https://gcc.gnu.org/bugzilla/show_bug.cgi?id=122776
Bug ID: 122776
Summary: vectorizable_simd_clone_call looks at
LOOP_VINFO_FULLY_MASKED_P during analysis
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: ---
if (masked_call_offset == 0
&& n->simdclone->inbranch
&& n->simdclone->nargs > nargs)
{
gcc_assert (n->simdclone->args[n->simdclone->nargs - 1].arg_type ==
SIMD_CLONE_ARG_TYPE_MASK);
/* Penalize using a masked SIMD clone in a non-masked loop, that is
not in a branch, as we'd have to construct an all-true mask. */
if (!loop_vinfo || !LOOP_VINFO_FULLY_MASKED_P (loop_vinfo))
this_badness += 64;
but LOOP_VINFO_FULLY_MASKED_P is always false until we call
vect_determine_partial_vectors_and_peeling at the very end of analysis.
This should result in always selecting a not in-branch variant of a simdclone
and thus disabled loop masking.
All LOOP_VINFO_{CAN,MUST}_USE_PARTIAL_VECTORS_P are transient during analysis,
so we probably need to discover both a "best" simdclone to use for the masked
case and a "best" one for the not masked case.
