ARM/THUMB long calls stubs

[Richard Earnshaw]

> Hi Andy,
> 
> > To anyone who can help(!),
> > 
> > I have hit an issue with long calls on arm gcc (version 3.0.4 with the
> > long-calls patch). I see someone else reported this issue on 30 April,
> > has there been a resolution?
> > 
> > The issue is as follows:
> > 
> > void FnHi() __attribute__((section("high"))) __attribute__((long_call));
> > void FnLo() __attribute__((section("low"))) __attribute__((long_call));
> > 
> > I can call FnHi() from anywhere within the "low" section, however I
> > cannot call FnLo() from the "high" section because the _call_via_rX
> > stub is located in "low", and hence the call to _call_via_rX gets
> > truncated by the linker.
> > 
> > Is there a way to get gcc to generate the stubs in all sections?
> 
> [Note - this is a GCC bug not a binutils bug, so I am CC'ing this
> reply to the GCC bugs list].
> 
> As things stand at the moment: "no".
> 
> Even if you could, it would not work, because the linker would just
> complain about duplicate definitions of the _call_via_rX symbols.  What
> you would need is for gcc to generate branches to functions called
> _call_via_rX_from_section_<foo> and then have it generate weak
> versions of these functions in every compiled source file that uses
> them.
> 
> You might be able to achieve a similar effect through judicious use of
> partial linking and special linker scrips.  ie you could perform a
> partial link of the "high" section code with libgcc.a to resolve all
> of the references to __call_via_rX there.  (Making sure that the
> script used places the object file containing the _call_via_rX symbols
> close to the "high" section).  You could then do a similar partial
> link for the "low" section.  The strip out any remaining occurrences of
> the _call_via_rX symbols and perform a final link.  It *might* work....

Hmm, I think this is a case where the linker needs to step in and "fix" 
the problem.  In this case it needs to manually insert a small trampoline 
so that the high call can work correctly.  I don't think inventing 
duplicate call_via_... functions is really the answer.  And it needs to do 
this for each section that can't reach the target section.

R.