Re: [PATCH][i386] Split not+broadcast+pand to broadcast+pandn.

[Andrew Pinski <pinskia@gmail.com>]

On Mon, May 24, 2021 at 11:23 PM Hongtao Liu <crazylht@gmail.com> wrote:
>
> On Tue, May 25, 2021 at 2:11 PM Andrew Pinski <pinskia@gmail.com> wrote:
> >
> > On Mon, May 24, 2021 at 11:03 PM Hongtao Liu via Gcc-patches
> > <gcc-patches@gcc.gnu.org> wrote:
> > >
> > > Hi:
> > >   This patch is about to do transformation like below.
> > >   Bootstrapped and regtested on x86_64-linux-gnu{-m32,}.
> > >   Ok for trunk?
> > >
> > > from
> > >         notl    %edi
> > >         vpbroadcastd    %edi, %xmm0
> > >         vpand   %xmm1, %xmm0, %xmm0
> > > to
> > >         vpbroadcastd    %edi, %xmm0
> > >         vpandn   %xmm1, %xmm0, %xmm0
> > >
> > > gcc/ChangeLog:
> > >
> > >         PR target/100711
> > >         * config/i386/sse.md (*andnot<mode>3): New combine splitter
> > >         after it.
> > >
> > > gcc/testsuite/ChangeLog:
> > >
> > >         PR target/100711
> > >         * gcc.target/i386/avx2-pr100711.c: New test.
> > >         * gcc.target/i386/avx512bw-pr100711.c: New test.
> > >
> >
> >
> > Does it make sense to make this more generic and have combine/simplify
> > rtx instead try:
> > (vec_dup (not)) to (not (vec_dup))
>
> Even w/ that, a combine splitter is still needed since we don't have
> any pandn patterns which contain op1 as vec_duplicate or "not" pattern
> for vector mode,  generic simplification only helps combine/forprop to
> match more possibilities, but not split pattern by itself.

Huh?  This is a 3->2 combining which definitely just happen.
You don't need a "not" pattern for the vector mode for this to happen
since the combining happens without an insn defined.

Thanks,
Andrew Pinski

>
> >
> > Thanks,
> > Andrew Pinski
>
>
>
> --
> BR,
> Hongtao