Re: Using binutils-2.12.1 on sparc64-sun-solaris2.8 to build gcc-3.1 results in relocation errors

[Teemu Torma <tot at trema dot com>]

David S. Miller wrote:
>    From: Alan Modra <amodra@bigpond.net.au>
>    Date: Tue, 16 Jul 2002 12:10:12 +0930
> 
>    On Mon, Jul 15, 2002 at 12:45:33PM -0500, Dana, Eric wrote:
>    > When building some of our 64-bit libraries in C++, we are seeing
>    > R_SPARC_DISP32 errors:
> 
>    Please try current binutils CVS.  If the problem isn't fixed there,
>    now would be the time to shout about it.  We're about to release 2.13.

I have exactly the same problem with C++ code and gcc 3.1 and newer, and 
binutils 2.12.1 and newer.  

> Actually this looks like some kind of code-model error.
> There should not be any R_SPARC_DISP32 relocations generated
> as no "call" instruction based function invocations should
> be emitted with Solaris's default code model.
> 
> "Call" is the only way R_SPARC_DISP32 relocations can be
> emitted.

As far as I can see, only "call" instructions are ever emitted by gcc. 
Nothing changes if I use -m32, -m64, -mcmodel=<any value>.

Am I missing something, or am I just ignorant?

> I really doubt this is a binutils bug.  It is either some user
> configuration error, or the libtool machinery is causing the
> wrong code model to be used when generating the c++ libraries
> and/or binaries.
> 
> If these folks are specifying the code-model explicitly on the
> command line, and linking with libstdc++ this could cause errors
> like the above as well.

In my projects no libtool used, just plain makefiles.  No code models ever 
specified, just -m64.  And everything with plain C works, the only problem 
is with non-trivial C++, where shared libraries are built and used.  I 
don't really know what is the point when the problem appears, every time I 
try to produce a test case, it works.

Teemu