Re: [PATCH v3] Bug 28191 - Interpret DWARF 5 addrx locations

[Giuliano Procida <gprocida@google.com>]

Hi.

On Wed, 24 Nov 2021 at 11:11, Giuliano Procida <gprocida@google.com> wrote:
>
> Hi Dodji.
>
> On Tue, 23 Nov 2021 at 17:15, Dodji Seketeli <dodji@seketeli.org> wrote:
> >
> > [I am adding Mark to the list recipients as I am going to ask him a
> > question below]
> >
>
> Indeed, I forgot to CC him on the patch. Hi Mark.
>
> >
> > Hello,
> >
> > Giuliano Procida <gprocida@google.com> a écrit:
> >
> > > This change uses libdw facilities to interpret location expressions
> > > instead of using libabigail's own mini-interpreter. With the fix for
> > > elfutils https://sourceware.org/bugzilla/show_bug.cgi?id=28220 in
> > > elfutils-0.186, abidw will correctly interpret Clang DWARF 5 symbol
> > > addresses. Without that fix many declarations will not be linked to
> > > their corresponding symbols due to the incorrect interpretation of
> > > location attribute data.
> > >
> > >       * src/abg-dwarf-reader.cc (die_location_address): Use
> > >       dwarf_attr_integrate, dwarf_getlocation and
> > >       dwarf_getlocation_attr to decode addreses, instead of
> > >       die_location_expr and eval_last_constant_dwarf_sub_expr.
> > >       * tests/data/test-read-dwarf/PR25042-libgdbm-clang-dwarf5.so.6.0.0.abi:
> > >       Refresh test reference output; two more symbols have types.
> >
> > Thanks for looking into that.
> >
> > It seems to me that we can also do away with the use of eval_last_constant_dwarf_sub_expr from
> > die_member_offset.  Would that work on your testing binaries?
> >
>
> I think I originally wrote the code just before a trip away and didn't
> have the time to look into the other callers then.
>
> >
> > From what I am seeing, dwarf_getlocation_attr and dwarf_getlocation are
> > present in elfutils 0.176 which is the oldest version that we need to
> > support, so we should be able to ditch eval_last_constant_dwarf_sub_expr
> > and its dependencies altogether.
> >
> > Or what am I missing?
> >
>
> I think it would be good if we could rely on elfutils for this functionality.
>
> I've had a go at die_member_offset and I think I have it working (at
> least with the test suite). I'll clean up what I have and post it.
> There is considerable overlap with the address function.
>
> I haven't been able to get die_virtual_function_index to work as
> dwarf_getlocation_attr fails.  There may be a simpler recipe to follow
> here.
>

My mistake, the problem is with base-class offsets, not vtable
indexes, so I don't have die_member_offset working after all. It's
possible an extra specialisation of the function for the right DWARF
"form" will fix this.

Giuliano.

> If/when the 3 existing callers are working, there may be some code in
> common to factor out as a new die_location_expr function.
>
> Regards,
> Giuliano.
>
> >
> > [...]
> >
> > Cheers,
> >
> > --
> >                 Dodji