Q: -mcpu= vs. -march= for VR41xx specific instructions

Q: -mcpu= vs. -march= for VR41xx specific instructions

[Johannes Stezenbach]

I am trying to implement power management for an embedded
device using a NEC VR4120 CPU core, which has the special
instructions "standby", "suspend" and "hibernate".

I use a toolchain based on binutils 2.12.90.0.1 and
gcc 2.95.4-debian.

To use that instructions I have to pass -march=vr4100
to the assembler. Unfortunately, -march and -mcpu cannot
be used together, so first I changed arch/mips/Makefile
from
  GCCFLAGS += -mcpu=r4600 -mips2 -Wa,--trap
to
  GCCFLAGS += -Wa,-march=vr4100 -mips2 -Wa,--trap

This works, but I am unshure what the effects of the
missing -mcpu switch are wrt the code generated by gcc.
AFAICS the kernel still works, but is the generated
code slower or subtly incorrect?

Then I the tried to compile the kernel with the standard
GCCFLAGS, setting -march=vr4100 only for the one
file which contains the standby/suspend/hibernate instructions.
This fails a link time with:

  mips-linux-ld: power.o: uses different e_flags (0x1100) fields than previous modules (0x1000)
  Bad value: failed to merge target specific data of file power.o

I looked at an old arch/mips/Makefile from
http://sourceforge.net/projects/linux-vr/, which has:
  CFLAGS += -mcpu=r4600 -mips2 -Wa,-m4100,--trap
This does not work with my toolchain.

So I think I need either:
- make gas accept -march=vr4100 along with -mcpu=r4600 (or -mcpu=r4100?)
- or have a ".set vr4100" directive to enable the vr41xx specific
  instructions where needed, without changing the flags in the
  ELF header
- or make the linker link modules with different (but compatible) e_flags
- or is "GCCFLAGS += -Wa,-march=vr4100 -mips2 -Wa,--trap" perfect?


Please, does anybody have suggestions how to solve this issue?


Regards,
Johannes

Re: Q: -mcpu= vs. -march= for VR41xx specific instructions

[Jim Paris]

> I am trying to implement power management for an embedded
> device using a NEC VR4120 CPU core, which has the special
> instructions "standby", "suspend" and "hibernate".

> So I think I need either:
> - make gas accept -march=vr4100 along with -mcpu=r4600 (or -mcpu=r4100?)
> - or have a ".set vr4100" directive to enable the vr41xx specific
>   instructions where needed, without changing the flags in the
>   ELF header
> - or make the linker link modules with different (but compatible) e_flags
> - or is "GCCFLAGS += -Wa,-march=vr4100 -mips2 -Wa,--trap" perfect?

For lots of discussion of this, see

   http://sources.redhat.com/ml/binutils/2001-10/threads.html#00504

then

   http://sources.redhat.com/ml/binutils/2001-11/threads.html#00001

I don't remember where things currently stand.

-jim

Re: Q: -mcpu= vs. -march= for VR41xx specific instructions

[Thiemo Seufer]

Johannes Stezenbach wrote:
[snip]
> I use a toolchain based on binutils 2.12.90.0.1 and
> gcc 2.95.4-debian.
> 
> To use that instructions I have to pass -march=vr4100
> to the assembler. Unfortunately, -march and -mcpu cannot
> be used together, so first I changed arch/mips/Makefile
> from
>   GCCFLAGS += -mcpu=r4600 -mips2 -Wa,--trap

If this gcc already supports vr4100 you could use

	GCCFLAGS += -mcpu=vr4100 -mips2 -Wa,--trap

> to
>   GCCFLAGS += -Wa,-march=vr4100 -mips2 -Wa,--trap
> 
> This works, but I am unshure what the effects of the
> missing -mcpu switch are wrt the code generated by gcc.
> AFAICS the kernel still works, but is the generated
> code slower or subtly incorrect?

I don't know what the compiler does then. I assume it defaults to
r3000 scheduling/opcodes.

The assembler will emit vr4100 code scheduled for vr4100. If you still
want r4600 scheduling (I doubt so), you could add

	-Wa,-mtune=r4600

to the GCCFLAGS line.

[snip]
> So I think I need either:
> - make gas accept -march=vr4100 along with -mcpu=r4600 (or -mcpu=r4100?)

-mcpu is deprecated and remains for backward compatibility to gcc < 3.0.
it should be replaced by -march/-mtune.

> - or have a ".set vr4100" directive to enable the vr41xx specific
>   instructions where needed, without changing the flags in the
>   ELF header

I would suggest the syntax

	.set march=vr4100

This is in my TODO list for gas, but don't hold your breath.

> - or make the linker link modules with different (but compatible) e_flags
> - or is "GCCFLAGS += -Wa,-march=vr4100 -mips2 -Wa,--trap" perfect?
> 
> 
> Please, does anybody have suggestions how to solve this issue?

The real fix is to use a newer compiler (gcc >= 3). :-)


Thiemo

Re: Q: -mcpu= vs. -march= for VR41xx specific instructions

[Johannes Stezenbach]

On Thu, Mar 14, 2002 at 07:10:31PM +0100, Thiemo Seufer wrote:
> Johannes Stezenbach wrote:
> >   GCCFLAGS += -Wa,-march=vr4100 -mips2 -Wa,--trap
> > 
> > This works, but I am unshure what the effects of the
> > missing -mcpu switch are wrt the code generated by gcc.
> > AFAICS the kernel still works, but is the generated
> > code slower or subtly incorrect?
> 
> I don't know what the compiler does then. I assume it defaults to
> r3000 scheduling/opcodes.

How bad is that compared to -mcpu=vr4100?

> I would suggest the syntax
> 
> 	.set march=vr4100
> 
> This is in my TODO list for gas, but don't hold your breath.

OK.

> The real fix is to use a newer compiler (gcc >= 3). :-)

Not too long ago people here told me that gcc 3.x is still
not ready for production use. Or is gcc-3.1-pre from CVS
ready for prime time?


Thanks,
Johannes