Excessive limitation on gas for MIPS?

Excessive limitation on gas for MIPS?

[Sean McNeil]

Hello all,

I was testing the CVS repository binutils against an older version of
GCC and I was getting some errors that didn't look right at all:

Error: Number (0x0) larger than 32 bits

where 0x0 was always smaller than 32 bits.

Looking into it a little, I think there is an excessive limitation in
load_register and the following patch fixed my problem:

RCS file: /cvs/src/src/gas/config/tc-mips.c,v
retrieving revision 1.197
diff -c -3 -p -r1.197 tc-mips.c
*** gas/config/tc-mips.c        12 Mar 2003 23:07:37 -0000      1.197
--- gas/config/tc-mips.c        18 Mar 2003 03:29:53 -0000
*************** load_register (counter, reg, ep, dbl)
*** 3582,3588 ****
   
    /* The value is larger than 32 bits.  */
   
!   if (HAVE_32BIT_GPRS)
      {
        as_bad (_("Number (0x%lx) larger than 32 bits"),
              (unsigned long) ep->X_add_number);
--- 3582,3588 ----
   
    /* The value is larger than 32 bits.  */
   
!   if (HAVE_32BIT_GPRS && ! dbl)
      {
        as_bad (_("Number (0x%lx) larger than 32 bits"),
              (unsigned long) ep->X_add_number);

I would appreciate any feedback on the validity of this relaxation.  It
appears that instructions are legitimate with the fix/patch.

Cheers,
Sean

Re: Excessive limitation on gas for MIPS?

[Thiemo Seufer]

Sean McNeil wrote:
[snip]
> *** gas/config/tc-mips.c        12 Mar 2003 23:07:37 -0000      1.197
> --- gas/config/tc-mips.c        18 Mar 2003 03:29:53 -0000
> *************** load_register (counter, reg, ep, dbl)
> *** 3582,3588 ****
>    
>     /* The value is larger than 32 bits.  */
>    
> !   if (HAVE_32BIT_GPRS)
>       {
>         as_bad (_("Number (0x%lx) larger than 32 bits"),
>               (unsigned long) ep->X_add_number);
> --- 3582,3588 ----
>    
>     /* The value is larger than 32 bits.  */
>    
> !   if (HAVE_32BIT_GPRS && ! dbl)
>       {
>         as_bad (_("Number (0x%lx) larger than 32 bits"),
>               (unsigned long) ep->X_add_number);
> 
> I would appreciate any feedback on the validity of this relaxation.  It
> appears that instructions are legitimate with the fix/patch.

It is wrong. HAVE_32BIT_GPRS checks if the CPU/the ABI allows to use
registers wider than 32 bits. The case you are observing should be
handled by the code immediately above the part you changed.

Could you provide the preprocessed assembly which triggers this case?


Thiemo

Re: Excessive limitation on gas for MIPS?

[Thiemo Seufer]

Sean McNeil wrote:
> > It is wrong. HAVE_32BIT_GPRS checks if the CPU/the ABI allows to use
> > registers wider than 32 bits. The case you are observing should be
> > handled by the code immediately above the part you changed.
> > 
> > Could you provide the preprocessed assembly which triggers this case?
> 
> I am using the following switches to an older GCC:
> 
> -G0 -mips4 -mlong32 -EB -ansi -fno-builtin -O -Wall
> 
> The assembler end up getting called with:
> 
> -G0 -EB -mips4 -O2 -v -mabi=32
> 
> even though cc1 was called with:
> 
> -mfp64 -mgp64 -mexplicit-type-size -G0 -meb -quiet -dumpbase
  ^^^^^^^^^^^^^
The compiler generates code for 64 bit wide registers. The assembler
shouldn't be called with -mabi=32 in this case. Adjusting the spec file
should help.

> bcm1250MacEnd.c -mips4 -mlong32
> 
> It turns out that there is a spec in the compiler that passes -mabi=32
> when no abi was specified.  It is this that is causing trouble as the
> compiler is generating assembly like:
> 
> 	dli	$3,0x5ee400000040000
> 
> the assembler is only displaying the lower significant 32 bits in the
> message, by the way.
> 
> so,
> 
> 1) an abi of 32 implies 32-bit registers?  I shouldn't think so.

It does. o32 is the ABI which was invented for the old 32bit MIPS
CPUs.

> 2) Doesn't a "dli" instruction state "double load" or 64 bits?

Yes.

> 3) This behavior has changed from previous gas versions.

Yes, gas expands now 64 bit macros (as it should).

> Thanks for the input.  I am not on the binutils list, so I had to check
> the archive to see your response.

A 'Mail-Followup-To:' mail header would have helped.


Thiemo

Re: Excessive limitation on gas for MIPS?

[Sean McNeil]

> It is wrong. HAVE_32BIT_GPRS checks if the CPU/the ABI allows to use
> registers wider than 32 bits. The case you are observing should be
> handled by the code immediately above the part you changed.
> 
> Could you provide the preprocessed assembly which triggers this case?

I am using the following switches to an older GCC:

-G0 -mips4 -mlong32 -EB -ansi -fno-builtin -O -Wall

The assembler end up getting called with:

-G0 -EB -mips4 -O2 -v -mabi=32

even though cc1 was called with:

-mfp64 -mgp64 -mexplicit-type-size -G0 -meb -quiet -dumpbase
bcm1250MacEnd.c -mips4 -mlong32

It turns out that there is a spec in the compiler that passes -mabi=32
when no abi was specified.  It is this that is causing trouble as the
compiler is generating assembly like:

	dli	$3,0x5ee400000040000

the assembler is only displaying the lower significant 32 bits in the
message, by the way.

so,

1) an abi of 32 implies 32-bit registers?  I shouldn't think so.
2) Doesn't a "dli" instruction state "double load" or 64 bits?
3) This behavior has changed from previous gas versions.

Thanks for the input.  I am not on the binutils list, so I had to check
the archive to see your response.

Cheers,
Sean