Re: [Libunwind-devel] FYI: new type of unwind info inaccuracy (x86-64, libm, rhel 6.x)

[Lassi Tuura]

Hi,

>> While testing the new libunwind version we came across a new type of unwind 
>> info inaccuracy on x86-64 RHEL6. See http://gcc.gnu.org/PR48597 for the 
>> details.
> 
> Looks like there is a gcc patch that fixes this now. But it'll be a
> while before distros package it. In the meanwhile libunwind will need
> to deal with these types of issues.

There now appears to be a patch for all of GCC 4.[456].x, so things look 
positive.

My feeling is that libunwind currently deals with these adequately, at least on 
x86-64. The omissions are not detectable in that unwind info is present and 
correct - just missing some info.

x86-64 unw_step() loop will handle the problems by returning an error code - we 
default to full paranoia memory checks, but it can be turned off if someone 
builds on a platform not subject to the problems.

I don't feel we'd need additional options, for example for fast trace. Clients 
can already choose full paranoia checks by using unw_step() loop, and adding 
them to fast trace will turn it into slow trace :-) I think the fast trace is 
reasonably useful as it is, as unfortunate as the crashes are.

Regards,
Lassi