On Wed, May 09, 2018 at 10:40:26AM -0400, Jason Merrill wrote:
> On Wed, May 9, 2018 at 4:55 AM, Jakub Jelinek <ja...@redhat.com> wrote:
> > On Tue, May 08, 2018 at 11:28:18PM -0400, Jason Merrill wrote:
> >> Maybe add a type parameter that defaults to size_type_node...
> >>
> >> > + ret = fold_convert_loc (loc, TREE_TYPE (expr),
> >> > + fold_offsetof_1 (TREE_TYPE (expr), op0));
> >>
> >> ...and then this can be
> >>
> >> fold_offsetof (op0, TREE_TYPE (exp0))
> >
> > Like this then?
> >
> > + ret = fold_convert_loc (loc, TREE_TYPE (expr),
> > + fold_offsetof (op0, TREE_TYPE (expr)));
>
> I was thinking that we then wouldn't need the fold_convert at the call
> sites anymore, either.
The patch only converts to non-pointer types, I'm not sure if it is
desirable to do the same with pointer types (and most of the other callers
don't use convert, but fold_convert which is significantly different, the
former is emitting diagnostics, the latter is just an conversion +
optimization).
If it is ok to use the final pointer type rather than the initial one, but
convert is not ok, then it would be something like:
if (!POINTER_TYPE_P (type))
return convert (type, TREE_OPERAND (expr, 0));
else
return fold_convert (type, TREE_OPERAND (expr, 0));
on the innermost constant and then indeed no conversion would be needed.
Jakub
