Re: RFC: Assign regions to orphan sections

[Daniel Jacobowitz <dan at codesourcery dot com>]

On Wed, Apr 28, 2010 at 03:29:59PM +0000, Joseph S. Myers wrote:
> On Fri, 26 Mar 2010, Daniel Jacobowitz wrote:
> 
> > On Thu, Mar 25, 2010 at 05:55:16PM -0700, Ian Lance Taylor wrote:
> > > Daniel Jacobowitz <dan@codesourcery.com> writes:
> > > 
> > > > I don't think the file header should be loaded in your case, either.
> > > > Anyone know what determines this?  It's not triggered by use of
> > > > SIZEOF_HEADERS, I don't think.
> > > 
> > > It's done by default if it can be done.  Look at phdr_in_segment in
> > > _bfd_elf_map_sections_to_segments in bfd/elf.c.
> > 
> > This bit:
> > 
> >       /* Deal with -Ttext or something similar such that the first section
> >          is not adjacent to the program headers.  This is an
> >          approximation, since at this point we don't know exactly how many
> >          program headers we will need.  */
> >       if (count > 0)
> >         {
> >           bfd_size_type phdr_size = elf_tdata (abfd)->program_header_size;
> > 
> >           if (phdr_size == (bfd_size_type) -1)
> >             phdr_size = get_program_header_size (abfd, info);
> >           if ((abfd->flags & D_PAGED) == 0
> >               || sections[0]->lma < phdr_size
> >               || sections[0]->lma % maxpagesize < phdr_size % maxpagesize)
> >             phdr_in_segment = FALSE;
> >         }
> > 
> > It seems to me that if the maximum page size is one byte, then D_PAGED
> > should be false.  There aren't any pages.  But that's really only the
> > case for *-elf.  Joseph, would tic6x-uclinux also have MAXPAGESIZE==1?
> > 
> > Anyway, a safe way to fix the test would be to add -z
> > max-page-size=256.  I think that will do it.
> 
> Adding -z max-page-size=256 isn't itself a solution - none of the -z 
> options are supported by elf32.em unless GENERATE_SHLIB_SCRIPT is set to 
> "yes".  Setting it for C6X (together with adding the -z option to that 
> test) does fix that failure - while introducing some other failures (the 
> linker support for shared libraries on this target isn't written, so it's 
> unsurprising for the linker to segfault if this option is blindly enabled) 
> - but isn't correct either for C6X at present or for the other targets 
> without shared library support that might be affected by this issue.  
> Should some subset of the -z options (or indeed all of them) not be 
> conditional on shared library support?

No, I don't think this should change.

Does ld -N do the trick?  There may be some other oddities with
objcopy, because BFD autodetects D_PAGED on input files with program
headers based on the maxpagesize / minpagesize.

I do think there's a real problem here, namely that this will copy ELF
headers onto the target system for no good reason.  We could make up a
page size, I suppose.  It shouldn't do any harm; ARM programs with the
standard 'page size' are run without an MMU all the time.

-- 
Daniel Jacobowitz
CodeSourcery