unwinding problem on android

[yichoi]

last make check-stage2-T-arm-linux-androideabi-H-x86_64-unknown-linux-gnu on f74b8f0d

failures:
    [run-pass] run-pass/cleanup-copy-mode.rs
    [run-pass] run-pass/fail-in-dtor-drops-fields.rs
    [run-pass] run-pass/tempfile.rs
    [run-pass] run-pass/terminate-in-initializer.rs
    [run-pass] run-pass/unit-like-struct-drop-run.rs
    [run-pass] run-pass/unwind-box.rs
    [run-pass] run-pass/unwind-resource.rs
    [run-pass] run-pass/unwind-resource2.rs
    [run-pass] run-pass/unwind-unique.rs
    [run-pass] run-pass/vector-sort-failure-safe.rs

has same result

fatal runtime error: Could not unwind stack, error = 9
Segmentation fault
stdout:
------------------------------------------

------------------------------------------
stderr:
------------------------------------------
task '<unnamed>' failed at 'explicit failure', /home/yichoi/rust_latest/src/test/run-pass/vector-sort-failure-safe.rs:69

[alexcrichton]

Hm, I find it odd that error 9 is getting returned, because I find any coorespondance to an error higher than 8... If you take a look for definitions of _Unwind_Reason_Code on android, does anything come up for the error 9?

[vadimcn]

In ARM EABI there's also _URC_FAILURE = 9 /* unspecified failure of some kind */

Hmm, also, looks like on ARM, _Unwind_Exception needs more room for private unwinder data. The total size of the struct should be 88 bytes on ARM.

At the moment I am not set up to test on Android (can this even be done without a real device?),
@yichoi, would you mind trying the following: in libstd/rt/unwind.rs, replace _Unwind_Exception's private_1 and private_2 fields with private : [_Unwind_Word, ..19] (or maybe 20), and see if that changes anything? Thanks!

[luqmana]

@vadimcn using the android emulator might be an option you can try.

[brson]

Changing private to [_Unwind_Word, ..20] appears to result in the tests failing to terminate. Odd...

[brson]

I'm trying to see if it otherwise works just using C++ throw but keeping all our other new machinery. Maybe __cxa_throw is doing something special on android.

[vadimcn]

trying to build android cross-compiler on windows, but no luck so far...

[brson]

I think the problem I was seeing is not that the tests are failing to terminate, but adb_run_wrapper.sh not coping with the following error, which itself is very peculiar:

CANNOT LINK EXECUTABLE: cannot locate symbol "_ZN2rt4util12dumb_println67h81a04d63b9f37141ed6a4d8d646e2997716a3d70842c8e2878dd5342a124ed28aV8v0.9.preE" referenced by "./unwind-box.stage1-arm-linux-androideabi"...

[brson]

Hm, I just did a successful test run on Android. I think something was causing my libstd on the emu to be stale.

[brson]

This branch makes unwinding work on android by using a C++ try function and using the g++ personality fn: https://github.com/brson/rust/tree/androidwip. I don't know why.

[vadimcn]

I thought you said you've got it working... Or was that using C++ throw?

[brson]

I'm not sure. I'm having a hard time understanding which configurations work. I believe it works when android just uses C++ for everything, plus the g++ personality function. I'll probably try committing a patch that uses C++ for just android until we figure something else out.

[brson]

No, I have a patch that just completely uses the C++ unwinder for Android. I'll probably try to clean it up and get it merged and leave fixing the underlying problem for later.

[vadimcn]

@brson: I think I have a fix for Android. I still need to clean-up the patch, make sure tests still pass of other platforms and whatnot Do you want to hold off your C++ unwinding patch?