Re: HEAD and gcc4 build fixes for MIPS

[Daniel Jacobowitz <drow at false dot org>]

On Thu, Feb 10, 2005 at 02:42:50PM +0000, Maciej W. Rozycki wrote:
> On Wed, 9 Feb 2005, Daniel Jacobowitz wrote:
> 
> > @@ -162,6 +164,24 @@ la_symbind64 (Elf64_Sym *sym, unsigned i
> >  # define La_regs La_ia64_regs
> >  # define La_retval La_ia64_retval
> >  # define int_retval lrv_r8
> > +#elif defined __mips__ && _MIPS_SIM == _ABIO32
> > +# define pltenter la_mips_o32_gnu_pltenter
> > +# define pltexit la_mips_o32_gnu_pltexit
> > +# define La_regs La_mips_32_regs
> > +# define La_retval La_mips_32_retval
> > +# define int_retval lrv_v0
> > +#elif defined __mips__ && _MIPS_SIM == _ABIN32
> > +# define pltenter la_mips_n32_gnu_pltenter
> > +# define pltexit la_mips_n32_gnu_pltexit
> > +# define La_regs La_mips_64_regs
> > +# define La_retval La_mips_64_retval
> > +# define int_retval lrv_v0
> > +#elif defined __mips__ && _MIPS_SIM == _ABIN64
> > +# define pltenter la_mips_n64_gnu_pltenter
> > +# define pltexit la_mips_n64_gnu_pltexit
> > +# define La_regs La_mips_64_regs
> > +# define La_retval La_mips_64_retval
> > +# define int_retval lrv_v0
> >  #else
> >  # error "architecture specific code needed"
> >  #endif
> 
>  AFAIK the agreed rule is to use _MIPS_SIM_ABI32, _MIPS_SIM_NABI32 and 
> _MIPS_SIM_ABI64 respectively.

All that, and you didn't even notice the _ABIN64 typo!  I was just
about to send a corrected patch.

>  And checking what's going on I've discovered, to my surprise, the 
> _ABIO32, _ABIN32 and _ABIN64 macros are still used.  Apparently an 
> incorrect patch was applied under the following ChangeLog entry:
> 
> 2004-11-05  Maciej W. Rozycki  <macro@mips.com>
> 
> See the thread starting at: 
> http://sources.redhat.com/ml/libc-alpha/2004-11/msg00034.html for details 
> and: http://sources.redhat.com/ml/libc-alpha/2004-11/msg00083.html for the 
> final patch.
> 
>  I've prepared a fix reverting the bad change and applying the intended 
> one that I will send separately once testing concludes.  It's a pity it 
> wasn't spotted before 2.3.4 was released.

Grr, and I just finished updating to the "new" style.

However, Maciej, I think we should stay with the _ABI macros.  Richard
Sandiford pointed something interesting out.  In the message you
referenced above you wrote:

> Unfortunately, after a closer look I've noticed these _ABI* macros get 
> actually defined only if they are going to be used for the value of 
> _MIPS_SIM, that is one at a time.  So after removing my addition of them 
> from <sgidefs.h> conditionals that refer to them suddenly stop working.  
> So let's just scrap my proposal as it does not make sense anymore.

But that's not true.  While only one will be defined, it will always
have a non-zero value.  So:
				-mabi=32	-mabi=64
#if _MIPS_SIM == _ABIO32	 1 == 1		 3 == 0
#if _MIPS_SIM != _ABIO32	 1 != 1		 3 != 0
#if _MIPS_SIM == _ABIN32	 1 == 0		 3 == 0
#if _MIPS_SIM != _ABIN32	 1 != 0		 3 != 0
#if _MIPS_SIM == _ABI64		 1 == 0		 3 == 3
#if _MIPS_SIM != _ABI64		 1 != 0		 3 != 3

The conditionals always come out right.

-- 
Daniel Jacobowitz
CodeSourcery, LLC