Re: Still can't make stfiwx actually store directly, but...

[Geoffrey Keating <geoffk at apple dot com>]

On 17/05/2005, at 5:40 PM, Richard Henderson wrote:

> On Tue, May 17, 2005 at 05:23:31PM -0700, Geoffrey Keating wrote:
>
> > Combine only knows about pseudos, not memory locations,  
> > unfortunately.
>
> You're wrong there.  For Alpha it's happy to combine
>
> (insn 13 8 15 0 (set (reg:DI 72)
>         (fix:DI (reg/v:DF 70 [ d ]))) 124 {*fix_truncdfdi2})
>
> (insn 15 13 16 0 (set (mem:SI (reg/v/f:DI 69 [ x ]) [3 S4 A32])
>         (subreg:SI (reg:DI 72) 0)) 239 {*movsi})
>
> into
>
> (insn 15 13 16 0 (parallel [
>             (set (mem:SI (reg:DI 16 $16 [ x ]) [3 S4 A32])
>                 (subreg:SI (fix:DI (reg:DF 49 $f17 [ d ])) 0))
>             (clobber (scratch:DI))
>         ]) 122 {*fix_truncdfsi_internal})
>
> I have essentially the same hardware constraints on Alpha;
> you should be able to make this work on powerpc...

Unfortunately, the powerpc operation is not fix:DI.  It's fix:SI but  
it puts a NaN pattern in the high bits.  rs6000 doesn't allow SImode  
values in FPRs.

Here's what I wish combine would combine, but it doesn't:

(insn 13 8 14 0 (parallel [
            (set (mem/i:SI (plus:SI (reg/f:SI 30 r30)
                        (const_int 24 [0x18])) [3 S4 A32])
                (fix:SI (reg:DF 33 f1 [ d ])))
            (clobber (reg:DI 122))
        ]) 223 {fix_truncdfsi2_internal_gfxopt} (nil)
    (expr_list:REG_DEAD (reg:DF 33 f1 [ d ])
        (expr_list:REG_UNUSED (reg:DI 122)
            (nil))))

(insn 14 13 16 0 (set (reg:SI 121)
        (mem/i:SI (plus:SI (reg/f:SI 30 r30)
                (const_int 24 [0x18])) [3 S4 A32])) 295  
{*movsi_internal1} (nil)
    (expr_list:REG_EQUAL (fix:SI (reg/v:DF 119 [ d ]))
        (nil)))

(insn 16 14 17 0 (set (mem:SI (reg:SI 3 r3 [ x ]) [3 S4 A32])
        (reg:SI 121)) 295 {*movsi_internal1} (insn_list:REG_DEP_TRUE  
14 (nil))
    (expr_list:REG_DEAD (reg:SI 3 r3 [ x ])
        (expr_list:REG_DEAD (reg:SI 121)
            (nil))))

I tried making the MEM a pseudo instead, but that doesn't work;  
reload doesn't know how to force a psuedo into memory, apparently.

Attachment: smime.p7s
Description: S/MIME cryptographic signature