Memory churn is high when swapping the ResourcesImpl of a Resources
object. Each time Resources#setImpl is invoked, all themes based on
that Resources object are assigned new ThemeImpl objects that are
created using the new ResourcesImpl.
ThemeImpls can only belong to one Theme object, so the old
implementation is discarded and the theme takes ownership of the new
ThemeImp.
This creates performance problems when framework overlays are toggled.
Toggling overlays targeting the framework causes all themes across all
processes to recreate and reallocate all of their themes. By rebasing
the ThemeImpl on the new ResourcesImpl without deallocating the native
theme memory, we reduce churn and produce less garbage that needs to
be garbage collected.
Bug: 141198925
Test: atest libandroidfw_tests
Test: atest ResourcesPerfWorkloads
Change-Id: I03fb31ee09c9cfdbd3c41bcf0b605607dab54ed7
Currently there is a limitation where ResourcesLoaders cannot be used
on Resources object not created through ResourcesManager. This change
creates an update handler for Resources objects that are not registered
with ResourcesManager.
The handler changes the loaders on the asset manager owned by the
Resources instance.
Bug: 151666644
Test: atest ResourceLoaderValuesTest
Change-Id: I5a89f686386bdb088dc964014e7becc0c2b4770f
Hard-coded values in styles can be copied between AssetManagers even if
the source package is not present in the destination AssetManager. Only
references and strings should be prevented from being copied over
because they would be invalid in the destination AssetManager.
Bug:126400561
Test: manual
Change-Id: I970a3e961763b2c003c15b950d864a9a0b615022
The docuemntation of setTo states that the resources two styles from
different AssetManagers have in common will be set in the destination
theme. This change adds this functionality. The package ids of the
attributes, the package ids of reference values, and the cookie of
attribute values have to be rewritten to match the destination
AssetManager. This change can later be made more generic if rewriting
references between packages is needed elsewhere.
Bug: 115897657
Test: libandroidfw_tests and manual test of app specified in the bug
Change-Id: Iee999ea2cc8473168cac11aaf3c34e14c958e5ae
This change brings Theme ApplyStyle down to 2x the original performance
and Theme attribute retrieval to less than the original performance.
Yay!
Benchmarks ran on marlin-eng
----------------------------------------------------------------------
Benchmark Time CPU Iterations
----------------------------------------------------------------------
BM_ThemeApplyStyleFramework 8540 ns 8500 ns 82105
BM_ThemeApplyStyleFrameworkOld 5280 ns 5258 ns 148849
BM_ThemeGetAttribute 8 ns 8 ns 88388549
BM_ThemeGetAttributeOld 11 ns 11 ns 63394463
ApplyStyle still takes some time, and the weird thing is that if I
switch the data structure of ThemeType to use an
std::vector<ThemeEntry>, the performance becomes better than the
original implementation! The issue is that std::vector<T> takes up 24
bytes, which would make Themes take up 8 more bytes per ThemeType, which
is unacceptable. Still trying to isolate where the performance gain is
coming from.
Test: make libandroidfw_tests && $ANDROID_BUILD_TOP/out/host/<host>/nativetest64/libandroidfw_tests/libandroidfw_tests
Test: make libandroidfw_benchmarks && adb sync system && adb sync data && adb shell /data/benchmarktest64/libandroidfw_benchmarks/libandroidfw_benchmarks
Change-Id: I0e7a756afd44b6aac1521e69c2b907258c262d3e
Theme copying should behave the way it did with the old AssetManager
(copy only the framework attributes when copying from a Theme object
from a different AssetManager).
Cleanup the dependencies on libziparchive in ApkAssets.
Test: make libandroidfw_tests
Test: out/host/<platform>/nativetests64/libandroidfw_tests/libandroidfw_tests --testdata=frameworks/base/libs/androidfw/tests/data
Change-Id: I973f7e6eb14ce311306e2ec66a623a4790c8d233
This adds support for shared resource libraries in the new
ResTable/AssetManager implementation.
The dynamic package map encoded in resources.arsc is parsed
and stored with LoadedArsc, and combined to form a resolved table
in AssetManager2.
Benchmarks show that this implementation is an order of magnitude
faster on angler-userdebug (make libandroidfw_benchmarks).
Test: libandroidfw_tests
Change-Id: I57c80248728b63b162bf8269ac9495b53c3e7fa0
The multiwindow model and Resources-per-activity
model that came in N puts greater demands on AssetManagers.
They are created whenever window dimensions change, which
can be frequently. There is a need to be able to cheaply
create a new AssetManager for each Activity, which shares
a lot of underlying state.
In order to make the creation of AssetManagers cheap,
we need a new implementation of the native AssetManager
and ResTable to support immutable representations of
APKs. This new data structure/class is ApkAssets.
ApkAssets have the same functionality of an AssetManager, except
that they operate on a single APK, and they do not do any caching.
Once loaded, they are immutable.
ApkAssets will be exposed as a Java object, with its implementation in
native code. The existing Java StringBlock will be owned by ApkAssets,
which means that Strings can be shared across AssetManagers.
ApkAssets can be cached by the ResourcesManager. Creating an AssetManager
requires only a list of ApkAssets and a configuration.
AssetManager2 (named with the suffix '2' for now while transitioning
to the new implementation) caches bags that are accessed.
Since ApkAssets are expected to be kept around longer, they do more validation
of the resource table, which cause slower load times. Measured on an angler-userdebug,
loading the framework assets takes 11ms with ApkAssets, and 2ms with the old
AssetManager implementation.
The tradeoff is that there does not need to be any security checks once an ApkAssets
is loaded, and regular resource retrieval is faster. Measured on an angler-userdebug,
accessing resource (android:string/ok) with many locales takes 18us with AssetManager2,
and 19us with AssetManager (this is per resource, so these add up).
Test: make libandroidfw_tests
Change-Id: Id0e57ee828f17008891fe3741935a9be8830b01d
Tests would expect parts of the APK to be unzipped and
maintained. Instead, we now decompress the required files
from the test APKs on test setup. This simplifies
test maintenance substantially.
Test: make libandroidfw_tests && libandroidfw_tests --testdata=frameworks/base/libs/androidfw/tests/data
Change-Id: I3d2100af22df913e02401dedcf9842cdb32b2a3b
