Re: Indy 64 or 32 bit?

[Ralf Baechle <ralf@oss.sgi.com>]

On Wed, Aug 08, 2001 at 04:12:39PM -0400, J. Scott Kasten wrote:

>     100012b0:   ffc20028        sd      $v0,40($s8)
>     100012b4:   dfc20028        ld      $v0,40($s8)
>     100012b8:   3c030807        lui     $v1,0x807
>     100012bc:   34630605        ori     $v1,$v1,0x605
>     100012c0:   00031c38        dsll    $v1,$v1,0x10
>     100012c4:   34630403        ori     $v1,$v1,0x403
>     100012c8:   00031c38        dsll    $v1,$v1,0x10
>     100012cc:   34630201        ori     $v1,$v1,0x201
>     100012d0:   fc430000        sd      $v1,0($v0)
>     100012d4:   dfc20020        ld      $v0,32($s8)
>     100012d8:   64430001        daddiu  $v1,$v0,1
>     100012dc:   ffc30020        sd      $v1,32($s8)
>     100012e0:   dfc20028        ld      $v0,40($s8)
> 
> Ouch that's a painfull load.  It's interesting that in both cases, it
> effectively makes the constant load a 16 bit operation, does the math 64
> bit, and stores 64 bit.

With GP optimization something like the following would have been created:

        dla     $gp, _gp + $8000

        ld      $reg1, %gprel(var)($gp)
        daddiu  $reg1, $reg1, 1
        sd      $reg1, %gprel(var)($gp)

_gp:
var:    .word   0

Gas is lame.  MIPS/ELF gas is more.

  Ralf