Re: .debug_frame section - incorrect relocations?

[Alexey Neyman <stilor at att dot net>]

On Friday, December 06, 2013 10:40:00 PM Alexey Neyman wrote:
> On Friday, December 06, 2013 07:44:32 AM Ian Lance Taylor wrote:
> > On Thu, Dec 5, 2013 at 11:38 PM, Alexey Neyman <stilor@att.net> wrote:
> > > So, the question is, is it possible to express the behavior described by
> > > DWARF spec ("insert an address that is offset of the specified symbol
> > > from the start of the specified output section") in x86-64 and i386
> > > relocations? If it is, shouldn't gas generate such relocations?
> > 
> > There is no straightforward way to express that in an object file.
> > 
> > > If not - are there any solutions to have .debug_frame at non-zero
> > > address,
> > > yet with DWARF-conformat CIE links?
> > 
> > I'm not aware of any.  DWARF sections are normally at address zero.
> 
> Thanks. Actually, this gave me an idea how to handle this using the AT(...)
> attribute in the linker script:
> 
>    __debug_frame_start = ALIGN(8);
>    .debug_frame 0 : AT(__debug_frame_start) { *(.debug_frame) }
>    . = __debug_frame_start + SIZEOF(.debug_frame);
>    __debug_frame_end = .;

Somehow this only works for a single sections, and it seems to be triggering 
an ld bug. Here is a linker script fragment:
[[[
    __debug_frame_start = ALIGN(8);
    .debug_frame 0 : AT(__debug_frame_start)
    {
        *(.debug_frame)
    }
    . = __debug_frame_start + SIZEOF(.debug_frame);
    __debug_frame_end = .;

    __debug_info_start = ALIGN(8);
    .debug_info 0 : AT(__debug_info_start)
    {
        *(.debug_info)
    }
    . = __debug_info_start + SIZEOF(.debug_info);
    __debug_info_end = .;
]]]

The linker complains:
ld: section .debug_info [00000000004440b8 -> 000000000050c361] overlaps 
section .debug_frame [00000000004440b8 -> 000000000044a907]
ld: lsk_d.elf: section .debug_info lma 0x4440b8 overlaps previous sections

In the map file, I have:
             0x00000000004440b8  __debug_frame_start = ALIGN (0x8)
.debug_frame 0x0000000000000000  0x6850 load address 0x00000000004440b8
             0x000000000044a908  . = (__debug_frame_start +
                                      SIZEOF(.debug_frame)
             0x000000000044a908  __debug_frame_end = .
             0x000000000044a908  __debug_info_start = ALIGN (0x8)
.debug_info  0x0000000000000000 0xc82aa load address 0x000000000044a908
             0x0000000000512bb2  . = (__debug_info_start +
                                      SIZEOF(.debug_info))
             0x0000000000512bb2  __debug_info_end = .

So the map shows that .debug_info is loaded at 0x44a908, but the error message 
says that LMA for .debug_info is 0x4440b8 - which is actually the LMA for 
.debug_frame, not .debug_info. As far as I can tell, the linker script does 
create separate LMAs for these two sections. Am I missing anything?

I've tried with GIT master branch and it is reporting this error.

Regards,
Alexey.

> 
> > If you are really only concerned about the .debug_frame section, you
> > should be looking instead at the .eh_frame section, which is loaded
> > into memory, and uses a slightly different format.  I don't know if it
> > will provide everything you need, though.
> > 
> > If you do need full debug info, then you'll have to tweak your DWARF
> > reader to adjust by the base of the debug sections as appropriate.
> 
> No, I've just started with .debug_frame, but I'll need other .debug_*
> sections as well. They have the same issue, as far as I understand, so the
> linker script above seems to be more general solution (than tweak all
> specifications of "offset from the start of the section").
> 
> Regards,
> Alexey.
> 
> > Ian