https://gcc.gnu.org/bugzilla/show_bug.cgi?id=85934
Richard Biener <rguenth at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
Priority|P3 |P2
Status|NEW |ASSIGNED
Assignee|unassigned at gcc dot gnu.org |rguenth at gcc dot
gnu.org
--- Comment #2 from Richard Biener <rguenth at gcc dot gnu.org> ---
The issue is that vector lowering does anything to the vectorized code, that
is, a disconnect between what the vectorizer thinks the target can do and what
vector lowering thinks. And then some bug in vector lowering generating sth
"stupid"
for mask vector lowerings. Ah, it runs into
/* If the vector operation is operating on all same vector elements
implement it with a scalar operation and a splat if the target
supports the scalar operation. */
tree srhs1, srhs2 = NULL_TREE;
if ((srhs1 = ssa_uniform_vector_p (rhs1)) != NULL_TREE
&& (rhs2 == NULL_TREE
|| (! VECTOR_TYPE_P (TREE_TYPE (rhs2))
&& (srhs2 = rhs2))
|| (srhs2 = ssa_uniform_vector_p (rhs2)) != NULL_TREE)
/* As we query direct optabs restrict to non-convert operations. */
&& TYPE_MODE (TREE_TYPE (type)) == TYPE_MODE (TREE_TYPE (srhs1)))
{
op = optab_for_tree_code (code, TREE_TYPE (type), optab_scalar);
if (op >= FIRST_NORM_OPTAB && op <= LAST_NORM_OPTAB
&& optab_handler (op, TYPE_MODE (TREE_TYPE (type))) !=
CODE_FOR_nothing)
{
for vector boolean operations to optimize ~{_88,_88...} to {~_88,~_88...}. And
then it ends up doing this to
mask_patt_63.10_95 = VEC_PACK_TRUNC_EXPR <mask__43.8_92, mask__43.8_93>;
but of course this trick doesn't work for that and the optab query is
"confused"
here because the vector mode is QImode and the component mode is QImode as
well.
The easiest fix is to hoist the VECTOR_BOOLEAN_TYPE_P check above this
transform.
