Re: [PATCH]i386: Optimize pmovskb on zero_extend of subreg HI of the result [PR98461]

[Hongtao Liu <crazylht@gmail.com>]

On Tue, Jan 5, 2021 at 3:20 PM Uros Bizjak <ubizjak@gmail.com> wrote:
>
> On Tue, Jan 5, 2021 at 8:04 AM Uros Bizjak <ubizjak@gmail.com> wrote:
> > >
> > > +(define_split
> > > +  [(set (match_operand:SI 0 "register_operand")
> > > +        (zero_extend:SI
> > > +          (not:HI
> > > +            (subreg:HI
> > > +              (unspec:SI
> > > +                [(match_operand:V16QI 1 "register_operand")]
> > > +                UNSPEC_MOVMSK) 0))))]
> > > +  "TARGET_SSE2"
> > > +  [(set (match_dup 2)
> > > +        (unspec:SI [(match_dup 1)] UNSPEC_MOVMSK))
> > > +   (set (match_dup 0)
> > > +        (match_dup 3))]
> >
> > Just write:
> >
> > (set (match_dup 0)
> >     (xor:SI (match_dup 2)(const_int 65535))
>

Yes, changed.

> BTW: This could be a universal combine splitter to simplify
>
> unsigned int foo (unsigned short z)
> {
>     return (unsigned short)~z;
> }
>
> Trying 7 -> 8:
>    7: r87:HI=~r88:SI#0
>      REG_DEAD r88:SI
>    8: r86:SI=zero_extend(r87:HI)
>      REG_DEAD r87:HI
> Failed to match this instruction:
> (set (reg:SI 86)
>    (zero_extend:SI (not:HI (subreg:HI (reg:SI 88) 0))))
>
> But combine does not "split" to one insns.

Yes, according to PSabi, the top half of the register is not
necessarily 0, so if you add the splitter, it just changes from notl +
movzwl to xor + movzwl, which doesn't look better?

>
> Uros.



-- 
BR,
Hongtao