On Wed, May 26, 2021 at 12:12 PM Andrew Pinski <pins...@gmail.com> wrote:
>
> On Tue, May 25, 2021 at 6:17 PM Hongtao Liu <crazy...@gmail.com> wrote:
> >
> > Update patch:
> > The new patch simplify (vec_duplicate (not (nonimmedaite_operand)))
> > to (not (vec_duplicate (nonimmedaite_operand))). This is not a
> > straightforward simplification, just adding some tendency to pull not
> > out of vec_duplicate.
> >
> > For i386, it will enable below opt
> >
> > from
> > notl %edi
> > vpbroadcastd %edi, %xmm0
> > vpand %xmm1, %xmm0, %xmm0
> > to
> > vpbroadcastd %edi, %xmm0
> > vpandn %xmm1, %xmm0, %xmm0
> >
> > Bootstrapped and regtested on x86_64-linux-gnu{-m32,}.
> > Ok for trunk?
> > gcc/ChangeLog:
> >
> > PR target/100711
> > * simplify-rtx.c (simplify_unary_operation_1):
> > Simplify (vec_duplicate (not (nonimmedaite_operand)))
> > to (not (vec_duplicate (nonimmedaite_operand))).
> >
> > gcc/testsuite/ChangeLog:
> >
> > PR target/100711
> > * gcc.target/i386/avx2-pr100711.c: New test.
> > * gcc.target/i386/avx512bw-pr100711.c: New test.
>
> This patch should not use nonimmedaite_operand at all in
There's no simplification opportunity for nonimmediate_operand, but
I'm not sure for other cases(not constants).
Reading from codes in case NOT of simplify_unary_operation_1, there
may be (vec_duplicate (not (plus X - 1))???
> simplify-rtx.c. Rather use !CONSTANT_P (XEXP (op, 0)) instead.
> And even then (not CONST_INT) will never be there anyways as it will
> always be simplified to a constant in the first place. So removing
> that check is fine.
>
> Thanks,
> Andrew
--
BR,
Hongtao
