1525510 - elfhack causes PGO'd Android builds to crash when loading libxul.so

Status: RESOLVED FIXED

(Firefox Build System :: General, defect)

Description

[Michael Shal [:mshal]]

I'm testing out PGO-enabled Android builds by generating an instrumented build (with -fprofile-generate=/data/local/tmp), and triggering the profile data writing by sending a signal to the native process via adb since the normal Android process shutdown doesn't call atexit() handlers. This gets us profile data for libxul.so, at least, which is then used on a subsequent build with -fprofile-use. Just mentioning this since these are differences from Linux PGO builds, which way or may not be relevant.

Unfortunately the final profile-use build crashes on startup when loading libxul.so, which glandium traced to a weird jump instruction in elfhack. Re-building the profile-use stage with --disable-elf-hack makes the crash go away, but we will ultimately need a fix in elfhack.

Comment 1

[Mike Hommey [:glandium]]

The problem is the address difference between the elfhack-inject code and the location of the original init function in libxul is too large for ARM instructions. Elfhack would need to detect this is the case and inject a trampoline, or something akin to this.

Comment 2

[Mike Hommey [:glandium]]

Attachment: Bug 1525510 - Add a trampoline to call original_init when it's not possible directly.

Comment 3

[Mike Hommey [:glandium]]

Can you test the attached patch?

Comment 4

[Michael Shal [:mshal]]

Yep, that works for me. Thanks for the quick fix!

Comment 5

[Pulsebot]

Pushed by [e-mail]:
https://hg.mozilla.org/integration/autoland/rev/94ab0cf7c69f
Add a trampoline to call original_init when it's not possible directly. r=froydnj

Comment 6

[Oana Pop-Rus]

https://hg.mozilla.org/mozilla-central/rev/94ab0cf7c69f