[Bug symtab/16581] GDB crash on inherit_abstract_dies infinite recursion

["brobecker at gnat dot com" <sourceware-bugzilla at sourceware dot org>]

https://sourceware.org/bugzilla/show_bug.cgi?id=16581

--- Comment #1 from Joel Brobecker <brobecker at gnat dot com> ---
>From https://www.sourceware.org/ml/gdb-patches/2014-01/msg00805.html:

For the record, we've actually seen the exact same behavior from
with GNAT a few years ago.  The problem occured when when we had
the following situation:

   procedure Outer (...) is
      [...]
      procedure Inner (...) is
         [...]
         -- Recurse by calling Outer..
         Outer (...);
      end Inner;
   begin
      [...]
      Inner (...);

In that case, when compiled at -O3, the compiler generated an Abstract
Instance Root (AIR) for procedure Outer, which owned/contained a DIE
defining procedure Inner (an out-of-line instance), which itself
contains a Concrete Instance Entry (CIIE) corresponding to the inlined
version of Outer's AIR. The CIIE has a reference to its AIR via
the DW_AT_abstract_origin attribute, hence the cycle.

Not being all quite sure how to make sense of the cycle in terms of
inheritance, we initially tried to fix in the compiler.  Although
the patch was initially approved in GCC, it was noted that the output
appeared to be conformant with the DWARF specifications (version 3,i
at the time, now version 4), particularly in light of the examples
in section D.7 of the specifications (3rd example).

Eventually, it was found that our GCC patch was causing some issues,
and thus was reverted. So far, we've kept the patch in AdaCore's
GCC tree, with a note to look into a GDB fix at some point.

We are indeed very close to the example from the DWARF specs cited
above, but it does not deal with recursion as we do here, so I think
that the DWARF specifications do have a hole when it comes to that.
Logically speaking, its seems that the sensible thing to do is to
inherit the whole Abstract Instance Tree (AIT) but excluding oneself.

-- 
You are receiving this mail because:
You are on the CC list for the bug.