Re: i386-rtems(i386-coff) failure in egcs-19980628

[Joel Sherrill <joel at OARcorp dot com>]

On Tue, 7 Jul 1998, Robert Lipe wrote:

> > version: egcs-19980628
> > host:    RedHat Linux 5.0 (i686-pc-linux-gnu)
> > target:  i386-rtems (based on i386-coff)
> >          (NOTE: i386-rtemself has built successfully)
> 
> Ha!  It's not that I enjoy seeing you specifically suffer, Joel.  It's
> just that I welcome somebody else into x86 assembler "what do you mean
> everything isn't linux?" hell. :-)

And everything i386 based is not a PC. :)

> It took me a couple of builds to duplicate this and I think I've spotted
> the offending code.  However, I have little idea what the _best_
> solution to this is.

I have a proposal below for i386-coff and i386-rtems.

> > The failure occured while compiling __floatdidf from libgcc2.c.  The
> [ munch ] 
> > _floatdidf
> > /tmp/ccekaQTS: Assembler messages:
> > /tmp/ccekaQTS:112: Error: Alignment not a power of 2
> 
> I was able to duplicate this only with the GNU assembler, but I see
> it fail on OSR5, too. i386-coff suffers from the same failure and
> has slightly fewer config files in the way.  Fortunately, OSR5 is
> immune from this, but I couldn't at all point to sco5.h as a model for
> anything.

:(

> For i386-coff, ultimately i386/gas.h gets included.  This combination of
> targets won't define ASM_OUTPUT_ALIGN, so we get the default one - the
> one that says the argument to .align is the power of 2 we should pad to,
> not the number of bytes that we should pad to.
> 
> This patch is absolutely wrong and should not be committed.  It will,
> I believe, get you going again, Joel.
> 
> If you'd like to fix only (and specifically) the rtems target, you might
> think about cabbaging the definition and comments for ASM_OUTPUT_ALIGN
> and stuffing them into i386-rtems.h.  I think that'll get you going.
> The comment about recent GAS versions sounds like somehing you should
> note.
> 
> This is too sticky for me.  I don't know the preferred way to make
> a common file like gas.h know which of the conventions are used.  I
> suppose it could be autoconfiscated...
> 
> 
> 
> Index: gas.h
> ===================================================================
> RCS file: /egcs/carton/cvsfiles/egcs/gcc/config/i386/gas.h,v
> retrieving revision 1.3
> diff -u -p -r1.3 gas.h
> --- gas.h	1998/05/21 19:12:18	1.3
> +++ gas.h	1998/07/07 04:38:49
> @@ -80,7 +80,7 @@ Boston, MA 02111-1307, USA.  */
>     doubt or guess work, and since this file is used for both a.out and other
>     file formats, we use one of them.  */
>  
> -#if 0 /* ??? However, not every port uses binutils 2.6 yet.  */
> +#if 1 /* ??? However, not every port uses binutils 2.6 yet.  */
>  #undef ASM_OUTPUT_ALIGN
>  #define ASM_OUTPUT_ALIGN(FILE,LOG) \
>    if ((LOG)!=0) fprintf ((FILE), "\t.balign %d\n", 1<<(LOG))

How about this for a proposal?

+ binutils 2.6 is older than anything I have on-line. :)

+ According to gas/Changelog binutils 2.6 was released around Nov 16,
  1995.

+ Anyone following the crossgcc instructions for building an i386-coff
  toolset should be using a newer binutils since the instructions tell
  you to do that.

+ i386-rtems was not even a patch for anything prior to binutils 2.8.

I propose that i386-coff.h be changed to include this version of
ASM_OUTPUT_ALIGN from go32.h: 

/* This is how to output an assembler line 
   that says to advance the location counter 
   to a multiple of 2**LOG bytes.  */

#undef ASM_OUTPUT_ALIGN
#define ASM_OUTPUT_ALIGN(FILE,LOG) \
  if ((LOG) != 0) fprintf ((FILE), "\t.p2align %d\n", LOG)

I don't know what to do about gas.h.  It is heavily used so I would be
scared to change it.  But I think i386-coff can use this and I KNOW this
is OK for i386-rtems.  Adding it to i386-coff.h fixes both targets.

It would not take much to convince me that gas.h should not support a
binutils that old.  If the target is using that old a version, then the
"good" definition should be overridden in a target specific file.  But
that is for someone else to decide on.  I am comfortable with the idea of
fixing i386-coff.h for embedded users.

Comments welcomed.

--joel