https://gcc.gnu.org/bugzilla/show_bug.cgi?id=98167
--- Comment #15 from Hongtao.liu <crazylht at gmail dot com> ---
(In reply to Andrew Pinski from comment #14)
> (In reply to Hongtao.liu from comment #13)
> > fold shulfps to vec_perm_exp, but still 2 shulfps are generated.
> >
> > __m128 f (__m128 a, __m128 b)
> > {
> > vector(4) float _3;
> > vector(4) float _5;
> > vector(4) float _6;
> >
> > ;; basic block 2, loop depth 0
> > ;; pred: ENTRY
> > _3 = VEC_PERM_EXPR <b_2(D), b_2(D), { 0, 0, 0, 0 }>;
> > _5 = VEC_PERM_EXPR <a_4(D), a_4(D), { 0, 0, 0, 0 }>;
> > _6 = _3 * _5;
> > return _6;
> > ;; succ: EXIT
> >
> > }
>
> So this is a bit more complex as not all targets have a good extract/dup
> functionary for scalars. So maybe this should be done as a define_insn for
> x86.
No need for extract/dup, if both perm indexes is the same, it can be c = a * b,
and vec_perm_expr (c, c, index}. it seems a quite general optimization which
could apply to all other operations.
