Re: Please try this testcase on Solaris.

["H . J . Lu" <hjl at lucon dot org>]

On Mon, May 20, 2002 at 07:14:42PM +0200, Martin v. Loewis wrote:
> "H . J . Lu" <hjl@lucon.org> writes:
> 
> > But the C++ testcase works fine on your Solaris? 
> 
> For me, this is only hearsay so far, I have not verified this
> myself. I believe Jason Merrill has, and came to the conclusion that
> even though ld.so appears to work the same, that RTTI matching still
> behaves different, without further investigating why.
> 
> I'll see whether I can get further details.
> 
> > Anyway, we have a strange situation here:
> 
> I agree.
> 
> 
> > 1. A has undefined symbol, foo, and B's behavior depends on the
> > definition of foo.
> > 2. B has DT_NEEDED for A and has a difinition of foo.
> > 3. C has DT_NEEDED for A and has a difinition of foo.
> > 
> > When both B and C are dlopened with RTLD_LOCAL, A may not work right
> > if the same A in memory is used for both B and C since foo in B may be
> > different from foo in A. I can come up with other testcases to show it.
> > I don't know if/how the dynamic linker should handle it.
> 
> I think it should: As David Abrahams suggests, the dynamic linker
> should resolve all references for foo to the definiton in B (thus
> ignoring the definition in C, just as it would with RTLD_GLOBAL).

It may not work right since C may want to use its own definition of foo.
I think we can do

1. A can't have undefined symbols which are defined in B or C.
2. Loading A with RTLD_LOCAL first should override definition in B/C
loaded with RTLD_LOCAL. That is Jason's #5. It doesn't seem to work
on Linux. I am trying to find a testcase in C.


H.J.