Re: Another unwinding problem on ARM

[Mathias Bauer]

Am 20.03.14 08:23, schrieb David Chisnall:
> On 19 Mar 2014, at 22:44, Mathias Bauer <mathias_bauer@gmx.net>
> wrote:
>
> > I'm curious about your opinion about this problem. The patch works,
> > but the question remains: should it be fixed in the runtime or in
> > the compiler?
>
> I think this is an LLVM bug.  I'll see Renato next week and talk to
> him about it then.
>
> I'm a bit hesitant to add the hack to the runtime, because it's
> possible that we'll have absptr encodings later that actually are
> absptr (e.g. in statically linked parts of the code or as a result of
> some rtld optimisation).  Given that we haven't yet had a release of
> LLVM with ARM EHABI support enabled by default, I'd be inclined to
> not officially support exceptions on ARM with anything before LLVM
> 3.5, and make sure that it works well in 3.5.

I agree with all of what you wrote. I only wrote this patch to point out 
the problem, I don't know what the compiler is expected to do on ARM. 
Meanwhile I've found some Android code where the runtime always assumes 
that the encoding is pcrel+indirect, but probably that's a hack also.

If I understood correctly, using DWARF structures for exception handling 
on ARM is not guided by any standard and it's just a sane choice. In 
that case there is no good reason to treat ARM differently in the way it 
uses these structures.

Regards,
Mathias