https://gcc.gnu.org/bugzilla/show_bug.cgi?id=110310
Bug ID: 110310
Summary: vector epilogue handling is inefficient
Product: gcc
Version: 14.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: tree-optimization
Assignee: unassigned at gcc dot gnu.org
Reporter: rguenth at gcc dot gnu.org
Target Milestone: ---
It looks like we apply some analysis only when transforming the main vector
loop. In particular vect_do_peeling does the following which elides a
vector epilogue after costing.
/* If we know the number of scalar iterations for the main loop we should
check whether after the main loop there are enough iterations left over
for the epilogue. */
if (vect_epilogues
&& LOOP_VINFO_NITERS_KNOWN_P (loop_vinfo)
&& prolog_peeling >= 0
&& known_eq (vf, lowest_vf))
{
unsigned HOST_WIDE_INT eiters
= (LOOP_VINFO_INT_NITERS (loop_vinfo)
- LOOP_VINFO_PEELING_FOR_GAPS (loop_vinfo));
eiters -= prolog_peeling;
eiters
= eiters % lowest_vf + LOOP_VINFO_PEELING_FOR_GAPS (loop_vinfo);
while (!vect_update_epilogue_niters (epilogue_vinfo, eiters))
{
delete epilogue_vinfo;
epilogue_vinfo = NULL;
if (loop_vinfo->epilogue_vinfos.length () == 0)
{
vect_epilogues = false;
break;
}
epilogue_vinfo = loop_vinfo->epilogue_vinfos[0];
loop_vinfo->epilogue_vinfos.ordered_remove (0);
}
vect_epilogues_updated_niters = true;
So for example for the loop
void foo (int * __restrict a, int *b)
{
for (int i = 0; i < 20; ++i)
a[i] = b[i] + 42;
}
we end up with no vectorized epilogue when using AVX512 but instead of the
AVX2 epilogue which is discarded we'd like to use a SSE2 epilogue. It
seems that vect_determine_partial_vectors_and_peeling as called from
vect_update_epilogue_niters should have been already determined when
analyzing the epilogue, but during the epilogue costing the loop_vinfo
still inherits the main loop NITER.
For the testcase at hand we're somewhat saved by BB vectorization but when
doing partial loop vectorization we unnecessarily get a AVX512 masked
epilogue here and the cost model doesn't get a chance to see the updated
known niter for the epilogue nor would there be a meaningful way to
do this when costs are compared because we have no way of estimating the
number of masked out lanes for example.
