The core problem I am trying to solve is making sure that stub
libraries in Bazel have a single apex available. In Soong, this will be
enforced using some graph walk which is not easy to port to Bazel.
However, we might need to revisit this when we build the enforcement
mechanism in Bazel. We likely need a `test_for` on the top level
apex_test Bazel rule so that the test apex is _allowed_ to link against impl
of the library it is trying to test.
(This CL retricts this to cc_library, I can expand this to other modules
if this is the right approach.)
Bug: 277651159
Change-Id: Iaeec22c5626df79a33785c766ed29102b1da403e
SortedUniqueStrings and FirstUniqueStrings dedupes repeating elements
and returns the deduped list. Currently, it also modifies the input list
in-place, which causes non-determinisitc failures like b/275313114
Operate on a copy of the input so that the input remains untouched.
SortedUniqueStrings is O(NlogN) and FirstUniqueStrings is ~O(N), so
creating a copy (O(N)) should not result in major performance regressions.
Numbers for this single unit test:
```
go test . -run TestStubsForLibraryInMultipleApexes -v -count 1000
Before: 174s
After: 172s
```
Test: go test ./android
Test: go test . -run TestStubsForLibraryInMultipleApexes -v -count 1000
Change-Id: Id859723b2c2ebdc0023876c4b6fabe75d870bad7
Modifying the list in-place causes some non-determinism.
Test: go test ./cc -run TestStubsForLibraryInMultipleApexes -count 1000
Bug: 275313114
Change-Id: Ia2519e146d97667ad5900cf68ab9935fcbaf08a4
If any of apexes in apex_available is an apex_test, then that name will
be propagated down from that apex to each apex variant. This metadata
will be used to enforce that stub libraries cannot have more than one
apex_available.
This logic is necessary so that bp2build can select the correct
stub/impl.
(To avoid replicating this complexity in Bazel, we should consider
dropping the test apexes in Bazel BUILD files, next CL)
Bug: 277651159
Test: go build ./apex
Change-Id: I63617c1dc2a2d5c9cd7758c416fec7b4db1f10a7
go/Android.bp mentions that ["//apex_available:platform"] is the
default, but currently it was []. This change does not create any
additional module variants.
(Noticed this for libz, I was expecting its apex_available to be
platform and not an empty list)
Test: TH
Change-Id: I9af06f813b1a1d7b716939874f469bd2e1ce4d14
This relands aosp/2457063. The original change broke T and U since those
branches still contain soong modules of type (kind+level). Those soong
modules have been cleaned up now
Test: Used go/abtd to test T and U branches with this change
Bug: 208456999
Change-Id: I0ef7933c055f88cb512a02108f1173e51156ef1c
min_sdk_version signifies device version and does not need an sdkKind to
describe it fully. Update the type and cleanup existing usages. As a
side benefit, we also get better error handling since users can no
longer enter something like `public_30` as a valid min_sdk_version in bp
files
Will do a similar cleanup for targetSdkVersion and maxSdkVersion in a
followup CL
Test: m nothing
Test: no change in ninja files (this should be a no-op)
Bug: 208456999
Change-Id: Ie6ae7e267d093c5e4787e82685daaca1021d202e
This reverts commit 502da3987a.
Reason for revert: b/274195633
```
In file included from out/soong/installs-aosp_cf_x86_pasan.mk:134984:
In file included from out/soong/Android-aosp_cf_x86_pasan.mk:981696:
In file included from build/make/core/soong_cc_rust_prebuilt.mk:76:
build/make/core/base_rules.mk:342: error: packages/modules/Uwb/service/uci/jni: MODULE.TARGET.SHARED_LIBRARIES.libuwb_uci_jni_rust already defined by packages/modules/Uwb/service/uci/jni.
```
Change-Id: Ic1ea6969e54c23a7d126eb0fb47ab6f2e44ee965
Currently, non-apex variants of modules that are in apexes are not
exported to make unless they're apex_available to the platform. This
means that you can't `m` those modules directly.
However, there is a workaround in the apex androidmk implementation that
emits make rules for the removed modules, but just redirects them to
build the apex itself. We want to remove that, but one of the problems
with doing so is that you can no longer `m` many modules afterwards.
To fix that, unhide the apex's dependencies from make. To ensure they're
not installed, call SkipInstall() on them, and update SkipInstall() to
be more strict by setting `LOCAL_UNINSTALLABLE_MODULE := true`.
Bug: 254205429
Test: Presubmits
Change-Id: Ib971981559f3b642ce6be8890679e994e1b44be0
Currently, non-apex variants of modules that are in apexes are not
exported to make unless they're apex_available to the platform. This
means that you can't `m` those modules directly.
However, there is a workaround in the apex androidmk implementation that
emits make rules for the removed modules, but just redirects them to
build the apex itself. We want to remove that, but one of the problems
with doing so is that you can no longer `m` many modules afterwards.
To fix that, unhide the apex's dependencies from make. To ensure they're
not installed, call SkipInstall() on them, and update SkipInstall() to
be more strict by setting `LOCAL_UNINSTALLABLE_MODULE := true`.
Bug: 254205429
Test: Presubmits
Change-Id: Ib094feb2c437ad50d8319c58caa997759e7ce32f
Also remove the soong hack that allowed us to build during the
transition
Bug: 243054261
Test: build + TH
Merged-In: I60bac1ec687f03a38f38240969a1c38a8e5bf92c
Change-Id: I60bac1ec687f03a38f38240969a1c38a8e5bf92c
The first hack has been introduced to facilitate the creation of the
com.android.btservices apex without the need to upload 100+ cl close to
the release date.
We now want to delete com.android.bluetooth apex and replace all the
apex_available rules to target the new apex name
It will be way much easier to do this in separate CL and without forcing
us into a single topic.
Bug: 243054261
Test: Build
Change-Id: Ia03698db839db4684871eb38c088f88d6bdcba4b
This hard-coded list causes inconsistent behavior regarding -target
triple and _ANDROID_APEX_MIN_SDK_VERSION_ macro for native modules.
Bug: 158059172
Test: m
Merged-In: Iae3ecb3bfaefc2ee73ed38a9268c68b6673f30c5
Change-Id: Iae3ecb3bfaefc2ee73ed38a9268c68b6673f30c5
(cherry picked from commit 870ab9c0f78327d92a871ac58c0337db8ae92245)
Set min_sdk_version/apex_available for modules used by
com.android.permission APEX.
Bug: 147364041
Bug: 158059172
Test: m
Change-Id: I93ff72ed765354d44a2ccb5105b5f2820f0358f5
Hard-coded min_sdk_version is not necessary for kotlinx-* modules
because they are already marked by updater script.
Bug: 158059172
Test: m nothing
Change-Id: I42cf26c9390244f93059af68027cca4f8518a828
When a library is included in two APEXes whose platform_apis settings
are different, two apex variants of the library is created: apex1000 and
apex1000_private.
This change was introduced with ag/15061306, especially by the commit
[1].
However, that part should be reverted because it actually creates
unnecessary variants. It's unnecessary because the two variants of the
library are compiled (excluding the linking) exactly the same. If a
private symbol of its dependency was actually used when compiling the
apex1000_private variant, then the other apex1000 variant wouldn't have
been built because that private symbol must have caused a linkage error.
[1] https://googleplex-android-review.git.corp.google.com/c/platform/build/soong/+/15061306/2..4/android/apex.go#b527).
Bug: 228785792
Test: m
Change-Id: Id58d3e98a51de5e628ca72ef86e9cd11b0ee8971
Inidividual boot or system server jars may have higher min_sdk_version
than the contianing apex, since the runtime respects the values of
min/max_sdk_version; e.g. runtime would not load a boot jar with
higher min_sdk_version. This allows shipping new boot jars via apexes
that target older platforms.
Bug: 190818041
Test: presubmit
Change-Id: I08ec0b4463a17bc8265b948fe09da55eb4e52ac3
This is to help following refactor for individual modules to have their
own version of checking where needed.
For example, apk in apex may want to enforce it's own version of
CheckMinSdkVersion.
Bug: 205923322
Test: presubmit
Change-Id: Ia2fad6c52af39e21f65385bcb283f1e3adab5548
Many modules requiring min_sdk_version have been used without setting
it, but hard-coded in allowlist.
Bug: 158059172
Test: m
Change-Id: Ibb09ddfdb11df0791a28eb8a0a49f1780084fd95
Many modules requiring min_sdk_version have been used without setting
it, but hard-coded in allowlist.
Bug: 158059172
Test: m
Change-Id: I6d11425f9c5db11cd52dd81f7500e4424555bfb5
The property is used to allow non-updatable APEXes to use platform APIs
(e.g. symbols marked as "# platform-only").
Bug: 191637950
Test: m com.android.virt com.android.compos
Merged-In: Id2410b4e38a78ec2146a42298840954381a7c472
Change-Id: Id2410b4e38a78ec2146a42298840954381a7c472
(cherry picked from commit fb63625a7f)
Neither InApexVariants nor InApexModules should have them. This allows
us to get rid of InApexVariantByBaseName as well.
Test: m nothing
Test: m nothing SOONG_CONFIG_art_module_source_build=false
Bug: 180325915
Change-Id: Icbe4e025ce1a4c8dd258ff95d326ca2f27905188
CopyDirectlyInAnyApex was documented to copy from child to parent, but
was copying from parent to child. It is unused, so reverse it to
match the documentation.
Bug: 183759446
Test: next CL
Change-Id: I950c9b5416d66e83d76ca489aeb5e0572e005d5d
Consider this case:
apex {
name: "com.android.foo",
native_libs: ["foo"],
}
override_apex {
name: "com.mycompany.android.foo",
base: "com.android.foo",
}
cc_library {
name: "foo",
}
There are two APEXes defined: "com.android.foo" and
"com.mycompany.android.foo" which is a copy of "com.android.foo" with
some properties overridden (e.g. signing keys).
The module "foo" is mutated into two variants by the apex mutator: the
platform variant and the apex variant. The former has the variation name
"" and the later has "apex<min_api_ver>" which usually is "apex10000".
Internally, the apex variant has an alias "com.android.foo".
ApexInfo.InApexVariants() returns only "com.android.foo" when called for
the module "foo".
We can see that the information that "foo" is also part of
"com.mycompany.android.foo" is completely lost. This is causing problem
when we compare the apex membership by their "soong module name", not
the "apex name". In the example above, the two modules have different
soone module names, but have the same apex name: "com.android.foo".
To fix that, this CL introduces a new field `InApexes` to the `ApexInfo`
struct. It has the actual name of the APEXes that the module is part of.
With the example above, `InApexes` is ["com.android.foo",
"com.mycompany.android.foo"].
Bug: 180325915
Test: m nothing
Test: m nothing on non-AOSP targets with ag/13740887 applied.
Change-Id: I4e7a7ac5495d2e622ba92a4358ed967e066c6c2e
.. in preparation for the upcoming change. This change doesn't alter any
behavior.
InApexes is a misleading name. People expects that it has the list of
soong module names of the APEXes that a module is part of. So, for
example, `core-oj` is a part of both `com.android.art` and
`com.google.android.art`. However, in reality, that's not true. The
field has `com.android.art` only. This is because the two APEXes
(android and Google) have the same apex name which is `com.android.art`.
That apex name is used in various places like the `apex_available` and
allows us to keep using the same name regardless of whether the APEX is
overridden or not.
However, this is causing problems in some cases where the exact list of
soong module names is required. The upcoming change will add a new field
to handle the case and the new field actually will get the name
'InApexes'. So, the existing field is renamed to a less misleading name
`InApexVariants`.
Bug: 180325915
Test: m nothing
Change-Id: I0c73361b452eddb812acd5ebef5dcedaab382436
ApexInfo is not part of the properties struct. It can handle structs
having private fields.
Bug: 1663140
Test: m
Change-Id: Ib07d4410f0ce187c9de347da34b84b814b2eb537
