So far, when a library `libfoo` has `stubs.versions: ["10", "11"]`, then
`shared_libs: ["libfoo"]` is linked to the version 11 of the stub.
This requires the author of `libfoo` to manually update the property
whenever a new version is introduced. Otherwise, clients are not able
to use the newly added APIs because the latest stub is for an old
version.
This change eliminates the need for manual updating. "current" version
is always implicitly added to `stubs.versions`. It is added even when
nothing is set on the property, if `stubs.symbol_file` is set. i.e.
```
cc_library {
name: "libfoo",
stubs: {
symbol_file: "libfoo.map.txt",
// no versions: [...] needed
},
}
cc_library {
name: "a_client",
shared_libs: ["libfoo"],
apex_available: ["myapex"],
min_sdk_version: "29",
}
apex {
name: "myapex",
native_shared_libraries: ["a_client"],
min_sdk_version: "29",
}
```
`a_client` links to the "current" stub of `libfoo` that has all symbols
shown in the map file.
Note that, above doesn't mean that the client has unlimited access to
APIs that are introduced even after the min_sdk_version of the client
(29 in this example). The use of such APIs still has to be guarded with
`__builtin_available` check.
Bug: N/A
Test: m
Change-Id: I70bb1600c18e74d36c6b24c3569d2149f02aaf96
Test behavior was changed a while ago so that tests by default ignore
non-existent source paths (unless they explicitly check for/rely on
them). Prior to that CheckSnapshot() could detect when files were
missing from the snapshot but it no longer can.
This change disallows non-existent source files in all the sdk tests
which means that they are disallowed when processing the snapshots as
they use the same preparers as were used to process the sources.
This caused a test failure which has been temporarily ignored and has
a TODO and bug associated with it.
Bug: 183184375
Test: m nothing
Change-Id: I969d8515d20ef5ae515f2b5f93d8ed4e4f8ede75
The sdk produces snapshots that are expected to be unpacked in one of
the Android repos. Often that can lead to issues due to conflicts
between the source and prebuilts. This change attempts to avoid those
conflicts by testing the different ways that those files can be used.
With the existing test to cover adding the snapshot
This change will cause the sdk tests to check the following:
1) Snapshot on its own (already done).
2) Snapshot plus original source where the original source is
preferred.
3) Snapshot plus original source where the snapshot is preferred.
It also adds the ability for tests to provide their own custom checkers
to verify the result of each of the previous tests.
This change reveals a number of bugs already present. Rather than
attempt to fix them this change adds the ability to specify error
handlers for the two cases that mix source and snapshot to allow those
errors to be temporarily ignored while allowing the majority of the
tests to benefit from this improvement. Each of those failures has a
TODO and bug associated with it.
Bug: 183184375
Test: m nothing
Change-Id: I105233195074dbe7a6422b6dfc5486e74398ea15
These packages have already been migrated to use per test build
directory so have no need for a FixtureFactory.
Bug: 183235980
Test: m nothing
Change-Id: I667d1d992caaf0f615de91f89efdae11c44986c2
This change is in preparation for removing testing.T from TestResult.
Bug: 181070625
Test: m nothing
Change-Id: I67535aff0d894e6e3d8456b75540f340af853355
This will allow the testSdkResult to be replaced with the TestResult
when switching sdk package to use test fixtures.
Bug: 181070625
Test: m nothing
Change-Id: Ieca63f4c189f5e804102aeb4073289ea03143b6e
Previously, it was assumed that generated headers must be arch specific
and so prevented the fields referencing the paths to those headers from
being automatically optimized by the sdk generation code. That is not
always the case, e.g. with headers generated from protos so this change
allows those fields to be optimized.
Bug: 180427921
Test: m nothing
Change-Id: Id2af419d58ae3c30ea6d9e87f71e33a9ff6ba13b
Previously, the snapshot handling code did not preserve the directory
structure of generated include directories and instead just copied the
headers into the same module specific directory and added that single
directory to the export_include_dirs (or similar) property.
That had a couple of issues:
* The include directory was repeated in the ..._include_dirs property.
* It did not work when the include directories overlapped.
In the latter case it had a couple of issues:
* Code which compiled fine against the source would not compile against
the prebuilt.
* Header files were duplicated in the output.
e.g. assume the following generated header file structure:
foo/
foo.h
bar/
bar.h
baz/
baz.h
When the sdk snapshot was passed include directories of "foo", "bar" and
headers of "foo/foo.h", "bar/bar.h", "bar/baz/baz.h" it would generate a
snapshot with the structure:
include_gen/
foo.h
bar.h
baz/
baz.h
And:
export_include_dirs: ["include_gen", "include_gen"]
However, when the include directories overlapped and include directories
of "foo", "bar" and "bar/baz" were passed in the directory structure
would be the same and the export_include_dirs would contain 3 usages of
"include_gen".
That meant that source code which used the following would build
against the source (because it would find "baz.h" in the "bar/baz"
include directory) but would fail when built against the prebuilts
because the "include_gen" directory did not contain "baz.h":
#include "baz.h"
This change preserves the input directory structure for generated files
in a similar way to how it does it for source files. So, the snapshot
structure looks something like this:
include_gen/
foo/
foo.h
bar/
bar.h
baz/
baz.h
And:
export_include_dirs: [
"include_gen/foo",
"include_gen/bar",
"include_gen/bar/baz",
],
Bug: 180427921
Test: m nothing
Change-Id: Id69eef8cf5eecd033841d3b7cd0c044a697ce404
A number of tests in cc_sdk_test.go use "include" for the input include
directory which results in the output containing "include/include"
(because the snapshot code uses "include" as the directory into which
native headers will be output). This change switches the inputs from
"include.." to "myinclude.." to differentiate between the two sources.
Bug: 180427921
Test: m nothing
Change-Id: Iba1e6d94647c74b31307b18254c03580e64c91cf
Uses new capability added in previous change to separate the testing of
the versioned and unversioned snapshots in sdk/cc_sdk_test.go. Any test
that generated an _snapshot module that simply listed the modules and
did not check a new type only tests the unversioned output to reduce
the duplication and cost of changing the tests.
Bug: 180479010
Test: m nothing
Change-Id: If11f82b3dd8ec79110b3a2f5adf193b6464000ab
Also fix //apex_available:platform that could get lost in the generated
apex_available properties. It only happened if a module didn't specify
an apex_available setting and then got extra entries from
apex.BaselineApexAvailable.
Test: Build ART SDK with libnativeloader_lazy in the SDK, and check
that the generated prebuilt module has //apex_available:platform
(before https://r.android.com/1480559).
Change-Id: I9df20aee63455932a7e558993f0f26769cb02792
This amends Idad7ef138cdbcbd209d390bf6c10ca8365d4619f. With the change,
when there is a member that returns IsHostOsDependent() == true,
the sdk having the member and the member itself are disable for host and
only the os that the member supports is explicitly enabled.
However, that change will cause a problem when we add the support for
the linux_bionic_arm64 target. The target is not enabled when building
sdk snapshots. The only linux_bionic target that is enabled is
'linux_bionic_x86_64'. However, since the granularity is os which is
linux_bionic, the snapshot is generated as follows.
cc_prebuilt_binary {
target: {
host: {
enabled: false,
},
linux_bionic: {
enabled: true,
},
linux_bionic_x86_64: {
srcs: ["x86_64/bin/..."],
},
// no srcs for linux_bionic_arm64
},
}
Above is a problem for linux_bionic_arm64 target because the target is
enabled (via linux_bionic.enabled: true), but srcs is not provided.
To fix the problem, the enabling of a target is done in a target
(os_arch) granularity, rather than os granularity. For example, above
now becomes ...
cc_prebuilt_binary {
target: {
host: {
enabled: false,
},
linux_bionic_x86_64: {
enabled: true,
srcs: ["x86_64/bin/..."],
},
},
}
Only the targets that the snapshot actually can provide srcs are enabled
and the rest of the host targets are disabled.
Bug: 159685774
Test: m nothing
Test: build/soong/scripts/build-aml-prebuilts.sh
runtime-module-host-exports
Change-Id: Ibca48c40f6dc4628b5f4bfa4ceb68ebe0973cc81
This reverts commit 323dc60712.
Reason for revert: Possible cause of test instability
Bug: 170513220
Test: soong tests
Change-Id: Iee168e9fbb4210569e6cffcc23e60d111403abb8
Putting t.Parallel() in each test makes them run in parallel.
Additional t.Parallel() could be added to each subtest, although
that requires making a local copy of the loop variable for
table driven tests.
Test: m checkbuild
Change-Id: I5d9869ead441093f4d7c5757f2447385333a95a4
This change ensures that each prebuilt in a snapshot explicitly
specifies its visibility even when that visibility is the current
default. This is done for two reasons:
1. It simplifies a follow up change that adds visibility rules to an
existing set of rules.
2. It ensures that the snapshots are independent of the current Soong
default visibility.
The latter is important because we intend to switch from modules being
visible to everyone by default (i.e. //visibility:public) to only being
visible to modules in the same package (i.e. //visibility:private). By
making the snapshots of modules that do not specify any visibility
explicitly specify that they are "//visibility:public" it ensures that
the snapshots will not need to be changed when the default changes.
Bug: 168301990
Test: m nothing
Change-Id: Ia034f4a1e5124c17f46d73b0e9a6c5f2a251038e
This is a quick fix to avoid build failures due to prebuilts lacking
stubs versions that the sources have, which causes dependencies on
older version stubs fail on master-art-host where the sources of e.g.
Bionic are no longer available.
More work is required to make the version handling of stubs DTRT - a
TODO comment outlines it.
Test: m nothing
Test: `m runtime-module-sdk`, check that the generated Android.bp
has the full stubs.versions list like the sources.
Bug: 169373910
Change-Id: I0ebfaf94f9d45a0e58d9785c40e7cea626f0ab83
liblog snapshot needs to sanitizer.address=false to avoid cycle in asan
builds. Adding that separately in library_sdk_member.go would start to
feel like whack-a-mole, so the snapshot generation is instead extended
to handle nested property structs.
This uses the BpPropertySet.AddProperty extension in
https://r.android.com/1423510, and common value optimisation now
recurses into non-anonymous structs, instead of comparing them as a
whole.
Test: m nothing
Test: `m SANITIZE_TARGET=address nothing` with prebuilts/runtime
present in the manifest and a fresh snapshot made with this
Bug: 151303681
Change-Id: I472554117a488e6c800045cb2ed59377778571a4
Needed for jni_headers which is marked as recovery_available because libchrome
is also marked and that library is used by others that are the recovery
partition. There may be other dependency paths that require jni_headers to be
recovery_available that was just the first one found.
Bug: 168301990
Test: m nothing
Change-Id: I916ecfe7700b9c153b29115e849da148fc7e7a76
This is necessary to not get dependencies on libclang_rt sanitizer libs
for the CRT objects which lead to cyclic dependencies in sanitizer
builds.
Test: `lunch aosp_blueline_hwasan-userdebug && m nothing`
in a tree that has the prebuilts created and unzipped from
`build/soong/scripts/build-aml-prebuilts.sh runtime-module-{sdk,host-exports}`
Bug: 151303681
Change-Id: I3f848a084280bdc3ade4b74df03e981d8cc61222
APEX variants that share the same SDK version and updatability
almost always use identical command line arguments to build but
with different intermediates directories. This causes unnecessary
build time and disk space for duplicated work.
Deduplicate APEX variants that would build identically. Create
aliases from the per-APEX variations to the new shared variations
so that the APEX modules can continue to depend on them via the
APEX name as the variation.
This has one significant change in behavior. Before this change,
if an APEX had two libraries in its direct dependencies and one
of those libraries depended on the other, and the second library
had stubs, then the first library would depend on the implementation
of the second library and not the stubs. After this change, if
the first library is also present in a second APEX but the second
library is not, then the common variant shared between the two
APEXes would use the stubs, not the implementation.
In a correctly configured set of build rules this change will
be irrelevant, because if the compilation worked for the second
APEX using stubs then it will work for the common variant using
stubs. However, if an incorrect change to the build rules is
made this could lead to confusing errors, as a previously-working
common variant could suddenly stop building when a module is added
to a new APEX without its dependencies that require implementation
APIs to compile.
This change reduces the number of modules in an AOSP arm64-userdebug
build by 3% (52242 to 50586), reduces the number of variants of the
libcutils module from 74 to 53, and reduces the number of variants
of the massive libart[d] modules from 44 to 32.
This relands I0529837476a253c32b3dfb98dcccf107427c742c with a fix
to always mark permissions XML files of java_sdk_library modules as
unique per apex since they contain the APEX filename, and a fix
to UpdateUniqueApexVariationsForDeps to check ApexInfo.InApexes
instead of DepIsInSameApex to check if two modules are in the same
apex to account for a module that depends on another in a way that
doesn't normally include the dependency in the APEX (e.g. a libs
property), but the dependency is directly included in the APEX.
Bug: 164216768
Test: go test ./build/soong/apex/...
Change-Id: I2ae170601f764e5b88d0be2e0e6adc84e3a4d9cc
Necessary to avoid problems with implicitly enabled host OS'es, e.g.
linux_glibc getting enabled when we only supply a linux_bionic linker
in runtime-module-host-exports. That will then cause a non-functional
prebuilt to take precedence over source if the prebuilt is preferred.
We don't do this for device since we so far only support a single
device OS (android).
This introduces the notion that SDK member types can be host OS
dependent or not. That way java members with host prebuilts don't get
restricted to a specific host OS.
Test: m nothing
Test: build/soong/scripts/build-aml-prebuilts.sh runtime-module-host-exports
Check that the generated Android.bp correctly disables the bionic
linker prebuilt for linux_glibc.
Test: art/build/apex/runtests.sh
on master-art with an updated runtime SDK snapshot
Test: art/tools/buildbot-build.sh {--host,--target}
on master-art with an updated runtime SDK snapshot
Bug: 160349757
Change-Id: Idad7ef138cdbcbd209d390bf6c10ca8365d4619f
Previously compile_multilib was ignored for device and defaulted to
"both", which could lead to an internal inconsistency in the snapshot
if a member was 32- or 64-only.
This also implements common property optimisation for compile_multilib,
by adding a struct that can be extended to other properties in the
future.
Test: m nothing
Bug: 151303681
Change-Id: I3e96222858e5e755173aca03508a23c431977e2a
There was a special case in cc_prebuilt_binary only, which resulted in
prebuilt libraries getting both 32 and 64 bit variants even when their
sources only had one of them, and the other variant would be defunct
since it wouldn't get any prebuilt artefact.
This moves the handling of compile_multilib completely to the common
update code, so that SDK members don't need to deal with it. It doesn't
take SDK member defaults into account, which means a bit more
boilerplate in the snapshots, but it's simpler and less error prone
(different SDK member types have different defaults).
This relands https://r.android.com/1359962 after disabling the SDK
tests on mac once and for all.
Bug: 151303681
Test: `go test -v ./sdk` in build/soong on mac and linux
Change-Id: I05f6603b2ac9b8676b25c3e297165ca23284e9cc
Break apart test helpers a bit to make it possible to enable
LinuxBionic in a single test, and add LinuxBionic support to
cc.GatherRequiredDepsForTest.
Test: m nothing
Bug: 160349757
Change-Id: Iace1024c964cee2308c130c945daef9e46c18c66
We frequently miss putting in the per-test skip condition, and since
mac isn't tested in presubmit this leads to build breakages and time
consuming reverts. To avoid that this blanket disables all the SDK
tests on mac. It's not used there and we got test coverage in many
linux-based builds.
Bug: 145598135
Bug: 161315642
Test: `go test -v ./sdk` in build/soong on mac and linux
Change-Id: I2aea92fef2c0f8c2742396fe36610501dc5a6f0f
There was a special case in cc_prebuilt_binary only, which resulted in
prebuilt libraries getting both 32 and 64 bit variants even when their
sources only had one of them, and the other variant would be defunct
since it wouldn't get any prebuilt artefact.
This moves the handling of compile_multilib completely to the common
update code, so that SDK members don't need to deal with it. It doesn't
take SDK member defaults into account, which means a bit more
boilerplate in the snapshots, but it's simpler and less error prone
(different SDK member types have different defaults).
Bug: 151303681
Test: m nothing
Test: build/soong/scripts/build-aml-prebuilts.sh art-module-host-exports
Check that the generated Android.bp passes Soong.
Change-Id: Ib73444c6788ee1c78480bdb103aa2b8ae8f2c63c
Test: m nothing
Test: `m` with prebuilts/runtime in the manifest (along with other
fixes)
Bug: 151303681
Change-Id: I450d476975c7ab4434228b8c4baf3af192142211
Test: m nothing
Test: Check on go/Android.bp that cc_prebuilt_library supports
unique_host_soname
Bug: 158743135
Change-Id: Ie02a5cae057fb9092c226d8c5b7f63e1f66ad066
The flags are necessary to avoid implicit dependencies on crt libs,
libc, and the linker (recursively). The reason cc_prebuilt_binary react
to these flags is that they can affect the install rules.
Test: m nothing
Test: art/test/testrunner/run_build_test_target.py art-linux-bionic-x64
on a master-art branch using snapshot built from runtime-module-sdk
and runtime-module-host-exports
Bug: 152255951
Change-Id: I6b2fe92d105d1f446fffd00bd3267a416f75efb7
Change since last attempt: Disable test that breaks on darwin.
Adds a test that fails with unknown property android.stubs.versions
and then fixes that by marking the field from which that property is
created with 'ignored-on-host' and implemented the isHostVariant on
*osTypeSpecificInfo.
Bug: 155628860
Test: m nothing
Change-Id: Ia4e744c9e799d5adaf8a2f761516f568ec363ad4
Revert submission 1302576
Bug: 156054601
Reason for revert: Presumed root cause of build break.
Reverted Changes:
Ifc8116e11:Detect invalid arch specific properties in snapsho...
I7ebd33307:Adds support for 'ignored-on-host'
I167b47a13:Fix snapshot of a host/device cc_library with stub...
Change-Id: Ibccce5286605bb71c6be3b3550ba86d8b7e24fa7
Adds a test that fails with unknown property android.stubs.versions
and then fixes that by marking the field from which that property is
created with 'ignored-on-host' and implemented the isHostVariant on
*osTypeSpecificInfo.
Bug: 155628860
Test: m nothing
Change-Id: I167b47a1374f541aa09d7e045972d740f1d9009c
Test: m nothing
Test: Build an SDK snapshot including libc, check that the generated
cc_prebuilt_library_shared for it is a stub that doesn't depend on
ld-android.
Bug: 152255951
Bug: 154310674
Change-Id: Ie5ec02aebdf00b75756a6eda40db821488d30065
This partially reverts https://r.android.com/1278193. The propagated
symbol files contain all versions, even though only one is applicable
in a given SDK snapshot.
It's uncertain what repercussions this might have, but one is that if
we were to update a snapshot for a fixed version then it might change
because the symbol file contains new versions that aren't applicable.
Since the symbol file isn't actually needed at this point it's better
to wait with this step until the use cases for it are more clear.
Test: m nothing
Test: Create an SDK snapshot with Bionic libs, drop it into a
master-art tree without bionic/ in it, build ART APEXes, and check
that the Soong phase completes (specifically that the stubs are
detected even without symbol files).
Bug: 152481980
Change-Id: Ic79f89bc6d11d0b6552fa20791f5680ff9a40c0d
Necessary to make the APEX build logic treat the libraries as API
boundaries rather than dependencies to bundle.
The .so files in the snapshots are the compiled stub libraries in this
case. They are strictly speaking redundant since they can be generated
from the .map.txt files in the snapshots, but doing that would require
extending the cc_prebuilt_library(_shared) module types with a full
compiler pass etc, and that would break a lot of assumptions in the cc
package.
Test: m nothing
Test: Create an SDK snapshot with Bionic libs, drop it into a
master-art tree without bionic/ in it, build ART APEXes, and check
that the Soong phase completes (specifically no errors about various
APEX libs requiring libc that is not available to them).
Bug: 152481980
Change-Id: I31b928e6261198b6dd6f6b17196e714f07b64172
Necessary to get correct prebuilts for many Bionic libs.
Cleaned up numerious "system_shared_libs: []" from test fixtures, since
they otherwise would need correction in the expected results, and it is
better to have a single test focused on testing system_shared_libs
propagation.
Test: m nothing
Bug: 152255951
Change-Id: If2e8a5296223e6281d833312660e8e9e4cd184c0
Previously, when an sdk snapshot only supported a single os type the
compile_multilib was set based on the multilib usages by the members
of that variant. After the change to support multiple os types per
snapshot the multilib setting was based on the multilib usages across
all the members of all sdk variants. That meant that if one os type
used only "64" and the other used "both" then they would both be
treated as "both" leading to missing variants when the snapshot was
unpacked.
This change tracks the multilib usages per os type and adds a property
for each one.
It intentionally changes a couple of tests:
1) Either by adding compile_multilib that is missing.
2) By targeting it at a specific os type instead of host.
The latter change is important to prevent the snapshot from being
used on a host platform (which will match the host target section)
but which is a different os type to the ones supported by the
snapshot.
Bug: 142935992
Test: m nothing
Change-Id: I883919b644292c3d019db223bb4fd5c11b39591f
Added link type support in module creation code as cc_prebuilt_library
has to provide both shared and static libraries.
Had to add some calls to FirstUniquePaths() around the include dirs
settings in library_sdk_member as otherwise the shared variant contains
duplicate include dirs (presumably because it gets one set from its
dependency on the static variant and one set of its own). That
difference in include dirs causes problems in the generated
cc_prebuilt_library.
Test: m nothing
Bug: 142918168
Change-Id: Ie7f23fc2341c83c7814cc98e3970df4f5d4c8423
Test: m nothing
Test: Add
sdk {
name: "runtime-module-sdk",
native_shared_libs: [
"libc",
"libdl",
"libm",
"ld-android",
],
native_objects: [
"crtbegin_dynamic",
"crtbegin_static",
"crtend_android",
],
}
to bionic/apex/Android.bp. Then:
build/soong/scripts/build-aml-prebuilts.sh runtime-module-sdk
Take the generated runtime-module-sdk-current.zip and unzip into a
master-art tree without bionic/, edit the generated Android.bp to
extend cc_prebuilt_* modules with:
nocrt: true,
stl: "none",
system_shared_libs: [],
apex_available: ["//apex_available:anyapex"],
recovery_available: true,
vendor_available: true,
ramdisk_available: true,
Then "m com.android.art.debug". This passes Soong but fails in the
build step because more members are required.
Bug: 148934017
Change-Id: I2ab8f6aadb1440b325697cae4a8ed761c62d15d2
Treehugger Robot