https://gcc.gnu.org/bugzilla/show_bug.cgi?id=111820
--- Comment #10 from Richard Biener <rguenth at gcc dot gnu.org> ---
(In reply to Hongtao.liu from comment #9)
> > But we end up here with niters_skip being INTEGER_CST and ..
> >
> > > 1421 || (!vect_use_loop_mask_for_alignment_p (loop_vinfo)
> >
> > possibly vect_use_loop_mask_for_alignment_p. Note
> > LOOP_VINFO_PEELING_FOR_ALIGNMENT < 0 simply means the amount of
> > peeling is unknown.
> >
> > But I wonder how we run into this on x86 without enabling
> > loop masking ...
> >
> > > 1422 && LOOP_VINFO_PEELING_FOR_ALIGNMENT (loop_vinfo) < 0))
> > > 1423 {
> > > 1424 if (dump_enabled_p ())
> > > 1425 dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
> > > 1426 "Peeling for alignement is not supported"
> > > 1427 " for nonlinear induction when niters_skip"
> > > 1428 " is not constant.\n");
> > > 1429 return false;
> > > 1430 }
>
> Can you point out where it's assigned as nagative?
> I saw LOOP_VINFO_MASK_SKIP_NITERS is only assigned in
> vect_prepare_for_masked_peels.
Yes.
> when LOOP_VINFO_PEELING_FOR_ALIGNMENT (loop_vinfo) > 0
> it's assigned as vf-npeel(will npeel > vf?)
npeel should be < vf
OK, so it should be positive indeed. But LOOP_VINFO_MASK_SKIP_NITERS
(when vect_use_loop_mask_for_alignment_p ()) means that the
first vector iteration only processes the first
vf - LOOP_VINFO_MASK_SKIP_NITERS scalar iterations, so a
for (i = start; i < end; ++i)
..
loop is executed as
for (i = start - LOOP_VINFO_MASK_SKIP_NITERS; i < end; ++i)
if (i >= start)
{
}
that is, the loop mask is used to mask out the first
LOOP_VINFO_MASK_SKIP_NITERS elements.
It's a bit difficult to force peeling for alignment on x86, but
usually -fno-vect-cost-model will peel a store (it will peel the
most used ref for alignment). With --param vect-partial-vector-usage=2
you should get AVX512 masked alignment peeling then I think.
