the version of the java.base module has to match the version of
the jlink tool which consumes it, the --module-version argument to the
jmod create call is also switched.
Test: make
Test: EXPERIMENTAL_USE_OPENJDK11_TOOLCHAIN make
Bug: 131683177
Change-Id: I606d22538165f309fe5537bd67a26390b1dd5771
This rule takes a moduleName parameter which only ever has the value
java.base. What's more, the assuption that the value will always be
java.base is baked into the rule in other ways (most significantly,
the module created with this name is passed to jlink, and jlink
requires a java.base module).
This change removes the parameter and hard-codes java.base everywhere,
on the basis that it's better to have it completely hard-coded than
partially configurable and partially hard-coded.
Test: make
Change-Id: Ib42cda92f623c5a1f929b5a99248ca8919e4700c
The check doesn't make sense for host.
Test: ALLOW_MISSING_DEPENDENCIES=true DIST_DIR=out/dist
./art/tools/dist_linux_bionic.sh -j110 com.android.art.host
is successful
Change-Id: Icdb4bb557a83b3c8044e6c73721dab47c56888a1
When SOONG_ALLOW_MISSING_DEPENDENCIES is defined in the build, don't
fail if an apex_available entry is not available. This is for smaller
Android manifests that may not include all APEX projects (e.g. master-art).
Bug: 142300164
Test: able to build from master-art manifest.
Change-Id: Ice1274f7dd143c4b1bf756e2cde93e217266affa
My previous change makes TMPDIR a temporary directory that's removed
before we exit. So instead of starting ckati to get this value to a
directory that no longer exists, return the semi-stable value that we
use during a normal build.
Fixes: 142277335
Test: get_build_var TMPDIR
Change-Id: I37d1219b1fda09f131dee17ef7b91e475d3e39ad
This allows us to change `logpersist.start` to `sh_binary` and unblocks
breakpad symbol uploading
Test: convert logpersist.start to sh_binary and verify symlinks are correct
Change-Id: I1b86c512df73a336205ca35216445a08e43bb879
Group aidl files into groups of 50 and compile them together into a
srcjar.
Bug: 124333557
Test: m checkbuild
Change-Id: I18e0858eab434071d3ff0039bb21e07c7311b601
Group protoc calls to generate java files into groups of 100
to avoid having a srcjar for every proto file.
Test: m checkbuild
Change-Id: I72a5efaabad24c975282df321d00ec2d7900645e
AIDL generates a deps file with no output listed temporarily pass it
through without error in the makedeps parser.
Bug: 141372861
Test: m checkbuild
Change-Id: I8c6740833dbc2ff3318dfc424ec497728cfc48d2
Instead of just using the host $PATH during dumpvars, use our path
configuration from the regular build. But instead of creating a ton of
symlinks to the interposer, just use a small directory of symlinks. This
only takes ~3ms (vs ~300ms), at the expense of error logging. Since we
do just about the same product configuration at the start of the build,
we can just rely on the logging there.
This fixes warnings printed by the Mac build, since we using the host
`date` instead of the toybox version:
usage: date [-jnRu] [-d dst] [-r seconds] [-t west] [-v[+|-]val[ymwdHMS]] ...
[-f fmt date | [[[mm]dd]HH]MM[[cc]yy][.ss]] [+format]
usage: date [-jnRu] [-d dst] [-r seconds] [-t west] [-v[+|-]val[ymwdHMS]] ...
[-f fmt date | [[[mm]dd]HH]MM[[cc]yy][.ss]] [+format]
TMPDIR was being shared with the main build, but we're allowed to run in
parallel. If a build is started while we're running, our TMPDIR may be
cleared. Instead, create our own temporary directory, which doesn't need
to be a stable value (as we never do caching here).
Fixes: 141893400
Test: (on mac) source build/envsetup.sh; lunch
Change-Id: I0ff536e71dc5649cae26daa0087eb80861704021
Host builds using libc++ used -nodefaultlibs to turn off the default
C++ runtime, and then added back all the other necessary libraries.
Clang supports -nostdlib++ since https://reviews.llvm.org/D47115
that removes the C++ runtime without affecting the other default
libraries. Use -nostdlib++, and remove the lists of default
libraries.
Test: m checkbuild
Change-Id: I722bd6596a1f3f5819f2767c29c0fa1e8b3ec0e8
Yi Kong