https://gcc.gnu.org/bugzilla/show_bug.cgi?id=121714
--- Comment #2 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
Or maybe we should try vector of QI mode for sizes larger than word_mode.
Like what is done for op_by_pieces_d.
```
/* Use QI vector only if size is wider than a WORD. */
if (can_use_qi_vectors (op))
{
machine_mode mode;
fixed_size_mode candidate;
FOR_EACH_MODE_IN_CLASS (mode, MODE_VECTOR_INT)
if (is_a<fixed_size_mode> (mode, &candidate)
&& GET_MODE_SIZE (candidate) > UNITS_PER_WORD
&& GET_MODE_INNER (candidate) == QImode)
{
if (GET_MODE_SIZE (candidate) >= size)
break;
if (by_pieces_mode_supported_p (candidate, op))
result = candidate;
}
if (result != NARROWEST_INT_MODE)
return result;
}
```
can_use_qi_vectors returns true for COMPARE_BY_PIECES.
And by_pieces_mode_supported_p is defined like:
```
static bool
by_pieces_mode_supported_p (fixed_size_mode mode, by_pieces_operation op)
{
if (optab_handler (mov_optab, mode) == CODE_FOR_nothing)
return false;
...
if (op == COMPARE_BY_PIECES
&& !can_compare_p (EQ, mode, ccp_jump))
return false;
```
Though for gimple level we should use:
if (expand_vec_cmp_expr_p (TREE_TYPE (op0), type, code))
return NULL_TREE;
Or:
get_vec_cmp_eq_icode (mode, TYPE_MODE (boolean_type_node)) != CODE_FOR_nothing
