On Mon, Jun 16, 2025 at 4:28 PM Hongtao Liu <[email protected]> wrote:
>
> On Mon, Jun 16, 2025 at 4:30 PM Hongtao Liu <[email protected]> wrote:
> >
> > >+enum redundant_load_kind
> > >+{
> > >+ LOAD_CONST0_VECTOR,
> > >+ LOAD_CONSTM1_VECTOR,
> > >+ LOAD_VECTOR
> > >+};
> > Perhaps rename to x86_cse_kind, X86_CSE_CONST0_VECTOR,
> > X86_CSE_CONSTM1_VECTOR, X86_CSE_VEC_DUP?
> > LOAD sounds a bit ambiguous.
> > Similar to ix86_get_vector_load_mode -> ix86_get_vector_cse_mode?
> >
> > >+ if (SUBREG_P (op))
> > >+ reg = SUBREG_REG (op);
> > Need to make sure subreg is a lowpart of op, it can't be a paradoxical
> > subreg.
> >
> > >+ /* Only single def chain is supported. */
> > >+ df_ref ref = DF_REG_DEF_CHAIN (REGNO (reg));
> > >+ if (!ref || DF_REF_NEXT_REG (ref) != nullptr)
> > >+ return nullptr;
> >
> > Could we just reuse df_find_single_def_src and add extra code to get
> > the insn and handle *insn_p = ***.
> > Also I notice df_find_single_def_src exclude DF_REF_IS_ARTIFICIAL
I will exclude DF_REF_IS_ARTIFICIAL. Since df_find_single_def_src
calls function_invariant_p, it doesn't work on
extern __m512i sinkz;
extern __m256i sinky;
extern char f;
void
foo(char c, int x)
{
c += f;
sinkz = _mm512_set1_epi8(c);
if (x == 2)
f += 3;
sinky = _mm256_set1_epi8(c);
}
> > which is not in your original patch, after excluding it, do below
> > codes still be needed?
> >
> I mean this part.
> + && (load->bb == bb
> + || kind < LOAD_VECTOR
> + /* Non all 0s/1s vector load must be in the same
> + basic block if it is in a recursive call. */
> + || !recursive_call_p)
>
> >
> > >+ ix86_place_single_vector_set (load->broadcast_reg,
> > >+ load->broadcast_source,
> > >+ load->bbs,
> > >+ (load->kind >= LOAD_VECTOR
> > Just load->kind == LOAD_VECTOR?
>
>
>
> --
> BR,
> Hongtao
--
H.J.
