https://gcc.gnu.org/bugzilla/show_bug.cgi?id=71310
Richard Biener <rguenth at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
Component|target |middle-end
--- Comment #5 from Richard Biener <rguenth at gcc dot gnu.org> ---
(In reply to Segher Boessenkool from comment #4)
> This happens because of
>
> /* Find the smallest nice mode to use. */
> for (mode = GET_CLASS_NARROWEST_MODE (MODE_INT); mode != VOIDmode;
> mode = GET_MODE_WIDER_MODE (mode))
> if (GET_MODE_BITSIZE (mode) >= bitsize)
> break;
>
> in stor-layout.c:finish_bitfield_representative .
This is what the C++ memory model requires. Thus the question is why
the store to skb->csum_valid does not honor it. Note the BIT_FIELD_REF
is probably due to fold-const.c:optimize_bit_field_compare (IMHO a premature
optimization) that in turn just uses get_best_mode.
On x86_64 we seem to use byte accesses, the BIT_FIELD_REPRESENTATIVE
is SImode (the bitfield has 24 bits, the representative can be max. DImode).
So it seems to me that optimize_bit_field_compare is to blame (not matching
what RTL expansion would do). I suppose instead of passing word_mode in
/* See if we can find a mode to refer to this field. We should be able to,
but fail if we can't. */
nmode = get_best_mode (lbitsize, lbitpos, 0, 0,
const_p ? TYPE_ALIGN (TREE_TYPE (linner))
: MIN (TYPE_ALIGN (TREE_TYPE (linner)),
TYPE_ALIGN (TREE_TYPE (rinner))),
word_mode, false);
it needs to pass DECL_MODE (DECL_BIT_FIELD_REPRESENTATIVE (field)).