WIP: Register JITTed functions with libunwind

[Keno]

I am extremely surprised that we didn't do this before - I guess libunwind
was just pretty good at figuring it out anyway, unlike the OS X unwind
library. However, for better backtraces, register the JITted functions
with libunwind, which also prevents libunwind from trying to access
invalid memory in a good number of cases.

WIP, because libunwind only supports 32 bit offsets between the .text
and .eh_frame sections, but MCJIT spreads the farther apart. I am in
the process of fixing this in libunwind.

[ihnorton]

🛅

[Keno]

thanks

[Keno]

I just mailed an appropriate patch upstream. Will update here once it's accepted.

[yuyichao]

Do you have a reference for it?

Also, do you have any idea if this could be the cause of #11691 ? I tried to enable the test on top of your branch and it doesn't fixed the problem but maybe it could with the libunwind patch.

[Keno]

Yes, quite possibly. The failure mode is is a ReadOnlyMemoryError. However, note that this is LLVM 3.7 only, because the old JIT did not emit the appropriate unwind info.

[JeffBezanson]

Yes, could you post a gist of the patch just in case?

[Keno]

https://gist.github.com/Keno/152e0eb513973dad5726

[carnaval]

@Keno any news ? libunwind often stops unwinding early with llvm-svn on jited code. I assume that it's because we often emit functions without frame pointers now, although I'm not sure what libunwind fallbacks to when proper unwind info is not provided.

[carnaval]

also if it's less bothersome than libunwind upstream we can probably fix that by overriding the default allocator in the SectionManager thing.

[Keno]

I reminded upstream about the patch. Let's see if it sticks.

[yuyichao]

Seems that the patch have been committed a few month ago? http://git.savannah.gnu.org/gitweb/?p=libunwind.git;a=commit;h=8afc33ce9f20538f048838e18680c120867ee0d2

[Keno]

Oh, huh. That's nice! Nobody told me about it.

[yuyichao]

There also seems to be other improvements on libunwind master since I'm getting much better backtrace with it even without yours or @carnaval 's patch. I'm wondering when will they make a new release (given that the last one is more than 3yrs ago).

[tkelman]

Hopefully soon after they merge the musl patch I sent them.

[carnaval]

Alright, what is holding this up ? Can we move on to libunwind master/carry the patch in our tree and merge this ? @Keno ?

We need this also to get inlined frame in backtraces if we start using DWARF for that since even though libunwind guesses correctly it certainly can't invent those ghost frames.

[yuyichao]

I just rebased this on current master (old branch backed up at yyc/kf/linuxunwind). Compilation passes with libunwind master but given this needs the UNW_INFO_FORMAT_IP_OFFSET format which is not available on the released version of libunwind, the linux CI will probably fail.

The libunwind master has a lot of features we need (much better unwind on x86 (and no segfault), aarch64 support, new dynamic unwind info format) and since there hasn't been a release since 3 years ago, and given libunwind isn't moving as fast as LLVM, I think we could try to ship the git master version of libunwind (fetch from git/make our own tarball) instead of the current released one. (Unless someone can convience libunwind to make a new release soon)

[yuyichao]

I pushed a commit to use my own tarball built from the libunwind master, let's see what the CI thinks.

[tkelman]

Can we move that somewhere more permanent? Or bother upstream into making a new release?