Re: [m68k] Fix and improve ColdFire function prologue/epilogue generation(respin)

[Bernardo Innocenti <bernie at develer dot com>]

Richard Henderson wrote:
> On Wed, Sep 03, 2003 at 11:32:58PM +0200, Bernardo Innocenti wrote:
>
> > +  if (from == ARG_POINTER_REGNUM && to == HARD_FRAME_POINTER_REGNUM)
> > +    return 0;
>
>
> Intuitively, this is wrong.  I can see how it worked before, but
> not after your change.

Actually, it works fine with and without -fomit-frame-pointer.
The reason is probably that GCC never invokes that function to
eliminate the ARG_POINTER_REGNUM from HARD_FRAME_POINTER_REGNUM.

When I first tested that code, I added some big constants
(1000, 2000, 3000 and 4000) and looked at the asm output produced
with several options.

I think the correct value should be:

 if (from == ARG_POINTER_REGNUM && to == HARD_FRAME_POINTER_REGNUM)
   return (frame_pointer_needed ? -UNITS_PER_WORD * 2 : -UNITS_PER_WORD);

The first case skips both the return address and the saved fp,
while the second skips only the return address.

>  To wit: 
>
> m68k.h:#define FIRST_PARM_OFFSET(FNDECL) 8
>
> That says that the offset from ARG_POINTER_REGNUM to the first
> actual argument is 8.  Which, when you have a frame pointer, is
> exactly right for m68k -- four bytes for saved frame pointer,
> and four bytes for saved return address.

What does GCC use that for? Its value gets assigned to the static
variable `in_arg_offset' at function entry time and seems to be
only used in instantiate_new_reg().


> With -fomit-frame-pointer, however, things get screwy.  The fact
> that we have a hard-coded 8 above is arguably a bug.

...a bug I'm afraid to fix until I've grasped what the
consequences are :-)


> It appears as if things were worked around by adding magic
> constants,
>
> > -		  asm_fprintf (stream, "\tsubq.w %I%wd,%Rsp\n", fsize + 4);
>
> like so.
>
> You've not preserved this bit of, ahem, style, and so as far as
> I can see, you've broken parameter passing without a frame pointer.

No, it works fine indeed. Tested with a full build of Linux kernel and
userland!

There must be some other way of computing offsets to the arguments.


> The correct way to do this is to make the argument pointer point
> to the arguments.  Then you do
>
>   if (from == ARG_POINTER_REGNUM && to == HARD_FRAME_POINTER_REGNUM)
>     return 8;

Ok, that's what I'm doing in the code above, taking care of
frame_pointer_needed too.


>   if (from == ARG_POINTER_REGNUM && to == STACK_FRAME_POINTER_REGNUM)
>     return stack_size + 4;

What's STACK_FRAME_POINTER_REGNUM? :-)

If you meant STACK_POINTER_REGNUM, that combo is already handled
correctly here:

if (from == ARG_POINTER_REGNUM && to == STACK_POINTER_REGNUM)
   return frame.offset + frame.size +
     (frame_pointer_needed ? -UNITS_PER_WORD * 2 : -UNITS_PER_WORD);

If you meant FRAME_POINTER_REGNUM, that pair is not included in
ELIMINABLE_REGS, not even in the i386 back-end.

--
 // Bernardo Innocenti - Develer S.r.l., R&D dept.
\X/  http://www.develer.com/

Please don't send Word attachments - http://www.gnu.org/philosophy/no-word-attachments.html