Duplicate .debug_lines (Was: [PATCH 5/5] Add --gdwarf-5 to ASM_SPEC)

[Mark Wielaard]

Hi,

On Sat, Aug 29, 2020 at 07:34:35AM -0700, H.J. Lu wrote:
> On Sat, Aug 29, 2020 at 5:24 AM Mark Wielaard <mark@klomp.org> wrote:
> > On Wed, Aug 26, 2020 at 04:38:21PM -0700, H.J. Lu wrote:
> > > On Wed, Aug 26, 2020 at 2:38 PM Mark Wielaard <mark@klomp.org> wrote:
> > > > Would it be possible to have something like the following in gas, so
> > > > that it doesn't try generating a .debug_line section if there already
> > > > is one, even when -gdwarf-N is given (unless the assembly also
> > > > contains .loc directives because that shows the user is really
> > > > confused)?
> > > >
> > > > diff --git a/gas/dwarf2dbg.c b/gas/dwarf2dbg.c
> > > > index e4ba56d82ba..c0c09f4e9d0 100644
> > > > --- a/gas/dwarf2dbg.c
> > > > +++ b/gas/dwarf2dbg.c
> > > > @@ -2626,7 +2626,7 @@ dwarf2_init (void)
> > > >
> > > >
> > > >  /* Finish the dwarf2 debug sections.  We emit .debug.line if there
> > > > -   were any .file/.loc directives, or --gdwarf2 was given, or if the
> > > > +   were any .file/.loc directives, or --gdwarf2 was given, and if the
> > > >     file has a non-empty .debug_info section and an empty .debug_line
> > > >     section.  If we emit .debug_line, and the .debug_info section is
> > > >     empty, we also emit .debug_info, .debug_aranges and .debug_abbrev.
> > > > @@ -2650,9 +2650,16 @@ dwarf2_finish (void)
> > > >    empty_debug_line = line_seg == NULL || !seg_not_empty_p (line_seg);
> > > >
> > > >    /* We can't construct a new debug_line section if we already have one.
> > > > -     Give an error.  */
> > > > +     Give an error if we have seen any .loc, otherwise trust the user
> > > > +     knows what they are doing and want to generate the .debug_line
> > > > +     (and all other debug sections) themselves.  */
> > > >    if (all_segs && !empty_debug_line)
> > > > -    as_fatal ("duplicate .debug_line sections");
> > > > +    {
> > > > +      if (dwarf2_loc_directive_seen)
> > > > +       as_fatal ("duplicate .debug_line sections");
> > > > +      else
> > > > +       return;
> > > > +    }
> > > >
> > > >    if ((!all_segs && emit_other_sections)
> > > >        || (!emit_other_sections && !empty_debug_line))
> > >
> > > I have run into this issue before.  "as -g" shouldn't silently
> > > generate incorrect debug info when input assembly codes already
> > > contain debug directives.  AS should either issue an error or
> > > ignore -g.
> >
> > Right, that is what this patch does for .debug_line.  gas already
> > doesn't generate .debug_info, .debug_aranges and .debug_abbrev if
> > .debug_info is non-empty, even if -g is given.
> >
> > > In either case, we need a testcase  to verify it.
> >
> > Right, and the documentation needs to be update.  But first we have to
> > know whether the gas maintainers think this is the right approach.
> 
> -g should be ignored in this case.

I am not sure what you mean by "in this case", or what precisely it
means to "ignore -g".

My proposal, and what my strawman patch implements, is that gas will
generate a .debug_line section when -g is given and the debug types is
DWARF (just as it does now). Unless there is a non-empty .debug_line
section already created by the input assembly and the input assembly
does not contain any .loc directive then gas will not try to generate
a .debug_line section itself but leaves the non-empty .debug_line as
is (currently gas will generate an error in this case). But if the
input assembly does contain both .loc directives and creates a
non-empty .debug line section gas will still generate an error (as it
does now, whether or not the input assembly contains any .loc
directives).

Does this sound sane?

Thanks,

Mark