Re: PATCH for i386: --no-copy-relocs

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

On Mon, Jun 18, 2001 at 06:14:16PM -0700, Ian Lance Taylor wrote:
> "Jose Luu" <jluu@mainsoft.com> writes:
> 
> > as outlined in message
> > http://sources.redhat.com/ml/binutils/2000-12/msg00183.html and
> > references, here is my patch to solve the aforementionned problem.
> > 
> > The patch has been extensively tested with our software (several
> > million lines) and has been shipped to customers since march, so I
> > am now trusting it. We compile it in such a way that the
> > --no-copy-relocs option is always on unless the -shared option is
> > active.
> > 
> > Whether we want this mode as default or not is a matter of debate:
> > This copying of relocations from shared libs into the main
> > executable seems to come from the i386 ABI, no other ABI seems to
> > need it, I don't know if it is a requirement that executables use
> > it. The problem I am solving is a byproduct of its usage.
> 
> If I read your patch correctly, the dynamic linker will be required to
> modify the text section, which means that it can not be shared when
> multiple processes run the executable.  I do not think that should be
> default.  Also, I think your patch needs to create a DT_TEXTREL if it
> does not already do so.

Definitely. Otherwise if it is true, the dynamic linker may be very
upset.

> 
> The COPY reloc is there in order to keep the text section read-only.
> The i386 is not the only ABI which uses it.  Several others do as
> well.

Is that for symbols in the .rodata (.text?) section? Can we tell gcc
to put such symbols in the .data section?  The SVR4 mips ABI has the
similar issue.


H.J.