| Age | Commit message (Collapse) | Author |
|---|
|
We used to look if there's any app image, however this can return true
when a shared library has an app image.
Test: test.py
Bug: 330564006
Change-Id: I0aa68edb14491db2fad6ea1d1c1b40d6c28eccff
|
|
Bug: 260881207
Test: presubmit
Test: abtd app_compat_drm
Test: abtd app_compat_top_100
Test: abtd app_compat_banking
Change-Id: Iacec568941c5baee57088e67ed410ba18f6abff2
|
|
This makes sure that we generate the best image.
Also add a couple of ScopedTrace to profile runtime app image
generation.
Test: test.py
Test: locally start an app, see image containing all classes.
Bug: 260557058
Change-Id: I1c2690c54a4c815ac575cada476f48948bd63131
|
|
Suppress madvise calls to non-primary vdex/oat files whenever possible.
Note that this suppression relies on AppInfo registration happening
relatively early in the app lifecycle, which is generally the case.
Until the registration is made, we madvise these artifacts as is
done with the current behavior.
This should avoid cases during app startup where Dynamite-like
dependencies (e.g., for GMS Core) currently trigger a full madvise
of all dependent vdex/oat files, which is generally unnecessary
and potentially wasteful for such cases.
Note that this change moves the oat madvise call out of OatFile's
constructor into OpenDexFilesFromOat. This makes it easier to reason
about ordering relative to other madvise calls (for dex/vdex), but
should be almost identical in terms of timing.
A/B perf tests shows that this slightly improves app startup time
across a suite of apps, with apps that don't load secondary APKs/dex
files not being impacted at all.
Bug: 235130726
Test: m + flash and verify GMS vdex/oat files not madvised by apps
Change-Id: I41374d5ae9f91f39e399ac21615a821307c56b6e
|
|
The compiler filter / reason reporting was not accurate for a variety
of reasons. (e.g. reporting was only done at startup, it was relying
on imprecise APIs and had errors in the logic).
In order to keep track of the precise optimization status, this CL
introduces the concept of AppInfo, which encapsulates the data about
the application / system server code paths, their optimization status
and possible other metadata (e.g. profiles).
To populate it, we rely on 2 distinct events:
1) The framework calling VMRuntime#registerAppInfo to inform the
runtime about the applications code paths and their types (e.g. primary,
split, secondary).
2) Class loading, when we can determine the actual optimization status
like filters, reasons, and whether or not we can load the odex files.
These events may happen in any order so we could deal with a partial
state at some point in time, but in the majority of cases they always
happen at Class Loading, followed by RegisterAppInfo.
This CL also deletes the OatFileManager#getPrimaryOatFile which was
a misleading API as it didn't work in most cases. It also adds more
tests to the metrics/reporting infra for previous missing or
unimplemented cases.
Test: gtest
Bug: 170149255
Change-Id: If0a7a25d06ff6fb89fe4861139b7dee61c05814d
|