Bug #17511: Segmentation fault when compiled with -O2 or higher on ARM Android

• Status: Closed
• Assignee: -
• Target version: -
• ruby -v: ruby 3.0.0p0 (2020-12-25 revision 95aff21468) [arm-linux-android]
• Backport: 2.5: UNKNOWN, 2.6: UNKNOWN, 2.7: UNKNOWN, 3.0: UNKNOWN

[xtkoba (Tee KOBAYASHI)]

To reproduce a segmentation fault, compile Ruby with CFLAGS=-Oz using Android NDK r22 (Clang 11.0.5), and input the following:

require 'rubygems'
pp

A segmentation fault also occurs with CFLAGS=-O2 when Up or Down key is pressed in IRB.

A workaround is to disable using __builtin_setjmp:

./configure --with-setjmp-type=setjmp

I have no idea if there are similar issues or not in non-Android Linux systems on 32-bit ARM.

Bug #14480 seems to suggest that there was a similar case on AArch64 Linux with GCC.

[mame (Yusuke Endoh)]

• Related to Bug #14480: miniruby crashing when compiled with -O2 or -O1 on aarch64 added

[mame (Yusuke Endoh)]

• Status changed from Open to Closed

Applied in changeset git|b9c1b3f8d82b6f2517786b2f9e06683909af34c3.

---

configure.ac: disable using __builtin_setjmp on ARM Android

A patch from Tee KOBAYASHI [e-mail]. [Bug #17511]

[mame (Yusuke Endoh)]

Thank you, I've imported your patch.

BTW, I'm curious for what purpose you are attempting to build Ruby for Android.
As I said, we currently have no CI or development environment for Arm Android.
So it is difficult for us to test the code for the environment.
I thought your patch for this ticket was very trivial, so I've imported it.
But if it was more complicated, such as fixing a test, it would be difficult for me to commit it.

If there is a practical use case for Android Ruby, I'm willing to try preparing a CI for the environment.

[xtkoba (Tee KOBAYASHI)]

Thanks for the merge!

My running Ruby on ARM Android is JUST FOR FUN :-)

I think it would be more productive to prepare a CI for AArch64 Android if you could, rather than for nearly out-of-date 32-bit ARM Android.

[xtkoba (Tee KOBAYASHI)]

• File ruby-ec-volatile.patch ruby-ec-volatile.patch added

I revisited this issue, and probably found another workaround: to add volatile qualifier to the ec argument of the functions rb_ec_tag_jump and vm_exec, as in the attached patch. I have no idea exactly what it means, and I don't expect it to be merged. After all, there is no problem unless __builtin_{set,long}jmp are in use.

[xtkoba (Tee KOBAYASHI)]

Just for the record, there turns out to be a Clang/LLVM bug in its __builtin_longjmp implementation for ARM Linux (more specifically for non-Darwin and non-Windows) for which there seems to be no reliable workaround on user's side.

https://bugs.llvm.org/show_bug.cgi?id=50202