Re: Support archives with 64-bit symbol table

[Mark Wielaard <mjw at redhat dot com>]

On Fri, 2012-08-03 at 01:50 +0200, Petr Machata wrote:
> Mark Wielaard <mjw@redhat.com> writes:
> 
> > On Thu, Aug 02, 2012 at 01:43:10AM +0200, Petr Machata wrote:
> >> these two patches implement support for 64-bit symbol table in .a
> >> archives.  As of recently, binutils' ar produces such archives on s390x
> >> (regardless of actual archive size).  This appears to be the same as
> >> "/", except all fields are 64-bit.
> >
> > Just a quick partial review, with some pedantic comments. Have to read
> > up on ar archives. But it looks correct to me. Are the magic entries
> > ("/SYM64/") described somewhere in a specification document or is this
> > GNU specific?
> 
> I went with the assumption that it's just like "/", except 64-bit where
> "/" is 32-bit.

Yes, I think the logic is sound.

>   I found something now, page 97:
> 
>   http://techpubs.sgi.com/library/manuals/4000/007-4658-001/pdf/007-4658-001.pdf
> 
> This is Irix documentation, 64-bit Irix is where this format comes from.

Thanks. And that confirms your assumption.

> >> -   Copyright (C) 1998, 1999, 2000, 2002, 2005 Red Hat, Inc.
> >> +   Copyright (C) 1998-2012 Red Hat, Inc.
> >
> > This should be 1998-2000, 2002, 2005, 2012. Only the first is a full
> > range. Similarly in some other places. (Yes, I did ask a lawyer.)
> 
> I'll correct this.  I recalled something with coalescing individual
> years into ranges, but wasn't actually sure what the rule was.

Assuming that the copyright notice is intended to be per-file in an
informational sense, it is allowed to coalesce individual years into
ranges, if all individual years are in that range. I think this is also
the most intuitive way. But we could adopt some other practice if that
is easier to implement and more intuitive. As long as we are clear and
consistent about it throughout the project.

> >> -      /* Now test whether this is the index.  It is denoted by the
> >> -	 name being "/ ".
> >> +      bool index64_p;
> >> +      /* Now test whether this is the index.  If the name is "/", this
> >> +	 is 32-bit index, if it's "/SYM64/", it's 64-bit index.
> >> +
> >>  	 XXX This is not entirely true.  There are some more forms.
> >>  	 Which of them shall we handle?  */
> >
> > Is this still true? Which other forms are there?
> 
> I don't know.  In Linkers & Loaders, Levine mentions several different
> formats for "/" and one for "__.SYMDEF".  The latter was reportedly used
> in original BSD and in a.out archives.  I'm not sure it's relevant
> anymore.

__.SYMDEF seems to not be used for ELF, it always uses "/". Looks like
only "/" (index), "/SYM64/" (64-bit index - Linkers and Loaders already
mentions it) and "//" (long names) are used in practice.

> >> +testfiles archive64.a
> >
> > Please document here how archive64.a was created.
> 
> Humm, I took a bunch of elfutils objects, but now that I think of it, we
> would have to ship the exact same source that led to their creation.
> Maybe instead I'll invent a couple trivial one-liners that can easily be
> pasted to run-test-archive64.sh.  Would that work?

Yes please. I really like being able to easily reconstruct our
testcases, so they can more easily be extended for new situations.

Thanks,

Mark