Re: MIPS patch to avoid lazy binding in la macros

[Richard Sandiford <rsandifo at redhat dot com>]

"Maciej W. Rozycki" <macro at ds2 dot pg dot gda dot pl> writes:
>  Well, it's a "bug by design" or IOW a feature.  Gccs used not to emit
> code sequences that would end up with an incorrect relocation.  If that
> changed meanwhile, blame for gcc for not accounting for the behaviour of
> gas.  The code is there since around 2.10, IIRC.

Indeed, historical precedent seems to be the main reason for not
changing things now.

> > >  Essentially you kill lazy binding for code as currently emitted by gcc
> > > entirely this way.
> > 
> > True.  I'm getting used to mips-rewrite, so I forgot what older gccs
> > (including trunk ;) do.  But more below.
> 
>  So what does exactly the branch do to select between GOT and CALL
> relocations?

Just to be clear, I'm talking about the gcc mips-3_4-rewrite-branch here.
It has two modes: one which uses assembler macros (as current gcc does)
and one that uses explicit %reloc() operators.  When using assembler
macros, it gets lazy binding by using "jal foo".

> > Yes, the problem I came across was with undefined weak symbols.
> 
>  What version of gcc?

mips-3_4-rewrite-branch, in the "assembler macro" mode.

> > As I understand it, the idea behind the old behaviour was that
> > gcc generated:
> > 
> >         la $25,foo
> >         jal $31,$25
> > 
> > and we wanted GAS to treat "la $25,foo" as part of a call sequence.
> 
>  Any "la $25,foo" actually as gcc used not to use the register for
> anything else when emitting SVR4 PIC code.

Are you sure?  What stopped gcc from using $25 for anything else?
Was it declared to be a fixed register?

As far as I'm aware, recent gccs have treated $25 as an ordinary
call-clobbered GPR.

> > Well, the code is still raw, but mips-rewrite does what you suggest,
> > at least for n32 and n64 PIC.  There's a patch (pending review ;) to
> > extend it to o32 PIC as well.
> 
>  Too bad you haven't sent it earlier here -- maybe someone would be
> interested and would fix the problem for real.  For now here is what I
> cooked up -- since it's been on my to-do list for at least two years, it
> certainly deserved being finally written. ;-)  For 2.13.2.1 for now.  I'll
> update that if that proves usable.

To be clear, I'm talking about a gcc patch here, not a gas patch.
CVS gas already has support for explicit relocations, see:

   http://sources.redhat.com/ml/binutils/2003-01/msg00449.html
   http://sources.redhat.com/ml/binutils/2003-02/msg00044.html

(...and various follow-ons ;)

Richard