Re: objcopy test failure on mips-linux

[Mark Mitchell <mark at codesourcery dot com>]

>>>>> "Andreas" == Andreas Jaeger <aj@suse.de> writes:

    Andreas> The problem is that with objcopy the text.rel section
    Andreas> from the original file ends in a text.rela section :-(.

Yes, that's not so good.  Logically, though, that doesn't seem a total
disaster, as long the resulting .o still means the same thing as the
first one.

There's a tricky problem here -- because BFD operates abstractly on
some internal notion of relocations, sections, etc., the operating or
reading in a BFD and then writing it back out is not always the
identity transformation, anyhow.

But, I see your point.

    Mark> The top-level problem, though, is that on IRIX6 we need both
    Mark> REL and RELA relocation sections simultaneously -- as you
    Mark> say both N32 and 64-bit objects use both at once.  I'm not
    Mark> surprised it creates empty sections; as I recall there
    Mark> didn't seem any way to know up front whether or not you
    Mark> needed the sections.

    Andreas> Ok, I understand.  Can't we add a check at write out time
    Andreas> to only write a e.g. REL section if the RELA section is
    Andreas> empty?  Currently we write an empty RELA section.

That sounds reasonable, but I think the check is that you should write
the REL section only if the REL section is non-empty, and the RELA
section only if the RELA section is non-empty.  Otherwise, with what
you wrote, you might write not write the REL section if both the REL
and RELA sections are non-empty.

--
Mark Mitchell                   mark@codesourcery.com
CodeSourcery, LLC               http://www.codesourcery.com