http://gcc.gnu.org/bugzilla/show_bug.cgi?id=60276
--- Comment #7 from Richard Biener <rguenth at gcc dot gnu.org> ---
(In reply to Richard Biener from comment #6)
> But you can clearly see that this isn't valid vectorization because of the
> read-after-write dependence of out[x].
Well, only after unrolling which happens here - then the introduced dependence
(we CSE the in-source one) becomes explicit:
tem0 = lp[target] - ((hp[target] + hp[target - 1] + 2) >> 2);
out[x] = tem0;
out[x - 1] = hp[target - 1] + ((out[x - 2] + tem0) >> 1);
tem2 = lp[target + 1] - ((hp[target + 1] + hp[target] + 2) >> 2);
out[x + 2] = tem2;
out[x + 1] = hp[target] + ((out[x] + tem2) >> 1);
thus
static bool
vect_analyze_data_ref_dependence (struct data_dependence_relation *ddr,
loop_vec_info loop_vinfo, int *max_vf)
{
...
if (dist > 0 && DDR_REVERSED_P (ddr))
{
/* If DDR_REVERSED_P the order of the data-refs in DDR was
reversed (to make distance vector positive), and the actual
distance is negative. */
if (dump_enabled_p ())
dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
"dependence distance negative.\n");
continue;
}
becomes invalid once we unroll if the distance is less than the
vectorization factor. Unless we apply the same CSE during vectorization
transform.