https://gcc.gnu.org/bugzilla/show_bug.cgi?id=85496
Jakub Jelinek <jakub at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
CC| |ebotcazou at gcc dot gnu.org,
| |jakub at gcc dot gnu.org
--- Comment #2 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
template <typename> class complex;
template <typename T> complex<T> operator* (complex<T>, complex<T>);
template <> struct complex<float> { _Complex float f; };
struct A { complex<float> g, h; complex<float> &foo () { return h; } };
void
bar (complex<float> a, complex<float> c)
{
A b;
b.foo () = c * a;
}
expand_normal returns an expression of the right mode:
(parallel:SC [
(expr_list:REG_DEP_TRUE (reg:DI 104)
(const_int 0 [0]))
])
but we do:
6987 /* Handle calls that return values in multiple non-contiguous
locations.
6988 The Irix 6 ABI has examples of this. */
6989 if (GET_CODE (temp) == PARALLEL)
6990 {
6991 HOST_WIDE_INT size = int_size_in_bytes (TREE_TYPE (exp));
6992 scalar_int_mode temp_mode
6993 = smallest_int_mode_for_size (size * BITS_PER_UNIT);
6994 rtx temp_target = gen_reg_rtx (temp_mode);
6995 emit_group_store (temp_target, temp, TREE_TYPE (exp), size);
6996 temp = temp_target;
6997 }
and starting from that temp has a wrong mode (DImode) rather than the desired
one (SCmode). I must say I really don't understanding this change, before that
temp_mode was GET_MODE (temp) except when it was BLKmode or VOIDmode.
Eric, can you please have a look?
