[RFA] mips, floating point arg passing

[RFA] mips, floating point arg passing

[Michael Snyder]

[-- Attachment #1: Type: text/plain, Size: 89 bytes --]

GDB is causing SIGFPE on target boards that have a FPU,
when compiled for a 32 bit ABI.


[-- Attachment #2: d3 --]
[-- Type: text/plain, Size: 1885 bytes --]

2006-05-31  Michael Snyder  <msnyder@redhat.com>

	* mips-tdep.c (mips_eabi_push_dummy_call): For 32 bit ABI, to
	decide how many registers it takes to pass a floating point
	argument, what matters is the size of a floating point register
	(not the size of a general purpose register).
	(mips_o32_push_dummy_call): Ditto.

Index: mips-tdep.c
===================================================================
RCS file: /cvs/src/src/gdb/mips-tdep.c,v
retrieving revision 1.393
diff -p -r1.393 mips-tdep.c
*** mips-tdep.c	31 May 2006 23:15:50 -0000	1.393
--- mips-tdep.c	31 May 2006 23:28:59 -0000
*************** mips_eabi_push_dummy_call (struct gdbarc
*** 2494,2500 ****
        if (fp_register_arg_p (typecode, arg_type)
  	  && float_argreg <= MIPS_LAST_FP_ARG_REGNUM)
  	{
! 	  if (mips_abi_regsize (gdbarch) < 8 && len == 8)
  	    {
  	      int low_offset = TARGET_BYTE_ORDER == BFD_ENDIAN_BIG ? 4 : 0;
  	      unsigned long regval;
--- 2494,2500 ----
        if (fp_register_arg_p (typecode, arg_type)
  	  && float_argreg <= MIPS_LAST_FP_ARG_REGNUM)
  	{
! 	  if (register_size (gdbarch, float_argreg) < 8 && len == 8)
  	    {
  	      int low_offset = TARGET_BYTE_ORDER == BFD_ENDIAN_BIG ? 4 : 0;
  	      unsigned long regval;
*************** mips_o32_push_dummy_call (struct gdbarch
*** 3118,3124 ****
        if (fp_register_arg_p (typecode, arg_type)
  	  && float_argreg <= MIPS_LAST_FP_ARG_REGNUM)
  	{
! 	  if (mips_abi_regsize (gdbarch) < 8 && len == 8)
  	    {
  	      int low_offset = TARGET_BYTE_ORDER == BFD_ENDIAN_BIG ? 4 : 0;
  	      unsigned long regval;
--- 3118,3124 ----
        if (fp_register_arg_p (typecode, arg_type)
  	  && float_argreg <= MIPS_LAST_FP_ARG_REGNUM)
  	{
! 	  if (register_size (gdbarch, float_argreg) < 8 && len == 8)
  	    {
  	      int low_offset = TARGET_BYTE_ORDER == BFD_ENDIAN_BIG ? 4 : 0;
  	      unsigned long regval;

Re: [RFA] mips, floating point arg passing

[Michael Snyder]

Ping?

Michael Snyder wrote:
> GDB is causing SIGFPE on target boards that have a FPU,
> when compiled for a 32 bit ABI.
> 
> 
> ------------------------------------------------------------------------
> 
> 2006-05-31  Michael Snyder  <msnyder@redhat.com>
> 
> 	* mips-tdep.c (mips_eabi_push_dummy_call): For 32 bit ABI, to
> 	decide how many registers it takes to pass a floating point
> 	argument, what matters is the size of a floating point register
> 	(not the size of a general purpose register).
> 	(mips_o32_push_dummy_call): Ditto.
> 
> Index: mips-tdep.c
> ===================================================================
> RCS file: /cvs/src/src/gdb/mips-tdep.c,v
> retrieving revision 1.393
> diff -p -r1.393 mips-tdep.c
> *** mips-tdep.c	31 May 2006 23:15:50 -0000	1.393
> --- mips-tdep.c	31 May 2006 23:28:59 -0000
> *************** mips_eabi_push_dummy_call (struct gdbarc
> *** 2494,2500 ****
>         if (fp_register_arg_p (typecode, arg_type)
>   	  && float_argreg <= MIPS_LAST_FP_ARG_REGNUM)
>   	{
> ! 	  if (mips_abi_regsize (gdbarch) < 8 && len == 8)
>   	    {
>   	      int low_offset = TARGET_BYTE_ORDER == BFD_ENDIAN_BIG ? 4 : 0;
>   	      unsigned long regval;
> --- 2494,2500 ----
>         if (fp_register_arg_p (typecode, arg_type)
>   	  && float_argreg <= MIPS_LAST_FP_ARG_REGNUM)
>   	{
> ! 	  if (register_size (gdbarch, float_argreg) < 8 && len == 8)
>   	    {
>   	      int low_offset = TARGET_BYTE_ORDER == BFD_ENDIAN_BIG ? 4 : 0;
>   	      unsigned long regval;
> *************** mips_o32_push_dummy_call (struct gdbarch
> *** 3118,3124 ****
>         if (fp_register_arg_p (typecode, arg_type)
>   	  && float_argreg <= MIPS_LAST_FP_ARG_REGNUM)
>   	{
> ! 	  if (mips_abi_regsize (gdbarch) < 8 && len == 8)
>   	    {
>   	      int low_offset = TARGET_BYTE_ORDER == BFD_ENDIAN_BIG ? 4 : 0;
>   	      unsigned long regval;
> --- 3118,3124 ----
>         if (fp_register_arg_p (typecode, arg_type)
>   	  && float_argreg <= MIPS_LAST_FP_ARG_REGNUM)
>   	{
> ! 	  if (register_size (gdbarch, float_argreg) < 8 && len == 8)
>   	    {
>   	      int low_offset = TARGET_BYTE_ORDER == BFD_ENDIAN_BIG ? 4 : 0;
>   	      unsigned long regval;

Re: [RFA] mips, floating point arg passing

[Joel Brobecker]

Hello Michael,

> GDB is causing SIGFPE on target boards that have a FPU,
> when compiled for a 32 bit ABI.

I've only worked with the n32 ABI so far, so I'm not familiar with
the EABI nor o32 ABI. But I presume you know what you're talking
about :-).

> 2006-05-31  Michael Snyder  <msnyder@redhat.com>
> 
> 	* mips-tdep.c (mips_eabi_push_dummy_call): For 32 bit ABI, to
> 	decide how many registers it takes to pass a floating point
> 	argument, what matters is the size of a floating point register
> 	(not the size of a general purpose register).
> 	(mips_o32_push_dummy_call): Ditto.

I'm not sure I really have the knowledge to approve but Daniel asked
for help in that department, so I'll give it a try.

This patch seems pretty logical to me, so go ahead and commit. Is it
possible to run the testsuite on either o32, or on a target board?

Thanks,
-- 
Joel

Re: [RFA] mips, floating point arg passing

[Daniel Jacobowitz]

On Fri, Jun 09, 2006 at 11:05:43AM -0700, Joel Brobecker wrote:
> Hello Michael,
> 
> > GDB is causing SIGFPE on target boards that have a FPU,
> > when compiled for a 32 bit ABI.
> 
> I've only worked with the n32 ABI so far, so I'm not familiar with
> the EABI nor o32 ABI. But I presume you know what you're talking
> about :-).

I can easily test on O32; is there a case in the testsuite that would
cover this?  I don't remember any argument passing tests failing.

Anyway, o32 has 32-bit floating point registers according to the ABI,
so I don't see how this could affect o32.

-- 
Daniel Jacobowitz
CodeSourcery

Re: [RFA] mips, floating point arg passing

[Mark Kettenis]

> Date: Wed, 31 May 2006 16:32:44 -0700
> From: Michael Snyder <msnyder@redhat.com>
> 
> GDB is causing SIGFPE on target boards that have a FPU,
> when compiled for a 32 bit ABI.

Hmm, aren't all registers supposed to be 32-bit for o32.  It seems to
me that we should hardcode that in the *_o32_* functions, instead of
trying to be register-size agnostic.

Mark

> 2006-05-31  Michael Snyder  <msnyder@redhat.com>
> 
> 	* mips-tdep.c (mips_eabi_push_dummy_call): For 32 bit ABI, to
> 	decide how many registers it takes to pass a floating point
> 	argument, what matters is the size of a floating point register
> 	(not the size of a general purpose register).
> 	(mips_o32_push_dummy_call): Ditto.

Re: [RFA] mips, floating point arg passing

[Michael Snyder]

Joel Brobecker wrote:
> Hello Michael,
> 
> 
>>GDB is causing SIGFPE on target boards that have a FPU,
>>when compiled for a 32 bit ABI.
> 
> 
> I've only worked with the n32 ABI so far, so I'm not familiar with
> the EABI nor o32 ABI. But I presume you know what you're talking
> about :-).
> 
> 
>>2006-05-31  Michael Snyder  <msnyder@redhat.com>
>>
>>	* mips-tdep.c (mips_eabi_push_dummy_call): For 32 bit ABI, to
>>	decide how many registers it takes to pass a floating point
>>	argument, what matters is the size of a floating point register
>>	(not the size of a general purpose register).
>>	(mips_o32_push_dummy_call): Ditto.
> 
> 
> I'm not sure I really have the knowledge to approve but Daniel asked
> for help in that department, so I'll give it a try.
> 
> This patch seems pretty logical to me, so go ahead and commit. Is it
> possible to run the testsuite on either o32, or on a target board?

I did -- but it was quite some time ago now, and I couldn't
really put my hands on them.  I'm pretty sure they got better.
Certainly they didn't get worse.

Committing.

Re: [RFA] mips, floating point arg passing

[Mark Kettenis]

> Date: Mon, 19 Jun 2006 11:50:11 -0700
> From: Michael Snyder <msnyder@redhat.com>
> 
> Joel Brobecker wrote:
> > Hello Michael,
> > 
> > 
> >>GDB is causing SIGFPE on target boards that have a FPU,
> >>when compiled for a 32 bit ABI.
> > 
> > 
> > I've only worked with the n32 ABI so far, so I'm not familiar with
> > the EABI nor o32 ABI. But I presume you know what you're talking
> > about :-).
> > 
> > 
> >>2006-05-31  Michael Snyder  <msnyder@redhat.com>
> >>
> >>	* mips-tdep.c (mips_eabi_push_dummy_call): For 32 bit ABI, to
> >>	decide how many registers it takes to pass a floating point
> >>	argument, what matters is the size of a floating point register
> >>	(not the size of a general purpose register).
> >>	(mips_o32_push_dummy_call): Ditto.
> > 
> > 
> > I'm not sure I really have the knowledge to approve but Daniel asked
> > for help in that department, so I'll give it a try.
> > 
> > This patch seems pretty logical to me, so go ahead and commit. Is it
> > possible to run the testsuite on either o32, or on a target board?
> 
> I did -- but it was quite some time ago now, and I couldn't
> really put my hands on them.  I'm pretty sure they got better.
> Certainly they didn't get worse.
> 
> Committing.
> 

Whoa wait, you didn't answer my questions about the patch...

Re: [RFA] mips, floating point arg passing

[Daniel Jacobowitz]

On Sun, Jun 18, 2006 at 01:35:53AM +0200, Mark Kettenis wrote:
> > Date: Wed, 31 May 2006 16:32:44 -0700
> > From: Michael Snyder <msnyder@redhat.com>
> > 
> > GDB is causing SIGFPE on target boards that have a FPU,
> > when compiled for a 32 bit ABI.
> 
> Hmm, aren't all registers supposed to be 32-bit for o32.  It seems to
> me that we should hardcode that in the *_o32_* functions, instead of
> trying to be register-size agnostic.

I am inclined to agree; what happened (IIRC) is that these
similar-looking functions were split from an identical parent at one
point to reduce changes for one ABI breaking another, and never cleaned
up.

-- 
Daniel Jacobowitz
CodeSourcery