https://gcc.gnu.org/bugzilla/show_bug.cgi?id=102281
Richard Biener <rguenth at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
Ever confirmed|0 |1
Last reconfirmed| |2021-09-13
Status|UNCONFIRMED |NEW
CC| |jakub at gcc dot gnu.org,
| |qing.zhao at oracle dot com,
| |rguenth at gcc dot gnu.org
--- Comment #1 from Richard Biener <rguenth at gcc dot gnu.org> ---
Confirmed, works with the C frontend. The issue is that we do
_mm_set_epi64x___q0.0_1 = _mm_set_epi64x___q0;
D.2371 = {_mm_set_epi64x___q0.0_1};
__builtin_clear_padding (&D.2371, 0B, 1);
return D.2371;
so __builtin_clear_padding marks the object addressable which in turn makes
it no longer a is_gimple_val and thus it cannot be returned directly.
I wonder if we can have a more friendly __builtin_clear_padding as
internal-function doing sth like
val = .IFN_CLEAR_PADDING (val, 0B, 1);
thus embedding a RMW cycle on an actual value?
