armeabi libatomic.a uses ifuncs

[DoDoENT]

When compiling our app with libc++ runtime and clang, as soon as first native method is called, we get a SIGSEGV with invalid stack trace (stack trace points to file and line number which can be reached only from completely different JNI method).

When using libc++ + GCC, app crashes when attempting to throw c++ exception.

GNU STL + clang and GNU STL + GCC combinations work correctly. We've tested on Samsung Galaxy Ace: armeabi ABI + Android 2.3 (API level 10).

[DanAlbert]

Most likely this is a dup of http://b.android.com/181517. I assume this was libc++_shared?

[DoDoENT]

@DanAlbert I cannot open your link. It was libc++_static.

Edit: http version of link works for me.

[DanAlbert]

Oops. Link fixed, but I guess not then...

[DoDoENT]

Well, the libc++ + clang version crashed with SIGSEGV immediately when calling native method and stack trace was completely wrong.

Hopefully I will have time to investigate this further next week and probably prepare some sample.

[DanAlbert]

E/AndroidRuntime( 5460): FATAL EXCEPTION: main
E/AndroidRuntime( 5460): java.lang.ExceptionInInitializerError
E/AndroidRuntime( 5460): at java.lang.Class.newInstanceImpl(Native Method)
E/AndroidRuntime( 5460): at java.lang.Class.newInstance(Class.java:1409)
E/AndroidRuntime( 5460): at android.app.Instrumentation.newActivity(Instrumentation.java:1021)
E/AndroidRuntime( 5460): at android.app.ActivityThread.performLaunchActivity(ActivityThread.java:1561)
E/AndroidRuntime( 5460): at android.app.ActivityThread.handleLaunchActivity(ActivityThread.java:1663)
E/AndroidRuntime( 5460): at android.app.ActivityThread.access$1500(ActivityThread.java:117)
E/AndroidRuntime( 5460): at android.app.ActivityThread$H.handleMessage(ActivityThread.java:931)
E/AndroidRuntime( 5460): at android.os.Handler.dispatchMessage(Handler.java:99)
E/AndroidRuntime( 5460): at android.os.Looper.loop(Looper.java:130)
E/AndroidRuntime( 5460): at android.app.ActivityThread.main(ActivityThread.java:3683)
E/AndroidRuntime( 5460): at java.lang.reflect.Method.invokeNative(Native Method)
E/AndroidRuntime( 5460): at java.lang.reflect.Method.invoke(Method.java:507)
E/AndroidRuntime( 5460): at com.android.internal.os.ZygoteInit$MethodAndArgsCaller.run(ZygoteInit.java:839)
E/AndroidRuntime( 5460): at com.android.internal.os.ZygoteInit.main(ZygoteInit.java:597)
E/AndroidRuntime( 5460): at dalvik.system.NativeStart.main(Native Method)
E/AndroidRuntime( 5460): Caused by: java.lang.UnsatisfiedLinkError: Cannot load library: reloc_library[1479]: 67 unknown reloc type 160 @ ( 379)
E/AndroidRuntime( 5460): at java.lang.Runtime.loadLibrary(Runtime.java:434)
E/AndroidRuntime( 5460): at java.lang.System.loadLibrary(System.java:554)
E/AndroidRuntime( 5460): at com.example.hellojni.HelloJni.(HelloJni.java:64)
E/AndroidRuntime( 5460): ... 15 more
W/ActivityManager( 106): Force finishing activity com.example.hellojni/.HelloJni

Relocation 160 is an ifunc. After some readelf and some objdump, tracked that down to libatomic. Android doesn't support ifuncs until M, so this is going to require a fix to GCC...

[DanAlbert]

A little more testing shows that this is specific to GCC, which is why the libc++ test suite didn't catch it (we don't run the tests with GCC yet because our version of libc++ is far too old for the test suite to be very GCC compatible).

[cjhopman]

Maybe the title should be updated, this is much worse than the title implies:

The failure isn't limited to libc++ runtime, and it affects android up until 5.0 (not just the rather old 2.3).

Would it be possible to recompile libatomic with HAVE_IFUNC=false? After a cursory look it appears to support that.

[DanAlbert]

Yeah, the fix is nearly done. The changes to our GCC were https://android-review.googlesource.com/#/c/210660/ and https://android-review.googlesource.com/#/c/212250/. I just need to actually pull the new prebuilts into the NDK.

Updated the title to reflect the issue more clearly.

[DanAlbert]

Just as an FYI, not clear if this will be in the r12 beta that I'll be kicking out in a couple weeks, but it will definitely be in the r12 release in June.

[DoDoENT]

@DanAlbert, we didn't experience that crash on Android 4.0, 4.1 and 4.4 - only on 2.3. Are you sure this affects all android versions up until 5.0?

EDIT: sorry, I didn't notice that this is solely an armeabi issue (i.e. not present in armeabi-v7a and x86, which we used for testing on 4.0, 4.1 and 4.4).

[DanAlbert]

Fix is in for r12.