Re: [Tarantool-patches] [PATCH luajit v4 8/8] OSX/ARM64: Fix external unwinding.

[Sergey Kaplun via Tarantool-patches <tarantool-patches@dev.tarantool.org>]

Hi, again!

On 01.12.22, Sergey Kaplun via Tarantool-patches wrote:
> Hi, thanks for the explanations!
> LGTM, then.
> 
> On 30.11.22, Maxim Kokryashkin wrote:
> > 
> > Hi!
> > Thanks for the review!
> >  
> > > 
> > >>Hi, Maksim!
> > >>
> > >>Thanks for the patch!
> > >>
> > >>LGTM, but I have a bunch of questions to clarify it.
> > >>
> > >>On 28.10.22, Maksim Kokryashkin wrote:
> > >>> Contributed by Edmund Kapusniak. For more info,
> > >>> see #698 and #757.
> > >>>
> > >>> (cherry picked from commit c38747b626b978555324504ec29a110f6b04902f)
> > >>>
> > >>> To allow compiler generate compact unwind info generation
> > >>> for Mach-O, fp must point to the saved fp, and the frame
> > >>> must be specified relative to fp+16.
> > >>
> > >>Is there any link to documentation or source code to inspect this
> > >>behaviour?
> > >Unfortunately, there are no official docs for that. However, there is
> > >a community effort to create one here[1]. Also, this header file from
> > >the Apple’s sources is quite useful. Added both of them to commit
> > >message for those who will get the masculine urge to dive into this.
> 
> Meh, not much info about it in the mentioned doc.
> 
> I suppose that aarch64 ABI requires to save fp and lr by analog with arm
> arch. See the following comment for arm (unfortunately, there is no such
> comment for aarch64).
> https://github.com/gcc-mirror/gcc/blob/master/gcc/config/arm/arm.h#L812
> 
> But I'm still struggling to find some relative docs about it.
> 
> > >>
> > >>> ELF unwind info has
> > >>> been updated to also use fp+16 rather than sp+CFRAME_SIZE.
> > >>
> 
> <snipped>
> 
> > >>> Offset to pointer to personality routine specified as @GOT-. rather
> > >>> than @GOTPCREL.
> > >>
> > >>Does it mean that we use incorrect encoded offset (I see encoding for
> > >>offset is still the same) for our personality routine?
> > >>If so, maybe the other changes are just refactoring?
> > >No, that is not correct. Offset has changed, because any 
> > >`func@GOT` expression is translated into offset.
> > >Moreover, I doubt it is possible to fill in offset to GOT by hand.
> > >What goes after the pointer to the personality routine
> > >is LSDA, according to the Apple’s source[2]. @GOTPCREL and @GOT
> > >are interchangeable most of the time, except for the cases when signed
> > >32-bit RIP references are not enough for you, which seems to be the case here.
> 
> So, GOTPCREL allows you to "recalculate" reference to a function to
> 32-bit value and use it? Do I get the idea correct?

Yes, this is the issue mentioned here:
https://github.com/LuaJIT/LuaJIT/issues/698#issuecomment-841645665

And usage of @GOT-. fixes the build for M1.

> 
> > >>
> > >>>
> > >>> Re-enabled LUAJIT_UNWIND_EXTERNAL by default on OSX.
> > >>>
> > >>> Maxim Kokryashkin:
> > >>> * added the description for the issue and the test
> > >>>
> > >>> Resolves tarantool/tarantool#6096
> > >>> Part of tarantool/tarantool#7230
> > >>> ---
> 
> <snipped>
> 
> > >[1]:  https://faultlore.com/blah/compact-unwinding/#unwinding-tables-dwarf-cfi
> > >[2]:  https://opensource.apple.com/source/libunwind/libunwind-35.3/include/mach-o/compact_unwind_encoding.h
> > >--
> > >Best regards,
> > >Maxim Kokryashkin
> > > 
> 
> -- 
> Best regards,
> Sergey Kaplun

-- 
Best regards,
Sergey Kaplun