https://gcc.gnu.org/bugzilla/show_bug.cgi?id=121349
Bug ID: 121349
Summary: scalar_cond_masked_set optimization fails to consider
all SLP lanes
Product: gcc
Version: 16.0
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: ---
We record a scalar operand in vect_record_loop_mask:
void
vect_record_loop_mask (loop_vec_info loop_vinfo, vec_loop_masks *masks,
unsigned int nvectors, tree vectype, tree scalar_mask)
{
gcc_assert (nvectors != 0);
if (scalar_mask)
{
scalar_cond_masked_key cond (scalar_mask, nvectors);
loop_vinfo->scalar_cond_masked_set.add (cond);
but with SLP and more than one lanes (and a non-uniform lane config) this
fails to properly key on the appropriate mask. Likewise the use in
vectorizable_opertation only queries for lane zero (or rather the def of
the SLP node representative):
if (loop_vinfo->scalar_cond_masked_set.contains ({ op0, vec_num
}))
{
We probably want to key off the SLP node itself? But there's users of the
same type in scalar tree-if-conv.cc
It might be possible to construct a wrong-code testcase from the above?
