I suspect it is a problem for you guys as well (unless you do your own
memory allocation and force things to be close together). In particular, I
wouldn't be surprised if you ran into problems if you put a safe trunc on
this line (that's a cursory guess, I haven't actually tested that):

https://github.com/dropbox/pyston/blob/60bab2eabd25f745c15eb967a7fd7d778581b9bc/src/codegen/unwinding.cpp#L92

Do you guys use no-fp-elim in LLVM? libunwind is pretty good about figuring
stuff out in those cases, so you may just not have noticed.

I will update the patch to address the review comments.

On Tue, Oct 6, 2015 at 6:27 PM, Kevin Modzelewski <[email protected]> wrote:

>
>
> On Tue, Oct 6, 2015 at 10:20 AM, Arun Sharma <[email protected]> wrote:
>
>> On Sun, Oct 4, 2015 at 10:30 PM, Keno Fischer
>> <[email protected]> wrote:
>> > Bump, does this patch look reasonable?
>>
>> Yes - looks reasonable to me. Wonder why it wasn't a problem for other
>> LLVM JIT users (eg: pyston?).
>>
>
> Now I'm curious too :)  Though I don't think our eh_frame and text
> sections are that far apart.  tbh I don't have a great understanding of the
> interaction between MCJIT and libunwind; for instance it also seems odd
> that the format that works is REMOTE_TABLE and not (non-remote) TABLE.
>
>
>>
>> > unw_dyn_remote_table_t, but
>>
>> I think you meant to say "Like REMOTE_TABLE, but ..".
>>
>> Perhaps call it IP_OFFSET instead of TABLE2?
>>
>> +  if (di->format == UNW_INFO_FORMAT_REMOTE_TABLE ||
>> +      di->format == UNW_INFO_FORMAT_REMOTE_TABLE2)
>>
>> Could you wrap this in is_remote_table()?
>>
>> > +    ip_base = segbase;
>>
>> Rename it something neutral that works for both the segbase and ip_base
>> cases?
>>
>>  -Arun
>>
>> _______________________________________________
>> Libunwind-devel mailing list
>> [email protected]
>> https://lists.nongnu.org/mailman/listinfo/libunwind-devel
>>
>
>
_______________________________________________
Libunwind-devel mailing list
[email protected]
https://lists.nongnu.org/mailman/listinfo/libunwind-devel

Reply via email to