Re: weak and strong aliases for global data fail

["Peter S. Mazinger" <ps.m@gmx.net>]

On Tue, 23 May 2006, Alan Modra wrote:

> On Mon, May 22, 2006 at 10:38:56PM +0200, Peter S. Mazinger wrote:
> > On Wed, 17 May 2006, Peter S. Mazinger wrote:
> > 
> > Could anyone tell me if this is a binutils bug (tested 2.16.1/2.16.92) or 
> > I should send it to the gcc list (tested with 3.4.6). It seems to me that 
> > it is rather an ld bug.
> > 
> > Thanks, Peter
> > 
> > > Hello!
> > > 
> > > Attached tests show a case, when a final binary is OK if 
> > > compiled w/ -fPIC (or -fPIE, it does not have to be ET_DYN, but can be), 
> > > else it fails if using aliases (both weak and strong) for global data.
> > > 
> > > Sorry for the "big" test cases, I couldn't make them smaller.
> 
> It's a user bug.  When the linker needs to generate a copy reloc, eg. as
> it does for x86 non-PIC code, for a variable like "optind_test" in your
> main app, then space is allocated in your main app and initialised from
> the shared object.  The main app uses this location for references to
> "optind_test", as do references from within the shared lib.  However,
> your shared lib is not referencing "optind_test" but an alias stored at
> a different location (in the shared lib).
> 
> With PIC code the linker can generally dispense with the copy reloc, so
> both the main app and the shared lib use the same location.

The app is a simplified version of getopt() used in glibc, but it happens 
with getopt.c from glibc as well. I have only renamed the internally 
used opt* to __opt* and provided a weak resp. strong alias to the 
externally visible version (also renamed to opt*_test so that it does not 
conflict with the libc provided getopt interface).

Peter

-- 
Peter S. Mazinger <ps dot m at gmx dot net>           ID: 0xA5F059F2
Key fingerprint = 92A4 31E1 56BC 3D5A 2D08  BB6E C389 975E A5F0 59F2