On Fri, Jun 27, 2025 at 02:00:26PM +0200, Richard Biener wrote:
> > + gcc_assert (RECORD_OR_UNION_TYPE_P (container));
> > +
> > + for (tree t = TYPE_FIELDS (container); t; t = DECL_CHAIN (t))
> > + {
> > + if (TREE_CODE (t) != FIELD_DECL)
> > + continue;
> > +
> > + tree byte_offset = DECL_FIELD_OFFSET (t);
> > + if (TREE_CODE (byte_offset) != INTEGER_CST
> > + || tree_int_cst_lt (offset, byte_offset))
> > + return false;
> > +
> > + tree bit_offset = size_int (tree_to_uhwi (DECL_FIELD_BIT_OFFSET (t))
> > + / BITS_PER_UNIT);
>
> That math seems to be wrong for actual bit offsets. I’d suggest to skip the
> field if it isn’t a multiple of BITS_PER_UNIT
>
> > + byte_offset = size_binop (PLUS_EXPR, byte_offset, bit_offset);
>
> it would be nice to avoid tree arithmetic here as well by using wide_int.
Or just use byte_position? That doesn't avoid tree arith, on the other side
is used everywhere.
Jakub
