Generally macrobenchmark is capable of measuring performance of a target app without changes to the app, but the big exception is the target's inclusion of profileinstaller, when using compiling with baseline profiles
Macrobench currently requests the app to dump its own baseline profiles via profileinstaller so they're visible to art, but this doesn't work if an app strips the receiver or doesn't use profileinstaller (which would otherwise be pretty reasonable if minApi >= 28)
To support apps without the profileinstaller reciever, macrobench could instead:
Unzip the target app's apk(s)
Extract out in-APK baseline profiles (legacy format)
Use profileinstaller to convert them to the appropriate format
Reinstall the target app together with converted baseline profiles, as DM
This has the downside of resurrecting the old issue of not persisting data across compilation, but would make it possible to measure performance of apps without modifying them.
Description
Generally macrobenchmark is capable of measuring performance of a target app without changes to the app, but the big exception is the target's inclusion of profileinstaller, when using compiling with baseline profiles
Macrobench currently requests the app to dump its own baseline profiles via profileinstaller so they're visible to art, but this doesn't work if an app strips the receiver or doesn't use profileinstaller (which would otherwise be pretty reasonable if
To support apps without the profileinstaller reciever, macrobench could instead:
This has the downside of resurrecting the old issue of not persisting data across compilation, but would make it possible to measure performance of apps without modifying them.