https://gcc.gnu.org/bugzilla/show_bug.cgi?id=98133
Richard Biener <rguenth at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
Status|ASSIGNED |NEW
CC| |joel.hutton at arm dot com
Assignee|rguenth at gcc dot gnu.org |unassigned at gcc dot
gnu.org
--- Comment #2 from Richard Biener <rguenth at gcc dot gnu.org> ---
#5 0x00000000019c9eeb in vectorizable_conversion (vinfo=0x3985f80,
stmt_info=0x39c0f70, gsi=0x0, vec_stmt=0x0, slp_node=0x0,
cost_vec=0x7fffffffc4f8) at ../../src/trunk/gcc/tree-vect-stmts.c:4686
4686 ncopies = vect_get_num_copies (loop_vinfo, vectype_in);
(gdb) p debug_generic_expr (vectype_in)
vector(8) short int
$1 = void
(gdb) p loop_vinfo->vectorization_factor
$2 = {<poly_int_pod<2, unsigned long>> = {coeffs = {4, 0}}, <No data fields>}
that can't work.
(gdb) p debug_gimple_stmt (stmt_info->stmt)
patt_10 = (a_15(D)) <<< ??? >>> (-30574);
the TREE_CODE of this is WIDEN_PLUS_EXPR (we seem to miss an entry for
tree_code_name and pretty-printing for this ... so much for using new
tree-codes).
vectorizable_conversion has
code = gimple_assign_rhs_code (stmt);
if (!CONVERT_EXPR_CODE_P (code)
&& code != FIX_TRUNC_EXPR
&& code != FLOAT_EXPR
&& code != WIDEN_PLUS_EXPR
&& code != WIDEN_MINUS_EXPR
&& code != WIDEN_MULT_EXPR
&& code != WIDEN_LSHIFT_EXPR)
return false;
so it's supposed to handle this case. I think the special-case is that
the whole stmt is invariant and thus not participating in determining
the vectorization factor (all ops are invariant). So we're running into
/* If op0 is an external or constant def, infer the vector type
from the scalar type. */
if (!vectype_in)
vectype_in = get_vectype_for_scalar_type (vinfo, rhs_type, slp_node);
now we should always be able to use vectype_out for determining ncopies,
we just have to correctly interpret it for WIDEN/NARROW or adjust it.
Leaving to arm folks to resolve.
