[RFA] PPC ABI compliance fix

[RFA] PPC ABI compliance fix

[Elena Zannoni]

Back in November, gcc changed the way structures <= 8 bytes are
returned on the PPC. The change was made to be compliant with the
SVR4 ABI.

http://gcc.gnu.org/ml/gcc-patches/2001-11/msg01468.html

The abi specifies that such structures are passed in r3 and r4.
Bigger structures are passed in memory.

Gcc was passing every structure in memory.  The change was made for
embedded targets, but not for natives (PowerPC Linux, NetBSD, and
FreeBSD).

Of course this change breaks binary compatibility with older gcc's.
Not sure what to do about that, if anything, the case it covers is a
corner case, anyway.


Elena


2002-03-21  Elena Zannoni  <ezannoni@redhat.com>

        * ppc-linux-tdep.c (ppc_sysv_abi_use_struct_convention): New
        function.  
	* ppc-tdep.h (ppc_sysv_abi_use_struct_convention): Export.
        * rs6000-tdep.c (rs6000_gdbarch_init): Use different
        structure returning convention for SYSV ABI case, but not
        for GNU/Linux, FreeBSD, or NetBSD.

Index: rs6000-tdep.c
===================================================================
RCS file: /cvs/uberbaum/gdb/rs6000-tdep.c,v
retrieving revision 1.38
diff -u -r1.38 rs6000-tdep.c
--- rs6000-tdep.c	2002/02/14 15:13:53	1.38
+++ rs6000-tdep.c	2002/03/22 18:46:53
@@ -2563,8 +2563,6 @@
   set_gdbarch_store_struct_return (gdbarch, rs6000_store_struct_return);
   set_gdbarch_store_return_value (gdbarch, rs6000_store_return_value);
   set_gdbarch_extract_struct_value_address (gdbarch, rs6000_extract_struct_value_address);
-  set_gdbarch_use_struct_convention (gdbarch, generic_use_struct_convention);
-
   set_gdbarch_pop_frame (gdbarch, rs6000_pop_frame);
 
   set_gdbarch_skip_prologue (gdbarch, rs6000_skip_prologue);
@@ -2576,6 +2574,27 @@
   /* Not sure on this. FIXMEmgo */
   set_gdbarch_frame_args_skip (gdbarch, 8);
 
+  /* Until November 2001, gcc was not complying to the SYSV ABI for
+     returning structures less than or equal to 8 bytes in size. It was
+     returning everything in memory. When this was corrected, it wasn't
+     fixed for native platforms. */
+  if (sysv_abi)
+    {
+      if (osabi == ELFOSABI_LINUX
+          || osabi == ELFOSABI_NETBSD
+          || osabi == ELFOSABI_FREEBSD)
+	set_gdbarch_use_struct_convention (gdbarch,
+					   generic_use_struct_convention);
+      else
+	set_gdbarch_use_struct_convention (gdbarch,
+					   ppc_sysv_abi_use_struct_convention);
+    }
+  else
+    {
+      set_gdbarch_use_struct_convention (gdbarch,
+					  generic_use_struct_convention);
+    }
+
   set_gdbarch_frame_chain_valid (gdbarch, file_frame_chain_valid);
   if (osabi == ELFOSABI_LINUX)
     {

Index: ppc-linux-tdep.c
===================================================================
RCS file: /cvs/uberbaum/gdb/ppc-linux-tdep.c,v
retrieving revision 1.13
diff -u -r1.13 ppc-linux-tdep.c
--- ppc-linux-tdep.c	2002/02/24 22:31:19	1.13
+++ ppc-linux-tdep.c	2002/03/22 18:48:39
@@ -414,6 +414,14 @@
    it may be used generically by ports which use either the SysV ABI or
    the EABI */
 
+/* Structures 8 bytes or less long are returned in the r3 & r4
+   registers, according to the SYSV ABI. */
+int
+ppc_sysv_abi_use_struct_convention (int gcc_p, struct type *value_type)
+{
+  return (TYPE_LENGTH (value_type) > 8);
+}
+
 /* round2 rounds x up to the nearest multiple of s assuming that s is a
    power of 2 */
 
Index: ppc-tdep.h
===================================================================
RCS file: /cvs/uberbaum/gdb/ppc-tdep.h,v
retrieving revision 1.6
diff -u -r1.6 ppc-tdep.h
--- ppc-tdep.h	2001/12/30 00:14:50	1.6
+++ ppc-tdep.h	2002/03/22 18:49:05
@@ -31,6 +31,7 @@
 int ppc_linux_frameless_function_invocation (struct frame_info *);
 void ppc_linux_frame_init_saved_regs (struct frame_info *);
 CORE_ADDR ppc_linux_frame_chain (struct frame_info *);
+int ppc_sysv_abi_use_struct_convention (int, struct type *);
 CORE_ADDR ppc_sysv_abi_push_arguments (int, struct value **, CORE_ADDR, int,
 				       CORE_ADDR);
 int ppc_linux_memory_remove_breakpoint (CORE_ADDR addr, char *contents_cache);

Re: [RFA] PPC ABI compliance fix

[Andrew Cagney]

> Index: ppc-linux-tdep.c
> ===================================================================
> RCS file: /cvs/uberbaum/gdb/ppc-linux-tdep.c,v
> retrieving revision 1.13
> diff -u -r1.13 ppc-linux-tdep.c
> --- ppc-linux-tdep.c	2002/02/24 22:31:19	1.13
> +++ ppc-linux-tdep.c	2002/03/22 18:48:39
> @@ -414,6 +414,14 @@
>     it may be used generically by ports which use either the SysV ABI or
>     the EABI */
>  
> +/* Structures 8 bytes or less long are returned in the r3 & r4
> +   registers, according to the SYSV ABI. */
> +int
> +ppc_sysv_abi_use_struct_convention (int gcc_p, struct type *value_type)
> +{
> +  return (TYPE_LENGTH (value_type) > 8);
> +}
> +
>  /* round2 rounds x up to the nearest multiple of s assuming that s is a
>     power of 2 */

Should this live in rs6000-tdep.c?

enjoy,
Andrew

Re: [RFA] PPC ABI compliance fix

[Elena Zannoni]

Andrew Cagney writes:
 > > Index: ppc-linux-tdep.c
 > > ===================================================================
 > > RCS file: /cvs/uberbaum/gdb/ppc-linux-tdep.c,v
 > > retrieving revision 1.13
 > > diff -u -r1.13 ppc-linux-tdep.c
 > > --- ppc-linux-tdep.c	2002/02/24 22:31:19	1.13
 > > +++ ppc-linux-tdep.c	2002/03/22 18:48:39
 > > @@ -414,6 +414,14 @@
 > >     it may be used generically by ports which use either the SysV ABI or
 > >     the EABI */
 > >  
 > > +/* Structures 8 bytes or less long are returned in the r3 & r4
 > > +   registers, according to the SYSV ABI. */
 > > +int
 > > +ppc_sysv_abi_use_struct_convention (int gcc_p, struct type *value_type)
 > > +{
 > > +  return (TYPE_LENGTH (value_type) > 8);
 > > +}
 > > +
 > >  /* round2 rounds x up to the nearest multiple of s assuming that s is a
 > >     power of 2 */
 > 
 > Should this live in rs6000-tdep.c?

Yes, There is a FIXME about that in the file. I am not sure why
sysv_push_arguments is in that file as well. I didn't want to change
it in case there was some real technical problem.

Maybe Kevin knows why?

Elena


 > 
 > enjoy,
 > Andrew
 > 
 > 

Re: [RFA] PPC ABI compliance fix

[Kevin Buettner]

On Mar 22,  3:06pm, Elena Zannoni wrote:

>  > > +ppc_sysv_abi_use_struct_convention (int gcc_p, struct type *value_type)
>  > > +{
>  > > +  return (TYPE_LENGTH (value_type) > 8);
>  > > +}
>  > > +
>  > >  /* round2 rounds x up to the nearest multiple of s assuming that s is a
>  > >     power of 2 */
>  > 
>  > Should this live in rs6000-tdep.c?
> 
> Yes, There is a FIXME about that in the file. I am not sure why
> sysv_push_arguments is in that file as well. I didn't want to change
> it in case there was some real technical problem.
> 
> Maybe Kevin knows why?

There's no real technical problem.  I wrote
ppc_sysv_abi_push_arguments() for Linux, so it ended up in the Linux
file.  I don't think there would be any real problem with moving it to
rs6000-tdep.c, but I would prefer not to since rs6000-tdep.c deals
more with the PowerOpen ABI (in addition to generic arch considerations).

In my opinion, they probably belong in a ppc-sysv-abi-tdep.c file.
But at the moment, there's no real incentive for creating this file
(aside from making things somewhat more tidy) because ppc-linux-tdep.c
needs to be a part of all GDB builds that target Power / PowerPC.

I've been thinking about Andrew's comments from a week or so ago in
the discussion regarding x86-64-linux-tdep.c.  As Andrew noted at the
time, the gdbarch machinery that we now have isn't set up to do
inheritance, and so as a consequence, we end up with the situation
that we have with the ppc*-tdep.c files in which you need include
linux support in a native build for AIX.  It would be nice if we could
restructure things so that there'd be a generic ppc-tdep.c file which
knows nothing about ABIs or OS considerations.  Under that, you'd have
ppc-poweropen-abi.c (or whatever IBM is calling their ABI these days)
and ppc-sysv-abi-tdep.c and perhaps even ppc-sysv-eabi-tdep.c.  Below
that you'd have the various OS specific tdep files.  Unfortunately,
we can't do it yet because the gdbarch machinery won't allow it.

Kevin

Re: [RFA] PPC ABI compliance fix

[Kevin Buettner]

On Mar 22,  2:28pm, Elena Zannoni wrote:

> Back in November, gcc changed the way structures <= 8 bytes are
> returned on the PPC. The change was made to be compliant with the
> SVR4 ABI.
> 
> http://gcc.gnu.org/ml/gcc-patches/2001-11/msg01468.html
> 
> The abi specifies that such structures are passed in r3 and r4.
> Bigger structures are passed in memory.
> 
> Gcc was passing every structure in memory.  The change was made for
> embedded targets, but not for natives (PowerPC Linux, NetBSD, and
> FreeBSD).
> 
> Of course this change breaks binary compatibility with older gcc's.
> Not sure what to do about that, if anything, the case it covers is a
> corner case, anyway.

Hmm...  I have a hunch that, eventually, the natives will need to
change too.

For now though, your change is okay.

(Alternately, we could take a really hard nosed stance and implement
the ABI exactly as written and just put up with the failures.)

Kevin

Re: [RFA] PPC ABI compliance fix

[Elena Zannoni]

Kevin Buettner writes:
 > On Mar 22,  2:28pm, Elena Zannoni wrote:
 > 
 > > Back in November, gcc changed the way structures <= 8 bytes are
 > > returned on the PPC. The change was made to be compliant with the
 > > SVR4 ABI.
 > > 
 > > http://gcc.gnu.org/ml/gcc-patches/2001-11/msg01468.html
 > > 
 > > The abi specifies that such structures are passed in r3 and r4.
 > > Bigger structures are passed in memory.
 > > 
 > > Gcc was passing every structure in memory.  The change was made for
 > > embedded targets, but not for natives (PowerPC Linux, NetBSD, and
 > > FreeBSD).
 > > 
 > > Of course this change breaks binary compatibility with older gcc's.
 > > Not sure what to do about that, if anything, the case it covers is a
 > > corner case, anyway.
 > 
 > Hmm...  I have a hunch that, eventually, the natives will need to
 > change too.
 > 

Gcc is explicitly disabling ABI compliance for the natives. I think
Aldy asked to the gcc mailing list if he could change at least Linux,
and there was a loud 'No'.

 > For now though, your change is okay.
 > 
 > (Alternately, we could take a really hard nosed stance and implement
 > the ABI exactly as written and just put up with the failures.)
 > 

Ironically enough the gcc implementation is not correct, yet. It
screws up where exactly in r3 and r4 it puts the structure if the
structure is strictly less than 8 bytes. :-( 
For instances 5 characters are stored like

R3        R4
. . . h   e l l o

Instead of 

R3        R4
h e l l   o . . .


I thought Aldy was taking care of this, but he has gone on vacation.

Elena


 > Kevin

Re: [RFA] PPC ABI compliance fix

[Elena Zannoni]

Elena Zannoni writes:
 > 
 > Back in November, gcc changed the way structures <= 8 bytes are
 > returned on the PPC. The change was made to be compliant with the
 > SVR4 ABI.
 > 
 > http://gcc.gnu.org/ml/gcc-patches/2001-11/msg01468.html
 > 
 > The abi specifies that such structures are passed in r3 and r4.
 > Bigger structures are passed in memory.
 > 
 > Gcc was passing every structure in memory.  The change was made for
 > embedded targets, but not for natives (PowerPC Linux, NetBSD, and
 > FreeBSD).
 > 
 > Of course this change breaks binary compatibility with older gcc's.
 > Not sure what to do about that, if anything, the case it covers is a
 > corner case, anyway.
 > 
 > 
 > Elena
 > 
 > 
 > 2002-03-21  Elena Zannoni  <ezannoni@redhat.com>
 > 
 >         * ppc-linux-tdep.c (ppc_sysv_abi_use_struct_convention): New
 >         function.  
 > 	* ppc-tdep.h (ppc_sysv_abi_use_struct_convention): Export.
 >         * rs6000-tdep.c (rs6000_gdbarch_init): Use different
 >         structure returning convention for SYSV ABI case, but not
 >         for GNU/Linux, FreeBSD, or NetBSD.
 > 


Committed.

Elena